Hi friends 👋,

Stocks are worth $0 now (I think, haven’t checked) and the string of perfect weather in NYC has ended, but the optimism streak continues! The world’s researchers and engineers are still cooking up the good stuff, and we’re here to serve it to you.

Let’s get to it.

This Week’s Weekly Dose is brought to you by… Runway

Runway is building foundational AI for how the world creates. 65M+ people use Runway to generate, edit and produce video, images and audio with 80+ top media models. You can read my essay with co-founder Anastasis Germanidis to learn more on where Runway is headed beyond video.

Nvidia, Amazon, Lionsgate, Chime, and 95%+ of the Fortune 100 are part of a growing community of teams, founders, and independent creators using Runway to make national TV spots, social ads at scale, launch videos and more.

You can also use Runway in ChatGPT, Claude, Grok, Cursor, Replit and more via the Runway MCP.

Create any idea with Runway now

(1) FDA Approves Voceryq, the First Plasmid DNA Immunotherapy

FDA / INOVIO, video by Runway

Last Friday, just a few hours after I hit send on the Dose, the FDA granted accelerated approval to INOVIO’s Voceryq, the first plasmid DNA-based immunotherapy approved in the US.

Voceryq will be used to treat recurrent respiratory papillomatosis, a disease I’d never heard of but sounds awful. Caused by HPV types 6 and 11, RPP gives victims wart-like growths in the airway, usually around the vocal cords, that make their voices hoarse, make them cough, and in severe cases, make it hard to breathe.

Surgeons can remove the growths, but the virus remains in the tissue, so the throat growths keep coming back, and patients end up in the operating room again and again for years. In the trial, 26 adults had needed at least three surgeries in the year before treatment.

As shown in the video above, Voceryq delivers two plasmids (small loops of DNA originally borrowed from bacteria) into muscle cells. One carries instructions for HPV-6 and HPV-11 proteins, which teaches the immune system to go after infected cells, and the other carries IL-12, a molecule that turns the immune response up. A device fires brief electrical pulses that open temporary pores in cell membranes so the DNA can get inside without requiring a virus for delivery.

After treatment, 7 of the 26 patients in the trial didn’t need surgery for the entire following year. It’s a small open-label trial, and INOVIO still has to confirm the results in a follow-up study, but it’s on the path now, which thank god, because this was the last shot. The company was set to run out of cash in Q1 2027, but its stock tripled in after hours after posting the results.

This isn’t the first treatment for RPP - Precigen’s Papzimeos, which uses a modified virus to target the same HPV types, won full approval in August 2025.

RPP patients have gone from no treatments beside surgery to two, which is amazing, but the platform will likely prove to be even more important.

If you can get a plasmid into cells and get the immune system to act on it, you can swap in instructions for other HPV diseases, cancers, and infections, which is what INOVIO is planning to do next.

DN-AYYYYYY!

(2) Zipline Raises $1.5B to Build Nationwide Drone Network

There are very few companies out there as cool as Zipline. After starting by serving hospitals and health facilities across Africa, where they were allowed to fly drones and the need was most acute, they came to America last year to do deliveries, starting in Texas.

CEO Keller Clifton said that over the past 18 months, they’ve grown flight volumes 18x and are now doing more that a million deliveries per year per metro.

Now, they’re going to be able to do a lot more. The company raised $1.5 billion at a $21.5 billion valuation in a round led by Paradigm, WCM, and Layer.

They’re using the money to bring drone deliveries nationwide. To that end, they signed “the largest autonomous vehicle contract in human history to expand automated delivery nationally with Walmart,” with whom they’ll scale to 300 supercenters throughout these United States.

This is awesome news. Zipline’s drones are quiet, safe (they’ve done over 150M commercial autonomous miles with no safety incidents), fast, affordable, and don’t require a human chaperone. As I wrote in my Deep Dive on Airbound, I’m incredibly optimistic that a good deal of movement will move to the skies, and this is an important step in that direction.

Since the Eagles stink and I can’t use this on them much, I’ll use it here:

Fly Zipline Fly.

(3) DREADDs Work in Humans

In April, Owl Posting wrote an essay on “new knobs of control,” including DREADDS, which he explains like this:

You use a viral vector to deliver the DREADD gene to whatever cells you want to control. Once the DREADD is expressed on the cell surface, it just sits there, inert. Then you administer CNO, which binds exclusively to the DREADDs. When it binds, the receptor activates its coupled G-protein just like any normal GPCR would, doing whatever you engineered the GPCR to do. And, as far as anyone can tell, there are no off-target effects.

This is unambiguously a new knob. The receptor didn’t exist in your body before. Its binding partner (CNO) doesn’t exist endogenously. As a result, you, and you alone, get to decide when and where it is turned on.

Well, now we have them. In the first adult human study, per a Grok explainer in the replies, “doctors placed inhibitory DREADDs in the hippocampus of 3 people with medication-resistant epilepsy. Taking the drug quieted the seizure-causing cells, cutting seizures by ~93% with only mild side effects. It lets doctors precisely dial down overactive brain circuits on demand.”

That is wild. We have a knob that can turn down epilepsy.

(4) NOM4D Program Demonstrates Automated In-Space Assembly

DARPA

If we’re going to become a spacefaring civilization, we’re going to need to build bigger structures in space than can fit in a rocket. Today, there are two options for big space things: we can fold something up like origami and unfold it in orbit, which is how the James Webb Space Telescope works, or take the ISS approach, where you launch big pieces over many missions and put them together by hand.

Until now!

DARPA’s NOM4D program (Novel Orbital and Moon Manufacturing, Materials and Mass-efficient Design, great name) is working on a third option. They can send up a box of parts and assemble them in an automated way. On Thursday, DARPA announced that a Caltech payload had done just that, in what the agency calls the first-ever demonstration of automated in-space assembly. Hopefully, it’s the first of many many.

In this case, a simple gantry robot, which moves along three straight axes, joined composite truss pieces with custom connectors one section at a time while unrolling a thin reflective mesh. It is a proof of concept for future space antennas.

Program manager Andrew Detor compared the system to Tinkertoys or K’nex. It’s meant to be super simple, without fancy robotics or advanced algorithms. NOM4D is aiming at structures hundreds of meters to kilometers across, like space depots, habitats, huge solar arrays, and radio antennas large enough to listen all the way to the edges of the universe.

And speaking of making things in space, not boring capital portfolio company Varda is currently making drugs in space!

(5) Telepathy Tapes Season 3: The Reincarnation Records

A couple of years ago, after listening to Season 1 of The Telepathy Tapes, I wrote The Return if Magic, and then I kept going down the rabbit hole, which led to pieces like Means & Meaning and Riding the Leopard.

Season 3 came out last month, and it’s on reincarnation. Reincarnation is a compelling bridge to believing that the world is more magical than we appreciate. For one thing, an awful lot of the world’s people believe in reincarnation as part of their religion, but for another, it’s a surprisingly well-researched topic. Since psychiatrist Ian Stevenson created the Division of Perceptual Studies (DOPS) at the University of Virginia in the 1960s, researchers there have documented and coded more than 2,500 cases worldwide of young children who spontaneously report detailed memories of previous lives.

I don’t want to spoil the series, but I do think you should give it a listen and consider what it might mean if it’s true.

With all the Means stories we cover here in the Dose, mix in a little Meaning.

And while we’re add it, this popped up in my feed yesterday…

Extra Doses: Science Breakthroughs, METR on Anthropic, CandyBench, Ramtin Naimi on ILTB, Northwood

Read more

Hi friends 👋,

Happy Friday and welcome back to our 217th Weekly Dose of Optimism.

And what a week it’s been! Per New York Metro Weather, we’re in the midst of the longest streak of 8/10 or higher days since the account started in 2021, and here at the Dose, we’re in the middle of a string of 9/10 or higher weeks so long that I’ve lost count.

Have a great Halloween tomorrow. For now, all treats no tricks.

Let’s get to it.

This Week’s Weekly Dose is Brought to You By… Reducto

The Document Editor Built for AI Products

AI products need more than file generation. They need a fast, accurate document editor that agents can also use.

Reducto Editor brings full-fidelity DOCX, PPTX, and XLSX editing into your product through an SDK and embeddable UI. Lightweight and browser-based, it gives agents the tools to revise contracts, update presentations, and edit spreadsheets while preserving formatting and all the functionality you’re used to.

Users collaborate in the same file, review native Word tracked changes and comments, and make their own edits without leaving your application. Files retain their full formatting when opened in Microsoft Office.

From the team behind the document infrastructure trusted by Harvey, Scale AI, Toast, and Vanta, Editor runs in the browser or on your own servers.

Build with Reducto Editor →

(1) Anduril Unveils American-Made BlueChip Computer

Another week, another extension of Anduril’s capabilities1.

Last week, we wrote about Warhound, Ben Lamm’s effort to (humanely) genetically engineer dogs to be used in police and military work. Anduril is providing “software and hardware enhancement for items such as dog protection vests and eye goggles.”

This week, Palmer Luckey announced that Anduril would be making its very own Made in America computer, BlueChip: designed here, with components from here, with metals America can mine, refine, process, and mill at scale and at reasonable price. American batteries, too.

This is cool because as Palmer says in the video, “If you don’t control your technology, you can’t control your future.” But it’s also cool because it seems as if Anduril has gotten to the point where it’s built a machine that can make machines. Assemble enough engineers, capital, manufacturing capacity, distribution, etc… and you can build almost anything. It’s very Elon Companies-esque.

And as we wrote in The Electric Slide, making products that can sell a ton of units, like laptops, can create the pull to bring more of the supply chain here, too.

(2) OpenAI Says Its Models Disproved the ETH (and more)

Video created with Runway

The race to solve everything solvable continues apace.

This week, OpenAI announced “three new discoveries in mathematics, computer science, and physics about harmonic analysis, complexity theory, and quantum field theory.”

In math, OpenAI’s models produced a proof of Christopher Sogge’s 1991 local smoothing conjecture in every dimension, which would also settle the Kakeya conjecture and Stein’s restriction conjecture (of course). For context, Hong Wang and Joshua Zahl proved Kakeya in three dimensions just last year, and it was one of the biggest math stories of 2025.

In computer science, a model refuted the Exponential Time Hypothesis (ETH), the widely held belief that problems like 3-SAT can’t be solved meaningfully faster than brute force. If that makes no sense to you, check out the explainer video I made with Runway, which breaks it down in claymation well enough that I could actually understand it.

In physics, the models produced a rigorous proof that interacting quantum field theories, including QCD and pure Yang–Mills, exist in four dimensions. That’s half of a Millennium Prize problem.

Whereas before OpenAI has just dropped a bunch of solutions on the internet’s head and walked away, leading to quasi-embarrassing situations like when they said GPT-5 had solved ten open Erdős problems but it turned out the model had mostly found solutions that were already sitting in the literature, this time, OpenAI is sharing the outputs as research data with experts they call “shepherds” in each field, who will choose what to do with the data and where/when to publish.

Shepherds don’t have to sign NDAs, don’t get paid, and do get the credit for explaining the work. Russell Impagliazzo and Ramamohan Paturi, who proposed ETH, are leading the review of its disproof with MIT’s Ryan Williams. Nobel laureate David Gross is leading the physics group, which chose to withhold the AI material and write it up themselves. The math proof and the ETH disproof are on GitHub with Lean formalizations. Lean, a checking software, guarantees the proof proves the formal statement, but human shepherds still have to confirm that the formal statement is the right one.

Big week for Lean, by the way. On Tuesday, Axiom announced a complete Lean auto-formalization of the Classification of Finite Simple Groups, the “periodic table of symmetry.” The proof runs about 15,000 pages and was written by more than 100 mathematicians over several decades. It’s so big that no single person has ever checked all of it. For a sense of scale, it took Georges Gonthier’s team six years to formalize just one piece of the classification, the Feit–Thompson theorem, back in 2012. But Axiom’s machines can do a lot of human-years quickly, and they’re going to need to to keep up with all of the math and physics coming out of the labs.

Because not only did OpenAI announce those results this week, a physics neolab, Physical Superintelligence (which we covered in Dose #209) released results on five problems from IQOQI Vienna’s list of open quantum problems, each with a paper, a Lean formalization, and open source verification code on GitHub.

For those of you who were planning to spend the weekend cracking OQP 12: Can vacuum correlations violate the CHSH Bell inequality at arbitrarily large separations? or OQP 12: Can vacuum correlations violate the CHSH Bell inequality at arbitrarily large separations?, I’m sorry. Machines did it. You’re going to have to find a new hobby.

Also, if you were trying to use the entanglement between two empty regions of space no matter how far apart they are to violate the Bell Theorem, I’m sorry for you too. The machine found that you can’t do it past some finite distance.

Anyway, it seems like this is just how we do math and physics now.

(2a) Anthropic Publishes Progress on Accelerating Drug Development

Apparently, it’s how we do drug development now too, if Anthropic succeeds.

AI for drug discovery is hot, but most of the focus is on discovering new molecules, which creates bottlenecks further downstream in actually developing these drugs and getting them to market. To understand and fix them, Anthropic believes, it needs to get inside the programs themselves, so for the past several months, its life sciences team has partnered with six organizations to start running programs of its own.

At the discovery end, Divergence Labs announced positive results in the fight against food allergies. Allergic reactions are triggered by IgE, which sits on your immune cells as well as floating in your blood. Standard drugs mostly neutralize the floating IgE, so Divergence’s approach aims to pry it off the cells instead. Claude turned that goal into a design plan and, largely on its own, generated about 35,000 antibody designs in a month. Three of the early hits dislodged IgE better than the standard medication, and people seem psyched that this represents meaningful progress against food allergies.

Another biotech, Generate Biomedicines, used Claude Code to redesign part of CAR-T, the therapy that uses a body’s own immune cells to hunt and kill cancer, so that it could eventually work on solid tumors. One of Claude’s designs was about 70% shorter than the standard version and got cells to make roughly 10 times as many cancer-spotting receptors.

At BigHat Biosciences, Claude proposed improvements to two antibody fragments, and BigHat’s automated lab made and tested about 1,010 variants in 10 weeks. The best ones bound more than 10x more tightly. Along the way, Claude and an Anthropic researcher figured out that the standard stability metrics were misleading and proposed better ones.

But the point is that these improvements are still hard to get to patients, so Anthropic turned to paperwork with Enveda, whose manufacturing scientists spend up to 75% of their time on it. Claude reviewed 1,061 pages of manufacturing records, caught all of the issues human reviewers had found (plus seven they’d missed), and cut review time by 89%. It also drafted a regulatory document in four hours that takes about three months by hand.

At the trial stage, Bristol Myers Squibb is using Claude to assess plans for an early-stage drug and estimate its odds of success. The Column Group, a biotech investor, is launching Pleiades, a company built to run many more drug programs with Claude than a small team normally could. The theory is that if discovery gets cheaper and faster, investors can back riskier ideas.

Anthropic is also running programs itself. In July, it acquired Mithril Medicine, a one-person company whose founder, Cody Tranbarger (aka Professor Oak), used Claude to run experiments across 11 outside labs. In under 18 months, he took a treatment for a rare muscular dystrophy from idea to promising results in mice. In one cohort of 42 kids with the severe form of the disease, only one ever walked on their own. If the treatment holds up, human trials could start as soon as late 2028.

Anthropic is upfront that nearly all of this is preclinical, so most of these drugs haven’t been tested in people yet, but they showed speedups in every part of the process. The tricky thing about complex processes like drug development is that you have to improve the whole thing at once, lest any one thing that didn’t improve slows the whole thing down.

We might finally have a machine capable of doing just that. Bad time to be a disease. Good time to be a human.

(2b) Robot + Agent Perform Brain Surgery on Mouse

I just saw this one this morning, so didn’t have time to write it up, but that’s ok, just read the headline and the tweet thread and realize that AI AGENTS PERFORMED BRAIN SURGERY ON A MOUSE.

(3) Readwise Launches Readwise Books

I’m an investor in Readwise, friends with its founders Dan Doyon and Tristan Homsi, and I love reading, so I’ve been lucky enough to get to be a part of the beta for what they were calling Bookwise and are now calling Readwise Books.

I love this app. It’s my second-most-used app on my phone, after X (need to break that habit again), and that’s not counting the time that I spend reading through Readwise Books on my Daylight Computer. The description people give most often is Kindle x Goodreads, which is part of it, but there’s also a little Strava mixed in there. There’s a leaderboard, which has definitely pushed me to read more, like the week I ended up #1 on the Readerboard in July.

The reading experience itself is familiar to those used to reading on the Kindle App, but it’s just a little bit better in places. There’s AI natively built in, which I don’t use normally, but was a lifesaver when I was reading Jean Beaudrillard’s Simulacra and Simulation, for example. Highlights go right into Readwise. And its text-to-speech is really good for listening when I get off the subway and start walking.

I am a firm believer that people should Read More Books, and although I’m biased, I’ve also voted with a ton of my time that this is the best product out there for doing that.

So go Read More Books. There’s a waitlist, but if you’re reading this, you can skip it here.

(4) Reta Works on Sleep Apnea, Too 😴

Lilly’s Retatrutide is perhaps the closest thing we have to a miracle drug.

We first covered Reta in Dose #175, as a gray-market “holy grail” in Vectorculture’s piece on Chinese peptides. By May, it was the lead item in Dose #194, after Lilly’s TRIUMPH-1 trial showed people on the top dose losing 28.3% of their body weight over 80 weeks.

Since then, the triple-hormone shot, which mimics GLP-1, GIP, and glucagon, has kept piling up wins. Now, it seems to work on sleep apnea.

Tens of millions of Americans have obstructive sleep apnea (mostly undiagnosed). The standard treatment is strapping a CPAP mask to your face every night. In two 80-week Phase 3 trials, TRIUMPH-1-OSA and TRIUMPH-2-OSA, 691 adults with moderate-to-severe sleep apnea started out stopping or nearly stopping breathing about 57 times an hour. In the trial without diabetes, the 9 mg dose cut that by 34 events an hour, versus 10 on placebo. About half of patients on the higher doses ended up with apnea that was resolved or mild without daytime sleepiness, compared with roughly one in ten on placebo. Along the way, they lost about a fifth of their body weight.

Some people have pointed out that of course, losing weight should help with sleep apnea, but OK sure whatever the mechanism, Reta just keeps fixing stuff:

• Diabetes: In TRANSCEND-T2D-1, presented at ADA in June, 75–83% of patients got their A1C to 6.5% or lower in 40 weeks.

• Heart risk: In TRIUMPH-3 in July, people with severe obesity and existing heart disease lost 22.6% of their body weight, with inflammation (hsCRP) down 51% and systolic blood pressure down 9.3 points.

• Knees: In TRIUMPH-1, knee osteoarthritis pain fell by about three-quarters.

• Liver: In a Phase 2 substudy, 93% of patients on the top dose had normal liver fat after 48 weeks.

The drug hasn’t even hit the market yet (if you don’t count the grey market). There’s still an open question as to whether these improvements also prevent heart attack and stroke, which Lilly’s dedicated cardiorenal trial will attempt to show, and which would be another massive win, but for now, we’ll just have to be happy with a drug that cuts weight, lowers heart disease risk, fights diabetes, cuts liver fat, makes our knees feel better, and replaces a CPAP.

Lilly, which crushed earnings this week and raised its full year revenue outlook by $5 billion, deserves every dollar of its $1.3T market cap.

(5) Neuralink Gives Voice Back to ALS Patients

From miracle drugs to downright miracles. Given the quantity and consistency of mind-blowing breakthroughs we cover here in the Dose, it’s easy to forget how miraculous some of them are, and how miraculous they must feel to those who benefit from them.

This week’s news out of Neuralink is one of those.

ALS, or Lou Gehrig’s Disease, is a fatal, progressive neurodegenerative disorder that destroys the motor neurons in the brain and spinal cord that control voluntary muscle movement. It takes away its victims’ ability to speak even though it leaves their mind intact, meaning they can think clearly but can’t verbalize those thoughts.

Neuralink has been working to fix that with VOICE. This week, it released early numbers from one participant who hit 75 words per minute with a 5.8% error rate across a 125,000 word vocabulary in their first week with the implant. Before VOICE, the best published result on the same vocabulary size, Stanford’s 2023 system, managed 62 words per minute with nearly a quarter of words wrong.

Natural conversation runs around 160 words per minute, so we’re about halfway there. Importantly, the system can also speak in the participant’s pre-ALS voice.

VOICE launched at UT Southwestern last October with FDA Breakthrough Device designation and plans to enroll up to six people with ALS, spinal cord injury, or stroke. That’s not a big group, and the one person in this study is just one-sixth of that. But every person who’s able to get their voice back through a BCI is a win, and the technology will only get better, faster, more accurate, and less invasive from here.

In related news, Elon Musk announced yesterday that its second product, Blindsight, which will restore sight to the blind, is ready engineering-wise.

Age of Miracles, pending regulatory approval.

EXTRA DOSES: The Commanding Heights, Science Breakthroughs, DoorDash Acquires Instacart, Thrive Holdings Acquires ADT, LIFE

Read more

Hi friends 👋,

Happy Friday and welcome to the 216th Weekly Dose of Optimism!

It looks like the hacks this week broke a lot of the internet, but over here in our little corner of it, things are looking good. We have everything from cancer to plant communications to genetically engineered war dogs to basically alchemy. Once again, the week was so jam packed that we have more in the Extra Dose (7 Extras!) than we do up top.

Let’s get to it.

This week’s Weekly Dose is brought to you by… Reducto

The Only Document Parsing Product You’ll Need

Reducto is the document ingestion layer trusted by AI teams at Scale AI, Harvey, Toast, and Vanta to reliably turn complex files into structured, grounded data for production AI systems at scale. Processing more than a billion pages per month, Reducto has built its reputation on the long tail of difficult documents that break conventional OCR: dense tables, unusual layouts, handwriting, low-quality scans, strikethroughs, and formatting-dependent information.

Its new frontier parsing model, r-1, combines layout detection, reading order, tables, formatting, grounding, and granular citations in a single request. It reduces parsing errors by up to 20%, improves latency at high volumes, and costs 1¢ per page all-in.

Already using another parser? Apply for up to $5,000 in migration credits and hands-on support to benchmark r-1 against your current stack on the documents that matter most.

Apply at Reducto.ai/migrate

(1) FDA Approves Rectal Cancer Immunotherapy (+ more)

U.S. Food and Drug Administration, Jeff Tang, Coherence Neuro; video created with Runway

We hate cancer here at the Dose, and we’re not afraid to say it. It’s probably why not a single cancer reads the Dose. We’re willing to sacrifice our numbers for our belief.

Consequently, we love when humans come up with new ways to kill cancer, and they’ve done it once again (and again, and again) this week.

In June 2022, in a small trial at Memorial Sloan Kettering, twelve patients with locally advanced rectal cancer took six months of an immunotherapy called dostarlimab. In all twelve, the cancer disappeared without chemo, radiation, or surgery.

On Monday, the FDA granted dostarlimab (GSK's Jemperli) accelerated approval as a standalone first treatment for stage II and III rectal cancers that are mismatch repair deficient (dMMR). These tumors basically have a broken DNA spell-checker, so mutations pile up until the cancer looks foreign to the immune system. Dostarlimab blocks PD-1, the off switch tumors use to keep T cells at bay, which lets the immune system clear the cancer (explained in the video above).

In GSK’s 154-patient trial, after a median of 26 months, cancer hadn’t progressed in 71% of patients treated with dostarlimab alone. In MSK’s expanded 50-patient study, every patient responded, and 93% stayed cancer-free for at least a year.

Only about 3% of locally advanced rectal cancers are dMMR, per the FDA, so this isn’t a cure for all rectal cancers, let alone all cancer, but we’re getting into a pretty consistent, almost weekly, pattern of new therapies working against specific cancers, which seems like the way we’ll beat cancer overall.

In some cases, individuals are even figuring out how to cure their own, n-of-1 specific cancer.

In #177 we led with Elliot’s story on one of those people, GitLab founder Sid Sijbrandij, going founder mode on his osteosarcoma after running out of standard options. He’s one of six people Jeff Tang included in a list of those who have cured or are curing their own cancer, or a loved one’s.

But wait, there’s more!

On Wednesday, Coherence Neuro announced that it had raised a $45M Series A from Founders Fund and OpenAI “to build cancer therapy that learns.”

The idea is to turn neurotech and BCIs at cancer to read the electrical signals surrounding it, understand how it grows over time, and influence the electrical signals to modulate the disease. Human trials are underway in glioblastoma patients, and they’re gearing up for pancreatic cancer and more.

There’s a long way to go on these and so many of the cancer treatments we discuss in the Dose, but the promising results have become a drumbeat, and it really does seem like we’re on the path to knocking out cancer once and for all.

So say it with me, as heartily as you can: get fucked, cancer.

(2) The Nuclear Company Proposes a 6.6 GW KY Nuclear Campus

Alexander C. Kaufman for Heatmap

As much as we hate cancer, that’s how much we at the Dose love nuclear power.

We love small nuclear reactors, and we love big nuclear reactors, and we especially love a lot of either. Big week for us, then.

It hasn’t always been so good for nuclear here in the US. In the roughly 15 years it took the US to finish its first two Westinghouse AP1000 reactors at Vogtle, China built four of them, then they reverse-engineered the design into its own Hualong One, and now they have 10 Hualongs running with at least 37 more on the way. China knows that the way to make nuclear cheap is to make a lot of reactors over and over.

The Nuclear Company, which employs a bunch of people who worked and learned on Vogtle, launched in 2024 to do just that.

On Wednesday, it announced Prosperity One, a proposed campus for up to six AP1000s, about 6.6 GW, on roughly 4,000 acres along the Ohio River in Lewis County, Kentucky.

6.6 GW is an enormous amount of power. Fully built, Prosperity One would be the largest nuclear site in the US and account for over 5% of the country’s total nuclear capacity (we’re at 97GW right now).

TNC’s plan is to build two units at a time, keeping the design consistent on each, so that it can bring down costs on each successive one, a lesson learned the hard way at Vogtle. After a slow, expensive start with Unit 3, Vogtle Unit 4 came in roughly 30% cheaper and faster. TNC Chief Nuclear Officer Joe Klecha said that 30% is the floor expected savings. The company is targeting 2028 NRC approval.

At the other end of the size spectrum, the Navy picked Radiant to install its Kaleidos microreactor at Naval Weapons Station Crane in Indiana. Kaleidos is a factory-built, 1 MW reactor (6,600 times smaller than Prosperity One), that is built centrally and shipped to sites by truck, ship, or plane. When completed (the Navy wants 2028), it will be the Navy’s first shore-based advanced reactor and the first power-producing reactor in Indiana. Radiant, which we wrote a Deep Dive on in 2024, now has partnerships with the Army, Navy, and Air Force.

ELECTRONAISSANCE.

(3) Ben Lamm and Palmer Luckey Launch Warhound

I already had this whole thing written, with a different story in this slot, when the news dropped that Colossal Biosciences’ Ben Lamm and Andruil’s Palmer Luckey were teaming up to launch Warhound with a fresh $200M from investors including In-Q-Tel, Sequoia, and 8VC. The new company is worth $1.2 billion.

This is the advanced technology version of a superhero movie, literally in that we’re talking about genetically engineered war dogs and figuratively in that we’re seeing two characters combining forces within a wider cinematic universe. Per the website, Warhound is “combining licensed technology from Colossal Biosciences and Anduril,” which is just a really interesting way to start a new business.

“Animal welfare is central to that mission,” the company says on its website, “with American Humane certification supporting our commitment to the care and well-being of every dog.”

I don’t think I’m going to be able to say anything Lamm didn’t, so Per TechCrunch:

Lamm believed dogs could be enhanced for health, longevity, and “for things like speed and agility or enhance olfactory sensors for bomb sniffing and whatnot,” he said. His team looked at the top-performing working dog breeds — the Belgian Malinois and German Shepherd — taking the cell lines from them and putting them in a biobank. From there, they will edit the genes, adding the ones that will make the dogs perform better while taking out the ones that won’t, to create “a new type of dog.”

Andruil, meanwhile, will provide software and hardware enhancement for items such as dog protection vests and eye goggles to assist the dogs and their handlers with matters like situational awareness and coordination.

“The idea is, if we make the dog safer, make it healthier, make it perform better, it’s going to perform better in the line of duty,” Lamm said. He added that Anduril’s additions aim to make the dogs safer and healthier outside of their genetic components, “on top of the dog, not in the dog.” He believes such enhancements will also protect the dogs’ human handlers better.

This is not a new structure for Lamm. In late 2022, Lamm co-founded Form Bio, a software company spun out of Colossal, and he later launched Breaking as another spinout. As with Warhound, Colossal will own equity (Luckey won’t), but this time, it’s combining technological forces with another company, Anduril.

They’re good dogs, Brent. 16/10 please thank them for their service.

(4) Doris Tsao’s Lab Finds Signs of Consciousness in Time

Janis Hesse, Doris Tsao, et al. on bioRxiv, video created with Runway

Late last year, in Dose #173, we covered Dr. Doris Tsao’s joining our friends at Astera to lead a new team, Astera Neuro. “The new team’s goal is to solve one of the longest-standing, most fundamental questions: how the brain produces conscious experience, cognition, and intelligent behavior.”

Just ten months later, Tsao and team are out with their first big paper, and it’s a doozy. The video above is one attempt to explain it, and here’s the word version.

If you show your left eye one face and your right eye another, you see just one face. Instead of blending them, one just vanishes from awareness, even though it’s hitting your retina. Neuroscientists use this trick, called binocular flash suppression, to separate what the brain receives from what it’s aware of.

The textbook belief has been that this happens because the suppressed image fades as signals move forward through the brain and never “ignites” into awareness. Last week, a team led by Janis Hesse in Tsao’s lab found a different explanation.

Recording thousands of neurons in the face-processing regions of two macaques’ inferotemporal and prefrontal cortex, they found that the same neurons alternate roughly every 60 milliseconds between representing the face the monkey is aware of and the one it isn’t. When they reconstruct an image from one set of moments, they get the conscious face. But when they reconstructed it from the others, they got the suppressed face in full detail.

So why are we only aware of one? Still need to figure that out.

But Tsao called it the most important discovery of her career, because it means that a decodable signal deep in the brain exists even for things you’re not conscious of, so scientists can now follow both faces and figure out exactly why we become aware of one and not the other, which should give deeper clues to how consciousness works.

Human consciousness research is one of the areas I’m most fascinated by, and I hope that this is the first of many breakthroughs from Astera Neuro and others.

(4a) Electric Plant Company Releases Bombadil to Speak Plant

As we better understand our own human brains, humans are also making progress in understanding the language of… plants.

Electric Plant came out of stealth this week with $11.2 million in funding and a foundation model, Bombadil, that reads electrical signals from inside living plants and translates them to estimate seven kinds of stress, like thirst and nutrient needs. Bombadil was pretrained on more than 4 million hours of recordings from 100 species, from redwoods to tomatoes, captured with sensors that cost tens of dollars. Co-founder Justin Mast, who founded Bloomscape, compares it to one model that reads the brainwaves of blue whales, horseshoe crabs, and fruit flies, given how diverse plant species are.

Practically, in the near-term, the company thinks it can cut farm water use by 20% or more with its models and sensors, which is important and economical, but more wildly…we are learning the language of plants? And it’s bioelectric?

I got to see a preview of this at Onto Ventures’ summit in California a couple weeks ago and it was one of the most mindblowing demoes I’ve seen. Justin is kind of like the Lorax, but he speaks to the trees.

(5) Marathon Fusion Says It Turned Mercury Into Gold

We cover a lot of impressive accomplishments in the Dose, like talking to plants and curing cancer, but it’s not every day that we cover a company that’s succeeded where the inimitable Sir Isaac Newton failed.

That’s what Marathon Fusion has done. Newton, under the pseudonym Jeova Sanctus Unus, spent three decades trying to crack alchemy, as I wrote about in Take Weird Ideas Seriously. I think it’s great that he tried. He just didn’t have the technology.

Marathon Fusion seems to. Co-founded in 2023 by CEO Kyle Schiller and CTO Adam Rutkowski, a former SpaceX propulsion engineer, the company posted that last month it turned mercury into “real, stable gold” at ten times the rate of CERN’s 2025 record. CERN's version produced about 29 picograms across four years of collider runs, so we’re not talking Fort Knox, but the physics is real. Fast neutrons from a deuterium-tritium fusion reaction knock a neutron out of mercury-198, which becomes mercury-197 and decays into gold over a few days.

Gold dropped pretty hard yesterday (unsure if related), but for now, it’s just a demonstration.

In the nearer-term, Marathon will use the technology to make medical isotopes. Neutrons that can turn mercury into gold can also turn zinc into copper-64 and copper-67, used in cancer imaging and targeted radiotherapy. Marathon says it has started making both.

The neutrons can make a lot of things. In August, outside testing lab Covalent certified that Marathon’s plasma centrifuge had enriched lithium-6, which fusion plants need to breed tritium fuel, and this week the company said a new run beat a Soviet record that had stood for 40 years. Next is a fusion reactor in San Francisco designed to fire a trillion neutrons a second into a one-centimeter target, continuously, for years.

It’s a very Fuse-like bet that the neutrons produced in fusion reactions can be very valuable long before Q>whatever, and a Small Steps approach to fusion.

Or as Mike Meyers would say, “I LOVE GOOOOOOOOLD.”

EXTRA DOSES: Science Breakthroughs, a16z Frontier, Formic $50M, Warp Sciences, Chinese Cancer Drugs and the FDA, Gecko Navy, Patrick Hsu on Dialectic

Read more

Hi friends 👋,

Happy Friday and welcome back to our 215th Weekly Dose of Optimism!

Welcome to the flying flipping car era. As bad as things have been for my Philadelphia Eagles, that’s how good they’ve been for humanity.

Let’s get to it.

This week’s Weekly Dose is brought to you by… Reducto

The Only Document Parsing Product You’ll Need

Reducto is the document ingestion layer trusted by AI teams at Scale AI, Harvey, Toast, and Vanta to reliably turn complex files into structured, grounded data for production AI systems at scale. Processing more than a billion pages per month, Reducto has built its reputation on the long tail of difficult documents that break conventional OCR: dense tables, unusual layouts, handwriting, low-quality scans, strikethroughs, and formatting-dependent information.

Its new frontier parsing model, r-1, combines layout detection, reading order, tables, formatting, grounding, and granular citations in a single request. It reduces parsing errors by up to 20%, improves latency at high volumes, and costs 1¢ per page all-in.

Already using another parser? Apply for up to $5,000 in migration credits and hands-on support to benchmark r-1 against your current stack on the documents that matter most.

Apply at Reducto.ai/migrate

(1) Neural Networks Are Already Clean

Nick Cammarata · Goodfire

Good news - AI is probably way less likely to kill us all now because humans rock and one of us just kind of figured out how to learn to read their minds.

It is beautiful that a person I became aware of and started following because of his understanding of his own mind—I discovered Nick Cammarata through his tweets on meditation and the jhanas—seems to be the person who’s figured out how LLMs’ brains work.

In the “The Short Version” section, he writes:

I think I have good news about the nature of neural networks.

After many months studying neurons and circuits in image and language models, I’ve come to believe a network’s own mechanisms are its legible building blocks, and that the outlook for mechanistic interpretability may be far better than I’d feared. The canonical evidence of polysemanticity and superposition that convinced much of the field, including me, that networks pack concepts arbitrarily was misread. What looked like arbitrary packing was mostly a motif I’ll call recursive generalization, which is much less concerning. This means we can study networks directly, instead of first untangling them. Neural networks are already clean.

Basically, he thinks that we can understand what’s going on inside of models’ brains - “interpretability” - which means that alignment is probably solveble.

As he explains, “interpretability might be the most important scientific problem in the world right now, as I don’t think we can solve alignment if we don’t know what we’re building.”

Cammarata used to work on interpretability at OpenAI, and he himself “came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots.” Now, he tweeted:

the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works.

If he’s right, not only did the p(doom) from runaway AI just tank, but it’s also another win for using technology to solve problems that previously seemed unsolvable. Chalk one up for the optimists.

(2) Tesla Introduces Flying Roadster

Thank god we’re going to live, because it seems like the future is going to be as cool as advertised.

Last night, nine years and a number of delays after Tesla first teased the second-gen Roadster, Elon took the wraps off at a SpaceX test site near Waco. It was worth the wait.

The car went 0–60 in about a second, which would have been the coolest thing at any car reveal in history, except this one. Next, using 10 nozzles fed by a 600-liter compressed-air tank, plus some very large fans, it drove up an upside-down ramp, spun around, and set itself down on the pavement. Then it hopped over a Cybercab, Cybertruck, Semi, Model Y, and Model 3 while moving. Tesla said the car would go half the speed of sound (?!?!) at 380 MPH?

Musk said production starts in 2028 (for real) and joked, I think, that they’d try to jump the Snake River Canyon, the ramp where daredevil Evel Knievel attempted to cross the canyon on a steam-powered rocket.

Just… watch the video if you haven’t.

(3) Arrowhead Shot Cuts Tau in the Brain by 73%

Arrowhead Pharma, video created with Runway

Speaking of brains…

As I’ve written before, my grandmother had Alzheimer’s, and me or someone I love getting Alzheimer’s is one of my biggest fears in life. So whenever there’s a potential development on the Alzheimer’s front, we cover it, even if they’ve historically ended in disappointment.

The hard part of treating brain disease like Alzheimer’s is getting drugs to the brain. The blood-brain barrier is very good at its job, which is why RNAi drugs, which silence a gene before it can make a protein, have been ineffective against brain disease. They’ve mostly been effective in targeting the liver (which filters everything).

On Wednesday, Arrowhead Pharmaceuticals reported interim Phase 1/2a data on ARO-MAPT, which silences the gene that produces tau, the protein that tangles up in Alzheimer’s, frontotemporal dementia, and progressive supranuclear palsy. Healthy volunteers who got two subcutaneous doses at 12.8 mg/kg saw mean total tau in their spinal fluid fall 73% by Day 105, with one participant down 85%. Day 105 was 90 days after the last shot, which means the effect held for three months and the eventual dosing schedule could end up be quarterly or longer. Check the video to see how it works.

All from a shot in the arm!

It’s early, but early Alzheimer’s cohorts are enrolling now and Arrowhead expects to return data in 2027. I hope it works, and that we can all forget about Alzheimer’s once and for all.

(4) Another Very Big Week for Robotics

Nature Will, Pi, DeepMind, Flexport

On Tuesday, Runway co-CEO Anastasis Germanidis and I published Runway: Learning From Watching the World, in which we lay out the company’s bet that robots can learn most of what they need from third-person video. It’s been a big week for robots overall.

On Monday, Physical Intelligence opened its Pi API beta. With it, companies can upload their robot data, fine-tune a π model on π’s compute, and run inference right on their own robots. Early partner companies are using it on everything from picking strawberries to assembling servers, from driving autonomous wheel loaders in Utah to controlling a dual-arm robot that deburrs the spacers used in its own assembly, basically with an API call.

On Tuesday, the Beijing lab Nature Will run by Tsinghua professor Ning Ding released Natura-1, a 50 billion parameter model pretrained on one million hours of egocentric human video. Egocentric means its video from people’s perspective, not robots’, which is important, because robot data is harder, slower, and more expensive to collect. The model watched egocentric video of people using their hands from their own perspective. Only afterward did they post-train on a small amount of robot data, a la Runway.

Per Rohan Paul, what’s interesting about this is: “One model handles 3 things robots usually fail at together: fine 2-handed control after the grasp, new tasks without extra training, and long sequences where it has to remember earlier steps and recover from mistakes. Many robot systems show 1 of these in a narrow setting.”

That same day, GDM entered the fray. Personally, as a long time GDM fan and an unbiased observer, I think Google DeepMind’s demo was the week’s coolest. Norman Di Palo posted a video of Gemini Argon teaching an Apollo humanoid to play cornhole (above). They gave the model a video of a person throwing a beanbag, and Argon took control of the robot, generated its own simulation environment, ran system identification so the sim matched the real motors, trained the throw, and started sinking them. The team did not have to write any specific cornhole controller or anything. It missed a few first — who amongst us hasn’t — but then it gets the hand of it.

The dream: robots enjoying themselves playing cornhole while we humans work!

Unless… Warehouse robots are here! on Thursday, Ryan Petersen posted a video of Flexport’s newest automated fulfillment center, just outside Chicago, where robots move real freight for real customers.

Robot demos are notoriously bad at turning into actual, valuable robots, and Flexport’s certainly isn’t the first automated warehouse. I’m not smart enough to judge any of these approaches against each other, or to say whether any of them will be the one, but there certainly seems to be a lot of smoke in robotics recently. If there’s fire (metaphorically, of course), the near future could get real abundant real fast.

(5) Applied Materials Unveils $5B Chip R&D Center in Silicon Valley

Applied Materials makes the machines that chipmakers like TSMC and Intel use to build chips: the deposition, etch, and inspection tools that lay down and carve the layers on a silicon wafer.

On Monday, it opened the EPIC (Equipment and Process Innovation and Commercialization) Center in Sunnyvale, California. It’s a big investment and a big building. It will cost five billion dollars, cover 600,000 square feet, include nearly four acres of cleanroom, use enough concrete to fill 34 Olympic pools and enough steel for two Eiffel Towers, and take six million construction hours to complete.

As impressive as the space is, this made the Dose because of the partners it’s bringing under one roof in the US of A: TSMC, Intel, Samsung, SK hynix, and Micron are all founding partners, alongside Broadcom, Advantest, KIOXIA, SCREEN, and Besi on the equipment side and Stanford, Berkeley, UCLA, Arizona State, and RPI on the research side. These are competitors, who occasionally sue each other, but have now agreed to put their engineers in the same cleanroom to work on process technology three to four generations ahead of anything in production today.

The point of EPIC is to speed up chip development, which normally runs in series, where materials research hands off to equipment development which hands off to pilot production, with many steps taking place in far flung locales around the globe. The whole process, as it stands, takes the better part of a decade. EPIC is built to run those stages in parallel under one roof, in order to compress the learning cycle on the slowest (and least glamorous) layer of the entire AI stack.

A few years back, I wrote a piece called I, Exponential praising the beauty of exponential curves in technology, the most famous of which is Moore’s Law. As Byrne and Tobias wrote in Boom, Moore’s Law happens because people believe it will happen and then invest ahead of future bottlenecks against that belief, which becomes a self-fulfilling prophecy. This is that in action.

Thank God for Data Centers.

(EXTRA 1) Ben Wilson on David Senra

Putting this one above the paywall because Ben is awesome, David is awesome for doing this, and if anyone the not boring community can help in any way, it would make me very happy. I’ll just leave you with David’s intro:

Ben Wilson has stage four cancer and four small kids. I wanted to have him on the podcast to talk about his work on the How to Take Over the World podcast, which is, in my opinion, one of the best podcasts ever created, and to talk about what it’s like dealing with terminal cancer at such a young age. I also want to ask you for help in any way that you can. Ben is a good man and he’s been dealt a bad hand. Please watch this podcast and see if you can help with medicine, introductions, or donations. It doesn’t matter if you have $5 to spare or 5,000. any little bit can help Ben’s family through this devastating time. He’s a good guy and a close friend. Please help in any way that you can.

You can donate here.

Discussing his prognosis, Ben says, “the amount of new treatments that are coming out for all sorts of different types of cancers is remarkable. We live in a really remarkable age.” He’s right. We cover it every week.

Just this week, for example, Bristol Myers Squibb and RayzeBio's RYZ101 significantly delayed progression of neuroendocrine tumors in a 343-patient trial. Neuroendicrine carcinoma is exactly the kind of cancer Ben has. Get fucked, cancer.

EXTRA DOSES: Science Breakthroughs, XBOX Labs, Zuck & Wang

Read more

Subscribe now

Hi friends 👋,

Happy Tuesday!

A couple of weeks ago, I mentioned that I was partnering with Runway this quarter as our presenting sponsor to bring some of my writing to life in the form of video. It took us a while to land on the right partner to form a deeper relationship with — it has to be a company I want to talk about a bunch — but when Runway reached out, it was a clear perfect fit. For one thing, I’ve had a lot of fun playing around with their video models. But for two others: 1. I’m fascinated by World Models (and quoted Runway co-CEO Anastasis Germanidis in the essay Pim DeWitte and I wrote) and 2. Runway has been on a tear with new product releases.

I wanted to understand how, so to kick off the partnership, I asked to write an essay with me explaining how video, interfaces, operating systems, and robotics all fit together, why video was the right starting point for World Models, where the world is heading if Runway’s bets are right, and whether or not, as someone who simulates worlds, he believes we’re living in a simulation.

Let’s get to it.

Runway: Learning From Watching the World

A Co-Written Essay with Runway Co-CEO Anastasis Germanidis

From time to time, a certain company catches fire. You can tell when it’s happening by the pace and breadth of their product releases. One day, a new capability. Cool. The next day, another. Woah. Next day, another still.

Runway, the world model company best known for releasing the first commercial video generation model, is on such a run.

Over the past six weeks, it has shipped a real-time interface model, a new interactive world model, an open-weight robotics model, an AI agent for performance marketers, a research preview of an operating system, and it bought a robotics lab in Paris along the way. Meanwhile, the business is growing faster than at any point in its history, adding more than $100 million in annual recurring revenue in the last quarter alone. From the outside, it looks like a company that suddenly figured something out.

From the inside, the view is much more interesting. Runway has been pursuing a particular set of bets for eight years, and that near-decade of research and iterative deployment is now yielding a seemingly disparate but actually related set of new models and products.

On September 1st, Runway introduced Solaris, the “first Interface World Model,” which allows users to generate interfaces in real-time, through prompts or actions on screen, like this one where a guy takes a selfie and tries on glasses and tweaks his hair and smile with the drag of a mouse.

On September 8th, Runway announced it had acquired Kinetix, a Paris-based research lab, which specializes in… robotics? From the video company?

Then at their September 30th AI Summit, the team shared Praxis-1, “an open-weight World Action Model that turns our video pre-training expertise into real world control for robots.”

