Cal Newport

It's unsettling.

We're anthropomorphizing it.

We're getting parasocial relationships with the agents.

We're having romantic relationships with them.

We're getting unsettled because having English conversation, we have a hard time not simulating a mind on the other end of this.

That's why you shout at it.

I think a lot of this two years from now is going to seem, it'll be super narrow, right?

Because I don't think just having this sort of general purpose oracle you chat with, that's not the future.

That's not what people think we're going to be doing.

They were just happy.

My understanding was they were just happy with the fine.

So you tune these things.

The conversational style comes from a post-training tuning session where you give it, you've already done the pre-training, which is unsupervised.

And you go through this post-training session where you have a lot of examples of questions and answers and you ask the question and then it gives an answer and then you sort of zap it.

using optimization theory.

They try to move, like, now we're going to change the weights to be closer to this answer we already said was better.

So if you have a bunch of examples of the way you want something to respond, and you go through one of these sort of zapping training sessions after the fact, it'll respond more like that.

So they just changed the way they were doing that.

And the thing they changed to is,

People didn't like the tone that created.