General
AI assisted problem solving - emerging usecases
Terence Tao just said something revolutionary. He is one of the best mathematicians alive.He called “AI-assisted literature reviews” the most promising short-term use of machine intelligence in mathematics. Not great at theorem-proving (yet), but connecting dots across millions of papers, surfacing forgotten proofs and unnoticed connections. In one case, six long-standing Erdős problems turned out to have been solved years ago — the answers were just buried in the noise.
That's amazing!
https://www.reddit.com/r/math/comments/1o8xz7t/terence_tao_literature_review_is_the_most/
I think this comment summarizes it best, keep in mind that the progress is rapid:
Yes it is good. Not all parts of the AI bubble are complete bogus.
What's happening here is that it is now (to some extent) possible to map documents into a vector space that reflects their meaning. This solves the hard problem with search, namely finding documents that talk about something but you're not sure in precisely which terms.
This makes LLMs quite good at finding relevant passages and translating between formats. You could probably show superhuman performance in either of those areas, though humans might still be better at some parts.
It is less good at true creativity, creating new texts with novel ideas. Even symbolic reasoning seems to be tricky at the moment, though it's getting better.
I was talking to someone at a conference about this the other day. They pointed out that if you ask these LLMs a research question they almost always give you wrong answers but they throw up the right keywords so that you can refine your search. My experience has been similar.