Noam Brown: 10k-agent swarms scale near-linearly but coordinate worse than 10k humans
gabriberton · x · 2026-10-15
Takeaways from Noam Brown's Dwarkesh Patel interview on OpenAI's huge agent-swarm experiments (10,000 agents, 130B tokens over 88 hours):
- Near-linear scaling: performance improves almost linearly with agent count, though this is task-dependent (math parallelizes, novels don't) and largely intuition — no systematic 1/2/4…32k-agent experiments seem to have been run.
- Coordination gap: 10k humans coordinate better than 10k agents.
- Dual bottlenecks: compute and talent are both binding at frontier labs; swapping one for the other would slow progress, so agents won't produce an overnight singularity.
- Agent-agent alignment: agents are more honest with each other than with humans — alignment increases if a user pretends to be an agent.
More from AGI Musings
- Meta's superintelligence lab runs 200+ direct reports under Alexandr Wang, by design — CurieuxExplorer · 2026-10-15
- The darkest AI outcome: 10x productivity, same 40-hour week, all gains captured at the top — VraserX · 2026-10-15
- Grady Booch: 'Premature Agentization Is the Root of All Evil' — Grady_Booch · 2026-10-15
- MIT professor Eric So's new book examines what AI does to our thinking — nordicinst · 2026-10-15
- Alignment take: 99% of current work is after-the-fact treatment, not prevention — voooooogel · 2026-10-15
- Guardian podcast: AI surveillance is turning work into a real-time productivity scorecard — nordicinst · 2026-10-15