Debate Erupts Over Agent Swarm's Contribution to Navier-Stokes Proof
A technical dispute has broken out in the research community over how much credit agent swarms deserve in the "AI discovers Navier-Stokes mathematical proof" story, centered on a judgment from OpenAI researcher Noam Brown: the swarm components may have contributed only about 10% of the discovery.
Confirmed
- Elie Bakouch relayed details of his exchange with Noam Brown: Brown believes the agent swarm's actual contribution to Navier-Stokes-type discoveries is low, with the corrected phrasing being "the contribution may be only about 10%."
- Bakouch also cited data showing that as agent count grows, parallel test-time scaling at the same budget is less efficient than serial scaling and has higher latency; he emphasized this is only data for 1→5 agents and cannot be extrapolated to 1→10,000.
- Counterpoints from scaling01, ChaseBrowe32432, and others: without scaling parallel TTC to 10,000 agents, the proof would never have appeared at all. Even assuming a single serial agent is 100x more efficient than a swarm, it would still take about a year to generate the tokens needed to complete the proof.
- scaling01 also extracted data from the disputed chart: the internal model gains about 0.17 logits per compute doubling, while Astra delivers about 0.18 logits/doubling — meaning parallel scaling contributes roughly twice as much as switching to a new model; he noted the chart's x-axis spans only about a 30x compute difference, yet was used by the other side to argue "one model-generation upgrade beats 10,000x parallel compute."
- ChaseBrowe32432 extrapolated model-generation progress from this: one generation later, about 16 sub-agents running 88 hours could solve the NS-related benchmark; two generations later, a single agent could handle it within 88 hours.
Unconfirmed
- "The swarm contributed about 10%" is only Brown's personal estimate, and the other side's token-generation-time and logits calculations are their own derivations — neither side has released complete raw data, so no authoritative conclusion exists.
Why it matters
- The dispute cuts to the heart of test-time compute scaling: whether serial depth or parallel breadth is the better deal. If 10,000-scale parallel agents really are the key, then massive agent swarms are a viable path to major scientific discoveries today; if model-generation upgrades are more efficient, resources should go into base model iteration instead. The conclusion will shape labs' compute allocation strategies.
2026-09-18 ~ 2026-09-19 · 11 related posts
- Episode 1: Debate Erupts Over Agent Swarm's Contribution to Navier-Stokes Proof(2026-09-18, 11 posts)
- Episode 2: Noam Clarifies Multi-Agent Systems Contributed Under 10% to Millennium Problem(2026-09-19, 3 posts)
Primary sources
- Noam Brown: agent swarm likely contributed just ~10% to OpenAI's Navier-Stokes discovery — scaling01 · 2026-09-18
- Insider rebuts claim that agent swarm contributed only 10% to Navier-Stokes proof — scaling01 · 2026-09-18
- Rebuttal: A Single Agent Would Need ~a Year to Generate the Tokens for the Proof — ChaseBrowe32432 · 2026-09-18
- [source] Navier-Stokes Proof Debate: How Much Did 10,000-Agent Parallel Scaling Contribute? — scaling01 · 2026-09-18
- Researchers Clash Over How Much a 10,000-Agent Swarm Actually Contributed — eliebakouch · 2026-09-18
- Sequential test-time scaling beats parallel multi-agent scaling at equal budget, data shows — eliebakouch · 2026-09-18
- [source] Blogger's logit math shows parallel 10,000x scaling beats the Astra model gap by 2x — scaling01 · 2026-09-19
- Scaling debate: 30x on the plot can't prove model gains beat 10,000x parallel compute — scaling01 · 2026-09-19
- Model generation gains could soon let a single agent solve NS in 88 hours — ChaseBrowe32432 · 2026-09-19
- [source] Noam Brown says the agent swarm contributed under 10% to the Navier-Stokes discovery — eliebakouch · 2026-09-19
1 near-duplicate retellings: scaling01