Letting programs learn through “open-ended” experimentation may unlock remarkable new capabilities, as well as new risks.