#...probably current ones would also need finetuning on ~examples of a mathematician making an error catching it and recovering from there
Text
Hypothesis: LLMs are bad at higher math (in particular novel proofs), and will stay bad at it, because higher mathematicians virtually never put to words the process of generating such proofs
#i think llms should be capable of it if given examples of ~verbal planning and verbalized attack on the theorem#smth smth proof discovery is hard but not that hard and you _can_ write down an approximate algorithm in human words#which will not get all theorems but will get a whole lot of them#and llms are not bad at faithfully executing instructions in human words#...probably current ones would also need finetuning on ~examples of a mathematician making an error catching it and recovering from there#it's just that almost nobody ever writes these kinds of texts on the internet#(and of course this will do nothing to the 'bullshit insight' proof technique)#(and there's the hard problem of discovering good concepts/definitions)#(but these could be tackled separately)#posts
8 notes
·
View notes
Last Seen Blogs
emo-lover77
Rayorak
28
(4/2/20) ❤️
benthemusicalbeard
Ben the Musical Beard
benjiitucnt
แค่ผ่านมาบ่น
laurenlandau
Untitled
spectralreplica
Supernumerary Rainbow