Terry Tao is one of the staunchest knowledgeable advocates of GPT models in mathematical research, and afaik he doesn't even bother trying to use the models for proof search. It's like trying to build a house with a box of shoes: sure, the shoe is technically more versatile because you can't use a hammer for tightening bolts (the shoe's sole has enough friction to do this) or foot protection (the shoe is the right shape for this) or electrical isolation (the bottom surface of the shoe is largely rubber), but please just use a hammer if you want to manipulate nails.
Somewhere inside openai is a reinforcement learning loop that looks like:
current_code = read(code_base_from_file.txt)
modify_prompt = "some prompt to modify code with expected outcome"
result = model.run(modify_prompt, current_code)
if check(result):
provide_positive_feedback(model)
else: provide_negative_feedback(model)
and its clear that not only does this work, it maybe the future of what unravels software engineering.the current models are being trained for coding, but theres no reason this couldnt be tried for other domains, like pure math.