Opinion: 'If I Can Build It, I Understand It' Is a Bad Rule for AI

Opinion.

On the September 21 episode of All-In, Naveen Rao gave host Chamath Palihapitiya a tidy epistemology: "I don't feel like we truly understand something until we can create it." It's a clean line, the kind that sounds like wisdom because it rhymes with the scientific method. It isn't the scientific method. It's an engineer's shortcut, and shortcuts are where the invoices hide.

Here's the distinction that matters. The scientific method says: predict, test, falsify, repeat, and your understanding is only as good as your model's ability to fail publicly. Rao's version says: if I can make the thing work, I understand it. Those are not the same claim, and the gap between them is exactly where a lot of venture capital currently sits.

Take the term of art the episode title leans on: the "energy wall." In AI infrastructure, that phrase refers to the point where further gains in model scale are constrained not by algorithms or chip design but by the physical availability of electricity — grid capacity, power purchase agreements, transformer lead times, the unglamorous plumbing that doesn't show up in a demo. It's a real constraint, and it's been attributed to multiple infrastructure operators and utility filings over the past two years, not invented by any one guest. The energy wall is a fact about voltage and megawatts. It does not care whether an engineer "understands" the model he built. The grid does not grade on epistemology.

That's the trouble with importing "I understand it because I built it" into a domain like biology, which the episode also flags — "beating biology," per its own title. Biology has spent roughly 3.8 billion years running an experiment with a sample size larger than any dataset a lab will train on this decade. Being able to synthesize a protein, or fold one with a model, or edit a genome, does not mean you understand the system you just perturbed. It means you found one lever that moves one outcome under one set of conditions. Pharma has generated multiple expensive lessons on that distinction, and none of them came cheap.

None of this is a knock on building things. Building things is good. Markets reward people who ship. But there's a difference between "I built a working system" and "I understand the system I built," and podcasts are an unusually poor venue for keeping that line straight, because the format rewards the confident aphorism over the hedged one. A host nodding along is not peer review. An audience of a few hundred thousand listeners is not a replication study.

The libertarian objection here isn't to Rao's ambition — it's to letting a builder's felt sense of comprehension substitute for the falsifiable kind, especially when that felt sense gets recycled into a fundraising narrative. Fewer rules, clearer numbers, no cheering section for founders means this: if the claim is that a model "understands" biology or cognition because it can generate plausible outputs, put a number on it — a benchmark, a held-out test, a dollar figure on what breaks if the claim is wrong. Chamath, to his credit, didn't adopt the thesis on air; he let Rao state it and moved on. That restraint is worth more than the aphorism was.

The energy wall, unlike the philosophy, comes with a meter. Power purchase agreements get priced in dollars per megawatt-hour, not vibes. That's the standard AI infrastructure claims should be held to — not whether a founder feels he understands what he built, but whether the thing keeps working when the power bill comes due and someone else tries to reproduce it.

Heard on All-In with Chamath, Jason, Sacks & Friedberg — Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology (2026-09-21).