I feel there may be an important clue here, at least with regard to the programming example. Peter asked Kevin Wang for some insights into his approach to progr

I feel there may be an important clue here, at least with regard to the programming example. Peter asked Kevin Wang for some insights into his approach to programming competitions, and one was this: [Kevin] I think specifically for AoC it's important to just read the input/output and skip all the instructions first. Especially for the first few days, you can guess what the problem is based on the sample input/output. [Peter] Kevin is saying that the input/output examples alone are sufficient to solve the easier problems; in the interest of speed he doesn't even read the problem description. From what I have read elsewhere, it seems that AlphaCode generates a great many candidate programs, and uses the given test cases to eliminate those which fail them. I wonder if that alone (without any reasoning from the problem statement to a solution) can account for its success rate on these challenges; Kevin's observation suggests this might be plausible. On the other hand, if the questions posed to AlphaCode are essentially as we see them, then its ability to use the test cases to screen its candidates seems impressive to me, just by itself (if it is a result of training, rather than an explicitly programmed behavior.)