To be fair a lot of the impressive Elo scores models get are simply due to the fact that they're faster: many serious competitive coders could get the same or better results given enough time.

But seeing these results I'd be surprised if by the end of the decade we don't have something that is to these puzzles what Stockfish is to chess. Effectively ground truth and often coming up with solutions that would be absolutely ridiculous for a human to find within a reasonable time limit.