Welcome to the 1,205 newly Not Boring people who have joined us over the past week! If you haven’t subscribed, join 194,036 smart, curious folks by subscribing here:
Today’s Not Boring is brought to you by… Secureframe
Prospects are asking about security earlier in the buying cycle. Companies need a way to get compliant that is faster and takes less employee involvement. With Secureframe you can get compliant in weeks—not months—so you can focus on growing your business.
Secureframe makes it fast and easy to get SOC 2, ISO 27001, HIPAA, PCI & GDPR compliant. With guided workflows and one-click integrations, the entire process is automated and backed by a support team of industry experts.
Get compliant fast with Secureframe.
Click here to set up a demo. Mention “Not Boring” during your demo to get 20% off your first year of Secureframe.
Hi friends 👋,
Happy Monday! I had a lot of fun writing this one.
While I’ve written about AI, I haven’t written much about OpenAI because there’s already so much attention on the company that it’s been hard to find much new to say. GPT-4 is amazing. There will be implications for white collar workers. Here are 75 things you didn’t know you could do with ChatGPT 🧵 👇 Are foundation models defensible? Are GPT Wrappers defensible?
Then, on Thursday, OpenAI introduced ChatGPT plugins, and the more I thought about it, the more excited I got. Through a series of seeming accidents, great product, and user pull, OpenAI has attracted attention, turned that attention into the must-use API, and with the release of plugins, is turning that attention into a platform unlike any platform that exists, one that ingests the capabilities of the apps built on it.
If OpenAI chooses to, it can build an Apex Aggregator and define the new best business model on the internet. We need to lay some groundwork to explain why.
Let’s get to it.
Attention is All You Need
The paper that kicked off the AI Revolution had a catchy title, as these papers go: Attention is All You Need.
Written by a team at Google Brain in 2017, the paper introduced the now-famous Transformer architecture that powers large language models such as OpenAI’s GPT-4.
As Chroma co-founder Anton Troynikov explained it to me in our first episode of Anton Teaches Packy AI, Transformers can do a lot just by paying attention to the right parts of the input. They don't need other types of neural network layers, like the ones used for convolutions or recurrent connections, to perform well. Attention is all they need.
That same year, in a follow-up to his canonical Aggregation Theory, Ben Thompson wrote Defining Aggregators, to catalog businesses that capture value by controlling demand for abundant resources. Aggregators have three defining characteristics:
Direct Relationships with Users
Zero Marginal Cost for Serving Users
Demand-driven Multi-sided Networks with Decreasing Acquisition Costs
Think Google and Facebook, the Super-Aggregators. Billions of users come to their sites, where they’re served content created by other people for free, and it gets cheaper to acquire users over time as demand attracts supply which attracts more demand which attracts more supply, and so on. Better experience leads to more users which leads to more suppliers and advertisers which leads to more users, and so on.
The takeaway from that post was essentially the same as the Transformer paper: to win big on the internet, attention is all you need.
OpenAI began its life in December 2015 as a non-profit. With over $1 billion committed by Elon Musk, Peter Thiel, Sam Altman, Jessica Livingston, Reid Hoffman, Ilya Sutskever, and others, it had no plans to make money. It didn’t need to support itself.
One prominent VC who looked at investing early on said he loved the idea and the team, but that with no revenue and no plans to make any, he couldn’t justify investing his LPs’ money in the startup. As recently as last summer, when DALL•E 2 brought OpenAI into the limelight, a team member told me that while their tech was excellent, they weren’t quite sure how they were going to monetize it.
Then, in December, OpenAI released ChatGPT. The team wasn’t expecting much – it was a “research preview” – but practically overnight, it caught fire. Its December release was one of those “I remember where I was when” moments for those of us who love this stuff (I was in a hotel lobby in LA and tweeted this after my first experience).
ChatGPT exploded. Two months in, it reached 100 million users. Nearly half of the YC W23 class is building apps using OpenAI’s APIs. GPT-4 became the most eagerly anticipated drop since the last Beatles album, and it delivered. Then last week, OpenAI released ChatGPT plugins. They downplayed their foresight on this one too, explaining the launch by saying: “Users have been asking for plugins since we launched ChatGPT.”
Since the launch on Thursday, a bunch of people have compared plugins to Apple’s App Store, which paid developers $60 billion in 2022 (and likely took $15 billion for itself). I think it has the potential to be much bigger than that, and I really don’t want to be yet another person who overhypes OpenAI, so I’ll spend this entire piece walking you through why I think so.
