Wondering if Modern LLMs like GPT4, Claude Sonnet and llama 3 are closer to human intelligence or next word predictor. Also not sure if this graph is right way to visualize it.
Wondering if Modern LLMs like GPT4, Claude Sonnet and llama 3 are closer to human intelligence or next word predictor. Also not sure if this graph is right way to visualize it.
Is it not the case that predicting the next word often requires reasoning about the next word in order to have any form of accuracy?
And that if you select for better and better prediction, you have to also select for reasoning?
This is true, but it’s specifically not what LLMs are doing here. It may come to some very limited, very specific reasoning about some words, but there’s no “general reasoning” going on.
Did you watch the video I linked?
It seems to be essentially about a way to trick them into doing general reasoning, and a direct response to your comment.
It’s not a direct response.
First off, the video is pure speculation, the author doesn’t really know how it works either (or at least doesn’t seem to claim to know). They have a reasonable grasp of how it works, but what they believe it implies may not be correct.
Second, the way O1 seems to work is that it generates a ton of less-than-ideal answers and picks the best one. It might then rerun that step until it reaches a sufficient answer (as the video says).
The problem with this is that you still have an LLM evaluating each answer based on essentially word prediction, and the entire “reasoning” process is happening outside any LLM; it’s thinking process is not learned, but “hardcoded”.
We know that chaining LLMs like this can give better answers. But I’d argue this isn’t reasoning. Reasoning requires a direct understanding of the domain, which ChatGPT simply doesn’t have. This is explicitly evident by asking it questions using terminology that may appear in multiple domains; it has a tendency of mixing them up, which you wouldn’t do if you truly understood what the words mean. It is possible to get a semblance of understanding of a domain in an LLM, but not in a generalised way.
It’s also evident from the fact that these AIs are apparently unable to come up with “new knowledge”. It’s not able to infer new patterns or theories, it can only “use” what is already given to it. An AI like this would never be able to come up with E=mc2 if it hasn’t been fed information about that formula before. It’s LLM evaluator would dismiss any of the “ideas” that might come close to it because it’s never seen this before; ergo it is unlikely to be true/correct.
Don’t get me wrong, an AI like this may still be quite useful w.r.t. information it has been fed. I see the utility in this, and the tech is cool. But it’s still a very, very far cry from AGI.
I’d also like to ask how you feel about this paper:
https://arxiv.org/abs/2402.12875
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.