I am these days a very irregular Codeforces visitor, but yesterday saw the post About Problem Coincidences by MikeMirzayanov and was wondering about a solution along the lines of a problem database like what was suggested some time ago by Um_nik. Lots of the comments there mention the idea of constructing a notion of problem similarity using solutions, which seems reasonable, but others countered that they don't want to have to write a whole solution just to tell if a similar problem exists. Given the current state of AI, could something like the following work now or in the near future?

1. For every problem statement in the CF, AtCoder, etc. universe, use AlphaCode or the editorial to get a solution.
2. Embed each of these solutions into a vector space. AFAIK this is straightforward to do.

Then when someone wants to see the most similar problems to their proposed statement, they either write a solution, or just write a statement and behind the scenes the database can have AlphaCode (or similar) generate a solution, embed using same method as in 2. and query for nearest vectors in the space. Of course, AlphaCode can't solve all problems, but maybe it at least writes something that "approximates" a correct solution well enough that the resulting vector is close to the "correct" one?

Of course, AlphaCode itself is not available to the public, but probably it or something similar will be soon. Thoughts?