Google DeepMind's AMIE & Co-Scientist: AI Beats Doctors, Makes Discoveries
Google DeepMind researchers Vivek Natarajan and Anil Palepu join the show to discuss two major papers published in Nature. The first covers AMIE (Articulate Medical Intelligence Explorer), which has now extended beyond diagnosis to outperform primary care physicians on treatment recommendations grounded in clinical guidelines, and is surpassing medical fellows in cardiology and oncology while remaining complementary to attending physicians. The second introduces Co-Scientist, a multi-agent AI system that succeeded on three open-ended scientific challenges—drug repurposing, therapeutic target identification, and bacterial drug resistance—independently converging on the exact drug-resistance mechanism that Google's academic collaborators had discovered experimentally but not yet published. The researchers emphasize that neither project relied on heavy custom fine-tuning; instead, both are built from general-purpose Gemini models via structured agent scaffolding, long-context reasoning (up to ~256K–1M tokens), tool use (web search, AlphaFold), and tournament-style pairwise evaluation of candidate hypotheses. They highlight three practical lessons for builders: structured/JSON-enforced reasoning outperforms free-form chain-of-thought, injecting new information or entropy (via search, tools, or human input) is essential to prevent mode collapse during extended self-critique loops, and Elo/tournament-style ranking is currently an industry-best practice for surfacing top ideas. Deployment is underway: AMIE is entering a clinical trial with real patients at Beth Israel Deaconess Medical Center (Harvard), and Co-Scientist is being rolled out through a trusted-tester program already engaging ~100 leading scientists, with ambitions for millions of users by year-end. The guests describe the path to AI doctors and "data-center geniuses" as increasingly an engineering challenge rather than a fundamental unknown, though they note incentive and benchmark-tradeoff frictions around deeply integrating specialist scientific modalities into frontier models.
Original sources
Podcast The Cognitive Revolution
YouTube The Cognitive Revolution
Video:New in Nature: Google Agents Beat Human Doctors, Make Scientific Discoveries – With Vivek and Anil ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.