Comment by freedomben
2 years ago
I think this problem needs to be solved at a higher level, and in fact Bard is doing exactly that. The model itself generates its output, and then higher-level systems can fact check it. I've heard promising things about feeding back answers to the model itself to check for consistency and stuff, but that should be a higher level function (and seems important to avoid infinite recursion or massive complexity stemming from the self-check functionality).
I'm not a fan of current approaches here. "Chain of thought" or other approaches where the model does all its thinking using a literal internal monologue in text seem like a dead end. Humans do most of their thinking non-verbally and we need to figure out how to get these models to think non-verbally too. Unfortunately it seems that Gemini represents no progress in this direction.
> "Chain of thought" or other approaches where the model does all its thinking using a literal internal monologue in text seem like a dead end. Humans do most of their thinking non-verbally and we need to figure out how to get these models to think non-verbally too.
Insofar as we can say that models think at all between the input and the stream of tokens output, they do it nonverbally. Forcing the structure of reduce some of it to verbal form short of the actual response-of-concern does not change that, just as the fact that humans reduce some of their thought to verbal form to work through problems doesn't change that human thought is mostly nonverbal.
(And if you don't consider what goes on between input and output thought, than chain of thought doesn't force all LLM thought to be verbal, because only the part that comes out in words is "thought" to start with in that case -- you are then saying that the basic architecture, not chain of thought prompting, forces all thought to be verbal.)
You're right, the models do think non-verbally. However, crucially, they can only do so for a fixed amount of time for each output token. What's needed is a way for them to think non-verbally continuously, and decide for themselves when they've done enough thinking to output the next token.
1 reply →
The point of “verbalizing” the chain of thought isn’t that it’s the most effective method. And frankly I don’t think it matters that humans think non verbally. The goal isn’t to create a human in a box. Verbalizing the chain of thought allows us to audit the thought process, and also create further labels for training.
No, the point of verbalizing the chain of thought is that it's all we know how to do right now.
> And frankly I don’t think it matters that humans think non verbally
You're right, that's not the reason non-verbal is better, but it is evidence that non-verbal is probably better. I think the reason it's better is that language is extremely lossy and ambiguous, which makes a poor medium for reasoning and precise thinking. It would clearly be better to think without having to translate to language and back all the time.
Imagine you had to solve a complicated multi-step physics problem, but after every step of the solution process your short term memory was wiped and you had to read your entire notes so far as if they were someone else's before you could attempt the next step, like the guy from Memento. That's what I imagine being an LLM using CoT is like.
1 reply →
> Humans do most of their thinking non-verbally and we need to figure out how to get these models to think non-verbally too.
That's a very interesting point, both technically and philosophically.
Where Gemini is "multi-modal" from training, how close do you think that gets? Do we know enough about neurology to identical a native language in which we think? (not rhetorical questions, I'm really wondering)
Neural networks are only similar to brains on the surface. Their learning process is entirely different and their internal architecture is different as well.
We don’t use neural networks because they’re similar to brains. We use them because they are arbitrary function approximators and we have an efficient algorithm (backprop) coupled with hardware (GPUs) to optimize them quickly.