Night Science 76 | Can Google’s Co-scientist project give scientists superpowers?
11 snips
Sep 8, 2025 
Alan Karthikesalingam and Vivek Natarajan from Google DeepMind delve into their groundbreaking AI co-scientist project. They discuss how this innovation mirrors human hypothesis generation while revealing its limitations. The duo highlights the intriguing idea of AI’s ‘tournaments’ of concepts and its quest to think beyond established facts. They also explore the potential for AI to generate novel research questions and the importance of maintaining human insight in scientific exploration, ultimately reshaping the future of discovery.
AI Snips
Chapters
Transcript
Episode notes
AI Augments Full Scientific Workflow
- The AI co-scientist is designed to assist each step of the scientific process from literature synthesis to experimental planning.
- The team built a multi-agent system to generate, evaluate, and operationalize novel hypotheses.
Tournaments Drive Idea Refinement
- The system uses tournaments and debates among agents to rank and improve hypotheses.
- This draws inspiration from DeepMind self-play but adapts competition into collaborative refinement.
User Test Led To Breakthrough Ideas
- Itai Yanai ran a five-day experiment by giving the system a one-page research goal and received ~30 detailed hypotheses back.
- Two hypotheses revealed unexpected mechanisms that reframed his thinking and suggested new experiments.