When AI is a better prover, perhaps our most important job is deciding what is worth proving.

It appears any theoretical domain will be dominated by AI, including theoretical physics, if not already. Those that will fall sooner are the domains with self-contained verifications done by computation rather than experimentation.

Maybe that is where our work moves next: creating questions for AI to solve.

The important questions may not be sitting in the training data, presumably challenging for modern AI.

But perhaps “not in the training data” is not enough to make a question beyond AI.

In a world where answers are monopolized by AI, can we still be the one defining questions?