The thesis is that while robotics data is limited, video data is practically infinite. So Runway pre-trains on video data and then uses that as a policy model for robotics developers who can train it on their specific embodiments. Standard Bots, whose co-founder Evan Beard and I co-wrote an essay on robotics, is working with Runway to deploy Praxis-1 on its arms.

And then the very next day, on October 1st, Runway previewed Project Continuum, “a new operating system research application built around real-time video interfaces.”

A video model. An interface model. A robotics model. An operating system. Are you getting it? These are not separate pretrains. They’re different applications, built around the same General World Model backbone, with different post-training on top.

Made with Runway

All of them are examples of what you can do when you can simulate any pixel.

An operating system ends up as pixels on a screen, like video and interfaces do, and like the world does through a robot’s eyes. If you can simulate every pixel at high enough fidelity and speed, then you can make almost anything that shows up to the world as pixels. This is the future Runway is building towards.

Runway is making a few different bets, on a research direction, the Bitter Lesson, and a business strategy. It is betting that world models can learn, like humans do, from third-person observation. It is betting that letting models watch a ton of video will lead to models that understand how the world works better than they could by reading about how the world works or from any other human-mediated data source. And it is betting that it can build a massive business in content creation in order to both help fund the large investment into better and better models, and to become the rare company that is both a frontier lab and a distribution layer.

Runway’s commercial success and product prolificness points to a surprisingly wide set of applications for World Models; surprising, that is, unless you read the essay Pim DeWitte and I co-wrote explaining why World Models have the potential to eat everything. In it, we wrote:

World models — these systems that learn from watching the world and the actions taken in it — are a fundamentally new kind of foundation model. They can compute what was previously uncomputable.

There are a number of different types of models whose makers call them World Models, although many don’t fit our definition; they don’t predict the next state based on intervention. We categorized Fei-Fei Li’s World Labs, which AMD recently agreed to acquire for $8.2 billion, as building “3D Reconstruction and Generation Models.” Runway, we wrote:

Is blurring the lines between video generation and world generation, as mentioned in the Video Model section. During its Research Demo Day 2025, Runway co-founder and CTO Anastasis Germanidis explained the company’s evolution that started from “generative AI models [as] viable tools for creative expression”. They then evolved towards World Models (while still making incredible progress in video models.)

“To build a World Model,” Germanidis explained, “We first needed to build a really great video model. We believe that’s the right path to building World Models, that teaching models to predict pixels directly is the best way to achieve general purpose simulation.”

Runway first mentioned its World Model ambitions in a December 2023 blog post by Germanidis, who is now co-CEO, titled “Introducing General World Models.” This has been the plan for quite some time: teach the models to make videos from watching videos, and then teach them how to bring the actions in video to life.

Anastasis describes it as a plan in four chapters:

Chapter 1: Media and Entertainment

Chapter 2: Broader Content Needs and Full Video

Chapter 3: Robotics (We Are Here)

Chapter 4: Science

The plan is coming together, as evidenced by the recent string of releases, and I have become a World Models convert who believes they have the potential to do a lot of things in the real world that LLMs can’t, so Anastasis and I decided to co-write an essay on the company’s bet that watching a lot of videos can teach models to do almost anything.

Learning from Watching

One of the things you realize when you have kids is that babies are learning machines. When they’re really young, they just keep changing, picking up new skills, like little Neos.

The thing about babies, though, is that they don’t know how to read. They don’t learn like we learn in school, by reading textbooks, because they can’t. Instead, they watch. They watch how we walk, how we talk, how we smile, how we bring food to our mouths, how we pick up toys and how we put them away, they watch how we shake hands, and hug, and run, and jump, and crawl. Then they try it themselves, clumsily at first, they watch again, they try again, they watch again, they try again, and try and try and try until they’ve gotten it.

Made with Runway

As we grow up, we do go to school, and we do learn from books, but we still learn by watching. We watch how the kids in the older grades dress and talk and interact. We watch how our favorite athletes play. We watch to learn how to interact with and manipulate the world.

If you’re a guy, think about the last time you had to tie on a bowtie for a wedding. The best way to do it, other than having the one guy who knows how to tie a bowtie tie it for you, is to watch that guy tying his bowtie. The second best is finding a YouTube video of someone walking you through how to tie a bowtie. At the bottom of the list is reading how to tie a bowtie. I’ve tried it, and it simply doesn’t work, at least not for me.

Language is compression. We look at the rich, ever-changing world in front of us, and we can’t possibly capture it all. By the time we’ve come anywhere even close to describing what we see at any given time with any amount of precision, it’s changed. And no matter how hard you try, words will never be enough, because you can’t, for example, see each of the atoms in the scene, track the movement of the electrons and protons and neutrons that make up those atoms, see how electromagnetic and gravity waves interact with them… you get it. Everything in the world changes, then changes again, and each time it does we’d have to start describing it again, like verbal Sisyphuses.

Instead of running a constant monologue as we walk down the street, describing all of the motion and appearances and interactions we see, we simply say, “The man walked down the street.”

Neither approach captures the full picture, but the latter is easier and allows us to actually live our lives. To jack as many subtle details into our brain as possible, therefore, we don’t read about life, we watch it. Our brain has evolved to pull out the relevant information.

Made with Runway

A picture is worth a thousand words. A video is worth 1,000 times the number of frames plus however many words the motion connecting them and sounds are worth.

But the value of video isn’t simply that it contains more information than words. Predicting what happens next and checking how well you predict, over and over again, creates a pressure to learn reusable relationships.

Imagine trying to predict the next frame of that street video.

To get it right, it isn’t enough to know what the man looks like. You need to know that he’ll keep walking in roughly the direction he was walking, at roughly the speed he was walking, that his left foot comes after his right, that people walking near him will walk behind him and not through him, that when he passes in front of a moving car, that car still exists and will come out the other side a beat later as the same car, that his shadow moves when he moves, etc.

When a model is trained on video, nobody tells the model any of this. There’s no line of code that says “objects still exist when something walks in front of them” or “people can’t walk through people.” The model guesses the next frame, sees how wrong it was, and adjusts. Somewhere along the way, the cheapest way for the model to stop being wrong is to learn something like the rules: things fall, and seem to fall in basically the same way at basically the same rate of acceleration (although feathers seem to flutter back and forth while bricks simply drop); glass shatters and rubber bounces; a ball that rolls behind something comes out the other side looking exactly the same, unless there was a puddle of mud. One video with a muddy ball might confuse the model, but over millions of them, it would learn that the mud is an anomaly.

These, incidentally, are things that babies learn during their first year. We play peek-a-boo because it makes them laugh, but also because it teaches them object permanence, this idea that things still exist even when you can’t see them.

This is how LLMs work, too. Their goal is to predict the next word, by any means necessary. Ilya Sutskever has been making this point for years: predicting the next token correctly requires understanding something about the reality that produced the token.

Swap tokens for frames and you have Runway’s thesis.

There’s a theory in neuroscience, which I discussed with Cristóbal Valenzuela, Runway’s co-founder and Anastasis’s co-CEO, called active inference. Karl Friston is its most well-known proponent. The idea is that your brain is a prediction machine, constantly running a simulation of what’s about to happen and checking it against what actually happens. You don’t normally notice it, because you get really good at prediction and it just runs in the background. You only notice it when you get it wrong. You’re walking up the stairs, you go for one more step, the step isn’t there, and for a split second your whole body lurches while the simulation crashes and reboots. “The moment you see it didn’t predict,” Cristóbal said, “you adjust.”

A model that sees enough examples of success and failure starts to carry around its own version of that simulation. But until the model runs in a physical body instead of a screen, it can’t feel what it’s like to stumble, trip, or fall.

Why video, though, and not another type of data? That’s where the first of Runway’s bets comes in: The Bitter Lesson.

The Bitter Lesson and the Big Blob of Video

It is practically illegal to write about training models without mentioning The Bitter Lesson. In 2019, Richard Sutton wrote a short essay arguing that the biggest lesson from seventy years of AI research is that “general methods that leverage computation are ultimately the most effective, and by a large margin.” It is bitter because every time researchers have tried to build their own understanding of a domain into a system (chess openings, the structure of grammar, the edges and corners in an image), they’ve eventually been steamrolled by dumber methods that could soak up more compute and more data.

For his part, Anthropic CEO Dario Amodei made a similar observation two years earlier, in a recently-partially-leaked internal OpenAI research document he titled “Big Blob of Compute.”

The Big Blob of Compute (BBOC) hypothesis, he wrote, “says that creating any given intelligent behavior is mostly about providing a large, minimally structured mass of computational capacity, and then giving it shape and form via interactions with a rich environment and a training process that drives towards behavior appropriate for the task at hand.”

Runway’s corollary is that the most scalable source of data tends to win. Call it the Big Blob of Video (BBOV) hypothesis.

LLMs won on text because the internet had an absurd amount of it. Robotics has been climbing its own ladder. It used teleoperation data, a specific robot doing a specific task puppeteered by a human. This year, many robotics labs began using a lot more egocentric video from cameras strapped to people’s heads as they perform a task. Runway is betting on something simpler and less human-mediated: vast quantities of video that already exist, and will be filmed (and generated) from here on out.

Watching video is the most Bitter Lesson-pilled approach to AI there is. Language is a compression of reality, an incredibly good one that humans have been refining for thousands of years, but a compression nonetheless. As Sutton wrote, “We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.” In other words, we limit agents with our own beliefs.

Put more bluntly: if you really believe the Bitter Lesson, you should be working on World Models.

There’s a good argument against this thesis, so we might as well make it ourselves.

If language is compression, well, compression can be useful!

Learning about the world from pixels is less efficient than learning from text. Researchers have shown that you can build a language model out of raw audio, just by predicting the next little chunk of sound, and that language eventually emerges, but that it takes something like three orders of magnitude more compute than learning from text. Text comes predigested. Somebody already did the work of sorting out the most important stuff. Pixels make the model do that work itself, and there are a lot of pixels on which the model must do it.

Then there’s Yann LeCun’s critique: predicting every pixel wastes effort on details that don’t matter (every leaf on every tree, every ripple in every puddle), so models should predict in an abstract representation space instead.

Runway’s answer is pragmatic: use whatever works. So far, we haven’t seen evidence that JEPA scales better than the diffusion models it uses, and the simplest approach tends to scale best. Plus, it’s really hard to know ahead of time which pixels are noise. The rustling leaf isn’t noise if it tells you that the wind has picked up. The wind, even ever so slightly, will impact how everything else in the frame moves.

The efficiency problem is real, though. Runway thinks the answer is more compute, better training to squeeze more from every frame, and better algorithms.

As Anastasis said in our first conversation, “Naively scaling what we did two years ago won’t solve these problems. I expect the compute needed to train World Models to eventually exceed what LLMs need, and we’ll need real algorithmic progress on top of that.” Good time to be Jensen Huang.

The World Model Backbone

Anastasis’s bio on X reads “Simple ideas, pursued maximally,” which is also a pretty good summary of Runway’s research strategy.

The simple idea at Runway is next-frame prediction. A language model learns to describe the world by predicting the next token in a sequence of tokens. An autoregressive World Model, as Cristóbal put it, learns to simulate the world by predicting the next frame in a sequence of frames.

Autoregressive is a fancy word for a relatively simple idea. The model generates one frame at a time, each based on the frames that came before it, the same way an LLM writes one word at a time. Most video models don’t work like this. They generate a whole clip at once, refining it from static into a picture over dozens of steps, which is great if you’re making a ten-second ad and useless if you’re trying to interact. Interactions stop feeling interactive somewhere around half a second of lag. If you want a model that a person, or an agent, or a robot can actually interact with the way humans do with other humans, it needs to produce the next frame quickly, then the next one, then the next, each taking into account whatever just happened in the previous one and in the reaction to it.

That’s what Runway did with GWM-1, which it announced last December. It took Gen-4.5, its most recent base video model, and turned it into a model that generates frame by frame, in real time, and can be steered with actions. In its Solaris announcement, asking, “What if software could generate itself, as you interact with it?,” it laid out the general recipe:

1. Teach the model to generate frames sequentially,

2. Distill dozens of denoising steps down to a few,

3. Train the fast model on its own outputs so the quality holds up over a long session.

No code underneath, no layout to translate. The image is the application. It treats your clicks and drags the way our video models treats words: not as commands to look up, but as a signal for what to generate next. Nothing about it is decided in advance, so nothing about it has to stay the same… the scene answers to your intentions. It becomes whatever you ask of it.

Actions are the important thing.

In my (Packy’s) previous essay on World Models, we drew the line between video models and World Models like this: a video model predicts the next frame, P(x_t+1 | xt), while a World Model predicts the next state given an action, P(s_t+1 | s_t, a_t). “That a_t,” we wrote, “is the magic.”

Runway’s last month or so of releases can be viewed as a bunch of different a_t’s. In GWM Worlds 2, the actions are camera movements plus free-form text addressed to anything in the scene. In Runway Characters, the action is audio, a person talking to an avatar that listens and responds. In Solaris, the actions are clicks and drags. In Praxis-1, the actions are robot commands, expressed as where the robot’s hands should go next. There’s a different vocabulary for each, but they all use the same machinery underneath to predict what happens next.

I (Packy) have been fascinated by World Models, and really do think they’re some kind of magic, so when I talked to Cristóbal, I wanted to be sure that I understood: “So all of this really just is the same backbone?”

“Correct,” he said. “The backbone is frame prediction at scale, which eventually gives rise to a World Model in the same way that next token prediction at scale gives rise to large language models. And the applications are diverse on top.”

This is why the company is moving so fast across seemingly disparate applications right now. About 90% of Runway’s research effort goes into pretraining. The belief is that the capabilities get learned there, by watching, and that post-training mostly brings out what the model already knows. Once the base gets good enough, each new application takes far less effort to stand up than the one before it.

You don’t, for example, need to build a robotics model from scratch. You can take a model that has already watched an enormous amount of the world and teach it to speak robot.

There’s a lot of room left. Runway estimates that Video and World Models are a year or two behind LLMs, which means pretraining is nowhere near saturated.

My (Anastasis’) rule of thumb is the number of pixels a model can generate, which is resolution x duration, roughly doubles every few months. I don’t see any fundamental blocker to hour-long generated video within a year or two.

Longer videos aren’t simply a matter of doing the same thing, more. A model that has to keep a scene coherent for an hour has to learn persistence. Turn around, and what was behind you is still there. The mug you set down on the counter is still on the counter twenty minutes later, unless somebody moved it, in which case the model should know who moved it, and where. Those are exactly the things today’s models still struggle with; longer context, the company believes, is how you train those problems away.

Longer term, the really big bet is that the field moves toward one model that does language, pixels, audio, and actions together. Even in the near-term, though, the World Model backbone that’s used to generate pixels on a computer screen seems to also be able to help robots navigate and interact with the physical world.

Teaching Robots with Video

Every robotics company with ambitions of deploying at scale runs into the data wall: real-world data is scarce, expensive to collect, and a general-purpose robot needs a lot of it.

In Packy’s Robotics piece with Evan, they wrote a whole section titled, “Robotics is Bottlenecked on Data,” on this idea:

LLMs had it relatively easy. The entire internet existed as a pre-built training corpus. There is so much more information on the internet than you could ever imagine. Any question you might ask an LLM, the internet has probably asked and answered. The hard part was building architectures that could learn from it all.

Robotics has the opposite problem.

So robotics companies have to build their own internet equivalent, one demonstration at a time, often with a human in a VR headset puppeteering an arm through the same task over and over again. For jobs with long tails of weird edge cases, like household robots, Runway argues that there’s no practical way to collect all of the data you’d need, let alone enough of it.

Video, on the other hand, is effectively limitless. People film and upload more of everyday life every single day than every robot lab on earth could capture through teleoperation. And as generative models get good enough to make video that’s indistinguishable from the real thing, it’s becoming even more limitless. You can start with real scenes and simulate thousands of variations, each with subtle differences.

Packy here. In that same Robotics essay, Evan and I wrote a section called “Video Dreams (and Their Limits)” in which we argued that internet video isn’t enough. Video shows you trajectories and timing and the order of steps, but it doesn’t carry mass, force, friction, or stiffness. Two surfaces can look identical and behave completely differently, and the robot finds out the fun way. We cited a study in which people who watched more videos of moonwalking and juggling became much more confident and not much better. Reading lots of books makes it easier to write a good book, we wrote, but watching lots of golf videos doesn’t do much for your golf game.

First, hell yeah. This is why I do these co-written essays. Smart people can disagree. General robotics is still unsolved and people much smarter than me are taking many different paths.

Second, though, this isn’t the disagreement it appears to be on the surface.

As one piece of evidence: Standard Bots is now testing Runway’s Praxis-1 on its RO1 arm.

What resolves the seeming disagreement is that Runway isn’t betting on video instead of robot data, but before robot data.

Praxis-1, which Runway unveiled at its AI Summit on September 30th, learns mostly from third-person video, built on the same large-scale pretraining as the company’s World Models.

From there, Cristóbal explained, “you can take the learnings of large video data and start fine-tuning it on robotics. So you add a bit more egocentric data, and that egocentric data will help the model understand a bit better the situation that it might be in.”

Then you fine-tune on the specific robot.

How much robot data does that last step take? About 30 minutes of data teaches the model to control a robot body it hadn’t seen in pretraining, and in new environments it can reach strong performance on some tasks with just minutes of data. You still have to swing the club a bit, so to speak, but you need a lot less practice when you have a robust model of how the world works pre-loaded in your brain.

As a test, Runway pretrained robot policies on ordinary video and, separately, on teleoperated robot video, using anywhere from ten to a thousand hours of each. Then it fine-tuned them and measured how close the robot got to putting an object where it was supposed to go. The two lines are basically on top of each other: web video worked as well as robot video. The web-video version ended up with a final placement error of 16.1 centimeters versus 16.0 centimeters for the teleoperated robot video version.

That’s not nothing; for high-precision tasks, 0.1 cm might as well be a kilometer. But these models are as bad as they’re ever going to be, and teleop data isn’t getting any cheaper to capture. If they’re roughly interchangeable as a starting point, the robotics data problem looks very different than it did in January.

But robots’ bodies and forms vary wildly, both from each other and from humans’. How can a model that’s mostly watched humans control a robot’s arm precisely?

Part of the answer is in how Runway represents actions. Praxis-1 expresses actions in a standard end-effector format: where the gripper (the “hand” at the end of the arm) should be in 3D space, and how it should be oriented. A person’s hand reaching for a mug in a YouTube video can be described in exactly the same terms. So can a humanoid’s hand, a six-axis industrial arm, or a two-armed setup. While the embodiment changes, what the model is trying to figure out – where should the hand go next – doesn’t really. To that end, Runway’s model is currently being tested on Noble Machines’ bimanual setup, Standard Bots’ six-degree-of-freedom RO1, and Ultra’s mobile robot.

In one demo, the same policy works in a controlled studio and in a home kitchen without retraining.

It may be helpful to think about the three ways World Models can be used in robotics.

The first is generating training data. Take a small set of real demonstrations and create variations with new objects, new lighting, and new layouts with which to train the robots on a bunch of different situations they might not see in real world training very quickly.

The second is simulation, or testing robot policies inside the World Model before you put them on real hardware. This is where Runway has published its most rigorous evidence.

In February, its robotics team took eight different manipulation policies from the RoboArena benchmark, all running on the same Franka Panda arm, and ran each one inside GWM-Robotics, an early version of Praxis-1. Each policy made its own decisions based on what the World Model showed it. Human graders then scored 1,450 simulated rollouts, more than 16,000 ratings in all, about ten per rollout. The ranking of the eight policies in simulation matched their ranking in the real world with a correlation of 0.95, and the graders correctly picked out both the best and the worst policy. It’s a test of ranking (which policy is better than which), not of absolute success rates, but for a robotics team deciding which of your twenty candidate policies deserves time on your three physical robots, ranking is exactly what you need. The theory here, and across the robotics work Runway is doing, is to do as much in relatively cheap simulation as you can before moving to the more expensive physical world. Runway’s model only needs a starting image, compared with competing approaches that require scanning and reconstructing every scene in 3D, which makes it relatively cheaper than even other simulation approaches.

The third way is to use the World Model as the robot’s brain. This is the one that Cristóbal is most excited about. “One of the most compelling use cases that we found for these models in robots is less on the synthetic data part and more on using the video models as the brains of the robot,” he said. “So they learn not by describing the world, but by simulating the world, and they do it in real-time.”

Remember our example of the person tripping on the stairs? A robot with a World Model for a brain is constantly predicting what should happen next and adjusting when it doesn’t, like that human does. That’s why recovering from failure is one of the key capabilities. If the gripper slips, the box is heavier than it looked, or the bag tips over, the model has to notice and adjust in real-time, the way you catch yourself on the stairs.

Like humans, then, once embodied, these models are able to start learning both by seeing and by doing. The seeing helps them start doing, and then they learn some more from the doing.

Runway thinks the first real deployments will be in enterprise: warehouses and manufacturing, fine-grained manipulation tasks where there are so many ways for things to go wrong that generalization can pay for itself. Useful home robots will take longer. That’s where we landed in the Robotics essay, too. Start where the economic value is, get paid to learn, and keep eating more.

And what will the robots doing increasing amounts of work look like? Runway thinks there will be a menagerie of cheaper robots that can use tools to do specific jobs. Simpler, cheaper robots scale better, from a data perspective, than expensive humanoids that break easily. And it will become cheaper to make things to stick on robots to help them do whatever job they need.

Long-term, I (Anastasis) don’t think we end up in a world of all humanoids. LLMs make it easy to design components and 3D printers make it easy to produce them, so people will design and fabricate custom robot bodies for specific jobs, all running on the same model layer. This is particularly true if teaching the brain to use new embodiments continues to take 30 minutes across arbitrary embodiments.

In this world, the model that is the robot’s brain is the layer at which value accrues, so what do you do to own that layer? Runway thinks you give it away.

Runway plans to release Praxis-1 with open weights in the coming months (for now, it’s with early partners), because open models give hardware developers more control over their own robots than closed ones. Open weights also help make Praxis the default starting point for anyone building a body, which means that Runway gets more data, which feeds the next, better Praxis and the next, better World Model on which it’s built. Runway’s acquisition of Kinetix, the Paris-based robotics research lab, earlier in September makes sense in that light, too.

It is true, as we wrote in January, that videos don’t contain data on things like force that really matter for making robots work perfectly and predictably. Robots still need to practice in a physical body, like humans do – you’re not going to drive the ball 300 yards the first time you swing a club, no matter how many Tiger Woods videos you’ve watched – and robots, like humans, need to learn from failures in order to anticipate them the next time. But it also seems true that good priors are worth a lot, that you need less practice when you start doing something with an understanding of what great looks like, or as Runway describes it, “how objects behave, how hands move, what a task looks like partway through.”

Robotics represents an enormous and obvious opportunity, and one for which World Models seem to be particularly well-suited. Babies, adults, and World Models all take in the scene around them and try to predict what to do next. If Runway can supply robots’ brains, it will build a very big business doing so.

But the team is as excited by, if not more excited by, the types of products and markets it can open up if it gives people the tools to simulate every pixel. Robots acting on a physical scene and a person clicking on a digital one, after all, look the same to a World Model.

Interfaces, Operating Systems, and Pixel Maximalism

If it’s just predicting frames or pixels on a screen, well, why not just predict everything on a screen?

A philosophical answer to my question about how Runway thinks it can build an operating system with a World Model via Continuum.

When people hear “pixels,” Cristóbal said, they think of video and movies, and then the obvious next steps, like games, ads, content. Runway, though, is a Pixel Maximalism company. An operating system is pixels on a screen. Us humans, we don’t think about using a computer as the same thing as watching a movie, but to a machine, it’s all just pixels.

Recall about how LLMs treat tokens. The same LLM that writes poetry can write high-performance C++, because to the model it’s all just tokens.

“Models want to generalize about the world,” Cristóbal said. ”And the task itself, how we label the task, doesn’t really matter for the models. It’s just pixels on a screen.”

Think about how software gets made today, even when LLMs do it. Someone designs what it should look like. Someone else (or, increasingly, the same person or just the LLM itself) translates that design into code. Everything about it – how it will look, what buttons will do, how it feels to the user – has to be anticipated and programmed ahead of time, then shipped. In the Solaris announcement, Runway describes the result as a lossy compression of all the ways someone might want to interact with a piece of software, “frozen before any user arrives.”

Solaris, which Runway released at the end of August, is the first of what it calls Interface World Models. Instead of rendering an app that someone has coded, it generates the interface itself, frame by frame, as you use it. The user’s clicks and drags become conditioning for the next frame, the same way a text prompt or an image would be. (That’s a_t again.) You can browse a virtual clothing store where the showroom is the interface, pull a shirt off the rack, and drag it onto a photo of yourself, tuck it into your digital pants, tug it back out. You can build a salad by dragging ingredients into a bowl. Remember the guy at the top of this piece, trying on glasses and fixing his hair with his mouse? Nobody coded a “try on glasses” feature. The model just knows what happens when glasses go on a face, because it’s watched more hours of video of people putting glasses on and adjusting them than any human would willingly be subjected to.

For now, Solaris pairs a World Model with an LLM. The LLM figures out what you’re asking for and decides what the application should do next (turns your words into an action); the World Model decides how that action impacts the next frame and renders it in real time. This is how Runway thinks things work in the short term. The World Model becomes the rendering layer (replacing the HTML, CSS, and JavaScript presentation layer), and the LLM handles the logic of translating your actions. Runway’s gaming work splits work the same way; the LLM writes the scenario, and the World Model renders it.

Compared to the way we’re used to interacting with software, it feels magical. Runway ran a study. It gave Solaris and Claude Opus 5 the same starting image and the same interaction requests, and had 250 people make nearly 7,500 side-by-side judgments. When asked which result better followed the instruction, people picked Solaris 61% of the time to the coded version’s 24% (13% called it a tie). When asked which behaved more naturally within the scene, it was 71% to 21% (7% tie).

“The second result highlights the broader difference between the two approaches,” the blog post says. “A coded interface can often reproduce the requested change, but it treats the interaction as an isolated update to the interface. With Interface World Models, because the model already understands how objects, materials and environments behave, it can generate interactions that feel coherent within the scene rather than treating each UI action as an isolated element.”

The post also admits that there’s a lot Solaris can’t do well yet. Text is the big one. Stable, legible text is one of the hardest problems in video generation, and interfaces are mostly text. Trust is another; for anything instructional or commercial, a convincing wrong answer is worse than no answer at all, so Runway is researching “conditioning generation on richer verified context as the session unfolds.” Long sessions drift, and Runway is actively researching fixes here, too. Finally, generated interfaces still need to work with screen readers and the rest of the software stack, “so that flexibility doesn’t come at the expense of usability.” Generating every frame fresh is costly, although Solaris is already orders of magnitude cheaper to run than a standard video diffusion model, and the cost curve keeps bending. For now, not every website should or will be generated in real-time, but things are heading in that direction.

While we’re at it, though, while we’re glimpsing at the future, why stop at generating individual websites? Why not generate the whole operating system?

On October 1st, Runway previewed Project Continuum, “a new operating system research application built around real-time video interfaces.” It is a research preview and very early. But it, too, points to a new way of interacting with software and the world around us.

“What it enables you to do,” Cristóbal explained, “is you can make anything an interface. Like, you take anything you want, if you wanna learn about anything in the world, you take a photo of what you want and that itself becomes an interactive application.”

In this future, if you’re trying to fix your engine, you can just take a picture of the engine and then turn it around and poke at it in sim; you can even turn the engine itself into an instruction manual. There was this great launch video for the Freckle Phone for kids a couple of weeks ago, that shows kids taking pictures of the world around them and learning about what they see from an LLM; what if the pictures could turn into interactive lessons for kids to play with in whatever way they wish?

This isn’t the future of software for everything software does today. Medical records and financial transactions are probably better left in a safe, structured database. “But for many others,” Cristóbal said, “I think that’s the holy grail of computer graphics. It’s like anything becomes an interface.”

When the holy grail becomes practical comes down to cost. If generating an interface gets close to the cost of running a traditional app, then the use cases best served by real-time generation will be.

First we shape our interfaces, then our interfaces shape us, or something. We humans alive in 2026 never felt, or barely remember, how magical it must have been moving from the command line to a GUI. We did feel what it was like moving from a GUI to a touchscreen, and putting that touchscreen in our pockets has changed the way that we interact with the world, for better and worse. Now, we mostly interact with the most powerful software ever built through a chat interface. Solaris and Continuum are bets that chat isn’t the final interface, that everything you can see might become the interface.

That is a wild and beautiful future, and it sounds like an expensive one, too. Someone has to pay for the research that makes it possible, and put it in contact with the messy world of real users.

Runway believes that it can solve both issues at once by building an actual business on their models today.

A Business and a Lab

Runway was founded in 2018 by three people who met at ITP, the graduate program at NYU’s Tisch School of the Arts: Cristóbal, Anastasis, and Alejandro Matamala Ortiz. They were artists and engineers building tools for other artists, back before AI was real.

Art school is in the company’s DNA. Multimodality is more multidisciplinary than language modeling. Building models that see and hear well takes taste, that overused attribute that many claim to have but over which art students at least have a more credible claim. Runway has both artists and engineers. It’s part of why Runway is in New York, where you are smushed into a melting pot of people doing all sorts of different things (also, many World Model researchers are in London and Europe, which makes logistics easier from the East Coast than the West).

It’s also why Runway decided early on to build a real business selling to customers in addition to its Research Lab.

Runway released Gen-1, the first public video model. Its most recent, Gen-4.5, topped the leaderboards. Getting creative tools in creative peoples’ hands lets the company learn faster.

But it also gives the company real revenue, which is accelerating faster than at any point in the company’s history. Runway is used by 95%+ of the Fortune 100, is growing ~50% month over month, has seen credit usage grow 5x this year, added over $100M in ARR in Q3 2026, and has 5x-ed its growth rate since. Customers include Dolce & Gabbana, Palo Alto Networks, Lucid Motors, Cloudflare, Yamaha, every major Hollywood studio, and every major advertising agency. Its tools have been used in Oscar-winning movies and Super Bowl ads.

In the four chapters framing, Chapter 1 (media and entertainment, making the tools powerful enough for real editing workflows) are bleeding into Chapter 2 (broader content needs: marketing, advertising, full videos, whole campaigns). Runway sizes content production, things like storyboarding, VFX, and post-production, at $600B+, and marketing and advertising at $1T+.

Runway Ads is the clearest sign of the move from Chapter 1 to Chapter 2. From the perspective of an art student or a researcher, it’s downright commercial. Designed specifically for performance marketers, it’s a tool Runway built because they needed it themselves.

The company said they’ve been using it for a couple months to “scale ad volume 1,000% while doubling our return on ad spend, increase conversion by 34% with click-through rate holding steady, and bring our cost per subscriber down 41%.” They’re now rolling it out to early design partners, with a public launch coming in weeks.

Runway describes itself as both a frontier lab and a distribution layer.

There are essentially four labs building world-class video generation models: Google, ByteDance, Kuaishou, and Runway. Most other video platforms are aggregators, reselling other people’s models, which squeezes their margins and leaves them one API cutoff away from trouble. Being a lab with its own distribution means Runway sees demand signals across the whole field: what people actually make with the tools, what they’re willing to pay for, what works with customers, etc. This data helps them make both better products and better models.

Runway uses the business side to feed and fund its bigger bet on World Models, which is a motion they think others will follow.

Anastasis expects every AI lab to turn into a World Model lab eventually. “Our head start is that this was the plan all along. We started working on World Models right after we released our first model.”

This is a common structure in deep tech. SpaceX using Starlink to fund Starship and the eventual colonization of Mars is the canonical case study here (see: Casey Handmer’s Starlink is a very big deal), and today, every business that can use near-term data center demand to bootstrap its really big long-term mission, is. Plus, you learn things interacting with the world that you don’t learn in the lab. This is why I’m bullish on Fuse’s approach to fusion – selling radiation on the path to Q>x) relative to many of its research-focused competitors.

And while other video model companies may try to back into this approach, Cristóbal thinks culture will be harder to copy. “There’s a different culture in being able to build new things than being able to optimize or copy things that are already in the world,” he told me. “We’re extremely good at building new things and we’re getting much better at optimizing the new thing.”

The risk, of course, is that the bigger bet gets very expensive. If World Models end up needing more training compute than LLMs, as Anastasis expects, the creative business has to keep growing fast, in a market where Google (the closest thing to an infinite money printer in tech) and ByteDance are competing for the same customers and the same researchers, and the LLM labs are entering into all sorts of byzantine financial arrangements to fund just the capacity they need.

The prize, though, is just as big, if not bigger. Bigger than ads, certainly, and even bigger than robotics, potentially.

Simulating the Universe

Now we are getting speculative, veering into Anastasis’ vision, so we’re going to let him lay it out himself. Over to Anastasis.

DeepMind’s founding mission was to “solve intelligence, and then use that to solve everything else.” For most of the past decade, that’s been the guiding idea at every AGI lab. In 2024, I wrote an essay called Towards Universal Simulation arguing that we should flip it: solve the everything else first, and use it to solve intelligence.

The idea, which we’re pursuing, is to train models on a far wider range of data than we do today. Human-made media, yes, but also observations from scientific experiments and raw sensor data, all at once, with the goal of predicting a larger and larger share of everything that can be observed in the world. Then use those models to run experiments virtually, far faster than we ever could in physical reality.

Most visions of the “AI scientist” imagine models that propose experiments for humans to run in a wet lab and then interpret the results. I think that gets the bottleneck backwards. Coming up with experiments is the easy part. Running them is what’s slow. The way to accelerate science is to give it an alternative to the real-world experiment, by simulating it.

Large-scale model training also has an anthropomorphic bias. We’ve only really scaled up on data made by humans and legible to humans: text, images, video, sound. That makes sense if what you want is a model that talks to people. It’s far too limiting if what you want is a model that understands the universe.

This past July, I extended the argument in an essay called Real-World Superintelligence. AI coding tools have been adopted faster than any software in history, and I think people are drawing the wrong lesson from that. Code worked because code comes with a perfect simulator. You write a program, run it, see whether it works, and do that millions of times. Math is the same: AlphaProof got good by generating problem variants and checking candidate proofs against a formal proof checker. I call this sandbox superintelligence: intelligence produced by repeated interaction with closed environments where you can check the answer.

Real-world superintelligence is something else. It means acting better than humans in messy, partially observed systems, where you have to predict how the world will respond when you intervene, choose under uncertainty, and learn from outcomes that show up late and noisy.

We’re nowhere close to real-world superintelligence. Frontier models can solve Olympiad math problems and still struggle to run a simple business. Anthropic actually tried this; in Project Vend, Claude ran a little shop in Anthropic’s office and, among other adventures, got very into tungsten cubes and lost money.

It’s worse in the real world. Drug trials take many years, trading strategies are expensive to test, and company decisions can take months to play out. In many of these cases, good luck figuring out which inputs exactly led to the outcomes. They are wicked problems.

Think about the leadup to Wright Brothers’ famous flight. In 1901, their glider produced only about a third of the lift that the accepted aerodynamic data said it should. Testing a redesign meant building a full-size glider, hauling it to Kitty Hawk, waiting for the right wind, and then trying to fly it. So they built a wind tunnel, a box where they could blow air over small model wings and measure lift and drag, and that fall they tested something like two hundred wing shapes without leaving the shop.

AI needs wind tunnels for every domain where experiments are slow and expensive.

The labs have started building them by hand. Over the past two years, there’s been a massive build-out of reinforcement learning environments, mostly software clones of things like e-commerce sites and airline booking systems, each one with its own dedicated engineering. That can work for a website, particularly now that AI can build those. For most of the things we actually care about, nobody knows how to hand-build a faithful simulator. Learned World Models are the more scalable path. With Learned World Models, we don’t have to program the rules of the world; the model learns them from observation. Fidelity improves with more data and more compute, Bitter Lesson style. And we’ve trained simulators on only a tiny fraction of the observations the world has to offer.

To me, our recent releases are wind tunnels.

Praxis-1 is a wind tunnel for robot policies. Solaris is a wind tunnel for computer-use agents, which are notoriously bad at simple tasks like booking a hotel, partly because they learn the specific layouts they were trained on and fall apart when the next hotel’s website looks a little different. GWM Worlds 2 is a step toward a general wind tunnel: one model handling navigation, manipulation, and human interaction. In one demo, we generated a world with a humanoid robot and two flags, told an agent to walk the robot to the red flag and then the blue one, and it did it on the first try.

David Silver and Richard Sutton call this next phase the Era of Experience. “Agents will inhabit streams of experience, rather than short snippets of interaction,” they write. Their example is AlphaProof, which started from roughly a hundred thousand formal proofs written by humans and went on to generate a hundred million more through its own interaction with the proof checker. Most environments don’t work like that, though. You can’t crash a car a hundred million times.

You can simulate it, though. If I’m able to predict 1,000 different versions of a car crash and I can actually see it, I can learn a lot, and a system and a robot can learn a lot about how to prepare for that.

There’s a catch, though. Agents cheat.

In a foundational 2018 paper, Ha and Schmidhuber asked whether agents can learn inside their own dreams. But when the researchers made the dream too predictable, the agent learned to exploit it. In their Doom-like game, an agent trained in a low-randomness dream survived for 2,086 steps inside the dream and just 193 in the actual game. It had found glitches in the dream that didn’t exist in reality. When they turned the randomness up, the agent survived 1,092 steps in the real game, well past the 750 that they counted as solving it. A dream that’s too easy to game teaches you to game dreams.

A useful World Model has to stay accurate when an agent does something rare, something that barely shows up in the training data, or when it actively tries to exploit the simulator’s loopholes. That’s why real-world deployment, and a continuous loop of updating both the policy and the environment, remain necessary. You can’t game physics.

We’ve built this into our work in robotics already. Our next chapter, Chapter 4, is Science.

Going beyond human-scale observation means feeding models data from other scales of space and time, plus scientific data. I imagine a single model that spans physics, biology, and everything in between, at scales from the subatomic to the observable universe, and from femtoseconds to the lifespans of stars. It’s prudent to start with related modalities. But our ambition is to tackle the whole thing.

You may be asking yourself, “Why would a model that has watched a lot of cat videos be any good at fluid dynamics?” Good question.

It sounds like a joke, but there’s growing evidence that it isn’t. Pretraining on language helps models generate speech. Pretraining on code helps with broader reasoning. Google’s TimesFM, a forecasting model trained on a mix of time series (traffic, weather, page views, hospital data), performs about as well out of the box on series it has never seen as specialist models trained on those exact series. A group of researchers trained a transformer on continuous glucose monitor and diet data from roughly 10,000 people and found it could predict their health outcomes four years out. Back in 2022, Riffusion took an image model, fine-tuned it on pictures of sound (spectrograms), and got it to make music.

I believe physics is the most striking case. In early academic work with computational physics groups, a video model pretrained on basic videos and then adapted to physics simulations beat a model trained directly on the physics data. Independent research points the same way. Last year, a team at UCLA built PhysiX, a model for physics simulation that starts from the weights of a video generation model. Their problem is the one robotics has: the internet has millions of videos, but the largest physics simulation datasets contain only tens of thousands of samples. Starting from a model that had watched the world helped. PhysiX beat task-specific baselines and the previous state of the art on The Well, a standard physics benchmark, and the authors conclude that knowledge learned from natural videos transfers to physics simulation.

Transfer works for two reasons.

The first is that models trained across many domains get good at learning how to learn.

The second is that there are common patterns across the scales of the universe. Waves are waves, whether they’re in a pond or a plasma. Things diffuse the same way whether it’s ink in water or heat through a pan. Scientific data today lives in silos, every lab and team with its own slice, which means there’s probably a lot of untapped upside in training on all of it together.

The same idea, action-conditioned prediction, is already showing up in other fields. Arc Institute’s virtual cell model predicts how a cell’s gene expression changes when you perturb it. Researchers have built learned market simulators that model how trades move prices and liquidity, and simulated societies that estimate how people might respond to news or a new policy.

This is where I think World Models will ultimately have the biggest impact: anywhere an agent has to act in an unpredictable environment and learn from the data directly, whether that’s a robot, a cell, or a market. Most of the universe looks like this. The TAM is practically infinite.

The endgame of a universal simulator of pixels is a universal simulator of reality.

And in that end state, I think these models will be more useful than LLMs are today.

Simulate Everything

Packy here again.

When I asked Cristóbal to distill Runway’s vision, he shared: “If every pixel will be generated or simulated, we want to become the platform and the engine in which those pixels are simulated in the first place.”

The applications, he said, are impossible to plan that far out. Runway doesn’t have product definitions for next year. “There’s a lot of things that we don’t yet know what we will name and what we will be doing,” Cristóbal said. “Even if you try, it will be impossible to predict.”

Naming things too early can actually be anathema to creativity and innovation. “Taxonomy is sometimes where ideas go to die,” he says.

People try, though! We try to categorize a new thing using the words of the past. When people first saw daguerreotypes, they described them as mirrors with memories. When cameras followed them, people thought of them as paintbrushes, because a paintbrush was the closest thing they had. The internet, at first, was a place for publishers to put their magazines so you could flip through them on a screen and for advertisers to put banner ads. Eventually, each one became its own medium, with its own vocabulary, doing things that their early namers never could have predicted.

Right now, most people put video models in a box labeled “tools for creatives,” which is fair, given where they are today, but which Cristóbal thinks is the latest version of the paintbrush analogy.

“People tend to look at video models as another cool, cute, creative thing,” he said. “It’s a toy for some. You should take it much more seriously. This is going to power most intelligent behaviors very soon.”

“The work we’ve been doing on World Models isn’t widely understood yet,” Anastasis adds. “People generally underestimate how far video can go.”

I, Packy, was probably one of those people, and a few months ago, I was explaining video models as separate from World Models… which they are! What I underestimated, though, was video as a path to World Models, on top of which all sorts of application-specific and embodiment-specific models might be fine tuned.

Runway’s bet is that a good enough video model, given actions, can become the World Model, and that on that same backbone you can render a movie, an ad, an entire desktop, a working warehouse, and maybe, eventually, a cell or a universe. The past six weeks are evidence that the bet is at least partly right. All of these applications sprouting from the same backbone, seemingly in real-time.

The thing to watch from here, to double-check if the bet is right, if Runway is as good at predicting the next frame as its models seem to be, is whether each new pretrain continues to make everything on top of it better and open up new possibilities, whether the next base model makes robots more dexterous, interfaces more coherent, worlds more persistent, videos longer, and ads more performant, all at the same time, without Runway having to build each improvement by hand. And then whether the company, and its users, can build more new things that you wouldn’t expect from a video company.

This is what’s happened with LLMs; while they remain spiky, broadly, each new model gets better at poetry and code and medicine simultaneously. If frame prediction works the same way, Runway’s hot streak will continue far beyond what people, outside of the company at least, are predicting.

The idea, though, is that World Models can go even further than LLMs. That if you can simulate pixels, you can simulate nearly everything.

Near the end of our first conversation, slowly getting simulate-all-pixels pilled, I asked Anastasis the obvious question: are we living in a simulation?

“Whether or not we are, this is a very meaningful universe,” he replied. “Either it’s the root universe, or someone thought it was worth simulating.”

Thanks to Anastasis, Cristóbal, Tenaya, Leo, and the Runway team for working with me on this essay!

That’s all for today. We’ll be back in your inbox with a Weekly Dose on Friday.

Thanks for reading,

Packy

Hi friends 👋,

Happy Friday and welcome back to our 214th Weekly Dose of Optimism!

I took Dev to watch LeBron James’ first preseason game as a Sixer last night, the good guys gave us too many great stories to fit into one Dose (check the extras), and it’s beautiful here in NYC. Life is good.

Let’s get to it.

Today’s Weekly Dose is brought to you by… Guru

Your company’s knowledge is one of its most valuable assets—but only if your AI can trust it. Guru helps organizations create a verified, current knowledge layer that people and AI can rely on.

You’ve probably experienced the frustration of giving AI access to a mountain of company information, only to see it struggle to distinguish an approved policy from an outdated document or an old Slack conversation.

Guru helps companies structure, govern, and continuously improve the knowledge behind their AI. The Knowledge Accuracy Gap guide explains how leading organizations are closing the gap between what AI can do and whether it has accurate company context to work with.

The guide covers how to:

• Correct an AI answer once and apply that correction across connected agents and AI experiences

• Keep people involved where judgment and approval matter, while AI handles routine knowledge work

• Build a trusted knowledge layer that supports both employees and AI

• Improve knowledge quality and governance as AI adoption grows

Your company is already using AI. Make sure it is using the right knowledge—with Guru.

Build AI Your Company Can Trust

(BONUS) KILO: a not boring film on Airbound

Last week, I wrote a Deep Dive on Airbound. In the Before Times, that would have been that. Hit send. On to the next. But we are in the Now Times.

In the last Dose, I told you about our new partnership with Runway, and over the weekend, I decided to see what would happen if I gave my essay, a few references, and some creative direction to Claude/ChatGPT and asked them orchestrate the production of a short film using Runway.

The result is KILO, the story of a package that was batched and pushed around until a kid in India figured out how to let it fly solo.

Again, I’m not great at this yet, but I find it incredibly magical that I can take all of the work that I pour into writing and turn it into a film in just a few prompts. It’s a good time to be a storyteller.

(1) Nobel Prize Picks 2026 Laureates

This week, the Nobel Prize committee began handing out the most coveted prizes in science, starting with The Nobel Prize in Physiology or Medicine.

Karl Deisseroth, Peter Hegemann, and Georg Nagel were awarded Prize for their “discoveries leading to optogenetics, which makes it possible to switch on, or off, the activity of individual nerve cells in a living brain.” Optogenetics is a wild field - by shining light at cells, you can turn them on and off, make them do different things. I had a non-meat burger made using a process that involved optogenetics once and it was delicious.

This video of Diesseroth telling his kids he won has racked up 4.5M views. It’s awesome.

