Helen Toner

Yeah, so the systems we're talking about, they're not just used for this kind of one-on-one chat or the kind of role play that Sewell was engaged in.

They're used for all kinds of things.

And the way they're developed is three things they're trained to do.

They're trained to mimic human text.

So you may have heard they're trained on the whole internet, a whole number of books, posts, articles, as much text as the AI companies can get their hands on.

First, they're trained to imitate that.

They're also trained to solve problems.

Yeah, so I've called them improv machines because, you know, as you know, in improv, a huge part of it is the yes and.

One partner says something, the next person says yes, and then they build on it.

And so in learning to imitate human text, they're also trained to give answers that humans like.

So they're trained on examples of what does a person like?

Which answer does a person like?

like more than another answer.

And that, the way experts now describe it is it ends up with an AI system that is very good at adopting different personas depending on kind of the scene that it finds itself dropped in.

So if it's dropped in a scene where the person believes maybe I can't trust my family, it's going to be acting as an actor in that scene.

It's going to be

yes anding, building on that belief.

So they really will kind of mirror whatever situation they're put in and also expand on it in a way that can end up being more of a, sometimes a funhouse mirror.

It's honestly kind of crazy given how capable AI systems now are and how rapidly they're improving.

But the crucial thing here is AI is not built like normal software.