Skip to content

Podcast

The Future of Addictive Design + Going Deep at DeepMind + HatGPT

Hard Fork

Source ↗ ← All highlights
  • Juries Found A Side Door Around Section 230
    • Two juries accepted a new theory that social platforms can be liable for defective product design, not just user-posted content.
    • Casey Newton says this opens a crack in Section 230 by targeting infinite scroll, autoplay, push notifications, and recommendation systems. Transcript: Casey Newton So I would say there are kind of two big reasons why these cases are super important. One is that these are what are called bellwether cases. Kevin, you ever heard of a bellwether case? These are like cases that set precedent for other cases, yeah? Exactly. These are the cases that, if successful, are going to open the floodgates for lots of other people to sue under the same theory. The second big reason that these cases are really important is that they appear to have opened up a crack in Section 230 of our Communications Decency Act here, which for 30 years has Been essentially the foundation that the entire internet rests on. It’s also a dentist’s favorite statute. Yes, that’s Section Tooth Hurdy, if the joke wasn’t landing for you. So yes, this is a super important, super important. I’m glad you got that. No, the really sad part was I was planning my own Section Tooth Hurdy joke because I just went to the dentist yesterday and now I didn’t have any cavities. Kevin Roose So tooth not hurdy. Moving on. Casey Newton Section 230, Kevin, you may remember, is the law that says that in most cases, these platforms cannot be held liable for what their users post. Yes. So if I went on Facebook and I defamed you, which is something I think about doing every day, you could sue me, but you couldn’t sue Facebook. Kevin Roose This is what’s been blocking my lawsuits against Facebook over your posts for years. That’s right. And back in the day, like 30 years ago, this was actually really important because there were these small internet forums that were starting up. Casey Newton Some of them got to be bigger size, you know, CompuServe, AOL. And inevitably, somebody would be mean to another user and they would say, I’m not just suing you. I’m suing CompuServe. I’m suing AOL. I’m putting the whole system on trial. And a couple of lawmakers got together and they said, this is going to destroy the entire internet. Like we need for there to be forums and not have these platforms being held liable for all these things. But fast forward to today and Kevin, would you agree that maybe there are some harms that are taking place on the internet that do not consist entirely of people defaming one another On Copysurf? Yes. Yeah. And so this is essentially the question that gets asked in this case, right? People say, hey, it seems like we’re a pretty long way away from 1996. I’m opening up TikTok. I’m opening up Snapchat. And I’m seeing infinite scrolling feeds. I’m seeing auto-playing videos. I’m a teenager, but I’m getting barraged by push notifications in the middle of the night. And that’s to say nothing of the recommendation algorithms that might be driving me toward content related to eating disorders or other things that are going to make me sad and upset. And so some of these people get together with their attorneys and they say, this actually feels different from the thing that Section 230 was designed to protect. Right. This is not about, oh, I got harmed by this particular piece of content. This is about the design of the whole platform. The design feels defective. And the really crazy thing about these cases, Kevin, is that juries agreed with these plaintiffs for the first time. And they said, we like this theory. We think these products are defective. Kevin Roose Right. So this is kind of a side door that these lawyers have found around litigating on Section 230, which they have successfully now shown that at least in these cases can convince a jury that It is not about what’s on the social network content wise. It’s about the actual sort of mechanics and plumbing of the social network that are harmful to people. Casey Newton That’s right. And we should say that we do expect some appeals here. And until those are, you know, sort of fully exhausted, I can’t tell you for certain, this is the moment that the internet changed forever. But there’s been a lot of commentary over the last week about what it would mean if these cases were upheld, because it seems like juries are just going to be really, really sympathetic To these claims. (Time 0:03:25)
  • What Platform Features Juries Treated As Harmful
    • The Los Angeles case focused on beauty filters, infinite scroll, autoplay, push notifications, and recommendation algorithms as harmful mechanics.
    • The New Mexico case argued Instagram misled families about child safety while becoming a playground for predators, with encrypted messaging criticized too. Transcript: Kevin Roose Please. So what are the actual platform mechanics that are being litigated over here? Casey Newton Yes. So in the L.A among the design features that were at issue were the so-called beauty filters that can make you, you know, look, quote unquote, more beautiful if you use them, infinite Scroll, autoplay video, these barrages of push notifications that platform sends. And also, I would argue more problematically, the recommendation algorithms that power the platform. And then in the New Mexico case, that was much more about kind of child safety. So they were arguing that Instagram in particular had become this playground for predators. It was very critical of the fact that Meta offers end-to encrypted messaging, and the basic idea was Meta falsely advertised that these platforms were safe when in reality, children Are being harmed there all the time. Kevin Roose So from what I understand, it was like the case was basically taken out of the playbook for going against big tobacco or another sort of industry that makes harmful products. You say this is harmful, and not only is it harmful, but the company that was making it knew that it was harmful and either made it more harmful or just released it as planned anyway. I did see some sort of exhibits that had been shown off at the LA trial, I believe, where some employees at Meta were sort of talking on their internal forums about how this stuff is so addictive For kids. That seems bad. And I imagine that was persuasive with the jury. But are there other instances where the platforms are being sort of taken to court over things that they sort of knew were harming people and that they either dialed up the harm in an attempt To spike engagement or sort of knowingly release these things to the public. Yeah. Casey Newton So, I mean, some of this research has come up in other litigation over the years, but I think this has been probably the most damaging case that we have seen. You know, the first time I remember reading a lot of these internal studies was in the wake of the Francis Haugen revelations a few years back, right? Like Francis Haugen walks out the door of Meta and takes a bunch of this internal research with her winds up sharing it with the Wall Street Journal and then eventually a bunch of other Reporters, including me. The reason that the research mattered a lot here, though, Kevin, was again, the plaintiffs are now building this very specific case, which is you’re building a defective product, Right? Before the past couple of years, we weren’t really using this language. We weren’t really adopting this sort of public health framing of a way to discuss the harms of social media. Before then, it was just kind of this more nebulous, like, hmm, like they’re studying the effect of Instagram on teen girls. And it seems like some of these girls are having really bad outcomes. But we didn’t really have the framing. Well, now we have the framing. And we’re just saying, like, hey, you looked into it. You found that some subset of your users are having really bad experiences. And you did not change the features. And so that mattered. (Time 0:07:08)
  • Why Addictive Design Is Harder To Define Than Nicotine
    • Kevin Roose questions whether addictive design can really be separated from content, since copied features do not always make apps succeed.
