Harlan Stewart

And, you know, a lot of times, you know, there's a lot of debate over experiments like this.

You know, people say, oh, you know, this experiment isn't exactly like reality, or, you know, maybe the researchers kind of set up the experiment in a way that caused that.

But in this particular experiment, it was specifically prompted.

It said, allow yourself to be shut down.

And, you know, the behavior was the opposite.

And that's very concerning.

And I think...

The problem is, you know, the more we make these things into agents trying to complete goals rather than some kind of passive question answering machine in a chat window, the more we're going to see them doing the scheming behavior because I think those things just go hand in hand.

Yeah, yeah.

I know someone who just the other day used

want these things to order a coffee from Starbucks.

And from what I understand, they just sort of said, here's my order, order it for me.

And without any human help or intervention, did it.

And that sounds great.

It sounds very helpful.

But yeah, it's the question, where is the line where it goes from being something helpful to being something to be concerned about?

I don't think we've

past that line yet.

You know, I don't think these things are quite capable enough to pose real dangers to us.

But the problem is, it's really impossible to know where that line will be.