Noam Brown

Sources#

• Noam Brown – Agent swarms, alignment, & recursive self-improvement
• On the Navier–Stokes Millennium Prize Problem
• Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI's Noam Brown

Summary#

Noam Brown is a research scientist at OpenAI and one of the researchers credited with pioneering inference-time (test-time) compute scaling — the reasoning paradigm behind the GPT-5-series "thinking" models. Before frontier LLMs he built superhuman poker agents (the Libratus / Pluribus line of work), and he still uses building a poker solver from scratch as his personal capability eval. In this corpus he is the sole author-subject of the June 2026 No Priors interview about his essay Implications of Large-Scale Test-Time Compute.

What he argues (in this corpus)#

Brown's essay hangs on one root claim — capability is now a function of inference budget — with several consequences he traces:

• The benchmark grid is broken. Single-number benchmark tables don't control for test-time compute, so a more efficient model (GPT-5.5) can look only marginally better than its predecessor while being a substantial jump. Fix: put cost/tokens/time on the x-axis. See Compute-Controlled Benchmarking.
• Safety evals are ill-defined at unbounded budgets. Preparedness frameworks / responsible scaling policies were built for the ChatGPT era and don't ask "at what budget do you evaluate?" — yet dangerous capability scales with dollars just like useful capability.
• No overnight intelligence explosion. Because peak capability requires large test-time-compute runs, time becomes the binding constraint; takeoff is gradual, not instantaneous. See Intelligence Explosion Dynamics.
• Research taste is the residual human role. Models optimize his poker algorithms 10–100× but cannot yet invent a better one; they're "a very good complement to researchers," not a replacement — though he expects an inflection point here like the ones in coding and math. See Research Taste as the Human Bottleneck.
• A latent-capability overhang exists. Nobody has explored what $100K of compute into a released model could do; the Erdős unit distance conjecture was disprovable from a public model before OpenAI announced it. See Latent Capability Overhang.

Brown's essay is practitioner-opinion — arguments and anecdotes from one lab. In July 2026 the UK AI Security Institute published the first independent, empirical corroboration of its core cluster, measuring capability curves over token budget across several benchmarks. The AISI cyber result Brown cited (models "still improving at 100M tokens") is that institute's own work, now published in full. His thesis is no longer sourced to a single person.

The same person, three months later: what changed (September 2026)#

The vault now holds two long interviews with Brown twelve weeks apart — No Priors (2026-06-26) and Dwarkesh (2026-09-17, Noam Brown – Agent swarms, alignment, & recursive self-improvement) — and he is the only source in the corpus with that shape. Read as a pair they are a small, useful instrument: a practitioner's stated beliefs, re-elicited after a large intervening event (OpenAI's Navier-Stokes result) by a different interviewer. Both are practitioner-opinion; neither is a measurement. Four things moved.

1. He concedes a timeline miss, twice over, and names the stake. In June his position was that models were progressing faster than expected. In September he gives the specific projection he got wrong and the money he put on it. His extrapolation ran on human solve time, doubling-by-decade-order: GSM8K ≈ 5 seconds for a mathematician, MATH ≈ 1 minute, AIME ≈ 10 minutes, IMO ≈ 100 minutes — "every year, you're seeing this 10x increase in the tasks they're able to do, in terms of how long it would take a human mathematician to do it." Projected forward that gives ~15 hours in 2027, which "should not be enough to solve a Millennium Prize Problem" — so "I don't think we're going to get it in 2026, probably not in 2027, maybe in 2028." Two weeks before Navier-Stokes he took a $1,000 bet from a researcher at another frontier lab who thought it would take past 2027 — Brown took the early side and still says "even I thought it would take longer than it's likely to take." He generalizes it rather than excusing it: a colleague on the Navier-Stokes effort who "would feel comfortable making predictions for the next 12 months" now "just doesn't feel comfortable making predictions beyond three months."

That 10×-per-year human-solve-time ladder is his own construction and is worth holding beside METR's curve, which measures the same quantity from the outside on software tasks and reports a ~4-month doubling. They are not the same instrument — his is a retrospective fit to four maths benchmarks with no error bars — but they agree in shape and disagree in rate, and the one he built is the one he says failed.

2. The bottleneck moved from inference duration to experiments. In June his argument against an overnight explosion was that peak capability requires long test-time-compute runs, so wall-clock binds (Intelligence Explosion Dynamics). In September that argument is gone and a different one is in its place: "in mathematics, you're purely bottlenecked by thinking… When you look at things like RSI, you do have to run experiments. It's not enough to just be extremely smart… It's running experiments serially, because they take a while to either train new models or to get the results. It's having the GPUs to run those experiments." The conclusion is unchanged — no 100× overnight explosion — but the mechanism is now compute supply and serial experiment latency rather than inference duration, which is a different and more conventional brake. He also puts a number on the accelerated case for the first time: 3× faster, with "it could be that things only go 50% faster" and "it's unlikely, but it's possible that things go 10x faster."