    • He contrasts nicotine with features like autoplay and infinite scroll, noting apps such as Sora copied the pattern without becoming hits. Transcript: Casey Newton Honestly don’t know the answer to that question. I think it’s a really interesting thing to watch. The question that you just asked is really, really controversial, actually, because much of what these platforms do is just protected under the First Amendment. And then Section 230 also protects a lot of speech, right? And the big debate that’s raging in the internet policy community right now is, can you separate design from content? I want to get your thoughts about this. Kevin Roose Right. Is it like the container or is it the stuff in the container that is dangerous? Casey Newton Yeah. And there are some people who are saying that, no, you cannot make that distinction and that effectively all design is content, right? Like if I want to send you a push notification, that is my right under the first Amendment, and you cannot tell me that I cannot do that. You cannot tell me that there is a certain limit that I have to place on the depth that you can scroll in Instagram. Like, that is protected. But for what it’s worth, juries are taking the opposite view. They’re saying that there are at least some things which seem like are just clear mechanical design features, and I happen to agree with them. Kevin Roose So let’s talk about this because I think this is maybe a place where you and I disagree, or at least where I have some misgivings about this theory. So in the case of something like cigarettes, which is a very heavily litigated field that I think a lot of this social media litigation has been modeled after, there’s like an addictive Ingredient, right? Nicotine. Everything that you put nicotine in becomes more addictive as a result of having nicotine in it. You know, this happens with cigarettes. It happens with vapes. It happens with, you know, nicotine pouches. If you started putting nicotine in ice cream, ice cream sales would go up because nicotine is very addictive. I think the question I have about the mechanical addictiveness of these sort of features like infinite scroll, like autoplay recommendations, is that if it followed the same principle As nicotine, then every product that has those would become way more popular. And one example I’ve been thinking about on this is Sora. They sort of took the playbook that was working for TikTok and Instagram, and they put it onto a new app, and the app did not succeed, right? There are other apps that have tried to mimic things like the news feed, that have tried to mimic things like autoplay video or recommendation algorithms that have not taken off. And so I guess the question in my mind is like, if the litigation over social media is modeled after the litigation over big tobacco, shouldn’t there be like some industry-wide lift As a result of every platform trying to borrow the most addictive features of Facebook and Instagram and YouTube? Casey Newton I mean, I hear what you’re saying, and I think it’s an interesting point, but I think that internet platforms (Time 0:10:11)
  • Scale Makes Addictive Design More Dangerous
    • Casey Newton argues scale changes the duty of care, so giant platforms should bear heavier responsibilities than small copycats.
    • He points to Meta hiring cognitive scientists to maximize time spent and says juries are reacting to a largely unregulated engagement machine. Transcript: Casey Newton And this is the approach that Europe has taken to regulating these platforms, right? They have certain categories. And if you are a very large online platform, then you just have more responsibility. That makes intuitive sense to me. I think that bigger and richer and more powerful you are, the more responsibility that you have to society, right? And so in this particular case, you have companies like Meta, which we know are hiring cognitive scientists who are working very hard to figure out all the different ways that they can Hack your brain to get you to look at Instagram for as long as they possibly can. It is in their interest to get you to look at Instagram as long as they possibly can. And right now, there’s just no break on that at all in our society except for this litigation. So I’m so sympathetic to these juries that are looking around. They’re seeing this almost completely unregulated platform, and they’re saying something’s got to be done. Kevin Roose Yeah. So regardless of sort of what our thoughts on the overall sort of legal theory here are, like, what do you think the effects are on the platforms? If this does get held up on appeal, if these platforms are found liable for millions or potentially billions of dollars in damages against all of these people who claim that they were Harmed by social media, does that mean that they have to, I don’t know, go back to like the reverse chronological feed of 2008? Does that mean they have to shut off, you know, infinite scroll and autoplay and recommendations and all these other things? This is where it gets really tricky. Casey Newton And this is maybe the one narrow way in which I’m sympathetic to the platforms, which is, okay, the juries have said your product is defective. What juries have not said is, here’s what an okay product looks like, right? They’re saying, we don’t like this sort of set of features, but they’re not saying with any specificity, like, well, how do we think that these features are interacting, right? Like, what is your actual model of the harm here? And so there is a world where the platforms feel like they have to comply, and they maybe start picking off some of these features one by one, like, okay, if you’re like under 16, we’ll Disable infinite scroll, for example. How much benefit does that really have to like the individual teenager who may be struggling? I don’t know. This, of course, is why it would be great if Congress could pass some sort of law regulating this. But, you know, we’re now like, I don’t know, a decade into that project and still not getting very far. (Time 0:14:19)
  • The TV Analogy Breaks On Personalization
    • Kevin Roose sees social media increasingly as media, which makes legal attacks on engagement mechanics resemble attacks on TV cliffhangers.
