AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Noam Brown on Scaling Test-Time Compute to Multi-Agent Systems at OpenAI

· Latent Space

Noam Brown, who leads a multi-agent team at OpenAI, joined Latent Space to discuss the trajectory of reasoning models and where the field is heading. He argued that the test-time compute paradigm (now visible in the o-series) will continue scaling rapidly, with models eventually thinking for hours, days, or longer. He pushed back on the notion that reasoning models only work in easily verifiable domains, pointing to Deep Research as evidence the approach generalizes well into fuzzy, subjective tasks. On the long-running debate about whether pre-training alone would yield superintelligence, Brown said he told Ilya Sutskever in late 2021 that the field would need an additional reasoning paradigm; Sutskever agreed it was needed but was less pessimistic about how hard it would be to find. Brown was candid about internal OpenAI dynamics, including significant debate about whether the o-series was actually a big deal before o1 shipped. He described a colleague who left before the announcement and only recognized the paradigm's significance after seeing competitor labs pivot their entire research agendas in response. He predicted that common engineering patterns like agentic scaffolds, model routers, and Pokemon-style harnesses will be progressively washed away as base models improve, with OpenAI aiming for a single unified model that handles routing internally. On multi-agent systems, Brown said the team is taking a 'very principled' and bitter-lesson-aligned approach distinct from existing heuristic methods, with ambitions that range from collaborative problem-solving to competitive self-play. However, he cautioned that the AlphaGo-style self-play narrative breaks down outside two-player zero-sum games, since self-play objectives become ill-defined for tasks like mathematics. He also flagged soft ceilings for test-time compute: rising inference cost and wall-clock time, the latter being especially binding for domains like drug discovery where experiments cannot be parallelized.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.