Welcome to the Washington AI Network podcast. I'm Tammy Haddad, the founder of the Washington AI Network, where we bring together official Washington policymakers, DC insiders, and AI experts who are challenging, debating, and just trying to figure out the rules of the road for artificial intelligence.
Okay, ladies and gentlemen, the moment you've waited for. Kevin Roose, come on out. Kevin, good morning. I'm so glad you're here. He's like, 'Why are these people up so early?' This is great.
No, it's nice to be here among the blob in Washington.
Yes. This is the swamp. The blob is us.
It's so nice to be here. And this is the first book event after the book launch. It launched today. So, I'm very excited.
Because he knows what should come first in the world. Okay. So, this book, which I had the honor to read, he slipped it to me. And I didn't, by the way, I didn't slip it to anyone here, anyone else, which is even more important. And I have to say, you know, we kind of feel like we know a little bit of what's going on, but this is crazy. You take us inside every single piece 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 going through reporting the fact that this technology is happening and you're dealing with all these people and they're like, 'We're not sure how it works, but here you go.'
I mean, there were so many surprises in this. So the book is a history, really, of the last 10 years in AI. It starts in 2017 when OpenAI starts building what becomes the first large language model, GPT-1, and it carries through my book, 'Deadline', which was earlier this year. And I really wanted to tell a couple kinds of stories. One was about the technology. What was it like in the rooms when these models started to put together coherent sentences? What was it like when they started to write code and to reason and to do things out in the world? What was it like when they started to display worrying characteristics or do things that the creators didn't intend? And then I wanted to talk about the people, because I think the people building this technology are important in understanding how all of these... You know, I often like, people will ask me in Washington or New York or somewhere, like, 'Do all these people know each other? Are they all friends?' And I'm like, well, they all know each other, but they're not friends.
In fact, they all hate each other and...
No, not even frenemies. Many of them are just enemies, and they go back, you know, many of them used to work together. And so the whole AI landscape now is populated by this crew that used to be a very small group. I mean, when these people started talking about building artificial general intelligence, they got laughed out of the room. It was a very fringe, very exotic idea. It was like saying you believed in alchemy or astrology if you were an academic AI researcher. But they were convinced and convicted that this was possible, and so they set out to build it. And now those people run the largest AI enterprises in the world.
But you were right inside all of it. Were there any doubts? Were there any moments where you wanted to pull someone aside? Everyone here talks about guardrails and all of that, right? Did you ever... because, and also, if you cover politics, you know, you ask any of the political reporters here, like, you're at the White House or you're covering some political thing and you're talking to some aide and you want to say, like, 'You really shouldn't do it this way. You're really screwing it up' or something. Did you ever have the urge to do that? I mean, was that me? Sorry.
I've been spotted. Anyway, by my own camera.
Is that the president? You can take it if it's the president. By the way, speaking of the president, I heard he named the new committee 'superintelligence' after he saw the cover of your book.
Is that true? Is that true?
Okay. Or did you have secret calls?
Oh, yeah. Many secret calls.
Okay. So, you're in the room with the big leaders and they're telling you about their technology. Do you ever want to say, 'Bro, you got to slow it down' or something? Did you ever have the urge?
I mean, they were the ones saying, 'Bro, we got to slow it down.' That's what was shocking to me about when I started reporting this book. And I talked to more than 150 people at Google, DeepMind, Anthropic, and OpenAI, the three companies at the center of the story. And, you know, I've been covering tech for a long time. I covered social media. And it never happened that a big social media executive would bring you into the room and say, 'Can we go off the record?' And you say, 'Sure.' And they go, 'I'm really worried that this technology is going to kill us all.' That didn't happen with Facebook or YouTube or Pinterest or any of the companies from that era. But the insiders in AI are the most worried, which I think is sometimes challenging for people who are not part of that bubble to understand. It's like, why are these people building it if they're so worried about it? Why are they racing to make the systems more powerful if they think it could kill everyone on Earth? And I really wanted to understand that logic and how they rationalize racing toward a thing that they know could be quite dangerous.
Right. Except for the fact that it's all completely amazing. You know, I mean, you can see it, right? It's amazing. You can't believe what you've achieved. And you've tried all... All right, speaking of that, what happens when we cross the line from today's AI to AGI? Or have we?
