This was a joy to read, and the compiler-versus-runtime distinction is the cleanest cut through the token debate anyone has made. Thank you both for writing it.
One piece of history makes the case even stronger. We tried turning businesses into rules once before. The expert systems of the 1980s were exactly this dream, capture the tacit knowledge in people's heads and turn it into something a machine runs deterministically. They died of two diseases. Extraction was brutally expensive, teams of knowledge engineers interviewing experts for months. And the rules rotted, because when the world changed, updating them was manual and slow. The field even coined a name for the first problem, the knowledge acquisition bottleneck. What's described here is that old dream with both fatal bugs fixed. AI collapses the cost of extraction, and the regenerate-and-retest loop cures the rot. Poetic reads less like a new idea and more like a forty-year-old correct idea whose missing piece finally arrived, which is usually what the best companies turn out to be.
There's also a strange macro consequence hiding in the thesis. If code does the doing and tokens only burn when the world changes, token demand stops tracking how much work the economy contains and starts tracking how fast the world changes. The labs end up long volatility. Quiet years starve them, chaotic ones feed them. That's a very different business from the one the current buildout is priced for.
So, this absolutely works, and I know this because this is what I do now. I have stumbled upon the same process by answering the very same or similar questions. What is the function we are trying to replicate? What documented process do we have? You spend 80% of your time talking to the people that do the thing, because the world is messier than the process implies. Code could never be adaptive enough previously. I embed AI at decision points of the process to help narrow what I call the probability space of what it is reasoning over. These tasks are cheap. They are specific. They compound.
The thing that I have seen already, being a non-coder before AI, is you target a manual process, and everything thereafter not only becomes better, it becomes easier, because what was a manual process with humans with tacit knowledge becomes a digital output packet that you can do absolutely anything with. It is the most fun you can have because it actually solves real world problems for real people.
I don’t get why more ppl don’t see this, Markie is 100% right here.
I recently exited my last company in the US, and now building in Italy, one of the biggest gaps is obviously automation. We just automate way less here.
But AI often isn't required to do the work. Once the logic is clear, most repeatable decisions should just be rules/code.
A lot of ppl are throwing tokens at stuff cuz it’s easier than doing the annoying part (turning the process into something that runs programmatically)..
Will it work most times? yeah. But it gets expensive fast and it’s never going to be 100% reliable
Thank you for writing this! It was an enjoyable read, and adds some nuance to the token spend / AI's ROT discussion. I'm working on a similar post too, but looking to calculate what it'd cost a business to use LLMs (albeit with some stylised math).
After some editing + verification of the math, I've finally uploaded my article (it's a very long read) titled 'What is the Return on Tokens?' Let me know what you think
I am motivated to write about this since LinkedIn connection Metis Analytics (St. Louis) appears to have just released their new product and I am reading a flood of posts about it but world models seem to be designed to handle this very problem with agents. I did not understand how agents did anything if there was no person coding what they were supposed to do but assumed that they did their best coding approximation of the natural language prompt.
The highest ROT case study found so far is not at a US startup. Doubao (ByteDance) hit 120 trillion tokens processed per day as of May 2026 -- that is the runtime Markie says nobody built yet. The capital structure is why: state-adjacent funding removed the fund life pressure that kills long-burn token experiments. ByteDance could run the find-what-humans-actually-pay-for loop until they found it. The hardware implication: Doubao at that scale drives Huawei Ascend procurement decisions that are already made. Compiler and runtime both exist in China -- I cover the supply chain daily at China AI Dispatch.
There is a dangerous illusion shaping the digital future. the belief that our data infrastructure is secure, resilient, and future-proof.
It is not.
At the center of this illusion lies tokenization a system designed to protect sensitive data, but one that is now becoming a hidden point of fragility at global scale.
Tokenization was built on a simple promise: separate sensitive data from its usage, reduce exposure, and contain risk.
But in reality, it has created a new concentration of power and vulnerability.
