I won’t pretend I understand all the math and the notation they use, but the abstract/conclusions seem clear enough.
I’d argue what they’re presenting here isn’t the LLM actually “reasoning”. I don’t think the paper really claims that the AI does either.
The CoT process they describe here I think is somewhat analogous to a very advanced version of prompting an LLM something like “Answer like a subject matter expert” and finding it improves the quality of the answer.
They basically help break the problem into smaller steps and get the LLM to answer smaller questions based on those smaller steps. This likely also helps the AI because it was trained on these explained steps, or on smaller problems that it might string together.
I think it mostly helps to transform the prompt into something that is easier for an LLM to respond accurately to. And because each substep is less complex, the LLM has an easier time as well. But the mechanism to break down a problem is quite rigid and not something trainable.
It’s super cool tech, don’t get me wrong. But I wouldn’t say the AI is really “reasoning” here. It’s being prompted in a really clever way to increase the answer quality.
This is a gross misrepresentation of the study.
That’s not their argument. They’re saying that they can prove that machine learning cannot lead to AGI in the foreseeable future.
They’re not talking about achieving it in general, they only claim that no known techniques can bring it about in the near future, as the AI-hype people claim. Again, they prove this.
That’s not what they did. They provided an extremely optimistic scenario in which someone creates an AGI through known methods (e.g. they have a computer with limitless memory, they have infinite and perfect training data, they can sample without any bias, current techniques can eventually create AGI, an AGI would only have to be slightly better than random chance but not perfect, etc…), and then present a computational proof that shows that this is in contradiction with other logical proofs.
Basically, if you can train an AGI through currently known methods, then you have an algorithm that can solve the Perfect-vs-Chance problem in polynomial time. There’s a technical explanation in the paper that I’m not going to try and rehash since it’s been too long since I worked on computational proofs, but it seems to check out. But this is a contradiction, as we have proof, hard mathematical proof, that such an algorithm cannot exist and must be non-polynomial or NP-Hard. Therefore, AI-learning for an AGI must also be NP-Hard. And because every known AI learning method is tractable, it cannor possibly lead to AGI. It’s not a strawman, it’s a hard proof of why it’s impossible, like proving that pi has infinite decimals or something.
Ergo, anyone who claims that AGI is around the corner either means “a good AI that can demonstrate some but not all human behaviour” or is bullshitting. We literally could burn up the entire planet for fuel to train an AI and we’d still not end up with an AGI. We need some other breakthrough, e.g. significant advancements in quantum computing perhaps, to even hope at beginning work on an AGI. And again, the authors don’t offer a thought experiment, they provide a computational proof for this.