Will Douglas Heaven

I don't want it to get, you know, higher than that.

You can even sort of, you know, tie it to little bits of wood to make it sort of go one way more than the other.

But you're not at any point saying that, you know,

You're not placing leaves in particular places.

It just grows within the structure that you've given it.

And I think that's quite a good way to think of LLMs.

The designers of them, they give it some parameters.

They give it some shape.

They say how big they want it and lots of other things like that.

But once you're training it, it's more like a plant growing than traditional software as we've come to know it.

Yeah, I mean, that was an example that I wrote about.

And it's a fun example because on the surface, that's so trivial, right?

And I think the assumption with these models is that when you ask what color is a banana and it says yellow, you probably assume it's sort of done that in the same way that a human would answer that question.

But when one group of scientists actually looked at what was going on inside when they asked the model what color a banana was,

they saw that two different parts of the model were in play, depending on whether you asked it a true force question, like is a banana yellow, versus what color is a banana.

I mean, I don't know about neuroscience, so maybe I'm going to get shouted at here.

But I would assume that in a human brain, there's a part of it that sort of processes yellowness of bananas.

And when you answer a question, you're sort of drawing on that.

Or at least that was sort of the assumption that I think...

these scientists who were building LLMs had made.