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

Hi, Nora.

Hey, Bob.

How are you doing?

Pretty good.

How about yourself?

I can't complain.

Let me introduce this.

I'm Robert Wright, publisher of a non-zero newsletter.

This is the non-zero podcast.

You are Nora Bellrose.

You are at Illuther AI, a nonprofit AI research group that has kind of an interesting history that maybe we'll talk about.

And you are head of interpretability, which is interesting to me because that means one thing you're doing is trying to figure out exactly how these large language models in particular, I guess, work, which is something

I'm interested in. And I'm interested in the fact that it's more of a mystery than one might

have imagined at one point, how these AIs actually work. I mean, more of a mystery, even to the

people who built them.

Also, interpretability is closely related to the field of alignment. The idea being that once,

the more we understand about how they work, the easier it may be to align AI with human values,

human interests. So reduce the chances that it will kill us all in a quest for world domination or cause harms in less dramatic ways.

So I want to talk about all that.

Now, one interesting thing about you, compared to some people who are doing alignment-related research,

is that you're not that much of a doomer, right?

...