BioML Seminar 4.3 - Vivek Natarajan on Advancing Science and Medicine with Collaborative AI agents

BioML Seminar 4.3 - Vivek Natarajan on Advancing Science and Medicine with Collaborative AI agents

by Vivek Natarajan

Formal & Physical Sciences AI — News & Applications UYComputer scienceUYQArtificial intelligence (AI)
🎙 Vivek Natarajan 👥 14K 📅 June 26, 2026 ⏱ 81 min 👁 38 📄 expert opinion 🧭 2026-08-16
Available in: English (current) Français

Keywords

AI co-scientistAMIEself-playreinforcement learningscientific discovery

Summary

In this seminar, Vivek Natarajan presents Google DeepMind’s efforts to build collaborative AI agents for science and medicine. He introduces the AI co-scientist, a multi-agent system based on Gemini, designed to generate and refine research hypotheses. The system uses self-debate among agents, scaling test-time computation rather than gradient-based learning. It accepts research goals in natural language and outputs research proposals. Natarajan highlights validated results, including drug repurposing for AML, targets for liver fibrosis, and a gene transfer mechanism. He also discusses AMIE, an AI co-physician that outperformed primary care physicians in simulated settings and shows promise in real-world validation. The talk covers the system’s architecture, the importance of generality, and the role of tools. Natarajan emphasizes that current models have ‘bookish’ knowledge and need real-world interaction for further progress. He draws parallels to AlphaGo’s self-play and the bitter lesson. The seminar includes Q&A on agent definitions, world models, and the balance between harness and model intelligence.

156 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the design and philosophy behind DeepMind’s AI co-scientist and AMIE. Natarajan articulates a clear thesis: general-purpose AI systems can accelerate science by scaling test-time computation and using self-debate, mirroring human scientific reasoning. He supports this with concrete examples of lab-validated results, lending credibility. The argumentation is coherent, addressing potential limitations such as imperfect world models and the need for real-world feedback. However, the talk is more of an overview than a deep technical exposition, and some claims are presented without detailed evidence. The Q&A section adds depth, clarifying agent definitions and the role of tools.

Scientific Rigor, Source Quality, Title Accuracy

The speaker is a credible authority, leading research at Google DeepMind. He references published work in Nature, Nature Medicine, Cell, and Advanced Science, which are high-quality sources. The title accurately reflects the content. The talk is a seminar, so it is not a peer-reviewed presentation, but the speaker’s expertise and the cited publications support its scientific rigor. No comments were provided for analysis.

179 words

Title / Content Match

The title accurately reflects the content: a seminar on AI agents for science and medicine, presented by Vivek Natarajan.

Quality & Reliability

8/10

The speaker is a leading researcher at Google DeepMind, directly involved in the projects discussed. The talk presents credible, published results (Nature, Cell) and transparently discusses limitations. However, it is a seminar talk with limited technical depth and no independent verification.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • AI co-scientist paper (Nature) — Mentioned as published results of the AI co-scientist, including drug repurposing and target discovery.
  • AMIE paper (Nature) — Mentioned as published results of AMIE outperforming primary care physicians.
  • Med-PaLM paper (Nature) — Referenced as prior work by the speaker.
  • Med-PaLM 2 paper (Nature Medicine) — Referenced as prior work by the speaker.
  • Cell paper on gene transfer mechanism — Mentioned as a lab-validated result of the AI co-scientist.
  • Advanced Science paper on liver fibrosis — Mentioned as a lab-validated result of the AI co-scientist.

Concurring Sources

  • AlphaGo paper — Supports the self-play and reinforcement learning approach.
  • The Bitter Lesson — Supports the argument for scaling computation over domain-specific methods.

Contribution & Novelties

The talk presents DeepMind’s approach to collaborative AI agents for science, emphasizing self-debate and test-time computation scaling. It offers a novel perspective on generalizing self-play to scientific discovery, with promising validated results. The discussion on the limitations of current models and the need for real-world interaction provides valuable insights for future research.

Pour aller plus loin :

  • AlphaGo paper — Foundational work on self-play and reinforcement learning.
  • The Bitter Lesson — Essay by Rich Sutton on the importance of computation and search.
  • Med-PaLM 2 — Prior work by the speaker on medical AI.
  • AMIE — Related work on medical AI agents.
  • AI co-scientist — Blog post by Google Research on the AI co-scientist.

113 words

Radar Profile

The radar profile shows high scores in quality and reliability, reflecting the speaker's expertise and published results. The lower score in technical depth indicates the talk is accessible but not highly technical. Overall, the profile suggests a credible and informative seminar.

Reliability 8/10