There's a lot of work showing that we can reliably get to or above human level performance on tasks where it's easy to sample at scale and the solution is cheap to verify.