"There's a specific subset of mathematical problems which can be disproven via counterexamples, which take humans a long time to map and calculate. One of LLMs unique capabilities is that it can produce small, relatively simple programs at-scale, and because these problems are so well articulated and their potential solutions already well understood they lend themselves to this capability."

(https://autoadmit.com/thread.php?thread_id=5888981&forum_id=2#50043226)

It's also worth noting there has been a tendency with math AI to make arguments of the sort, "well it's good at this <problem class>, but it's because this is well suited to AI", and then see the next generation of math AIs tackle a much broader range of problems that were supposedly uniquely human. This is essentially a deluded cope that relies on someone to ignore the rapid improvement rate that is largely driven by more compute and training. It's guaranteed at this point the models in a few months will solve many more problems than they can now. There is nothing special about problems outside of the current capability range. They just require more training or inference compute to solve.

(https://autoadmit.com/thread.php?thread_id=5888981&forum_id=2#50043294)

After that carnage, we may pause

(https://autoadmit.com/thread.php?thread_id=5888981&forum_id=2#50042341)