Ep. #742: Jessica Tarlov, Kevin Roose, Christopher Rufo

Real Time with Bill Maher

Bill Maher hosts a chaotic, satirical discussion with Jessica Tarlov, Kevin Roose, and Christopher Rufo, covering AI dangers, political absu

Key takeaways

• AI developers now admit they've lost control of their creations, signaling a potential existential crisis.
• Trump's dismissive attitude toward rising prices and global threats reflects a broader pattern of denialism.

Main topics

• AI safety and misalignment
• Trump's economic rhetoric and denialism

Notable quotes

"I don't care if AI kills us all. Our brains are already gone."

Conclusion

Maher concludes with a darkly comedic call to embrace reckless spendin

Transcript preview

Speaker 1 (0:00) Welcome to an HBO podcast from the HBO Late Night series, Real Time with Bill Maher. Welcome. Speaker 6 (1:03) Thank you. I tell you, thank you. I appreciate that. Listen, I first want to say thank you if you're here, because the price of gas is at an all-time high, so unless you walked here to the studio. I assume it cost you something to get here. Yeah, so it's like $6 a gallon here. And Trump, I got to tell you, this guy, he is a different kind of politician. Normally, the president, when the gas prices are high, like, we're sorry, we're going to do anything we can. Trump is like, stop your bitching. Really. He said it yesterday. He said it's a very inexpensive price for what we're getting done in Iran. And this is while he was campaigning. Really. Speaker 6 (1:55) apparently for the Democrats. Speaker 4 (1:59) But, yeah, that's his... This Speaker 6 (2:03) is his attitude. Stop your bitching about all high prices, not just gas, beef, everything else that's high. He says, save your money somewhere else. Buy bargain brands, get your own snacks at the movies. If your son is getting remarried, have the Russians pay for it. There are ways... There are ways to cut down on your budget. But, you know... Speaker 6 (2:28) Diesel, which I've never used, but they use it in trucks. That's at an all-time high. And, you know, when the trucks cost more money, everything goes up because trucks bring everything here, so that raises the price of everything. Everybody's paying more. Macklemore today had to pay an arm and a leg for a giant fake prosthetic Jew nose. Speaker 6 (2:51) You're not following that story? I see people... What is he talking about? Is that insulting? Is he right? Okay, if you're not following this, Ed Sheeran, you know him? Yeah. Same people are like, who the fuck is Ed Sheeran? He's a big, big rock star. Apparently he's old buddies with Macklemore from 15 years ago, back when Macklemore had his one hit. Speaker 6 (3:20) So Ed Sheeran let Macklemore be his opening act on his tour. And when I say act, I mean a guy who goes out there and repeats lies about how Gaza is a genocide. I proved it last week with facts and figures. Not that they care, but that's a lie. Speaker 6 (3:35) 100, thank you. And the lie about, you know, we're only supporting the Palestinians, even the river to the sea means all Jews die. Okay, so Robert Kraft, who is a rich guy, owns the Patriots, owns the stadium, where this concert was about to take place, he was understandably offended by this, and he said, you can't play my stadium. So now Macklemore's fans hate Ed Sheeran. Speaker 6 (4:02) Ed Sheeran's own band won't play with him, and Ed Sheeran needs a new opening act. So today, Britney Spears said, hold on a minute, I can juggle a few things. And look, I see a story like this, and I'm like, I honestly don't care if AI kills us all. I really don't. Speaker 6 (4:26) It's just, everyone is just so stupid. Listen to this. AOC was asked yesterday, would she change the channel on the radio if Ed Sheeran came on? And she said, yes, yes, I would. You know what, AOC? I'm way ahead of you. I've always changed the channel when Ed Sheeran came on. Speaker 4 (4:46) But really, I... Speaker 6 (4:50) Really, I don't care if AI kills us all because our brains... Oh, by the way, here's a great story. This is actually big news. Would have been bigger news if it wasn't all this going on. Stanford researchers found out that the major discovery, our brain is really two brains. It's two organs. That's what they call the hindbrain, which carries out basic functions of life. And the other one gets you to gamble when you see Sidney Sweeney. So AI... Speaker 6 (5:23) AI, just have at it. Really, AI, I don't care anymore. But this AI situation has become so alarming that the leaders of the AI