Everyone now has the same intelligence on tap. The comfortable conclusion is that this levels the playing field. The true conclusion is the opposite. As input converges towards being universal, the only thing left to vary is whether your organization is allowed to reorganize around it — and large organizations are, almost by definition, built to forbid exactly that. The scarce resource in 2026 is the permission to act on intelligence, and that's the trap which only flat organizations can escape.
I learned how money actually works at 13, selling Diet Coke and candy to my schoolmates.
The product was nothing. Anyone could buy a can of Coke. What compounded was not the candy but my reputation as a generous lender: I let people eat first and pay later, so more and more of them came to me to do exactly that. That meant taking on credit risk with people I barely knew, and the way I solved it was the second lesson — leverage. I paid a tall Russian friend ten dollars in candy to stand next to me whenever someone new wanted to buy on credit, and to chase the debt if it came to that. He's in the New Zealand Army now, a genuinely good guy. The school was small enough that everyone knew him. I never wrote off a single bad debt.
When demand outgrew me, I hired classmates to sell under my name, using the same method, and scaled it to about $600 a month in profit. We all carried the same colour of bag — that was the brand. The school eventually shut it down, because the on-site tuck shop was quietly losing its customers to a swarm of kids in matching bags.
That was the most useful business lesson I have ever had, and it is this: most products do not differentiate. The raw thing you sell is usually a commodity, and when it is, you do not compete on the product — you compete on trust, brand, distribution, and borrowed leverage. It also taught me that building in public is frequently a mistake. The matching bags were good branding and bad operational security. Visibility is what got us caught.
Then I stopped compounding for a long time.
I did not understand at thirteen that I had just grasped the whole thing — that money is value transfer, and the rest is mechanics. So I drifted. I played League of Legends semi-professionally, talked my way into a university lab to chase brain–computer interfaces because a film had hooked me, went to London to study, and ran a handful of side projects, one of which reached $20k MRR. By the conventional scoring, this was a loss. The people my age who are winning have spent a decade compounding deep domain knowledge at some technical frontier. At twenty-five, I have no such moat.
But the conventional scoring is wrong here, and seeing why is the entire point of this essay.
The question "how do I build something that compounds without a decade of my own expertise buried inside it?" is the same question as "how do I build something that compounds without me at all?" I was forced to ask the first question because I had no decade to offer. In 2026, the second question finally has an answer. The weakness turned out to be the brief.
When I raised venture money in 2023, I already knew the shape of what I wanted: a company that could eventually run without me, and grow at least with the industry in the years I was absent. In 2023 this was impossible. Context was lossy. It lived in people's heads, or it was compressed into rigid database schemas, and even a carefully written instruction was read five different ways by five different people. There was no medium that could hold the company's actual reasoning.
Here is what I had not noticed: the problem I was now facing was the candy problem again. A frontier model is a commodity — everyone can rent the same intelligence, the way anyone could buy the same can of Coke. When the raw input no longer differentiates, the edge moves to the system you build around it. At thirteen that system was trust and borrowed leverage. This time it would be something the candy could never do — improve itself.
Late in 2025, frontier models crossed a threshold, and the medium appeared. With the right scaffolding — above all, explicit rules for when and how context is handed off — you can suddenly retrieve accurate context and move information entropy around at will. We kept memory in markdown files, ruthlessly prioritized and kept current, and let the model run internal functions: PR notes, then engineering tickets, then customer-facing documents. The surprise was not that it could do the work. The surprise was that, given enough well-maintained context, it spotted patterns I had missed and proposed strategy I would not have reached on my own.
So here is the definition worth fixing in your head. A recursive company is one whose core loop — observe, decide, act, then rewrite its own instructions — runs more times each week without a human stepping into the iteration. The human's job does not vanish; it migrates. You stop running the loop and start curating it: deciding which context is non-negotiable for each class of decision, and letting the loop handle everything downstream of that judgment.
Running this way, we grew one product's customer base from 1,000 to 4,500+ in a matter of weeks, with no increase in headcount.
None of this is a private discovery. Tom Blomfield was describing the same loop at a YC event this week; friends at several hedge funds have spent the past year trying to rebuild their analysts the same way, with mixed results. We happen to have been running it in production since February. When unconnected people converge on the same structure from different directions, that is usually a sign the structure is real — not that any one of us has it figured out.
Now the part that matters for anyone choosing what to build.
The obvious objection is that a large company will simply do this too, with more data and more capital. It will not — and the reason is structural, not a matter of speed or competence. To make a company recursive you must remove humans from decision loops. But a bureaucracy is not a company that happens to have humans in its loops. A bureaucracy is the humans in the loops. Their sign-off, their judgment, their headcount, their political capital, their path to promotion — all of it is constituted by being a necessary node. So the people who would have to approve the transformation are precisely the people the transformation makes unnecessary. You are asking an organization to vote for its own deletion. It declines. Instead it buys AI and bolts it onto the org chart as a tool that makes each existing node slightly faster — which is not a recursive company, only a faster bureaucracy.
A startup has no veins to slice. There is no one whose career depends on standing inside the loop. This is the rare advantage that grows larger the smaller and younger you are. Almost every other edge decays as you scale; this one is strongest before you have an org chart worth defending. That asymmetry — not the models themselves, which everyone can rent — is the real alpha of 2026, and I do not see nearly enough founders treating it as one.
I will not pretend the position is fully secure. Today's edge is partly scaffolding: the markdown memory, the handoff rules, the discipline of context. A fair question is whether the models eventually absorb the scaffolding and leave nothing proprietary behind. Some of it, certainly. But the last mile is not a temporary inconvenience waiting to be optimized away — it is structural. Deciding which context is mandatory, owning the consequences of a decision, sitting at the messy seam where the model meets the world: those do not disappear as the model improves. They relocate. Anyone claiming certainty about where they land is selling something.
The cliché says AI is the great equalizer. It is not. It is the great divider — between organizations that can act on intelligence and organizations that can only approve it. For now, that gap is wide open, and walking through it is the cheapest alpha a startup will ever be offered.
This piece started as a conversation with Kareem on the podcast. Tom Blomfield's YC talk is worth watching too.