Most k-sat problems are easy, or can be reduced to a smaller and easier version of k-sat. Neural networks might be able to write Lean or Coq to feed into a theorem prover.