Riding the AGI

A lead in this technology now expires in weeks, so the working move is to live a bit ahead of it rather than defend last month’s edge. Which layer stays scarce is contested, and that hedge sits next to the claim rather than under it. The duration of exclusivity is the new variable.

The stack, held loosely

The room’s picture has three layers. Hardware is treated as a commodity a manufacturing-scale state already owns: anything that can be made in the US can be made there more cheaply, one tiny cable has thousands of manufacturers in a single city, and production is subsidized to capture scale. Manufacturing scale is real. “Owns hardware” and “no one beats them in a decade” are partisan, and they are not load-bearing for the operators that follow.

Software is treated as commoditized the moment an agent can one-shot a specification — the economic consequence of Software 3.0, which owns natural language as a programming medium. The completed tense overstates it. The direction is the usable claim.

Model-building is treated as the one layer that is not a commodity. Nobody builds a frontier model in a garage: clusters, proprietary data, scarce researchers, and a regulatory field that Regulatory Capture via Doom-Marketing owns as the way an uncommoditized layer concentrates. Fine-tunes and small models are garage-accessible. The hedge is on frontier training. Doom-era democratization never arrived at that scale.

This is the claim with the shortest shelf life, and the one The Margin Moves to the Serving Layer contests in full. A sibling episode three weeks later puts durable money in applications above and cloud and chips below. This page does not settle the layer. It points.

Open source, on the same picture, rarely surrenders a lead it takes. Once an open project pulls ahead, an ecosystem accretes — the maintenance-lead pattern of a kernel or a mobile OS, used here as analogy, not proof. A closed lab then cannot justify burning runway to re-pass something free. Open-source maintenance leads are real. Open-source frontier-model leads, as of 2026, are not established. Keep it as a tendency the panel believes, not as a law.

China’s open-source flood is offered as deliberate state strategy: a cheap software layer keeps Chinese hardware competitive, and funding labs to stay number two or three and publish anyway would be rational statecraft. Complementary-goods logic is real economics. The intent is unshown. No funding evidence was offered. Theory, marked.

Two labs earning revenue directly off their models, with user bases feeding reinforcement trajectories, were named as the remaining kings where five had stood two years earlier. That snapshot will rot. It is color, not the spine.

The strongest open models already run cheaper and, given a good harness, close much of the usable gap. The multiples named on air are unsourced. The direction travels. Power users reach for open or jailbroken models to shed the refusal-and-tone-policing tax. That is a pull, not a market share.

Live in the future

The same smartest model can be exclusive for two to four weeks and then not exclusive at all. The prior world let an edge be operationalized with some durability. That window is contracting toward zero. The response is not to defend the edge. It is to keep moving to the next one. positional decisions are how those bets get priced when the compounding surface is that short.

Live in the future, here, means paying to work under next year’s conditions and building backwards from what that reveals. The heuristic is old and public. It is a method, not a budget. Spending on the order of a hundred thousand dollars a year on tokens, in the room’s illustration, buys the working conditions of a normal user two years later. That figure is a 2026-07 illustration, not a price forecast and not a commandment. The open question is the smallest spend that actually puts a year of conditions ahead.

The present lever is context. A long window can hold a whole corpus. A person holds a handful of items in mind at once. The contrast is corpus-in-context against handful-in-mind, not a claim about how many facts a company can be known by. A dated product illustration — a million-token window, three novels — will move. The contrast will not.

Orders of magnitude more inference change what can be done, not only how fast. A volunteered multiple for the next two to three years was offered with a hedge of a couple of orders of magnitude and is not a finding.

A named tool in December 2025 was the room’s tipping point. The pattern is the keepable thing. Dismiss the tool at capability N, and the jump to N+1 makes the earlier verdict wrong. The product name is not the spine.

What survives whichever layer is scarce

Advantage decaying in weeks, the compounding-error reason to pay for intelligence, and the context lever survive whichever layer turns out to be scarce. The rest is operating intelligence from inside a room selected for people for whom this is already working. Timelines rot fastest. Geopolitics is partisan and not load-bearing.

The AI Industrial Revolution is the sibling field report: factory rather than artifact, the smartest-model logic. Those compressions stay there. The AI Productivity Curve asks whether the compute build-out shows up in productivity. This page does not steal that question.