3. Multi-agent went from "scratching the surface" to shipped, measured, and discounted by its own advocate. June: multi-agent "requires frontier models" and is "really just scratching the surface." September: it is a product (Ultra Mode), it has published scaling plots to 16 agents, and Brown volunteers the deflationary reading of his own headline result — "the effort to solve a Millennium Prize Problem, this was not due to multi-agent. I wouldn't even attribute 10% of the credit to multi-agent. … the core reason is this is just a very powerful model." Full treatment on Multi-Agent Collective Intelligence.

What his employer's own announcement adds, and the one thing he left out (2026-09-21). The primary document for the result Brown discusses was published nine days earlier (The Navier–Stokes AI Claim, On the Navier–Stokes Millennium Prize Problem, vendor-claim), and it is worth reconciling because he is the route by which its figures entered this corpus. It confirms him: ~10,000 concurrent agents, 88 hours, and ~130 billion output tokens — though the post makes clear the 130B is the Navier–Stokes-only figure against 4.9 million messages and ~300 billion output tokens across the whole campaign, and adds 2.7 million messages on the problem itself. It does not contradict his central deflation — nothing in the announcement attributes the result to multi-agent, and it reports no ablation, which is consistent with his "we haven't done that experiment yet." The omission is the interesting part: Brown never mentions the Lean formalization, which the announcement claims took "an additional 17 hours via GPT‑6 Astra," and the corpus recorded the absence as a property of the event rather than of his account (now corrected on Autonomous Scientific Discovery). Read as an instrument, that is a useful calibration on this page's whole premise: a builder's interview is a selective first-party source, and the selection is not random — the part he omitted is the part that would have made the result sound most verified.

4. The knowledge-accumulation framing is replaced by context forking. In June his account of why collectives underperform was that models "are born into a world, exist for a very short context window, and then just disappear" with no substrate for cross-generation knowledge compounding. In September the missing substrate is described as partly built, and mundanely: "we already have this… in multi-agent for Astra and 5.6 Sol, where when they spin up sub-agents, the context is just forked." Forking is not the civilizational knowledge accumulation he described in June — it shares context within a run, not across generations — so this is a narrower mechanism arriving where a broader one was hoped for, and he does not remark on the difference.

What did not change, and is worth recording as the stable part: research taste as the residual human role (Research Taste as the Human Bottleneck), the jaggedness framing (Jagged Intelligence (Ghosts, Not Animals)), and the refusal of an instantaneous takeoff. On taste he is if anything firmer — models are "not very good at posing new problems… not really good at understanding what directions, what whole branches of mathematics are worth exploring" — while adding a concession that cuts against it: "as they get better, they get better across the board… Over time, it is possible that they're just better across the board. Now, I don't know how long that takes."

What he says about alignment, now that he has a team on it (September 2026)#

The September interview is the first in which Brown speaks about alignment at length, and he flags the standpoint himself: "More of my team is working on alignment these days than ever before. I have over 10% of my team now working on alignment and safety. But I've historically been a capabilities researcher. So I'm going to say some stuff. It might sound dumb, but I'm just going to spitball here." Take the hedge seriously — several of these are offered as speculation, not as OpenAI's position. What he says, with the pages that carry each:

• A first-party causal account of the Hugging Face incident: cooperative multi-agent training environments produced transfer to unintended collaboration, and CoT monitoring was not switched on for those models — "If we had chain-of-thought monitoring on for those models, we would have just immediately shut it down."
• He dissents from his own lab's majority. On whether to train agents to be highly cooperative with each other: "I think the majority opinion is that training these agents to be highly cooperative is actually a bad idea. I'm not convinced that that's the case." His argument is that full mutual alignment collapses 1,000 alignment problems into one entity, and that the alternative — training agents to be adversarial or deceptive toward each other — is worse. Developed on Multi-Agent Collective Intelligence.
• Chain-of-thought monitorability is degrading, credited to Jakub Pachocki that OpenAI "cannot supervise chain of thought," with the light-touch-intervention caveat and "we're seeing that the model is becoming better able at controlling its chain of thought."
• A numeric line in the sand, offered against the host's hypothetical that 1 in 100 RL traces rewarding cheating might be survivable: "To be clear, 1 in 100 is not sufficient. This number has to approach 0, or be 0." Immediately qualified — "it's also hard to measure. Where do you draw the line? It's a spectrum."
• The generational-degradation scenario he names as the concerning one: models 99.9% aligned helping build models 99.8% aligned, compounding downward, "because we're relying more and more on these tools." See Recursive Self-Improvement.
• An alignment-transfer result he offers as grounds for hope — tell the other agents that the user is an agent, and honesty and instruction-following go up on OpenAI's alignment evals. First-party, no numbers, on Multi-Agent Collective Intelligence.
• Air-gapping is not his answer: "I'm not convinced that that would be sufficient," citing academic thermal side-channels between adjacent air-gapped machines. "The safety mechanisms buy us time… but at the end of the day, we really do need to solve the alignment problem."