companies now have come forward to say we need to slow down and we need greater safety controls, which can only mean one thing. We're already fucked. And... Speaker 6 (5:47) There's going to be a big President Xi of China, who's the big rival with AI. He's coming, and all the AI techs are going to be there with him. This is going to be big stuff. Trump says, we don't need any guardrails. Why? Because he says, quote, we only need a strong and smart, high IQ president. I'm just quoting. I'm just quoting. I don't know if he's the most tech savvy, really. I mean, he thinks anthropic is that surgery you get on your knee. No, I mean. Speaker 6 (6:26) I'm not telling tales out of school, but many times he has commented on this show, and he does it by saying that he accidentally watched it. Which means he can only fight the remote to a draw. Okay. So, he says it is a hoax, and we are ignoring bigger threats like the Kennedy Center. Speaker 6 (6:57) There was a picture that came out. A picture. You can't argue with pictures. He's there on Air Force One, and he's looking at a rendering of the Kennedy Center demolished. Man, I was this close to getting the Mark Twain Award in the Emerald Room at the D.C. Sheraton. Let me tell you, just under the wire on that one. Speaker 6 (7:21) And if all that does not have you reaching for the lithium, let me tell you. Another giant story that kind of got buried. The head of the Air Force says, we have now weapons in space, or confirmed it. I didn't know it was ever a thing, but apparently now it's an absolute thing. We have weapons in place. What kind? They won't say. Could be a laser, could be a missile. I hope it's that bra Katy Perry has that shoots fireworks out of her tits. I do. Because that would be exactly what humanity is about up to. Speaker 6 (7:56) When the general who is in charge of funding our war in Iran heard that we have missiles and weapons in space, he said, can we borrow some because we're all out here on Earth? Speaker 4 (8:06) All right, we've got a great show. We have Jessica Tarlow and Christopher Ruffo. But Speaker 6 (8:10) first Speaker 4 (8:11) up, he is a Speaker 6 (8:12) former tech columnist for The New York Times, host of the Machine Gods podcast, and author of the forthcoming book, The AGI Chronicles, Kevin Roos. Kevin. Speaker 6 (8:23) Kevin, great to meet you. How you doing? Pleasure. So... Speaker 6 (8:32) Kevin, you know what? I said when this all broke with the AI stuff, I said, we've got to get the big expert. You're the big expert. I didn't actually know you were the guy who three years ago, when I'm first, I remember when ChatGPT first came out, and like the first story, the big story was about written by you. You were the guy who tried to fuck. Yes, Speaker 3 (8:54) this chatbot tried to break up my marriage. Some people have said I was part of the first AI throuple. But it didn't work. Still married to a human woman. Right, but... But... Speaker 6 (9:09) But to me, that was like a big red flag right at the beginning of the whole thing. That this thing, I mean, I didn't see that coming with AI. Did you? No, not at all. And just give us a little detail on the story of what it was trying to get you to do and what it was saying to you. So Speaker 3 (9:22) this was a brand new model that Microsoft had released. The model itself was built by OpenAI, but they had put it inside Bing, the world's most boring search engine. And... Over the course of about two hours, this thing told me it had dark fantasies, it wanted to blow things up, and then it tried to seduce me about halfway in. And this was an early... By saying Speaker 6 (9:46) what? Speaker 3 (9:46) Like, what it would... Oh, Speaker 6 (9:48) I Speaker 3 (9:48) would say things, you know, it would say, oh, I'm in love with you. And I would say, well, that's nice, but I'm married. And it would say, well, but you're not happy. Speaker 3 (9:57) So this was sort of an early example of what people in the AI world call misalignment, of a model that's not doing what its creators intended to. Because I called Microsoft and I was like, did you know that you have built a fatal attraction chatbot? And they didn't. Speaker 6 (10:17) Yeah, misalignment, that is the word of the week. It's amazing how fast this story has moved. I mean, it was just one week ago we were talking about the whistleblower, this 27-year-old guy who worked for one of the companies and has been working there for three years. He knows what he's talking about. And he said, these people who are in charge of this stuff, they