Noam Brown – Agent swarms, alignment, & recursive self-improvement

The episode opens with the fact that reorganises the whole conversation: OpenAI announced, the week before this recording, that a system of ten thousand separate agents had solved one of the Millennium Prize Problems — 130 billion tokens burned over 88 hours. Dwarkesh's framing is that Noam Brown was two or three years early on the previous wave, arguing that scaling inference compute lets you see the future, because the base capability of models a few years out is already legible in the scaling curve today. The question now is whether agent count is the same kind of telescope.

Brown's version of the curve is simple and empirical: plot test-time compute on the x-axis against performance on basically any reasoning benchmark, and the line goes up. The analogy is the SAT — five minutes versus five hours. The model spends the time in monologue, enumerating cases, ruling out possibilities, building on its own earlier discoveries.

The limit is latency. Nobody waits three years for an answer, so you do what organisations do: parallelise. A team of people, a company of people, ten thousand agents. Multi-agent is test-time compute scaled in parallel rather than serially — less efficient, because no single agent owns the whole context, but effective if executed well.

Dwarkesh flags that he is going to ask naive questions because the model is unreleased and the qualitative properties are genuinely unknown. Then he does the arithmetic that anchors the episode: 130 billion tokens is 4,000 years of a single human thinking full-time, eight hours a day, from ancient Sumeria to now — compressed into 88 hours.

What surprises him is the absence of a larger parallelisation penalty. Ten thousand agents collaborating at that scale should degrade badly, unless agents coordinate better than humans do, or unless the penalty is real and simply hidden behind the raw volume. Brown takes the second framing: we do not have good science on multi-agent scaling at this size.

The published data comes from the 5.6 release, the first model with a proper multi-agent system shipped as an option — Ultra Mode, four agents by default, configurable higher. The blog post plots one, four, and sixteen agents on several benchmarks. Four agents finish a problem twice as fast for twice the cost; sixteen continues the pattern with somewhat worse efficiency.

The speedup is slightly sublinear, and the variance across domains is the interesting part. Mathematics parallelises well, though not maximally. Web search — a Deep Research report sifting many sources — parallelises extremely well. Writing a novel would not: ten thousand agents on a novel would fail for the same reason ten thousand people on a novel would fail.

The science stops at sixteen agents because the experiments get expensive. Ten thousand is one data point, and a single data point cannot tell you what a single agent would have needed to solve Navier-Stokes, because that experiment has not been run. A proper ablation at that scale is unaffordable, so the field will have to settle for methodical work at 64, 128, 256 and extrapolate — and will never know for certain what the marginal 9,000 agents bought.

Brown then spends real effort deflating his own headline. The Millennium Prize result was not a multi-agent achievement; he would not give multi-agent ten percent of the credit. What actually happened is that OpenAI trained a very powerful model, got it to operate over very long horizons, and got it to think in parallel. Multi-agent is flashy and new, and therefore collects disproportionate attention.

The generalisation is what startles Dwarkesh. These systems are trained by RL against checkable synthetic problems, none of which resemble a Millennium Prize Problem, yet the competence transfers to a problem of that ambition and to a parallel effort of that size.

Brown partly agrees — the models are trained on very hard problems, and there is a measurable gap between training difficulty and demonstrated capability — but raises the structural worry. As models get smarter, the questions you can pose become too easy. That is the strongest argument for why LLMs might not follow the AlphaGo curve.

The mechanism is the curriculum. Self-play in Go is an infinite curriculum: you always face an opponent of equal strength, so the difficulty ratchets automatically. RL on an LLM gives the model a fixed problem; if it solves it in a second, it learns nothing. If the supply of problems hard enough to challenge the model runs out, progress slows — a plausible scenario, not yet a wall, and one Brown thinks would be worked around if it became serious.

Dwarkesh supplies the contrast for the audience: Go went from beating a European champion ranked around fiftieth in the world to beating the world champion to being orders of magnitude beyond any human, inside a single year. Mathematics could take that trajectory. Brown thinks it plausibly will not.

The practical question — if people have access to these systems in six months, how should one model what it is like to hire one — requires explaining that OpenAI's architecture is deliberately unlike the multi-agent systems most people build. The common pattern is a scaffold: a coordinator delegates to children, children return answers.

The scaffold's failure modes are structural. Two children given overlapping tasks usually cannot talk to each other, which is exactly the inefficiency a human team would never tolerate. A child with a clarifying question must choose between returning the question unsolved and guessing at the parent's intent. Every scaffold bakes in some of these limits.

OpenAI's choice was the opposite extreme: bake in as little structure as possible, give the agents primitive tools, and let them discover coordination. The core primitive is a message tool — an agent can send a message to another agent whenever it wants, and the message is inserted into that agent's context. That is essentially the whole mechanism.