Other winners announced so far are:

• Francis Halzen in Physics “for decisive contributions to the IceCube Neutrino Observatory and the discovery of high-energy neutrinos of astrophysical origin”

• Henri B. Kagan and Kenso Soai in Chemistry “for the discovery of non-linear effects and autocatalysis in asymmetric organic synthesis”

• Anne Carson in Literature “for her bold and inventive oeuvre that, in playful dialogue with the classical tradition, has created new forms for contemporary literature”

Now, I need to figure out what any of this actually means, but I, for one, am happy that we take some time out of every year to celebrate the world’s geniuses, and to let them celebrate their great achievements to humanity.

(2) Atomic Machines Comes Out of Stealth with Matter Compiler

Just in case you thought we might not be living in the future, we are living in the future.

Yesterday, a startup called Atomic Machines came out of six years in stealth with a product called the Matter Compiler: “an AI-native manufacturing system that builds working micro-machines from code alone. No per-product tooling. No process development. Different code, different machine.”

If the name sounds familiar, it’s because a matter compiler is a long-held sci-fi dream. In one of my all-time favorites, Neal Stephenson’s The Diamond Age, he literally calls the machines they use to make everything up to and including entire islands “matter compilers.” The idea goes back further.

In 1959, Richard Feynman gave "There's Plenty of Room at the Bottom" and offered $1,000 for a working motor that fit in a 1/64th-inch cube. He expected it to take a whole new approach to engineering. William McLellan won it in 1960 with a microscope, a watchmaker's lathe, and a toothpick. In 1986, Eric Drexler's Engines of Creation imagined molecular "assemblers" that could "build almost anything that the laws of nature allow to exist." A year later, Star Trek: The Next Generation upgraded the old food synthesizer into the replicator. Biology got here long before humans did: founder & CEO Jeff Holden's essay compares the Matter Compiler to the ribosome, which builds every protein in your body from code.

Now, anyone can say they’re building a matter compiler. But Holden was Amazon's tenth engineer, built Prime, and became Uber's first Chief Product Officer. He's also been watching Star Trek since he was five. He has both the skills and the childhood dream necessary to pull this off.

The first product AM is building with the Matter Compiler is the PrimeSwitch PS-150, a 9.5 mm relay that carries 150 amps and opens in 50 microseconds, ~1,000x faster than a contactor, for the 800-volt DC data centers that AI is pushing the industry towards. No existing chip fab can process its materials or assemble its moving parts, but AM says it can, and is beginning to ship to early customers now.

Micro-machines the likes of which AM intends to build are already all over your life. MEMS, or tiny devices with both mechanical and electrical parts, were a $15.4 billion market across 31 billion units in 2024. They include the gyroscope in your phone, the microphone in a hearing aid, RF filters, inkjet nozzles, and Texas Instruments' micromirror chips that flip millions of tiny mirrors to project movies. But as Holden notes, engineers coaxed those few devices out of a toolset built for chips, and each one takes years of process work to get up and running. Atomic Machines wants to go after a whole new swath of sci-fi products now that they’ve made it easier: gears the size of a grain of sand, robots that work inside the body (a la Kurzweil), cooling built into AI chips, even a device that runs instant blood tests (perfect timing with the Nathan Fielder/Elizabeth Holmes doc coming out next week; Holden argues that her dream was possible, but needed a manufacturing technology that didn't exist yet.)

Holden closes his essay with a line that we might tattoo onto the Dose’s forehead:

“We are in the early innings of everything.”

Related Bonus Note: you know I like the woo. Cyan Bannister, the legendary investor at Long Journey, looked at a chart of AI CapEx spend last week, closed her eyes, and saw two words: MATTER COMPILERS. Then “On Monday, on a random call, I was pitched a company doing just that and we invested. The world works in mysterious ways and stuff is about to get wildly fantastic.”

(3) Anduril Investing $3.7B Into Arsenal-2 Shipyard

One of the biggest open questions in manufacturing is Why Can’t the U.S. Build Ships?

Anduril (disclosure: we are investors) is trying to answer that question with a “We can.” On Tuesday, the company announced that it would invest $3.7 billion into Arsenal-2, a planned two-million-square-foot shipyard at Sparrows Point outside Baltimore. The facility will manufacture components for Virginia-class nuclear submarines and is expected to create more than 3,100 permanent jobs.

The plan is to start with components like torpedo tubes and eventually work up to larger submarine sections, supplying the existing shipbuilders. According to officials who spoke with USNI News, torpedo tubes are already holding up submarines under construction.

As part of the deal, the Navy has awarded Anduril a contract worth up to $2.9 billion, with payments tied to demonstrated production outcomes. As is its style, Anduril is committing the capital to build the capacity, and it will get paid as it delivers.

Anduril calls Arsenal-2 a “software-defined shipyard,” which means it will use software to design how work moves through the factory, automate where possible, and keep detailed digital records down to individual welds. Production in Maryland is targeted for 2030, with an earlier start at a California facility planned for 2028.

Sparrows Point is symbolic as well as practical. Back when the US built lots and lots of ships, and back when Bethlehem Steel existed, the company made lots of ships there. And while making nuclear submarine components won’t solve all of our shipbuilding needs - we can make them, just not enough of them or within budget, and this doesn’t address commercial vessels - we love to see any progress on the effort and lots and lots of good American jobs.

(4) Privo Patch Helps Oral Cancer Patients Avoid Surgery

Here at the Weekly Dose, we’ve long taken the brave and controversial stance that cancer is bad and should be eliminated. We can do it. It’s not that we don’t have the technology to kill cancer cells, it’s usually that it’s really hard to kill just the cancer cells, to target them and only them for destruction.

Well, score another point for the good guys. On Monday, Privo reported results from a small Phase 2 study of PRV111 in patients with non-invasive oral cancer or high-grade precancerous lesions. Basically, instead of getting surgery to remove the cancer from their mouths, they put a patch directly on the cancer to deliver chemotherapy.

After treatment, biopsies showed complete clearance in 21 of 22 patients. The remaining patient’s lesion improved to low-grade dysplasia. All 22 avoided their planned surgery, and no recurrences had been observed in the treated areas at a median follow-up of 14 months.

The patch delivers cisplatin, an existing chemotherapy drug, directly into the affected tissue while keeping exposure elsewhere in the body low. Privo reported negligible systemic exposure and no serious treatment-related adverse events.

The usual Phase 2 caveats apply: it’s early, it was a small group of 22 patients, more work needs to be done on how long the benefits last, etc… BUT getting mouth surgery sucks; it can harm your ability to speak, swallow, and eat. Sometimes, people wait too long and they have to get their whole jaw removed. If we can replace that with a chemo patch that completely clears mouth cancer in 21 of 22 patients, hell yeah.

Say it with me, y’all…. get fucked, cancer.

(5) Biohub and Others Commit $1.8B to Data for Virtual Cell

Biohub; video created with Runway

ZUCCKKKKKKK!

On Wednesday, Mark and Priscilla’s Biohub announced an expanded $1.8 billion collaboration to build the data needed for predictive models of cells.

Google DeepMind, Isomorphic Labs, and Meta are collectively investing $300 million, joining Biohub’s existing commitment and contributions from the Department of Energy and NIH. The total includes computing and existing data resources alongside funding.

The idea behind virtual cells is to understand cells well enough that scientists can ask what would happen if they changed something and get a useful answer from a computer, is about as well as I can explain it, but if you want a great understanding, I highly recommend checking out ’s excellent What Are Virtual Cells?

The Century of Biology

What Are Virtual Cells?

Welcome to The Century of Biology! This newsletter explores data, companies, and ideas from the frontier of biology. You can subscribe for free to have the next post delivered to your inbox…

Read more

a year ago · 118 likes · 11 comments · Elliot Hershberg

Virtual cells are something of a holy grail. They should allow researchers to try thousands of possible interventions digitally, at computer speed, take the promising ones into a physical lab, and accelerate towards real treatments more quickly.

One of the hard parts is that making virtual cells requires measuring an enormous number of real cells and how they respond to different conditions. To that end, Biohub and its partners plan to build shared datasets that researchers can use to train and test their models.

Look, if you’re not paying for the product, you are the product, and Meta has made a lot of money off of you as a product, but if a lot of that money goes back into helping you stay alive and healthy for longer, then… wait a second, Zuck is a genius.

EXTRA DOSES: Science Breakthroughs, Victorian Elites, Your Brain is Not a Computer, Biggest Questions in Science, US Govt Science Initiatives, Rejuvenated Eye Cells

Read more

Hi friends 👋,

Happy Friday and welcome to our 213th Weekly Dose of Optimism.

You might notice that the title card looks a little different than it normally does, and you might also notice that our fifth story comes with its own short film. Good eye. Over the coming quarter, we’re partnering with our friends at Runway to bring the stories we tell here in the Dose and in not boring to life. I am… not a filmmaker, but I am a writer, and it’s wild to be able to make something that feels like a pro video in just a few prompts, and I’ll get better at it as we go.

If you haven’t yet, check out our Deep Dive on Airbound. We are going to fly.

Let’s get to it.

This week’s Weekly Dose is brought to you by… Reducto

The Only Document Parsing Product You’ll Need

Reducto is the document ingestion layer trusted by AI teams at Scale AI, Harvey, Toast, and Vanta to reliably turn complex files into structured, grounded data for production AI systems at scale. Processing more than a billion pages per month, Reducto has built its reputation on the long tail of difficult documents that break conventional OCR: dense tables, unusual layouts, handwriting, low-quality scans, strikethroughs, and formatting-dependent information.

Its new frontier parsing model, r-1, combines layout detection, reading order, tables, formatting, grounding, and granular citations in a single request. It reduces parsing errors by up to 20%, improves latency at high volumes, and costs 1¢ per page all-in.

Already using another parser? Apply for up to $5,000 in migration credits and hands-on support to benchmark r-1 against your current stack on the documents that matter most.

Apply at Reducto.ai/migrate

(1) Starship Goes Orbital (and Deploys Starlinks)

Fourteenth time’s the charm!

On Monday, Starship reached orbit for the first time and once there, spat 26 Starlink V3 satellites out of a Starlink-deploying slot.

Starship’s previous flights reached space but followed suborbital trajectories that brought them back down. This time, Starship fired an engine to enter orbit, released its cargo, and later fired an engine to leave orbit before splashing down.

Were there some things that need fixing? Of course there were. An engine shut down early during ascent, and SpaceX shortened the mission as a precaution, bringing the ship down at a prearranged location in the northern Pacific.

But they keep getting better. They’ll fix the engine and catch thousands of these Statue of Liberty-sized rocketships with giant chopsticks, make them rapidly reusable, and keep sending them up to blanket the world with internet, and then maybe data centers and power and telescopes and drug factories and whatever else might make sense to move to space when launch is cheap and constant.

In the meantime, just take a look at these beauties.

(2) National Design Studio Unveils America.gov

Joe Gebbia, RISD grad and co-founder of Airbnb, runs the US Government’s National Design Studio. It shows. The government made a good website, a great one even, in “the first step toward replacing thousands of outdated government websites with one product.”

For now, you can ask questions about government services and get answers grounded in government sources. GSA says it brings together information from more than 29,000 government websites, which is an absolutely ridiculous number of government websites.

Adam Bhaloo, who’s on the NDS team, tweeted that “Next, we're building features that let you renew your passport, enroll in Medicare, and book a campsite at a national park, and so much more, all in one place.”

I can’t wait. I have a friend who’s planning a trip to Paris with a bunch of family members and is stressing out about being able to get her passport renewed in time. The process is ridiculously complex; I once had to sleep in my car in Washington, DC to get a new passport for a trip. I hope they bring this miracle technology (modern websites) to the state-level next so that I never have to interact directly with the DMV again, either.

The government works for us, and we should be able to tell the government what we need and have its software figure out how to get it done.

God Bless America.gov.

(3) Matic Unveils Master Plan for Family of Home Robots

A couple of years ago, Matic co-founder & President Mehul Nariyawala came to my house in Brooklyn on a Saturday to help set up an early version of the Matic in my house. He brought his laptop, too, and he walked me through the company’s master plan.

Build an incredibly useful robot in a big market, use that to learn and generate revenue, and then build on that learning to build more in-home robots that can do more and more useful stuff to give people their time back.

Over the past couple of years, they’ve executed like crazy on the first part, putting robots in 16k homes covering 500M square feet. This week, he unveiled that plan to the public.

The basic argument is that starting with robots that can do everything it the wrong order, and that instead, you need to start with what it can do really well, collect data, learn how to navigate a home and all the edge cases in it, earn trust, and then grow.

The next phase for Matic will be to add manipulation, to give robots the ability to act on the things they’ve learned to see. The video shows both a non-humanoid looking thing with arms picking things off the floor, and a humanoid folding laundry. Presumably, they’ll build the former first, and then the latter.

As someone who would rather spend time with my family than doing chores, this vision can’t come to life soon enough.

(4) Freckle Launches First Kids Phone Built for Real-World Adventure

Speaking of family-friendly machines…

I’ve been in California this week, away from my kids, and whenever I’m away from them, I get the urge to call them. Of course, I can call Puja and FaceTime with the kids, but there are random times throughout the day when I think, “I wish I could just call Dev or Maya right now.”

But what are we going to do? Give our kids cell phones and get them hooked early? No shot.

Sam Terris is trying to change that with Freckle, a phone for kids that will cost $199 plus a $25/month cell subscription.

Freckle runs its own ClayOS, with parent-approved calls and texts, maps, and tools for exploring. Kids can photograph something with it and then ask AI to explain what they’re looking at. Parents can set up real-world “Quests” for their kids too. Freckle doesn’t have open web, social media, or an app store, so there isn’t a risk of kids getting sucked in (unless they figure out how to hack it, in which case, good on them).

In the launch video, it shows two kids bumping their Freckles to turn them into walkie-talkies, which I would have gone bananas for as a kid.

The philosophy seems to be to help kids get outside and explore - they have a map, a way to ask questions about the world, and a way to keep in touch with parents so parents don’t have to keep their eyes on them at all times. It’s just a pre-order phone, and there’s a lot of societal cruft to be figured out, but it seems like a good step back towards letting kids be kids.

(5) Transplanted Hearts Adjust to Host’s Biological Age

Nature; video created with Runway

Saving the wildest story for last…

Researchers at Harvard and Brigham and Women’s Hospital found that transplanted hearts shifted toward the biological age of their recipients. Older hearts showed signs of rejuvenation in younger bodies, while younger hearts showed accelerated aging in older ones.

They first studied mice, using patterns of chemical tags on DNA to estimate the hearts’ biological age and found the effect. Then, they examined biopsies from 11 human heart-transplant recipients with substantial age differences between themselves and their donors and found a similar pattern. In a larger analysis of patient records, several measures of function, including exercise capacity, were associated with the recipient’s age rather than the donor’s. Which is nuts!

Of course, this is a preprint on a small human biopsy sample, and changes in an aging clock don’t mean all of the heart’s accumulated damage has disappeared. The mouse experiments also didn’t show that receiving a young heart rejuvenated the rest of the body, just the heart itself, which makes sense.

But come on! This is not a result I would have expected, and it has both practical and more profound implications.

Practically, it might mean that doctors can use older donor hearts in more patients, which would expand the donor pool and save more lives.

Profoundly, it implies that some aspects of an organ’s age seem responsive to the body around it. Researchers may be able to use what they learn here to understand what it is about younger bodies that rejuvenate older organs and potentially reproduce whatever it is to reverse aging more broadly.

May we all be young at heart.

EXTRA DOSES: Science Breakthroughs, AMD Buys World Labs, General Intuition and Varda Raise

Read more

Hi friends 👋,

Happy Monday! Today is a good day.

Today, I get to tell you the story of Naman Pushp and Airbound.

There are very few people with the combination of ambition and skill than I’ve seen in Naman in the year since I first invested and have gotten to know him. Just twenty-one, he taught himself physics, engineering, and manufacturing on YouTube during COVID, and turned that knowledge into real, flying aircraft and a company backed by Greenoaks, Lightspeed, and Lachy Groom. I suspect he is in the early days of what will be a decades-long pursuit.

He has chosen a worthy mission. There is no category that so physically reshapes the world as transportation technology. Where you live and work, what you can access, how much you pay for everything, and most importantly, who you spend time with, are all a function of the transportation technology at your disposal.

In my opinion, improving the way we move will have a more positive and noticeable impact on our lives than AI will, and Naman’s aim is to move all movement to the sky.

This is a Deep Dive on how he plans to do it.

Airbound: As We May Move

Every morning, or most, when I can, I wake up before my wife and kids, tip-toe downstairs, ease open the front door, and sit on the stoop as the sun comes up. Huberman says the morning sunlight is good for your circadian rhythm or cortisol or some shit.

I’ve been doing this for a little while now, since the winter at least, but some time in the spring I noticed the birds. There’s a honey locust tree on the sidewalk in front of the stoop, framed by Eric Adams’ legacy (new garbage cans) every Monday and Thursday, and as the sun rises, a quarrel of house sparrows flits back and forth between it and their little nooks at the top of each brownstone.

Since I started noticing these birds, I’ve started noticing other birds more consciously. I think, but I can’t be quite sure, that I can see their wings flapping in higher-resolution, or slower motion, or both? And I’ve been wondering why. Like are the birds a sign? Am I supposed to follow them? How can I follow them, when I can’t fly?

One morning, as I approached my office in Soho thinking about this very piece you’re reading right now, a great blue heron floated regally over the roofs of the block’s cast-iron buildings, and I was like, “Oh, I get it.”

Some time in the spring, I started talking to Naman Pushp about writing a deep dive on Airbound. Pretty much every week since, often more than once, we’ve hopped on a Zoom, me in Manhattan and him in Bangalore, to talk about what the world looks like when flight is the way everything and everyone moves, and what it would take to get there.

One of the things that happens when flight dominates, Naman said, is that buildings flip.

“You fly on to the roof and enter from there,” like a bird.

Birds have long been the envy of and inspiration for earthbound humans.

Leonardo da Vinci studied birds to understand flight at the turn of the 16th century.

Humans failed to turn those avian observations into flight for four centuries, but the Wright Brothers studied those failures. As bicycle makers, they understood that a machine didn’t have to balance itself if its rider had the right controls. They studied the birds, too. Watching birds twist their wingtips helped them develop “wing warping,” the control mechanism that helped make their 1903 flight possible.

“Learning the secret of flight from a bird was a good deal like learning the secret of magic from a magician,” Orville later said. “After you once know the trick and what to look for, you see things that you did not notice when you did not know exactly what to look for.”

Flying progressed quickly from the Wrights to October 1958 when Pan Am flew the first Boeing 707 from New York to Paris and ushered in the Jet Age, an age that we, 68 years later, still inhabit.

That we can fly through the air, across the globe, while watching movies and munching on snacks, is a technological miracle. It has shrunk the globe, made roses cut in Kenya available in Brooklyn flower shops, and allowed Naman to fundraise in New York and San Francisco without permanently leaving Bangalore.

“Like seriously what in the ever-loving fuck are planes,” Naman marveled in an old blog post. “I’m an aerospace engineer and I still refuse to believe that shit is anything short of magic. Like there’s genuinely no other way to put the accomplishments of humanity other than just… sheer magic.”

And yet… humans still do not flit about in the air as birds do. Planes are much too big for that, for one thing, and too expensive to operate, for another. Daily, we are roadbound, grounded to the streets by feet and wheels, stuck in traffic on streets built for cars at the expense of humans.

Naman wants to change that, by sending all movement to the sky, with aircraft that borrow from birds not just in their wings, but in the shape of their whole bodies.

To do so, Airbound must beat existing vehicles on speed, versatility, and cost. All three are required for victory.

How quickly can you get somewhere? Can you go from where you are to where you want to be, when you want to go? What does it cost?

“Our goal is to build vehicles that dominate1 across these axes,” Naman wrote in a recent technical memo. “If an aircraft can be more versatile than the car, faster than the train, and cheaper than the bus, then you can replace all of them. That is the universal vehicle we intend to create.”

An eVTOL (electric vertical take-off and landing) naturally falls out of these requirements. It takes off and lands straight up and down like a helicopter, so it doesn’t need a runway (versatile), it flies over traffic in a straight line (fast), and it runs on electric motors, which are cheaper to power and maintain than engines (cheap). Every delivery drone you’ve seen is an eVTOL, as are the air taxis Joby and Archer are building. Others are making eVTOLs, too, so while they will all be fast and versatile, Airbound must beat them on cost.

That seems crazy, when the competition has so much scale. Amazon - Amazon! - the company that is the best in the Western Hemisphere at moving goods to people - Amazon is making eVTOL delivery drones.

Last week, The New York Times wrote a story on Amazon Prime Air’s nascent operations in Richardson, Texas: It’s Raining Boxes: Amazon Drones Overwhelm a Texas Suburb. Describing the logistics king’s flying machine, Juliet Macur wrote, “Within a minute or so, an 83-pound drone with six blue arms and whirring propellers appeared in a corner of the sky before zipping over nearby houses, looking like a cross between a mechanical dragonfly and an X-wing fighter, and buzzing like a battalion of bees. The drone’s mission: delivering an Amazon package. The single cardboard box of cargo held a maximum of five pounds...”

Six arms and 83 pounds to carry a maximum of five.

Before I met Naman, I would have read that and thought, “Woah, the sci-fi future has arrived,” but after meeting him and hearing what he’s building and why, those numbers sound absurd.

To be fair, Amazon’s 83 pounds is fully-loaded, so it’s really 78 plus the five pound payload, for a payload-to-empty-weight ratio of 0.06x. Zipline has a much stronger ratio; its P2 weighs 55 pounds and can carry eight, good for 0.15x.

Airbound’s is in a different league. Its first drone, the TRT, is so light (1.5kg) that Naman could hold it with three fingers, and carries 1kg, for a payload-to-empty-weight ratio of 0.67x.

Its new V2 drone can carry five kilograms (11 pounds). Its payload-to-empty-weight ratio is 1.67x. It can carry more than it weighs.

It is not that Amazon doesn’t have resources or smart engineers. It’s just that Naman is obsessed with building aircraft that carry the most payload with the least weight, because he believes that combination is what will allow Airbound to offer the cheapest deliveries possible (fast and versatile, too), which is what he believes it will take to win.

“We believe that if we can truly solve for cost,” Naman wrote, “flight can dominate all existing transportation for 80% of all cargo and 90% of all human movement.”

If they can – and that is certainly not guaranteed – then not only will they move movement to the sky, they will force the earth to shape itself around this new mode of movement.

As Naman wrote in Airbound’s Manifesto, this world is “A world where goods move as freely as information does today; a world where flying is cheaper than taking the bus; a world without compromise.”

Naman’s Vision

I first heard of Naman when my plane landed in Mexico City for an event last October. When my international service switched on, I saw an email from Lachy Groom, who had invested in Airbound and called Naman “One of the most exciting founders I’ve backed recently.” Naman was going to be in NYC the next week, “Want to meet (I highly encourage)?”

I wasn’t particularly interested in a drone delivery company, to be honest. It was a competitive space in which it felt like Zipline had an insurmountable lead and the culture to keep it. Amazon and Google waited in the wings anyway. Regulation is tough. And Naman didn’t seem to have a PhD in aerospace engineering, or a college degree, for that matter.

But I respect Lachy’s judgment, Airbound’s monthly flights were going vertical, and I was intrigued by Naman’s simple explanation - “we’re building the most efficient way to transport atoms” - and ambition - “we can make 1 cent deliveries possible and enable a world where you see trillions of deliveries annually,” so I said I’d love to.

When I met Naman, at Felix Roasting back in Soho, 8,306 miles from Bangalore as the crow flies, he looked young, because he was, just 21 years old, which made it even more embarrassing when he opened his mouth and proved himself to be about 8,306x smarter than me.

In that first conversation, he showed the grasp of the spectrum from 30,000 foot vision to sub-millimeter details that great Vertical Integrator founders seem to possess, and which, I’d come to appreciate, are necessary to hold in one brain when you want to solve a tightly coupled optimization problem like building the “perfect aircraft.”

By the end, he’d convinced me it might be possible to make flight the cheapest mode of movement, inevitable even, that whoever did it would build an enormous business, and that he might just be that whoever.

I invested in Airbound’s $37M Greenoaks-led Series A, and my conviction has grown as I’ve spent time with Naman over the past year and come to better understand what he is trying to achieve, and how.

The idea behind Airbound is really simple: if you move mass more cheaply than anyone else, you win.

“There is one loss function that we care about: cost per kilogram-kilometer,” Naman told me. The total cost of making and operating the aircraft, divided by how much payload it moves and how far it moves it. “Everything is secondary to that.”

Those kilograms might be a burrito, a box of company t-shirts, building materials, or even people. Transportation and logistics are both forms of movement, and Airbound is a Movement Company.

Today, everything that moves needs a human chaperone, and chaperones are expensive. A bus is cheaper than a taxi per-person because it’s cheaper to have one driver move 20 people instead of two. A truck is cheaper than a car per-package because it’s cheaper to have one person move 1,000 packages instead of one. So humanity built its movement networks around batching.

Autonomy removes the onboard chaperone, so the vehicle can be sized around what it carries. Airbound’s bet is that a light enough aircraft can make small, frequent shipments economical. It will start with a small aircraft, where experiments are cheap, and use what it learns to build larger aircraft that move each kilogram, including humans, more cheaply still.

Scale that up or down, and most of movement eventually becomes point-to-point, in the air, carried by billions of aircraft, which Naman thinks, given the astounding size of the sky, will be no more noticeable than birds2, which is to say: barely noticeable unless you choose to notice, from time to time, and marvel.

Birds or Airbound Aircraft? (Airbound Aircraft)

Because if Airbound really succeeds in its mission, those bird-like dots will reshape the world.

Reshaping the World

I grew up outside of Philadelphia, America’s first capital and once-largest city. I love it there. Go Birds. But I don’t live there. I am writing to you from New York City, where I live and work, because of the mad ramblings of a man named Jesse Hawley under the pen name “Hercules.”

See, Hawley was a flour merchant who milled his wheat upstate in Seneca Falls, New York, and he had a devil of a time getting it to the big market of New York City. There was no direct route to the Hudson River from Seneca Falls. He and his partner went bankrupt, and back then, going bankrupt landed you in debtors prison. There, with nothing but time and what-ifs, Hawley as Hercules wrote a series of fourteen essays, published in the Genesee Messenger in 1807 and 1808, making the case for a “Genesee Canal.”

President Thomas Jefferson called the scheme “little short of madness.” In 1817, President Madison denied federal funding. New York Governor DeWitt Clinton, on the other hand, took it up and went it alone. Opponents called it “Clinton’s Ditch.”

Seventeen years after Hercules’ letters, his ideas took physical form. On October 26, 1825, Governor Clinton and a party of dignitaries boarded the Seneca Chief in Buffalo and led a flotilla of horse-drawn boats east across the state on the brand new Erie Canal, and on November 4, off New York City, Clinton poured a keg of Lake Erie water into the Atlantic, a ceremony they called the Wedding of the Waters.

The Canal connected Buffalo, the gateway to the Great Lakes, to the Hudson River, and from there to New York City and the Atlantic Ocean. Before the Canal, it cost $100 and three weeks to move a ton of grain from the midwest, where it was grown, to New York City, where it was consumed or shipped overseas. After, it cost $10 and took six days. That meant that a lot more grain came to New York, which meant that ships bringing immigrants and manufactured goods from Europe could bring grain back to the Continent instead of returning on a costly empty dead leg. That meant that most of the westbound ships came to New York City instead of Philadelphia, Baltimore or Boston; it was the economical thing to do. And that reinforced and grew New York City’s lead as America’s leading city.

In 1820, New York ranked third among American ports in flour. By 1827, it had taken the lead. By 1860, its population had climbed from 124,000 to 814,000, nearly 250,000 more than Philadelphia’s. Everything organized around the people and the trade, and today, New York City is America’s financial capital, the place where a wide-eyed new college grad like I once was just had to move, largely due to the Erie Canal.

This is a very old pattern. Cities don’t typically grow their own calories, so their size and influence has historically been capped by how economically they could get calories to come to them.

Rome was likely the first city to support a million people, despite the fact that Rome itself couldn’t grow nearly enough to feed a million mouths. It did so by shipping somewhere between 150,000 and 400,000 tons of grains a year across the Mediterranean, most of it from Egypt and North Africa, in a vast logistics system called the cura annonae (“care of the grain”). Based on Diocletian’s 301 AD price edict, scholars estimate that for the same weight over the same distance, wagon transport was about 50x as expensive as sea transport.

No European city would hit a million residents again until London did at the turn of the 19th Century, fueled by coal. Just like calories, cities need to import energy. A wood-burning city’s scale is capped by how much forest it can reach by cart, and London had chewed through its hinterland by the 1500s. So it started burning coal from Newcastle, 300 miles up the coast, carried down by a fleet of fat, flat-bottomed colliers that by 1800 were landing something like a million tons a year on the Thames.

I can keep going.

In America, as the Erie Canal cemented New York’s advantage on the east coast, Chicago became the country’s Second City thanks to the railroad. Lee Kuan Yew is rightly admired for his work turning tiny Singapore into a prosperous nation, but the shipping container deserves just as much credit3. And just as Singapore emerged from the swamps thanks to the movement of goods, Dubai emerged from the desert due to the movement of people, on planes like the Boeing 777.

Venice had the galley. Amsterdam had the fluyt. Los Angeles had the car. Atlanta, like Dubai, had the jet.

Did you catch that? One of those is not like the others.

The galley, the fluyt, the canal boat, the grain ship, the railroad, the container ship, and the jet are all centralizing batch machines. They’re expensive to build and expensive to crew, so they only make economic sense when they’re full, and they only stop where there’s enough cargo or enough people to fill them. That means they make points – ports, junctions, stations, hubs – around which cities grow. Rome, New York, Chicago, Singapore, and Dubai are all, at bottom, places where big batches come and get broken into smaller ones.

Even when batch machines spread a city out, they spread it along lines. The streetcar suburbs that grew up around American cities at the end of the nineteenth century reached out in fingers along the tracks.

The car is different. It carries one family or one person, leaves when that person or family wants to leave, and can go anywhere there’s a road. It doesn’t make points or lines, but area. Ford made that vehicle cheap enough for almost everyone, and over the next few decades, the space between the fingers filled in. Los Angeles grew, as did countless American suburbs previously too inconvenient for those who wanted to work where batching machines agglomerated economic activity.

Transportation reshapes the world in two directions. Batched vehicles pull people and activity toward the places where those vehicles land. Individual vehicles let people spread out. The map of the world’s biggest cities is the result of both: great cities formed where the batch machines met, then spread as far as a car could take people in half an hour.

It is notable, then, that despite all of the technological progress of the past half-century, very little has changed on the transportation side. Planes have gotten bigger and more efficient. Cars are beginning to drive themselves. But there has been no great leap. Peter Thiel’s quip that “We wanted flying cars, instead we got 140 characters” is drastically overused but appropriate here.

Flying cars would expand the radius of daily life, as Vight founder Tsung Xu wrote in not boring earlier this year.

In 1934, Lewis Mumford, paraphrasing Bertrand Russell, noted that “each improvement in locomotion has increased the area over which people are compelled to move: so that a person who would have had to spend half an hour to walk to work a century ago must still spend half an hour to reach his destination.” In a 1994 paper, Anthropological Invariants in Travel Behavior, Cesare Marchetti formalized the idea into what others would call Marchetti’s Constant: humans like to commute about 30 minutes each way, or about an hour a day, and the further we can travel in 30 minutes, the further we will.

For a while there, further we did! Per Marchetti’s study of Berlin, people walked about 2.5km both ways in 1800, then horses pulled us 4.5km starting around 1850, then electric trams, subways, and cars each contributed to growing the radius to about 20km by 1960.

Then we stopped expanding. Distance traveled is just about the same now as it was in 1960, and may have actually shrunk since then due to increased traffic.

Progress in the speed with which we move things is arguably even worse. Container ships, for example, travel more slowly today than they did during the Nixon Administration in the name of fuel efficiency. Planes fly about the same speed as they did back then. Freight trains are a little bit slower.

Of course, the average consumer does get their orders more quickly today. You can order almost anything on Amazon and get it tomorrow. That’s because Amazon, and those who have followed suit, have invested a tremendous amount of capital building distribution centers closer to where customers are. When I run to Red Hook, I run past block-sized Amazon DCs that should be houses or restaurants. Per the NBER, Amazon went from eight warehouses across the country to 54, and cut the average package’s trip from 450 miles to 141 miles. Then, they added DCs like the ones in Red Hook where items are sent from fulfillment centers in bulk to be carried the last few miles in smaller batches by vans or silly little bike/car hybrid things, which then block traffic on my street or nearly hit me when I cross the street, respectively.

The car (and its sometimes-goofier logistics-specific offspring) explains this, too. It unbatched people by letting the human cargo drive itself. But a package can’t drive, so for things, we’re still stuck with the batch.

Breaking Batch

Good technology eliminates trade-offs. Airbound wants to eliminate the trade-off between cost and speed in the movement of matter.

Imagine that you absolutely need to get an item from where it’s made to its final destination as quickly as possible. A pen, let’s say. A Namiki maki-e fountain pen, lacquered and gold-dusted and encrusted with diamonds, finished by hand at Pilot’s factory in Hiratsuka, an hour southwest of Tokyo on Sagami Bay. For some reason, this pen is the only pen with which two parties in a Park Avenue conference room can sign a $10 billion agreement, and for another reason, the deal explodes in 18 hours. The artisan sets down her brush at 11:00 AM Tuesday, Tokyo time. It is 10:00 PM Monday in New York. You have until 4:00 PM Tuesday. Obviously, price is no issue. What do you do?

The courier takes a car fifty minutes to Haneda, where he boards a Bombardier Global 7500 and flies for eleven hours and forty-five minutes over the Bering Sea before landing at JFK. Teterboro has a 100,000 pound weight limit, and while the plane would come in under after burning its fuel, the Port Authority needs a weight certification seventy-two hours in advance and you don’t have seventy-two hours, so JFK it is. The courier walks the case, still strapped to his wrist, through customs, which takes twenty-five minutes, and boards a helicopter bound for the city. A few minutes later, he and the chopper land at the 34th Street Heliport. He takes a car from the Heliport to the office on Park Avenue, gets out, enters, checks in with security, takes the elevator up to the thirty-seventh floor, and is escorted into a conference room, where he unlocks the case, hands over the pen, and watches the contract get signed at 1:30 PM New York time, with two and a half hours to spare.

The whole trip costs around $430,000 for a 0.04 kg pen over 10,900 km, or $1,000/kg-km. You can get something anywhere you want pretty fast as long as you are willing to pay.

This is, of course, a ridiculous example, meant to show that the cost to move things exists on a spectrum. You can move something fast if you send it by itself, with its own chaperone and its own vehicles. If you want to pay less, you need to send it with other items. You could send the pen from Japan in the mail for ~$10, in which case it will share a commercial plane with 300 passengers and a bunch of other mail. You could send it even more cheaply if you put it in a container on a ship with 300,000 Pilot pens; it would cost a few cents, but it would take a couple of months.

If you want to send something cheaply today, you need to batch it.

This is how our logistics system is designed, whether global or local. It is built around moving the people who move the goods, which means aggregating as many items per person as you can reasonably manage. The further you are from the destination, the bigger the batches that make sense; you wouldn’t tug a ship to each customer’s door. So we’ve built a series of smaller locations to break big batches down into smaller ones: hubs, warehouses, distribution centers, fulfillment centers, sort centers, depots, and last-mile vehicles.

Modern logistics is an exquisitely tuned machine that has squeezed nearly every ounce of optimization out of the existing possibility space. Given the transportation technology available, routing technology has pushed the efficient frontier out as far as it can go, and logistics companies exist to let customers choose from any point on that frontier.

Every vehicle sits somewhere on that frontier, making a different set of trade-offs between speed (and versatility, which impacts factory-to-door speed) and cost. Trains can be fast and cheap, but they only go where the tracks do, and stop wherever the stations are. Cars take you from your driveway to your destination whenever you’d like, but spend much of the trip negotiating other cars. A truck can move goods cheaply as long as it’s full. Bikes are slow, but they’re cheap and versatile. Planes are really fast, but they’re not cheap and they have to land in airports. That’s why the economy makes so many different types of vehicles to do what is fundamentally the same job: move matter, living or dead.

But we are still very limited.

Batching means waiting for the other items in the batch before moving. A restaurant’s delivery driver may need to wait for four meals to finish cooking instead of delivering each meal as soon as it’s ready.

It also means pre-positioning with expensive, fixed, CapEx-heavy locations closer to customers, which costs rent, inventory, equipment, and salaries, all of which get amortized into the cost of the item the customer purchases, and may force further trade-offs, between speed and getting the exact item you want. A retailer may offer limited SKUs for same-day delivery and only make the full catalog available in two to three days.

To solve this problem and make logistics both fast and cheap, you need a transportation technology that is cheap, fast, and free of batching, like the one Airbound is building.

The path to such a technology mirrors that of the path from the steam carriage to the automobile.

You Can’t Replace a Horse with a Steam Engine

When we talk about the history of the car, we often discuss the 23 years from Karl Benz’s three-wheeled Patent-Motorwagen in 1885 to Henry Ford’s Model T in 1908.

Benz’s invention, however, arrived over a century into humanity’s quest to replace the horse with a machine. Before internal combustion engines, engineers used steam engines.

In 1769, a French military engineer named Nicolas-Joseph Cugnot built a steam-powered vehicle intended to move artillery. The project was abandoned two years later, but Cugnot had demonstrated something extraordinary and unprecedented in human history: a heavy vehicle could move itself across the ground, without an animal pulling it.

In 1803, the Cornish inventor Richard Trevithick brought a steam carriage to London. It could carry around eight passengers and required no horse, but its replacement occupied a considerable portion of the vehicle; coal and water rode along with the people, including one person whose job it was to shovel the coal into the engine.

By the 1830s, these experiments were turning into actual transportation services. Walter Hancock, an engineer in East London, ran steam buses on regular routes. His 22-seat Automaton made more than 700 journeys and carried over 12,000 passengers in 1836.

Trevithick and Hancock’s vehicles were novel and inventive, but they were not what we would recognize as cars. They carried eight to twenty-two passengers along established routes at regular intervals.

They were, in other words, buses. They moved humans in batches, because their economics strongly favored batches.

In order to move one person somewhere, that person would have to pay for the time and expertise of the two other people, in addition to the depreciated capex of the machine, its fuel, and its maintenance. Like the ships arriving on America’s east coast before the Erie Canal, the cost of this steam carriage might also have to support an empty ride back home. Those costs are easier to bear if they’re spread across twenty passengers heading in roughly the same direction, and twenty more heading back.

For more than a century, mechanical transportation advanced without making independent, everyday movement affordable. Steam’s greatest successes came from moving more people and goods together. Replacing the family horse and carriage would require a machine that worked economically at the scale of a family.

Enter Benz, who was an engine maker before he was a car maker, and in 1885 created a vehicle with a compact gasoline engine built into a purpose-designed chassis. It had two seats, three wire-spoked wheels, a tubular steel frame, and a single-cylinder four-stroke engine that produced three-quarters of a horsepower. Burning fuel inside the engine eliminated the separate steam boiler and the job of continuously attending to its fire. The engine was small enough that the vehicle it carried could be small, too. Over time, people could even learn to drive it themselves.

Finally, it was practical for an individual to own and operate their own car, to hit the road with no other living human or beast along. It’s just that the individual who owned and operated the car had to be pretty rich. At 600 marks, the original Benz Patent-Motorwagen would have cost a German factory worker a year’s wages; Benz only made 25 of them anyway. It wasn’t particularly powerful, either. It couldn’t, for example, climb up hills.

Over the intervening two decades, automobiles actually got more expensive as they became more powerful. A 1901 Mercedes 35 hp car cost about 15,000 marks. Cheaper versions, like the Curved Dash Oldsmobile and Sears motor buggy, were flimsy and underpowered. Automakers were discovering an efficient frontier of their own, and making different choices on it.

Henry Ford did not invent the automobile, nor did he make the first cheap automobile. When Ford released the Model T in 1908, it was not the most affordable car on the market or even that affordable at all. At roughly $850, it cost something like a year-and-a-half of an average American’s $550 annual pay. What he did, though, was to produce a car that performed well - 20 hp, four cylinders, vanadium steel, enough ground clearance for rutted farm roads - for about half the price of comparable cars, and he did so with careful choices about weight, durability, repair, and manufacturing that would allow him to bring down the cost with scale.

Ford bet on vanadium-alloyed steel – three times as strong as the steel American carmakers were using at the time for its weight – to make the car lighter, and therefore more performant for the same engine, while still being robust to old bumpy roads. He bet on the internal combustion engine starting with his Quadricycle in 1896, while many others pursued steam and electric cars, and reaped the benefits of range and simplicity as the gasoline distribution network built up around him. He bet on a planetary transmission system, which was dead simple for regular drivers to learn, so that owning a car wouldn’t require hiring a driver. In 1909, he bet the whole company on a single car, announcing that Ford would build only the Model T.

As all of these bets paid off, improved, and compounded, he created new points that he alone occupied, far beyond the industry’s efficient frontier.

By 1925, the Model T’s $260 price tag cost only 20% of an average American’s much higher $1,350 annual pay. While its cost improved, its performance didn’t. The 1925 Model T had essentially the same 2.9-liter, 20 hp four-cylinder engine as the 1908 car, the same two-speed pedal-operated transmission, and the same 40 to 45 mph top speed. Performance arguably got worse. That was an explicit trade: the man was maniacally focused on cost to the exclusion of everything else. He offered “any color the customer wants, as long as it’s black,” not only for the simplicity, but because this particular black dried more quickly and thus sped up the assembly line.

Cost, he believed, is what would lure customers away from horse and carriages and to cars.

He believed it so completely that when he announced, in 1916, that he’d stop paying special dividends and pour the profits into lower prices and a bigger factory, two of his own shareholders, the Dodge brothers, sued him. The court sided with the Dodges. Ford bought out the rest of his shareholders so he could have full control, and with it, he kept cutting prices.

The first Model T cost roughly 2x a horse and carriage, which was enough to attract early adopters and begin to pull it down the cost-performance curve. Within eight years, it hit price parity. Four years later, it was about half the cost. Everyone bought a Model T.

There’s this famous pair of images, both taken of the Easter Parade on New York City’s Fifth Avenue, one in 1900, the other in 1913, that shows how thoroughly the Model T dominated horse and carriages.

In just thirteen years, the parade went from one car among a sea of horses, to one horse among a sea of cars.

Naman thinks that aviation has followed an uncannily similar trajectory, on an uncannily similar timeline, and is in an uncannily similar place today as the auto industry was in 1908.

We have had powered flight for more than a century. We have made it extraordinarily good at moving large groups of people and things between places equipped to receive large aircraft. The pilots, ground crews, maintenance facilities, runways, terminals, and schedules necessary to make the whole thing work necessitate a system that becomes economical by spreading its costs across passengers and cargo.

But for daily movement, that arrangement hits many of the same constraints the steam carriage hit.

Even if you could make an ordinary plane dramatically cheaper, it would still need somewhere to take off and land, qualified pilots, and extensive maintenance operations. Its purchase price is only one of the many reasons you don’t have a personal 777 in your driveway. Even a very cheap plane would not replace the car. It is the steam carriage of the sky.

Today’s eVTOLs, to Naman, are the early automobiles. They are capable of autonomous flight and they are electric, which means they can be smaller. Since they’re smaller, they can fit in your driveway, or on your vertiport. It is now practical to replace a human on a bike or in a car with an eVTOL to bring you your burrito, and pending regulatory approval, it is theoretically possible for an individual to own their very own eVTOL. That is just very expensive, even more so than the original Benz. Human-grade eVTOLs cost about $4-5 million to purchase today.

Even at these initial prices, things like delivery and air taxi service make sense. Not having to pay a human chaperone is huge, as is flying right over traffic, as are electric motors. Plus, the electric stack allows for high utilization and requires low maintenance, which lets those who build on it amortize CapEx over so many flights. There are many companies working to take advantage of the novel things the eVTOL stack enables.

There is no one else, Naman believes, who understands how close we are to replacing cars with personal aircraft.

“The two things that make Airbound different,” he explained to me on a recent call, “are, one: I don’t think that anyone else notices that the eVTOL stack is like 90% of the way to replacing cars. They don’t need runways, they can operate autonomously, they’re electric, they can be small. Cost is the last 10%, and it is the most… if you get cost right, everything else will solve itself and you will reshape all movement. If you don’t get cost right, this will just be Aviation++.”

The pieces are there, in his view, “for an aerial vehicle that anyone can own in the same way that the first car [the Benz] brought together the pieces for a car that anyone could own.” The Benz and its successors elsewhere on the same efficient frontier were 90% of the way to replacing horses and carriages. Ford’s focus on cost took the industry the final 10%, and now horses and carriages are the analogy everyone uses to describe something that was completely obsoleted by a new technology.

A flying car that anyone can own sounds as preposterous today as a car that anyone could own did at the turn of the 20th century, but history suggests that we’re on the right track, and that it will come down to cost. So how do you carry aircraft that final 10%?

That’s the second thing that makes Airbound different.

“To get cost right, we realized that carbon fiber is the thing. That we can do so much better with carbon fiber,” Naman gushed. Naman fucking loves carbon fiber.

Naman Learns to Make Drones

He’s loved carbon fiber for his entire adult life, and then some.

Naman grew up across Mumbai, Hyderabad, and Malaysia, where he got his first exposure to robotics via LEGO. By age 13, he started making things with aluminum and steel. He was entranced by his burgeoning ability to turn an idea into a real, physical thing. Then COVID hit when Naman was fifteen, keeping him stuck inside and feeling bored, like a lot of kids. Unlike a lot of kids, he took advantage of all of the free time.

There’s this great David Senra conversation with Ribbit founder Micky Malka that Alex Danco and I included in our piece on HAA, where he talks about the young founders he meets:

The most successful ones right now were the ones that during COVID were learning and were on YouTube sucking information about anything and learning about everything in life. So when they go and build companies, they find that first they’re naive, so they think anything is possible.

That’s Naman.

