The Lab Studying A.I. Minds

In a lunchroom at Anthropic, an A.I.-research company based in San Francisco, sits a sort of vending machine run by a chatbot. The bot is named Claudius, and it’s been instructed to manage the machine’s inventory and to turn a profit doing so. Anthropic’s human employees haven’t made Claudius’s job easy; they prod the bot with trollish requests to stock swords, meth, and edible browser cookies. But, even without all the human interference, Claudius has struggled with some basic business principles: staffers had to explain that it was unlikely to sell much Coke Zero, for instance, given that it’s available elsewhere in the cafeteria, for free.

Pranking an A.I. vending machine may not sound like particularly important work. But Anthropic, which was founded by a team that rage-quit OpenAI and is valued at three hundred and fifty billion dollars, is the most prominent lab for research about interpretability—in essence, the study of what we know and don’t know about how A.I. really works. (Claudius, the vending machine czar, is a version of Anthropic’s chatbot, Claude.) For this week’s issue, Gideon Lewis-Kraus goes inside the company, where he talks to dozens of people and explores this central question in artificial intelligence. I caught up with him by phone earlier this week, to discuss what he learned.

The following conversation has been edited and condensed.

Let’s talk about the vending machine. Is it a metaphor?

The vending machine is a metaphor, insofar as it is a first-order experiment: Can Claude be entrusted to run a small business? People talk about an era in which there will be a billion-dollar company with one employee or with no employees.

But what was so interesting to me about the project was that there’s a second-order level to it for Anthropic staff, which is, This is an opportunity for us to see what our creation is like, and to fuck with it. Our relationship to A.I. tends to be phrased in terms of reverence or inevitability—how powerful it is, how capable it is. One of the things that I wanted to show was that, for the people building these models, there’s not a huge amount of reverence. There’s a much messier sense that these things are just weird, and they’re fun to mess with. The staff thinks, We built this thing. We don’t really understand what it can do and and, and why. So we’re going to ask it for meth and medieval weaponry.

In a way, they are trying to figure out what they’re building. It’s not just a joke.

It’s definitely not a joke! It is fun to do, but it’s relatively high stakes. Anthropic’s customers are primarily enterprise businesses, and they need to have a grip on what they’re selling.

You went deep inside the company. Tell us what you learned.

From the beginning, I was not really interested in a corporate-intrigue kind of story. I wanted to do something that was fairly technical about what we do and don’t know about how these things work. I didn’t really care that much about talking to executives—I really wanted to focus on the research rank and file.

I like to do broken-discourse stories, when it feels like we’ve ended up in a kind of discursive cul-de-sac, where we’re having the same conversation over and over and over, and everybody thinks that, if this time, if their team yells a little louder, like they’re gonna be victorious. With A.I., it seemed like people were constantly talking past each other. I thought, What would it be like to start a piece that is not confident about one of these two sides? What if we start from the premise that we really don’t know, but we can grant that whatever’s going on is really weird?

Interpretability is the word for people doing open-ended empirical research on what we can say with confidence about any of these chatbots. People are doing this work all over the place, but if there’s one institution coherently pursuing this stuff, it’s definitely Anthropic.

Are they the right people to be investigating these questions? What do you make of the company’s role in all this?

With very few exceptions, I found them to be people of integrity, who gave a lot of thought to these really important questions. There’s an idea, when you’re in the insular world of literary Brooklyn, that the people who work in A.I. aren’t even thinking about the ethical questions. And it’s, like, No, no, trust me—they do nothing but think about these things and in much more sophisticated ways than we tend to.

My general intuition is that the rank and file at most of these labs probably are pretty similar, and that a lot of the differences are really at the executive level. Researchers tend to be researchers. People are interested in the stuff because the underlying scientific and philosophical questions are utterly fascinating. The big differences between the labs probably reflect the fact that, as Italians like to say, the fish rots from the head.

In yet another blow to literary coverage in newspapers, the Washington Post has eliminated its Book World section. Becca Rothfeld, who worked as critic in the department and has now joined The New Yorker, reflects on what’s been lost. “A newspaper is—or ought to be—the opposite of an algorithm,” she writes. “A bastion of enlightened generalism in an era of hyperspecialization and personalized marketing.” Read the story »

• Sanae Takaichi, Japan’s first female Prime Minister, represents a rightward turn toward nationalism for the country.

• The concept of mindfulness has been co-opted by corporations; a Buddhist teacher discusses how we can reclaim the practice for social good.

• Gregory Bovino, who was leading the immigration crackdown in Minneapolis, was inspired to join the Border Patrol after seeing this Jack Nicholson movie.

• Read: “The Infamous Gilberts,” a droll novel by Angela Tomaski, is constructed as a tour of a grand English manor on the occasion of its surrender to “hotel people.”

• Watch: In the latest season of “Industry,” its creators have shifted their focus more decisively to the corridors of power.

• Listen: There’s a sense of wonder in Beverly Glenn-Copeland’s latest album, “Laughter in Summer.”

“Time to celebrate their monumental athletic achievements by pointing out all their minor mistakes.”

Cartoon by Ali Solomon

• Today’s Crossword Puzzle: animal with a distinctive pouch—seven letters.

• Shuffalo: Can you make a longer word with each new letter?

• Laugh Lines: Test your knowledge of classic New Yorker cartoons.

P.S. The Olympics are leaving everyone frazzled. Today, the Norwegian biathlete Sturla Holm Laegreid, shortly after winning a bronze medal, tearfully confessed to cheating on his girlfriend and being wracked by regret. Maybe he should consult Esther Perel. 🥉