In under a year, OpenAI has gone from under-the-radar and unsure-how-to-monetize to one of the most potentially dominant strategic positions and business models in the history of tech. It happened almost by accident. While the technology deserves the praise it receives, its translation into a juggernaut business was far from certain until tens of millions of people rushed to use its new consumer-friendly interface. Since then, the dominos have fallen rapidly.
When it comes to turning mind-blowing technology into an equally mind-blowing business opportunity, it turns out, you guessed it: attention is all you need.
As it stands today, if OpenAI chooses to, it can build the Apex Aggregator by building an Action Engine. The Action Engine subsumes search and any number of products that let users do things, and it does them for users with nothing more than a simple prompt. If Aggregators controlled demand and commoditized supply, the Apex Aggregator can control demand on multiple fronts, turn attention into actions, disaggregate any supplier that feeds it, and even aggregate the aggregators. It could make Apple, Google, and Facebook’s models seem soft and cuddly… until they fight back.
I’ll explain what’s gotten me so excited about OpenAI’s accidental business model, but first, we need to do a quick refresher on the current tech business model kings.
If you’re reading this, I assume you know a lot about how Google and Facebook’s business models – the greatest cash-generating machines known to humanity – work, so I’ll keep this section very brief.
Facebook (and its family of apps) use billions of users to attract billions of users, collect a lot of data on those users, and let advertisers target those users based on that data. If you’re a company that wants to get in front of a lot of a certain type of buyer because you think that type of buyer might like your product, Facebook is the place to go. Then, you just serve them a colorful ad, a certain number of them will click on it, and a certain number of those will buy. Because Facebook has the most social attention – over 3 billion users – it attracts more people, and because it has the most people and the most data on them, it attracts the most brand advertising dollars. Both demand and supply cost it practically nothing to acquire, and ads cost it practically nothing to serve.
Google’s model is even more direct. Billions of people go to Google explicitly to tell the site exactly what they’re looking for – from information to hotels in certain cities to accounting software – and Google auctions off well-placed links to advertisers who offer exactly what those people are searching for. Those people have expressed intent, and Google rakes in cash from advertisers intent on capturing it. Because Google captures the most search attention – 93% of searches worldwide go through Google – it attracts the most advertising dollars of any business on earth. Both demand and supply cost it practically nothing to acquire, and ads cost it practically nothing to serve.
In 2022, Facebook/Meta did $116.6 billion in revenue and Google/Alphabet did $282.8 billion in revenue, the vast majority of which, in both cases, came from serving ads. It’s hard to think of better internet business models. But I think OpenAI just did by accident.
Attention with Intelligence
OpenAI’s potential business model is hard to comp directly because it’s so many things at once: an aggregator, an API, and a platform.
Imagine that a bunch of apps used Google Search as their core feature, Google was the main place to find those apps, and those apps worked better when you used them in concert with each other through Google, and you’re getting close to what OpenAI might be able to do.
OpenAI hits its partners from all angles:
Attention Grabber. ChatGPT is the front-end for OpenAI, the “aggregator” to which over 100 million users flocked within the first two months. ChatGPT will let OpenAI control demand for abundant resources, which are themselves tailored to OpenAI.
Intelligence API. A number of new products are being built around OpenAI’s API and a bunch of incumbent products are re-architecting themselves around OpenAI’s API. Unlike most APIs, which serve non-core functions, the OpenAI Intelligence API is core to many of these product experiences.
Plugin Play. Products, whether built on OpenAI’s APIs or not, can now be plugged into ChatGPT itself (and potentially into OpenAI’s Intelligence API over time), turning the product into both platform and aggregator, and bringing more attention to ChatGPT. Unlike an App Store, which just lists apps but doesn’t ingest its capabilities, this essentially turns any plugin into an API that gives ChatGPT its company’s full capabilities.
By attracting so much user attention with a genuinely excellent product, OpenAI has also attracted developers to build with its Intelligence API and on its plugin platform. In the short-term, this is great for developers – they get more capabilities and more users – but it’s a sort of Faustian bargain. Developers building on the Intelligence API are relying on OpenAI for their core functionality, and plugin developers are handing their core functionality to OpenAI. In both cases, OpenAI gains more attention, reach, and power.
