Mark Williams-Cook

When they see that string of tokens, the kind of bell curve of

you know, answers that they've seen is sort of very steep.

It's like almost everything says it's high consensus.

Whereas if you say what happened in the news today, every time they're going to come across that in training is going to be different.

So it's going to be, I guess, flat and then like, okay, we need to go look for an answer.

Yeah.

But there's all the really interesting things, right.

Of, um, arguably things that don't have an objectively correct answer.

which is a lot of questions.

Oh, it's a lot.

It's a lot of questions.

So then how do you train an AI to deal with that without making it the most vapid experience ever where it just replies like, oh, there's many different viewpoints to this.

You know, those type of answers you get every time.

And it's just like, oh, like, I know.

Just tell me which one's best.

There's lots of ways to define best.

Like, just tell me what the best one is.

Well, I did a talk recently and I asked the audience who regularly shops on Amazon and pretty much everyone put their hand up.

And then I said, you know, keep your hand up if you think Amazon is kind of a good ethical company you should support.

And