But it does produce value. We now have an explicit solution and even a proof. Humans can then work on clarifying why it's true. Unsurprisingly, not all that different from software, where models can generate working code just fine. The details will all be there and all correct, but the architecture is currently not ideal, so a human guiding it can greatly improve the proofs.

I actually found this to be the case with some basic linear algebra notes I was recently doing in Lean (without using mathlib). The model could generate working proofs, but they obscure the basic ideas (actually I wonder somewhat if this is because the Lean code that's out there to train on doesn't make a huge effort to read like textbook proofs, which was my motivation in the first place). I give it a skeleton of a couple lines of `calc`, letting it fill in the reasoning for each line, and it does much better. Then ask it about making some macros to simplify "trivial" or "obvious" things, and it does even better. etc.

I suspect there's a good workflow where a big SOTA model makes an impenetrable proof (or code) and then a human works with a FIM model to simplify it (with the larger gnarly proof right there in context for FIM), but unfortunately everyone seems to only care about agents right now.