The token vault the core of this system is now one of the most critical and sensitive assets in the digital economy. If it fails, is compromised, or becomes inaccessible, entire systems can collapse. Financial networks, healthcare systems, identity infrastructures all depend on mappings that, if disrupted, can render data unusable or exposed.
This is not theoretical.
It is a structural risk.
At the same time, the rapid expansion of AI, cloud computing, and real-time data systems is pushing tokenization beyond its original design limits. The need for constant access, processing, and scale is forcing repeated detokenization increasing exposure, latency, and attack surfaces.
We are building faster systems on top of fragile foundations.
And the consequences of failure are no longer local . they are systemic.
Write on Medium
If tokenization architectures are improperly dismantled, migrated, or compromised without robust safeguards, the result could be catastrophic:
Permanent data loss
Large-scale exposure of sensitive information
Collapse of critical digital services
Loss of public trust in digital systems
This is no longer just a technical issue.
It is a matter of global stability.
The path forward demands urgent attention:
First, acknowledge that tokenization is not a complete solution. it is one layer in a broader security architecture.
Second, invest in resilient alternatives and complements. strong encryption, decentralized architectures, zero-trust systems, and secure computation models.
Third, establish global standards for the safe migration, decommissioning, and destruction of tokenized systems ensuring that data remains protected even as architectures evolve.
Finally, recognize that security decisions made today will define the resilience of tomorrow’s digital world.
The question is not whether our systems will be tested.
It is whether they will survive when they are.
The responsibility lies with you.
Act before the system is forced to reveal its weaknesses.
The ratio looks clean until you ask what got removed to make it work. Lower token spend is good if the workflow improved. It is dangerous if you just cut out judgment and called it efficiency.
This was a joy to read, and the compiler-versus-runtime distinction is the cleanest cut through the token debate anyone has made. Thank you both for writing it.
One piece of history makes the case even stronger. We tried turning businesses into rules once before. The expert systems of the 1980s were exactly this dream, capture the tacit knowledge in people's heads and turn it into something a machine runs deterministically. They died of two diseases. Extraction was brutally expensive, teams of knowledge engineers interviewing experts for months. And the rules rotted, because when the world changed, updating them was manual and slow. The field even coined a name for the first problem, the knowledge acquisition bottleneck. What's described here is that old dream with both fatal bugs fixed. AI collapses the cost of extraction, and the regenerate-and-retest loop cures the rot. Poetic reads less like a new idea and more like a forty-year-old correct idea whose missing piece finally arrived, which is usually what the best companies turn out to be.
There's also a strange macro consequence hiding in the thesis. If code does the doing and tokens only burn when the world changes, token demand stops tracking how much work the economy contains and starts tracking how fast the world changes. The labs end up long volatility. Quiet years starve them, chaotic ones feed them. That's a very different business from the one the current buildout is priced for.
Yes! Excellent point
"token demand stops tracking how much work the economy contains and starts tracking how fast the world changes" faxxx
So, this absolutely works, and I know this because this is what I do now. I have stumbled upon the same process by answering the very same or similar questions. What is the function we are trying to replicate? What documented process do we have? You spend 80% of your time talking to the people that do the thing, because the world is messier than the process implies. Code could never be adaptive enough previously. I embed AI at decision points of the process to help narrow what I call the probability space of what it is reasoning over. These tasks are cheap. They are specific. They compound.
The thing that I have seen already, being a non-coder before AI, is you target a manual process, and everything thereafter not only becomes better, it becomes easier, because what was a manual process with humans with tacit knowledge becomes a digital output packet that you can do absolutely anything with. It is the most fun you can have because it actually solves real world problems for real people.
This is very cool to hear! Once you see it, hard to imagine more won’t move this way.
I don’t get why more ppl don’t see this, Markie is 100% right here.
I recently exited my last company in the US, and now building in Italy, one of the biggest gaps is obviously automation. We just automate way less here.
But AI often isn't required to do the work. Once the logic is clear, most repeatable decisions should just be rules/code.
