Noam Brown: lab results show scaling agent swarm size improves performance — possibly a new scaling law
Afinetheorem · x · 2026-09-09
- Noam Brown confirmed lab results on scaling agent network size exist in the GPT 5.6 and Astra blogs, though only for relatively small numbers of agents; larger-scale science is expensive but in progress.
- He notes swarm scaling is less efficient than serial CoT, but the gap depends on the technique.
- Kevin Vallier argues this is among the first lab evidence that scaling agent count scales performance — mirroring how humans scaled intelligence via cooperative networks, and potentially a new scaling law.
Related event: OpenAI Researchers See New Scaling Law in Parallel Agents(2 posts)→
More from AGI Musings
- From GPT-4 failing basic addition to 10,000 agents solving a Millennium Prize problem in 3.5 years — MartinSignoux · 2026-09-09
- If AI proved P=NP, it would imply insights humans lack, argues jessi_cata — jessi_cata · 2026-09-09
- Roko's Bet: If AI Solves P vs NP Soon, It'll Be P=NP Via an Ugly Algorithm — jessi_cata · 2026-09-09
- "We're the First and Last Generation to Code for a Living" — tlakomy · 2026-09-09
- Plan alleged to lure academics with limited AI compute, then run their ideas in-house — ns123abc · 2026-09-09
- Brundage: my critique indicts the industry and policymakers, not just OpenAI — Miles_Brundage · 2026-09-09