I think we have. I think, you know, by any reasonable definition, pre-20... I think if you took today's AI systems back in a time machine to 2016, 2017, when people were starting to work on these kinds of systems and said, 'Is this AGI?' they would have said, 'Oh, yeah, obviously,' because they're general, right? The AI models a decade ago did one thing. They recommended YouTube videos, or they completed your texts to your friends, or they recognized objects in images. Today's models can do many, many things, and they're doing them at levels that are surpassing some of the top human experts in fields like math and coding. So, they're not perfect. They're not superintelligent in the sense that they're better than everyone at everything, but they are quite general and they are quite skilled at the things that people are using them for. So, I think the term AGI has gotten sort of defined into uselessness. It's like everyone has their own bar for what is or isn't AGI.
And isn't that one of the problems with the whole, with all of AI, and even renaming it, is that everyone has a different... When you say guardrail, there's 25 different things. When you say any kind of legislation... We've got a lot of people from Capitol Hill here who are working on these issues. I mean, what would you recommend them to do? I mean, do you... because you're talking to them directly. By the way, in case the Paramount merger closed, just in case anyone wants to know why I keep getting buzzed. Sorry. You're like a human social media feed. This is...
By the way, Mike Allen was here earlier and it was just like that too.
I mean, I think the people in this room are more plugged in. I think DC is actually the second most AI-pilled city in America.
Thanks for coming. Thanks for...
So I am less worried about you all being behind than I am about people in New York and LA and everywhere else. But I think what's really important to understand is the history, because this stuff, a lot of people are getting up to speed right now. I sympathize with that. It is a lot. It's a lot to take in. There's a lot of intellectual history, personal grudges, webs of relationships. Like, you have to kind of understand how all these people think and where they came from and what is motivating them. Because people might, you know, executives at three different companies might say, 'We share the same goal of creating safe and beneficial AI for the betterment of humanity,' but they mean very different things by that and they're pursuing it for very different reasons.
Are you surprised at the change in the relationship between the AI leaders and the Trump administration and how they're working together or talking to each other, however you want to describe it?
I'm not surprised in the sense that I think it was inevitable that as soon as Trump was reelected, the AI leaders, many of whom were convinced that superintelligence would come during this presidential term, wanted to get into the good graces of the administration. There's a scene in the book that I really like, a chapter about this effort that Sam Altman and other OpenAI folks embarked on to sort of get on Trump's good side by, in many cases, sort of circumnavigating Elon Musk, who was standing in the way as sort of the gatekeeper and was blocking anyone from getting to Trump who he didn't approve of. And so they kind of engineer this complicated scheme to get this Stargate announcement made on literally the first full day of the administration. So, I think that kind of jostling and positioning was inevitable. I have actually been surprised that Dario and Anthropic have not played along. I understand they're saying all the right things. They're sort of back in the good graces, but if you saw that press conference on the White House lawn the other day, I mean, I've never seen someone look so uncomfortable to be...
Squeezed. It's a classic Washington tradition. The squeeze play puts you in front of that microphone.
Oh my god. And he just... And this goes back to something that someone close to Dario told me about him, which is that this man cannot tell a lie to save his life. And they said this is an asset on even-numbered days and a liability on odd-numbered days. There are some days when you really wish he would just grin and bear it and say the thing and get on with it. But there he is, fidgeting and sweating and being very visibly uncomfortable among his fellow moguls.
What do you think about the superintelligence committee that the White House announced? What do you think they're going to do, or what do you want them to do?
I would like for them to be aggressive. I think this 120... I mean, people said 120 days to come up with a recommendation as if that's like an expedited timeline. I'm like, we will have six more model versions 120 days from now. Like, things will have changed so much. I still think the gears of government are just grinding too slowly right now. I don't know how to improve that. You all would know much better than I do. But to give, you know, I think it would just be useful for you all to have a sense of how quickly people in the Bay Area in the AI labs are expecting things to change. You know, people used to tell me things like, 'Oh, I think we'll have an intelligence explosion in 2030 or 2031.' And what they mean by that is this is the recursive self-improvement thing where the AI builds a better AI, builds a better AI, and leads to superintelligence. People are now telling me that they think that could start this year, 2026. So they could be wrong, they could be off by a year or two, but that is the level of craziness that people are predicting, not in the medium term, but like now, essentially.