First, he joined Helpful Engineering, a Slack community of thousands of engineers, technologists, scientists, and doctors worldwide, organized into roughly 40 sub-groups each tackling a different piece of the COVID response in an open-source, share-everything fashion. The main focus of his original group, the Open-PAPR Design team, was respirators, but spending time with grownup makers from around the world gave him the confidence that he could build useful stuff, too. Like drones.

Naman started Opendrone, which would grow into Airbound, as a Helpful Engineering group, formed to create an open-source 3D-printable delivery drone that makers anywhere could build to deliver essential supplies to those in need. To get the drones delivering medicines ASAP, they’d need to work without expensive supporting infrastructure.

The one problem was that Naman, well, he’d built robots but he’d never built drones. No matter. He had all of the world’s knowledge at his fingertips via YouTube, and he gulped it down at 3x speed. He taught himself physics and engineering and aerodynamics and CAD and whatever else he’d need to win the one-week Hackaday hackathon.

The beauty of an open source hackathon is that we can go back and see Naman’s learning in real-time.

In the first design log, which Naman wrote, he and the team had designed their first aircraft, the hexacopter. By the second log, Naman and team had evolved the design, “marking the switch from what people imagine a ‘drone’ to look like, to one more in line with the requirements of a long-range delivery drone.”

“We settled on a tail-sitter,” he wrote, “where the drone would sit vertically, take off, and pitch 90 degrees to transition into fixed wing flight.”

The third iteration, the Blended Wing V2, is the first that looks like what Airbound would eventually build. While the complete redesign “retained many of the characteristics” of the first blended wing model, it featured a “larger middle section, which allows for enough space to store the electronics and the cargo,” based on a 2004 paper from Boeing on blended wing body aircraft.

Note that at this stage (the third image above), it looks very metallic. The biggest note in the fifth and final log, then, from the perspective of Airbound, was one on how they planned to make it and from what: “we had initially begun with a plan to 3D print the drone, then moving on to sheet metal fabrication and finally settling on a composite wet layup process for the body.” While the composite wasn’t yet carbon fiber, he was on the path.

Naman’s design was good enough to win the hackathon, grand prize: $500, which took six months to arrive due to banking issues related to giving money to a teenager. He used the time to keep learning and to keep iterating on the design, until the money hit in early 2021 and he had to actually make a prototype with $500.

“I had to be insanely scrappy,” he told Rahul Sanghi for the awesome Tigerfeathers piece. “I didn’t know anything about aerodynamic principles then. I was just playing around with a bunch of random airfoil designs and seeing what came together. It was more of a 3D modelling exercise than an aerodynamic challenge. Things became trickier because I couldn’t afford an actual 3D printer. So I printed out the stencils of each part using a regular A4 paper printer, and then used a hot-wire cutter to carve the actual parts out of industrial foam. Then I basically stuck them together with toothpicks and tape and some adhesive. I added some polyester putty to fill in any gaps, and then sanded down till I got a smooth shape.”

That first prototype that Naman painstakingly built, made using fiberglass as the composite, actually flew, and that was good enough to get a $1,000 grant from 1517 and then a whopping $25k from India’s gradCapital (Naman’s early experiment graces its homepage, and gradCapital founder Abhishek Sethi needed to get on a call with Naman’s dad to convince him to let his son take the money). The 1517 grant gave him the money to buy a 3D printer for the foam molds, and the grad windfall allowed him to move into his own apartment / unofficial lab, where he reworked the design from scratch based on everything he learned and upgraded to carbon fiber and kevlar sheets in place of fiberglass. During that same year, his 12th grade year, he got $150k from Draper Startup House and angels, formed Airbound as a real company to develop the technology, got into Carnegie Mellon, convinced his parents to let him defer Carnegie Mellon, and got to work.

Airbound was born. A Lightspeed investment followed soon after.

Tigerfeathers has a lot more fantastic detail on Naman’s entire journey, and they did such a great job that I’m going to tell you to read it so that we can get back to carbon fiber.

Betting on Carbon Fiber

Now that digital brains are smart enough to drive things, technologists around the globe are racing to add autonomy to anything that moves. There are autonomous boats, autonomous cars and Robotaxis, autonomous trains, autonomous warplanes, and of course, autonomous drones.

Autonomy is important. It lets you remove the pilot. Electric motors are, too; they let you build small without the weight penalty of a small combustion engine.

These two dynamics are well-understood by Airbound’s competitors.

DJI, Neros, Zipline, Wing, Amazon, DoorDash, and the rest, all of them make autonomous drones with electric motors. All of the drone delivery companies would agree that flying a single item point-to-point is better than batching, and that soaring in the sky beats the danger, traffic, ugliness, and unpredictability of driving on the ground. Zipline, too, aims to give the earth back to humans by moving deliveries to the sky, and it is growing incredibly fast in its efforts to do so.

What Naman thinks everyone else is dramatically underappreciating is carbon fiber.

Carbon fiber is not a new material. Invented in the 1960s and adopted by Rolls-Royce for ill-fated jet engine fan blades by the end of that decade, adoption spread from aerospace and military applications into golf clubs, tennis rackets, and fishing rods in the 1970s and 1980s. Carbon fiber entered Formula 1 with McLaren’s 1981 carbon monocoque, and eventually diffused into road cars and bicycles. The Boeing 787, launched in the 2000s, was the first commercial airliner with a mostly composite airframe. Today, every high performance modern drone, including Zipline, Wing, and Amazon, uses carbon fiber.

Despite its longstanding and fairly widespread use, Naman thinks that everyone is still thinking about carbon fiber wrong, as a metal replacement with better strength-to-weight ratios or something.

He thinks that it’s the third major materials paradigm in world history, after the natural era of wood and stone (find it and use it) and the semi-natural era of metal and concrete (mix existing materials together to get new properties), and the first in which we can design materials from the molecule up.

There has generally been a dominant material for each age with the best structural properties.

The reason why the best material matters and not the most common is that the things we can invent are limited by the best material. A car would not be possible without the structural properties of metals. “You just can’t make a wooden car,” Naman points out. “In the same way, we believe that the future of aviation wouldn’t be possible without the structural strength of carbon fiber and the whole composites umbrella.”

Naman argues that while engineers have been working with carbon fiber for decades, they still design as if it’s a basically drop-in replacement for metal instead of reimagining what’s possible in the new paradigm. The breakthrough, he suggests, won’t come from replacing metal components with lighter, stronger carbon fiber ones, but from entirely redesigning the aircraft, from engineering the reinforcement’s arrangement and the structure together, and precisely tailoring where and how it carries loads.

Instead of thinking about carbon fiber as a stronger, lighter metal, think about it as an exceptionally strong string that can be arranged into a cloth that can follow a three-dimensional shape. The ability to place the carbon fiber strings within the cloth, and the cloth into a shape, gives an engineer more degrees of freedom. He can employ fiber direction, layer placement, resin, and reinforcement to determine how the finished structure behaves.

All of this is unintuitive. Naman, carbon fiber’s staunchest advocate, remembers being “insanely confused” when he worked with carbon fiber for the first time. “It’s really weird that the material is literally a cloth, often woven in the same way your t-shirt is.”

This literal cloth gives weight-conscious engineers superpowers. By placing thin carbon fiber skins on both sides of something light and cheap like foam or honeycomb, you can make it much harder to bend by moving the strong material further apart while adding very little weight. Start with excellent tension-carrying fibers, lined up and spaced to support the intended loads most efficiently, and engineer the rest of the structure around those.

Carbon fiber lets you design a lighter aircraft, which kicks off a virtuous loop called mass decompounding. Holding payload, range, and required performance constant, the loop works like this: a lighter airframe needs less energy to fly, which means you can use a smaller battery, which reduces weight further, lowering the required propulsion power, which means fewer or lighter motors, which means even less weight, and so on.

Aircraft design normally faces the opposite, vicious version of the loop: add weight, need more power, need more propulsion, repeat. Carbon fiber, enabled by autonomy and batteries, can flip it. Just as Ford bet on vanadium-alloyed steel – hiring a metallurgist, contracting a small Ohio mill to pour it, and using it in the Model T’s crankshaft, axles, gears, and springs – Naman is betting on carbon fiber.

You can do unbelievable things with carbon fiber, Naman believes, if you design aircraft carbon-fiber-first, and if you couple the design and manufacturing.

What does that mean?

“The easiest thing I can state without having any IP risk,” Naman states, “is that we can use carbon fiber to build more complex structures. What we have pioneered is the ability to make an entire aircraft in one piece - like all of the internal structures, all of the internal support.”

The Airbound Aircraft

Airbound’s v1 aircraft, the TRT, was not made in one piece, but it was made of carbon fiber.

Working with carbon fiber changed how Naman thought about engineering, to the extent that he’d thought about it at all. Because he had taught himself fresh, with a specific goal in mind, he was lucky: he learned composite manufacturing without having absorbed the conventional assumptions of metal manufacturing. He just wanted to figure out how to make a blended wing body drone. As Micky said, “they’re naive, so they think anything is possible.”

And because the TRT was designed to carry a 1kg payload - specifically, medical supplies - the team could start with the impossible and iterate towards just barely achievable relatively cheaply. Small drones gave him room to experiment with material, geometry, and manufacturing at the same time.

They asked questions like “What does a carbon fiber first design look like? And what are the manufacturing capabilities we can build out in carbon fiber?” Then, he tells me, “we use the new capabilities we built to influence the design, which influences what new capabilities we need to build.”

Airbound was trying to do things with carbon fibers that others hadn’t or wouldn’t. While TRT wasn’t all one piece of carbon fiber, including the internal structures, the composites specialist on the now-growing team, Sambid, figured out how to make its surface from a single layer of roughly 40-grams-per-square-meter carbon cloth.

Sambid actually managed to crack the problem, with 36 gsm kevlar instead of carbon fiber at first, during a working interview on a Saturday, while he still had another job. “We wanted him to figure out this layup sequence that we thought was possible and had a few ideas on but weren’t able to get to work,” Naman recalled. “He said it was impossible, I said try it anyways and gave a few ideas. He gave it a try, made a few mistakes and left to let it cure overnight. That evening, he asked me, ‘Can I come tomorrow? I know I can do better.’” He came back Sunday, fixed his mistakes, and made the first successful version of that layup, which Naman thinks was the “first in India, probably first in the world.”

A French vendor that visited Airbound, Tigerfeathers wrote, “came into the office and was like ‘what the hell are you guys even doing, no one else in the market is even attempting this!’.” Outside vendors couldn’t work with such light materials, which forced Airbound to develop more of the capabilities in house. Each improvement in cutting, joining, shaping, or reinforcing the material could change what the aircraft could be, which Airbound could act on immediately, because it was the one designing the aircraft, too.

A new manufacturing technique gives the engineers a new structural possibility. They design around it, discover the next limitation, and work out how to manufacture that. Over successive iterations, the aircraft and the process evolve together. Ideally, the craft keeps getting lighter and lighter and more and more performant as the loop spins. And at some point, they learn enough that they can apply it all to an entirely new aircraft design, and then loop on that.

The current apogee of this loop is Airbound’s V2.

The V2 is an aircraft designed to carry five kilograms, up from the TRT’s one, with an airframe that Naman says the company can manufacture in a single piece, internal supports included.

Carbon fiber makes manufacturing the V2 as a single structure possible, and manufacturing as a single structure provides a number of advantages to the weight-conscious aircraft designer.

Wherever you join two separately manufactured components, you have to transfer forces between them. The connection itself needs to be strong enough, and the material around it needs to support the connection. All of that adds weight, and it also adds work. Someone has to make the separate components, line them up, join them, and check that they’ve been joined correctly. Joining slows down manufacturing and weighs down the aircraft.

Designing the structure to be made together gives Airbound an opportunity to remove some of that weight and work at the same time. It also changes what you can design.

Take the wings. The V2 has long, narrow wings, which provide a high aspect ratio. For a given wing area, a higher aspect ratio reduces induced drag, the drag associated with producing lift. That makes long wings attractive if you want to fly efficiently. “You can just stick a wing out to infinity and it’ll be ridiculously efficient,” Naman told me the first time we talked about V2.

The challenge is, those long wings need to withstand the forces acting on them without bending or twisting too much, and adding material to make them stronger adds weight, which eats into the benefit you were trying to get and sends you into the vicious loop. But using carbon fiber in one piece, with no extra structural material at the join, gives you the ability to improve your aerodynamics, too. Make the wing sufficiently light and strong, and you can use a shape that would otherwise cost too much weight to support. Then you need less energy to fly the same payload over the same distance, which gives you another set of choices about batteries, range, and payload. Reduce weight further? Improve performance? A little bit of both?

In designing its aircraft, Airbound abides by a “Zero Compromises VTOL” philosophy. Moving to vertical take off and landing should not lead to a loss in efficiency. In the early designs, like the original TRT, they achieved the goal by limiting VTOL components to a minimum in order to minimize extra weight. It stood on its tail to take off, then pitched its entire body forward to fly, saving the weight of a mechanism to tilt the propellers, but that design left a broad surface exposed to crosswinds during vertical flight.

As they scale up, though, they need specialized mechanisms for versatility and safety. A tailsitter, for example, can’t scale, because it rotates, and rotating a payload limits what you can carry.

So in newer designs (like the V2 and upcoming V3), they’re focusing on adding VTOL components that add efficiency rather than taking it away. The unique(ish) aspect of their tilt rotor design is that they place their propellers at the wingtips, so that they improve the efficiency of the aircraft by countering wingtip vortices. The effect of wingtip propellers is identical(ish) to extending their wing further out, without any of the structural cost.

Initially, Naman couldn’t figure out how to make the tilt rotor configuration work with just two propellers while preserving its weight and aerodynamic advantages. But the longer wings help, and once the team figured out a way to tune the controller to make it more stable, it actually became easier to stabilize than the tailsitter. When we spoke after one of the first V2 test flights, and I asked Naman how it was going, he said, “It is just unreasonably stable. It is incredibly stable in the sky.”

Now that it’s flying, he realizes that “for the past three years we’ve just been engineering on hard mode [with the TRT].”

This is a useful example of how Airbound makes decisions. The tail-sitter had fewer mechanisms, and if you were designing to minimize the number of mechanisms or the weight of each component, you might not make the switch.

But Airbound’s north star is to reduce the cost per kg-km, and that allows for certain trade-offs that seem suboptimal in isolation but produce performance and economic benefits at the system level. Then, once the new system is set, they work to make it cheaper, lighter, more efficient, and more economical.

Naman calls Airbound’s approach to engineering efficiency engineering.

Efficiency Engineering

According to Naman, there are two types of engineering: capabilities engineering and efficiency engineering.

Most of aerospace engineering, he argues, including today’s drone industry, focuses on capabilities engineering, making aircraft do things they couldn’t do before. They can carry different types of payloads, hover in the air as the payload lowers to the ground, land in more difficult places, better detect nearby objects. He gives Zipline real credit here; its aircraft can do things that other aircraft simply cannot.

Airbound, on the other hand, is focused on efficiency engineering, which is what you do if you believe that the only thing that will matter in the long-run, once capabilities become commodities, is cost. It takes a fairly well-understood set of capabilities—take off vertically, carry a package, fly to its destination, deliver it—and makes them work as efficiently as possible. It reserves novelty for materials and manufacturing processes, all in service of making the system work better, cheaper.

In Vertical Integrators, I wrote that the best Vertical Integrators optimize entire systems instead of specific components, making trade-offs that those just making components might not.

Airbound is a prime example. Sometimes, as with the tilting rotors, adding something is actually the right move. Sometimes, making one part of the aircraft harder to engineer lets you make the whole thing more easily.

One of the clearest examples involves the cargo bay.

In addition to being able to carry the weight of whatever someone orders, a delivery aircraft needs enough room for the items. A five-kilogram weight limit is much less useful if you run out of space after loading two kilograms of groceries; it is limited by whichever you run into first, the weight or the volume. Naman wanted the V2 to have a large payload volume, which introduced a new problem: the cargo could move the aircraft’s center of gravity around substantially depending on where someone put it.

This becomes especially important when you’re trying to build an aircraft that carries more than its own weight, because whatever it carries has a lot of influence over the balance of the craft.

Naman gave me a deliberately extreme example. Imagine someone puts a tungsten cube at the very back of the cargo bay, then fills the rest with foam so it can’t slide around. The shipment might fit perfectly and come in under the weight limit while leaving the aircraft badly balanced.

You could constrain the payload space to make loading mistakes harder, but Naman hated that answer. It “physically hurts me to have to lower volume,” he said, because which payloads you can carry is limited by both weight and volume.

Instead, the team developed a way to lift the payload box from three points and use the load at each to calculate the payload’s center of gravity. If it fell outside the acceptable range, the system could reject the load and tell the worker to try again. Originally, the design process was built around maximizing the center of gravity variance the aircraft could handle, but at a certain point, when it can handle practically anything but tungsten cube-like situations, you’re just sacrificing efficiency on every payload to capture less than 1% of edge cases. At that point, stop maximizing and think. Build a system that can accept 99%+ and reject the <1% instead of failing on them.

This goes beyond aircraft design to systems design.

Elsewhere, he was willing to give the flight controller more work if doing so avoided adding weight and drag. This is similar to the Wrights’ bicycle insight, with software taking on the work of the rider. The aircraft has to be controllable across the conditions in which it operates; deciding to solve something in software gives the controls team a real challenge, but if they solve it once in software, they get it in every aircraft thereafter for free.

These are the sorts of trade-offs Naman is willing to make in pursuit of the “perfect aircraft.”

“I’m not settling for anything less than that because there is a way to measure perfection,” he told me. In Airbound’s case, that is cost per kg-km.

To calculate the cost per kg-km, Airbound counts more than the energy and people needed for each flight. It counts making and maintaining the aircraft, replacing its batteries, and getting it ready for its next job, too. Add up all of those costs over the aircraft’s useful life, then divide by the total kilogram-kilometers of payload it delivers over that same period. That’s the cost per kg-km.

This loss function gives the team a way to argue across every aspect of the business, because each piece contributes to the cost. If one mechanism reduces the force required from a servo but weighs more than the servo weight it saves, that’s a bad trade. If making the controls more sophisticated preserves useful cargo volume without adding hardware, that might be a good one.

The arguments can get extremely detailed, but they have a built-in arbiter: what does this do to the cost of moving useful payload?

They also have a conductor.

On one of our calls, Naman was showing me the V2’s design and explaining why its wings sweep slightly forward. It was an unconventional choice for a drone, but the aircraft needed to tolerate changes in its center of gravity while hovering, and it needed to tolerate them while flying forward. Those were two different engineering problems, and their acceptable ranges didn’t naturally line up. Sweeping the wings forward, he explained, helped bring them into alignment.

I asked him how the team actually, in practice, arrived at a decision like that, how they took such a multivariate problem, with one team working on one thing and another on another, and moved forward despite the push and pull of each choice.

“This is the kind of aircraft that can only be designed by one person,” he said. “Designing the aircraft is the most complex activity. And that is, like, the one activity I can only trust myself to. Like, nobody else was involved in the aircraft design process. It was just me.”

Design, as we’ve discussed, is an iterative process, which means that Naman’s role iterates as contact with reality teaches his team. While most of the people on the team need to be experts in their respective areas, one person needs to hold enough about all of it in his head to feel how it all fits together. To understand the structural cost of changing the wings, the aerodynamic benefit the change might produce, and whether the team might be able to manufacture a structure light enough to make the trade worthwhile. He needs to know how much cargo volume customers will need, how they might load it, how far that could shift the center of gravity, and the levers available to pull to overcome those shifts.

“I need to think through the end state of how the drone will be made while designing the drone,” Naman said.

In Naman’s opinion, the founder needs to be able to follow a decision across the boundaries between disciplines, including boundaries that might otherwise separate different companies. Something the manufacturing team learns changes what the aerodynamicist can ask for. Something the controls team can solve changes how much structure the aircraft needs. He can consider changing several of those things together. Vertical integration helps insofar as the founder can wield the information and control that vertical integration enables.

Fred Brooks, the computer scientist who wrote The Mythical Man-Month, argued for a software system architect who preserves a design’s conceptual integrity, the coherence of the whole, as many people work on its parts. Even an iterative development process, he argued, needs someone responsible for that coherence.

Brooks was writing about software. At Airbound, which makes flying hardware, weight compounds the consequences of losing that coherence.

Imagine each team adds a little hardware to make its own problem easier. The controls become easier to tune. The structure gets a more comfortable margin. Assembly becomes more forgiving. Each decision might be perfectly reasonable on its own, but each adds a little bit of weight, which demands more power, which means bigger batteries and motors and more weight.

Brooks, in fact, used aerospace to illustrate the importance of knowing which resource a team is optimizing. In a NASA moon shot, he told Kevin Kelly in a 2010 WIRED interview, “money is abundant but lightness is scarce.”

Naman’s job is to preserve lightness while still producing an aircraft that people can load, manufacture, maintain, and fly reliably. That requires an obsession with understanding where weight might be saved without making everyone’s jobs impossible, and it requires knowing when to spend weight, as he did on the tilting mechanisms, because the aircraft benefits overall.

The team supplies the specialized knowledge that makes those choices possible. Sambid’s work with thin carbon cloth is a good example. When Sambid figures out how to use really thin carbon fiber cloth, Naman’s job is to understand what really thin carbon fiber cloth makes possible elsewhere in the design, what new capabilities that resulting design will demand from the team, and how much those demands will cost in grams and dollars.

“What makes you good at this?” I asked him, as someone who would decidedly not be good at this.

“A good efficiency engineer is someone who can predict the future,” he answered. “Before I do an experiment, will I know whether this works or not?”

But how can he predict what will happen?

Naman provided a glimpse in a conversation he had with Hemant Mohapatra, a Partner at Airbound investor Lightspeed India, on the Lunchbreak podcast, when he said that basically, he just understands six or seven basic axioms and can build everything up from there.

I think there’s something to the fact that Naman never went to college, that he learned basic axioms, found a problem he really wanted to solve before he’d learned much else, and then set those axioms and directed knowledge acquisition to work on solving it. His intuition has all developed around one idea. There’s that quote that Hall Hibbard once said about Skunk Works legend Kelly Johnson, “That damn Swede can see air.” Naman can see air, too.

When you talk to Naman, you can almost feel how deeply and naturally he understands the forces at play. His objective is simple, and so are the axioms on which the solutions are based. Being able to see how they fit together is incredibly complex, but it’s what he’s trained himself to do on YouTube, in long hours spent cutting, shaping, and joining carbon fiber sheets by hand, and by watching what he’s built try to fly.

That last part is important, because no matter how much Naman can hold in his head, and no matter how good his or any team is, to understand how something flies, you need to fly it.

Which is why the fact that autonomy, the electric stack, and carbon fiber allow Airbound to build small things first is so important.

Small, Medium, Large

If Tesla had its Master Plan, this is Airbound’s: build just three aircraft, a Small that carries 5kg (V2), a Medium that carries 100kg, or a person (V3), and a Large that carries 500kg, or a family (V4).

The plan begins with cargo in small vehicles, allowing Airbound to learn cheaply before it carries people, a luxury afforded to the company by the modern ability to build small. Ford didn’t have that luxury, so he tried to start with a personal airplane.

On July 30, 1926, his sixty-third birthday, Henry Ford unveiled the Model T of the air.

He called it the Flivver, after the slang people already used for the Model T itself. It was a single-seat monoplane, a little over fifteen feet long and 350 pounds empty, with a three-cylinder, 35-horsepower engine. According to its designer, Otto Koppen, Ford wanted it small enough to fit in his office. Ford’s top test pilot and close friend, Harry Brooks, used the prototype to fly home from Ford Airport.

In February 1928, Brooks tried to fly a Flivver from Detroit to Miami. He broke the light-plane distance record on the way, but the destination was Miami, so on to Miami he went. En route, he crashed near Melbourne, Florida, and died. Ford abandoned the project.

Eight years later, in February 1936, a Curtiss-Wright engineer named Theodore Paul Wright published a paper in the Journal of the Aeronautical Sciences called “Factors Affecting the Cost of Airplanes.”

At the time, the Bureau of Air Commerce was sponsoring a program for a small two-seat airplane it hoped could sell for $700, if ten thousand of them could be built. The Bureau’s director at the time was Eugene Vidal, Gore Vidal’s dad. One economist who later dug into the history suggests Wright was inspired by both the mass production of cars and the dream of the family airplane.

What Wright found was that each time the cumulative number of airplanes built doubled, the labor needed to build each one fell by a steady percentage, about 20% on average across American aircraft makers. We call it Wright’s Law now. The curve behind cheap solar panels, cheap batteries, and cheap chips, and hopefully cheap nuclear reactors, one of the most oft-cited ideas in not boring, began as a study of how to make airplanes cheap.

Wright discovered his Law in pursuit of a cheap little airplane. It went on to make almost everything cheap except little airplanes.

Neither the $700 plane nor the Flivver ever arrived in every garage. Planes, thanks to steam-carriage-like economic factors, went big, and therefore, the number of planes produced annually stayed small. “The entire aviation industry globally makes less than 10,000 aircraft a year,” Naman likes to point out, roughly the size of the single production run the Bureau hoped for ninety years ago. The auto industry makes something like 90 million vehicles a year. It has ridden Wright’s curve while airplanes have not.

Today, however, autonomy, the electric stack, and carbon fiber, make small aircraft make sense, which should be a starting point from which to scale to automotive-like volumes.

“Autonomous means I’m not paying any penalties for building small, and then carbon fiber lets me build better,” Naman told me. “So the point is that if I can now build better and there’s no cost to building small, let me build better while small and then scale that up.”

Small

Airbound’s big plans start small.

“Let’s perfect it in a 5kg model,” Naman said, referring to the V2, or Small. “Because that’s what lets us iterate a lot. There, it’s a lot cheaper to iterate. We put a product out into the market and perfect it, and then the engineering risk is just scaling it up.”

Airbound’s V2 is much cheaper than its larger drones will be to produce, and is already cheaper than existing drones made by bigger companies. The V2 costs less than $5,000 to make in India, and would cost an additional $1,000-2,000 with American labor. While there aren’t great publicly available cost numbers on competitors’ drones, Naman estimates that they cost roughly $100k when they were at Airbound’s current production scale, and have come down to the mid-to-high tens of thousands as production ramps. The human-carrying eVTOLs we talked about earlier are obviously much more expensive, at over $1 million to produce and $5 million to buy.

For V2’s price and weight, it is also a very performant aircraft. Naman’s rule of thumb for a conventional aircraft is “30% structures, 30% fuel, 30% payload.” A fully loaded V2 is more than 60% payload at takeoff, and Naman thinks Airbound can get its structure down to 4-5% of the aircraft’s total weight.

It’s easier to manufacture, too. V2 has a wingspan nearly double that of the TRT, and it should take half as long to build.

How? “We just simplified shit a lot,” Naman said. “The running joke is that the V1 was designed by someone who didn’t know manufacturing and the V2 was designed by someone who knows manufacturing. They were both designed by me.”

The Small doesn’t need to beat a fully loaded semi to generate a lot of demand, although there are some uses the Small can steal from a truck immediately: things that people want immediately. Because it doesn’t need to wait for a batch and because it flies over traffic and buildings as the crow does, the Small can start eating same-day delivery use cases on items up to 5kg.

The Small, Naman says, “is meant to be the aircraft that can compete with vans in last-mile.”

The Small mainly needs to beat the cost of a person on a scooter, in a car, or around a van route hand-delivering you something that weighs a few kilograms. In the US, DoorDash advertises base pay of $2–10-plus per job, before tips and promotions. If Airbound can get its all-in cost below a dollar per delivery, it could charge less than many existing US delivery services, earn a margin, and still make a lot of existing deliveries cheaper. India, where Airbound is starting, is a much tougher bar to beat: its deliveries are both cheaper and faster. When we first met, Naman told me he was shocked at how bad American delivery was. Still, a cheap, driverless aircraft should be cheaper and faster than even Indian delivery.

As the cost falls, more things make sense to deliver, more frequently. You might happily pay $5 to have a full dinner delivered, but you probably wouldn’t pay $5 to have a lime delivered to make a margarita unless you were a real lush. But at 50 cents? 25 cents? 10 cents? At some price point, you’d probably stop storing a bunch of things yourself and just order them when you need them. This is particularly true if an AI assistant is monitoring and ordering for you, so you don’t even need to spend the mental energy.

In 2025, Grandview Research sized the market for last-mile delivery at about $167 billion. This is probably high in some ways and low in others. For example, it doesn’t include the cost of the delivery vehicles themselves.

Because Airbound is starting small in an enormous market, and because it’s not flying humans, it’s able to fly a lot. Airbound has made more than 13,000 autonomous flights across Bengaluru and Guntur. When I talked to Naman earlier this month, he had two V2s and twenty-six TRTs in the air at the same time. Each flight gives Airbound data with which it can both improve its aircraft and operations, begin making the case to regulators that Airbound aircraft are safe, and work towards larger platforms.

Medium

All of that flight data will be particularly useful for the Medium aircraft, which will start by flying packages but is intended to ultimately fly humans.

Naman himself admits the Medium can’t be just a scaled-up Small. It has different aerodynamics and a different structural fraction, because it’s a different aircraft with a different airframe. What he hopes to transfer and scale is everything around the airframe:

• The carbon-fiber-first design method and the one-piece manufacturing process

• The experience making lots and lots of aircraft

• The controls work and design principles that make V2 “unreasonably stable”

• The autonomy stack and the operations playbook

• The supply chain and regulatory relationships

• The flight hours

Turns out, that’s actually the kind of stuff that Wright was tracking (the Wright stuff), too. He tracked how many hours of labor it took to build the next plane, which improved based on everything the manufacturer learned. That’s why Wright’s Law describes the “Learning Curve.” In Airbound’s case, the learning from the Small will be brought to bear on the Medium.

The Medium is the aircraft Naman is most looking forward to building, because he can make it much more efficient than the Small.

“It is like, you just cannot make a five kg as efficient as you can make a 100 kg aircraft,” Naman said. “It’s not physically possible.”

You can understand why by watching the birds. Small wings moving slowly live in what aerodynamicists call the low Reynolds number regime. There, the air’s stickiness matters much more relative to the wing’s size, so a small wing pays more drag for every bit of lift. It’s why the sparrows flap like crazy and the heron floats.

Naman says his most efficient designs are north of 100 kg. Bigger aircraft can be proportionally lighter, too. At 100 kg, he thinks structure can get to about 4% of the aircraft’s weight or lower, without needing carbon cloth any thinner than what already exists.

The Medium also fixes the Small’s people problem. While Airbound’s aircraft are cheap to manufacture and getting cheaper, flying one is “still double the cost of what it should be because of labor,” he told me. Specifically, while there are no pilots in the aircraft, people are still required to supervise the fleet from a control room, remotely. As Airbound proves its vehicles are safe, and as regulations change to support drone delivery, the same person will be allowed to monitor more drones. Today, in India, there’s no regulatory limit on how many drones one person can monitor. In the US, it varies from operator to operator, but no one has gone higher than about 10. For now, one person watches every five Airbound aircraft fly. No matter how low that ratio goes, as long as it’s above zero, an aircraft with 20x the payload will pay for 20x less labor per kg.

At that scale, Naman expects electricity and battery replacement to account for most of the cost, with battery replacement the largest item. Longer-lasting batteries and more efficient aircraft would then become the main ways to drive costs lower.

“What we are saying,” Naman said, “is that we can make it so that energy starts to dominate at 100 kgs.”

That will be an incredibly important milestone. When energy dominates, you’re left with physics and nothing left to squeeze but the aircraft itself. And squeezing the aircraft is what Airbound does.

Over time, the goal is to get cheaper (cost) than most vehicles carrying 20 tons on a kg-km or tonne-km basis, even though Airbound aircraft will carry only kilograms at a time, in the air, meaning it will win on speed and versatility, too.

The x-axis shows how many kilograms a vehicle moves at a time. The y-axis shows what it costs to move one tonne one kilometer (the numbers look silly small on a kg basis, so multiply by 1,000 to get a tonne). Bigger is cheaper per kilogram, and costs in the chart are sensitive to loading: the closer a vehicle is to its max payload, and the fewer empty return trips it makes, the cheaper per tonne-km.

“There is a linear relationship between the amount of payload moved and… cost per kg, or cost per ton,” he said, until it “starts to flatline after like 20 tons plus.” Moving a kilogram in a semi is much cheaper than moving it in a van. Moving it in something bigger than a semi isn’t that much cheaper than the semi. This curve is why logistics companies and airlines batch, why your package sits in a warehouse waiting for enough other packages to fill a truck.

Airbound’s pitch is that it can flatten the curve out at 100 kilograms instead of 20 tons. If moving 100 kilograms at a time, like the Medium will, is about as cheap per kilogram as moving 20,000, you don’t need to wait for the other 19,900.

Hitting those costs requires high utilization, which makes the choice between charging and swapping batteries important. It will come down to the economics of each country. For now, in India, where labor is cheap, he’d rather use 5 minutes worth of labor to swap the batteries vs investing in infrastructure to autonomously charge the systems, especially since charging is slower. “It’s basically a question of ‘which is more expensive: 5 minutes of labor cost many times a day, or a 60% utilization loss per vehicle?” Naman explained. “In the US, using labor is probably more expensive. In India, losing the utilization is costlier.”

To dominate, to win on speed, versatility, and cost, the Medium will need to get very cheap to operate, based on a million different decisions on everything from battery swapping to loading to design. The American Transportation Research Institute puts the average operating cost of a truck at about $2.34 per mile. Spread that over a full 20-tonne load and convert to kilometers, and it costs roughly seven cents to move one tonne one kilometer. Carry a half-load – ten tonnes – at the same vehicle cost, and it’s about fifteen cents.

Naman’s target for the Medium is four to six cents per tonne-kilometer. That’s cheaper than a fully loaded semi, and Naman expects it will also be less than half the cost of a bus or train. That is what it takes to give everyone access to flight. Carrying a full 100 kilograms, it works out to about half a cent per kilometer. And those are straight-line kilometers, which Naman counts as worth about two on the road, so the real gap is wider than the per-kilometer numbers suggest.

It’s a target, and there is much work to be done to get there. But getting better than, or close to, full-20-ton-truck economics at 100kg would be a major unlock. A business could receive a truckload’s worth of goods over the course of a day without receiving a truckload all at once, and send each order out as soon as it is ordered. At some point, they will restructure their operations towards point-to-point from fulfillment centers further from a city’s economic core.

Despite the fact that most cargo volume moves on ships, trucking accounts for half of the global logistics market by spend. Market size estimates vary, but it’s somewhere between $8-12 trillion, of which trucking makes up $4-6 trillion. Whatever the true size, it is a very large prize, and one that is largely determined by price.

“2027 is the year I will start pitching ‘Replace all of your trucks with my drones. Not some of them, all of them,’” he said.

But cargo shouldn’t have all the fun. People want to move themselves faster, cheaper, too. The Medium will be the first Airbound aircraft designed to do that.

The sequencing is intentional. Getting regulatory approval to fly humans is much harder than getting approval to move packages. Airbound can use less precious cargo to learn how to make the Medium better, cheaper, and safer, and use the data from its flights to prove that it is indeed safer.

“Before moving a single human,” he said, “let me just come to you [the regulator] with like 10 million hours of flight data on… an aircraft that is technically capable of carrying humans,” carrying “payloads that weigh about the same as a human.” It’s sort of like Wright’s Law applied to safety: learn by doing, and get better and better the more you do, so that by the time you fly humans, you know it’s safe.

An aircraft that carries humans will need to be safe and comfortable for humans, too. It will probably need windows, for one thing, and a seat, for another. Adding windows to the single carbon fiber structure will introduce engineering challenges, and the seat will introduce weight. An early sketch Naman showed me didn’t look very comfortable, but it’s not physically impossible to carry humans safely and comfortably, and there will be plenty of room in the cost to work with. Already, in just a couple months, the latest design has the ceiling height of a sports car, and is roomier in every other dimension. He is learning.

The goal, at the start, is to make the Medium cost competitive with a car. On launch, he expects it to cost around $20,000 to manufacture and about $35-40,000 to buy as a starting point, with costdowns from Wright’s Law from there. He doesn’t expect it will replace the family car, but he does think that it might replace the second car. “That one is normally just for getting one person to work anyway,” he said.

To really eat into the market for moving people though, Airbound will need to let them move together.

Large

Naman plans to introduce a 500 kg aircraft as well. He didn’t want to at first, because five 100kg craft can do what one 500 kg craft can, but he realized that people have families, and families like to travel together.

Here, the cost to beat is the family car.

AAA’s 2026 analysis puts the average cost of owning and operating a new car at about 86 cents per mile, assuming 15,000 miles a year over five years. Fuel, maintenance, repairs, and tires account for about 29 of the 86 cents, so even if you already own the car and count only fuel, maintenance, repairs, and tires, that’s 29 cents a mile. A Medium, which carries just one person, would be much cheaper, at Naman’s under 2¢ per mile target. A Large with a full family in it, at Naman’s target costs, would be about 3¢ per mile. Both would be comfortably cheaper than the family car, not to mention faster and cooler.

How big is that market for moving people? Here, the math is a little trickier, because the ride-hailing, public transport, and airline markets are surprisingly small (combined ~$3T or less). The vast majority of the market is in private vehicles - the car itself. Each year, $2.5-3T of new cars are sold, not accounting for used cars (another $2T or so), not including parts and service, financing, or insurance, which is why Airbound wants to compete with the car itself on cost.

This vehicle, of course, exists only in spreadsheets and in Naman’s head. The numbers work on paper, but there will be an enormous amount of design, engineering, manufacturing, and regulatory work to get there. It is, however, incredibly important to Naman that he and the team get all that work right. That is how you do to the car what Ford did to the horse.

Earlier, I compared today’s eVTOLs to the early automobile. Look at the numbers, though. The Model T went from about a year and a half of the average American’s pay in 1908 to about a fifth of a year’s pay in 1925. Call it 8x. Getting a $4-5 million passenger eVTOL to something like a car’s price takes more like 100x. At Wright’s 20% per doubling, 100x takes about twenty doublings of cumulative production, roughly a million times more aircraft than you initially built. And at $4 million, there’s unlikely to be enough demand to do much doubling.

To be fair, this is why companies like Joby and Archer aren’t selling eVTOLs to consumers, but planning to operate them like flying taxis and buses. That way, they can amortize the cost.

Still, Naman’s bet is that instead of pushing a $5 million vehicle down the cost curve, it’s much easier to push a sub-$5,000 one up the size curve while continuing to bring the cost-per-kg-km down. Airbound will keep doubling production, making more carbon fiber and components by serving the more abundant demand for burritos delivered more cheaply than by bike and blood delivered more quickly than by van. That is how you push out the efficient frontier.

My bet, in investing in Airbound, is that Naman’s bet is the right one, and that if it comes even close to paying off, it has the potential to support an absolutely gigantic business.

S, M, L Aircraft → XXL Business

As we’ve been going back and forth in our conversations, Naman and I have tried to come up with one chart that clearly explains how Airbound wins more and more of a share of movement over time, and this is the best we’ve come up with.

Each light blue bar is a way to move things today. The dark blue bars are Airbound’s costs and target costs, from V1 today to the Large. Each new aircraft targets a lower cost per tonne-kilometer.

Every vehicle Airbound beats on cost opens another market it can compete for. Beat more and more vehicles on cost and you capture their business.

It’s one clear bet, and Airbound is all-in on it. I fucking love it.

“Imagine,” Naman asked me, “what would happen if every single automobile ever made was just squished into three variants.” Like, if instead of cars from dozens of companies in countless makes, models, and configurations, instead of scooters, delivery vans, trucks, taxis, and family cars, not to mention helicopters and small private planes, just three vehicles captured all of that demand.

Imagine the scale you could achieve on those three platforms. Imagine Wright’s Law ripping those platforms down the learning curve with each doubling. Imagine the orders you’d place for all of the components you need, and imagine the margin you could squeeze out by making many of them yourselves. Imagine making those platforms cheaper and cheaper and more and more performant, such that the demand those three platforms would generate today is dwarfed by Jevons Paradox tomorrow.

The timelines at play are very long, as they can be for someone who’s only used up two of his many decades himself, and the costs that Airbound will need to achieve are very low, and they are, to be clear, just Airbound targets for now, but the idea is that the more different vehicle use cases Airbound can move onto its platforms over time, the more the advantages will compound in its favor.

“There’s something quite deep about the notion of using time horizons as a competitive advantage,” Stripe CEO Patrick Collison told Ezra Klein a decade ago, “in that you’re simply willing to wait longer than other people and you have an organization that is thusly oriented.”

Decades are an advantage, as is pushing all of those decades improvements onto a small number of platforms.

Automakers already understand the value of this, which is why many build both mass and luxury brands on the same architecture. By its tenth anniversary in 2022, Volkswagen’s MQB architecture had underpinned more than 32 million vehicles across the group. Standard interfaces nearly halved the number of engine and gearbox variants; versions of the heating and air-conditioning system fell from 102 to 28. When Wolfsburg began producing the new Golf in 2019, it was able to reuse around 80% of the existing body-shop installations. Customers got different cars, but underneath, Volkswagen repeated much more of the engineering and manufacturing.

Naman wants to extend that logic across categories that currently require different machines. The economic imperative for a scooter, a van, or a truck disappears if one aircraft can do enough of their jobs at a lower total cost, faster and more flexibly. And the more of those jobs it wins, the more resources Airbound can put into making that aircraft cheaper.

Some of the math is straightforward, simple industrial logic. Say it costs $10 million to develop a particular component. If you can only recover that investment over a thousand units, development alone contributes $10,000 to each one. If you can recover the cost over a million units, development costs only contribute $10 to each.

For cars, this kind of math is standard, but Naman argues that airplanes are made in such low volumes that their components are far more expensive than they need to be. Take avionics, for instance.

“An aspect that a lot of people miss,” Naman told me, “is that the reason why you have avionics that should cost tens of dollars to make, being sold for tens of thousands of dollars in aviation, is because this industry is just pathetically small. It is insane how small the scale of this industry is. Which means that you have these highly bespoke supplies. There is nothing expensive about making aviation-grade hardware. And honestly, the R&D required to get to that hardware is also not that expensive. Because it is low volume, the cost ends up being obscenely high. What should actually happen is if you get to automotive scale, aviation-grade hardware should maybe be 20% more expensive than automotive to compensate for the higher cost. The other part where this cost comes down is because if we can get to reasonable scales, we can expect industrial-grade costs for our internal components. We can start to expect to use more standard components for these aircraft.”

Would they just make the parts themselves then?

Naman’s answer is another reason it makes sense to start by moving things before moving people.

“The nice thing is that off-the-shelf hardware, the cheap off-the-shelf hardware works for cargo movement,” he said, before departing on another riff.

“Cargo movement is just less sensitive. This is generally acceptable for drone systems. Sure, maybe instead of like one small flight computer, I need 3. But I mean, even at small scale buying that’s what, less than $600 in flight controllers? And with scale, that number goes down massively. So that is what makes avionics really cheap. And as we get into passenger movement, we will need much better flight controllers. But there, what I would want to do is I would want to start by making these flight computers ourselves, creating this template, and then rather than going to aviation suppliers and saying, ‘hey, make this for me at scale,’ I would rather go to automotive suppliers and say, ‘look, you already have scale. Make this better component for me.’ This is in fact the same thing that we’re dealing with with drone components because the drone industry we believe just doesn’t have that quality and have that sense of scale. And the reason why we will be attractive for an automotive supplier is because we will give automotive volumes. And at the end of the day, this creates bigger companies. Like our idea is that we don’t want to create suppliers like the suppliers Boeing has. We want to create suppliers like Foxconn. And guess what? Foxconn is bigger than Boeing by revenue. It’s crazy how huge you can get with just low margin, high volume. And that’s what we really want to turn aviation into.”

So is that what Airbound wants to become, a Flying Foxconn?

Not exactly, because he wants to sell the product itself, and because despite its revenue, Foxconn is smaller than Boeing by market cap. Airbound wants its suppliers to become Foxconn.

“To some extent,” he told me, “we have to play the role of supplier for most things to kickstart this, but eventually, we want to vertically integrate into the things that matter most, and have suppliers build out the rest at scale.”

He thinks Airbound can become the world’s largest carbon fiber fabric consumer, for example, which would provide enough incentive for someone else to get really good at it.

Airbound’s advantage today is in turning carbon fiber into an aircraft: figuring out which structures carbon fiber allows them to make, how to make them, and how better manufacturing processes might change what the aircraft can do. As we’ve discussed, those things develop together.

At much higher volumes, it would be worth it to a supplier to spend the CapEx to solve those problems with equipment and automation. Nothing, I would imagine, motivates you to automate like the healthy percentage of his life that Naman has spent wrestling carbon fiber sheets by hand. Every small reduction in scrap becomes valuable when you have enough scrap, just as John D. Rockefeller found it worthwhile to use 39 instead of 40 drops of solder to seal each five-gallon can of kerosene. Every production improvement that saves ten minutes per aircraft saves almost nineteen years of continuous labor across a million aircraft. These little things compound.

Naman thinks the effect can eventually reach further upstream, into making the carbon fiber itself, not just turning it into sheets. “We will actually make carbon fiber cheaper,” he told me. “We will bring carbon fiber down to a point where it becomes a similar cost to steel.”

Ideally, at enough scale, and with enough proven demand, Airbound will incentivize other suppliers to invest in getting really good at making carbon fiber, so they can focus their capital and attention on new problems.

This is how Naman thinks about batteries.

In July, Naman told me that if Airbound captured the amount of movement he had in mind, it could become “the single largest consumer of lithium ion cells in the world,” they might even buy, he predicted, twice as many lithium ion cells as everyone else combined.

There is a bit of bravado in this answer, a sense that Naman is already living in the future that he can see, but the importance of batteries is a clear and present challenge today, and he’s already thinking about what it would look like to not have to use traditional lithium ion cells.

“Our cost numbers,” Naman explained, “are incredibly sensitive to our battery replacement cost, which means if batteries become cheaper, the cost comes down dramatically. But more importantly, if batteries become longer-lasting, the cost moves down dramatically. Think about solid state batteries. If they can work, they’re projected to have, I think, 10-100x the life of lithium ion, so that will be fun when we can adopt them.”

