I still really doubt the ability for LLMs to develop new knowledge and experiment to prove its accuracy. They do a below average job at spitting out existing knowledge but creating new information that isn’t already on Wikipedia or elsewhere on the internet is something else beyond their scope.
There are certain types of problems they are better at solving. Basically if we already have all of the needed information, but you need to synthesize a huge amount of it into the final answer, they’ll be excellent at that kind of issue.
True, they can be a useful tool. One of the big things that really annoy me is how they are being pushed as a solve everything miracle rather than just a tool that is useful in certain cases.
I guess that isn’t as profitable and they have to push it on everyone for things it does poorly just in the name of forcing widespread adoption.
All of the needed words. Gotta be careful here. LLMs don’t deal in information. But yes, they are good at stringing together tokens if that’s all you need.
While I had similar concerns, I do think about this a fair bit now. A nontrivial math problem was proved by Aristotle (in Lean, a proof assistant, so we’re relatively sure it’s correct). Alpha Evolve then generalized. GPT pro did some writing and visualization work.
It’s basically the quality of work I aspire to. Slightly cheaper, and substantially faster. Not the strongest PhD, but a solid graduate student.
I still really doubt the ability for LLMs to develop new knowledge and experiment to prove its accuracy. They do a below average job at spitting out existing knowledge but creating new information that isn’t already on Wikipedia or elsewhere on the internet is something else beyond their scope.
There are certain types of problems they are better at solving. Basically if we already have all of the needed information, but you need to synthesize a huge amount of it into the final answer, they’ll be excellent at that kind of issue.
True, they can be a useful tool. One of the big things that really annoy me is how they are being pushed as a solve everything miracle rather than just a tool that is useful in certain cases.
I guess that isn’t as profitable and they have to push it on everyone for things it does poorly just in the name of forcing widespread adoption.
All of the needed words. Gotta be careful here. LLMs don’t deal in information. But yes, they are good at stringing together tokens if that’s all you need.
While I had similar concerns, I do think about this a fair bit now. A nontrivial math problem was proved by Aristotle (in Lean, a proof assistant, so we’re relatively sure it’s correct). Alpha Evolve then generalized. GPT pro did some writing and visualization work.
It’s basically the quality of work I aspire to. Slightly cheaper, and substantially faster. Not the strongest PhD, but a solid graduate student.