The papers created by AI on math, right now, are torturous reads. Figuring out what is actually going on in them, is the work of months or even years.

I’ve seen multiple conference submissions that read like English, but take hours of effort to disentangle and figure out what they mean.

There is a limited number of known interesting problems that are worthwhile to engage in. New ones are found and understood, but it is a slow process and it is dependent on understanding and verification, not generation of content.

Saying AI will create new ones, strengthens the strip mining analogy. The idea that AI will find other problems parallels the idea that firms can cut down a forest and plant them elsewhere.

The conversation on HN is driven by non mathematicians, to whom math is a subject, like “nature”. It’s not their core business. The original post by Tao is a better source of information on what he means.

It can't consistently produce new, interesting-to-humans conjectures faster than it solves open problems, no. AI-generated natural-language proofs are more or less human-unreadable and are solely concerned with the process of constructing whatever object serves as the counterexample. Natural-language counterparts of formally verified proofs are often unfaithful to the formal component. All of this combines to create less-than-fertile ground for future study.

Does creativity need sentience?