A lot of ppl are throwing tokens at stuff cuz it’s easier than doing the annoying part (turning the process into something that runs programmatically)..
Will it work most times? yeah. But it gets expensive fast and it’s never going to be 100% reliable
Thank you for writing this! It was an enjoyable read, and adds some nuance to the token spend / AI's ROT discussion. I'm working on a similar post too, but looking to calculate what it'd cost a business to use LLMs (albeit with some stylised math).
Glad you enjoyed Les! Looking forward to yours
After some editing + verification of the math, I've finally uploaded my article (it's a very long read) titled 'What is the Return on Tokens?' Let me know what you think
https://lesbarclays.substack.com/p/what-is-the-return-of-tokens?r=rq26d&utm_campaign=post&utm_medium=web
I am motivated to write about this since LinkedIn connection Metis Analytics (St. Louis) appears to have just released their new product and I am reading a flood of posts about it but world models seem to be designed to handle this very problem with agents. I did not understand how agents did anything if there was no person coding what they were supposed to do but assumed that they did their best coding approximation of the natural language prompt.
Agentic AI: Is The Emperor Naked? https://technbiz.blogspot.com/2026/06/agentic-ai-is-emperor-naked.html
The highest ROT case study found so far is not at a US startup. Doubao (ByteDance) hit 120 trillion tokens processed per day as of May 2026 -- that is the runtime Markie says nobody built yet. The capital structure is why: state-adjacent funding removed the fund life pressure that kills long-burn token experiments. ByteDance could run the find-what-humans-actually-pay-for loop until they found it. The hardware implication: Doubao at that scale drives Huawei Ascend procurement decisions that are already made. Compiler and runtime both exist in China -- I cover the supply chain daily at China AI Dispatch.
There is a dangerous illusion shaping the digital future. the belief that our data infrastructure is secure, resilient, and future-proof.
It is not.
At the center of this illusion lies tokenization a system designed to protect sensitive data, but one that is now becoming a hidden point of fragility at global scale.
Tokenization was built on a simple promise: separate sensitive data from its usage, reduce exposure, and contain risk.
But in reality, it has created a new concentration of power and vulnerability.
The token vault the core of this system is now one of the most critical and sensitive assets in the digital economy. If it fails, is compromised, or becomes inaccessible, entire systems can collapse. Financial networks, healthcare systems, identity infrastructures all depend on mappings that, if disrupted, can render data unusable or exposed.
This is not theoretical.
It is a structural risk.
At the same time, the rapid expansion of AI, cloud computing, and real-time data systems is pushing tokenization beyond its original design limits. The need for constant access, processing, and scale is forcing repeated detokenization increasing exposure, latency, and attack surfaces.
We are building faster systems on top of fragile foundations.
And the consequences of failure are no longer local . they are systemic.
Write on Medium
If tokenization architectures are improperly dismantled, migrated, or compromised without robust safeguards, the result could be catastrophic:
Permanent data loss
Large-scale exposure of sensitive information
Collapse of critical digital services
Loss of public trust in digital systems
This is no longer just a technical issue.
It is a matter of global stability.
The path forward demands urgent attention:
First, acknowledge that tokenization is not a complete solution. it is one layer in a broader security architecture.
Second, invest in resilient alternatives and complements. strong encryption, decentralized architectures, zero-trust systems, and secure computation models.
Third, establish global standards for the safe migration, decommissioning, and destruction of tokenized systems ensuring that data remains protected even as architectures evolve.
Finally, recognize that security decisions made today will define the resilience of tomorrow’s digital world.
The question is not whether our systems will be tested.
It is whether they will survive when they are.
The responsibility lies with you.
Act before the system is forced to reveal its weaknesses.
ROT has the same trap as revenue per employee.
The ratio looks clean until you ask what got removed to make it work. Lower token spend is good if the workflow improved. It is dangerous if you just cut out judgment and called it efficiency.
Which, if any, of the current "business into software" solutions are promising?