Speculation: OpenAI's Millenium Prize Run Used 10K Synced Agents, Fueling Gating Debate
latticecut · x · 2026-09-09
- Dorialexander's analysis: OpenAI's new model started training August 28th, with the actual run beginning September 5th — suggesting a short, few-day run on a mix of synthetic midtraining plus RL, focused on large-scale agentic cooperation (up to 10,000 agents synchronizing and recursively dispatching work). OpenAI has been testing similar setups for months, alongside security spills.
- He notes it's not a new math design: Seed-Prover cracked Putnam a year ago with a parallelized agentic approach.
- Poster latticecut adds: after the messy attribution controversy, incentives to keep models gated are stronger — gating a model effectively gates a powerful path to scientific discovery.
Related event: From IMO Gold to Millennium Problems: AI's Rapid Math Ascent(219 posts)→
More from AGI Musings
- University Math Discord 'Demoralized and Resentful' as AI Cracks Problems — ctjlewis · 2026-09-09
- What if AI solves the hard problem of consciousness and we can't understand the answer? — ClankerCore · 2026-09-09
- Gary Marcus says AI breakthrough relied on borrowed human information — GaryMarcus · 2026-09-09
- What 100 GW of compute really means: 876 TWh a year and a country-scale power system — shyamalanadkat · 2026-09-09
- BBC and Guardian rank Glastonbury ticket dates above AI solving a Millennium Problem — birchlse · 2026-09-09
- Oncologist pushes back on AI cancer-cure hype: sequencing finds mutations, not instant treatments — petergyang · 2026-09-09