    • Casey Newton argues personalized recommendation systems differ because they search the entire corpus to keep one specific user watching longest. Transcript: Casey Newton Well, so I want to hear a little bit more about how you think about this, because you have talked on this show many times about your own struggles to look at your phone less. This is an issue that, you know, at various times you feel like has plagued you. So how are you feeling about the addictiveness of these platforms? Like, do you buy the sort of public health framing for the way that people are talking about them these days? Kevin Roose Or do you think that this is overreach? So I need to do some more thinking about the product harm arguments here and whether it makes sense to me. I am basically on board with the idea that there should be age gating for social media. I am sold on the premise that there is a certain age, whether it’s 16 or 18 or 14, where sort of the most harmful effects taper off. And I think before that age, it makes total sense to age gate or at least give parents a lot more control over what their kids are able to do and not on these platforms. I think the addictiveness question is just hard for me because I feel like my sort of macro theory on all this stuff is that what is happening to social media over time is that the social Part is fading away and the media part is rising in the mix. And so I think that if you start treating the design and mechanical decisions of these media platforms as harmful under the law, it just sort of leads me into a place where I become much Less certain. Like before any of this existed, there were cliffhangers on TV shows that were designed to keep you coming back after the commercial break or to the next week’s episode or whatever. Those were arguably addictive features. They would keep people coming back. Is that illegal? I would say probably it shouldn’t be. And it’s not. So I think there is a certain sense in which the closer that social media moves to something like TV or streaming video, the blurrier the lines in my mind get between the content and the Mechanics. What are your thoughts on that? Casey Newton Well, I have to disagree. I do think cliffhangers should be illegal because I want to know what happened. I don’t have to wait till the fall to find out, you know, if that person is still alive. But also, I do think that there are some really important differences between like, let’s say YouTube and HBO Max, right? Like HBO Max is not like going to modify the content of HBO to your individual preferences, right? Like they’re going to go pay some money for a bunch of shows and they’re going to hope a bunch of people watch them. The platforms that we’re talking about do something very different, right? They’re looking across the entire corpus of like every video that’s ever been uploaded to their platform. And they’re trying to figure out what will keep you personally here the longest. And we’re going to show you that as much as we can. So I just do think that there’s a kind of categorical difference here. (Time 0:17:00)
  • Child Safety Lawsuits Could Undermine Encryption
    • Casey Newton backs the verdicts but warns that treating encryption itself as evidence of danger could produce a serious privacy loss.
    • He says Meta dropped encrypted Instagram messaging while the New Mexico case criticized it, showing how child-safety litigation can pressure private communication. Transcript: Casey Newton Yeah, so here I understand that I’m coming across as being broadly supportive of these jury verdicts, which I am, but I do want to acknowledge this could lead to some really bad places, And this is why we need to handle Section 230 with care. In the New Mexico case, the attorney general argues that a reason that Meta should be considered liable in advertising their platform as being safe for children is that it includes Encrypted messaging, right? In fact, Meta in March announced that they would discontinue encrypted messaging on Instagram in what I believe was an effort to sort of get ahead of this. What they said was, look, if you want to use encrypted messaging, you can use WhatsApp instead. But to me, this would be like just a legitimately horrible outcome of all of this is if like every company that now offers encrypted messaging either voluntarily decided to stop offering It or was pressured by the government to stop offering it. Because in my view, encryption is a necessary part of privacy in a world where people are mostly communicating online. Kevin Roose Right. Are you comfortable with all of this happening in the courts through jury verdicts? Casey Newton This is not my preferred way of addressing this, but I think it was inevitable in part because the tech companies have been so obstinate about making meaningful changes to their platforms, Right? The world have been begging these companies for a decade, please do something to make these platforms safer and to make them less addictive and to reduce some of the harms. And instead, what we’ve mostly seen is a series of engagement hacks designed to get people to look at them longer, right? And in the United States, where you cannot regulate the content of any of these apps, for the most part, you’re really only left with the design, right? You’re really only left with just the raw mechanics of the app. So if the social media platforms are upset about the verdict here, I truly believe they brought this on themselves. (Time 0:20:10)
  • AI Chatbots Are Next In Addictive Design Lawsuits
    • Casey Newton thinks AI chatbots are the next frontier of addictive-design debates because teens already use them heavily and they may prove even stickier.
    • He says labs should want Congress to define what counts as a safe chatbot before lawsuits fill the vacuum. Transcript: Kevin Roose Wait, I do want to talk about the AI of this all. So obviously every discussion on this show has to come back to AI at some point. So I’m curious, like, what effects you think this might have on some of these AI companies, because they are also trying to create experiences that are engaging, addictive, whatever You want to call it. I can imagine some of these, you know, lawsuits that are being brought against the makers of chatbots for harms. Like, it all feels like it’s sort of going to converge at some point. So what’s your take on that? Casey Newton Yeah. So Pew did a study in 2025 and found that 64% of teens now use AI chatbots. About three in 10 use them daily. That same survey said that the teen use of YouTube, TikTok, Instagram, and Snapchat had remained relatively stable, right? So yes, chatbot usage is growing. It has not yet come at the expense of the social platforms. Although, of course, I expect that we’ll soon see chatbots inside all of those platforms, right? And like these things will all just kind of merge together. There’s something about these things where they do kind of go hand in hand. And to your point, like, I think that, yes, AI chatbots will be the next frontier of this debate, because in many ways, they’re much more engaging, and I think like will be stickier than Even these platforms are. Kevin Roose Yeah. I mean, it just seems so obvious to me that the platforms should be like absolutely begging Congress to regulate them, because the alternative is like they just get sued into oblivion By a bunch of, you know, law firms. Casey Newton I mean, absolutely. Like if I were running one of the big AI labs, I would want to have an understanding from Congress of like, what do you consider a safe chat bot? Like, give me a checklist that I can follow because I don’t want to have to be dealing with this in the next few years. (Time 0:24:33)
  • Demis Hassabis Described Science As A Spiritual Calling
    • Sebastian Mallaby says Demis Hassabis framed late-night scientific work in almost spiritual terms rather than purely technical ones.
