Ladies and gentlemen, the moment you've been waiting for — Kevin Roose. Please come on up. Kevin.
I'm so glad you're here. He's wondering why did these people wake up so early? This is crazy.
Really. It's great to be here among this elite in Washington.
This is the elite, the swamp, the elite.
Yeah, welcome to us. It's really wonderful to be here. And this is the first event for the book after its launch. It launched today. So I'm very excited.
Because he knows what should come first in the world. Okay, this book that I had the honor of reading, he handed it to me secretly, and by the way, I haven't handed it to anyone else here, and that's more important. And I have to say, as you know, we feel somewhat that we know a little about what's happening, but this is crazy. You take us inside every detail of it, okay? But I have to ask you, I have to start here. I don't want to ask you what surprised you. I want to ask you how shocked were you while covering the fact that this technology is happening, and you're dealing with all these people saying: we're not sure how it works, but go ahead.
I mean, there were many surprises in this. So, the book is a real history of the past ten years in AI. It starts in 2017 when OpenAI began building what became the first large language model, GPT-1, and continues to the book's deadline, which was earlier this year. And I really wanted to tell two kinds of stories. The first was about the technology. What was the mood in the rooms when these models started assembling coherent sentences? What was it like when they started writing code and analyzing and doing things in the real world? What was it like when they started showing alarming traits or doing things the creators didn't intend? Then I wanted to talk about the people because I believe the people who build this technology are important to understanding how, as you know, people in Washington or New York or somewhere often ask me: do all these people know each other? Are they all friends? And I tell them: well, they know each other, but they're not friends.
Yeah. You make that clear.
Frenemies?
They're... no, they're not even frenemies. Many of them are just enemies, and they, as you know, used to work together in the past. And so the entire AI scene is now populated by this crew that was previously a very small group. I mean, when these people started talking about building artificial general intelligence, they were mocked and everyone dismissed them. It was a very marginal idea, and very strange. It was like saying you believe in alchemy or astrology if you were an academic researcher in AI. But they were convinced and believed this was possible, so they went off to build it, and now these people run the biggest AI companies in the world.
But you were in the middle of all that. Did you have any doubts? Were there any moments you wanted to pull someone aside? Everyone here is talking about safety guardrails and all that, right? Did you ever do that? Because, also, if you cover politics, you know, ask any political reporter here, ask "Kat" for example, if you're in the White House or covering a political event, and you're talking to an aide, and you want to say: "You really shouldn't do it this way, you're really messing things up," or something like that. Did you ever feel the urge to do that? I mean...
Was that me? Sorry. I've been exposed. Anyway, why am I the only photographer here?
Is this the president? You can take a picture if it's the president.
Yes. By the way, speaking of the president, I heard he named the new committee "Superintelligence" after seeing your book cover.
That's, that's true. Yes.
Okay. Did you make secret calls?
Oh yes, lots of secret calls. Yes.
Okay. So you're in the room with, you know, top leaders and they're telling you about their technologies. Does everyone say: "Man, you need to slow things down"? Or something like that. Did you ever feel that urge? I mean, they themselves were saying: "Man, we need to slow things down." That's what was shocking to me when I started preparing this book and talked to more than 150 people at "Google DeepMind", and "Anthropic", and "OpenAI", the three companies at the heart of this story. And you know, I've been covering tech news for a long time. I've covered social media news, and never has a big executive at one of those companies taken you to a side room and said: "Can we talk off the record?" And you say: "Of course." And they say: "I'm really worried that this technology is going to kill us all." That never happened with Facebook or YouTube or Pinterest or any of the companies that emerged in that era. But the people in the AI field are the most worried, which is something I think is sometimes hard for people outside that bubble to understand. It's like asking: why are these people building this technology if they're so worried about it? Why are they racing to make these systems more powerful if they believe they could kill everyone on earth? I really wanted you and I to understand this logic, and how they justify racing toward something they know could be extremely dangerous.
Right, except for the fact that it's totally amazing. You know, I mean you can see that, right? It's amazing. You can't believe what it's achieved and what it's tried to do. Okay, speaking of that, what happens when we cross the line from today's AI to artificial general intelligence? Or have we already done that?