But the bullets above are based on what’s been announced to date. I’ll save speculation for later. For now, let’s go a little deeper into each leg of the stool.
Attention is all you need. And ChatGPT is an attention grabber.
Just two months after its launch, it reached 100 million users, the fastest ever to reach that mark by an order of magnitude. While OpenAI hasn’t updated that number, and doesn’t release daily active user numbers, it’s not hard to imagine that in a year, hundreds of millions of people will start their internetting with ChatGPT each and every day.
In Defining Aggregators, Thompson wrote, “The key characteristic of Aggregators is that they own the user relationship.” If people go to your site to find anything on the internet, you can tell them where to go. If people open your app to ride, you can tell them whose car to get in. Owning the customer relationship lets aggregators commoditize their suppliers and capture more value from each transaction.
ChatGPT (and GPT more broadly) takes this a step further and completely disintermediates its “suppliers.” Google can show you any website its algorithm wants, and it can summarize the website in a prominent box up top, but its job is still to send people to websites. ChatGPT, on the other hand, reads most of the available content on the internet and spits out complete answers to any prompt right in the chat. It’s not just a starting point, but a destination for an increasing percentage of the things a user might want to do online.
When ChatGPT came out, there was a fair amount of harrumphing from AI gray hairs that it wasn’t really a technical advance, just a shiny interface on an old model. That missed the point on the product side – interfaces are crucial – but it especially missed the point on the business model side. ChatGPT’s interface was an attention grabber, and once you have attention, you can direct it in any way you choose.
ChatGPT didn’t just capture consumer attention, it captured developers’ imaginations.
While GPT-3 was already available via API, and while some developers were already building with it, the buzz around ChatGPT pushed seemingly every developer to think about how to incorporate GPT into their products. Something like half of the current YC batch is building on OpenAI’s APIs.
These companies are derisively called “GPT wrappers,” and many have wondered about their viability as venture-scale businesses. What worries me about these businesses is how much they’re handing over to an API.
In APIs All the Way Down, I pointed out that there seems to be a sweet spot for API-first companies:
Strong API-first businesses sit in this sweet spot: they provide mission critical but non-core functionality to their customers, like accepting payments, providing cloud security, or sending communications to customers.
Plugging in an API makes a ton of sense when APIs deliver cloud hosting or payments or text messages or any number of things that every company needs to do but that don’t provide a competitive advantage.
But what happens when APIs deliver intelligence?
OpenAI’s APIs are essentially Intelligence APIs. Write a few lines of code, and your product can do a bunch of things that a pretty smart human could, like answer customer service chats or discover drug formulations or tutor a student. Of course, most of the products that use OpenAI’s APIs do other things besides just delivering ChatGPT in a different interface, and they certainly involve some fine-tuning to better address the customer’s specific needs, but for most of these products, it really feels like they’re breaking the Cardinal Rule of APIs: don’t outsource the thing that makes your beer taste better.
But what’s a company to do? Spend hundreds of millions to train their own models in the hopes of producing something different enough to matter? Not make their products intelligent? This is the Chinese Finger Trap that’s captured so many companies before. As I wrote in Shopify and the Hard Thing About Easy Things:
Here’s the hard thing about easy things: if everyone can do something, there’s no advantage to doing it, but you still have to do it anyway just to keep up.
In just a couple of months, we’re already at a point at which there’s no real competitive advantage to using OpenAI’s Intelligence APIs, but there’s a disadvantage for not using them.
From OpenAI’s perspective, however, this is an incredible situation. Developers have to pay to use OpenAI’s APIs, and each company that writes those few lines of code to incorporate intelligence into their products makes it even more table stakes for the next one. Plus, it gives OpenAI more attention: now, there are armies of marketers pushing the value of OpenAI’s products on its behalf.
Until a week ago, the situation was tenable for GPT wrappers. Some people would be happy to just use ChatGPT and get an ~80% experience on that specific thing – get a good recommendation for a travel itinerary but go execute it themselves – while others would be happy to pay $10/month for an app that gave more tailored recommendations and booked the trip.
Then OpenAI announced ChatGPT plugins.
Where this really becomes an Ouroboros is when products built using OpenAI’s APIs plug in to ChatGPT itself.