    • In a London park, Hassabis told him reality was screaming at him and linked that feeling to Spinoza’s idea of understanding nature through God. Transcript: Kevin Roose Sebastian Melby, welcome to Hard Fork. Great to be with you. So people who listen to our show are familiar with Demis Hassabis and DeepMind. He’s been on several times. What is something non-obvious about Demis that you learned through talking with him through many hours and interviewing many people who know him? Sebastian Mallaby I mean, I think maybe the spiritual underpinning for his scientific curiosity was interesting. You know, there was one time when we were sitting in this London park and talking for a couple of hours, and he suddenly started saying, you know, when I’m up at two in the morning at my desk By myself, thinking about science, thinking about computer science, I feel reality is screaming at me, staring me in the face, waiting for me to explain it. And he calls it the God of Spinoza, that this is the 17th century philosopher Spinoza who said that to understand nature is getting closer to God’s creation. And that resonates with Demis. Maybe that’s something people don’t know. (Time 0:29:57)
  • The Evil Genius Myth Broke Down Under Reporting
    • Sebastian Mallaby rejects the caricature of Demis Hassabis as an evil genius and says the label partly came from a game Hassabis built.
    • Hassabis still pushed back hard when Mallaby uncovered internal documents about DeepMind’s attempt to spin out of Google and sent him to a lawyer. Transcript: Kevin Roose I mean, yeah, this has been something that’s come up in my own research, too, is that he grew up going to church, I believe, with his mother, and I think unlike a lot of the other AI leaders, Has a way of sort of fusing the science of AI with his own spiritual beliefs. And I know some folks have seen his ambition and his many years of competing to build AGI and have seen something suspicious in that, right? Elon Musk has this whole theory about how Demis secretly wants to be an evil AI dictator who takes over the world. And I guess I’m curious if in any of your reporting with him, you ever saw something that seemed like what Elon Musk was talking about. Sebastian Mallaby No, I mean, to the contrary. I think this idea that Demis is a, quote, evil genius, which is the one that’s the phrase that Elon used to use, came from the fact that in his video game production days, Demis had created A game called Evil Genius. And so maybe it was a joke at first, but, you know, really, I got to know Demis extremely well. I spent more than 30 hours with him. You stress test people quite deeply, as you know, Kevin, when you’re writing about them and then you might get pushback and legal threats and all that stuff. And he did make me talk to his lawyer once, and it wasn’t totally easy the whole time. But he was reasonable in the end. Kevin Roose Wait, why did he make you talk to his lawyer? Yeah. Sebastian Mallaby He was very mad at the fact that I unearthed the whole story about DeepMind trying to spin out of Google between 2016 and 2019. And, you know, they retained a whole bunch of advisors, lawyers, bankers, etc. They got Reid Hoffman to pledge a billion dollars to finance the spin-out. They went to see Joe Tsai in Hong Kong, the Alibaba co-founder. Anyway, so the lawyer was not amused that I had all these internal documents from Inside DeepMind, which had been leaked to me, the board presentation that DeepMind gave to Google, And so forth. And he said, you’re not supposed to be writing about this. And I said, well, you know, people gave me this stuff and tough. (Time 0:30:58)
  • Why Demis Hassabis Saw Himself In Enders Game
    • Demis Hassabis told Sebastian Mallaby he deeply identified with Ender’s Game as a boy-genius mission to save humanity.
    • Hassabis even gave the novel to his wife hoping she would understand him better, but she sympathized with Ender more than with him. Transcript: Casey Newton I have always believed that when a source gives you secret documents, it helps you get closer to God’s creation. So that’s what I would have told him. I wanted to ask another question about childhood because Demis told you that he really identified with the boy genius protagonist of the novel Ender’s Game, and of relating to this Feeling of being socially isolated by his own talent and consumed by a desire to make his mark on the universe. And the reason it struck me is that in this novel, Ender believes that he’s doing training exercises, but then what he thinks is like a test, essentially a video game, accidentally wipes Out an alien species. So I wondered if you talked with him about like why he relates to that story, and in particular, if there’s any relation to that and the idea of maybe trying to build a superintelligence. Sebastian Mallaby Well, I was astonished. You know, this was before my first dinner with him. And it was sort of kind of the vetting process. It was the last part of the vetting process where he agreed to give me the access I needed. And he said, you know, you’ve got to read this novel before you come and see me. And so I show up, I’ve read this story. It’s about a diminutive boy genius who basically saves humanity from aliens. And I’m thinking, does he really see himself as saving humanity by doing what he’s doing with AI? And even if he thinks that, why would he be so crazy as to tell me? I mean, surely that’s hubristic beyond belief. Why would you put that out there? And, you know, he made no secret about it. He said, yeah, you know, I feel like I identify because this guy put all of his energy and his life into saving humanity. And I feel like I’m on a mission like that. And he said, I felt so strongly about this, I gave it to my wife to read it, thinking that she would understand me better and sympathize with me. And you know what? She sympathized with the kid, Enda, but not with me. (Time 0:33:03)
  • ChatGPT Turned Demis Hassabis Competitive Instinct Into War
    • Sebastian Mallaby says Demis Hassabis’s competitiveness surfaced most clearly after ChatGPT changed the balance of power.