I think we have. I think that by any reasonable definition before 2020, if you took today's AI systems through a time machine to 2016 or 2017 when people started working on these systems, and asked them: "Is this artificial general intelligence?" they'd say: "Well, yes, obviously." Because they're general systems, right? AI models a decade ago did one thing only. They suggested YouTube videos, or completed your emails to your friends, or recognized objects in images. Today's models can do many, many things, and they do it at levels that surpass some of the top human experts in fields like math and programming. So, they're not perfect. They're not superintelligent in the sense that they beat everyone at everything, but they're very general and very skilled at the things people use them for. So, I think the term artificial general intelligence has been defined in a way that makes it useless. It's like everyone has their own scale for what counts as artificial general intelligence and what doesn't.
Isn't that one of the problems with everything about AI and even renaming it? That everyone has a different concept, so when you say "guardrails," there are 25 different things for that. When you talk about any kind of legislation, we have a lot of people here from Capitol Hill working on these issues. I mean, what do you advise them to do? I mean, because you're talking to them directly. By the way, in case the Paramount merger deal goes through, just so everyone knows why my phone keeps getting alerts. Sorry.
Oh, you know, like a human social platform. That's great.
By the way, Mike Allen was here earlier and it was the same.
Um, I think the people in this room are more informed and engaged. I think Washington, D.C. is actually the second most AI-interested city in America. Um.
Thank you. Thank you for coming.
So, I'm less worried about you falling behind compared to people in New York and Los Angeles and everywhere else. But I think the really important thing is understanding the history, because a lot of people are trying to keep up with these things right now. I sympathize with that. It's a lot. It's a lot to take in. There's a lot of intellectual history, and personal grudges, and relationship networks. You have to understand how all these people think and where they came from and what motivates them. Because people, you know, executives at three different companies might say: "We share the same goal of creating safe and beneficial AI to improve humanity," but they mean completely different things by that and are pursuing it for very different reasons.
Are you surprised by the change in the relationship between AI leaders and the Trump administration? And how they're working together and talking to each other, whatever way you want to describe that.
I'm not surprised in the sense that I see it as inevitable once Trump was re-elected, that AI leaders, many of whom were convinced that superintelligence would arrive during this presidency, would seek to win the administration's favor. There's a scene in the book I really like, a chapter about these efforts by Sam Altman and others from OpenAI to get close to Trump, by going around Elon Musk in many cases.
He was standing in the way as a gatekeeper, preventing anyone he didn't approve from reaching Trump. So they devised a kind of elaborate scheme to literally announce "Stargate" on the first full day of the administration. So, I think this kind of jockeying and positioning was inevitable. In fact, I was surprised that Dario and Anthropic didn't play along. I understand they're saying all the right things. They've somewhat returned to the fold. But if you saw that press conference in the White House Rose Garden that day, I've never seen anyone look less comfortable being in the White House than him.
It's a classic Washington tradition, the lobbying style. They put you in front of that microphone. Yeah.
Oh my God. And it's simply, and this goes back to something someone close to Dario told me about him, which is that this man cannot tell a lie even if it would save his life. And they said this is a virtue on even days and a flaw on odd days. There are some days you really wish he would just put on a smile and say what needs to be said and move on. But there he is squirming and sweating and looking clearly uncomfortable among his fellow tech titans.
What do you think about the Superintelligence Committee announced by the White House? What do you think they'll do? Or what do you want them to do?
I'd like them to be bold. I think this 120-day period, people are saying 120 days to submit a recommendation as if this is an urgent timeline. I say we'll have had six more model releases 120 days from now. Things will have changed a lot. I still think the gears of government are moving very slowly right now. I don't know how to improve that. You know much better than I do. But to give you an idea, I think it would be useful for you to understand how fast people in the San Francisco Bay Area and AI labs expect things to change. People were telling me things like, oh, I think we'll see an intelligence explosion in 2030 or 2031. And what that means is this thing about recursive self-improvement where AI builds better AI, leading to superintelligence. People are now telling me they believe that could start this year. 2026. So, they might be wrong. They might be off by a year or two. But this is the level of craziness people are predicting not in the medium term, but right now.