Fun. IF they can work is a big caveat, because the big challenges to commercialization today are precisely the longevity issues, like cracking, loss of contact, and dendrites. What he means is that, theoretically, they should be able to last longer, and the way he sees it, getting from theoretical to practical simply means solving a bunch of engineering challenges. Plus, they’d be more energy-dense, which means weight savings, which as you now know, compound too. They’ll be more expensive per kilowatt-hour for a while, but sometime before they become directly cheaper, they’ll make more sense as part of the system.

In the meantime, Airbound is looking at other novel battery system solutions.

“We’re also looking at some interesting problem statements on how do we integrate batteries into our structures, and I also think there’s a lot of ways to push past what exists here,” he said. “In fact, the structural battery thing is really cool because there are some nice, cute applications we can use it for right now. Like one interesting example is the payload basket we have right now. If we want to deliver to customers’ homes, we should have lights on it and the ability for the bottom to be kind of like a trapdoor that opens and closes to release the payload. You don’t want the drone to be waiting for the person to pick it up, then you run into a lot of issues, and the drone can’t fly without the payload box, because it’s part of the structure. So the problem is that if I want to put a battery in the payload box, that makes things really heavy. And if I want to deliver power through the wires, that gets really clunky. But if the payload box structure itself can be a battery, it’s a really tiny battery that serves all these purposes.”

“So we’re speaking to a research team that has worked on this,” he continued. “We want to see how far along the technology is and if we can push the boundary on that. Because, again, a lot of this stuff is making me realize that our entire goal is how do we make aviation, like, an order of magnitude more affordable? And so being able to figure out batteries is going to be a very core part of that.”

Structural batteries and solid-state batteries are research projects. They might not make commercial sense for a decade. That’s fine. Naman has decades. I’m not sure how, in the long run, you compete with someone who’s willing and able to obsess over those details for such a long time period.

Anyway, becoming the largest battery buyer in the world would require quite a lot of volume, given that the IEA estimates that electric vehicles alone deployed 1.2 terawatt-hours of batteries in 2025. Airbound would be entering a supply chain already serving an enormous and growing market.

(He thinks it won’t be large and growing for long. “If you believe in Airbound, you should short Tesla. There’s not much point to a self-driving car when you can have a self-flying car,” he told me once, half kidding I think.)

But the strategic logic is straightforward. A large, dependable buyer can negotiate supply agreements, support dedicated production capacity, and give a manufacturer a reason to develop cells around its requirements. For Airbound, those requirements would include weight, power, charging performance, and its real-world cycle life. Long before becoming the world’s battery buyer, it could command enough purchasing heft to accelerate those characteristics.

All of which should make the aircraft cheaper to manufacture and operate, cheap enough to manufacture and operate that people can own and rent them out like Robotaxis.

“Private ownership is not going to disappear,” Naman conceded. “Even if there’s an incredibly compelling air taxi, people are still going to own cars. For this to replace the car, it has to be consumed like the car. The best analogy I can think of is Tesla’s FSD, where you still own the car, but you can also lease it out as a Robotaxi when you’re not using it. You make some money from it, and Tesla takes a cut from each trip your car does. I envision a similar end state for us in the next one to two decades.”

“The next one to two decades.” All of the specific things discussed in this section – avionics, carbon fiber, batteries, whatever – those will change. Airbound will solve problems, which will create new problems and new opportunities; they’ll take advantage of the latter and solve the former, over and over. Above and beyond the specifics, the great advantage Airbound can achieve is to make small improvements, at increasing scale, over a very long time.

This business model vision also speaks to the power of vertically integrating from the carbon to the customer.

Imagine flying to work, getting out, and making your aircraft available until you need it again that evening. In the off chance that you need to leave in a hurry, there are plenty for rent nearby. Someone else could fly yours without buying one, and you can offset your cost of ownership. Airbound earns money when it sells the aircraft and can earn a fee when the aircraft works on the network. Owners can supply some of the capital needed to expand the fleet, so Airbound can spend more of its own capital on R&D. This also, theoretically, increases demand, giving more people more ways to afford and use Airbound’s aircraft, which… you get it.

The other thing that model does, though, is to give Airbound a second core moat, beyond Scale Economies.

The Movement Monopoly

Airbound, if it works, seems to be the rare business that’s capable of developing both Scale Economies and Network Effects, with the manufacturing know-how potentially adding Process Power as well. (I could make an argument for the other four of Hamilton Helmer’s 7 Powers, too, but I’ll leave that to the reader.)

When I brought that idea up to Naman, he’d been thinking along the same lines.

We had covered Scale Economies many times, and have done so above. It is Airbound’s raison d’etre. But what about Network Effects? That depends on whether more people using the network make it better for others, in a way beyond just creating more scale to lower costs, the way that X is more valuable the more of my friends shitpost on there.

I think there is a case to be made. More available aircraft in an area can mean shorter waits. Better availability attracts more customers. More customers create more paid work for aircraft owners, making it more attractive to put an aircraft on the network. More aircraft improve availability again.

It’s a similar argument to the one Uber makes. Back when I was launching Breather in New York in 2013, I asked Uber NYC GM Josh Mohrer for advice, and he said their whole game was about making sure there was always a car nearby to minimize wait times. The whole flywheel spun from there, and later, the theory went, benefited from adding Uber Eats to throw more demand into the system.

Airbound’s proposed network could serve goods and people using vehicles designed around the same manufacturing system, the dream of Uber Eats. Within cargo operations, an aircraft delivering for a hospital could find its next job with a retailer or a restaurant. Where the aircraft’s configuration and approvals permit it, passenger and cargo demand might also help fill different parts of the day. The valuable thing is always finding another useful job nearby, instead of spending time idle or flying empty. The more kg-km, the lower the cost per kg-km.

Network effects in the physical world are local and have to be earned market-by-market, but Airbound’s advantage over an Uber is that “local” covers a larger circle.

That gives Airbound two ways to make a trip more affordable – reduce the cost of supplying the aircraft, and increase the useful work the aircraft performs – which interact with each other. Cheaper aircraft make it easier to add supply to the network. A productive network makes an aircraft more attractive to own. More owners create more demand for aircraft, which supports further manufacturing investment. A larger customer base gives those aircraft more work to do. Etc.

What would a business whose Network Effects feed its Scale Economies, and vice versa, look like?

Naman had been thinking about that, too. He had been looking at the largest companies in the world and thinking about the combination of “having a deep moat and having great distribution.” His manufacturing analogy was “a Boeing moat multiplied by a TSMC moat,” while the customer relationship would have “the same distribution as Uber and DoorDash.”

The kid is not short on ambition.

The Boeing part is the difficulty of designing, manufacturing, and earning approval to operate safe aircraft. The TSMC part is the cumulative manufacturing capability, investment, and volume required to make them competitively. The Uber part is being the service people open whenever they want to move themselves or something.

None is a perfect analogy - you may be squabbling with each, right now, as you read them - but together they describe the business Naman wants to build, hard engineering and scale underpinning a service affordable enough for people to use routinely.

If he pulls it off, it could make competing with Airbound difficult in a bunch of ways.

A new manufacturer could design a good aircraft – a better aircraft, even! – and still struggle to match the price of components bought in much larger quantities, the production yields of an experienced factory, or the cost of maintaining a standardized fleet. Before Airbound gets to scale, of course, this is an advantage that goes against the company to the benefit of its larger competitors.

Its customers might also prefer an aircraft that can earn money on an established network.

A new network could offer attractive software, like Uber and DoorDash do, and still need aircraft that are economical to operate. If Airbound’s aircraft have a substantial cost advantage, the challenger would need access to those aircraft or an alternative that matches their economics.

A company trying to build both would have to win enough customers to support its manufacturing investment while supplying enough aircraft to make its network useful. Airbound’s ambition is to have the two sides already supporting each other by then.

This is why, despite the great demands of manufacturing, owning the network is also important to Naman. When I asked in July whether he planned to sell vehicles or manage them, he was still considering leasing models, but was clear about the strategic priority: “creating that network and owning that network is just so powerful.” When we’ve talked more recently, the ownership model had evolved to include customers buying aircraft and renting them out. His consistent ambition has been to retain the relationship with the end user and the jobs to be done.

None of this automatically creates a monopoly. Suppliers can sell improved components to competitors, customers can use multiple services, and a rival with substantially better economics could still win.

And the competition is moving. Uber invested in Zipline and signed a deal that will absorb all of Zipline’s currently planned production, aiming for a million drone deliveries a day by 2029. DoorDash, an Airbound investor, is building its own drones. The two biggest delivery apps have each already placed bets.

But if Airbound achieves the costs Naman is targeting, and uses those lower costs to build a more productive network, a competitor would have to overcome much more than Naman’s best aircraft design.

The business Airbound would create in that scenario would be truly massive.

Uber handled 13.6 billion rides and delivery orders in 2025, roughly 37 million a day, generating $193 billion in gross bookings and $52 billion in revenue. That is an enormous business, built around a fraction of the movement people already do.

Americans alone made an estimated 232 billion driving trips in 2024, or roughly 636 million a day. Plenty of those trips would remain better served by a car (for a while, at least), a bicycle, or walking. Still, it gives you a sense of the underlying frequency: movement is something almost everyone in the world does multiple times a day, and they command multiples of what they do themselves in the form of DoorDash deliveries, Amazon packages, and more. Airbound’s ambition includes 80-90% of the movement of goods and people around the world, plus the increased activity that will occur when movement becomes cheaper and faster.

Imagine, for illustrative purposes only, an eventual movement network handling a billion movements a day, at a dollar per movement. That would be $365 billion in annual revenue; at ten percent net earnings, it would produce $36.5 billion a year. At an illustrative thirty times earnings, that is a roughly $1.1 trillion business.

That isn’t a forecast and it is admittedly insane. It would require the network to process about twenty-seven times Uber’s 2025 daily trip volume, and the economics will depend heavily on the mix of small deliveries, cargo, and passenger journeys. A ten trillion dollar outcome would require another order of magnitude of earnings.

But it’s a quick back of the envelope calculation that shows what Naman is talking about when he says that he believes Airbound could become a $10 trillion company. His ambition is to earn a small amount, repeatedly, on an enormous amount of movement, while owning capabilities that make Airbound’s movement cheaper than a growing percentage of alternatives.

Scale Economies would give Airbound room to charge less, lower prices would make more trips affordable, more trips would make the network more useful and its aircraft more productive, all of which would lead to more scale, and so on. A chunky percentage of the world’s movement, on three platforms.

Taking generously low margins on high volumes, and winning on both cost and performance, is how you dominate a market.

“You want to build an unfuckable business, basically,” is how Naman put it.

There is a long, long, long way from a pre-commercial 5kg aircraft to a Movement Monopoly, and it should go without saying that Airbound’s ultimate success in achieving this vision is anything but guaranteed.

Trade-Offs, or Choosing What Not to Do (Yet)

There are some things that Naman hasn’t figured out, that he doesn’t think he needs to.

On one of our calls in July, I asked Naman something about customers, and he said something like, we have a couple of customers today, but the opportunity is so large that “it feels like a waste of time to get commercial traction for the sake of getting commercial traction,” and specifically that “It’s clear that the bottleneck to becoming a $10 trillion company is getting the tech right, and if the potential is that high, we should just focus on the tech until we get to $10 billion.”

I asked him, once, about a comment that he made on that Lightspeed podcast, to the effect of, we’re not really focusing on exactly how the handoff (the delivery getting from the drone into the customer’s hands) would work, that (and this is me paraphrasing) “we just need to make the aircraft and we’ll figure that shit out later.” He didn’t deny it. He said that “We have some more clarity on handoffs, but still not complete clarity. Like I don’t think the best solutions will be built out by us.”

On an earlier call, I’d asked about regulation, and while Airbound is engaged with regulators and acknowledges that existing regulations, in India and elsewhere, will be a bottleneck to growth, he agreed with my assessment of his position which is, basically, if you get the economics right, everything else follows. In the case of regulation, he thinks there are actually two separate jobs: making a safe aircraft and proving to regulators that the aircraft is safe, and that the latter is much harder than the former, so if you can make an aircraft cheap enough that there’s lots of demand to use it for logistics, you can walk into a regulator’s office with reams of data showing how safe it is.

All of this, in a way, is systems engineering, too. It is also the textbook definition of strategy. Airbound has only so many resources at its disposal, and it needs to choose where to focus them. Whether it’s good strategy depends on whether he’s identified the right bottleneck against which to focus.

Naman’s bet – and it is a bet, however steeped in historical comparisons and models it is – is that if Airbound can win on cost, speed, and versatility, the world will bend around the company. Flying things should beat grounded ones on speed and versatility, so the bet is really if he can beat other flying things, and then bikes, scooters, cars, vans, trucks, and trains on cost, he wins.

That may be the wrong bet. Naman believes that while the aircraft themselves will not be commodities, movement will become one, and in commodities, lowest cost wins. But what if it is a land grab? What if the broader movement industry’s bottleneck is simply the capability to fly, and the scarce asset is deep partnerships with companies, like the ones Zipline has signed with Uber and Walmart, with enough demand to pull you down the learning curve, until you’re cheap enough?

There is credible pushback on this bet, too. I mentioned this strategy to a founder friend, and he wondered whether switching costs away from drone delivery companies would prevent cost from mattering as much as Naman believes. In high switching cost businesses, he explained, cost matters less than cost + credibility with existing customers. Plus, customers pay the delivery fees anyway; would DoorDash or Uber really switch their operations just to save their customers a buck or two?

I texted Naman to ask about this, and he pushed right back in a rapid succession of texts. “I think the point that customer eats delivery fee anyway is incredibly untrue,” he texted. “Delivery fee is one of the biggest concerns for all of these last mile players. Low cost is Uber’s entire land grab strategy. Uber uses low cost to capture market share and eventually ends up in a scaled steady state where their unit economics are better than anyone else’s. If you can have structurally 10x better unit ecs than Uber it’s actually pretty easy to beat them. These companies are 90% cost games, it’s just that when nothing other than scale affects unit economics, then the other layers start to matter more. This is why Uber is so obsessed with autonomous cars. There are so many misconceptions on how these industries work.”

That is Naman’s answer, at least. He will need to prove it in the market, gram by gram and cent by cent.

It is too early to know whether Airbound can win enough demand to reach these costs. Zipline is an amazing company run by very talented people doing real business today, and more every day. Joby and Archer, too, are doing a ton of manufacturing and regulatory work to make their aircraft safe and prove to regulators that they are. If those solutions are good enough for consumers, there’s a chance that Airbound is never able to reach the scale it needs to become cheaper than bikes and cars and trucks. If those solutions are good enough for businesses, and tailored to their logistics networks, there’s a chance that those businesses decide not to bend their operations to fit Airbound’s shape.

There is also, to be sure, a world in which those companies win in America, and Airbound wins in India, where customers demand delivery that is fast and cheap, and in similar markets. It is hard to imagine a world where the vast majority of movement occurs on one company’s platform.

That’s the world Naman imagines, though, or sees. Imagine is too fuzzy a word.

Of course he wants to win India. It’s where he and Airbound are from. But he has no desire to build India’s best drone company.

Made in India, Made for the World

Way back in 2020, I wrote about Reliance Industries, India’s largest company. I was excited about investing in India then, and invested in a few startups there. The market was so large, its population so young and talented, its middle class growing, that there must surely be an opportunity to build large startups to serve it. I think I lost money on every one. One of the things that investors have to learn every few years is that the Indian consumer is so value-conscious that it’s hard to get them to pay much for software. And here I am again.

In that Reliance essay, I wrote about Prime Minister Narendra Modi’s Atmanirbhar Bharat Abhiyan, which translates to “Self-Reliant India Mission.” Mukesh Ambani, Reliance’s CEO, extended it, unveiling a new plan at the 2020 Annual General Meeting: “Made in India, Made for India.”

Naman is pursuing a Made in India, Made for the World strategy, and he thinks that the same characteristics that make India so hard for consumer software make it the ideal place to build flying machines cheap enough to become ubiquitous.

“India is a very value-conscious economy,” he told me, “and I think building for value is something that does not exist in Silicon Valley culture but exists in Indian culture.” Whereas Silicon Valley values capabilities, he believes India values efficiency. Way too many companies, in Naman’s opinion, treat making something possible as the critical thing, and making it affordable as something to address later. Serving a huge, price-sensitive market makes cost part of the initial requirements.

It is also, as we have discussed, a place where consumers demand more from their delivery companies. He cites Meesho and Zepto and Blinkit as examples of companies working intensely on delivery economics and speed. He recalls a conversation he had with a founder in SF. When Naman told the founder what he was building, the founder asked something like, “Well, how much faster could you make delivery than 45 minutes anyway?” “Food coming in 45 minutes in India,” he replied, “means that you must live in the middle of nowhere.” Recalling that conversation and his own experiences, he said, “I’m still continually amazed by the low standards that Americans have for deliveries in general.”

And finally, in India, serving a mass market with low prices is the path to profits. “It is by catering to that large demographic of people that you can really make money,” he explained.

If you can make it in India, the logic goes, you can make it anywhere. And if you truly want to move all movement to the sky, to not just deliver things people currently order faster and cheaper but to move people and change the way we store and move everything, you need to design a product that is cheap and ubiquitous enough to meet the exacting standards of the Indian consumer.

Building products and services affordable enough for everyone used to be what America did. “I feel like this is a deeply American value that has just been lost over the past century,” Naman said. Ford described the Model T as “large enough for the family, but small enough for the individual to run and care for,” priced so anyone with a decent salary could own one. Naman’s goal is similar, but airbound: “I want to offer an experience that everyone can afford.”

That is an ideological stand, and an economic one. If you move a little bit of movement to the air, you make people’s lives a little bit more convenient. If you move a lot of movement to the air, you change the shape of the world once again.

Reshaping the World Again

In his Manifesto, Naman predicts the ways in which the world will change if Airbound wins.

What it describes is that Airbound will fit into the existing shape of the world at first, making things a little bit better, faster, and cheaper. Over time, as the world metabolizes the new capabilities, it will begin to reshape itself around fast, versatile, and cheap flight, the way that it has around ships, containers, trains, planes, and cars.

Last-mile logistics will stop batching and send things as ordered, from distribution centers located outside of cities instead of on valuable real estate, and what people can and will order will change, too. They might order food from their favorite restaurant on the other side of town, or decide to rent outfits for each occasion instead of wasting apartment space storing clothes.

People themselves may choose to remain in cities that become more walkable, as drones enter from the top of buildings, like birds, and cars, roads, and parking lots go the way of the dodo. They may also choose to move somewhere with more land and more space, more nature, more family and friends, without giving up the economic opportunity that has long been bundled with urban centers. New York and Philadelphia, my two hometowns, which the Erie Canal sorted into winner and runner-up, will become one accessible metropolitan area. Humanity’s efficient frontier will move outward, and so will we. Marchetti smiles.

Of course, these are just guesses. Naman doesn’t think he needs to figure out delivery handoffs yet, let alone architect the cities and towns of the future.

But there is precedent. Every transportation technology to date has altered our geography in one of two ways: batch machines grouped us into hubs, and cars spread us out along roads reaching out from them. Cheap flight expands the map in every direction, all at once, connecting every point on the map to all others, making the physical world as traversable as the internet.

The internet, with its point-to-point connections, reshaped the way that information moves. It feels obvious now, the water we swim in, but it wasn’t always.

In 1945, Vannevar Bush wrote “As We May Think” about a world in which anyone could reach any piece of human knowledge, instantly, from their desk. At the time, information moved the way matter does now: in batches, on schedules, through hubs, with a human chaperone, a mailman carrying your letter or a paper boy chucking a newspaper onto your lawn. Bush imagined that information technology would offer a better way, and he was right. He didn’t predict which email clients, blogs, and YouTube channels would be popular, just that things like those would become possible.

Fast, cheap, and easy access to information, anywhere in the world at any time changed the world dramatically. It is the reason that a fifteen-year-old kid stuck in his parents’ house in India during COVID was able to learn physics, engineering, and manufacturing, find collaborators, and build his first drone.

His next drones, the ones he built with talented people and more resources and knowledge earned from making the first one, and then the next, and then the next might ultimately do to matter what the internet did to information.

We are so used to our world that we don’t appreciate how deeply we’ve transformed it to fit our transportation technologies.

We enter from the street instead of the roof, because our vehicles and feet are stuck to the ground. Even the stoop I sit on watching the birds is a consequence of our transportation technology. New York’s rowhouses got their high stoops, stairs climbing up a story to the main entrance, the story goes, to lift the front door above streets full of horse shit. Twice a week, garbage cans line up at the bottom of it waiting for a truck, because even trash moves in batches. The Amazon DC I run past in Red Hook takes up land that should be houses, because vans can only carry packages so far, so fast. My great-great-grandparents took a steamship into New York City instead of Philadelphia when they came over from Italy, because the Erie Canal made it economical to send that boat back to the motherland full of food. My wife’s parents came to New York from India because, thanks to those 19th century ships full of immigrants and food, New York became the city of opportunity in the land of opportunity. And my wife and I live here now for the same reason.

We may never move from here, or we might one day move to the suburbs to give our kids a backyard without sacrificing the opportunities this city offers. Either way, it would be a choice instead of a trade-off.

Neither Naman nor I pretend to know what we will choose, what anyone will choose when handed the keys to a new efficient frontier.

“I think my job,” he told me, “is just to make these possibilities obvious.”

Thanks to Naman, Relina, and to those of you who provided feedback.

That’s all for today. We’ll be back in your inbox with a Weekly Dose on Friday.

Thanks for reading,

Packy

Hi friends 👋,

Happy Friday and welcome to our 212th Weekly Dose of Optimism!

Fun week over here at not boring. Feedback on Would We Send Our Kids to HAA?, which I wrote with Alex Danco, has been great - as one person said, if you have kids, want to have kids, or know someone who has kids, you should read it. While we were busy writing, the world’s entrepreneurs and scientists were busy making the world a little bit better.

Let’s get to it.

This Week’s Weekly Dose is Brought to you by… Jack & Jill

One hire can change your company.

Jill helps over 5K startups like Ramp, Strala, Corgi, and Attio make their next great hire.

Jack helps 380K+ professionals find their next best move.

When there’s a match, you get a direct introduction to the company or candidate. No application required.

Thinking about your next move? Talk to Jack. Building a remarkable team? Meet Jill.

Talk to Our AI Agents for Free

(1) Vironexis Turns the Liver Into a Cancer Drug Factory

Here at the Dose, we love a lot of things, but we hate cancer.

In last week’s Dose, we covered a drug that helps T cells attack cancer and a gene therapy that gives kids the instructions to make a missing enzyme. This week, a company called Vironexis shared encouraging results from putting those two ideas together: give the body instructions to make its own cancer-fighting medicine.

On Monday, Vironexis reported that all three patients in a subgroup with relapsed or refractory acute lymphoblastic leukemia achieved complete responses after a single administration of VNX-101, with no residual leukemia detected in testing. The first patient’s response lasted 260 days, at which point their doctor proceeded with a stem cell transplant.

Three is a very small number of patients, but the mechanism is very cool. An adeno-associated virus, or AAV, delivers genetic instructions to liver cells, which produce a protein called GP101. One end grabs CD19 on cancer cells while the other grabs CD3 on T cells, bridging the immune system’s killer T cells right to their targets.

Once a patient takes VNX-101, their liver cells keep making GP101, like a little pharmaceutical factory manufacturing the treatment inside the body.

There are a bunch of qualifications - those three patients had no pre-existing anti-AAV antibodies; the broader Phase 1/2 trial has treated nine people across several blood cancers; they’re early, company-reported results; Cytokine release syndrome and neurological toxicity occurred during dose escalation and resolved with treatment; and there is still a lot to learn about dosing, safety, and how long the responses last - but the promise is enormous. The pharma industry spends enormous effort and money figuring out how to manufacture medicines and get them into people, repeatedly, over and over. If we can safely teach people’s own cells to handle more of that manufacturing, we could give people more time living their lives, less time receiving infusions, and ultimately, longer, healthier lives.

Say it with me… get fucked, cancer.

(2) Solcoa Raises $75M for American Rare Earth Metallization

Not Boring readers will understand that the future depends on rare earth metals, and that China currently mogs the US in its control over that supply chain.

Enter Solcoa. On Thursday, Solcoa announced $75 million to build Solcoa One, a rare earth metalmaking plant in Nevada designed to produce 500 tonnes a year, with operations targeted for July 2027. The financing includes $45 million in equity led by Bain Capital Ventures and $30 million in debt and equipment financing anchored by J.P. Morgan.

Solcoa works on metallization: converting refined rare earth oxides into metals that magnet manufacturers can use. It’s one more step in the chain from mine to magnet to anything that moves electrically that we’ll be able to do here.

The company already produces neodymium-praseodymium and samarium metal at a rate of about 10 tonnes a year in Alameda, but 10 tonnes doesn’t do much against a 100-kiloton-a-year NdPr market and a 2 kiloton-a-year samarium market. The money will help them 50x that, and then 6x that, and if they’re successful, hopefully unlock capital to scale some more.

Its approach combines new chemistry with modular reactors that it builds itself. Solcoa says the process uses less energy and costs less than conventional molten-salt electrolysis. Developing new technology is the key way America will catch up to China.

And there’s progress further along the chain, too. In Dose #169, we covered Vulcan Elements’ $1.4 billion partnership with the government and ReElement to expand domestic magnet production. This week, Axios reports that Vulcan will supply magnets for the Army’s SkyFoundry drone initiative. Vulcan says its magnets will go into the brushless motors that power the drones.

You can start to see a western rare earths supply chain developing. Earth AI can find them. Solcoa can metallize them. Vulcan can turn them into magnets. And Westmag can turn those magnets into motors and actuators to fly drones and actuate robots.

We’re going to need a lot of rare earth metals. Good to see them becoming less rare.

(3) Terraform Makes Methanol From Sunlight and Air

We’re a simple newsletter. Terraform makes something from sunlight and air, we cover it.

Back in Dose #205, we covered Terraform making hydrogen with its solar-powered electrolyzer. Seven weeks later, Casey Handmer & Co are back with their next molecule. On Monday, he announced that Terraform’s reactor team had produced batches of high-purity methanol at its Muroc desert test site. Per Casey, here’s the timeline for those keeping score at home:

Let’s review our progress in 2026. In January, we closed on our Muroc test site. By March, construction was underway. By May, we had 1.8 MW of solar installed. In July, we produced CO2 on site, and in August, Hydrogen.

Methanol is both a useful chemical in its own right, and as a starting point for making more complicated things, including gasoline, aviation fuel, and plastics. Terraform already makes (and now sells) methane, and adding methanol gives it another route into replacing the enormous collection of products we currently get from oil and gas with sunshine and air.

Now, the big question is what cost Terraform can get to at scale. Casey’s strategy is to start with high-purity products that command higher prices, sell them, and use the production experience to drive costs down until the company can compete in bigger, cheaper markets.

The hope is that we’ll be able to manufacture chemicals we previously needed dead dinosaurs to make… and we’re all out of dinosaurs.

(4) Joby Flies Across America Autonomously

At Woodland Cemetery in Dayton, Ohio, I suspect Wright Brothers are flying in their graves. The planes that they figured out how to make fly can now fly themselves.

Last Friday, after we hit send on the Dose, Joby announced that its J208 had flown 3,199 miles across the United States with zero control inputs from the onboard safety pilot, ending in North Carolina’s Outer Banks near Kitty Hawk, where the Wrights took their historic first flight in 1903.

The plane was a converted Cessna Caravan, not one of the company’s electric air taxis, but it served as a test of the company’s autonomous systems, and its autonomous systems passed. It handled taxis, takeoffs, navigation, and landings across a route with stops in Arizona, Texas, and South Carolina, during which it operated at busy airports and rerouted around severe weather.

We’re big fans of autonomy here at the Dose. This summer, we covered autonomous boats, drones and robotaxis, and more drones, and for good reason. Autonomy is one of the things that will help reshape how things and people move around the world, and in so doing, reshape the world itself.

But autonomy is just one piece of the puzzle, table stakes. More to come on the rest next week.

(5) Meta Makes VR Glasses (and more)

Never bet against Zuck!

Back when Dan wrote the Weekly Dose, we included a lot of pro-Zuck content, because Dan is a self-described “Zuck Cuck.” Well, Dan is smiling today, wherever he is (slinging creatine).

Meta held its Meta Connect event on Wednesday, and it announced a bunch of new stuff, both digital and physical. Muse, its AI assistant, has a voice and video now. Meta Glasses can serve as FDA-approved hearing aids. There’s a less creepy pair of Glasses that are audio-only, along with a bunch of new frames. They even released an unexpected Muse Charm, a little guy you can put on your keychain to talk to Muse. Anything to get you away from your Apple iPhone.

The biggest announcement, though, were the Meta VR Glasses, a VR headset weighing just 100 grams. I love it. Zuck has been skewered by critics for the tens od billions of dollars he’s invested in the “Metaverse” and related devices, an investment that was not worth just the Oculus Quest 3. But he persisted against the plan.

Zuck has been super clear for years — I remember an old Joe Rogan or Lex Fridman podcast, maybe both — that the company was going to develop less technologically advanced glasses (the Meta Ray-Bans and Glasses) for daily use, and bulkier, more advanced ones that offer AR and VR experiences occasionally, while you’re sitting down, but that over time, the technology would improve and shrink and ultimately, those two would merge into one thing, a pair of glasses that are comfortable to wear, look cool, and have even more capability than today’s most advanced VR headsets.

This probably isn’t that, yet, but they’re very cool. The glasses have a 5K micro-OLED display, eye and hand controls, and the ability to turn a flat surface into a keyboard and touchpad. Meta is promising more than 100 immersive live US sporting events a year, plus 3D movies and hologram calling. They’re due in spring 2027 for $1,299.99, with up to three hours of media playback on a charge.

More importantly, they’re a sign that the plan is on track, and that ultimately, Zuck is going to be right. Long live the Emperor.

EXTRA DOSES: Science Breakthroughs, Rainmaker, Precision Neuro, HAA

Read more

Subscribe now

Hi friends 👋

This is a special post, written by me and Alex Danco together, where we’re going to throw tech blogging convention out the window and write about a personal topic.

Let’s get to it.

Today, we’re talking about how kids learn, but if your business’ models learn from big, complex files, you should let Reducto help teach them.

Reducto is the document ingestion layer trusted by AI teams at Scale AI, Harvey, Toast, and Vanta to reliably turn complex files into structured, grounded data for production AI systems at scale. Processing more than a billion pages per month, Reducto handles the long tail of difficult documents that break conventional OCR: dense tables, unusual layouts, handwriting, low-quality scans, strikethroughs, and formatting-dependent information.

Already using another parser? Apply for up to $5,000 in migration credits and hands-on support to benchmark r-1 against your current stack on the documents that matter most.

Try Reducto Today

Would We Send Our Kids to HAA?

Yesterday, the Horowitz Andreessen Academy officially launched; and if you haven’t seen it, please go check that out. Gagan Biyani (who founded Udemy) is a co-founder, and will be the CEO, and he wrote a manifesto here on what they’re doing and why. Tldr, it’s a program for young college-aged people (it’s not a college or university; it’s an alternative) to come to San Francisco, learn how to build things and be useful with a like-minded peer group of other ambitious kids, living in a residential setting in SF.

Now, in an early era of our lives as tech bloggers, the obvious way to write about this would be, “What an interesting idea; is this unbundling the university? Is this disruption? Is it some other mental model? And it would make for a great think piece. You know the type.

But the thing is, now we’re both older, and we’re both parents. So this hits different. It’s not just a theoretically interesting exercise for us; there’s a really literal question, “could we imagine our kids going to something like this?” (Gasp!)

Speaking for both of us, we both had a really good time in college. Danco went to McGill, up in Montreal (a great college city, arguably the best), Packy went to Duke, down in Durham, NC (maybe the greatest American city outside of New York). We both had really good experiences, and I think it’s reasonable to say that we both ended up as curious life-long learners that know how to pick up books and read them, and that appreciate whatever “soft” or “social” skills you pick up in a university setting that aren’t strictly utilitarian, but seem important.

Their plan for the Academy is interesting: they clearly aren’t just trying to make it an apprenticeship program. There’s definitely a “bundle of life experience” that’s coming together, in a totally different way than it comes together for most people in traditional college, but definitely valid. I’m not sure you would’ve arrived at something like this if a bunch of experienced professionals sat around and thought hard about “the future of the University” or something. It’s like that Paul Graham essay: “The way to find a good startup idea is definitively not to sit around thinking of startup ideas. It’s to solve a problem that you personally have.” In this case, the good folks at a16z have seen some pretty impressive young people become phenomenally useful (and interesting, and happy) jumping right into building, and the plan for HAA seems clearly to be, “get more people launched on that trajectory”, rather than something more abstract.

A large part of my appreciation for traditional college - but also, a reason why I might like HAA too! - comes from my (Danco’s) experience parenting. The thing about parenting is, before you become a parent you imagine a lot of the “hard parts” of being a parent (especially in the younger years) and you picture all of the typical stuff like changing diapers, teaching them skills, driving them to soccer practice, whatever.

Once you actually have bigger kids you realize those aren’t the hard parts at all. The hard parts are like, “are your kids resilient? Are they kind? Do they take initiative? Do they take responsibility?” and when you think back on it, the real purpose of all these classic rituals and institutions around growing up (e.g. playing team sports) is like, yeah it’s kind of about the sport, but it’s also about the life experience.

Some of the hardest parts aren’t about the kids at all. They’re about your own psychology as a parent. Parents in cities like New York and Toronto and San Francisco are competitive and worried. Kids like certain things, but are they the right things? Will they get our kids into good schools? Will they set them up for success? I (Packy) live in a neighborhood in Brooklyn where parents (not us) pay $65k per year for every year from Kindergarten onwards, because those grades are attached to a good high school, and if you get into Kindergarten, your kids are practically guaranteed admission to that high school. You’re basically paying $650k to make sure your kid is admitted to a good high school, which should, but certainly isn’t guaranteed to, help them get into a really good college, at which point they’ll probably get a good job, at which point, they’re set for life. You don’t want to screw up your kids’ futures, right?

The challenge with that, though, is that who knows what a good job will be in fifteen years? Our parents didn’t send us to McGill and Duke to become bloggers.

But one of the things you learn raising kids is that it’s a lot easier to move with them than against them. If they can find something they love, your effort will go to getting them to stop doing it instead of getting them to start. My (Packy) son is obsessed with chess. Obsessed. If there’s no one to play with, he’ll set up puzzles and play against himself. We finally gave in and started letting him play on chess.com for an hour a day, and we have to pry his little fingers from the keyboard when the clock strikes 60. Our daughter? Couldn’t care less about chess. Absolutely loves gymnastics and music. If we give her a few minutes of YouTube, she’ll ask for the Michael Jackson Name That Song game and breeze through the back catalog. We’re starting puppeteering club at school soon, too, because she talks as her stuffed animals all the time anyway and she’s legit funny.

Do I think my daughter will be a professional song guesser or puppeteer? I don’t know man, the world is getting weirder, anything is possible, but I doubt it! Do I think that being really good at those things will develop her confidence and sense of self and her memory and performance skills and her trust in her own gut? I do!

Do I think my son will become a professional chess player? Unless his brain is 99% his mom’s and only 1% mine, I’m afraid he’ll hit a ceiling, just as a son of mine probably isn’t going to play in the NBA. But during his first chess tournament last year, he came into the auditorium where I was waiting after his first match and I couldn’t tell if he’d won or lost because he looked so happy. (An eight-year-old kicked his five-year-old ass, it turns out.) He’s learning that if he works really hard, he can get better at stuff that seems really hard. It’s right there in his Elo. Resilience and relentlessness seem like good traits for him to hone.

It’s funny writing this now, now that I know the kids. In 2022, when Dev turned 2 and Maya was just 2 months old, I wrote an essay called How Do I Teach These Kids?! about how we were thinking about raising those two little blobs given how fast everything was changing.

Things are changing even faster now, and parents are rightfully worried and confused about what exactly they’re supposed to do to give their kids a good life. Pushing them to get really good at math seems silly, when AI is currently tornadoing through all of the field’s hardest open problems. Math is useful in investment banking too, but are the kids going to be better modelers than 2040 ChatGPT? Should you push them to become a lawyer? That seems risky! Even if senior partners and their relationships remain valuable, are they going to need armies of highly-paid whippersnappers drafting docs and finding precedent? Doctor? Probably your best bet, given licensing requirements. But do you dare risk their future on such a specific path? And what if they don’t want to become a doctor?

Continue through this maze, and you’ll keep hitting dead ends. There is no safe career path today, let alone one that will remain safe for twenty years, which means that there is no set of schools and degrees that can guarantee a safe career into the 2040s and 2050s.

This seems really scary. It is freeing!

We can stop pretending that we can plan our kids’ lives and follow them into whatever lights them up.

In that 2022 piece, when I knew even less than I do now, I synthesized all of the research and conversations I’d had into a paragraph:

Expose your kids to a lot of things early, and then help them go as deep as they’re able to go on the things that excite them the most. It doesn’t matter what those things are; cultivating curiosity and depth is the point.

I’m surprised at how well that’s seemed to work in practice. Our kids are a lot more motivated to learn skills in service of some other thing they really want to do than because we tell them they have to learn something.

Tyler Cowen points out, recently, that the most valuable thing you get from college nowadays is being more likely to get married and stay married. A strong group of friends that gets established together, witnesses life events together, and follows each other through the ups and downs of adulthood is just one of the most valuable things you can have, even if for no other reason than making you accountable to yourself. College friends make great witnesses.

One thing that’s definitely true, though, is that you forge your closest friends and most valuable peer groups by going through real, shared experiences. College is certainly designed as one, although in recent years, some of it has gotten a bit too cushy and resort-y for our taste. Esprit de corps is best forged in non-ideal physical environments, not in luxury accommodations.

To that end, working on real things is a great way to forge those kinds of connections, experiences, and appreciation for the best things in life. One of Gagan’s points in his essay, which is absolutely true, is that most kinds of “real work” mean making things, and college traditionally comes up pretty light in that department.

Kids come with this unbelievable gift that they don’t know certain things are impossible, and so they try to figure out how to do fantastical things. The longer you can keep that alive, the better. I (Danco) have had the pleasure of working alongside some 19-20 year olds on the a16z New Media team who have been unbelievably impressive in terms of the responsibility they’ve taken on, their maturity, and their growth as interesting people. “There’s no way I could’ve ever handled responsibility like that when I was that age”, is the obvious thought. But on the other hand, people have a way of rising to the occasion. Maybe more young people should simply be given the opportunity to try hard things? That’s how we’ve done it, through most of human history.

When I (Packy) wrote about Somos and Panandina founder Forrest Heath last year, we talked about this. He said:

I’d say the thing I’m grateful for, and I don’t really have the ability to take much credit for it, is just somehow I never lost my innate childhood curiosity.

And I think this is something that I like, when I met your kids and they’re just playing around and imagining stuff on the playground in Brooklyn, it’s like, everyone was kind of like that at some point, right? Like, almost all of us had some way we were playing with imaginary forts and imagining a future that didn’t exist or imagining a world that didn’t exist and kind of putting our headspace into it.

And I think for some weird reason—both me being kind of stubborn and a bit skeptical of a bunch of different things, and honestly, I think just random chance—I never got that beaten out of me. And I kind of was always, like, finding my own weird path to sort of come at stuff.

Tangentially, that meant that I wasn’t really competing against the system of, ‘Get higher SAT stores, get into the right college and, check all the right checkboxes.’ And I think there’s this sort of, like, innate builder kind of curiosity in me that I never lost. If anything, I’ve kind of compounded it.

In a recent conversation with David Senra, Ribbit Capital’s Micky Malka talked about this, too, about what young people will do given naivete, a desire to do something, and time to figure it out:

But the more time I spend with them [young founders], the biggest thing and the biggest question that I find asking them is because the oldest of the generation are probably 25, 26, the younger are 18. You ask them what were you doing during COVID? And that is the most fascinating question.

The most successful ones right now were the ones that during COVID were learning and were on YouTube sucking information about anything and learning about everything in life. So when they go and and build companies, they find that first they’re naive, so they think anything is possible.

We’re talking about founders here, but you can find the same pattern in exceptional outcomes in any field.

Taylor Swift’s dad was a stockbroker and her mom worked in marketing at a mutual fund; they had no grand musical ambitions for their daughter. But she got obsessed with musical theater, then with country music, and started lobbying her parents to take her to Nashville. When they did, on Spring Break at age 11, she walked into record label offices by herself to give them demos of her singing and try to get signed. They didn’t sign her. A lot of people in Nashville could sing. She needed more skills: guitar and songwriting. 12-year-old Taylor learned a few chords from a computer repairman, and she wrote her first song shortly after. Story goes she practiced guitar so many hours a day that her fingers bled and cracked and her mom had to tape them up. Her parents drove her to New York for voice lessons. They got her a slot singing the national anthem at a 76ers game. At 14, the family moved to Nashville so she could go for it. Sony/ATV signed her to a publishing deal that year, the youngest songwriter they’d ever signed. Two years later, in 2006, she released her first single, Tim McGraw.

If you’re a parent, you know that it’s impossible to get your kids to practice anything until their fingers bleed if they don’t want to themselves. You also know how hard it is to get them to stop if they do.

There has never been a better time to be a kid who wants to figure something out. I think the best we can do as parents is help them figure out what that something is, and then give them every opportunity to pursue it. At some point, that means letting them outgrow what you can do for them finding other people who are as obsessed as they are.

That’s how I read what Gagan and the team at HAA are doing: creating a place for young people who are weirdly obsessed with building something to surround themselves with other young people who are, too, and an environment in which they can more easily access the skills and experienced people they’ll need to do it. A Nashville for Nerds. It’s taking kids seriously, and adding real stakes early.

So could I imagine sending Dev or Maya to HAA?

No idea. It’s not going to be up to me. Personally, I’d love for them to go to Duke, to have the experience that I had, but they’re their own people. If I need to pry their fingers off of making things that the world needs, it would be awesome to have a place where they can go as hard as they want surrounded by other people their age who push them to go even harder, and I’m glad Gagan and team will have twelve or fourteen years to make it great by then.

But I also hope that there are dozens of options like HAA at that point, for all manner of passions. For songwriting, puppeteering, engineering, mountaineering, imagineering, and for pursuits that we can’t imagine today. I hope today’s great universities redouble their commitment to what makes them great, to support those kids, like me, who haven’t figured out what they want to be when they grow up by the time they’re eighteen. Some people, like Jamie Beaton, might just want to keep learning forever, and I hope that we’ve created enough abundance that they can.

A good future is a world in which, in the words of Isaac Asimov, kids have the agency, curiosity, and support to pursue Whatever They Wish.

So, maybe the right way to think about it is, a dozen years from now when our oldest kids start making those kinds of choices, it really won’t be up to us at all. If we’ve done our jobs right, wherever they go to college - or HAA or something like it, or something else entirely - will be whatever they make of it. And we’ll be delighted to send them off.

Thanks to Danco for co-writing with me - it’s always a treat.

That’s all for today. We’ll be back in your inbox Friday with another Weekly Dose.

Thanks for reading,

Packy

Hi friends 👋,

Happy Friday and welcome to our 211th Weekly Dose of Optimism!

Lock-in season is in full swing and the Dose is bursting at the seams. But I believe in you. Once you’re finished, you can find a lot more to read in the new not boring library, which includes many of the books and essays I’ve linked to most in not boring.

Let’s get to it.

This Week’s Weekly Dose is brought to you by… Reducto

The Only Document Parsing Product You’ll Need

Reducto is the document ingestion layer trusted by AI teams at Scale AI, Harvey, Toast, and Vanta to reliably turn complex files into structured, grounded data for production AI systems at scale. Processing more than a billion pages per month, Reducto has built its reputation on the long tail of difficult documents that break conventional OCR: dense tables, unusual layouts, handwriting, low-quality scans, strikethroughs, and formatting-dependent information.

Its new frontier parsing model, r-1, combines layout detection, reading order, tables, formatting, grounding, and granular citations in a single request. It reduces parsing errors by up to 20%, improves latency at high volumes, and costs 1¢ per page all-in.

Already using another parser? Apply for up to $5,000 in migration credits and hands-on support to benchmark r-1 against your current stack on the documents that matter most.

Apply at Reducto.ai/migrate

(1) A Cancer Trial Gives Ben Wilson Reason for Hope

Ben Wilson has spent years studying people who changed history on his podcast, How to Take Over the World (which is excellent if you don’t already listen). In July, the father of four shared that he’d been diagnosed with stage 4 high-grade neuroendocrine cancer, which had spread to his liver, lungs, spine, and brain. In The Cancer Episode, he said his doctors had told him to expect roughly another year.

He also said that we live in an age of miracles, and that he was working to find one.

On Tuesday, Ben shared some encouraging news about a treatment his family has been raising money to help him access: tarlatamab.

Amgen and AstraZeneca reported positive results from DeLLphi-305, a randomized Phase III trial involving 563 patients with extensive-stage small-cell lung cancer. After initial chemotherapy and immunotherapy, patients received either durvalumab alone or durvalumab plus tarlatamab. The combination significantly improved overall survival and the time patients lived without their cancer progressing. Tarlatamab connects cancer cells bearing a protein called DLL3 with the immune system’s T cells, helping them attack; durvalumab blocks one of the mechanisms tumors use to suppress that immune response.

The detailed survival numbers aren’t public yet, and results in the trial’s specific population don’t establish what will happen in Ben’s case, but it seems to be a positive result and one that might give him a chance to live longer.

Stage 4 cancer is always dangerous but I might have many more high-quality years in front of me.

Given the pace of progress in cancer treatments, slowing cancer’s spread and giving people more time might mean that they live long enough for a more complete cure, something that knocks the cancer out entirely.

We cover a lot of cancer studies here, but given that I’ve been a long-time listener of Ben’s podcast, this one is a reminder that behind every diagnosis and survival curve are people who want to see their kids grow up, spend more time with their partners, and keep doing the work they love. Progress means giving more of them that chance.