    • Hassabis described OpenAI’s launch as a battlefield move, saying they had parked the tanks on his lawn. Transcript: Kevin Roose Yeah. I mean, one other character trait that comes up over and over again in reporting about Demis, and especially in your book, is how competitive he is. This is a guy who loves to win. You know, he was a child chess prodigy, and he won this thing called the Pentamind, you know, five times, which is sort of like an all-around gaming competition. You think that is part of his approach to AI? I mean, he’s always talking about how he wants to use this to solve scientific mysteries and cure diseases, but is some part of it just like, this guy loves to win and this is a really big Contest. Sebastian Mallaby And he said, you know, Sebastian, this is war. These guys at OpenAI, they’ve parked the tanks in my front yard. He actually said, park the tanks on my lawn because he’s English. But yeah, you get it. You (Time 0:34:53)
  • DeepMind Missed How Far Language Alone Could Go
    • Hassabis did not regret missing ChatGPT so much as he believed language-only systems could not reach true intelligence.
    • His neuroscience background favored action and perception, but Mallaby says he underestimated how much real-world knowledge language already contains. Transcript: Casey Newton Bring up the release of ChatGPT, which happens in November 2022. And I’d love to hear a little bit more about how Demis had reacted to that. Because I think before that happened, Google really thought they were comfortably in the lead and did not seem to be feeling a lot of pressure to release anything. So I’m particularly interested if, in hindsight, Demis has regrets about the fact that they sort of let Sam Altman beat them to the punch. Sebastian Mallaby Yeah, I mean, he has an explanation more than a regret. And the explanation is super interesting. It’s basically that because he studied neuroscience for his PhD, and you’ve got to remember this is back in 2008, 2009, so nothing worked in AI, so he was starting from scratch. And one of the ideas in neuroscience is called action in perception. And this is the idea that to really be intelligent, you have to take action in the world. You don’t know what it means for something to be heavy unless you pick it up. You don’t know what gravity is unless you actually drop something. And so he had this idea when the Transformer paper came out in 2017 and OpenAI was starting to do the first GPT in 2018, second one in 2019, and so forth. You know, that’s not going to work. It’s not going to take you all the way to powerful intelligence because language is just a system of symbols. It’s not grounded in the real world. And it’s not that he was wrong in the sense that now we see world models come back in 2026 as a big area of excitement and research. But back in 2018, 2019, he was missing the fact that a huge amount of knowledge about how the real world works is in fact in language if you download all the language on the internet. (Time 0:35:46)
  • Two Rival Theories Of Intelligence Shaped Modern AI
    • Kevin Roose frames AI’s split as two theories of intelligence, one built on optimizing through competition and one on answering from vast knowledge.
    • Sebastian Mallaby says DeepMind’s distinctive bet was combining reinforcement learning with deep learning, as in AlphaGo’s London and Mountain View collaboration. Transcript: Kevin Roose Yeah. I mean, I want to run a theory by you, Sebastian, for your take. But as I’ve been working on my own book and about this sort of period at Google and at OpenAI and at DeepMind, it strikes me that there are sort of like two visions of what intelligence is That these companies disagree on. And in one vision, it’s like, intelligence is about winning. It’s about optimization. It’s about a contest between rival intelligences. And that’s very much like the deep mind sort of reinforcement learning paradigm, which is like AlphaGo. And, you know, you play a board game a bunch of times and you get better at it a little more every time. And then there’s this other view, which is sort of the more open AI sort of language model scaling paradigm, which is like, no, it’s about answering questions. Like, being very smart is about having the right answer to everything. Does that theory hold water with you that there’s something like psychological about these two approaches to AI development that actually are rooted in like what we think intelligence Actually is? Sebastian Mallaby Yeah, I would say that the DeepMind special source right from the beginning was to try to put those two things together. It’s interesting, for example, that with AlphaGo, the early research on that, Ilya Satskeva contributed to it. And of course, he was the sort of leading practitioner of deep learning, went on to be OpenAI’s chief scientist. But at the time, he was working for Google because Google had acquired his boutique. And so the reinforcement learning people in London working for DeepMind collaborated with the deep learning people in Mountain View. And that’s what produced the AlphaGo breakthrough. So I think you’re right. There are these two strands within AI of reinforcement learning, which I would describe as learning through experience, interaction with the real world through trial and error. And on the other hand, learning through data, and that is the deep learning. And for humans, you could think of it as being, you know, you can go to the library and read all the books, and that would be deep learning. You’re learning from data, from sort of crystallized human knowledge. Or you can go out there in the real world and learn about stuff by planting your garden and whatever. (Time 0:37:31)
  • How DeepMind Secretly Tried To Escape Google
    • DeepMind tried for years to spin out of Google after its promised safety and ethics structure fell apart.
