You're listening to Robert Wright's Non-Zero podcast. You're listening to Robert Wright's Non-Zero Podcast.

Hi, Jeffrey.

Hey, Bob.

Thanks for having me on.

Well, thanks for coming on.

When we introduce us, I'm Robert Wright, publisher,

The Non-Zero Newsletter.

This is the Non-Zero podcast.

You're Jeffrey Laddish, Executive Director of Palisade Research, which studies AI risk.

And in particular, ways that AI could be used offensively by bad actors.

And before that, you were that, you were on the security team at Anthropic, which is one of the big,

I think of it as one of the big three large language model makers along with Google and

Open AI.

And of those three, the one whose identity is kind of most closely associated with concern

about AI risk, I would say.

Now, I'm really excited about this because there's a lot I want to talk about.

For starters, researchers at Anthropic actually published some important findings a few weeks ago about what's

going on inside the mind of Claude, their large language model, which I'm a fan of. And that has

implications for both risk and, you know, alignment that is keeping the AI in sync with human values and interests.

And then there was a paper that got a ton of attention in AI circles called situational awareness by Leopold Ashinbrenner, who was fired from OpenAI purportedly for just closing, you know, important

information. I get the sense that that was more of a pretext for firing a guy that Sam

Altman found annoying, but maybe I'm reading too much into it. But anyway, that paper, a lot of people found alarming.

...