Are LLMs Solving Math?
Are LLM agents solving all known math problems from low to high complexity consistently now or are LLMs "only" synthesizing solutions for ripe problems given humanities current math knowledge (by synthesing existing knowledge via analogy - running experiments at massive scale)?
LLM intelligence is opposite to how we normally think of (human) intelligence: very grounded, methological thinking based on first pricinples. It is like having someone who knows a lot but is slightly drunk throwing out ideas. But with enough guidance and verification you can extract useful output. - paraphrased from Terence Tao interview.
I can see LLMs marching on to solving math only by creating a loop of continual learning - where experiments produce (verified) knowledge that is used to guide the next round of experiments.
Value of human understanding?
Tao also emphasizes that LLM written proofs show lack of human understanding: the proofs stretch trivial (uninteresting) parts too long and keep the most interesting parts too short. LLMs lacking this ability, may indicate that truly autonomous AI scientists are still further away than AI labs like to think..