An ‘AI Scientist’ Is Inventing and Running Its Own Experiments Letting programs learn through “open-ended” experimentation may unlock remarkable new capabilities, as well as new risks.