Encoded Founder — Chapter VI

The Compound

Where it lives, what it produces, and what we cannot predict

In the spring of 2025, a fraud detection system at Stripe flagged two hundred near-identical payment requests. Same low-entropy user agent. Rotating across proxies. Arriving roughly every forty seconds. Individually, no single feature triggered an alarm. The card numbers were valid. The amounts were plausible. The IP addresses were distributed. By every traditional metric, these were two hundred ordinary transactions.

The system saw an island. Stripe had launched Radar in 2016. For eight years, it learned from every transaction on the network, tens of billions of them, fifty thousand new ones every minute, $1.4 trillion in annual payment volume. The traditional machine learning models trained on discrete features worked well enough, reducing card-testing attacks by eighty percent over two years. Then in 2024, Stripe built something new on top of that decade of accumulated data: a transformer-based Payments Foundation Model. They treated each transaction as a token and payment sequences as context. Same architecture as a language model, except the language was money.

The old system looked at individual transaction features. The new one, trained on a decade of compounding operational data, understood the semantics of money: transactions from the same issuer cluster together, transactions from the same bank cluster closer, transactions sharing the same email cluster closest of all. The two hundred fraudulent requests did not flag on any single feature. They flagged because they appeared as an island in an embedding space that only exists because the system had been watching money move for eight years.

Detection of card-testing attacks on large merchants jumped from fifty-nine percent to ninety-seven percent. Overnight.

Emily Sands, Stripe's Head of Data and AI, named the mechanism: "The real advantage isn't the raw size. It's the compounding loop." Ninety-two percent of cards a merchant sees for the first time, Stripe has already seen on another merchant. Each transaction improves the model, which attracts more merchants, which adds more data, which improves the model again. The loop has been running for a decade. A competitor can build the same transformer architecture in a quarter. They cannot buy the decade of transaction data that makes it see the island.

The compound, running, looks like this: a system producing intelligence that nobody programmed, from data that nobody else has, improving with every cycle. The encoding lives inside it. The architecture carries it. Infrastructure gives it a home, and time makes it unreplicable.

The Transfer Needs a Home

And right now, for most people who have made it this far, the encoding lives in a prompt.

A prompt inside a tool that will update its interface next quarter. A set of instructions inside a platform that changed its API twice this year. A document that someone wrote last month and nobody has opened since. The expertise is real. The structure underneath it is not. The encoding runs when you remember to trigger it and improves when you manually review the output and manually rewrite the instructions and manually feed the corrections back into a system that has no memory of what it learned yesterday.

Call it what it is: a person doing a slightly different version of the same work they did before the encoding existed.

The compound begins at the moment the system closes a feedback loop without anyone touching it. The output of cycle one becomes the input of cycle two, automatically, and cycle two produces something measurably better than cycle one did, and nobody had to intervene for that improvement to happen. Before that inflection, you have a tool. After it, you have a system. The tool requires maintenance. The system requires direction.

What Exists Right Now

The cost of running an encoded AI system dropped by an order of magnitude in a single year. Open-source models closed the capability gap with proprietary ones to the point where choosing a provider is a preference, not a constraint. You can switch providers without losing your Layer 1 decision logic, can run on your own hardware if you decide no external company should hold their data. The lock-in that defined the first wave of AI adoption is dissolving. The cost objection is dead.

The software story is encouraging, but the physical story is something else entirely.

Data centers in the United States are running at ninety-five percent capacity. The chips that power AI workloads are sold out through mid-2026. AI infrastructure consumes electricity at a scale that would make it the fifth-largest power consumer on earth if it were a country. Virginia, the largest data center market in the world, faces moratoriums because the grid cannot keep up. New power generation takes four or more years to build. The five largest technology companies committed nearly seven hundred billion dollars in capital expenditure and it still cannot close the gap between demand and what the physics of construction allows.

The software gets cheaper every quarter. The atoms underneath it obey a different clock.

