But instead, what exactly happened? I mean, what led to Hegseth labeling Anthropic a supply chain risk, and what does, what does that mean? [laughs]

Well, that's a really good question.

Now, typically, this is a restriction that has only ever been applied to, um, companies, typically infrastructure companies that have ties to our adversaries.

So for example, Huawei.

Like we know that Huawei has a close relationship to, um, the CCP.

Huawei is a telecoms infrastructure company.

It makes sense for the government to say, "We don't want anyone we're working with to be using, you know, Huawei infrastructure because we don't know what it's gonna be sending, exfiltrating and sending back to Beijing." So-On one level, there's, there would've been, okay, any government contractor, while in the course of doing their government business, cannot simultaneously be using an Anthropic product like Claude.

I guess I wonder if like there are just like a lot of kinks that are still being ironed out, but that there will one day be a point at which we can regulate it or, um, sort of treat it the way that we would a gun or a tank or a phone.

Can you explain what that is?

I use the idea of interpretability pretty broadly.

That, like, typically, when people talk about interpretability in a narrow sense, they're talking about what's often called mechanistic interpretability, which means looking at the level of, like, the actual substrate.

Like, looking at, like, the way that individual neurons and groups of neurons relate to each other as this is processing information.

So s- kind of on the level of neuroscience.

Like, they often compare it to, like, a sort of protobiology of machines.

But that only gets you so far.

What are the kinds of things they're doing? Walk us through, like, how they're experimenting with this thing that they've built.

Mm.

But then I think that on the deepest level, the reason that, like, we are doing this, for the people who are the most candid, is, like, because we can.

That, like, if you are capable of building something like this, you're just gonna do it because it's, like, really fucking interesting to do, and it's interesting for technological reasons.

It's interesting for what it m- may reveal to us about ourselves or about learning or about consciousness or about thinking, that, like, all of a sudden, we just, like, have this, like, other entity that can talk, and we've never had that before.

And in some ways, it seems sort of like us, in other ways, it seems nothing like us, but, like, the fact that this other thing exists as a point of comparison, like, just opens up a lot of really, really interesting questions.

And, like, that is one of the things that was on my mind a lot over the course of reporting, is that, like-...

It was a real emotional rollercoaster, for kind of lack of a better word, that, like, there would be times where I would come back from San Francisco with, like, a feeling of, like, total despair, and other times that I would come back with, like, feelings of, like, exhilaration.

When you would come back from San Francisco feeling depressed, what were you seeing that was making you feel that way?

So what does that tell us about the difference between Claude acting and basically Claude deciding?

Well, when this experiment came out last spring, there were kind of two responses to it.

You know, one response to, like, Claude's propensity to, to blackmail is just to say: "No, it didn't.

It didn't happen." You know, which is the, the kind of thing that a lot of people, for, you know, very good reasons, who don't trust these companies and don't trust these systems, they wanna just believe that this whole thing is fake, it's all just a marketing ploy.

And, you know, it's kind of easy to rebut that by just saying, like: "Well, no, this actually happened.

Like, this was an experiment that happened," and that we should be applauding Anthropic for publishing these kinds of results, which, you know, theoretically could run against its bottom line, since it wants to be selling these systems to companies that will have Claude reading their emails.

The more sophisticated criticism is to say: This really doesn't prove very much, because Claude recognized that it had been put in a position that resembled a kind of kitschy '90s corporate thriller, and it picked up on all of the clues.

And to a certain extent, this is what happens with advancements, but is this progression different?

Um, what do you mean when you say that the people designing them don't understand why they work the way they work?

work?...

Well, because they're not like traditional programs that say, "If X happens, do Y." They're not explicit instructions that way.

What they do is that they seek out patterns in huge volumes of data, and then they make inferences based on the regularities that they observe.

So they're sort of like, uh, it's, it's not too extreme to say that they're a little bit like children who are learning to generalize about the world, and they make generalizations based on everything that they have inhaled, which is virtually all written material on the internet.

And then they are able to extemporize and create text or code, or follow the patterns that they have observed in their training data.

And we don't always know exactly what that's gonna mean.

What does that mean?

[chuckles]

... people will believe me." And then, the stuff that started to get me to pay attention again, like, little, m- maybe about a year and a half ago, a little less, was research that was coming out of interpretability groups, which are groups looking into, like, how these models work, and, and alignment science groups about, like, what, you know, what are the values reflected in these models? And there was stuff about how, you know, models might-...

fake, to like pretend that they were aligned in one way in order to, like, get through training, to then be deployed.

