Can AI Discover Physics?

The optimistic case says that a sufficiently large model, trained on the right data, will re-derive Newton, then Einstein, then whatever comes next. The pessimistic case says taste — the choice of what to look at — is a distinctly human act, and taste doesn’t emerge from imitation.

Both cases assume discovery is a single act. It isn’t. It’s a long sequence of small acts: noticing an anomaly, choosing a formalism, discarding a beautiful theory when reality disagrees.

The interesting question isn’t whether AI can discover physics. It’s which of those small acts we’re willing to delegate, and which we’ll keep for ourselves.