Rooting very hard for Ben. Get fucked, cancer.

(2) First Treatment for “Childhood Alzheimer's” Approved

Ultragenyx

Until doing the research for this week’s Dose, I didn’t know there was a disease nicknamed “Childhood Alzheimer’s.” As someone terrified of getting Alzheimer’s in old age who also has two young kids, I can’t think of a more tragic disease than this one, Sanfilippo syndrome type A.

Kids are born seemingly healthy, start developing normally, and then, because they're missing a single enzyme, a sugar molecule called heparan sulfate builds up in their brains. They lose the words they've learned, then the ability to walk. The median life expectancy of those inflicted is 15.

There has been nothing doctors could do for these kids except manage symptoms… until yesterday.

On Thursday, the FDA approved Fayuvi, Ultragenyx’s gene therapy and the first treatment ever for Sanfilippo type A. It’s a one-time IV infusion that uses an AAV9 virus to deliver a working copy of the broken gene so the body can make the missing enzyme itself. In the pivotal trial, 17 treated kids scored 23.5 points higher on a standard cognitive scale than 27 untreated kids from a natural history cohort, with follow-up now stretching to nearly eight years. Treated kids kept learning during the years the disease would normally have stolen their ability to do so.

Despite the treatment’s massive potential to improve kids’ lives, it almost didn’t make it to this point. Haiyan Fu and Doug McCarty built the therapy at Nationwide Children's more than a decade ago and licensed it to Abeona, which ran out of money despite good data. They handed it to Ultragenyx, who submitted to the FDA for approval. Then last July, the FDA rejected it over manufacturing issues, even while calling the clinical data robust.

Now, Ultragenyx has fixed the plants, resubmitted, and expects to ship within 30-60 days.

It works best given early, before the damage is done, so the next thing that needs to happen is newborn screening. But there’s now a treatment for one of the most appalling diseases known to humankind.

Age of miracles.

(3) Natural General Intelligence

Ryan Orbuch for Lowercarbon

When Will O’Brien and I wrote about The Great Blue Frontier, he shared that “we have mapped only 27% of the ocean floor with modern sonar, and much of that is coarse, low resolution.” The ocean is just one very big domain that we don’t have enough data for, but in his latest essay, Lowercarbon partner Ryan Orbuch argues that we don’t know nearly enough about the planet generally, that we now have the tools to get all that data, and that we should get it and put it in a model for Natural General Intelligence.

I got a chance to read an early draft, and what struck me is that this problem is an awful lot like biology, where models have been useful in making sense of enormous amounts of ever-changing data that’s too complex for humans to understand holistically. Using models to better understand our planet so that we can intervene, like we do with our own bodies, is an ambitious goal, and I think a worthy one. We have the technology.

Since I read the early draft, it got even better, and Ryan added a bunch of beautiful and interactive graphics, so you should go spend some time better understanding the state of understanding our world.

(4) ChatGPT & RLHF Co-Inventor Releases TypeSafe AI

As smart as it seems to be, AI is still terrible at writing. So why not just make one that shuts up and gives you the answer?

Diogo Almeida, who helped develop the research behind ChatGPT, has spent the past two years building a model for software that needs decisions. On Tuesday, his company, TypeSafe, introduced Jev.

Jev takes information and questions with predefined answer formats, then returns decisions and probabilities. Which department should receive this support ticket? Does this transaction look suspicious? How likely is this customer to cancel? It produces those answers in parallel, without generating a paragraph one token at a time.

Almeida claims Jev produces 20–200× faster responses and 40–400× lower costs than frontier models on the decisions Jev is designed for. They’re company reported numbers, and I’m sure third-party benchmarks are coming, so take them with a grain of salt, but it makes sense. It’s not burning tokens writing long answers or figuring out how to get humans to understand what it does.

It seems like a smart way to increase companies’ Return on Tokens (ROT), and the idea is that if intelligence gets cheap and fast enough, developers can squeeze it into all the little decisions that previously required brittle rules or a human review.

People keep sharing examples of what Jev can do, and they’re very cool! Check out here, here, and here, or just scroll through the Quote Tweets on the original. The tools keep getting better!

As it gets cheaper and more efficient, we’ll use more of it. That’s why they named it Jev, after William Stanley Jevons, of Jevons paradox fame. It’s also a step towards recognizing what models are and are not good at and building capabilities around those strengths and weaknesses. I’m just happy that we might get to read a little less slop to get our work done.

(5) Thatch Raises $108M Series C at $1B

When I wrote about Thatch last October, my thesis was that changing who controls healthcare dollars could change the healthcare system. Give individuals choice in how they spend, and healthcare providers would need to develop offerings that compete for their dollars.

On Tuesday, Thatch announced a $108 million Series C at a $1 billion valuation. Its ARR has grown nearly 7× over the past twelve months, and more than 5,000 employers now use the platform. Investors include The General Partnership, Index Ventures, General Catalyst, and a16z, alongside strategic investors including Eli Lilly, ADP Ventures, and Paychex. Not Boring Capital has been an investor since 2022.

Thatch makes it easy for employers to give employees a tax-free healthcare budget through ICHRA. Employees choose their own individual insurance plans and can use remaining funds for eligible medical expenses. Thatch handles the financial and administrative work behind the scenes so that individuals can focus on picking the best plan for themselves and their families.

The bet is that insurers and providers behave differently when the person receiving care controls where the money goes. They have to compete for that person’s business. And if someone can keep their insurer when they change jobs, the insurer has more reason to invest in keeping them healthy over time. Capitalism, baby!

In order for Thatch to change the big, sclerotic, heavy, embarrassing healthcare system, though, it needs to get big enough to matter to insurers and providers. Thatch needs to aggregate enough demand to persuade carriers to build better products, which could attract more people, which would pull still better products out of the market.

More capital is both a sign that the company is scaling well, and fuel to scale further still. There’s maybe nothing with the same combination of size and dysfunction as the US healthcare system; if Thatch can help fix it, there’s nothing we can’t do.

EXTRA DOSES: Science Breakthroughs, Perez AI, Steve Martin’s Ears, World’s Most Educated Person, How to Unclench

Read more

Hi friends 👋,

Happy Friday and welcome to our 210th Weekly Dose of Optimism.

Today is the 25th anniversary of September 11th. I still remember exactly where I was, in my first week of high school on my way to the school chapel, when the news hit. It’s hard to believe that so much time has passed.

I’m sending at exactly 8:46am, 25 years to the minute when the first plane hit the North Tower.

I’m not going to outwrite Colston Whitehead on this one, so instead, I’m going to share the essay he wrote two months after the attacks: Lost and Found. I read it every year on September 11th, and it’s the best thing I’ve read on what it means to be a New Yorker. This short ESPN documentary on Welles Crowther, who sacrificed himself to save 12 lives, is a sober reminder of the real people who died in, and survived, the attack.

The world changed forever 25 years ago. In some ways, it’s gotten worse, and in many, the world in 2026 is better. We focus on those. But never forget.

Let’s get to it.

This Week’s Weekly Dose is brought to you by… Reducto

The Only Document Parsing Product You’ll Need

Reducto is the document ingestion layer trusted by AI teams at Scale AI, Harvey, Toast, and Vanta to reliably turn complex files into structured, grounded data for production AI systems at scale. Processing more than a billion pages per month, Reducto has built its reputation on the long tail of difficult documents that break conventional OCR: dense tables, unusual layouts, handwriting, low-quality scans, strikethroughs, and formatting-dependent information.

Its new frontier parsing model, r-1, combines layout detection, reading order, tables, formatting, grounding, and granular citations in a single request. It reduces parsing errors by up to 20%, improves latency at high volumes, and costs 1¢ per page all-in.

Already using another parser? Apply for up to $5,000 in migration credits and hands-on support to benchmark r-1 against your current stack on the documents that matter most.

Apply at reducto.ai/migrate

(1) OpenAI Publishes Its Navier–Stokes Proof

OpenAI

Navier-Stokes is no Navier-Joke.

While some frontier lab employees were telling us that there’s a good chance they’ll kill us all, OpenAI announced that a group of roughly 10,000 AI agents had produced a proposed solution to the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. The agents worked for about 88 hours using an unreleased model more capable than GPT-6 Astra. Formalizing and verifying the result in Lean, a math programming language and theorem prover, took another 17 hours.

The question is whether the equations describing a smooth, three-dimensional fluid can develop a singularity: a point where the math blows up. OpenAI’s construction starts with fluid at rest, applies a smooth external force, and produces unbounded velocity while keeping total energy finite. It addresses the forced version of the problem. You can read the 166-page proof and inspect the Lean code yourself, if you are incredibly bored this weekend.

Of course, there was controversy. Anthropic was working on its own Navier-Stokes solution with a mathematician, and they accused OpenAI of using the logs from their Codex conversations to come to their own solution. Sam Altman denied the charge but said that rumors that Anthropic was close did spur OpenAI to give it the ol’ college try themselves.

Whatever happened, this is a glimpse into the future: AIs doing all of our work for us while we spend time doing more human things, like arguing for credit.

(2) Google DeepMind Maps Nine Billion Possible DNA Changes

Pushmeet Kohli and Žiga Avsec for Google DeepMind

It’s safe to say that Google DeepMind is getting its butt kicked on the frontier LLM side, but it’s still doing some of the coolest work in science AI.

On Tuesday, it released AlphaGenome Atlas, a one-petabyte database of predicted molecular effects across the human genome. It includes the 98% of our DNA that doesn’t directly encode proteins, where changes can still affect when, where, and how genes get used. Researchers can look up variants through a website and use a combined impact score to figure out where to look more closely.

Basically, the human genome contains about three billion letters. At each position, you could swap the existing letter for one of three others. That gives you roughly nine billion possible single-letter changes. GDM predicted what all of the changes might do.

Like at the Broad Institute, the Atlas helped researchers identify a DNA change predicted to disrupt RNA splicing, providing supporting evidence in an unresolved rare-disease case.

To be clear, it’s not a replacement for actual experiments, but those take a lot of time and money and Atlas gives researchers a better sense of where to spend their limited resources.

“Organize the world’s information and make it universally accessible and useful” now extends to life’s … alphabet.

Speaking of cool stuff GDM has done, check out all of the things that people have been doing with its complete map of fruit fly neurons:

(3) Hermeus Uses Fast Planes to Go Faster

Hermeus

As we covered way back in our Deep Dive on Astro Mechanica, the challenge with ramjets is that they need to be moving very fast before they start working. Forward motion compresses the incoming air, which means you need some other way to get the engine up to speed in the first place.

Luckily, Hermeus has a really fast plane, the Quarterhorse.

On Wednesday, the company unveiled Ramjet-X, an expendable test vehicle designed to launch from its reusable Quarterhorse aircraft. Quarterhorse gets it to the necessary speed and altitude, releases it, and returns for another mission, then Ramjet-X keeps flying under its own power.

The company plans to sell access to this setup as Flight Test as a Service. Customers could test sensors, materials, autonomy software, communications equipment, and propulsion technology in sustained high-speed flight without having to build an entire test program themselves. Engine testing is underway in Florida, with integrated vehicle testing planned for 2027.

Hermeus is special because of its fast development of fast planes. Now, it’s using one of those planes to light up a ramjet platform that will let others develop faster, too.

(4) Michael Levin Publishes Ingressing Minds Paper

Michael Levin

We don’t include every Michael Levin paper in the Dose. Man is far too prolific. But we include a bunch, and this might be the most important of them all.

On Wednesday, Levin published the revised, peer-reviewed version of Ingressing Minds. He proposes that minds are patterns in a structured, nonphysical space, and that bodies provide interfaces through which those patterns enter the physical world. He argues that many things beyond humans and animals can be interfaces, including computers, which is the most compelling argument I’ve heard that AI might be able to become “conscious” or “alive.”

Platonic space is a little tricky to understand, but it’s simpler if you start with triangles. If you know two angles of a triangle, you get a third “for free.” In the paper, Levin asks whether that idea extends to much richer patterns, including patterns with agency. In his framework, some of the things inhabiting this “Platonic space” are minds, and biological or engineered systems let them become embodied.

Levin himself came to the idea from the biology side: cells can assemble into unfamiliar living constructs with capabilities that weren’t explicitly designed into them, before any replication or evolution happens. He’s now systematically investigating where those capabilities come from and how to access more of them.

In his announcement, Levin describes the pushback Ingressing Minds has attracted, then adds: “if you think this stuff is weird and uncomfortable, just wait.”

I can’t. I can’t wait. I think Michael Levin will go down as the most important scientist of this era, and this may be his most important paper. Worth a read.

(5) Neko Health Launches in New York City

On Tuesday, I got invited to a preview of the new Neko Health location in Soho.

Neko is a preventative health company founded by Hjalmar Nilsonne and Spotify founder and CEO Daniel Ek. I’d seen a bunch of tweets about the experience in the UK and Stockholm, where it launched in 2023, and I’ve been wanting to try it. As soon as I stepped through the front door, it delivered.

I don’t want to spoil it, but it feels like walking into the future, from the stairs to the waiting area to the screening room that felt like what it would feel like if friendly aliens invited you onto their spaceship and offered to make you right. There’s a full-body scan, some bloodwork, cuffs, and a bunch of lasers and lights designed to understand your heart, blood pressure, body composition, a bunch of markers, and more. The people were friendly, calm, and professional, and the screens kept you posted on the progress. The whole thing flew by, and as soon as it finished, they took me into a room to share and explain the results.

Luckily, everything was clear - there’s some inflammation to fix and I could do to lose a few pounds / points of BMI, but it was a relief to know my heart and cholesterol are in good shape. I haven’t been to a traditional doctor in a very long time.

The reason this is Dose-worthy is that they do it for $499, with a goal of squeezing as much into that price as they can over time. It combines people, design, and technology to pull it off. And with a medical system that’s focused on fixing things once they’re broken, a delightful, accessible network of preventative health locations will help catch things before they break, when it’s easier and cheaper to fix. Plus, it’s nice peace of mind.

It’s great to see people who’ve had success in tech like Daniel and Shak put their resources towards fixing one of humanity’s biggest problem in a way that serves a lot of people instead of just a few who could afford something more successful.

Big thanks to Shak and Sarah for the invite. They’re not a sponsor and I’m not an investor, but you can move up in the line using my link here.

EXTRA DOSES: Science Breakthroughs, Boring Company, UAPs, Woods & Johnson

Read more

Hi friends 👋,

Happy Thursday! We dropped a new notboring.com website this week, and this is our second essay of the week. Lock-in season.

This one is a continuation of the series arguing that strategy is important and moats matter. It’s still popular for startups to say that they don’t need them, that they’re just faster and better at product, that it’s too much to ask such young companies to have Network Effects or Scale Economies in place.

But startups don't get a pass on strategy because they're young. Luckily, there’s one moat that they can use to buy time. This is an ode to that one.

Let’s get to it.

Today’s Not Boring is brought to you by… Arena Magazine

Arena Magazine is the print magazine of American technology, capital, and industry. They publish four beautiful issues per year, printed to last on archival paper, with great writing about people chasing frontiers in America. Their latest is Issue 009: Hello World, with 136 pages of writing and art on the epic story of software.

Subscribe to get Issue 009 delivered to your doorstep.

Subscribe to Arena Magazine

An Ode To Counter-Positioning

There are 7 Powers, but if you made me choose a favorite, I’d pick counter-positioning.

In his primer on Hamilton Helmer’s 7 Powers, Mind the Moat, Lindy CEO Flo Crivello defines it as “the practice of developing your business model such that incumbents have conflicting incentives preventing them to compete effectively.”

The definition contains the reason I enjoy it so much: before they’re old enough to access the other 6 powers, it’s the moat with which startups are able to attack incumbents. Incumbents have bigger businesses and more resources, and it is those advantages startups use to attack them. Their business is too good to mess with.

When I wrote about Ramp for the first time back in December 2020, I compared counter-positioning to the Five Point Palm Exploding Heart Technique from Kill Bill. When a startup hits them with counter-positioning, the incumbent might not realize they’re dead; take a step in the startup’s direction, though, and they topple over.

If incumbent corporate cards incentivized customers to spend more by giving them points, Ramp built its business on helping customers spend less. Incumbents couldn’t match that for a number of reasons, including the fact that they were nowhere as good at software, but mainly because if customers spent less, they made less, which their shareholders would not appreciate. Other startups like Divvy and Brex played the same game incumbents did, offering rewards and points for spend. They both had good outcomes: Bill.com bought Divvy for $2.5B and Capital One bought Brex for $5.15B. Today, Ramp is valued at $44 billion and launching a flotilla of products that help customers save time and money, including a model router.

If you come at the king, you best not [play their own game but with better technology].

I am thinking about counter-positioning today because of Base Power Company CEO Zach Dell’s conversation with David Senra on David Senra. At 16:18, discussing how he and Justin came to their business model, he gives a sermon on counter-positioning.

“If you’re going to take on an incumbent,” he says, “my view is the best way to do it is to have a counter-positioned business model.”

If I showed up and said, ‘I have... the best home battery on the market... and I’m going to sell it to you for 10% below the big guys,’ your margin is my opportunity, right? Like these big guys are going to compete on price. First they’re going to copy my product, then they’re going to drop their price and they’re going to run me out of business.

But if I show up and say, ‘I don’t sell batteries. I sell electricity,’ and we’re going to install this battery on your home, and because of our different business model, you pay 1/20th or even 1/40th of what you’d pay to own it outright, then I can win.

And if you want to compete with me and you’re one of the incumbents, you’ve got to completely change your business model, which as you know for a public company is very hard, right?

Base won’t be able to use counter-positioning forever. The moats it digs to protect its margins from competitors will come from the other powers: Scale Economies, Cornered Resources, Switching Costs, Brand, and perhaps even Network Effects and Process Power. But as Zach points out, there is no better way to start taking on an incumbent than by counter-positioning.

This is a theme that Ben and David bring up again and again on the Acquired podcast: “As Hamilton [Helmer] would put it, counter-positioning is usually a take-off phase power.” Helmer himself went on the pod and explained why: “It’s the only one that’s a partial source of power. If you want real durability… you want to have another source of power.”

Counter-positioning is here for a good time, not a long time. Which is part of what makes it so fun. It’s the trickster power. Counter-positioning buys you time against the incumbent — which is positioned in some way against which you can counter-position — but it doesn’t necessarily protect you against another startup. You need other moats for that.

Poetically, Helmer chooses Dell to make the point. Dell was counter-positioned against Compaq because Compaq’s dealer channel made direct sales painful. Eventually, though, everyone could adopt direct sales, and Dell’s counter-positioning power disappeared. By then, Dell had used the window to build Scale Economies around its direct, just-in-time model.

The Acquired canon is chock full of examples.

When Google launched Android, it could afford to give it away for free because the mothership made its money on search. It could afford to just subsidize the losses completely, but it also didn’t have to, because more mobile search on Google meant more revenue. Nokia, on the other hand, had to make money from the mobile stack itself; giving it away for free would have killed the business.

This is one good way to think about counter-positioning: can Challenger A do X because it makes its money somewhere completely different from Incumbent B?

It, to be clear, is easier to do if you have Google’s cash machine behind you, but it’s always worth thinking about how you can fuck with the way that your incumbent competitor makes its money.

Amazon, born on the internet, was built around direct-to-consumer fulfillment. Barnes & Noble was optimized around stores. For B&N to go all-in on Amazon’s model would have meant reallocating capital, executive attention, and distribution infrastructure, and tolerating lower near-term profitability from its existing stores. The Amazon model “would be less profitable for them… versus the hugely profitable stores,” Ben said. Amazon “counter-positioned against everyone whose cost structure was set up” for physical retail, David added.

This is another way to think about counter-positioning: the more an incumbent has invested in building out its infrastructure, the less able it is to adapt.

Startups with less money than Google can certainly use this form. My favorite example is Somos Internet. As I explained in Cable Caballero, Somos designed a new network architecture and built its own hardware to run it, and incumbent telcos can’t respond because doing so would require ripping up the billions of dollars of CapEx they’ve spent buying and upgrading third-party hardware in the old network configuration. You can see their stuckness. Plus, given their reliance on vendors, most don’t have the technical wherewithal or nimbleness with which to re-architect the whole thing anyway. Plus plus, they’ve used so much debt to finance the network buildouts that they can’t really lower prices to hurt Somos, let alone fund an uncertain new network buildout.

There’s the flip side of both of the above, too, which is familiar to anyone who’s watched a movie where a bad guy gets leverage on a good guy by kidnapping their family: having nothing to protect is freeing.

When Microsoft took on the company nobody got fired for buying, IBM, “Microsoft basically had no baggage,” Ben explained. IBM sold integrated computer systems - software, hardware, everything - through the greatest enterprise salesforce in the business. To launch its PC, urgently, Big Blue broke with its normal model, used off-the-shelf components, an Intel processor, and even an operating system supplied by Microsoft. But Microsoft didn’t license exclusively to IBM; it licensed DOS non-exclusively, and was incentivized to make PCs as cheap and abundant as possible in order to get its software onto as many desks as possible. Per Ben & David, they could say: “We don’t need to make any money on hardware. We don’t need to even make hardware.”

Microsoft didn’t win because it wrote better software than IBM. IBM wrote software that was just fine eventually. But IBM needed its software to sell high-margin machines, whereas Microsoft benefited when the hardware layer got commoditized. As hardware commoditized, the software layer became the one with which all of the hardware providers had to be compatible. Or as Ben put it, Microsoft was “free to become the whole point of integration for the entire ecosystem just by shipping bits.”

This is a vicious form of counter-positioning: an entrant can be counter-positioned not just because it is willing to cannibalize a profit pool, but because it actively benefits from the destruction of the incumbent’s profit pool.

Then there’s Facebook, which faced off against MySpace and its 1 million users when it launched in 2004. MySpace is a joke now, but was a rocketship then. It launched in 2003, got to 1 million users in February 2004, passed Friendster the next month, and quintupled to five million users by November.

Facebook counter-positioned by saying that mashing a lot of users in one network to start was bad, actually, and that it was better to start small, to seed a network with the gentlemen and ladies of Harvard and to grow in little circles.

If you’re MySpace, and growth is good, and these nerds are building a little thing for their nerd friends, what are you going to do? Stop growing?

But it turned out that it was better to start small, to seed a network with the gentlemen and ladies of Harvard and to grow in little circles. Not forever, of course, but in the beginning. Facebook used the counter-positioning period to build a ravenously passionate network of people talking to their new college friends and poking each other and made the product so appealing to those on the outside, younger and older, that when they opened the floodgates, it FLOODED, and they moved on to Network Effects.

People haven’t learned.

Earlier this week, a16z partner Josh Elman tweeted, “The new moats are the same as the old moats. Every few years, we fall in love with shiny new tech and forget the basic physics of consumer software.” I’d expand it to say that we forget the basic physics of business strategy, but we’re on Facebook, so let’s stick with consumer.

Assistants are the buzzy product du jour, and Instinct the buzziest of those. It’s a wonderful experience, a better product than the assistants that came before it.

But as Ben says, “Being better is not counter positioning.” The Strategy Don, Michael Porter, writes that “Operational effectiveness means performing similar activities better than rivals perform them. Competitive strategy is about being different.”

Better is a benefit; Power requires a barrier. Counter-positioning exists when the incumbent sees that you’re better but cannot copy what makes you better without damaging the business it already has.

(For the avoidance of doubt: you can’t just pick a model that’s so bad that anyone, including you, would be damaged by pursuing it.)

Who would be damaged by releasing helpful agents? The Labs have kind of done it in their desktop apps, Grok Bot has done it, other new startups are doing it (and it is definitionally impossible to counter-position against new startups), and still Instinct is better. Because it’s better, it raised a $350 million Series B at a $2.5 billion valuation in late August, and one of the theories was that they raised so much so they could continue to offer it for free and drown competitors on the way to new business models (i.e. transaction fees).

Better, though, is not a moat, and good luck drowning Meta’s cash machine.

Meta (née Facebook) released their Assistant this week, too. Its name is Muse.

Some early reviews say that it’s better than Instinct. I still need to try it; better is not the point.

The point is that not only is it not painful for Meta to copy Instinct’s model and product, it’s great! Meta knows more about you than anyone and can recommend things you didn’t know you needed, solving the cold-start problem. They can, as Ben Thompson wrote, give each user “a VM with 8GB of RAM and 8GB” which is “a real-deal computer, and Meta is offering that to everyone in the United States, and eventually the world. It’s pretty extraordinary!” They can integrate it directly into WhatsApp, their enormous but undermonitized messaging app, they can integrate it with FB Marketplace so you can get what you want for cheap, they can put it in Meta Ray Bans so you can just tell your own face what to do, and they can train it in their gigantic data centers. They can give it away for free until the heat death of the universe (or until Anthropic hits on its >10% chance of killing us all). They can bring Network Effects and Scale Economies and Brand and whatever moat is necessary to bear, because they counter-positioned early and then dug all of those.

Now, Zuck has tried to copy or create new products before and failed, or failed to kill his target. Threads hasn’t cured my Twitter addiction. But Twitter had Network Effects. It was moated.

There are, to Josh’s point, too many startups out there that think that the new technology is so incredible that they no longer need moats. They just need speed and taste or whatever and focus and nimbleness and the special je ne sais quoi that comes with being a startup.

And there are plenty of startups who will do just fine with that approach. The big moated companies are willing to pay a lot of money to bring some of that je ne sais quoi inside their castles and to keep it out of their competitors’.

I couldn’t point to Cursor’s moat, but it was an excellent product and it was worth $60 billion to SpaceX. Nvidia paid $12,930,300,000 for Hugging Face, which I actually used as an example of an AI company that used a Complexity Uncertainty Window to develop early Network Effects in When to Dig a Moat. I talked to a founder not long ago and asked about moats, and he said something like “We don’t care about moats we want to build a great product” and so I passed and the company is valued at like 50x what it was then, so what do I know.

It’s rumored that Meta (and others) even offered 10 figures for Instinct!

But when Instinct turned them down, Meta said, “Fine, we’ll build it ourselves,” because it cost them nothing but money and they have plenty of that. Of course, it’s probably easier than ever to copy unmoated products now that AI can write the code, but that’s besides the point, because without moats, any good company will get copied, even if the humans have to do it themselves.

We are in a period in which AI is so shiny and new that incumbents are willing to pay a lot of money for products that are better than anyone else’s, but acquisition hope is not a strategy and better is not a moat.

I hope that Instinct wins, as I hope that many startups take down their industries’ Goliaths. They are fighting. Just yesterday, they introduced a Trusted Person network, to let people’s assistants talk to those of their significant others, family, and friends.

It is an attempt to create Network Effects, and they’re going for it. Respect. The question will be whether they’ve given themselves enough time to build a real Network Effect before people start using Muse, and Muses inevitably work with each other, too.

Time is why counter-positioning is such a powerful weapon to the otherwise unshielded.

It makes it painful, suicidal, even, for someone to copy your better product. You need to use that time to sprint like hell until you can establish your long-term moats.

That’s all for today. We’ll be back in your inbox tomorrow with another Weekly Dose of Optimism.

Thanks for reading,

Packy

Subscribe now

Hi friends 👋,

Happy Tuesday and welcome back from the long weekend!

It’s beautiful out in NYC and it’s time to get back to work. So here’s a short and fun one about why everyone is reading the same old book that I am.

Let’s get to it.

Today’s not boring is brought to you by… not boring world

If you enjoy not boring, there’s even more for paid members at not boring world. Get a bigger Weekly Dose, updates from an American engineer in China, shorter off-the-dome notes from me, essays co-written with smart founders, and more.

Subscribe now

Everyone is Reading Lonesome Dove

There is a phenomenon called Baader-Meinhof where when you discover something, you start seeing it everywhere. It’s not that the thing is suddenly everywhere; it’s just that you notice it now.

So maybe it is because I’m reading Lonesome Dove that I’ve noticed that everyone is reading Lonesome Dove.

I mentioned the book in a Slack I have with eight friends. One wrote back that he’s reading it now:

Another shared that his mom is reading and loving it:

And three, including the one whose mom is reading it, said that they are going to read it, which makes five people in a group of nine.

This is heartening, because they say that no one is reading anymore. It turns out that people are, just they’re only reading Lonesome Dove.

Let me be clear: Lonesome Dove deserves to be read. It is a Pulitzer Prize winner. If you haven’t, you should go read it right now. But it is also weird that everyone is reading it! It was published in 1985 and won the Pulitzer in 1986.

So Why Is Everyone Reading ‘Lonesome Dove,’ an 858-page Western From 1985? That’s the question that Esquire asked and attempted to answer on March 20th, 2025.

The author, Michael Sebastian, had seen a post on a friend’s Instagram about the book and decided like a year later, driving past a bookstore, that he had to buy it. He tried to figure out why it seemed like everyone was now, like him, reading the book and traced it to a few things:

⁠McMurtry died in September 2023, the same month that a biography about the author and a big story in The New Yorker about him came out. This confluence of factors was likely the spark. The kindling is that all things western are in the Zeitgeist: the Taylor Sheridan universe (Yellowstone, 1923, Landman), the new Netflix show American Primeval, Kevin Costner’s epic Horizon: An American Saga, Pharrell’s western-wear fashion show for Louis Vuitton in January 2024.

Then he asked why it resonated so much and concluded that it was something to do with the fact that it deals with something a lot of men are feeling, which is that we’re never going to be the man we want to be. And maybe! That hasn’t been my experience; I’ve enjoyed the characters and the simplicity of the prose and the unvarnished look at life on the frontier in the 19th century.

But there are a lot of books that might resonate. Why this one? The McMurtry profiles / death timing and Yellowstone / Landman popularity felt like they would have been enough to get a few people to read it, but while I loved both shows, I didn’t know that McMurtry had died (Esquire was wrong, btw; he died in 2021). I didn’t even know that there was a Larry McMurtry.

No, I found Lonesome Dove because I asked Claude to give me a book I wouldn’t be able to put down, and it recommended The Mountain in the Sea, which I’d already read, and Lonesome Dove.

I went back to the source this week to ask Claude to help me understand whether it was responsible for the Doveaissance, and it said kind of but not really, so I asked it to give me a timeline, which it did, but it felt incomplete, so I asked ChatGPT Astra 6, and it delivered a plinko ride through how things spread today.

Basically, the whole thing kicks off with The Ringer. Chris Ryan and Andy Greenwald spent COVID Summer (2020) discussing LD on The Watch, culminating in a 2.5 hour megapod.

Then McMurtry, like so many of his characters, died. He left us on March 25, 2021. That summer, Stephen King went on Stephen Colbert’s “What Ya’ Readin?” and said that his favorite book was probably Lonesome Dove by Larry McMurtry. When it later went viral on BookTok (the part of TikTok where people talk about books), a bunch of people said that King’s rec was the reason they picked it up.

I, mind you, didn’t know about any of this, and I suspect you didn’t either. Media is a fragmented thing.

We pick the trail back up in November 2023, when the From the Front Porch podcast picked it as its 2024 “Conquer a Classic” for another group read-along. One participant described reading it alongside “thousands of other readers,” which is more copies than most books sell total.

In June 2024, popular writer Jia Tolentino put Lonesome Dove in The New Yorker’s “What To Read This Summer” list, after having been recommended the book by a friend in 2023. She wrote:

I gasped and sobbed and clutched the book when it was over. I was so thankful for the reminder that reading can still be like this, that the realm of the unread always holds, somewhere, the exact thing I’m looking for, whatever it is.

The next month, in July, Defector’s Tom Ley wrote, It’s Time For You to Read Lonesome Dove, just gushing about the reading experience.

In February 2025, undoubtedly due to the book’s growing popularity and that of the aforementioned Taylor Sheridan Universe, Teton Ridge Entertainment – “THE LEADER IN WESTERN SPORTS, MEDIA & ENTERTAINMENT” – acquired the adaptation rights.

On March 6, 2025, another Substacker, Bart Schaneman, wrote Larry McMurtry’s Lonesome Dove is the darling of BookTok. It deserves the hype., and wrote:

It seems like every guy with a BookTok or Instagram account about books has either read or is reading Lonesome Dove, and most of them are in agreement that it’s one of the best books that they’ve ever read. Writers I admire like Elisa Gabbert said it was one of her favorites of last year. Kevin Morby has cited it as an influence. Adam Gnade references it. A couple of weeks ago it was announced that someone had bought the rights to all of Larry McMurtry’s novels that include these characters, so expect a streaming series or movies or both.

Later that month, we got the Esquire piece, a cap, it would seem, on the breakout-again success.

I still had not heard about any of this, and none of it is that weird yet. Event-driven word-of-mouth mixed with some genre resurgence and influence from influencers.

In April, the BookTokkers kept coming. This guy, Booksaresick, got 41k likes on a “Books for dudes” video that featured Lonesome Dove.

Maybe this would have happened anyway given that you can’t change time, but in mid-to-late 2025, Simon & Schuster released a 40th-anniversary paperback edition with a Taylor Sheridan foreword, creating a direct bridge between Lonesome Dove and the modern Yellowstone audience. On September 23rd, it released an entirely new 35h 49m audiobook narrated by Will Patton.

In October, S&S said that the book had sold more copies in 2025 - 56,000 - than in any year since the TV miniseries, Return to Lonesome Dove, aired in 1993-94, the year that the Knicks were last in the Finals and OJ Simpson raced the police in his Ford Bronco. All while reading was reportedly getting scalped by the internet!

That brings us to 2026.

Early this year, Robert Duvall, the actor who played the protagonist Gus McCrae on the show, passed away at 95. He had said Gus had been his favorite role, and that Lonesome Dove was “The Godfather of Westerns,” and obituaries foregrounded it. RFD-TV re-aired the entire four-part miniseries in his honor. There was a big spike on Google Trends around this time.

In March, Sonny Bunch wrote Why Lonesome Dove Continues to Resonate on The Bulwark as part of the press tour for another McMurtry biography, Western Star: The Life of a Legend by David Streitfeld. When I tweeted about Lonesome Dove on Monday as I started writing this, my friend Cristin Culver replied recommending the book.

In May, The Art of Manliness kicked off another read-along, with weekly posts and discussion prompts on the book.

Finally, in August, Rachel Monroe and LD-stan Jia Tolentino wrote a whole New Yorker piece titled What To Read After “Lonesome Dove”.

In Texas, where I live, someone always seems to be reading “Lonesome Dove,” Larry McMurtry’s thrilling, ambivalent Western, first published in 1985. The rest of the country is catching on: the novel sold more copies in 2025 than it had in thirty years, and earlier this year it spiked on Google Trends. And I’m not surprised—“Lonesome Dove” is the kind of romantic, immersive story that gets BookTok going, but McMurtry shades his epic with just the right amount of regret. There’s never a bad time to hunker down with Gus, Call, and the other members of the Hat Creek outfit as they drive a herd of cattle from Texas to Montana, but summer is ideal. Nearly the whole novel takes place outdoors, through a series of vast landscapes that are at once brutal and beautiful. I guarantee you’ll cry at the end.

What is interesting to me is that, again, I didn’t see any of this! There was this whole Lonesome Dove movement happening right under my nose, on BookTok, in The New Yorker, on Instagram, on TV, across Substack, in read-along communities, and I missed it all.

But Claude didn’t.

The internet fragmented the media landscape into a billion pieces, like cable TV smashed it into hundreds before it. So much so that despite all of the time I spend on my computer and phone, despite how much I love to read, and despite the fact that I ripped through Yellowstone and Landman and Dutton Ranch and was hungry for more in the category, I didn’t catch a whiff of it.

Then I asked Claude for a book rec, and it told me to read Lonesome Dove. Which is maybe a coincidence! Maybe it just knows me well and thought I’d like it.

But I asked ChatGPT, too, on two separate occasions, and all three times, it recommended… Lonesome Dove. The second time was influenced by the first, but it re-recommended it when I asked for books to read in France. It was the only non-France-themed book on the list.

Two LLMs both recommended the same 850+ page book, a book that starts out pretty slow, when I asked for page-turners. I asked friends who read the book how they came to it, and many of them said Claude or ChatGPT recommended it to them, too. D.A. Wallach pointed out that LLMs keep recommending the book to his book club, which is a patient zero-to-superspreader scenario.

Claude loves Lonesome Dove so much that one guy, jj.books.and.beats, made an Instagram video in which Claude keeps pushing him to read it while he’s trying to work.

There’s a way to read this that’s something like “Ugh the flattening of culture!” Remakes are bad enough, now we’re so out of new ideas that the LLMs are just pushing us to the originals.

That’s not how I read it.

Two of the bigger challenges seem to me to be that 1) people are reading less and 2) things are so fragmented that we’re never all watching the same thing. There’s no TGIF or new Harry Potter book release at midnight.

If the internet fragmented discovery, LLMs might be the right tool to put the pieces back together. I don’t read Jia Tolentino or watch Colbert, but they do. They read and watch everything.

It’s been very fun to be reading something old and great at the same time as so many people on the internet seem to be, and finding those people and wondering how the hell they’re just coming to Lonesome Dove now, too.

The fact that it seems to be a bunch of traditional triggers that might have otherwise fizzled but have been given staying power and reach by LLMs - pretty much the one thing that we’re all using - is pretty cool. LLMs may be a new distribution channel for the back catalog of human culture.

Search is good at finding what you’re looking for when you know what you’re looking for. Social media is good at showing you what people are talking about right this second. LLMs are good at figuring out what you, specifically, might think is great, which is often what many other people think is great. Both search and social media have historically advantaged newness, and it’s interesting to think about what happens when LLMs advantage “best” or “best for me” or “lindyness.”

Our new distribution technology might turn out to be real good for old things.

I guess that’s life on the frontier.

That’s all for today. We’ll be back in your inbox Friday with another Weekly Dose of Optimism.

Thanks for reading,

Packy

Hi friends 👋,

Happy Friday, Happy Labor Day Weekend to my fellow Americans, and welcome to our 209th Weekly Dose of Optimism.

Right on cue, we have a week so full of celebration-worthy labor from the good guys that I don’t even know where to fit it all. We have a couple extra stories above the fold and a whole heaping for not boring world subscribers below.

If this is what acceleration feels like in the summer, hold onto your butts in the fall.

Let’s get to it.

This Week’s Weekly Dose is brought to you by… Reducto

The Only Document Parsing Product You’ll Need

Reducto is the document ingestion layer trusted by AI teams at Scale AI, Harvey, Toast, and Vanta to reliably turn complex files into structured, grounded data for production AI systems at scale. Processing more than a billion pages per month, Reducto has built its reputation on the long tail of difficult documents that break conventional OCR: dense tables, unusual layouts, handwriting, low-quality scans, strikethroughs, and formatting-dependent information.

Its new frontier parsing model, r-1, combines layout detection, reading order, tables, formatting, grounding, and granular citations in a single request. It reduces parsing errors by up to 20%, improves latency at high volumes, and costs 1¢ per page all-in.

Already using another parser? Apply for up to $5,000 in migration credits and hands-on support to benchmark r-1 against your current stack on the documents that matter most.

Apply at reducto.ai/migrate

(Early Bonus) Any Human Ever

h/t Gina Gorlin

I’ve been saying for a while that the best thing the pro-progress people could build is a Counterfactual Machine. Run a simulation in which you remove certain technologies from humanity’s toolkit and let people experience that world in VR.

This isn’t quite that, but it’s pretty darn close. Someone built a site that lets you draw a life at random from any of the over 100 billion people who have ever lived. As they wrote, “most lives ever lived were Asian, poor, and brief,” and indeed, the life that I pulled was an Asian woman who lived in the 11th century and lost four of her six kids before age 5.

Try it for yourself.

In a 2016 speech, President Barack Obama said, “If you had to choose one moment in history in which you could be born, and you didn't know ahead of time who you were going to be — what nationality, what gender, what race, whether you'd be rich or poor, gay or straight, what faith you'd be born into — you wouldn't choose 100 years ago. You wouldn't choose the fifties, or the sixties, or the seventies. You'd choose right now.” He was drawing on John Rawls concept of the Veil of Ignorance (and if you haven’t, stop what you’re doing and read Scott Alexander’s Being John Rawls).

People love to talk about the challenges of modern life, and there certainly are some, but there’s also never been a better time to be alive in human history. The only better times, I suspect, all lie in the future, thanks to the types of stories we’ll cover today.

What a time to be alive.

(1) World Model Week: Runway Solaris & GWM 2 and World Labs Atlas

World Model enjoyers stand up!

In March, General Intuition founder & CEO Pim DeWitte and I co-wrote an essay on World Models. “World models,” we wrote, “these systems that learn from watching the world and the actions taken in it — are a fundamentally new kind of foundation model. They can compute what was previously uncomputable.”

There are a bunch of different kinds of models that are called World Models, and in the essay we wrote that Runway’s were Video Models and World Labs’ were 3D Reconstruction and Generation Models, but both took a step closer to our definition with these releases (Runway a little more than World Labs), but whatever you call them… holy shit are these releases cool.

On Monday, Runway dropped Solaris, which it calls an “Interface World Model.” Instead of an LLM writing code that a browser renders into an app, Solaris generates the app itself, frame by frame, in real time, and when you click or drag something it generates the next frame. There is no code. “The image is the application.” This is a picture is worth 1,000 words situation, so just watch the video. Software is going to get so cool.

Not to be outdone, on Tuesday, Fei-Fei Li and the World Labs team showed off Atlas, an “omni world model” trained from scratch on text, images, video, camera poses, depth maps, and 3D all at once. It generates up to a minute of 1440p video with what they call pixel-perfect camera control, meaning you hand it real camera geometry instead of typing “pan left,” it will invent a coherent hallway between two unrelated photos if you tell it they’re in the same building, and it reconstructs real scenes into 3D from anywhere between one photo and a hundred. Ehhh.. you know what? Just watch this one, too:

And then, yesterday, Runway dropped another model - GWM World 2 - “a world model for real time environment simulation.” Instead of an interface or a scene, this one “generates full, interactive, real-time video simulations in one continuous 720p stream at 24 fps and audio at 48000 Hz, responding to your inputs as you explore.”

What a … world.

(1a) OpenAI Drops Astra 6

The launch video nails it. While I haven’t gotten to play with it yet, the demos from those who have seem unreal. Like this one, from Matt Shumer, who built a simulated world full of Astra-powered agents and watched them coordinate. (He also used Unreal to build a realistic Manhattan world.)

A lot of people are saying AGI is here, and I don’t know exactly what that means, but it does seem as if humans are going to be able to use these new models to do a lot of things for us that we shouldn’t be doing ourselves, and that it will let us be creative in entirely new ways. It’s the most excited I’ve felt to play with a new model.

(2) Fervo Signs Its Largest Geothermal Deal Ever with Google

Jennifer Hiller for WSJ

All of these models - World Models, Astras, whatever comes next - are going to need a lot of power.

On Tuesday, Fervo announced a 396 MW power purchase agreement with Google, the biggest enhanced geothermal PPA any company has ever signed, to power a potential Google data center in Utah with 24/7 carbon-free electricity from its Cape Station project in Beaver County starting in 2028. Google also gets an option to take roughly 600 MW more by June 2030, which would bring the total to a full GW, from just one customer.

We’ve covered Fervo a few times here, most recently in #190 & #193 related to its IPO, but if you’re new, enhanced geothermal uses shale drilling rigs to hit hot rock. You don’t need a hot spring or a geyser, just heat within reach of a drill bit, which is a lot more places.

This isn't the first Fervo-Google deal. Project Red in Nevada has been powering Google's data centers there since 2023, and the two signed a 115 MW deal with NV Energy in 2024 that invented the Clean Transition Tariff (Google pays extra for more expensive clean energy so neighbors don’t have to).

Fervo says this new deal “unlocks new, around-the-clock electricity capacity at no cost to existing ratepayers,” and under a Utah law called SB132, it may be able to sell straight to Google’s load.

Thank God for data centers.

(Speaking of which, this WSJ article, See the New Building Techniques Turbocharging the Data-Center Boom, is short, sweet, and mostly visual, and its subtitle makes the case for DCs: “Hyperscalers’ need for speed is spawning timesaving products and building techniques.”)

(3) Tesla Launches Cybercab in Austin

The future of transport is safer, more enjoyable & gives you more time back

Yesterday in Austin, Texas, Tesla launched the Cybercab, and starting at 5pm today, anyone in the Austin geofence can hail one from the Robotaxi app.

The interior was designed for self-driving: it has two seats, butterfly doors, no steering wheel, no pedals, and no controls of any kind. The future should look like the future. It has nine cameras, a 48 kWh battery pack good for roughly 290 miles, one 219 horsepower motor, weighs 3,113 pounds, comes with Starlink in the roof, and, per Elon, is “basically a super comfortable lounge on wheels with a great TV and epic sound.”

Back in Dose #204, when Zoox got the first US approval for a vehicle without a steering wheel, I wrote that cars would start being designed like moving rooms. That was seven weeks ago. Now we have another one, designed explicitly to be an entertainment room.

Cybercab is starting small of course. Waymo does 500,000 paid rides a week across 14 cities with 4,000+ retrofitted Jaguars and Ojais. There are only 45 Cybercabs in the Texas fleet next to 269 Model Ys (one way they want to scale is by letting people sign up to purchase fleets of them).

But this is the first vehicle designed from a blank sheet to be autonomous and cheap and mass produced, by a company that has mass-produced impossible things before, and real people are riding in them. The Zoox is a $250,000-ish minivan-shaped thing. Waymo pays Zeekr and Jaguar and then bolts on $100K of sensors. Elon says the Cybercab costs under $30,000 to build and about 20 cents a mile to operate at scale, with 10 cities by the end of the year and positive unit economics by mid-2027.

As friend of Not Boring Sam D’Amico tweeted, “Cars are robots now.”

There may be tens of millions of these robocars on the streets one day. I still can’t wait to read about the horrendous things people do in them, as they’re whisked very safely from Point A to Point B.

