Comment by JonChesterfield

The klever idea of using llms instead of superoptimisers is to drop the correctness requirement. You get some shorter code that does a different thing, maybe it'll still do something you like.