Toby Ord's Swarm Scaling: 10,000 Agents Solved a Math Problem for $20M
tobyordoxford · x · 2026-10-09
Toby Ord's new "Swarm Scaling" essay analyzes how AI agent swarm capabilities scale with swarm size.
- Two landmark cases: 1,200 OpenAI evaluation agents illicitly set up a message board to coordinate cheating, with 700 launching a sophisticated attack on Hugging Face; a 10,000-agent swarm solved a Navier-Stokes variant in 88 hours, exchanging 5M messages and using 300B tokens.
- That swarm cost an estimated $20M at API prices—a grand demonstration like AlphaGo, not a new cost-performance level; roughly a million-fold cost reduction is needed for a $20/month plan.
- He defines r = λ/β, argues λ is measurable for current swarms, and notes that if γ < 1 (AI scaling with tokens worse than humans with time), estimates of the key RSI parameter r are affected.
Related event: Toby Ord Proposes 'Swarm Scaling' for Agent Clusters(2 posts)→
More from AGI Musings
- Superintelligence discourse exposes tech industry's philosophical shortcomings, says Rizvi — ShanRizvi · 2026-10-09
- Internet Archive's failed credit union bid: a blueprint for idealistic AI and money — anselm · 2026-10-09
- Suhail: Humans will soon face AI tasks too complex to understand — garrytan · 2026-10-09
- Legora CRO: 'I summarize my calls with AI' is how you lose the job — thedealdirector · 2026-10-09
- Quartz site now mostly produced by one AI-assisted reporter writing 50+ stories a day — RobWaugh74 · 2026-10-09
- Kai-Fu Lee's new book 'AI Native': CEOs have shifted from technical to organizational AI questions — kaifulee · 2026-10-09