John Platt explains the role of artificial intelligence in scientific discovery

Summary

In the Latent Space podcast, John Platt (inventor of SVMs, Oscar winner) discusses Google's ERA system. This LLM-based, MCTS-inspired evolutionary optimization tool automates 'scoreable' scientific problems. The interview also touches on the system's applications in climate science and beyond.

Why it matters

ERA offers a concrete, deployable AI-assisted discovery model that goes beyond chatbot Q&A by focusing on iterative code optimization. Its success in solving a counterfactual problem regarding condensation trails over two years suggests that AI can contribute to high-risk, multi-variable science in the near term. The discussion on Goodhart's law and the distinction between predictive and descriptive models serves as a current safety measure for researchers using optimization-heavy AI. Platt's deep domain expertise and insistence on 'climbing the mountain' provide a useful counterbalance to the prevailing 'write a prompt and publish' culture.

Highlights

• ERA's UCB-guided notebook mutation (a close relative of MCTS) showed a sharp leap between Gemini 2.0 and 2.5. The system moved from being non-functional to a level capable of producing 10 or more papers.
• Condensation trails contribute approximately 1% to human-caused warming. ERA solved a counterfactual heat retention model that had stumped the team for over two years.
• Platt's core warning: ERA is 'a powerful tool that can cut your fingers.' Predictive models are not automatically descriptive, and reward hacking (Goodhart's law) is a real risk.

Critical take

The text reads more like a letter of admiration than a critical analysis. It does not discuss ERA's failure scenarios, computational costs, or how well the 'scoreable task' framework generalizes beyond well-defined optimization problems. The 'Automatic Kaggle' framing may be overstated, as real science often lacks a clean, agreed-upon scoring function. This is precisely the difficult part Platt acknowledges. This indicates that the system's upper bound is more uncertain than the enthusiasm implies.