Where intelligence is worth paying for

Loop a model many times and a middling-reliable one collapses. A much more reliable one degrades slowly. Independent trials make the arithmetic ordinary: a nine-in-ten success rate, run a hundred times, is nearly zero; a 999-in-a-thousand success rate, run the same hundred, is still about nine in ten. The input percentages on air were invented. The arithmetic is not. A large error-rate gap is worth paying for on high-leverage judgment. Cost-sensitivity belongs on cheap repetitive tasks. Accuracy Before Speed owns that compounding-error arithmetic.

The same compounding surface, here time-contracted from years to weeks, is what The Age Of Nonlinear Returns owns at full depth.

The bitter lesson is the finding that general methods that ride more compute beat specialized, hand-built systems. The startup implication is the panel’s, not the finding’s: vertical AI software is exposed to a frontier model with tool use. Small leveraged teams reaching large revenue is real. Selling software that is itself commoditizing is a shrinking base. The structural risk is ordinary antitrust intuition: if the harness war collapses to one provider, or labs nationalize, startups face a monopolist. Dates and revenue figures attached to that risk are the room’s.

What stays human

If capability reaches expert level but not a system that supplies its own aims, humans remain the motivated element — handlers of a fleet, supplying taste, judgment, accountability, and desire. Desire is the one input a model cannot supply. Higher-Order Generativity vs Higher-Order Judgment grades the writing argument that follows. Human vs AI Capability Lens owns what stays human on two axes. A Motorcycle for the Mind already owns tutor-at-your-level and the full desire-as-human-input treatment. This page keeps desire as the remainder, not as a restaged essay.

Displacement, on this page’s bet, concentrates on people who refuse the tools, not on those who pick them up. Automation and the Job Iceberg owns the displacement record. Complementary-automation evidence is mixed. The bet is pointed there, not proven here.

Writing and speaking are the output of thinking. Outsourcing them wholesale atrophies the thinking. Prose a human never compressed wastes the reader’s time. Agent-to-agent text with neither human in the loop is the degenerate end. The Right vs Wrong Way to Work With AI owns the offloading line. The counter already on the table: a large enough personal corpus plus an eval harness can make a “skill file” indistinguishable from the person. The harder claim from the room — good writing is novelty a next-token model cannot produce — grades badly against the generativity-versus-judgment split. The grade stands. The split is not resolved the other way.

The room is three founders and a host talking their own book, selected for people for whom AI is already working. That interest mark is part of the report.

The strongest internal counter is already inside the same picture. The commoditization and concentration they describe drive toward an oligopoly, where “just ride it” is a bet on a window staying open. A sibling episode three weeks later compresses the exclusive-to-commodity cycle and puts durable money above and below the model. Newest wins. The model layer stays contested. The recap lives on The Margin Moves to the Serving Layer.

The same weeks-long exclusivity is still the condition. Riding it is remaining the motivated element while each lead dies. Desire is the input that does not arrive with the model.

Related

• The AI Industrial Revolution — sibling field report: factory-not-artifact, smartest-model logic
• Software 3.0 — natural language as a programming medium; this page is the economic consequence
• Regulatory Capture via Doom-Marketing — how the uncommoditized layer concentrates
• The Margin Moves to the Serving Layer — the counter-position on which layer is scarce
• The Age Of Nonlinear Returns — the compounding surface, here time-contracted
• The AI Productivity Curve — whether the compute build-out shows up in productivity
• Accuracy Before Speed — compounding-error arithmetic
• Automation and the Job Iceberg — the displacement record this page’s refuse-the-tools bet points at
• Higher-Order Generativity vs Higher-Order Judgment — grades the AI-writing argument
• Human vs AI Capability Lens — what stays human, two axes
• A Motorcycle for the Mind — desire as the human input; tutor-at-your-level stays there
• The Right vs Wrong Way to Work With AI — the offloading line
• positional — how to price bets when the compounding surface is short

Open Questions

• Which layer is scarce — model-building, or applications above and cloud and chips below — stays open.
• What is the smallest spend that actually puts a year of working conditions ahead.

Sources

• Naval Ravikant, Garry Tan, Daniel Francis, Farbood Nivi. “Riding AGI, AI Anxiety, Who Funded COVID, Defending Taiwan, and California Empire.” nav.al/future, 3 July 2026. The operators, the duration claim, the stack held loosely, the live-in-the-future method. Talking their books.
• All-In Podcast, episode 282 (25 July 2026). The counter-position on which layer is scarce. Developed on The Margin Moves to the Serving Layer, not recapped here.
• Richard S. Sutton, “The Bitter Lesson,” 2019. General methods that leverage computation beat specialized, hand-built knowledge.
• Paul Buchheit. “Live in the future and build what’s missing.” Long-standing public heuristic; the method, not a budget.