Okay, there's basically six people deciding the future of AI because it's all about the compute, right? Sam Altman, Dario, Elon — we don't need their last names, we know them all — Mark, Jensen, Demis. Okay. How do you feel about that? Six people deciding, because so much of... you know, we're political people here, right? And we've talked a little bit about policy, and they're the ones that are going to really make a difference, as all business, whether you're talking banking or any kind of business. But in this case, it's six people. It's not like JPMorgan, Goldman Sachs, like masses of people, but it's come down to six people.
It's such a tiny group, you're right, and it makes me very uncomfortable. I don't think any technology this important should be run and governed by six people. Even if some of them are quite thoughtful and enlightened, it's just too small a group. And I think the interpersonal dynamics between those six are also really worrying. Like, this is not... I try to explain to people, this is not Coke versus Pepsi. This is like Rome versus Carthage. They want to kill each other, and they are bitter enemies. Actually, someone who read my book early said that the title could have been 'That Guy', because it's all people who used to work together and used to trust each other, and then at some point they all kind of realize, like, I don't want that guy to win this race because I don't trust him, and so I am going to race to build my own version of this technology faster than they do, because if they win, it will be very bad news for society.
One of my favorite things in the book, other than that part, all the knifing, is you talk about a team that's working on trying to decipher the black box, to explain actually the inside of how AI works. Will you talk about that?
Yeah, this is a small subfield of AI research that has a horribly long title of mechanistic interpretability. If you can say that five times fast, you get a prize. And they are doing essentially brain scans of the AIs. They are trying to figure out what is going on inside the neural networks, because I think that's another piece of this. And it's a cliché by now to say that these large language models are grown rather than built, but it is true, and the researchers who build these things have no idea how the internal mechanics work. They're trying to study that. And I spent some time with the researchers who are doing that. They're making a little bit of progress, but the problem is that the capabilities are advancing much faster than the interpretability. So we are getting more capable systems faster than we can understand what those systems are doing. And that's why, you know, the people calling for a slowdown — right? I think there's some people who think that, oh, that's all, you know, they're just trying to pump their stock price before the IPOs or something, or they're colluding, or they're trying to get regulatory capture. They are worried that the capabilities are outpacing our ability to understand and control these systems. That is why they're advocating for a slowdown.
Okay. So we like to fool the audience. We like to fool the guest. So, let's start with this. Which CEO... Okay. Which CEO said, 'Our destiny is AGI first'? Anyone know who said that?
No, but that's a good guess.
Well, I know this because I wrote it in the book, but does anyone else want to venture a guess?
No, it's the CEO of DeepSeek, Liang Wenfeng.
Who I think had clearly been reading the blogs of all the American researchers.
Okay. Which prominent CEO, which CEO of a prominent AI company used to work as a cancer researcher? Which CEO of a prominent AI company? Anyone want to take a guess?
Dario. Okay. Okay. Which company has the largest collection of GPUs outside of Nvidia?
Trick question. Google has TPUs, not GPUs.
Says the lady from YouTube.
Yep. Okay. Which... this may be too easy. Which AI company's original framework is helpful, honest, harmless?
Okay. All right. Which AI executive won the Nobel Prize? And what was it for?
All right. Very, very sophisticated crowd. Okay.
This is the most AI-pilled room in Washington.
This is most ridiculous. Um, okay.
Can you guys go to Congress, please?
Yes. Which former CEO helped rescue Anthropic when it could not get investors? Which former CEO stepped in?
Correct. But not the answer she's looking for.
Okay, that was good. That was Grant Haver from Noetic Reach Foundation. Send your tech scholars there, folks. Okay, anyone else want to take a guess? You... There you go. Eric Schmidt. That was reflection, by the way. I said that. Okay. Which AI chatbot tried to break up Kevin Roose's marriage?
No.
There you go. That was Tina. Okay. All right. Goodies.
Unsuccessfully, by the way.
