Noam Brown: multi-agent gave under 10% of the millennium problem win

AI寒武纪 · wechat · 2026-09-20

A recap of Noam Brown's 80+ minute interview on OpenAI's multi-agent direction.

A single task consumed 130B tokens of reasoning — roughly 4,000 years of one researcher's thinking — completed in 88 hours. Multi-agent systems are fundamentally about parallelizing test-time compute rather than stretching chain-of-thought serially.
With 4 agents, some math benchmarks ran nearly twice as fast at only 2x compute cost; scaling to 16 shows mild sublinear returns. Parallel gains depend on task type: math, code, and deep research parallelize well; long-form writing doesn't.
Brown insists the multi-agent architecture contributed less than 10% to solving the millennium prize problem — the decisive factor is the underlying base model's long-horizon reasoning.
OpenAI dropped centralized orchestrator topologies entirely: each agent gets only a basic messaging tool, and coordination rules emerge from RL training, producing human-like asynchronous collaboration resembling mathematicians debating on Slack.
Math capability leapt from GSM8K (2024) to an IMO gold medal (2025) to a millennium problem (2026) — far ahead of the 10x-thought-per-year trend that predicted 2028+.
Brown sees this transferring to recursive AI R&D self-improvement: top researchers already spend $7,000–8,000/day on Codex internally, with 3x faster development cycles, though real-world GPU experiment latency remains a bottleneck against exponential takeoff.

More from AGI Musings