How do humans succeed in tasks like proving Fermat's Theorem or predicting the Higgs boson?

I discuss issues of inverting feasibly computable functions, optimal discovery algorithms, and the constant overheads in their performance.