Your layer, the part where the encoding actually runs, tells the sharpest story. Nearly nine in ten AI agents never reach production. They are built, demonstrated, and shelved. The fraction that survives shares one trait: narrow scope, deep encoding. Specialists carrying one expert's judgment for one domain, not general-purpose assistants answering any question. The ones that fail do not fail because the hosting went down or the API changed. They fail because nobody encoded the decision logic that would have made the output worth trusting.

The infrastructure was ready, but the expertise inside it was not.

If you have been building through this and feel behind, stop. You have probably sat through enough vendor demos and platform launches by now to recognize the vertigo, the sense that the ground keeps moving before you finish laying the floor. That feeling is accurate. But the reader who built through the previous five chapters is ahead of ninety-five percent of the market, because construction at this scale has outpaced the encoding layer everywhere. The compute is abundant. The models are capable. The physical layer is constrained but expanding. And encoding, the layer that determines whether any of it produces value, is where almost everyone has done the least work. The question, if you have done it, is what you build on that position before the ground shifts again.

Evidence

The Infrastructure Stack (March 2026)

What Is Forming

When the encoding has a home and the feedback loops start closing, something happens that you will not expect. The system stops needing them for the part they thought was irreplaceable. Not the judgment, but the execution. The judgment stays. The execution compounds on its own, each cycle producing data that the next cycle uses to produce better data. Someone who served twenty clients by working now serves two hundred by checking. The bottleneck does not disappear. It moves from the hands to the eyes.

The compound forms in what the encoding produces when it runs long enough to accumulate assets nobody else can replicate without running for the same duration.

Bloch's filter resolves this. AI compresses the time it takes to do things. It does not compress the time it takes for things to happen. Building software, producing content, running analysis, maintaining integrations: all of it accelerates. Building a proprietary dataset from thousands of real client interactions, accumulating network density across an industry, earning regulatory permission through years of compliance history, compounding operational knowledge through edge cases that only surface in production: none of it accelerates. The speed of doing collapsed, but the speed of happening held.

That filter separates what compounds from what expires.

Every technology that transforms an industry eventually disappears into the infrastructure. Dishwashers automated domestic labor and nobody thinks about dishwashers. GPS externalized spatial memory and nobody thinks about navigation. Search externalized factual recall and nobody thinks about retrieval. The trajectory runs one direction: the tool becomes invisible, the capability it provides becomes the default, and the default captures everything.

The pattern is older than any living industry. The printing press did to monastic knowledge what AI is doing to expert knowledge. Before Gutenberg, a scribe's value was inseparable from the scribe. The expertise lived in the hand, in the monastery, in the person. The press encoded it into a system that transmitted without the scribe present. The scribes who adapted became editors, publishers, scholars. The ones who insisted the hand was irreplaceable became footnotes. Five centuries later, the same displacement runs at a different clock speed on the same logic: the expertise trapped in a person becomes encodable, and the person who encodes first captures the production layer of whatever that expertise touches.

The destination the paper has been building toward is the same disappearance applied to the current era. The system you encode your judgment into does not stay visible to your customer. It becomes the surface they interact with. They never see the architecture underneath or the encoding. They see output: the deliverable, the recommendation, the product, the answer. It arrives with a quality they associate with you, produced by a system they never think about. The expertise became the default. The default captured the market. And whoever set it first captured the position.

The data on defaults is not theoretical. In countries where organ donation requires opting out, consent rates reach ninety-eight percent. In countries where it requires opting in, fifteen percent. Eighty-three points of difference from changing who has to take the action. Most people do not change defaults. Most businesses do not switch infrastructure. If you become the default, you do not need to be better forever. You need to be first, and adequate, and present when the decision is made.

If you built through the first five chapters, this process is already underway. Your encoded systems run, your data accumulates, and your clients experience the output and stop thinking about how it was produced. Every month the system runs is a month of compounding that no competitor can buy back by starting later with a better model or a bigger team. The model will be better. The team will be bigger. The accumulated time will not be there.

