I was more thinking of Github Copilot, but ChatGPT is similarly good at programming. Programming is arguably even less error-forgiving than providing a written

I was more thinking of Github Copilot, but ChatGPT is similarly good at programming. Programming is arguably even less error-forgiving than providing a written answer to math problems (I was more thinking of problems that ask for an analytical rather than numerical answers, which is also what most university-level math courses ask for). I wouldn't have expected those models to perform nearly as well at programming as they do, so I've updated my prior. Of course you'd need to find a math corpus of a similar size to github's repositories first. Overleaf has a bit of a similar position to github in STEM research, so that would be an interesting place to start looking (if it's legally possible). It's probably still too small, but would be interesting to see what a GPT version trained on the Overleaf corpus could do.