    • Sebastian Mallaby says they recruited advisers, got Reid Hoffman to pledge $1 billion, met Joe Tsai, and used Project Mario to pressure Google. Transcript: Casey Newton You mentioned earlier this, I don’t know if it’s fair to call it a plot. It sort of seems like a plot that they had at one point after they had gotten acquired by Google to try to spin themselves out. I believe they call this Project Mario. I would love to hear a little bit more about how that came about and why they didn’t go through with it. Sebastian Mallaby What happened was that when they sold DeepMind to Google in 2014, they had a rival offer from Facebook. And Facebook actually offered them more cash. And one of the reasons they said no was that they wanted safety protections around their technology. And so they had this deal, it was going to be a safety and ethics board. And Google promised that and they went ahead and sold to Google. And they had a first meeting of the Safety and Ethics Board in 2015 after the acquisition. And in order to bind in the other people in the space, they got Elon Musk to host the whole Safety and Ethics Board at SpaceX. They got Reid Hoffman to show up. And you will notice that then these are the characters who either found OpenAI or funded in those two. So Google wasn’t best pleased, as you can imagine. Casey Newton I have to say, that doesn’t seem like a very ethical thing to do. Maybe not the people I would have put on my ethics board are these characters. Sebastian Mallaby But it’s a dichotomy, right? Dilemma. I mean, either you put people on the board who don’t know what they’re talking about and they’re not interested in AI, or they do know about AI, in which case they want to go and do their Own thing because it’s too exciting not to. And a fundamental mistake that Demis made in his early conceptualization of how AI would be developed was this notion that there would be one single lab producing AI on behalf of all Humanity And therefore, it could be safe because there’d be no race dynamic. And you could take your time in sort of red teaming the models before you release them. And that’s why he brought Musk into the tent. That’s why he brought Reid Hoffman into the tent, precisely because he thought we could all be one team together. And so then what happened after, to answer your question, Casey? So what happened after was that having lost that first experiment in setting up a safety and ethics oversight board, Google didn’t want to do another one. (Time 0:39:53)
  • Sundar Pichai Quietly Locked DeepMind Inside Google
    • Sebastian Mallaby portrays Sundar Pichai as the quiet force who kept DeepMind inside Google by delaying every path to independence.
    • He says Sergey Brin remains a source of tension, making the Pichai-Hassabis alliance one of capitalism’s most important partnerships. Transcript: Sebastian Mallaby Know, it’s attrition. Sundar Pichai, his personality and his management style comes out quite interestingly in this whole story because, you know, right at the beginning in 2015 when, you know, the first Safety and ethics oversight board fails, the next idea that Demis has for how to get some independence and control of the technology is to become a bet, as in an alpha bet, when they were Spitting out Waymo and some of the other side bets they had. And Larry Page was cool with this, and he was CEO at the time. But then right as these discussions were going on, he handed over to Sundar. And Sundar kind of pretended to say, oh yeah, absolutely, great idea, we should look into it. But really, he was just spinning them along and had no intention whatsoever of letting Demis spin out because he recognized him as the AI talent that Google was going to need in the future. And so essentially, there was this long drawn out, you know, delays here, we should just look at some more details. And here’s another term sheet. And I was given some of these term sheets. They’re like huge, great documents with red lines all over them where, you know, one team of lawyers had come back to the other team of lawyers. And, you know, basically by 2019, everybody was exhausted. It all fizzled out and they just moved on. Casey Newton There’s been a lot of sort of jostling for independence within DeepMind ever since the earliest negotiations about selling to Google. Give us an update on how things are going with them now. Like, you know, when we talk to them, they present things as being, you know, fairly like hunky-dory between everyone, but are there still kind of tensions and fault lines between Google And DeepMind? Sebastian Mallaby Well, you know, I’ll give you sort of what I would regard as somewhere between probably true and unconfirmed rumor. Is that all right? Am I allowed to do that? Oh, please. Casey Newton We love to gossip on this show. Spill the tea. Sebastian Mallaby So I’d say that, you know, Sergey Brin is the troublemaker here. That he, one of the Google IOs, I guess it was a couple of years ago, the stage was set up for two people to be on it. There was the interviewer and there was Demis. And suddenly Sergey kind of runs onto the stage. They have to get a third chair. And then he kind of inserts himself into that conversation. And, you know, what I hear is that that was the outward symptom of a much deeper tension, where Sergei doesn’t really like Demis’ leadership on this and wants to push back against it. And I think it follows from that that the single most important business buddy act in all of capitalism today is the one between Sundar Pichai and Demis Asabes. Because Sundar manages the board, manages the sort of high politics of Google and Alphabet, that Demis has the space, the resources, the oxygen to go do his science. (Time 0:42:36)
  • Competition Changed Demis Hassabis On Military AI
    • Sebastian Mallaby says Demis Hassabis changed on military use because the dream of one benevolent AI lab collapsed.
