It turns in top-level performance on original, out-of-distribution problems given in international math and programming competitions, but it's "not comparable to a human level." Got it.