Okay, there are basically six people deciding the future of AI. It's all about computing power, right? Sam Altman, Dario, Elon, we don't need their last names. We all know them. Mark, Jensen, Demis. Okay. How do you feel about that? Six people deciding, because we're here very interested in politics, right? And we've talked a bit about policies. And they're the ones who will make a real difference in all businesses, whether you're talking about banking or any kind of business. But in this case, it's just six people. It's not like J.P. Morgan or Goldman Sachs where there are huge crowds, it's narrowed down to six people.
It's a very small group, you're right, and that makes me feel extremely uneasy. I don't think any technology this important should be run and controlled by six people, even if some of them are thoughtful and wise. It's just a very small group. And I think the personal dynamics between these six are really concerning too. I try to explain to people that this isn't like a Coca-Cola versus Pepsi rivalry. This is more like a Rome versus Carthage conflict. They want to destroy each other. And they're bitter enemies. In fact, someone who read my book early told me the title could have been "Damn That Man." Because they're all people who used to work together and trust each other, then at some point realized: "I don't want that man to win this race because I don't trust him, so I'm going to race to build my own version of this technology faster than them, because their winning would be very bad news for society."
One of my favorite things in the book, besides that backstabbing part, is your talk about a team working on trying to crack the black box to explain how AI works from the inside. Can you talk about that?
Yes, this is a small subfield of AI research with a long and annoying name called "mechanistic interpretability." If you can say that five times fast, you get a prize. And they're actually doing brain scans of AI systems. They're trying to understand what happens inside neural networks, because I believe this is another part of the story. It's become cliché to say these large language models are grown rather than built, but it's true, and the researchers who build these things have no idea how their internal mechanisms work. They're trying to study that. And I spent some time with the researchers doing this. They're making modest progress, but the problem is that capabilities are advancing much faster than the ability to interpret. So, we're getting more capable systems faster than we can understand what those systems are doing. And that's the reason behind people calling for slowing down, right? I think there are some people who think it's all just an attempt to pump up their stock price before an IPO or something like that, or that they're colluding or trying to capture regulation. They're worried that capabilities are outpacing our ability to understand and control these systems. And that's why they're calling for a slowdown.
Okay, we love to trick the audience, and we love to trick the guest. So, let's start with this. Which executive, okay? Which executive said our destiny is artificial general intelligence first? Does anyone know who said that? No, but it's a good guess.
Okay, I know this because I wrote it in the book, but I don't know if anyone else wants to risk guessing.
It's the CEO of DeepSeek, Liang Wenfeng.
Who I think was clearly reading all the American tech CEOs' blogs.
Okay, which prominent AI company executive worked as a cancer researcher? Which AI company executive? Does anyone want to guess?
Dario. Okay. Okay. Which company has the largest cluster of graphics processing units (GPUs) outside of Nvidia? I can't hear you. Tell them.
Trick question, Google has tensor processing units (TPUs), not graphics processing units.
The lady from YouTube says.
Yeah. Okay. This might be very easy. What AI company's original framework was to be helpful, honest, and harmless?
Claude. Ding ding ding, Claude. Ding ding ding, Claude. Okay. Okay, which AI executive won a Nobel Prize and what did they win it for? And what did they win it for? It's the AlphaFold protein. AlphaFold. Okay, a very sophisticated audience. Okay.
This is the most AI-filled room in Washington.
This is very silly. Okay.
Can you go to Congress, please?
Yes. Who is the former executive who helped save Anthropic when it couldn't find investors? Which executive...
A former executive intervened in the matter?
Right, but that's not the answer you're looking for.
Okay, that was good. That was Grant Teeple from the Noble Rich Foundation. Send your tech researchers out there, guys. Okay, does anyone else want to guess the answer? We have an answer. Eric Schmidt. By the way, that was a reflection. I didn't say that. Okay. What AI-powered chatbot tried to destroy Kevin Roose's marriage? There you go. That was Tina. Okay. Okay, good answer.
By the way, unsuccessfully.
Still married to a human woman.
Yeah. What was her reaction, by the way?
Oh, okay. When was that night when you had that conversation with Bing? I think it was around one in the morning. I came to the bed and she was asleep, so I said, 'You won't believe the conversation I just had with the chatbot.' She said, 'That's nice.' So she's not a tech-obsessed person. She never cared, but she asked, 'Should I, like, close my bank accounts?' As if to say, 'Is there a rogue AI hunting me now?'
Oh my God.
But what was the prompt that you entered?