And I sort of think, like, well, that's just very bizarre, and it struck me that, like, one of the ways to kind of, like, get around people's instinctive defense mechanisms about these things, where either they're so sure that they're, like, powerful and gonna be super intelligent, or they're so sure that, like, it's all hype, was to say, like, "Well, maybe we could all take a step back and agree that, like, whatever's going on, setting aside anything speculative about the future, like, just what we have right now, I think we could all agree that, like, it's pretty weird, that, like, whatever's going on is weird." And I thought, like, I think that there's a way to, like, go back into this and say, like, "Look at this research that's being done, that even if you don't wanna grant all these other speculative things, you can grant that, like, something bizarre is happening." And so, um, there's a guy, one of the seven co-founders of Anthropic is a guy called Chris Olah, um, whom I had met when he was, like, basically a child working at Google 10 years ago, and he's done a- you know, he's considered kind of the godfather of what they call mechanistic interpretability, which is like a, you know, looking at the individual neurons and, like, how it all works on the, on the level of the substrate.

And I wrote to him, and I was like: "Look, this is not gonna be a story about corporate power or the consolidation of corporate power.

It's not gonna be a story about geopolitics or regulation.

Like, those are all important things for other stories, but I'm interested in a story about, like, what can we say with any kind of certainty about how these things work and even really what they are?" And I think the other reason they were interested is that I said, like, "So much of the conversation about AI ends up being about, like, what this executive said or that executive said," and I said, like, "You know, with all due respect to the executives, executives are executives, and like, I'm really interested in the kind of, like, rank-and-file researchers who do this kind of thing, who I think don't get enough attention, um, and who tend to be, like, very thoughtful about these things." And so, to my surprise, Anthropic was like, "Yeah, you know, like, why don't you come hang out?" And I said to them, like, "I want- I wanna do this over seven or eight months.

What does it mean for an AI model to be safe? Like, is it just i- if you ask it, you know, for help in, like, creating a cocktail of drugs that will kill you, it'll refuse to do that? Is it like this idea that AI might replace us? And so I guess I wonder, like, when we talk about safety, what we're actually focusing on.

Are there policies that work? And are you hearing any of those policies being raised by the Trump administration?

So the general consensus among demographers seem to be that if you spend a truly staggering amount of money, you can raise the birth rate by perhaps a fifth of a baby per woman.

So a place like France spends something like 5% of its GDP, and they're certainly higher than the OECD average, but they're lower than the replacement rate that you would need to keep a stable population.

But then again, Hungary spends 5% of their GDP and their birth rate continues to fall.

So, so far, and there are all sorts of different things that people have tried.

People have tried matchmaking services.

People have tried, you know, more normal sounding things like subsidies and our extended parental leave and attempts to change social norms around parenting, especially egalitarian parenting.

Is there anything you can do to make it easier when they do?

South Korea is a country that has more deaths than births every year, and it's not an easy country to immigrate to. So its population is getting older and smaller. Um, and so it's sort of seen as almost like this bellwether for where the rest of us are headed.

And so I got there, and y- y- I, you know, I got into the center of Seoul, and I went to the subway at rush hour. And you saw no children anywhere. But at first I thought, like, "Oh, well, you know, in New York, like, would I take my kids on the rush hour subway? Like, probably not." But then pretty immediately, like, you really just don't, you know, you don't see playgrounds, or, like, the handful of playgrounds that I saw were completely empty at any time of day. And you just don't see a lot of children. And there were these no-kids zones everywhere. There were signs on restaurants and other establishments that said, "No kids here." And I, I mean, so much of it is about a rapid shift in cultural norms about kids. So in, in fact, I met with this young economics reporter who writes about this for a living, and she was in her late 20s/early 30s, and she said, like, "I understand all of this stuff on, on a deep economic level, but when I write about it, I think, like, 'Well, what would change my mind?' And the answer is nothing. There's nothing that would make me want to have kids because it's the norm to not want to have kids."

Right. It does feel like children are disappearing.

Yeah. And, um, something like 200 nursery schools have been converted into retirement homes because there's, like, a radical dearth of retirement homes. And you can see that some of these nursery schools that have been turned into retirement homes, they've kept the same directors and they had kept the same, like, rubberized play floors for the old people. And they even had, um... Actually, my fact-checker, Emily, found this. When she talked to them, she was like, "Not only do we have the same rubberized play floors. We have the same crayons." That, like, the kids used to use the crayons, and now we just, like, the seniors use the crayons.