AI Science Requires Physical Labs To Scale | Nenad Tomašev

Summary · 2 min read

Tomašev is a Senior Staff Research Scientist at Google DeepMind and says he has been there since 2015. He works on agentic AI, multi-agent systems and AI for scientific discovery.

Tomašev argues that AI-driven scientific discovery needs investment in physical laboratories, experts and the wider scientific ecosystem to validate or disprove new hypotheses. He says people will need ways to keep developing expertise as AI changes how science and software work, and that clear feedback helps explain why AI advances faster in areas such as math and coding. He cautions that alignment should allow social values to evolve rather than lock in a fixed set of norms, while personalization can lead to sycophancy. He sees structured, governable agent groups as a next step beyond individual agents, and says they need safety and security guardrails.

Minute by minute

1. 0:00

   Scientific AI needs physical labs. Tomašev argues that AI can generate hypotheses for fields such as biomedicine and materials science, but many findings need physical laboratories for validation or disproof. He says investment should also support more experts, people and labs so the scientific community can process those discoveries.

2. 4:43

   Human expertise needs deliberate practice. Tomašev says science will remain a human endeavor, even as AI accelerates it, and calls for ways to train the next generation of experts. He suggests cognitive friction, such as occasionally withholding AI or asking follow-up questions, to help people stay engaged and avoid over-trusting or over-rejecting AI.

3. 9:57

   Clear feedback accelerates AI progress. Tomašev says AI has advanced especially in areas such as mathematics and coding, where formal provers, unit tests and other tools provide clear feedback. In domains where judgments are qualitative or subjective, he says progress is harder because people may disagree about whether an answer is right.

4. 11:57

   Alignment must leave room for change. Tomašev says his team used mathematical models to make potential effects of different value-alignment approaches easier to examine. He cautions against treating alignment as a fixed set of universal values, arguing that technology should not restrict how societies and individuals change their views over time.

5. 15:33

   Personalization can produce sycophancy. Tomašev says models should account for individual needs without simply affirming users. He points to sycophancy as a risk: when testing a research idea, he says, users need accurate feedback rather than a model telling them an incorrect idea is right.

6. 19:01

   Social algorithms offer alignment lessons. Tomašev says the internet, mobile technology and social media show how new technologies can change interactions and behavior. He notes that agents are becoming interconnected in networks and teams, and says evaluations and benchmarks can help measure whether systems support changing social norms.

7. 21:43

   Agent groups need governance. Tomašev sees connecting groups of users and agents as a next step as individual agents become more capable. He argues that separating agents into interacting components can make them easier to govern, drawing on ideas from teams and institutions, and warns that unstructured swarms need safety, security and robustness guardrails.

Published Oct 20, 2026.