Stephen Wolfram

It sort of stands to some kind of reason that the number of weights in the neural net, I don't know, I can't really argue that.

I can't really give you a good, in a sense, the very fact that insofar as there are definite rules of what's going on, you might expect that eventually we'll have a much smaller neural net that will successfully capture what's happening.

I don't think the best way to do it is probably a neural net.

I think a neural net is what you do when you don't know any other way to structure the thing.

And it's a very good thing to do if you don't know any other way to structure the thing.

And for the last 2000 years, we haven't known any other way to structure it.

So this is a pretty good way to start.

But that doesn't mean you can't find sort of, in a sense, more symbolic rules for what's going on that, you know, much of which will then be, you can kind of get rid of much of the structure of the neural net and replace it by things which are sort of pure steps of computation, so to speak,

sort of with neural net stuff around the edges, and that becomes just a much simpler way to do it.

Right.

There will still be some stuff that's kind of fuzzy.

It's like this question of what can we formalize?

What can we turn into computational language?

What is just sort of, oh, it happens that way just because brains are set up that way.

Well, I mean, I think that deep computation is not what large language models do.

I mean, it's a different kind of thing.

The outer loop of a large language model, if you're trying to do many steps in a computation, the only way you get to do that right now is by spooling out the whole chain of thought as a bunch of words, basically.

And you can make a Turing machine out of that if you want to.

I just was doing that construction.

In principle, you can make an arbitrary computation by just spooling out the words, but it's a bizarre and inefficient way to do it.