Expanded afternoon edition — September 20, 2026.
In the September 17 Dwarkesh Podcast, Brown attributes less than 10% of OpenAI’s reported mathematics achievement to the multi-agent method itself, emphasizing the underlying model and long-horizon work. He is candid about the missing comparison: published scaling evidence reaches roughly 16 agents; a much larger successful run cannot establish how many agents were necessary. Exact episode and transcript
The interview’s opening 15 minutes are the best entry into an approximately 80-minute conversation. Beyond raw scale, it raises an appealing design question: how can a group expose mistakes early and preserve disagreement long enough to learn from it? A final answer alone tells us little about whether collaboration contributed. Failed approaches, changed minds and discarded work would make the demonstration much more informative.
That is a substantive reason to listen beyond the enormous agent count. Later sections consider AI organizations and the relationship between mathematical progress and AI research. The discussion includes speculation, with enough explicit uncertainty to give it some resistance.