On Thursday, while I was on a plane with messaging but no WiFi, my PBD chat with Ben Rollert and Dror Poleg lit up after Ben texted a link to a Twitter that I couldn’t open:
Ben: [twitter link I couldn’t open]
Ben: This is the new platform shift..
Ben: Last time I felt this excited was as a kid playing with internet w/ 28k modem
Packy: I’m on a flight with only messaging, what’s the Twitter link?
Dror: ChatGPT plug-ins from other apps
We weren’t the only ones excited. When my WiFi finally started working, I found Twitter lit up with awe and fear.
If you missed the announcement, OpenAI is introducing plugins directly into ChatGPT, including:
Browsing: “An experimental model that knows when and how to browse the internet.”
Code interpreter: “An experimental ChatGPT model that can use Python, handle uploads and downloads.”
Retrieval: “The open-source retrieval plugin enables ChatGPT to access personal or organizational information sources (with permission).”
Third-Party Plugins: “An experimental model that knows when and how to use plugins.”
If the Intelligence API lets companies bring OpenAI’s intelligence into their products, then plugins let companies turn themselves into APIs that feed OpenAI’s Intelligence. Any product, entire big and small companies’ products, essentially become APIs that OpenAI’s users can automatically hook into their personal Action Engine.
All of a sudden, with the right plugins, ChatGPT can do many of the things GPT Wrappers can. Thanos snapped.
Now, it’s very early. OpenAI is launching with just eleven pre-vetted partners:
With the big caveat that this is just an announcement, and there’s a long way between here and what I’m about to write, this is a massive upgrade for a few reasons.
First, it fills in gaps that GPT has had to date. Namely, it provides access to up-to-date information, making it an even more viable alternative to search for a number of queries, and the Wolfram integration gives ChatGPT “computational superpowers.” A few weeks ago, ChatGPT couldn’t really add; with Wolfram, it can do complex mathematical functions, run algorithms, function plotting, and even genealogy.
Second, as a bunch of people have pointed out, it’s OpenAI’s move into becoming a platform and building its own App Store.
But the platform and App Store analogies fall short, because the App Store isn’t a destination in and of itself, it’s just a place that people go to find apps, which they then use separately.
What OpenAI is building is another, third thing.
Most importantly, plugins will turn ChatGPT into one destination for nearly everything: search, discovery, travel planning, restaurant booking, gift shopping, first drafts, research, you name it. Importantly, unlike a traditional platform, through which users can find and download apps, ChatGPT consumes plugins and absorbs their capabilities into the main product. It’s a product as a platform, or a platform as one product.
With Browsing, it will do everything that search does, plus customization. With third-party plugins, it becomes a platform on which all of the suppliers add more functionality to the platform itself, and in the process, send their demand to OpenAI’s attention grabber.
Are you going to go to Kayak.com to book your flight, or ask ChatGPT to plan a trip and book you the best flights and let it deal with Kayak? Are you going to go to OpenTable.com for a dinner reservation, or ask ChatGPT to book you a table for 8 next Wednesday somewhere great and let it deal with OpenTable? I’d imagine that when people have something very specific in mind, they might still go to the legacy websites, but for most situations, the situations where we just have a rough idea of what we want, ChatGPT will work best. Even typing that last sentence, though, if I know exactly what I want, why not just tell ChatGPT to go do it for me?
Google has Google Flights and Google Shopping, which gets closer than just traditional search to recommending the best flights and products and cutting clicks, but there’s still clicking, browsing, and distraction. Plus, when people are already in ChatGPT anyway, it’s easy to switch from researching a presentation to asking it to book your travel to the conference. In the current version, booking travel through the Kayak plugin will be slightly easier, it’ll shave a couple clicks. But in the near future, booking could be as easy as sending a message if you trust ChatGPT enough.
More wildly, are you going to go directly to Zapier.com to painstakingly handcraft a bunch of Zaps to connect all of the different systems your business uses, or are you just going to tell ChatGPT, “Hey I want customer support tickets from Zendesk to create an Airtable entry and ping the @customer-care channel in Slack”?
Plugins turn manual actions that take a bunch of searches and clicks to complete into something as simple as having a conversation with a smart person who understands what you’re going for, even if you’re not entirely clear yourself. Removing clicks is the lifeblood of online commerce, and that’s an important piece of the plugin puzzle, but being able to flexibly type (and soon speak) any desire and have it come true is a game changer.
