@tao@mathstodon.xyz How much of this do you think is an existential threat to mathematics as we know it, and how much is a relatively straightforward fine-tuning problem that could be addressed with another round of RLHF? This situation reminds me of complaints people had with early coding models: they produced sloppy code, they were "lazy," they had poor comments, and in general they tended to produce a vibe-coded ball of wax that "works for no reason." This seems almost exactly parallel to that. With software development, some the answer was simply to train better models and teach people how to use them more effectively. Do we expect the same with math? Thanks, and great essay.