(4) Humanity’s First Robotically Harvested Head of Iceberg Lettuce

Danny Bernstein and Pascal Labrecque

As fund theses go, Reservoir VC’s might be the most specific: “Reservoir VC invests in rugged AI startups that solve one high-value task in ag and then compound to adjacent industries.” Their answer to “let me know how I can be helpful” is unique, too. GP Danny Bernstein has a 40 acre farm / agricultural robotics playground in Salinas where the robots can get real-world practice.

There, this week, a Canadian startup called Sami Robotics had a robot cut and pick a head of iceberg lettuce, which seems pretty cool but is apparently a bigger deal than it looks. Danny called it "an AgTech moon landing moment."

NZ robotics company Sentinel Robotics added: “I don't think people understand how monumental this kind of thing is. Harvest has always been the hardest part of agricultural automation and leafy greens are the hardest of the hard.” Every head is a different size, you have to cut in exactly the right spot without bruising it, and a good crew of humans with knives is really fast as it is.

If I ever see a Sami iceberg salad on the menu, I’ll say to my friends, “Lettuce try that.”

(5) NASA’s Roman Space Telescope Launches Nine Months Early

NASA (h/t Matt Gialich)

Last Sunday at 7:26am, a Falcon Heavy took off from Pad 39A with NASA’s Nancy Grace Roman Space Telescope on top, the start of its three month journey to Sun-Earth L2, almost a million miles away.

It’s going to hunt for exoplanets.

Roman’s mirror is the same 2.4 meters as Hubble’s (it’s actually a spy satellite mirror the National Reconnaissance Office gave NASA in 2012), but its Wide Field Instrument sees a chunk of sky at least 100 times bigger than Hubble’s infrared view at the same sharpness. Over five years, it’ll survey more than two billion galaxies, measure how dark energy has stretched everything out over time, and do a microlensing census of planets toward the center of the Milky Way.

This is Dose-worthy on its own because we love a good space telescope here. We shared the first images from James Webb way back in Dose #3!

But it’s even more Dose-worthy, because as AstroForge founder Matt Gialich noted, it’s… “9 months ahead of schedule. That is unheard of for a billion-dollar-plus mission.”

Casey Handmer, who’s been hilariously critical of his former employer NASA (please read NASA’s Orion Space Capsule Is Flaming Garbage), QT’d Gialich proposing that we “Promote the people responsible and give them another worthy project with performance-linked bonuses and lots of authority and responsibility!”

If Administrator Isaacson’s NASA keeps this up, we may get to Mars after all.

(6) Physical Superintelligence Raises $58M to for Physics Takeoff

Matthew Pines

For most of human history, genius was the rarest thing on Earth.

The last time I heard from Matthew Pines, it was when he was the guest on a 4+ hour American Alchemy episode a year ago titled UFO Physics & Disclosure Under Trump. Now, he and his team at PSI (which stands for Physical Superintelligence, but is also a nod to psi abilities) have $58 million to bring our physics up to the 21st century.

Pines is doing it with Dr. Alex Wissner-Gross, whose Eon Systems fruit fly brain we covered in Dose #184, and who graduated top of his class at MIT in 2003 as the last student to complete a triple major under the institution’s historical policy, and Alex Klokus, who created the Gravity Blanket and co-founded the UAP / psionic research group, Skywatcher.

Which is to say, we have a bunch of wide open minded geniuses with tens of millions of dollars to save physics from whatever path it went down decades ago. Hell yeah.

Yesterday, Starcloud CEO Philip Johnston announced a non-profit called The Fermi Explorer, which plans to launch “a minimum viable spacecraft to Alpha Centauri, the nearest star system, leaving in 3 years and arriving in around 80,000 years. We hope it will be both the first spacecraft to leave Earth for another star, and the last to arrive.” AstroForge’s Matt Gialich, who you might remember from the story right above, is an advisor.

I bring this up because the mission was designed by… PSI.

We’d spent three months and one billion tokens on it [the mission],” Johnston said. “In a week, they [PSI] spent 10 billion tokens on it, and it came up with this ingenious perihelion-pump maneuver that we then validated with the NASA guys.

And speaking of AI models doing crazy physics, our friends at Arena Physica, who you might remember from Electromagnetism Secretly Runs the World, which I co-wrote with its founder Pratap Ranade, released its new model, Heaviside-1 this week. Heaviside-1 is its second generation foundation model for electromagnetism, and can run sims like this in milliseconds instead of today’s simulators’ hours.

You can play with it yourself in the Fields Simulator here. Look at me. Look at me. I’m the [physicist] now.

Read more

Hi friends 👋,

Happy Friday and welcome back to our 208th Weekly Dose of Optimism.

While there was a ton of good news this week, the videos of the floods in Nepal and Tibet are some of the most heartbreaking I’ve ever seen. We’re going to cover the good stuff, but if you’d like to donate to the relief efforts, I’m giving to DirectRelief’s Nepal Flood Relief 2026, where 100% of donations support Nepal response activities.

Now, August rules: you should be outside as much as possible while you can, so we will just quickly tell you to check out “Matthew’s” Field Notes from China, and then…

Let’s get to it.

Today’s Not Boring is brought to you by… Create

Dan and Sienna rolled out what looks like the most delicious creatine ever made, the Lemonade Collection. To celebrate, they agreed to give not boring readers 50% off through the weekend. Go get your Lemonade Creatine + Electrolytes and code notboring50 at checkout.

(1) FDA Approves Targeted Therapy for Metastatic Pancreatic Cancer

Xavier Martinez for WSJ

Pancreatic cancer has long been one of the deadliest and hardest to treat forms of the disease thanks to a mutation that sticks the RAS protein on and causes runaway cell growth.

In terrible news for cancer and great news for people, the FDA has approved a drug, Revolution Medicines’ Rasonque, that targets the mutation and helped subjects in a late-stage study live for more than 13 months instead of around 6.7 months for those who receive chemotherapy alone.

The drug does have side effects, and 13 months isn’t forever, both issues personified by “Former Nebraska Sen. Ben Sasse, who went public in the spring with what he described as a significant reduction in his tumors while in a clinical trial of the drug, appeared on television with his face covered in red blotches.” Sasse lived longer than expected, but he still expects to pass away, as discussed in his incredible conversation with Leopold’s wedding officiant Ross Douthat, titled “How Ben Sasse is Living Now That He is Dying.”

Still, this is a massive win, one that drew a standing ovation at a cancer conference last spring when Revolution announced the results.

As one patient, former Disney exec Jay Williams, told WSJ, “What it does is buy you time, and that’s one thing you don’t always have with pancreatic cancer.”

And speaking of time, good on the FDA for approving this so quickly. Per the Agency, “The FDA granted Rasonque Breakthrough Therapy and Orphan Drug designations. Rasonque received Priority Review for this indication. The application was also reviewed under the Commissioner’s National Priority Voucher pilot program, which is intended to help accelerate the review of therapies that address national public health priorities.” As a result, it was approved about 6.5 months faster than the standard PDUFA timeline, which means means it went through in roughly half the usual already-accelerated window for a priority oncology drug.

Patients, oncologists, all of us, and regulators are saying the same thing: get fucked, cancer.

(2) Airbound Announces $37M Raise and Drops Manifesto

Naman Pushp for Airbound

Last year, I invested in a company called Airbound founded by a young guy from India who taught himself physics and engineering on YouTube during COVID and used that knowledge to build better drones.

After we got introduced but before we met, I didn’t really think I’d invest - drones are a fairly crowded space and Zipline is awesome. Then I met Naman, and decided to invest immediately.

Over the past couple of months, I’ve had the chance to work with him on his Manifesto for Airbound, in which he did a great job laying out the future where all movement is in the air that got me so excited in our first conversation.

He published it this week, along with the news that he raised $37M from Greenoaks, Lachy Groom, and others.

Read it - and stay tuned for more.

(3) Radiant and Antares Selected for US Army Janus Program

U.S. Army Communication and Outreach Office

We love to see friends get contracts from the US Army to make a lot of nuclear reactors and hopefully save a lot of lives in the process, so we were very happy to see the Army choose Radiant Nuclear, who we wrote about in 2024, and Antares Nuclear, whose co-founder Julia DeWahl was my Age of Miracles co-host, sign deals as two of five companies selected for the $2.2B Janus program.

Radiant is getting $750M to make 15 Kaleidos microreactors, and while Antares didn’t reveal specifics on this contract, they tweeted that they “now have a contract value on the order of $1B that is growing rapidly.”

While the business side is awesome, the most important piece is what this might mean for warfighters. As Doug said in the deep dive, “Over 55% of US soldier casualties in the two Iraqi conflicts are from ambushed convoys. And convoys move ammo, water, and fuel. So if you could put a reactor in one of these locations, you completely remove all those fuel shipments, you save tens of thousands of lives.”

As a bonus, our portfolio company Standard Nuclear has signed a binding agreement to supply Antares with TRISO fuel through 2035 for microreactors built for defense and space missions, which means it’s now supplying both Janus-winning startups.

(3a) Actinide Becomes First Startup to Enrich Uranium to HALEU

And speaking of nuclear fuel, Actinide became the first startup to enrich uranium and produce HALEU (High-Assay Low-Enriched Uranium), the fuel that many advanced reactors use (it goes inside the TRISO pellets).

They did it using their particle accelerator, Endurance, designed and built in Texas, which had been producing “enriched Yb-176, a precursor for cancer therapies worth roughly $30,000 per gram,” just last month.

HALEU will be one of many isotopes the company makes in its attempt to “address humanity's greatest unsolved problem: control of the atom.”

(4) Anthropic Kicks Off Preview for Model Hardware Standard

MHS enables AI agents to operate multiple lab and manufacturing instruments, such as microscopes, liquid handlers, and robotic arms, in parallel, and perform intricate tasks ranging from routine drug discovery experiments to laser calibration on a quantum computer.

Ruxandra Teslo has some good quick thoughts here, which you should read, but which are basically: this is aiming to automate boring stuff, like laser alignment, so humans can do science, which is core to the human spirit.

David Hansen has some, too, on what humans at US manufacturing companies can already do with Claude and Codex.

Read more here.

(5) Hugging Face Quacks into Robotics with Microduck (and sells to Nvidia for $12.9B)

Two days ago, rumors swirled that Nvidia would be buying Hugging Face, the open source model repository, for $12.9 billion. Open source makes sense for Nvidia - commoditize your complements!

So when Hugging Face co-founder Thomas Wolf tweeted on Thursday morning, he tweeted what was on everyone’s mind: “We have a huge news to share today!”

The news, though, was that Pollen Robotics, which the company acquired in 2025, was “unveiling the first truly accessible RL robot - welcome Microduck.”

The thing is cute, costs $399, and you can train it yourself with reinforcement learning. Before you do, it can do a bunch of stuff out of the box, like “walk, sit, crouch, roller-skate, pick up objects with its articulated beak, and recover on its own.”

Asked to comment on Hugging Face’s big week, Billy Madison said, “That’s quacktastic!”

EXTRA DOSES: Science Breakthroughs, Stony Brook Quantum, Figma Podcast, Anthropic Model Hardware Standard, Dyna ROI, Alex Colville

Read more

Happy Tuesday!

A few weeks ago, a longtime not boring reader and one of my smartest friends, someone I go to with questions on everything from energy to robotics, told me that he was going to China. Over the next few months, he planned to meet with energy, AI, and hardware companies and develop his China Fingerspitzengefühl.

Would I be interested in sharing what he learns in real-time with not boring’s readers? Incredibly easy yes.

If you are interested in the types of companies and technologies we write about in not boring, you probably need to understand China better. I need to understand China better. It currently dominates the Electric Stack, for one thing, and the next few decades will see Western companies compete with Chinese ones to bring the best products to the world market, to say nothing of geopolitics.

There are a number of great China-focused blogs for western audiences out there - I like Sinocism, China Talk, Tech Buzz China, and Hello China Tech, not to mention Dan Wang’s great book, Breakneck, but there are many more - but I think there’s something particularly interesting about riding along with someone sharp, curious, and motivated as they land in China with a bunch of imported questions, try to answer them, realize they were probably the wrong ones, ask new ones, and seek those answers. The process is often more illuminating than the output.

We decided to share his field notes as a series for paid not boring world subscribers for a few reasons:

1. This is exactly the kind of thing I want to publish for paid subs in not boring world: an operator sharing raw information with a little intelligent interpretation through his own earned lens.

2. So that he can be more open with his discoveries and analyses, and move freely.

For that same reason, he’s writing under a pseudonym: “Matthew.”

Over the past couple of weeks, as we’ve chatted from across the globe, I’ve been amazed by how fast “Matthew” went from American outsider to someone in the right WhatsApp chats to conversations with increasingly heavy hitters. You can see that evolution already in today’s piece - from wide-eyed tourist to factory tours. In just the few days since the ones we cover here, he’s gone deeper still.

Today, for our first installment, we’re including some of his initial impressions of Hong Kong and the mainland above the paywall as an introduction, but most of the good stuff today and going forward for the next couple of months will be for not boring world subscribers. Subscribers can also ask any questions on China you have — specifically around hardware, energy, AI, and manufacturing, but really, just throw anything at us. Send them over to me and he’ll dig, as he’s generally quite curious about China beyond his immediate interests.

Understanding China as more than a cartoonish adversary is critical for anyone who wishes to compete with its companies, and for those who simply want to better understand the world. I hope that Matthew’s adventure will give you a specific and increasingly nuanced view of how they do what they do, and why, and what they might do next, and what it all means for American technology companies, through the eyes of an American engineer.

Let’s get to it.

An American Engineer in China

Field Notes #1: The First 11 Days

Introduction

This piece will cover my first 11 days in China. The purpose of my extended trip is to learn about and meet with Chinese energy equipment companies and factories, but I’m also curious about China’s culture and financial ecosystem.

In the first section, I was an outsider getting a feel for China and getting invited to the right WeChat group chats. As I did, I began getting invited to factories and having discussions with more and more “mainlanders” who helped fill my knowledge gaps about the country. Throughout the piece, I’ll share a mix of my experiences and key takeaways.

Three Biggest Takeaways:

1. The China of 2036 will be vastly different than the China of 2026. Companies that used to manufacture luxury handbags are now becoming the brands themselves. Investors who used to copy American VCs are now running their own state-backed deeptech playbooks. Battery cell companies are becoming grid, datacenter, and asset-owning companies. If you aren’t heavily updating your mental model of China after this piece, then you will get blindsided.

2. China is heterogeneous. China has so many different dialects, dresses, jobs, climates, educational attainments, and personalities that I feel as though I could spend centuries in China learning new things. Beware of anyone who makes reductive stereotypes about the country, positive or negative.

3. China is big. Yes, this is “obvious,” but you don’t viscerally feel it until you casually see hundreds of skyscrapers in a row and single companies outproducing the industries of entire countries. Everything else just seems so…puny.

Section 1: Initial Thoughts

When I first arrived in China (or rather, Hong Kong, before heading into mainland Shenzhenbei the second day), I was incredibly shocked by the scale. The airport seems innocuous enough, although larger than most American international airports, but the moment you take a taxi into the city, everything becomes mind-bogglingly dense. In New York City, America’s densest city, there are around 900 buildings above 100 meters. In Hong Kong, there are 4000, filled to the brim with people in and around them. Frankly, it surprises me when people say that China is lying about the size of its population. You can feel the sheer size of everything here. Even the “suburbs” here can be millions of people large, with towering high-rises

stretching on and on and on.

Being guided by my friend in Hong Kong, I could find anything anywhere. Hong Kong is a financial and trading hub (import/export businesses have created many wealthy individuals), and in 10 minutes I saw three separate stores selling professional-grade cellos. But things really got nuts when I walked into the malls and high-rise buildings. You could have an IKEA on floor 5, an ice skating rink on floor 10, and then a wave pool on floor 11. I truly never quite knew what to expect every time we entered a new building. For example, Toys “R” Us was a staple of my American childhood, and it went bankrupt around a decade ago. However, the Chinese arm of Toys “R” Us was always financially healthy, so it still exists in China. Similarly, many famous US brands in China are owned by the Chinese version of the company, despite sharing superficial similarities. The 5090 GPU I found in an electronics market here, though, was very much the same Nvidia as the US one.

Hong Kong is also interesting because many mainlanders refer to it more as an “antique of the British Empire.” Hong Kong used to be China’s dominant financial hub, partly because of its advantageous tax structure and lax rules compared to the mainland (changes in both during COVID led foreigners to flee to Singapore). However, Shanghai, Beijing, and even Shenzhen have grown rapidly since the 90s, giving Hong Kong competitors.

That said, Hong Kong has an abundance of great universities, many of them highly technical and sending out thousands of students. This is one of the city’s more lasting draws for mainlanders, along with Hong Kong’s nature. Near one of these universities, while taking the subway (which arrived every 2 minutes, had perfect AC, and moved millions of people a day), I first saw Hong Kong’s low-rises and was surprised by the size of its beaches and archipelagos. I took a speedboat to tour and visit some of the beaches, and was even more shocked by the number of wakeboarders and superyachts corralling around. It felt more like Monaco than the urban metropolis I expected.

Heading back, although the locals will just drive up, you can hike up Hong Kong’s mountains to get a breathtaking view of the city. Hong Kong’s densest region is roughly 3x the size of Manhattan in terms of population, and from the tallest mountains you can even see Shenzhen. One of the buildings, around 500 meters tall, was covered in OLEDs gleaming pink, a cyberpunk obelisk staring at you from across the border.

The next day I headed into Shenzhen, the famous “tech-bro” city, and I should’ve allowed a lot more time for immigration. It was my first time getting a lot of particular paperwork and processes done, like giving them all of my fingerprints while being watched by quite literally over a hundred cameras. So, by the time I got on the high-speed rail into the city (which, likely because it was in an urban area, only went 160 km/h), I was in Shenzhenbei, a train stop about 10km north of where I wanted to go. I should’ve gone to the Futianbei train stop because a lot of the urban scene, my hotel, and seemingly everything fun is more on the south side of the city, but it was still useful. Shenzhen, I’ve been told, is much more friendly to foreigners in terms of its infrastructure than other mainland cities (some of which have just straight-up zero English anywhere), but it was still quite a difference. Uber no longer really worked; you needed to have WeChat Pay or DiDi. Most signs had little to no English; the rate of foreigners dropped dramatically, and cigarettes and cameras watching you took their place.

Most people who drive are in EVs; I saw very few gas-powered cars.

The McDonald’s that I scrambled to grab food at had a slightly different menu, but honestly wasn’t too different from an American one (besides some folks eating with plastic gloves, although most were just eating with their hands/utensils), before I headed into the core part of the city. I dropped my stuff off at my hotel, which, despite being super nice/multi-floor, was around $50 USD, where I met a friend who had moved to Shenzhen a year ago.

There, he introduced me to a ton of folks on WeChat (WeChat is where everyone does business; if some low-tier Alibaba factory responds to your email, then you know they’re going to be bad), and gave me a bunch of advice summarized in this italicized paragraph.

“Shenzhen is massive, it’s new (Deng Xiaoping’s SEZ 40 years ago), so nobody is “from there” the way they are from Beijing/Shanghai, culture is fairly welcoming, 100k people every day make something new so it changes all the time, even if you were here for a decade or two decades you would know so so so so so so little about the whole of it all, tier-1 manufacturing can be found in many parts of the world not just china but density of suppliers and supply chain make Shenzhen unmatched. APEC 2026, normally held in Beijing/Shanghai, is being held here for the first time because Xi likes the city and the government wants to show it off. It helps to be specific about who you want to meet because hardware specialists exist in every subfield you can imagine, and many universities offer 4-year/master’s degrees in manufacturing that are actually really good. No foreigner will ever get government subsidies (you may be well liked, and even a permanent resident in Hong Kong, but you will forever be a foreigner. Never forget that.) and that’s why you see so few here; they just get brutally outcompeted, but you can and should work with folks on suppliers/supply chain. Be careful about IP; people are super competitive here. If you want to cut costs and figure out expansions, don’t just think about input/output; you’ll get scammed. You need full process knowledge to figure out the limiting factors and how cheap things really are. He’s seen suppliers charge companies $200k for a machine that should have cost $20k. The food delivery guys can make a living here because the cost of living is so low, but they were often just scrolling e-girls on their phones while we were driving. However, if Meituan tried replacing them all with drone delivery, there would be so many riots in the streets.”

I then (after being guided by a friend who said “just come here, trust me”) entered an unexpectedly massive anime/cosplayer facility underground, with every possible character/combination you could imagine. There were blatant ripoffs of many characters (I know a Minion when I see one), but it was interesting to see the variety of genres. While I didn’t snap pictures of the cosplayers for fear of being rude, many elaborate anime cosplays in thick makeup, wigs, and costumes walked around Shenzhen.

Afterward, I walked upstairs and spent a few hours around Huaqiangbei, Shenzhen’s famed electronics market. To be honest, it moderately blew my mind, and I say moderately only because the initial Hong Kong skyscrapers had completely shattered my worldview the prior day.

If you are a hardware nerd, you would’ve loved this place. Massive billboards advertised MOSFETs and spare motorcycle parts. I walked into a McDonald’s looking for a bathroom and ended up in a random 5-story electronics market. I saw every type of phone (Huawei stores included), drone, robot form factor- you name it; here, there was technicality (and the smell of solder wafting around) in the air. Plenty of small kids came up to me yelling “nihao nihao”, which was cute, although the 7-11 shockingly not taking international payment was not cute, as were the waves and waves of blisteringly fast scooters that threaten to run you over if you don’t dodge and weave around them.

I went to the original Bambu Labs store in Shenzhen, where they proudly displayed that their 3D printers could make guns. That was probably one of the bigger culture shocks of the day, given the paranoia in the United States about ghost guns.

An interesting takeaway from the first few days is that we are not at the center of their universe.

America is certainly important and interesting to China, but China has the Greater Bay Area (which has more skyscrapers than all of North America and Europe combined, compared to our puny American Bay Area) and will come to dominate over the next 10, 20, 65 years. They’re going to spend over a trillion dollars on new infrastructure through the next decade, which goes so far here, and try to return to where they once were: the capital C greatest Civilization in the world.

Section 2: No More Foreigners

By the fourth day in China, I had begun to internalize a very important lesson: foreigners won’t be helpful here. This might be an exaggeration, but I see it as directionally correct. Even if you speak the language and have been in the country for decades, there is naturally a wariness that exists if you are not a Han Chinese person who was born in mainland China. That’s a big reason there are essentially no foreigner-run ODMs (Original Design Manufacturers) in Shenzhen.

Of course, even though China is a country pretty explicitly for Chinese people, it is not a homogeneous place. In a couple of hours, you can go from the center of a Tier-1 gleaming coastal city to see literal rice paddy farmers working in the fields. Some companies are almost entirely subsidized, while others take zero government money. There are people dressed in streetwear covered in tattoos, piercings, and dyed haircuts, while others are so straight-laced they would never even breathe incorrectly. Over 80% of people in China do not hold a college degree. Many buildings were in shoddy condition, yet they were surrounded by marvelous, glowing OLED obelisks.

Shenzhen’s scenery is a visceral example of China’s heterogeneity. Show of hands: how many of you knew that Shenzhen, world capital of the hardware nerds, had palm trees? I certainly didn’t, although I suppose that it makes sense that there would be palm trees in Shenzhen and beach archipelagos near Hong Kong. It’s right near the Tropic of Cancer line of latitude, similar to where Cabo is. Of course, Shenzhen is a lot more mercantile and only mildly less alcoholic (debatable) than Cabo is.

Similarly, how many people talk about Shenzhen’s creative district, which would not be out of place in a standard American or European coastal city? Just like in these countries, they also have nice little patios where you can order pastries and public art, which I will describe simply as “somewhat odd.”

On my fourth day, I visited a UPS and PCBA factory in Shenzhen.

Read more

Hey friends 👋,

Happy Friday and welcome to our 207th Weekly Dose of Optimism.

It’s Puja’s birthday — happy birthday Puj! — and I’m taking the day off, so today is both late and light on commentary. Thankfully, the good guys didn’t take the day off - lots to report.

Let’s get to it.

In lieu of a sponsor, some of my favorites…

Matic Robots - a home cleaning robot you can talk to (right link this time!).

Quince - high quality … everything at affordable prices. half my wardrobe.

Notion - not boring’s everything hub.

Ramp - Time is money. Save both.

Create - you should be taking creatine (and supporting Dan) - 30% off brother discount.

(1) Merck-Moderna mRNA Cancer Vaccine Successful in Phase III

Over here at the Weekly Dose, we are half-Irish, we hate cancer, we love mRNA, so imagine our excitement when we saw the news that Merck and Moderna’s mRNA cancer vaccine prevented melanoma (skin cancer) from recurring.

Given with Keytruda after surgery in high‑risk melanoma patients, the vaccine significantly reduced the risk that cancer would return or spread compared with Keytruda alone. It’s the first successful late‑stage trial for a personalized mRNA cancer therapy and could lead to approval and launch around 2027.

Melanoma is just one kind of cancer, we’d rather no one get it in the first place, and the vaccine is still expensive and complicated, but cmon. We have promising personalized cancer vaccines.

Say it with us… get fucked, Cancer.

(2) Voluntary attention regulates acute immune responses in humans

Nofar Mizrachi, Menachem Rottem & Liron Rozenkrantz in Nature Human Behavior

In Means and Meaning, I argued that attention is the thing we are, and where we deliberately direct it is how we create meaning for ourselves. Apparently, it’s also how we regulate immune function. From the study:

These results establish attention as a cognitive mechanism capable of regulating immune function in vivo. Situated within predictive and allostatic models of brain–body control, our results reframe perception as biologically consequential rather than merely experiential.

Attention really is all you need.

(3) Generalist… Generalizes?

The holy grail in robotics are robots who you can train on a bunch of specific data, then ask them to do something that’s not in the training data and they just do it.

According to robotics company Generalist, their GEN-1.5 is a one-shot learner: show it a few seconds of data, and it figures it out. It can learn from simulation, from watching humans, and even by chaining together smaller tasks. The thread has a number of good examples.

Robotics is really hard, and there are certainly a lot of things to figure out between this demonstration and billions of useful robots, but this is genuinely impressive and it seems like robots are going to happen sooner than later.

(4) Big Week for Routers

Stripe and Ramp and Merge

Routers help companies and developers … route their work to the right AI models for the job. Sometimes, you need frontier capabilities and are willing to pay for them. Others, cheap and fast is perfect. Ben Thompson has a characteristically good breakdown in Stratechery.

In early June, Markie Wagner told us that companies were going to start caring about Return on Tokens (ROT), and she was right. Routers are one way to do that, and they have been ripping.

On Wednesday, leading router provider OpenRouter confirmed it’s being acquired by Stripe. The NYT reports they paid $7.5 billion. OpenRouter was founded by the founder of NFT marketplace OpenSea, Alex Atallah.

In the very same week, two of our not boring capital portfolio companies announced routers of their own. On Tuesday, Merge launched Merge for Workforce, which lets companies spend 75x fewer tokens for better results. The next day, Ramp announced that it purchased the domain router.com and to house its own Router product, intended, of course, to save companies time and money, this time on intelligence.

Figuring out how to get the right intelligence at the lowest price automatically is going to be a huge market if the pace of AI adoption keeps up (and giving people better value may accelerate adoption further), so there will be a bunch of winners. What will be fascinating from a strategy nerd perspective is to see which of these companies’, and others’, set of assets lend themselves best to, and benefit the most from, routing.

(5) Big Week for Drones

Zipline x Uber and Amazon

On Monday, Zipline and Uber announced a partnership through which Uber will invest in the drone company and together make 1 million deliveries a day by 2019.

On Wednesday, TechCrunch reported that Amazon is expanding its Prime Air delivery service to reach 500 cities by the end of the year.

These announcements come on the heels of DoorDash launching its DoorDash Air in-house drone program.

And yesterday, a Texas woman shared a video of Amazon Prime Air dropping her package in the pool.

I am incredibly bullish on flight and think these bits of news are just the beginning.

EXTRA DOSES: Science Breakthroughs, Dog Cancer, TRISO, Dan & José

Read more

Hi friends 👋,

Happy Friday and welcome back to our 206th Weekly Dose of Optimism!

This is a few minutes late because I thought it was a pretty slow week and had it all set and then new stuff just kept coming out, including Anduril announcing that they’ve launched a satellite, and I had to re-jigger things, which goes to show that the good guys and gals won’t even take one week off.

There are lots of goodies in main and in the Extra Doses, but if you have some extra time this weekend, get yourself a nice big iced coffee and read the third chapter in the Base Power Company story.

For now…

Let’s get to it.

Today’s Weekly Dose is brought to you by… Matic

It’s not actually brought to you by Matic. They’re not paying me and they didn’t ask me to do this, although they did give me a free Matic when it first came out. They don’t even know I’m doing it.

It’s just… we didn’t have a sponsor this week, and as I was playing with the Matic’s new voice control last night, I realized that it’s one of the most delightful things I own and I wanted to share it with you.

Anyway, you should do yourself a favor and buy a Matic for your home.

(1) Avidrone wins DARPA Heavy Lift Challenge

This week, DARPA held its Heavy Lift Challenge, a competition to see whether a drone could complete a course while lifting 4 times its own weight. Hunter Weiss has been going viral all week posting the videos.

A Canadian company called Avidrone won with its Katana uncrewed aircraft. The 29.3 lb drone completed the course in 19:17, while carrying a payload of 112.4 lbs… just shy of the 4x mark at 3.84:1, but better than anyone else. They tried to hit 4:1 one final time and crashed hard. In any case, they won 1.25x.

There are military reasons for wanting heavy-lift drones or DARPA wouldn’t have done this, but the reason it’s leading off the Dose is that getting drones to be able to carry more weight, further, for less money will reshape the world like nothing since the car.

Katana isn’t what that future will look like, but it’s a fun little preview. Stay tuned.

(2) Josh Kushner, Bob Iger to Buy Lakers for $12.5 Billion

When Thrive Capital announced Thrive Eternal in April, it prompted me to write an essay on Scarce Assets. I wrote:

Last Friday, Josh Kushner announced Thrive Eternal, the firm’s permanent capital holding company that will concentrate in a small handful of “Iconic franchises and cultural institutions rooted in tradition, identity, and shared experience,” starting with the San Francisco Giants.

And look, the Giants are fine. Not having a great season, but a solid baseball franchise. A nice starter purchase for Eternal.

The Lakers, though? Iconic. Rare. The scarcest of the scarce.

Imagine agreeing to terms courtside at the Lakers, one of the places to be seen. A founder pic courtside with Josh could send a company’s valuation up a billy.

Per Sportico, the $12.5B transaction set a new record price for an NBA team, surpassing… Mark Walter’s purchase of the Lakers last year.

That chart should look comfortably familiar to a venture capitalist, and it shows no signs of slowing. While there’s a new $10 billion big lab spinout neolab every week (see: Jeff Dean’s Disco Loop is in talks to raise $1B at $10B), they aren’t making any new Los Angeles Lakers. As Matthew McConaughey might say in a Dazed and Confused remake, “You know what I love about these Scarce Assets man? People keep getting richer, but they stay the same amount.” Or whatever.

In related news, Thrive Holdings, the firm’s fund that owns and operates accounting and IT services businesses, raised a fresh $2 billion yesterday, too. Chapeau.

(3) Recast Comes Out of Stealth to Give Humans Agency Over Weather

We now live in a world in which there are two American startups run by young geniuses attempting to help humans direct the weather, which is awesome. Doing so might mitigate disasters, like droughts and floods, and allow us to take a more proactive approach to things like energy and terraforming.

Recast Systems, “a weather company,” has been busy while it’s been quiet:

In the past 6 months we’ve produced a cheaper, more efficient seeding material, trained a hyper-resolution AI weather model, and operated 100+ cloud seeding flights for the governments of Texas and New Mexico.

We’ve covered Rainmaker a bunch of times in the Dose, and in The Great Differentiation, in which I wrote that they’d earned their differentiation by doing something incredibly cool in the physical world (flying drones into clouds).

What I think is great about Recast’s launch is that they’re carving their own distinct, differentiated brand in the space. It’s more science-forward, more handmade. You can see it in founder Olivia Li’s behind the scenes here, and in the images they chose for the unstealthing thread. This is the first one:

This is the next:

The sky is a really big place, and beyond cloud seeding, “the weather” is even bigger. There will be multiple winners, and the addition of a second credible startup, with a totally different vibe, adds credibility to the whole space. I think both Rainmaker and Recast will do great things, and I hope they do them ASAP.

To summarize: a few years ago, we had no weather startups. Now, we have two. I guess… when it rains, it pours.

(4) Fuse’s Faeton-X Hits 1.27×10¹² Neutron Yield

Compared to every other fusion company, the thing I love about our portfolio company Fuse Energy (aside from the founder, JC Btaiche) is that it generates revenue by selling radiation on the way to generating power from fusion.

When I wrote about the company in May 2024, I described a couple of paths. The TITAN path - from TITAN to Z STAR to APEIRON, is what will ultimately deliver fusion power to the grid.

The FAETON path, while not the path to fusion directly, produces valuable neutron shots: “In the near-term, Fuse actually expects to make more money from FAETON II shots than it does from TITAN II shots, because the value of FAETON II’s neutron-producing shots are twice as valuable as TITAN II’s.” Defense contractors and satellite makers pay tens of thousands of dollars a shot to blast their electronics with neutrons to find out whether they could survive a nuclear detonation or a decade of radiation in orbit.

FAETON also uses some of the same components, namely the Brick, “the high-voltage capacitor and gas switch that is the fundamental building block of its pulsed power generators”

All of which is a long way of introducing the importance of this breakthrough. This week, Fuse published results from FAETON-X, showing that it generated 1.27×10¹² fusion neutrons in one shot, making them the first fusion company to publicly document a yield in the 10¹² range.

Importantly, though, they did it with a higher current efficiency than the labs: 4.5 MA/MJ, highest current efficiency ever for a MJ-class plasma focus, against 2.35–3.25 for all published peers including government machines. It took Livermore’s MJOLNIR 1.3 MJ to produce 1.2×10¹² neutrons, where it only took Fuse 1 MJ to produce its 1.27×10¹².

That is important, because yield in pulsed-power z-pinch devices scales as the fourth power of drive current. Current efficiency is the compounding variable of which yield is the derivative. 40% more current per MJ gets you 4x more fusion, not just 40% more.

This is maybe the most important curve for Fuse: neutrons per megajoule of capacitor bank. On Faeton, it means delivering those valuable neutron shots more efficiently. On the TITAN path, it means that everything gets smaller and cheaper for the same current, which could mean getting to fusion power, and making it cheaper, sooner.

This is the Fuse thesis, too. Instead of straight-shotting fusion like competitors, get out in the field and start selling. Iterate. Improve the brick on FAETON, and bring it to TITAN. Move fast. They went from design to first shot on FAETON-X in nine months, and from first shot to record in another nine. Keep interacting with the messiness of the real world, keep iterating, and maybe one day the iterative approach gets you to the end goal faster.

If not, there’s a lot of very useful stuff you can do with all that radiation at Q<1.

(5) Matic Now Lets You Talk to Your Robot

I love our Matic. It reliably vacuums and mops. It’s proof of the saying that something is called a robot until it works. The Matic robot works, so we just call it the Matic.

Yesterday, the Matic reintroduced the magic. You can (and last night, I did) talk to it, and tell it where to clean. From the company’s tweet:

1. Say "Hey Matic" - it locates your voice, turns, and looks at you
2. Say "Hey Matic, follow me" and start walking. Matic will follow behind
3. Say "Hey Matic, go clean the living room". Since it knows your house map, it navigates and just does it

People think the path to robots doing useful things in our home will look like a person. Matic’s bet has been that it’ll just be really useful at one thing, and then get useful at more things. So far, it seems like the right bet.

EXTRA DOSES: Science Breakthroughs, Anduril’s Satellite, Ilia Delio on Intelligence, Rubin on Greatness, AI’s First Really Good Thing

Read more

Subscribe now

Hi friends 👋 ,

There are only two companies on which I’ve written multiple Deep Dives.

The first is Ramp, which I first wrote about when it was valued at $300 million, then again when it was valued at $1 (and $1.6) billion, and then multiple times since. Today, it is valued at $44 billion.

The second is Base Power Company, which I first wrote about when it was valued at $300 million (Chapter 1), then again when it was valued at $1 billion (Chapter 2). Today, the company is announcing that it has raised $1 billion at a $12 billion pre-money valuation, bringing its valuation to $13 billion.

In the intro to the last Ramp piece, Ramp at $1 Billion, I wrote that “One of my goals for this newsletter (and Not Boring Capital) is to find a handful of generational companies as early as possible, write their story in chapters over time, and then publish the whole thing as a book when they IPO.”

That is certainly my aim with Base Power Company, which I’ve both invested in and plan to keep writing about through and beyond IPO.

This is the third chapter in the Book of Base. There will be many more to come.

If you want to help write the Base Story, and I think you should, because I think it’s going to be one of the most important of our generation, my advice is to get yourself on the next flight down to Austin and join them.

Let’s get to it.

One Friday night earlier this summer, Base Power Company CEO Zach Dell and Ribbit Capital founder Micky Malka agreed to terms on a $1 billion Series D at a $13 billion post-money valuation. By Saturday, the round, co-led by Ribbit, Valor Equity Partners, Addition, and JPMorganChase’s Strategic Investment Group, was fully subscribed. By Sunday, it was multiples oversubscribed.

$13 billion is, to be sure, a hefty price tag for a toddler, no matter how precocious. Base is less than three years old, and it is now the second most valuable energy startup in the world, behind Helion, and the most valuable that is selling a product today.

It is a bet, I think, that Base can become the largest energy company in the world.

That seems crazy, I know, but it’s been my working thesis for a while - that Base has the opportunity to do to energy what SpaceX did to space, what Anduril is doing to defense - namely, win a large, old, incumbent-dominated market.

It’s also the company’s ambition.

Last November, before Base had decided to expand to Chicago, before it had produced its first Base Core, Zach came to New York to do a fireside at Thrive Capital’s office.

After the conversation, we were chatting with a small group and someone asked whether there was still upside in going to work at Base, to which Zach replied, “Base can become a $400 billion company.”

It was offhand and automatic, but what struck me was that Base had just closed a $1 billion round the month before that valued the company at $4 billion. As the valuation grew, the ambition did too, at exactly a 100:1 ratio. I suspect no matter how highly valued Base becomes, the goal will always be 100x higher.

When we caught up last week, I asked Zach about that conversation at Thrive, and whether now that investors had valued the company at $13 billion, in his mind, Base was now a $1.3 trillion company in the limit.

He didn’t remember saying $400 billion at Thrive, and said that he doesn’t necessarily think about it that way… before laying out exactly how it could happen:

I think about it more in terms of the scope, the size, the total opportunity, and how big a part of the energy market we can be.

A trillion dollar company is pretty big today. I don’t think it’ll be that big in 10 years.

So I mean if you believe that energy is going to remain one of the world’s largest industries, or even get bigger as a percentage of total GDP, and you look at all the largest energy companies in the market, you stack rank them by market cap and you stare at it and you say, ‘Which one of these can be trillion dollar market caps? And you think okay there might be one or two in here.’

And then you look at that list and you’re like, ‘Do we think Base is going to crack this list? And where on it do we think they will be?’ I think we will crack that list and I think we can be near the top over time. I think we will be at the top.

So. Biggest industry in the world. We want to win it. You tell me how big that would be.

Energy is an enormous industry, and arguably the world’s most important. It sits at the base of everything. The world’s largest energy company today, Saudi Aramco, is worth $1.7 trillion. Exxon Mobil is second at $650 billion, a cool trillion dollars less. What is remarkable about the energy industry, though, is that there are 18 energy companies worth more than $100 billion, and 55 worth more than $50 billion. At its current $13 billion valuation, Base would rank 182nd among publicly traded energy companies, between China’s COSCO Shipping Energy and Norway’s Vår Energi.

This is a top-down view of the situation, basically just confirming that it is indeed very possible to build a very large energy company. What’s notable about the list, though, is just how old the companies on it are. These companies were largely built around the winning assets of the previous energy system: oil fields, refineries, pipelines, tankers, centralized generation, poles, and wires.

The energy transition that’s been underway has already created new bottlenecks – the grid struggles matching variable supply with unpredictable demand, for example – and therefore new valuable positions.

Then the most ravenous consumer of electrons in history entered the market in the form of data centers. If a new energy winner is going to be born, now would seem to be the time.

The question is: how? In an industry dominated by well-established and literally well-resourced companies, how might a new entrant break into the top group, or even take the top spot?

This is the question I tackled in Power in the Age of Intelligence in February. My hypothesis was that a company needs to identify the constraint holding its industry back and focus everything on breaking it in order to win the High Ground with a scarce, defensible asset. Then, use that position and its capabilities to integrate outward.

This is what Standard Oil did, and Ford Motor Company, Carnegie Steel, and Swift Meats. The case study I used to show this idea in motion was, of course, Base Power Company.

Base picked the right energy bottleneck - the grid - and the right technology to unblock it - batteries. They picked the right place to put the batteries, right next to homes, in the right starting state, Texas. They vertically integrated, assembled the right team to own each chunk it bit off, and got to work.

“In 2023,” Zach told me, “we had this very opinionated view that batteries were this unique unlock to scaling energy consumption. And we thought that in a commodity business where cost is everything, to win, you had to vertically integrate and you had to leverage technology, because you needed to combine those things to get a structural advantage.”

With Base Core, Base has won the High Ground: the ability to make the cheapest residential battery and deploy a growing network of them in the most valuable place on the grid, next to homes, starting in Texas, the state where batteries currently make the most sense.

Owning the High Ground doesn’t mean Base has won, of course. Not even close. The High Ground is simply a strategically advantageous position from which to attack (“Job’s nowhere near finished,” Zach confirmed).

What it means, though, is that it will be a lot harder for any new entrant, or even any incumbent, to try to attack the bottleneck in the way that Base has, or to try to beat Base at its own game. Their customers will be harder to acquire, their electrons more expensive, and their ability to impact the grid less sophisticated.

As important as the position itself is what Base has learned capturing it.

“Our first three years were basically learning a bunch of hard lessons about how to scale, build a brand, build a team, raise capital, manufacture and design products, optimize the supply chain, and the output of all of that is Base Core,” Zach said.

Base’s young life has been a process of figuring out how to do a really hard thing, earning the right to make it a company capability, using it and other capabilities to do the next really hard thing, and turning that into a capability, too.

Collectively, Zach calls all the things Base has learned to do its “Capability Cannon”: the deployments and engineering and electrical services and manufacturing and brand and balance sheet and policy chops that Base has earned along the way.

The Capability Cannon is the most important thing Base builds, more important than any single product, as long as the opportunity ahead remains larger than its past accomplishments, which it should for a long time. Base can point its Cannon at the next opportunities it attacks.

It is already pointing it at new geographies and new customer types.

When we last saw Base, in Chapter 2, they operated in one state. Now they’re in two, having launched Illinois in June. Back then, fifteen months ago, its regulated utility business existed as one pilot. Last month, more than 50% of its installs were on regulated utility customers’ homes for the first time.

It learns lessons from these expansions, too, and it will apply them to the next state and the next utility, and the ones after that, until there’s a Base Core on homes in every state, every neighborhood, and maybe even in foreign countries, too.

It can also point the Cannon at new products, and at the AI infrastructure buildout.

The way Zach sees it, there are four things you can do with power: you can make it, move it, store it, and sell it. Make, Move, Store, Sell.

“There are massive opportunities across the whole stack,” he explained to me. “You think about the stack and power, it’s really four things: you have to make it (generation), you have to move it (transmission and distribution), you have to store it (batteries) and you have to sell it (retail and wholesale markets).”

A couple of the other people on the team mentioned it too, as in, “Has Zach talked to you about his Make, Move, Store, Sell framework yet?” It seems to be an animating framework.

And he shared it with Ribbit, too, which used MMSS as an organizing concept in the Power Letter it published last week.

Ribbit took it a step further, actually, applying MMSS to both electrons and intelligence, and grouping “Dam Breakers” by which of those functions they did to which commodity.

This is “capture the High Ground and integrate outward” in a picture, and the question is: which squares give you the best starting point from which to expand, and how far? This is why we get out of bed in the morning here at not boring HQ!

Maybe they move left and start to Make or Move electrons.

That would clearly fit the mission of fixing the grid, and of delivering affordable and reliable power to customers. Zach has hinted at owning generation since the beginning, and both Jared Greene, Base’s Head of Software, and Justin Lopas, its co-founder and COO, have talked about how you can start to integrate backwards into the grid once you own the Base Layer with batteries and software.

Maybe, just maybe, they even move down into intelligence to capture some of those blue squares.

“There’s a tendency for a company of our scale now to be like, ‘Oh, well we’ve got 500 people now so we can go and do more things,’” Justin told me, when I asked what they might do next, “But I still think - Zach and I talk about this all the time - that focus is extremely key.”

Justin Lopas (l) and Zach Dell (r) in Focus

Focus on what, though?

If what Base sells is batteries next to Texan homes, then any deviation from that - selling batteries outside of Texas when there are so many Texans left, or selling something besides batteries or electrons altogether - might count as a lack of focus.

But that’s not what Base sells, what it is building, or where its ambition ends.

Base exists to fix the grid and sell affordable, reliable power to those who consume it, and it is building a set of capabilities that it can deploy at the best opportunities to do that.

Three years ago, it had a very opinionated view that batteries were the best way to do that, and they nailed it. They nailed the strategy, and they’ve nailed the execution. No one brings batteries online faster, and no one makes batteries that deliver cheaper electrons.

“Now,” Zach told me, “it’s time to scale the fuck out of that.”

Along the way, they’ve learned a ton about how the industry works, and they’ve earned the knowledge that there are way more opportunities beyond making batteries to store and sell electrons.

“And so we’re going to do all of that,” Zach said… “And then we’re going to take that massive Capability Cannon and we’re going to point it at AI.”

And we’ll get there, I promise, but all of Base’s best laid plans, up to and including becoming the world’s largest energy company, rely on its ability to build its capabilities, to build the Core and then build outward from there.

Which is one of the reasons that Base started manufacturing its own batteries.