That whooshing sound you hear is the attention users paid to a number of different apps rushing into ChatGPT. If attention is all you need, this is a major step in capturing a ton of it.
Again, it’s early. A small group of users (I’m sadly not one of them yet – hook it up Sam!) has access to a small number of plugins. OpenAI hasn’t said how it’s planning on monetizing integrations, although it’s pretty easy to imagine.
If Google built a world historically amazing business by handing companies high-intent users, OpenAI could build something at least as good by turning that intent into action. It should be able to capture a fee on the transactions it drives, and in many cases, capture higher API fees on the other side as its Intelligence-API-using partners see more volume. You could even imagine that OpenAI could expose plugin capabilities in the Intelligence API, bringing plugins to all of the apps that build on top of it, turning it into a smart platform of platforms.
That would be good for plugin partners – being built into more products would increase demand – and good for OpenAI – more users will attract more plugin builders. Even if it means siphoning off some attention from ChatGPT, it may be worth it to grow the ecosystem early. Everybody wins.
All of that is if OpenAI decides to play nice with its partners.
If OpenAI optimizes for its ChatGPT users, though, it’s going to disintermediate a ton of businesses and force them into changing how they operate.
Plugging in the Apex Aggregator
No longer do distributors compete based upon exclusive supplier relationships, with consumers/users an afterthought. Instead, suppliers can be commoditized leaving consumers/users as a first order priority. By extension, this means that the most important factor determining success is the user experience: the best distributors/aggregators/market-makers win by providing the best experience, which earns them the most consumers/users, which attracts the most suppliers, which enhances the user experience in a virtuous cycle.
- Ben Thompson, Aggregation Theory
Earlier, we touched on Google and Facebook’s models. They serve up ads that they think will be relevant to users, and users click on those ads and decide whether and what to buy.
The early glimpses at OpenAI plugins look similarish, but on steroids. The third-party plugins demo video showed an example Instacart transaction: a request for a vegan recipe transforms into a pre-loaded Instacart shopping cart, which the user can click on, explore, and purchase.
While the demo was incredibly impressive, and leaps and bounds beyond doing the same thing on Google, I can’t help but feel that it’s still in the skeuomorphic phase, or the “attract” phase on the partner side, to steal two ideas from Chris Dixon. It’s also still just learning – maybe I don’t like that recipe suggestion, or I do but I want to change a couple ingredients in the cart. In the current instantiation, I’m able to do that (and I probably always will be).
It’s not hard to imagine, though, that once there’s been enough RLHF (Reinforcement Learning from Human Feedback), ChatGPT will cut out a bunch of those clicks in order to optimize the user experience.
“Hey, I want a vegan meal on Sunday, something with tofu, and I’ll be home all morning.”
“Great, your ingredients will be delivered on Sunday between 10-11am. I’ll send you a recipe then.”
Maybe that order will go to Instacart. Maybe it will go to whichever product is able to deliver the best groceries for the cheapest price in the tightest delivery window. And “tightest delivery window” won’t just mean what each particular plugin says its delivery window is, but which plugin, over thousands then millions of orders, actually delivers closest to when it says it’s going to.
There’s this other kind of aggregator, not exactly like the Ben Thompson kind, that sits on top of a bunch of apps in a particular category and exposes data from all of them so that customers can make the best decision. If you watch TV, you’ve likely seen commercials for some famous ones, like Trivago and Priceline in the hotel category.
You’ve seen these commercials because the whole point of these businesses is to attract demand and then partition it out to whichever platforms offer the best prices or the best whatever the customer is looking for. They’re essentially marketing businesses, as reflected by the fact that Booking Holdings, the $94 billion parent of Priceline, spent $7.8 billion or 45% of revenue on Selling & Marketing expenses in 2022, twice as much as it spent on the G&A part of SG&A.
You see where this is going. I asked ChatGPT what the existence of aggregators like Booking Holdings has done to the profits of the other companies in the travel & hospitality value chain:
This is a pretty good summary of what’s coming for every industry that plugs into ChatGPT: increased competition, pressure on margins, shifts in consumer behavior, and a pressure to innovate on the things that matter in this new paradigm.
ChatGPT is the Apex Aggregator – it will aggregate the aggregators. And because it owns the customer relationship, because it has attention, it won’t have to spend the $7.8 billion per year that Booking Holdings has to!