Framework

The Five Time-Bottlenecked Assets

The five percent of companies that have achieved material AI returns generate 3.6 times the shareholder value of the other ninety-five percent. Technology was not the differentiator. The technology is the same for everyone. The differentiator was whether anyone built the architecture, encoded the expertise, and gave the compound enough time to start forming. The five percent did. The ninety-five percent installed AI on top of their existing structure and wondered why nothing changed.

What We Cannot Predict

Everything above has been precise about what is measurable. What follows is precise about what is not, because honesty about where the map runs out matters more than a performance of certainty.

Nobody knows whether AI capability will continue climbing at the rate it has climbed. Three quarters of the researchers building these systems believe the scaling approach that produced the last four years of gains has reached a ceiling. The people running the labs disagree. Both camps are guessing from different vantage points with different incentives. The leading forecasting organization in the field admits its predictions carry roughly half a significant digit of precision. Half a digit. The honest answer is “we do not know, and neither does anyone else.”

The model in this paper holds under both scenarios. If capability plateaus, encoded expertise becomes the primary differentiator because the model underneath it stops improving and the only variable left to optimize is what you feed it. If capability explodes, encoded expertise becomes the verification layer that determines whether the explosion produces signal or noise. Either way, anyone with deep encoding is positioned and anyone without it is exposed. The risk is symmetric in a way that favors building.

The timeline to artificial general intelligence spans decades depending on who you ask. Some of the people building it say five years. Some say thirty. The median estimate from the research community falls somewhere around 2040. If it arrives and can generate its own domain expertise the way AlphaGo Zero surpassed all human knowledge of the game without studying a single human move, then the thousand hours of closed loops that Chapter 3 described become a sunk cost rather than encoding capital. The bottleneck migrates from encoding what you know to encoding what matters. The expert does not vanish but shifts into a different role, the person who defines the objective rather than the person who executes it. A different job. It may be a more valuable one. But it is not the same one.

One counterargument deserves more weight than any other. A randomized controlled trial measured experienced developers using AI coding tools. They believed they were twenty percent faster. They were measured as nineteen percent slower. The confidence of acceleration masked the reality of deceleration. In aviation, the same pattern killed two hundred and twenty-eight people on Air France Flight 447 when pilots who had relied on automation for years could not manually fly the aircraft after the automation disengaged. In medicine, nurses who relied on AI predictions and encountered a wrong prediction performed worse than nurses who had never used the AI at all. Nearly twice as badly.

The pattern is deskilling. The tool that augments the expert's capability simultaneously erodes the capability it augments, because the expert stops practicing the judgment the tool was built to carry. Cal Newport's research sharpens the diagnosis: introducing AI tools into knowledge work increased administrative tasks by over ninety percent while reducing deep work effort by nearly ten percent. The tools accelerated the wrong layer. The encoding flywheel is a partial answer, because the expert who shifts from executing to verifying maintains contact with the domain. But verifying is not doing. The skill that built the expertise was doing. Whether verification preserves the judgment that execution created is a question the paper cannot answer with the evidence currently available.

A related problem sits underneath. AI is automating the entry-level work that trains the next generation of experts. The junior associate who used to spend three years doing the tedious pattern work that built their intuition now has a system that handles it in minutes. The tedium was not overhead. The tedium was the training. If the pipeline that produces experts breaks, the current generation of encodable expertise may be the last one with full-depth domain knowledge. The paper names this problem because pretending it does not exist would violate the first variable the paper was built on.

The honest position is that encoding is a bet: the evidence favors it, but the evidence does not guarantee it. Three conditions would collapse the thesis:

If AI achieves autonomous domain expertise through self-play or simulation, the way AlphaGo Zero surpassed all human knowledge of Go without studying a single human game, then the thousand hours become a sunk cost and K loses its advantage. If the deskilling trap degrades the expert pipeline faster than the encoding flywheel can capture what remains, then the current generation of encodable expertise is the last one with full depth, and the model's shelf life is measured in decades, not centuries. If regulatory capture or infrastructure consolidation locks incumbents in before new entrants can encode, then the window described in Chapter 2 closes differently than predicted, favoring capital over competence.

