Comment by edot
3 months ago
Math is verifiable. LLMs will continue to make strides in such search spaces where all they need to do is try->verify->repeat. You can't train an LLM on how to end the war in Iran or Ukraine, nor how to solve hunger, nor climate change, etc. Real issues.
And LLMs still suck at art and writing.
Three years ago, people were saying "LLMs just generate plausible-sounding text, they don't understand the notion of truth so they can't do verifiable work like math proofs."
They still can't. But very smart humans constructed ways to use the monkeys with typewriters (with a statistical advantage) to find correct answers to problems where they already knew how to verify the answer.
Which problems are not amenable to that approach?
3 replies →
You say it like it's not an achievement.
1 reply →
Right, which was true at the time. So hundreds of billions of dollars have been poured into making LLMs better at these tasks via pretraining, RL, RLHF, post training, etc. again all with something verifiable in the loop. In order to improve the thing in the loop, the loop itself needs to be verifiable.
There have only been a few thousand wars, and they’re all different and all different in the world in which they occurred. The dimensionality is absurd, which is not a problem for LLMs if there’s enough data, but in this case there isn’t.
> they don't understand the notion of truth so they can't do verifiable work like math proofs.
Nobody that understands automated proof checking was claiming that.
You understand that those are different kinds of "truth", right?
> they don't understand the notion of truth so they can't do verifiable work like math proofs."
a) No one ever said that.
b) Your comment shows a lack of understanding of the notion of truth.
Yeah but... Both things are true. LLMs are doing things we would have considered absolute AGI god tier magic 5 years ago. They also still can't count to 100 yet.
There are definitely LLMs that can count to 100
If you think all they need to do is "just that" (try, verify, repeat as you say) then I think you're very, very mistaken.
There is no way for the LLM to bruteforce the search space any better than a human. What it can do better, tho, is to make connections between seemingly (for us) unconnected notions and join them, then verify if that's right.
Your view is not only wrong but also condescending in this day and age.
That’s literally how AlphaEvolve works. And presumably how OpenAI’s research loop works if they’d ever publish it.
They put an LLM in a loop, with a reward function, and keep trying to get a higher score. The reward function for AlphaEvolve is “did this code get a better score or faster”, for math research it’s “did it write a LEAN proof”.
I’m open to hearing that I’m wrong but I don’t think I am. I agree that LLMs make connections that humans wouldn’t, but eventually you need to verify those because otherwise the connection it made may as well be a lie unless proven otherwise.
Claude had some pretty good ideas about how to end the war in Iran when I asked it a while ago (The US should sue for peace, I believe was the conclusion).