Prospective travelers won’t go to Google and search “hotels in Mexico City,” intent that Booking Holdings must pay to acquire; they’ll go to ChatGPT and say, “I like up and coming neighborhoods, nice accommodations, authentic food, lively atmospheres, and first class flights as long as they’re reasonable. My budget is $8k, and I’m traveling with my husband. Can you plan us a 7-day trip in Mexico City and make all the bookings?”
Today, in the very early innings, some of that will happen through Kayak (a Booking Holdings company), and users will still need to click to book the travel themselves and have a Kayak account and enter their credit card information. In the beginning, this will be a boon for plugin partners – Kayak will be able to get the booking without having to pay Google, and it might be able to reduce its brand marketing spend, too. Just plugin, sit back, and let the bookings roll in.
But if OpenAI chooses to go full throttle – and I think they will for the customer experience benefits, even if they don’t want to kill intermediaries – Kayak will just be one of many potential paths. Once users don’t have to specify plugins, once they all come out of the box and ChatGPT searches among all of them for the product that best suits its users’ needs, Kayak will be crunched, in the same way that Kayak has crunched intermediaries in its own value chain.
But wait, there’s more. One of the reasons that aggregators work is that humans can’t parse a thousand different options themselves across all of the features that matter to them. So they go to Booking.com or Priceline or Kayak or whatever, find the flight that’s cheap enough in the right time window, and the hotel that’s 4.5 stars or above for less than $500 a night kinda near where they want to be, and they’re happy about it.
ChatGPT, on the other hand, can happily look through a million options and weigh all of the dimensions that might matter to its user. Instead of going through aggregators to reduce complexity, ChatGPT can let all of the world’s hotels and airlines and restaurants build their own plugins. Since those hotels and airlines and restaurants won’t have to pay fees to the aggregators when they go direct to ChatGPT, maybe they’ll be able to offer lower prices or better perks to people who book through ChatGPT. Maybe, when they have more confidence that they’ll be matched based on what they offer instead of simply the lowest price, they’ll focus on making their offering more differentiated so that they can appeal to the people looking for exactly a certain thing that they can offer.
Taken to its logical conclusion, ChatGPT plugins will provide infinite choice without the Paradox of Choice.
And it’s not just travel & hospitality. Ridesharing is a pretty obvious example. I don’t give a shit if I’m taking Uber or Lyft. “Get me to the airport by 7pm as cheaply as possible. But seriously, I can’t be late.” So ChatGPT will find the best price, and it will double-check that the service it chooses actually usually sticks to the time windows it says it will across all of ChatGPT’s users who’ve booked rides recently, and it will book that one. It might be Uber, it might be Lyft, it might be another local service that’s really excellent at getting people to the airport on time for a low price.
This dynamic will extend to software products, and ecommerce, and education, and and and.
OK, I’m getting myself all excited as I write this, so let’s pause and take a deep breath, because the implications are going to get even wilder.
Like what happens to brand as a moat? Rex Woodbury recently tweeted that “brand is a stronger moat than ever”:
He pointed out that “Today, marketing is often *all* that distinguishes a product,” and I agree with him. It’s one of the main conclusions I came to in that Shopify & the Hard Thing About Easy Things piece. The differences among so many products come down to a little feature here and there, which are often drowned out by marketing and brand, because humans are more drawn to brands than we are to caring about little feature differences.
This is What Clayton Christensen Got Wrong, according to Thompson:
That is the problem: Consumers don’t buy aircraft, software, or medical devices. Businesses do.
Christensen’s theory is based on examples drawn from buying decisions made by businesses, not consumers.
The reason this matters is that the theory of low-end disruption presumes:
Buyers are rational
Every attribute that matters can be documented and measured
Modular providers can become “good enough” on all the attributes that matter to the buyers
All three of the assumptions fail in the consumer market, and this, ultimately, is why Christensen’s theory fails as well.
But hold up! With ChatGPT:
Buyers can be rational, or can let ChatGPT be rational on their behalf
There’s more incentive than ever for suppliers to document and measure every attribute that matters and plug them into ChatGPT
Modular providers will have the incentive to be “good enough” on the attributes that matter to particular buyers, knowing that they can be matched to those buyers.
In this world, I think I actually come to the opposite conclusion from Rex and Ben: Brand matters less than before. Actual differentiation matters more.