Under any of these conditions, the model fails. A falsifiable claim. The paper holds itself to it. If you build with that distinction in mind, if you build aggressively but carry the humility to know which parts of the foundation might shift, you are who this paper was written for.

Why the Clock Is Running

The infrastructure exists, the models are capable, the methodology is proven, and the compound is forming for those who started. None of these are arguments to act; they are conditions. The argument to act is scarcity.

Talent is scarce. Three open positions exist for every qualified candidate in AI implementation, and the demand grows faster than the training pipeline can produce. The founders who secured their architects and system builders in 2025 are not sharing them. Anyone who starts looking in 2027 will find the market picked over and the remaining talent priced beyond what most mid-market companies can afford.

Physical capacity is constrained. Data centers in major markets are full. New power generation takes four years to build from permit to production. The largest data center market in the country has imposed moratoriums because the grid cannot support additional load. Software runs on hardware, hardware runs on electricity, and electricity obeys the speed of construction rather than the speed of innovation. If you secure infrastructure access now, you operate inside a constraint that tightens every quarter.

The regulatory window has a public deadline. August 2, 2026. The EU AI Act's high-risk system requirements go live. Anyone building before that date architects compliance into the foundation. Anyone building after it retrofits under enforcement pressure, at multiples of the cost, with regulators watching. The nineteen US states with enacted privacy laws are not reducing their scope. The surface area of regulation expands because the capability it regulates expands. Permission to operate becomes harder to earn with every month that passes.

The market is consolidating. In the first vertical where AI agents reached maturity, the top three captured seventy percent of the market within eighteen months, a preview of what comes next. If you encode first in your industry, you set the default. Encode second, and you compete for the remaining margin. Encode third, and you discover the position was taken while you were planning.

The second-mover argument says: wait. The tools get cheaper. The models get better. Let someone else make the mistakes. Correct about the tools. The tools will be cheaper next year. Wrong about everything the tools run on. The encoding that the first mover built while the second mover waited does not get cheaper. The data that accumulated does not transfer. The network density, the regulatory history, the operational advantage: none of it relocates or resells. Every month the first mover ran the compound is a month the second mover cannot purchase at any price.

The Closing Argument

Dario Amodei, who is building what may be the most powerful AI system on earth, describes this moment as an adolescence: a species handed almost unimaginable power before the institutions that govern it have the maturity to wield it. Leopold Aschenbrenner, who was inside the lab before he wrote the paper that defined the conversation, puts years on the timeline and says the endgame is on. Tristan Harris, who warned about social media in 2017 and watched every prediction come true, calls the only viable path a narrow one, not acceleration or caution but discipline at every level. The narrow path between building something that compounds and building something that collapses runs through every variable in this model.

If you hold both realities, if you build aggressively within the constraints of what is actually responsible, if you encode with depth but verify with discipline, if you move fast on the parts inside the frontier and keep human judgment on the parts outside it, you build things that last. Speed without discipline creates fragility that compounds. Caution without action creates irrelevance that also compounds. The path is narrow. The model tries to map it.

The compound hasn't fully formed for me yet either. This paper is the first proof, not the finished product.

Somewhere around the four hundredth demo I stopped hearing the objections and started seeing the structure. Three years ago. The structure became a model, the model became a paper, and the paper became the first test of whether the thesis holds: expertise encoded into a system that transmits without the author present, distributed through infrastructure that compounds with every reader who acts on it. If the thesis is correct, this paper is its own first proof. If it is wrong, the paper is the most detailed record of a bet that did not pay. Either way, the signal is clean. It originated at the source. Everything in it is what I actually believe. The rest is up to the compound.

Your expertise is either a compounding asset or an expiring commodity. The window is open. The clock is running. The compound has started for the people who built before you. The only question left is the one this paper cannot answer for you.

What you do with what you know.