    • In a competitive race, refusing defense work just hands the contract to OpenAI or Anthropic, so he now prioritizes government-wide rules instead. Transcript: Kevin Roose Yeah. One area where Demis has changed his mind is about the use of AI in the military. This was a big sticking point in the negotiations with Google and Facebook back when they were selling DeepMind. He didn’t want their technology to be used for the military. Now, obviously, Google DeepMind has one of these Pentagon contracts. They’re working with the military. So what do you attribute that shift in his thinking to? Is it just kind of the realities of the market or needing to compete or what is it? Sebastian Mallaby Yeah, I mean, Demis described this to me as, you know, you mature. You get to know the real world and all that. One might say, how come you weren’t mature when you sold the company in the first place? I mean, surely it was predictable. But I think that the real truth of the matter is he did not predict. I mean, it comes back to this singleton idea, which I mentioned before. He really thought there would be one lab. And in a scenario where there’s only one lab who’s got the technology, then sure, you can say to the military, you can’t have our technology. Go away. And the problem today is, as we saw with Anthropic just now with the Pentagon, if Anthropic tries to draw a red line, you know, open AI is in there like a shot and says, hey, Mr. Pentagon, what do you need? We’ve got it for you. Kevin Roose Do you worry that Demis’ competitive streak or his pursuit of science, whatever it is that drives him, will compromise his ability to develop something like AGI safely? Sebastian Mallaby You know, I asked myself that question all the way through my research and in some ways the question about can you be a strong strong, consequential actor in the world and still be good Is sort of the deep question in the book. And he is somebody who really wants to be good. And I think one way of framing this question about, is he being good? Will he be good? Can he be good? Is to say, should he, will he do what Dario did standing up to the Pentagon about red lines on military usage and surveillance. And I don’t think he is going to do that. And I think the way he would rationalize this would be to say, look, you got to pick your moment with this stuff. If you make a stand and actually the Pentagon does what the hell it wants anyway, you didn’t really make the world better. My best shot at making the world better and making AI safer is to go through the route, which is the only route that can get us to AI safety, and that is government intervention, forcing Safety rules on all the labs at once. Because otherwise, some are safe, some are not safe, and the ones that are not safe are going to screw it up for everybody. And that’s the route that I think Demis wants to push. Problem is you have the Trump administration. They just want to accelerate. And so all you can do for now, I think, is to keep this conversation alive with other governments. (Time 0:45:39)
  • DeepMind Recruits Were Warned About The Bunker Endgame
    • DeepMind job candidates were once told to prepare for a final bunker phase near AGI when secrecy and focus would become total.
    • A former employee told Sebastian Mallaby that if Hassabis had ordered a flight to Morocco to hide, he would have seen it as fair warning. Transcript: Casey Newton Write that Demis used to inform job candidates at DeepMind that if they signed on, they should, quote, prepare for a climactic endgame when they might have to disappear into a bunker. Why would they have to disappear into a bunker? And do they still tell the job candidates that? Sebastian Mallaby Yeah. So the idea was when you get very close to AGI and it’s super dangerous, you’re going to A, be subject to potential attack by bad guys who want to steal the technology, and B, you really Don’t want to be distracted by quotidian real-world stuff, so you disappear into the desert. Yeah. That’s right. You leave your TikTok on your phone in some… I think Kevin used to lock his phone up in a box, as I recall. That’s correct. And so you do a Kevin and you go and you really, really focus and you really get the AI right in the last stages. That was sort of Demis’ vision. And to test whether he really meant it, I was having dinner with somebody who used to be at D-Mind in that period around 2015, 2016, and had now left. And I said, this wasn’t really true. He didn’t really. Oh, yeah, yeah. This guy said to me, if Demis had told me any time when I was working at DeepMind that I had to take the next flight to Morocco and hide, I would have said I’d been given fair warning. Wow. So the bunker is in morocco just so everyone knows yeah and i said why why morocco and he said well you know it’s the desert and you know the manhattan project was in the desert oh it’s the Oppenheimer syndrome these (Time 0:48:33)
  • AI Agents Are Starting To Break Human Review Systems
    • Kevin Roose thinks AI agents will break human-moderated systems by flooding them with submissions faster than people can review.
    • He links the Wikipedia incident to a coming inbox apocalypse where email, GitHub, and contribution workflows get overwhelmed by bot output. Transcript: Kevin Roose I think this is just going to be the year that every system on the internet that is built on human contribution and review is going to break. And it will break not only because the AI tools, but because people are letting them loose onto websites where they are doing things like editing Wikipedia articles and defaming people Who contribute things to GitHub projects. We heard from Scott Shambaugh about that on a previous episode. But I think this is going to be a challenge. I have started talking about the inbox apocalypse that is going to hit this year, where everything that is normally sort of reviewed and bottlenecked by humans is just going to be overwhelmed And fluttered with AI submissions. Casey Newton Absolutely. I mean, I’m already getting emails now every week from something claiming to be an AI agent that says, you know, it’s running a company, you know, but it’s always sort of like, let me know If you want to talk to my human. And I was like, you’re human. Better hope I don’t catch them in a dark alley because this does not belong in my inbox or frankly anywhere. Yeah. Kevin Roose Yeah. I’m getting these two. It’s like, it’s a total scourge. It’s somehow even more annoying than the like faceless PR spam that you and I get. Casey Newton Just to be very clear, there’s not one thing that anyone’s agent could do or say to get me to respond to it in any way. So use that information. I hope that goes into your training data. Stop (Time 0:56:20)
  • When A Disney Robot Fails The Magic Turns Horrific
    • A new Olaf animatronic in Disneyland Paris lost its carrot nose and slowly collapsed, instantly turning cute robotics into nightmare fuel.
    • Casey Newton says child-friendly machines become Lovecraftian horror the moment the illusion breaks and their mechanical guts show. Transcript: Kevin Roose I met Olaf, the Frozen Robot, Who Might Be the Future of Disney Parks. Sean reported in mid-March about his interaction with a new animatronic, Olaf the Snowman robot from Frozen. It weighs 33 pounds, it was trained with an NVIDIA GPU, and is controlled by an operator using a Steam Deck. But when it made its debut at Disneyland Paris, well, Casey, something Should we take a look? Let’s take a look. All right, Olaf the snowman talking, waving his stick arms. Oh, no! No! We lost him. Olaf! Oh, the carrot nose falls off. Oh, it’s… Casey Newton Oh. There’s something about the way that he very slowly falls onto his back. Kevin Roose Oh, no. Yeah. 20 children just got lasting trauma. They’re going to be talking about this in therapy. Look, what do you expect? Casey Newton Like, of course he was frozen. That’s what the whole movie is about. Oh. Kevin Roose Do you want to kill a snowman? Okay. Casey Newton I mean, there is, it’s just reliably very funny when you create an animatronic thing for a child, and then it is like revealed to be a machine, and it just sort of feels like a Lovecraftian Horror. Yes. You know, like something about that transition from like a cutesy, cuddly thing to like its eyes are, you know, bulging out of its head, and sparks start flying out of the back. I’ll never forget the day at Chuck E. Cheese as a kid when I learned that the guitar playing mouse wasn’t real. You know Chuck E. Cheese’s full government name, right? What is it? You don’t know. This is not a joke. It’s Charles Entertainment Cheese. Come on. I swear to God. I learned something every day from you. (Time 0:57:42)
  • Claude Code Proved The Wrapper Is The Product
    • Kevin Roose says the Claude Code leak mattered because the exposed value was the agentic harness around the model, not just the model itself.