I’ve written a few times recently that we’re going to see an explosion in the number of Small Apps created thanks to the abundance of APIs and products like Replit that make it easier than ever for people to create software products. I missed the biggest catalyst. ChatGPT both makes it easier for people to build Small Apps – it can create a barebones website from a sketch, or walk users through how to code up sites – and makes it easier for those Small Apps to find customers as long as they differentiate hard enough on certain attributes.
This is most clear in software. Developers can build products that are really excellent at one particular thing people want, plugin to ChatGPT, and not worry about marketing the product or building a brand. They can charge less, capture more demand for the one thing they do really well, and let ChatGPT worry about orchestration with other products that each spike on one particular feature (maybe through the Zapier plugin, for now).
But as ChatGPT connects to the physical world, it might begin to reshape other types of businesses as well.
There’s this famous phenomenon where a lot of restaurants started looking the same – neon signs, green walls, attractive dishes – because of Instagram. In The Premium Mediocre Life of Maya Millennial, VGR captured it beautifully: “Premium mediocre is food that Instagrams better than it tastes.” If I’m right, ChatGPT plugins might have an even bigger impact in the opposite direction.
If enough people prompt, “Make me a reservation for whichever restaurant in New York has the best French Onion Soup,” there’s incentive for a restaurant to focus on making really fucking great French Onion Soup. If people prompt, “I’m going to Mexico City for 7 days, and I want to eat at one restaurant each night that’s known for being the best at a different type of local cuisine,” the restaurants that do that one thing really well win.
I can play this game for a lot of industries, online and offline, but I think you get the point, and I need to save something for future posts.
The takeaway as I see it is that, again if OpenAI chooses to go this route, a lot of industries and value chains are going to be shaken up, and the winners will be the companies that focus on differentiation, on doing a specific thing really, really well, as opposed to those who do a lot of things pretty well and pump money into brand.
Of course, this won’t be true in every category. Luxury fashion will still be valuable as a status signaling mechanism. Software products that organize complex workflows for companies will still retain a lot of customers and may find a lot of new ones through ChatGPT. I think Zapier is in a much stronger position than Kayak, for example, because it manages a ton of complexity under the hood and ChatGPT gives it a smoother, more intuitive interface.
I know that I’m getting overexcited and that the gap between the theory as written here and the practical reality is wide and will take time to work through. The impact may not be as huge and immediate as I’m making it sound.
And of course, incumbents will try to fight back. No one likes being disintermediated. But, like, what do you do? Not plugin to ChatGPT? Not mold your product to better respond to prompts?
At the rate it’s growing, ChatGPT may quickly become the firehose for demand. The way websites have had to contort themselves to please Google’s SEO, brands won’t have a choice but to shoehorn their products to fit ChatGPT’s recommendations.
Some will resist. They’ll offer discounts if you go direct through the app, the way restaurants offer discounts if you call them instead of ordering through Seamless or DoorDash. They’ll choose not to use OpenAI’s Intelligence APIs and build on another foundation model, or try to build their own.
And some of it might work! The fun thing about complex adaptive systems is that I’m only looking a couple moves ahead here, and the mass of business owners and developers will come up with unpredictable ways to evolve their businesses and products to stay a step ahead.
It seems practically guaranteed, though, that this will be a boon for consumers: lower prices for commodity products and more differentiated offerings, easily discovered. It will likely be deflationary, too, as companies compete to deliver the best product for the lowest price even more extremely than capitalism has forced them to up to this point.
And as I hope has become clear, OpenAI has stumbled into the best business model on the internet, if it chooses to pursue it.
It’s funny that the company that started its life as a non-profit has a path to becoming the world’s most profitable.
As the Apex Aggregator, it can capture massive amounts of attention, turn that attention directly into transactions for whichever plugins best suit its users’ needs, and capture a fee on every one of those transactions.
If it chooses to, it can disintermediate any number of suppliers, push prices lower, and push differentiation and quality up. It can steal share from search and from aggregators alike, and no one but those companies’ shareholders would shed a tear. It can learn what users like, and even put out RFPs (in a sense) for builders to fill in gaps, often using OpenAI’s products to do so.
The question is: will it?
The answer is: I don’t think it has a choice.
OpenAI's mission is “to create and promote friendly AI for the betterment of humanity.” Aggregators, according to Thompson, win by “providing the best experience, which earns them the most consumers/users, which attracts the most suppliers, which enhances the user experience in a virtuous cycle.” Those ideas are aligned. The best experience for users is the one that gets them exactly what they want with as little friction as possible. If it comes down to a choice between plugin suppliers and users, I think OpenAI will side with the users.
