OpenAI Researcher: Large-Scale AI Science Is Costly and Less Efficient Than Serial CoT Scaling
polynoamial · x · 2026-09-09
Responding to Nathan Lambert, an OpenAI researcher shared that results with GPT-5.6 and Astra blogs cover only a small number of agents, that larger-scale science remains very expensive, and that parallel agents are less efficient than scaling serial CoT—though the gap depends on the technique. Useful first-hand context on the economics behind the Navier-Stokes effort.
Related event: OpenAI Researchers See New Scaling Law in Parallel Agents(2 posts)→
More from AGI Musings
- AI will hand us breakthroughs we can't understand, like alien answer sheets — sterlingcrispin · 2026-09-09
- Mathematician littmath calls to abolish science prizes, citing the fraught Poincaré conjecture saga — littmath · 2026-09-09
- Sam Altman: OpenAI will hunt for room-temperature superconductors with 1000s of AI agents — Dr_Singularity · 2026-09-09
- X debate: is tasking an agent swarm to 'solve alignment' the worst idea ever? — jd_pressman · 2026-09-09
- Matthew Berman: more nervous about AI than ever in 3+ years of coverage — MatthewBerman · 2026-09-09
- Employees Who Ignore AI May Be the First to Be Laid Off — kevinsurace · 2026-09-09