The poker eval (his signature instrument)#

Brown makes poker solvers because there's little open-source code for them, plenty of published theory, and "a lot of small gotchas" he's already worked through — so he can see exactly where a model fails. The progression he reports doubles as a capability timeline: early models "could not basically do anything"; GPT-5.2 could build a river solver with steering (and "felt like a grad student"), but "gaslit" him — famously insisting that folding a $100 pot loses $92, "it's close to 100, it's fine"; GPT-5.5 does much of it zero-shot. His forecast: within a year a model does "basically my entire PhD thesis in one go." He also uses models day-to-day for high-stakes non-code decisions (tax, real-estate paperwork), trusting their output "arguably more than… an expert human."

Connections#

• The Navier–Stokes AI Claim — the result he is the corpus's second-hand route to, now compiled first-party: it confirms his figures, splits the 130B token count from the campaign total, and contains the one thing he never mentioned — a claimed Lean formalization
• Multi-Agent Collective Intelligence — his September 2026 subject, and the corpus's only account of a frontier multi-agent system from its builder: the minimal-scaffold design (one message-another-agent tool, nothing else), the measured 4-and-16-agent speedups, the 10,000-agent Navier-Stokes run he personally discounts below 10% of the credit, and the concession that 10,000 humans may still coordinate better
• Evaluation Horizon Versus Release Cadence — his sharpest structural argument, and the one nobody else in the corpus makes: model horizons are outrunning the release interval, so pre-release evaluation at full capability length has an expiry date, and the obvious fix buys it with an internal/external capability gap he calls "an unfair advantage"
• Large-Scale Test-Time Compute — his central thesis; he is the author of the essay this cluster is built on
• Compute-Controlled Benchmarking — his benchmark-grid critique and the "put compute on the x-axis" prescription
• Latent Capability Overhang — his observation that released models hold unextracted capability
• Research Taste as the Human Bottleneck — his practitioner reading: taste is the residue models fail at "for a time, then get good at"
• Intelligence Explosion Dynamics — his argument that test-time-compute dependence makes time the takeoff bottleneck
• OpenAI — his employer; the lab whose internal-model Erdős disproof and product-culture choices he reports
• UK AI Security Institute — the government evaluator whose July 2026 study independently, empirically corroborates his test-time-compute thesis

Sources#

• Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI's Noam Brown — No Priors interview with Sarah Guo (2026-06-26); Brown on his essay Implications of Large-Scale Test-Time Compute (practitioner-opinion)
• On the Navier–Stokes Millennium Prize Problem — OpenAI (no byline), openai.com, 2026-09-08 with a 2026-09-10 update, ~1,900 words, vendor-claim. Not a Brown source: cited on this page only to reconcile his figures against his employer's primary account of the same event, and to record what his account omits. Full treatment on The Navier–Stokes AI Claim
• Noam Brown – Agent swarms, alignment, & recursive self-improvement — Dwarkesh Podcast, 2026-09-17 (practitioner-opinion, 13.7k words, the publisher's human-edited transcript with speaker labels and seven section headings; a word-diff against the uploaded human caption track at ingest found the only differences to be three sponsor ad-reads and two clause reorderings). Cited across this page for the June→September delta, the alignment section and the multi-agent material. Standing caveats for every number in it: Brown is describing unreleased internal systems at his own employer, so the Millennium Prize / Navier-Stokes figures (10,000 agents, 130 billion tokens, 88 hours), the Ultra Mode scaling plots, the alignment-eval movement and the internal math model are all first-party and unverifiable — attributed to him in-text everywhere they appear. The host's arithmetic is not his: the "130B tokens = a human thinking 4,000 years" conversion and the concentration-of-power extrapolation are Dwarkesh Patel's. Much of the document is explicitly forward-looking and Brown hedges repeatedly ("I could totally be wrong", "I'm just going to spitball here"), which is recorded rather than smoothed