AI could crack unsolvable problems — and humans won’t be able to understand the results