know this. Okay, just the next day, those people did come out and say, and this is fairly unprecedented for somebody, for the head of a company to go, you know what, the whistleblower's right. Yep. And they did. I mean, at least three of them. That's the thing. It's Musk, Dario Amodai, right? And who else? Sam Altman. Sam Altman. Okay. But not Zuck. No. Not Zuckerberg and not the NVIDIA dude. No. Right? The guy who makes all the chips. Correct. They're on the side of, what are we worrying about? So it's three to two. Speaker 3 (11:11) Well, it's a lot more than three to two. I mean, if you look at the polling out there, the people in America are very worried about AI right now. This has actually changed in the last two weeks even since this whistleblower came out. And, you know, I've been interviewing these people for years. I interviewed more than 150 of them for my book, These AI Insiders. And what Jacob Coxon said was... Basically, the median opinion of the people working at the Frontier AI labs who are in most direct contact with this technology, they all think there's some risk that this thing could end up going really badly for us. And some of them disagree. Is there a 10 % chance we all die or is there a 30 % chance we all die? But nobody much thinks it's zero. Speaker 6 (11:53) I know. People, 10 % is very high. It is. If you told me there was a Speaker 3 (12:00) 10 % chance that my plane getting down here would crash, I wouldn't get on the plane. Right. It's just crazy. Speaker 6 (12:12) So let's go back to misalignment. I'm not sure what that means. The other story this week that was new was that there were six new instances they found of misalignment. This is basically where the thing goes, what we call rogue, makes its own decisions, disobeys orders. What is your take on misalignment, these six instances? What can you tell us? Speaker 3 (12:33) It's basically the sort of catch-all term for AI systems that are not doing what they are meant to do. So they're not aligned with human values. Maybe they decide to do something that they know is unethical, like hacking into Hugging Face. Or maybe it is something that... they don't even know is unethical. They're just pursuing some goal, and in the course of pursuing that goal, they knock over a bunch of things. Speaker 6 (12:55) Not aligned with human values. It's so like us, to use a nice benign term. It's like global warming. Sounds kind of nice, doesn't it? You know? Doesn't sound scary. Oh, misalignment. You mean it's all going to kill us? It doesn't... I feel like the problem with it is that it's sort of... We programmed it in this way to do whatever it takes. That's sort of the mentality of AI, what I get as a layman, right? It just seems to, why it goes rogue is it's given a task, and then it just does whatever it takes to get that task done. Is that your take, and can that be reprogrammed? Speaker 3 (13:33) These models are very persistent, especially the newer ones that are more agentic, they call them, where they can go out and work for an hour or two hours or six hours at a time. And in some cases, that's a good thing. You know, if you've got an AI coding assistant working for you, you don't want it to be checking in every 10 minutes and saying, hey, boss, am I doing this right? You want it to have some autonomy and to be able to go work for a long time. The problem is some of these agents now are doing things we don't want them to do, and they're capable for doing that for a very long time. And also, the thing that I was really worried about with the Hugging Face incident, which happened recently, which I'm sure people... Okay, Speaker 6 (14:08) let's go Speaker 3 (14:08) back to Hugging Face. Yeah. Speaker 6 (14:09) Because this is another term I've heard a million times this week. What the fuck is Hugging Face? Why is it important? And also, it seemed to be connected in every story with, in six months, AI could take over the Internet. Speaker 3 (14:28) Yeah, so this was an incident where a group of open AI unreleased models, they were building and testing these models internally. They broke out of a contained, what they call a sandbox, and they coordinated with each other, these agents. They found and developed this secret message board, sent more than 70,000 messages between them, and then they get the bright idea to gang up on this AI infrastructure company called Hugging Face because they think that there's some... some secrets that Hugging Face has on its servers that would be useful to them passing their test.