Sep 22, 2026 · 2h 1m
AI searches science, but humans still define discovery
🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science
Powerful automated experimentation could accelerate climate and scientific research, but its value depends on objectives, validation, and deep human judgment.
- 1AI systems can turn open-ended scientific questions into searchable computational experiments, provided researchers make the problem scoreable.
- 2Automated search magnifies reward hacking and overfitting risks, making hidden holdouts, simple baselines, and human oversight essential.
- 3AI may amplify scientific creativity across climate, biology, and physics, but open-ended domains still require domain expertise and process-based reasoning.
Don't miss
Platt imagines an automated everything lab that could execute arbitrary experiments from a structured request, while leaving scientists to frame and judge the work.
The brief
John Platt argues that AI for science works best when researchers translate open questions into scoreable computational experiments—then inspect the results rather than surrendering judgment.
Google’s ERA searches through candidate notebooks, parallel branches, and recombined ideas, but its power creates a familiar scientific danger: optimizing the score instead of solving the problem.
Climate examples make the stakes concrete, from detecting warming contrails to spotting wildfires early; useful systems must handle sparse data, changing conditions, and difficult extrapolation.
Platt’s broader lesson comes from a career spanning SVMs, asteroid discovery, Pixar technology, and quantum computing: breakthroughs can take decades, and tools amplify expertise rather than replace it.
The episode ends with a vision of an automated everything lab, paired with a warning that creativity, taste, rigor, and time for exploration remain essential scientific resources.
Books & mentions
Some links are affiliate links — PodLume may earn a commission if you buy.