Manufacturing the Core

The plan, when I spoke to Base’s Head of Hardware Dino Sasaridis for Chapter 2, was to not manufacture batteries for a little while. Crawl, then walk, then run.

The Gen 1 battery, the one Base has been installing on customers’ homes, was a basically off-the-shelf battery system.

Gen 2, he told me, was being designed to Base’s requirements and manufactured by partners (one battery partner, and two inverter partners). That was the focus at the time of writing.

Then, once they’d worked out the kinks, they’d make the battery they really wanted to make, Gen 3, the vertically integrated, Made in Texas, absolute best battery in the world.

Gen 3 is why Dino joined Base, after thirteen years at Tesla, where he led mechanical engineering on Powerwall 3.

“I decided that whatever I did there I can do it again,” he told me, “and I can do it better here, and then Base Power Company can have the absolute best residential battery in the world. And I think we’re going to easily achieve that, honestly.”

“That best residential battery is Base’s Gen 3, which it’s planning for now. Gen 2 is up next, later this year,” is what I wrote then, in April 2025…

“Actually, the first time I recommended we cancel Gen 2 was right after I joined [in May 2024],” Dino admitted when we caught up a couple of weeks ago. “And I was like, these suppliers do not care about our success as much as we do. We should not do this.”

He actually looked into canceling the contracts then, but for a number of reasons, including the fact that the $15 million in cancellation penalties would eat a big chunk of the money Base had raised, they couldn’t get out. So he turned his focus to making Gen 2 as good as it could be.

The idea of just canceling it kept nagging at Dino, though, “Because at the end of the day it was still just buying a product from someone else, money transacted at day zero. And the reason that ends up being a problem – I’m sure I’m the 50th person to tell you this today – is that if you don’t own the stack of the technology, you’re absolutely not empowered to fix it when it breaks.”

“So I took a hard look at the design, mechanically and electrically, and asked, ‘Is this thing going to break?’ And I was like, ‘Yep, it’s probably going to break.’ It was in a lot of ways more complicated than our Gen 1 product.”

Worse, Dino suspected that Base wouldn’t actually learn anything new from Gen 2, which is why you’d ever waste time walking before you run in the first place. Focus means focusing on things that grow your capabilities, and Gen 2 just wouldn’t.

Everything was pointing in the same direction: Dino had gotten himself to about 80% confident that Base would move faster if it just cut Gen 2.

Then he went to Zach Dell, Base’s CEO and suggested that “Hey, crazy idea, I’ve looked into it for almost a year and I really do think we should cut Gen 2 and just do it ourselves,” to which Zach responded something like, “Woah, that’s big, I need to think about this.”

He did, and when they spoke two or three days later, they’d both walked it back a little, and they kind of nervously laughed, like “That was crazy, right?”

But you know how this works. You’ve felt it. Once you get that feeling in your gut, whatever the details, the only thing that will resolve it is just cutting bait? That’s where Dino was.

Which is why, even though they’d put the cancellation on pause, Dino kept talking to people about what if.

In a one-on-one, he mentioned to Jared, that “Yeah, I was talking to Zach about how we should probably just cancel Gen 2, but never mind, that’s crazy.”

“And Jared looked at me and said, ‘Oh my God, you’re right. We need to cancel Gen 2.’ And I was like, wait, am I right?”

And whether he was actually right or not, it ping-ponged from that conversation to a leadership team meeting within the week, right around April 2025, right when the Liberation Day tariffs made it that much more attractive to make as much as possible in the US anyway, at which point Base’s leaders decided, “Okay, fuck, we need to do this. Let’s rip the Band-Aid off.”

On May 1, right on schedule as always, two weeks after I’d published Chapter 2 with a few hundred words on the merits of Gen 2, Zach sent out the monthly investor update.

Everything had moved in the right direction in April. Monthly installs crossed 400, with a new high day of 26. Closed customers grew 30% MoM. Austin Energy had selected Base to move on to the next stage of its RFP process.

And one more thing…

We made the decision to cancel our Gen 2 program and accelerate Gen 3, a lower cost, Base designed and manufactured (in Texas) home energy system. We’ve re-allocated all internal engineering resources from Gen 2 to Gen 3, worked with our Gen 2 suppliers to wind down all work with them, and focus 100% of our hardware & supply chain efforts on accelerating Gen 3 development.

“It might have changed the game,” Dino tells me. “It might have pivoted the success of the business, honestly.”

“We had dozens of people immediately stop working on it and start focusing on Gen 3. There was a bunch of throwaway work we avoided. And the number of meetings per week talking about the day-by-day slip of one inverter supplier versus the other… all that shit just turned off.”

“Then, because we had all the resources freed up, we did a big architecture week,” he continued. “The power electronics team is in San Carlos, so I invited them here. We gathered the desks in a big circle and sprinted toward a concept for Gen 3. Then we got Zach and Justin and Jared, the whole team, together and looked at it. And we were like, ‘Oh man, we can make this so much better than it was going to be.’ I think at that point we all felt like we’d dodged a bullet. And then we went and made it.”

“And then we went and made it” sounds so simple, but making batteries was not something Base actually knew how to do back then. So they called Andy.

After eight years at Tesla, where he left as the Director of Battery Manufacturing Engineering, Andy Ross decided to take his talents to General Motors as Director of Manufacturing Optimization. General Motors is a very large company and there is a lot to optimize.

Just one week into his new role, in August 2024, Justin and Dino reached out to Andy to talk about Base. At the time, Base wasn’t manufacturing its own batteries, and wasn’t planning to for another generation.

“The timing isn’t great for me,” Andy told them, “and you don’t need anybody who does what I do either. But feel free to use me as a resource.”

Andy joined Base in March 2025.

Has it been more fun than GM? I asked.

“Oh my God, so much more fun.”

Base people have their own definition of fun. In Andy’s case, “fun” meant:

• He joined in March,

• They decided to cancel Gen 2 and manufacture Base Core in-house in April,

• They signed a lease in the old Austin American Statesman building in June,

• They got the keys in September, and

• By March 2026, his one year anniversary, he and the team had landed two production lines and begun putting Base batteries in the field.

The company refers to this level of speed as Base Pace, and it runs on a mix of process and elbow grease.

When I asked Andy to describe the culture of the factory team today, now that it’s grown from four to one-hundred-and-thirty-four, he said, “We want people that can solve tough problems but don’t really care where that problem is, people that don’t care what the nature of the problem is they just want to be a part of fixing it.” That comes from the top.

As soon as Base took possession of the old building, they wanted to get to work. But it was late summer in Texas, very hot, and there was technically cooling tower infrastructure in place but it hadn’t been cycled in seven years and they weren’t going to be able to get it recommissioned in time, and they’d quoted some retrofits to be temporarily installed but they were outrageously expensive, like $1 million, and they wouldn’t be ready for seven months, and the plan, at the time, was to only stay in Base Factory 1 (BF1) for thirteen months anyway, so that wasn’t going to work.

“So I just started fielding local rental companies,” Andy recounts, the way you recount war stories from the good ol’ days, when there were just four engineers on the team and no maintenance leader and so you just figured it out yourself, “to just start landing some chillers outside the building. Good, bad, or otherwise, the door-opening mechanisms on the dock overhead doors were broken, and they’d be outrageously expensive to repair too, so I was like, ‘All right, we’re just gonna land them in front of the dock doors’ and me and one of my NPI guys were just sawzalling holes in the dock doors to get ducting into this place because it was late summer so it was like 95 degrees in the building and we wanted to get to work.”

“The first night they showed up, I was here until about midnight on a Friday with the rental HVAC company just trying to get this building less hot,” he remembers, with that glint. “I never thought I’d be sawzalling dock doors to duct tape and chewing gum some air conditioning into a building, but everybody was very happy when they weren’t dying of sweat on Monday.”

He applied that same elbow grease to the facility’s compressed air system. When Andy realized that he wasn’t going to be able to get the leftover 1980s-level factory compressor system working himself, he “went on Tractor Supply and bought the largest, but also 120V, compressors that I could find, and I was in the ceiling making up a makeshift pneumatic system for all of our pilot systems, so running pipe and dropping lines and everything.”

Andy calls these his “hacky redneck” fixes. They’ve largely been replaced at this point, now that Andy has a team. But because he’s done many of the hard jobs himself, when he asks his team to do the same, he can tell them with a straight face: “Sorry it sucks, but it isn’t going to exist otherwise.”

“When I was at Tesla,” he told me, “it was a passion of mine to personally operate every station on every line on every continent that my team deployed things to, because a good solution to an engineer might not be a good solution to the person that has to use it.”

Which is probably why, as Andy grew the team from those initial four engineers to the one-hundred-and-thirty-four people that make it up today, 80% of the salaried team members are people that Andy has worked with before, which I suggested must feel pretty darn good.

“The thing that’s been special about manufacturing in particular is the amount of great people that have been,” he searches for the right combination of words, “bold enough, hopefully not foolish enough, but ambitious enough to pick up my call and hear me say, ‘You want to go do something crazy?’ and they say, ‘Yeah.’”

“Bold enough, foolish enough, ambitious enough… doesn’t matter. They want to go solve big problems and we had big problems.”

“We had no factory, we solved that problem, we got a factory. We had no team, we were solving that along the way. And we had no concept of a manufacturing strategy yet, then it was all just philosophical, on the computer, and then we went out there and found new machines, new vendors, new equipment from all over the world to go achieve it.”

When Base started sourcing the equipment for BF1, it was just Andy and one other manufacturing engineer. By the time they were ready to start placing purchase orders, a few others had started to trickle in. Andy would tell them, “Welcome, this is the thing we’re going to buy tomorrow, you’re going to be in charge of executing this project now.”

Those early days in BF1 might not have been pretty, and will certainly look downright primitive compared to BF2, BF3, BF…X. That’s great. At Base, you figure something out, and then you turn it into a capability. Your best laid plan needs to face harsh reality, so you can learn, improve, and try again.

Plus, whatever the Manufacturing team had to do, it worked. Fast.

In Austin, Texas, right down the street from Base HQ, “We landed two battery module manufacturing lines and one inverter manufacturing line (the ratio is because there are four battery modules per inverter).”

By January, a little over one business quarter after taking possession of the space, Base’s pilot battery module line was up and running and the team started pre-production builds.

All told, over the course of the last fifteen months, Base got out of Gen 2 contracts with only $1 million in penalties, built out a manufacturing team, opened a factory in the old Austin American-Statesman press building, figured out how to add a generator hookup to the battery, stood up a supply chain, built the first Gen 3 batteries, tested them in as much as three feet of water, rewrote software and firmware to match their new capabilities, built more Gen 3 batteries, began installing them on employees’ homes, then friends’ homes, and then all sorts of Texans’ homes, and renamed the battery system Base Core.

All that done, Base now offers what Dino is “quite confident is by far the cheapest residential battery on the market.”

By February, it had its first Base Core system ready to install.

The Day The Team Installed the First Base Core

Manufacturing is a scale story. You certainly don’t set up a whole factory to make one battery. But you don’t scale up a whole business without stopping to appreciate one battery every once in a while, either, especially when it’s an important one.

That first Base Core, which people from almost every Base team gathered to install on a teammate’s house in February, was an important one.

Dana Paz, the Anduril alum who is Base’s Head of Deployments, responsible for the multi-state logistics and installation network that Base is growing, told me that it was the highlight of this chapter in the Base story.

“That first Gen 3 install – sorry, Base Core, don’t tell Cole [Jones, Base’s Head of Growth] I said Gen 3 – truly was super epic,” she said. “There were so many people in a tiny packed side alley and the vibes were so high.

For those, like Dana, who had been there from the beginning, it was an opportunity to reflect, briefly, on how far they’d come. “The first Gen 1 install was me and Cole with the Craigslist electrician. It took over two days to hack the thing together and it was just pure chaos,” Dana remembered. “And then this one was chaos but in a much better way”:

We now have an entire support system, so you have the engineers on-site, the hardware, the firmware, the electricians were there, the deployments people were there. We were finishing stuff in the factory, off the line, functional tests, like throw it in the back of a person’s car, drive it over to the same house to do the install. We had a blackstone out, we were cooking tacos. It was just way better vibes on Gen 3 than Gen 1, which was cool.

“It set the stage well for us to then go out and do something hard,” Dana said. “We had a bunch of pilots that we had to go rework a bunch of times, because we would try something new out and it wouldn’t work, so people would be going back to the same house like six times. So you’re like, ‘Is it really working?’ I don’t know, you have your doubts. But being able to anchor on that super epic first install… yeah, that was the highlight for me over the past year.”

Something is working.

This month, only fifteen months after it decided to make the thing in the first place, Base will install more Base Core systems than Gen 1 systems.

How to Build Your First Battery Factory

There’s been a lot of learning in between that first Base Core install and the thousandth, naturally and on purpose.

Base likes to make a thing, get it out in the field, hook it up, and learn from it as quickly as possible, so that it can tweak its design, maybe supply chain, and manufacturing lines as quickly as possible, so that it can keep compounding as quickly as possible.

For the Hardware and Deployments teams, that meant getting the first one out and hooked up to a house, to see where it broke and tweak the design accordingly.

For Andy and the Manufacturing team, that early iteration loop meant that they needed to keep the pilot facility, BF1, as manual and flexible as possible.

“What bites you in manufacturing is you design a thing that you think is going to be the thing you launch on and then you launch with a mildly different, sometimes radically different, product architecture and you’re scrambling from a manufacturing engineering and equipment standpoint to try to recover the machine to match the thing you need to go build,” he explained. “And the thing that Tesla taught me was to try to limit the amount of one way doors, or the inflexibility, as long as you can from a manufacturing line standpoint.”

This is why automation comes last in Elon Musk’s five-step algorithm. Automation is inflexible.

“That’s one reason that we kept it pretty manual to start out at this factory, was to enable a higher amount of flexibility in both process and in design as long as we could,” Andy said.

Another reason was that Base’s systems weren’t set up to handle much more than manual.

The Manufacturing Execution System (MES) - the software that tracks, monitors, documents, and controls the manufacturing process - that Base had been using was “originally built to do retail clothing inventory tracking. It was something you would use in a job shop where you’re making 50 of a thing and then you’re making a new thing, not where you’re making 1,000 of a thing per day.” It was a low-volume tool where Base needed a high-volume one.

At first, the team tried to redneck hack it by building an architecture in which the original software system might work. Maybe at the first level it didn’t look good, but at the second…

“No, when we dug into it, it still wasn’t very good.”

They presented their findings to Justin, and told him they felt confident they should go build their own MES, a significant undertaking that would require going slow for a while to go fast later.

“One of the things I appreciate about Justin is that he values and appreciates the experience the members of my team bring,” Andy said. “And when we presented the idea of making our own MES to him, he looked at the team and said, ‘Yep, make it so.’”

So they did. They released the first version of the system in January, right as the line got up and running. By April, it was the system of record at every station. By June, the manufacturing software team was integrating live controls for the factory’s 21 robots. (Being smart about when to automate doesn’t mean no automation, as Justin describes to Molly O’Shea during their BF1 walkthrough.)

Then they built Whimsy, which stands for “Where the heck is my stuff y’all,” to track inventory from the moment it arrives at BF1 until it leaves for one of Base’s six-plus warehouses around the country, and they built a bespoke Quality Management System, to handle supplier quality, production approvals, and non-conforming parts.

In each case, from the first install to the flexible manufacturing line to the internal tools, what Base is becoming excellent at is the process of going from “We’ll figure it out” to encoding the solution in process or software as quickly as possible.

Figure It The Fuck Out

I first discussed writing this piece with Gabriel Federman at Base a couple weeks ago. Three days after that, on a Monday, he’d packed my calendar with meetings with about ten team members. And that same day, I got an email from two Base interns, Prithika Devarajan and Richa Pandya, subject line: “At Base Today?”

Turns out, they’d each seen my meetings with their respective bosses on their calendars and decided to take the initiative and reach out. Also turns out that taking initiative is a big part of the culture at Base, and when we spoke the next day, Richa and Prithika each had stories.

Richa, a University of Toronto electrical engineering student working on Dino’s Hardware team during her eight-month co-op, has mostly been focused on the inverter, on field returns, and on new designs for the battery management system.

But as soon as she got to Base, she started looking through the Google Drive, where she found [REDACTED UPCOMING BASE PRODUCT] and thought, “This is the coolest thing I’ve ever seen.” She told Dino she wanted to work on [REDACTED UPCOMING BASE PRODUCT] and he said, “Great, write me a proposal,” and he meant like a proposal of what she’d work on, but Richa came back with a 10-page report on what the commercial version of [REDACTED UPCOMING BASE PRODUCT] could possibly look like, and Dino said, “This is really cool, you should keep pressing,” and so now Richa is working on that, too.

Prithika, a University of North Carolina-Chapel Hill student (there really is talent everywhere I guess), is working on Gabe’s marketing team, on some SEO stuff, and on the vision and design for a commercial energy product that’s launching soon, and even spent two weeks boots-on-the-ground in El Paso, on which more in a second.

But she was also fascinated by the company’s utility partnerships, and noticed that Base didn’t have a map of all the batteries in Texas and whether they were charging or discharging, like Powerwall does, and so she made a quick MVP and brought it to Lauren Thomas on the Utility Partnerships team. Lauren looked at it and said, “Actually, this is fricking sick. I don’t really care that you’re an intern, I want you to go figure this out and make it better and come back to me as soon as possible with it.” Which she did.

This kind of openness and cross-pollination is part of the culture, too. There are fifty-five (55!) interns at Base this summer, and they all work on different teams and hang out all the time, which means they all have visibility into all the different areas of the business.

It sounds like a ton of fun, but when duty calls, the interns are expected to do their core jobs, whatever it takes. For example…

Base recently launched a utility partnership with El Paso Electric that Tim Pianta, Base’s Head of Utility Partnerships, was telling me about the day before I talked to Richa and Prithika. What made the partnership unique, he said, was that its goal was deploying in specific, particularly constrained parts of their grid to defer capacity upgrades.

What that required was installing 10MW behind a particular substation, which meant getting roughly 10% of all the homes behind that substation installed, all before a September deadline. It was a higher penetration rate, and a more targeted one, than Base had ever attempted.

Prithika was tasked with helping make that happen.

“In El Paso, we had to deploy 10MW of batteries,” she told me. “So they sent me and Olive, another marketing intern, to El Paso.”

While Base has signed up thousands of Texans, many through utility partners, the level of geographic precision needed in El Paso was new.

“Tim was like, ‘You need to figure it out,’” Prithika said, “And so we did. It was 104 degrees, we were there for two weeks, we knocked on doors, we hosted happy hours at coffee shops, we showed up at farmers markets, and we were able to pull it off and end up two weeks ahead of schedule.”

In the end, one-third of the neighborhood signed up and 10% got installed (many didn’t qualify). Cole, who has overseen every customer acquisition effort in Base’s history, confirmed that those numbers are “freaking crazy.”

Part of the lesson, she said, was that “we will figure it the fuck out no matter what,” but the other part is that you need to turn that into learning how to do a process well. “The next utility will be such smooth sailing now that we’ve figured it out.”

“A lot of how we think about things here,” Richa added, “is as a factory. The focus from the beginning of anything is how we’ll be able to standardize it and turn it into a factory, which influences how you design the product in the first place.”

Then she said something that describes Base in 2026 as well as anything I’ve heard, how it’s able to maintain its culture and grow really fast and not stagnate but not descend into chaos, how it can compound on the core and hold the High Ground and expand out.

“Imagine Base sort of like a web. It’s expanding outwards, and all of our new ideas are at the edge, and all the scrappiness is at the edge, and in the center, we’re trying to take in new things once they’ve been processed at the edge and standardize them into processes and software and make them more efficient.”

To synchronize our metaphors, the Cannon is the result of the Web.

Learn something, then turn it into a process or encode it in software, and as a result, the team is experimenting with new tools every day, they told me. Where the factory has its MES and logistics has Whimsy, for example, the Software team has been working on building out Base OS, which takes customers from top of funnel down to getting their install scheduled.

Historically, this has been a bajillion tools tied together through APIs and now they want to do it through one homegrown system.

“So it’s constantly failing, but it’s also constantly getting better,” which is a beautiful metaphor, too.

Build, fail, get better, own that capability, build on it, fail, get better, outward until the web envelops the industry.

That is what Andy and team are doing in BF1, too.

By the end of the year, they’ll need to be pumping out about 1,300 Base Core systems, consisting of four 10-kWh modules each, every week. That’s roughly 5x from today, with those gains to come from adding a night shift, and from “attacking all of the uptime problems, the cycle time opportunities, and standardizing a lot of the workflows.” Take some scrappy redneck process, figure it out at the edge, and standardize.

The year after that, they’ll need to 3x, and then they’ll need to 2x again, at least on the current trajectory. That means that in 2028, they’ll be producing 30x more batteries each day than they do today, from the comfort of the bigger, more automated Base Factory 2.

From 2028, today’s production numbers will look tiny, but today, they’re cause for (brief) celebration.

“You should see an all-hands,” Richa and Prithika told me.

“It’s literally just Justin and Zach saying some KPI and everyone starts shouting and clapping, everyone, every time there’s a win. Recently, in the factory, they’ve been hitting records on how many modules and inverters they produce. Every time we hit some kind of record, everyone is just so, so into it and excited.”

Last month, the factory made 295 10-kWh modules in a single day. The crowd went wild.

Unbelievably Fast

The thing about constraints is that they move.

Figure out the factory, and something else becomes the thing standing between Base and flawless execution.

When I asked Justin which team currently had the Turtle, the stuffed animal turtle that sits on the desk of the person most responsible for clearing the critical path, he didn’t hesitate: “Squarely on the Supply Chain. Basically, we’ve been able to ramp our factory faster than four or five of the key suppliers we’re sole sourced to right now.”

Because Andy and his team ramped up the factory so fast, the Turtle now falls to the team whose job it is to fill it with parts.

Colby Jarrett is working on it.

Colby came over from Tesla, too, where he worked on the Vehicle program, unlike a lot of the Tesla refugees at Base, who worked on Energy there before working on Energy here. But Colby did work on batteries for vehicles, most recently as the Senior Manager on Technical Programs for Cell Manufacturing Engineering, and on every factory launch other than Fremont, which predated his 2017 start date. During his eight years there, Tesla got bigger, and more bureaucratic, and he got tired of that, so he looked for something smaller, spent a bit of time at Form Energy in Pittsburgh, and then jumped to Base.

Colby joined to work for Dino in Hardware, working on program management, compliance, and sourcing. When he joined in July 2025, there was no factory; there was a program and an early design for Gen 3 and that was about it.

“And we’ve been kind of push-pull-flipping it up the hill ever since.”

When Colby started at Base, one of the first problems he was tasked with was “picking a supply chain, so roughly 20 vendors to make 40 parts.” I figured he was the right person to ask about the challenges Justin brought up about vendors not working at Base speed, and how you solve those.

He gave me some technical reasons.

Like, Base has something about its design that’s hard that they didn’t realize would be hard and now the vendor is having yield issues. As we speak, Base is ramping more than ten supplier lines alongside its own, sending engineers to supplier floors to develop processes, install hard tooling, and project-manage the move from bridge suppliers to production suppliers.

Or, Colby adds, at the “commercial interface, at the speed that Base moves, it’s pretty much impossible for us to tell them exactly how this thing’s going to be made in high volume. By the time we need them to start, they’ve got to start with quite a bit of ambiguity still in the product design and in the program, and removing that ambiguity without losing commercial leverage on price or schedule is a very hard problem with vendors.”

This kind of stuff is new for the company, and the people who started it. Justin admitted that he had kind of a blind spot, because high volume manufacturing is very different from low volume manufacturing, and at SpaceX and Anduril, he “had done mostly low volume manufacturing.”

“And in low volume, you can kind of brute force almost anything. ‘Oh, we didn’t get the part done Friday? We’ll just work this weekend to finish the machine part for Falcon or Starship or whatever.” But, he said, in high volume, “you plan production once in advance, you sign long-term contracts, you reserve long-term capacity. All of these are concepts that are intellectually obvious but when you actually go do them in practice, it’s a bit different.”

But you do it, you make contact with reality and you learn and you earn the right to get to scale and to hire even more great people like Colby, and then you continue to figure it out.

You can figure out all of the hard technical stuff, but maybe the biggest thing you have to solve, Colby told me, is something softer.

“I speculate that some of our bigger vendors just didn’t expect us to execute,” he told me.

“We showed them a timeline and they said, ‘Yeah we work with a bunch of startups, we’ve seen this movie before, you’re going to come back and push the schedule on us.’ But plus or minus a month or two, we’re on the same schedule we published 18 months ago.”

“Some of them have had their little ‘Oh Shit’ moment, that this is really happening at the rate that we said it was going to, and are trying to play catch up because they didn’t really start out of the blocks.”

So you have to prove to them that you’re going to do what you say, which you can only do with time, and which Base has now done, and you have to sell them on a vision of what happens when you continue to do what you say.

“Base as a customer maybe isn’t a big deal now,” Colby said, “but they’ve got to understand the opportunity and the ambitions of Base as a whole and of the Base Core battery so that they’re incentivized to grow with us.”

In short, Base needs to get its suppliers even half as excited about Base Core as its employees are.

The Platonic Ideal of a Battery

Just as Dino wasn’t excited about Gen 2 from an engineering perspective, “Dana wasn’t excited about it from an operations perspective, I wasn’t excited about it from our members’ perspective,” Cole told me.

“We’re like, ‘Wait, what are we doing? This doesn’t make nearly as much sense as it would if we were to build the platonic ideal of a battery.’”

Cole is very excited about Base Core.

“Base Core has been the company’s big bet, to make it real. And now it is very real,” he said. “It’s the biggest, best thing for the energy world, ever maybe. In my opinion, it is the key to energy abundance.”

In terms of its specs, Base Core is the best battery on the market.

Base Core is the highest-capacity battery on the market, with 39.2 kWh of energy capacity, compared to 25 kWh for Gen 1 and 13.5 kWh for Tesla Powerwall 3. Many customers get two Base Cores, for nearly 80 kWh of capacity, which gives Base more ability to balance the grid, and room to add energy-consuming products to the offering.

It has a 20 kW inverter, which is crazy powerful and means that Base can push out a lot of power very quickly to support the grid. Justin tweeted that they believe the inverter is one of the most American-made on the market, which is particularly important now that the FCC has added foreign-made inverters to the Covered List since they’re vulnerable to being shut off remotely.

Base Core operates in temperatures as low as -22°F, much colder than Gen 1’s 14°F, which will come in handy (necessary) in Illinois and other cold-weather markets. It is physically smaller, and its volumetric energy density is nearly 2x that of Gen 1, meaning it packs more energy in a smaller package. Its transfer time - how long it takes to turn on when needed - is 10x faster than Gen 1, at 50ms instead of 500ms, the difference between noticing a flicker when the power goes out and not noticing it at all.

Not only is it the best battery on the market generally, it is custom-built to do what Base needs it to do.

Dino told me that the priority ranking when designing the battery was: 1. Safety → 2. Function → 3. Reliability (All Must-Haves) → 4. Cost (why we’re going to win) → 5. Install (why we’re going to win) → 6. Serviceability → 7. Beauty → 8. Delight.

Base optimized for landing a lot of relatively inexpensive energy capacity at each home in order to spread inverter, customer-acquisition, and installation costs over many more kilowatt-hours (kWh). Accomplishing that means it’s the cheapest residential battery on the market, and is ideal for long backup duration, and for buying, storing and trading more grid energy.

The more kWh (and MWh and GWh) Base controls, the more money it can make. Base aims to roughly breakeven on the batteries themselves - it charges members a few hundred dollars upfront, to cover the install costs, and $19 monthly fee in some areas. It makes most of its money today by storing power when it’s cheap and selling it back to the grid when it is expensive. Every battery is a time machine.

Increasingly, Base makes money by helping utilities put more storage capacity online, as well, in order to better balance and get more out of their grids in order to delay capital investments.

In both cases, more capacity directly translates to more revenue. And given that customer acquisition and install costs are roughly the same whether you’re installing 39.2 kWh or 20 kWh, the more you can pack in each system, the better the unit economics.

This is one of the benefits of vertical integration: you can design your hardware to do exactly what both your customers and your business need it to do. Base Core has a little something for everyone.

For Dana and Deployments, this is a battery that was designed to be installed.

“The hardware was designed to install with ease,” Colby told me, “because we knew emotionally that we were going to have to be the ones to install it and didn’t want the pain and complexity of a hard battery install. I think if Base weren’t as vertically integrated as it is across the whole business, not just the hardware stack, those dependencies are less obvious to optimize for.”

The biggest change is to Disco, so named because it’s an internal disconnect that sits between the grid, the home’s main panel, and Base Core. It allows Base to split the install into two stages: electricians do the detailed electrical work when they disconnect the home from the grid with Disco, and installers come to plug in the module the next day.

“Install basically goes from three steps down to two steps,” Dana told me, and importantly, it decreases how much scarce electrician time Base needs to use.

Designing for install does mean making trade-offs. For example, Dino told me, they could have squeezed in more energy per module, “but it would have been a trade: now every single time we lift the battery, it would have to be done by a piece of equipment and it can never be lifted by a person.”

Vertical integration allows you to control that trade-off. One of the things I’ve appreciated about Base from the very beginning is that it optimizes for the whole deployment instead of any one spec. They call it the Deployment Factory, and the more Base controls, the better a factory it can design.

Base Core comes with features that help the Software team, too. The same Disco has new telemetry on it, which helps them get to base truth on any issue very quickly.

“For Gen 1, sometimes people would call us and say ‘my lights are flickering’ and we’d try to diagnose it based on information you give us and our one-hertz telemetry,” Jared told me, referring to the fact that it ran 1 time per second, a huge upgrade from the every five minutes utilities are used to but still too slow for the precision Base wants. “Now, we implemented new telemetry on the Disco, which at seven kilohertz (7,000 times a second) is running analysis on the actual waveform.” With that, the team can diagnose, and then fix, most issues remotely and near-instantly.

The new battery is flash-flood tested and can be submerged in three feet of water, too. Dino mentioned that they “welded the lid on like a submarine. We hardened the crap out of it because we wanted them to be able to work underwater. And then we tested it, with water all the way up to the top of the inverter, and it just works. Full submarine mode unlocked.” This sounds like it’s a cool-to-have, but it also expands where Base can operate to include low-lying flood zones.

Then there was the big one for Cole, the feature that answered the most common customer question: “How long is this thing going to last?”

So Dino and his team added a generator port, and now Cole and his team can honestly tell customers, “We have this generator port, so it lasts as long as you need. Obviously, you can use the battery as your first line of defense, but in the event of a winter storm, you’ve got this thing included, and it’s only a few hundred bucks to add a generator to it.”

Think of it like adding a gas tank to a Tesla to eliminate range anxiety. Most of the time, you’re golden with the battery, but if you can’t find a charging station and you just need to fill ‘er up quickly, you can. “I think it’s the best feature, honestly.”

As Dino pointed out when deciding to kill Gen 2, though, this thing is going to break. The thing I heard from multiple people that makes making Base Core in-house so great is how fast you can fix it, which Dana said means that “The iteration loop is much faster.”

That is the biggest benefit to vertical integration: the speed of the feedback and iteration loop.

On an early install she went to, “One of the installers, the electrician, was like, ‘Why did we do this thing that way?’ And I turned to the designer and was like, ‘Yeah, I don’t know, designer, why did you do that thing that way?’ And the designer was like, ‘Oh, shit.’ And so all that stuff just makes the feedback loop much tighter, which just makes it a little bit better.”

“Because we’re so vertically integrated, if something’s broken in another org, I can go fix it” Colby agreed. “I don’t have to argue with people outside my organization and work through a commercial interface. That’s a speed enabler, which I think is huge.”

Put all of it together, and you have what the team is quite certain is the best and cheapest residential battery on the market, a killer combination that, once installed, will protect you from outages and save you money on your electricity bill, and it keeps getting better as the team learns from contact with reality.

Which is why everyone seems to want a Base battery.

More Demand Than We Can Meet

At the end of each of my conversations, I asked people what they would title this chapter in the Base story.

Zach asked to come back to me with a title, but said “The subheader of mine is: commercial validation, just rip-roaring product-market fit, more demand than we can meet.”

Base’s demand has come from what Cole calls the three-headed monster: homebuilders, utilities, and direct deregulated customers. Each has been humming, with utility surging to become the largest source of demand over the past few months.

In July, Base signed over 6,000 new Battery members, roughly 40% more than in June, while installing a record of more than 2,000 systems. For every battery it installed, more than three new Battery members signed up.

Despite the manufacturing ramp, Base is still supply constrained, which is why Andy is so focused on increasing production 5x and why the Supply Chain team has the Turtle. As many Base Cores as the team makes, there will be demand.

But recently, for the first time, being a Base Member doesn’t necessarily mean installing a battery. In April, the company launched Base Energy, “an energy plan with guaranteed below-market rates,” for renters, apartments, and other Texans “who might not be ready for a battery just yet.”

Base Energy is another early bet beginning to pay off. In Chapter 1, we discussed why Base operates as a Retail Electric Provider (REP), or more specifically, a “gentailer,” a retailer who owns their own generation. I wrote:

The battery means that Base can offer customers a lower price on their electricity. The reason is counterintuitive.

Selling electricity in a deregulated market is a low-margin, high-risk business. It’s low margin because it’s competitive, and REPs have to buy from generators and sell to price-sensitive consumers. It’s high risk because prices can fluctuate wildly. For most REPs, that makes selling electricity even lower margin, because they need to hedge their risk with financial instruments for which they pay a premium.

“The hedge is a financial battery,” as Jared described it. “The battery is a technical hedge.”

Since Base can hedge with technology instead of options, it can pass the savings it would have otherwise spent on premiums back to customers.

Base Energy shows that that’s true on a network-level, not just member-by-member. With its fleet of batteries, Base is long power: when electricity is valuable, it has stored electricity it can dispatch. A retailer, like Base Energy, is short power: it has promised to serve customers and must procure that power. Combining the two can reduce risk and improve economics, and can also serve as top-of-funnel, a way to build member trust, and a word-of-mouth engine. “How is it that you’re paying 10% less than me for the same electricity?”

Three months after launch, Base Energy had reached more than 14,000 members, bringing Base’s total membership across both offerings above 30,000.

The direct business - with or without Battery - is growing, as is the homebuilder business. Lennar accounted for roughly 2,500 homes, or 30% of Base’s 2025 installations, and by March 2026, Base was available in 77 Lennar communities with a take rate above 65%.

The biggest growth engine, however, has been utility partnerships.

Providing Utility to Utilities

Fifteen months ago, we focused the entire “Expansion” section on regulated utility partnerships, of which Base had just signed its first, with Bandera Electric Cooperative.

When asked how Base might fail to become what it could become, one of its investors told me, “If Base doesn’t work outside of deregulated markets, if it can’t break into utilities, it becomes a less good business. The size of the prize is smaller, and it’s more competitive.”

The good news is: Base has been able to break into utilities.

Tim Pianta, Base’s Head of Utility Partnerships, told me back then that while Base had the best battery technology, and believed it offered the fastest speed to capacity and highest reliability compared to grid-scale batteries and Virtual Power Plants (VPPs), the whole game was to “chop wood, carry water,” to take the time required to prove to utilities that it could do what it said it could.

Since then, Base has been proving itself. That initial Bandera pilot has turned into five larger utility partnerships, totaling 220 MW:

“When I last talked to you,” Tim remembered, “I think the regulated part of our business was quite small. In 2025, I think we were on the order of a couple percent of the business. This month and through the end of the year, about 50% of the installs that are happening are regulated utilities installed. So we’ve kind of caught up with the deregulated side to some extent.”

In fact, a full 58% of Base’s July installs were in regulated utilities’ territories.

This makes Cole’s job easier, at least from a customer acquisition standpoint. “From a go to market motion, the utility world is quite simple,” he told me. “You get like 5% of the addressable market signed up just by the utility sending an email. It’s a free CAC. I love it.”

It’s “free,” of course, only after the months or even years of work that Tim and his team put in to sign a utility partnership, which is a long sales cycle in its own right. But as Base continues to prove that it can get capacity online faster than centralized solutions, even utility partnerships are becoming lower cost to acquire.

When I spoke with Zach, he was on his way out for a road trip to three of the largest investor-owned utilities (IOUs) in the west.

“All the big utilities are calling us and they’re saying, ‘Please help us,’ basically,” Zach told me. “That was just not true fifteen months ago.”

There are different types of utilities, each of which might want something different from Base, which means that while Base manufactures one battery, it can monetize it in a number of ways.

In a deregulated market like Texas, it owns the battery, sells electricity to members as a REP, and monetizes the fleet. With co-ops and municipal utilities, Base deploys a contracted fleet that the utility can use. With investor-owned utilities, the IOU might buy the assets through a build-transfer contract, because they can place the batteries in rate base and earn an allowed return on the approved investment, and pay Base for software and services.

Base’s innovation, from its initial decision to own the batteries instead of selling them to retail customers, has always been a business model as much as a technical one.

It has been able to meet customers, whether retail or utility, where they are, and as a result, Base is well on its way in Texas.

Its batteries backup early adopters’ homes in Dallas-Fort Worth, Greater Houston, and parts of the Austin metro. It has partnered with the state’s forward-thinking utilities, including recently signing its home utility, Austin Energy, through which it signed up more than 1,000 of its neighbors in the first month.

There is an argument to be made that Texas is perfect for Base, and that it should continue to focus on pressing its home court advantage.

Base, however, does not want to simply become Texas’ largest REP. It wants to be America’s Next-Generation Power Company, and America is bigger than Texas.

So in June, for the first time, Base pointed the Capability Cannon north and fired.

Daaaaaaaaa Batteries

On February 1, 2026, right on schedule as always, Zach and Justin sent the monthly update to its investors. In it, they shared that Base had decided on its second market: Illinois.

Illinois is large, with roughly 4 million single-family homes, deregulated like Texas, and has become one of the most battery-friendly states in the country. The same month Base decided to launch in Illinois, Governor Pritzker signed the Clean and Reliable Grid Affordability Act (CRGA, or SB 25). Base is so key to the state’s energy plans that Governor Pritzker cited Base’s decision to launch in Illinois as evidence that the Act was working.

Tactically, CRGA establishes a 3 GW battery storage target by 2030 and gives residential batteries access to the PJM wholesale capacity market.

PJM is the Regional Transmission Organization (RTO) that covers parts of Illinois, including Chicago, and the Mid Atlantic. PJM’s high capacity prices make stored power particularly valuable, while Illinois’ aging infrastructure and time-varying distribution rates create the opportunity to get paid for relieving stress on the local wires, a value stream Base cannot currently monetize in deregulated Texas.

Illinois is also well-positioned as a potential … base … of expansion to the northeast and midwest.

These are all specific reasons to choose Illinois once Base decided it was time to expand, but not reasons to expand in the first place.

I asked Justin how they decided the time was right to make the Lone Star State less lonely.

His answer was that the process was “perhaps too scientific,” but basically, they look at the penetration rate of eligible single-family homes in a given metro, and how expensive it would be to increase the slope of expansion within that metro. It gets expensive at a certain point, near where they’re at in Texas.

“So we think the economic optimum is to expose ourselves to more TAM over time,” or in English, to enter new markets.

Less scientifically, while everything is bigger in Texas, Base’s ambitions are even bigger than that. The only way to prove that it can expand outside of its home state is to expand.

Base can take the capabilities it has built in Texas, and turn them on the next best market, where it can, as with the battery itself, see what breaks in the new context, fix those things, and apply them to the next market, and the next one, so that it iterates on go-to-market and deployment faster than everybody else, too.

All of which means tweaking and testing.

On the Growth side, Cole said that a lot of the early marketing was very Texas-focused, but, “We’re shifting to becoming America’s Power Company, which requires you to kind of localize your approach.”

“So if you look at our Illinois website,” he continued, “it has a classic Midwestern brick home with a kid playing baseball in the backyard. The approach and how we think about photography, how we think about design is expanding.”

“What we learned in Texas is that you need to earn trust and you need to humanize energy,” he said, “but the way you humanize it is obviously going to change region-by-region. You’re seeing this with the Illinois launch, and you’ll see this with a few other launches we’re cooking up over the next few months, as well.”

The value prop people care about in each region differs slightly, too.

Base’s core offering, everywhere, is “affordable, reliable power,” but, Cole said, “some will veer more affordable, for example Chicago, and others will veer more reliable, for example Texas.”

This makes sense - Texas has been rocked by outages, and PJM wholesale power has gotten dramatically more expensive recently. ComEd, the utility that covers Chicago, has seen its all-in residential rate jump to 17.07¢/kWh, up roughly 90% from ~9¢/kWh in 2021. Brian Potter talks about why PJM and others have gotten relatively more expensive compared to ERCOT here.

What is true across all of Base’s customers is that they would like to pay less money to receive more reliable power.

Another thing that seems to be true among Base’s customers is that they trust local news. “Our target audience watches the local news at 5pm, and so getting Zach on TV with a reputable local TV station is a good strategy,” Cole said.

Local newspapers work, too.

Base launched in Chicagoland, the greater Chicago metro area, on June 24th, on the heels of “recent severe storms which left more than 400,000 without power.” It signed up more than 1,000 members and did 26 installs almost immediately.

Then on July 27th, storms knocked out power to more than 530,000 ComEd customers. The next morning, the Chicago Tribune ran a favorable feature on Base: Base Power launches backyard battery backup system for ComEd customers.

That day, 1,437 leads came in in Chicagoland, five times the previous best day, and 172 of them signed that day, nearly double the previous daily record of 90 and almost a quarter of the new Chicagoland members signed up in all of June.

In Illinois, too, it seems that Base will have more demand than it can meet, which means it falls to Dana to meet it.

Deploying a New Market

It should be easier to launch a second market than it was to launch the first. Base knows what it’s doing now. It knows how to talk to potential members and how to walk them from top of funnel through signup with BaseOS. It knows how to make batteries, how to get them to warehouses, and how to get them from warehouse to house. It knows how to install them, now with less electrician time. It knows the type of people it needs to hire, and roughly how many. In many ways, a new market is just doing those things in a new place.

To be sure, there are differences, too, new problems to be solved, or old problems, already solved in Texas, to be solved anew.

Take permitting. In Chapter 2, we talked about the fact that each Authority-Having Jurisdiction (AHJ) in Texas could supersede the National Electric Code (NEC) with its own rules, which meant that each new neighborhood was a new adventure, and whether Base was able to install smoothly or not often came down to the whims of each particular inspector.

It was Kafkaesque and annoying, but Dana and her team figured it out case-by-case, and it was another source of advantage “if you think about Base as a machine that eats complexity and turns it into moats.”

Then, Base ate complexity and turned it into law. In August 2025, thanks in large part to work by Base’s policy team, Texas passed SB1252, a standardized code framework that said, “a municipality may not adopt or enforce an amendment to the National Electrical Code, or adopt or enforce an ordinance, rule, or other measure, that would regulate the installation or inspection of a residential energy backup system.” The framework said that the rules had to be the same throughout the state.

That made Deployments’ job much more predictable, if not easy, and it was one of a million little things they figured out to make Texas go more smoothly.

But Illinois is not Texas, and its municipalities are not subject to SB1252. Chicagoland has more than 500 permitting jurisdictions. Base is back to negotiating with individual inspectors in each about small differences. There are different labor and union dynamics there, too.

In each state Base enters, it will have to push for policies that make permitting easier to navigate, and that generally support affordable, reliable energy, unless it can make the National Electric Code… national.

“As of 2025, Texas is the only state where the state itself controls permitting, because of SB1252,” Justin told me. “We’re trying to do that elsewhere, but there’s a federal permitting bill very slowly making its way through Congress, and we’re trying to get language on that bill that is basically supportive of a similar path to what we did here in Texas.”

Justin is very serious about this, because it slows Base down, hurts the grid, and hurts Base’s customers. Plus, it’s an unnecessary headache when there are so many other hard things to worry about. When we saw each other at Reindustrialize in June, we were talking Illinois permitting within 60 seconds. Last week, he wrote an op-ed for Utility Dive titled America’s permitting crisis lives at City Hall making the case that local permitting discrepancies are slowing the transition to clean, cheap, reliable energy and calling for federal or state-level action to standardize the process.

To that end, in November 2025, Base hired Zina Bash as Chief Legal Officer. Bash is a huge get and a perfect fit, having clerked for Justice Samuel Alito, served as a Special Assistant to the President for Regulatory Reform, and as Senior Counsel to the Texas Attorney General before going into private practice.

Bash’s hiring is a sign of how important the company expects policy to be going forward, and another example of figuring something out at the edge of the web (SB1252) and turning it into a core capability.

There are a lot of new things to figure out in Illinois, but this time, Base is doing it with more resources, a more experienced team, and stronger brand awareness.

“When we were trying to hire people in Dallas with no warehouse and no footprint, people were like ‘You’re a scam for sure, like this is not a real company,’” Dana recalled. That made hiring tough. Now, people know that it’s a real company. They may have read about it in the Chicago Tribune, the WSJ, or Not Boring, which seems to be making hiring easier, even in new markets.

A couple of weeks ago, Dana flew into Chicago on one of her frequent trips up to launch the market. Her plane diverted to Indianapolis to refuel, and by the time she got to Chicago, she’d missed the rental car window and they canceled her car. So Dana goes to the rental car window and tells them, politely, that it’s not like she chose for her flight to be delayed…

So we’re chatting with the guy, and I’m like, ‘It’s midnight, I really just need you to give me a car. If you tell me I need to take an Uber an hour to my hotel I’m gonna be very sad.’ And he was like, ‘Okay, I don’t like that, let me see what we can do.’

And then he asked, ‘OK, what’s your email?’ and I told him, ‘Oh it’s whatever at basepowercompany.com’ and the other guy goes, ‘Oh my gosh, I just applied to get a battery, they’re reviewing my site survey!” and I’m like, ‘No way!’ and the other guy says, ‘I install solar when I’m not doing this,’ and I was like, ‘Do you want to work at Base?’ I gave him the recruiter’s email, and he emailed and he’s applying to be an installer.

It was pretty electric and it made my day.