Other factors will come into play and new questions will emerge:
Will OpenAI choose to take transaction fees from plugin suppliers, or will it simply charge users $20/month for ChatGPT Plus?
How will users set up accounts and pay for everything without clicking through and purchasing at each plugins’ main site? (Sam Altman did start another company that’s building “the world’s largest identity and financial network as a public utility”... )
Which companies won’t be ripe for ChatGPT’s disintermediation? Certainly, there will be many companies that will thrive independently, unbothered, moisturized, in their lanes. What characteristics make up those companies?
What about the regulators? If what I’ve written is half true, it will make OpenAI the most powerful company the world has ever seen. Is that a good thing or is it dangerous, even if that power is put to work on behalf of users?
Regulators aside, is it good for one or even a few companies to have even more powerful business models than they do today? Ezra Klein wrote a good piece exploring that issue, pre-plugins.
Things are always messier in reality than they are on paper. This is just my attempt to play out what’s happening today into the future, and I’ve gotten predictions very wrong before!
And what about competition from all of the other foundation models, and from dominant companies like Google, Apple, Meta, Amazon, and Microsoft (half-partner, half-competitor)?
That last question is particularly interesting, and I think it’s a matter of speed and willingness to adapt. Google, Apple, Meta, and Microsoft all dwarf ChatGPT’s user bases. They, too, control attention. They all have strong AI/ML teams, and LLMs (in-house or partnered), and vectors of attack (Google: search bar, Apple: iPhone/Siri, Meta: more personal data and social, Amazon: Alexa and recommendations, Microsoft: enterprise strength and the new Bing).
Can ChatGPT become the default Action Engine before its bigger competitors can adapt? Can it build an ecosystem of plugins that makes its Intelligence more intelligent before the others alter their models to leverage their strengths and relationships in head-to-head battle? As in the Third Browser Wars, the giants’ advantages can also be their Achilles Heels. All that’s certain is that with trillions of dollars at stake, and no one’s going to go down without a fight.
If this is the beginning of the next big platform shift, and I’m more convinced by that now than I was even when I began writing this piece, there are going to be all sorts of questions to answer, competitive moves to be made, and surprising twists and turns along the way.
All that said, I do think that OpenAI is in the driver’s seat and has the potential to build a very big business if it goes all the way on adopting the newly-available best business model.
I don’t, for the record, think there was anything devious about its initial messaging or its change of tack since. I really do think that the emergence of this potential business model was a happy accident, and that the company’s intentions are good. OpenAI hasn’t indicated that it’s even going to take a cut of any of the transactions that go through its interface, or that it will ever abstract away the companies behind the plugins. Sam Altman supposedly doesn’t even own any shares in the company!
But on the road to building AGI, OpenAI caught attention in a bottle, and everything that happens in the coming months and years will flow from there.
Now, perhaps more than ever, attention is all you need.
Thanks to Dan for editing, Anton for teaching me about Attention is All You Need, and Ben Thompson for the insights that keep on giving (and congrats on 10 years!)!
That’s all for today. If you found today’s post useful, share it with a friend or two. If you didn’t, thanks for making it all the way down here anyway :) We’ll be back in your inbox on Friday with another Weekly Dose of Optimism.
Thanks for reading,
I'd be really interested to learn about the implications for finance and investing? ChatGPT plug in seems very similar/analogous to "smart order routing", where Wall St. traders are able to leverage AI to find out where the best trading venue and supply is for their particular trade. In other words, it matches buyers and sellers, cutting out the middle man (broker).
There seems to be a lot of use cases for ChatGPT in finance..."Find me the best performing large cap growth mutual fund over the last 15 years"..."Put me in contact with three mortgage lenders offering the lowest rate currently"..."How many treasury futures do I need to buy to close my duration gap" (SVB could have used this).
Would love to hear your ideas on potential implications for finance/investing/wealth management. Thanks Packy!
Great read. But the idea of an Apex Aggregator is frightening. Agree that inserting themselves into money flow is a path but centralized control soon follows when someone inserts themselves in the money flow (ex, Apple's 30%?). I think OpenAI is on the path of running the Web2 growth playbook...shit, YC practically wrote it.