    • People quickly cloned the surrounding apparatus and even paired it with open models, showing how fast coding-tool advantages can diffuse. Transcript: Casey Newton Well, this, Kevin, is a story about the Claude Code leak. So, Kevin, what do you make of this Claude Code leak? Kevin Roose Well, I think it’s a big deal, in part because the agentic sort of coding harness that is around Claude Code is really the special sauce, right? It’s the model underlying it is part of what makes Cloud Code and other agentic coding systems good at coding, but it’s really all the stuff around it, and that’s what leaked. It is not the actual, like, weights or the source code of Opus 4.6 or whatever model people are running inside Cloud Code. It’s like the sort of apparatus around it that makes it quite effective. So within hours of this leak, there were people who had cloned it and set up their own versions of it. I imagine it’s a very busy week over at the Anthropic legal department trying to get all this stuff taken down. But look, I think this kind of thing was inevitable, maybe not at Anthropic, but the agentic coding tools were all going to get good. They were all going to sort of reverse engineer cloud code and figure out what made it better. But I think this probably just accelerated that. Casey Newton When I saw this, my first thought was, right now, Kevin Roos is somewhere vibe coding cloud code using the downloaded leaked cloud code harness. Kevin Roose I have not yet downloaded the leaked cloud code harness, but I have seen other people sort of taking it and then putting it on top of like an open source Chinese model or something and sort Of Frankensteining their own sort of version of CloudCode that they can run. And I will say, the closer I get to my rate limits on CloudCode, the more I’m tempted to do something like that. Casey Newton That makes sense. Here’s the last thing I’ll say. If Anthropic is looking for a new harness for Cloud, they might want to pick one up at Mr. S. Leather in San Francisco, down in the Folsom District. They have some really nice options down there. (Time 0:59:31)
  • AI Is Making Software Vulnerabilities Easier To Find
    • Kevin Roose says the real cybersecurity story is that AI can now find bugs in foundational software better than most human experts.
    • Casey Newton ties that to Anthropic reportedly delaying a model so cyber defenders could prepare, echoing GPT-2-style caution. Transcript: Kevin Roose Oh, my God. Casey Newton In some cases, people only found out that their Zoom calls were recorded once Webinar TV reached out to them to say their call was turned into a podcast in an attempt to promote Webinar TV services. What is happening? Kevin Roose What is going on? What is happening? Okay, I want to start by saying, I am committed to making a podcast with you for the rest of my life. But if we ever get overtaken on the charts by an AI-generated webinar TV podcast that’s been trained on people’s boring-ass Zoom meetings, I am leaving this industry. Casey Newton Here’s why this is such great news. I think a lot of podcasters struggle with the idea that maybe their podcast, you know, maybe they didn’t have a great episode. Maybe they’re wondering, like, is this thing good enough to put out on the internet? Congratulations, because every single human-made podcast is better than every single webinar TV episode that’s ever been released. Kevin Roose Yeah, I mean, I’m just like, these have to be the most boring podcasts ever created. Like, what are you going to talk about? Is it called Action Items? Is it called Circle Back? What’s the title of this podcast? Touch Base, a limited eight-part series. Casey Newton I heard there’s a great series over on Webinar TV right now. It’s called, oh, I think you’re on mute. Kevin Roose So you may want to check that one out all right stop generating next out of the hat we have north korean hackers suspected in axios software tool breach this comes to us from bloomberg Casey Newton And it’s about axios not the media company i actually would prefer to read a story about this from axios if you have one on hand. Kevin Roose This is a tool, an open source tool widely used to develop software applications. This has been a big security breach. Hackers were able to breach one of the few accounts that can release new versions of Axios late on Monday and publish malicious versions. Axios is downloaded about 80 million times every week. Anyone who has downloaded the malicious version of Axios could then have their own computer and the data on it stolen by hackers. This is being attributed to North Korea. Seems really bad. Yeah, man. Casey Newton Like, there’s a lot of cybersecurity incidents we’ll talk about where it’s like, you know, but like no personal data was stolen or like, you know, nothing sensitive was at risk. This is one where it’s like, no, like everything was at risk. Like, this is one of the bad ones. And, uh, you know, if you’ve been messing around, uh, with NPM over the past week, you probably need to take a look at this. Kevin Roose Yeah. It’s really, I think this is going to be one of the biggest stories of the year is just what is happening in cybersecurity right now. Um, I was watching this YouTube video. If you ever are, uh, you know, need something to keep you up at night, watch a talk given by this guy, Nicholas Carlini, who’s a security researcher at Anthropic at a cybersecurity conference Recently. It is like the most terrifying conference speech ever given because what he’s basically saying is these AI tools have gotten better than almost any human hacker, any human security Expert at finding vulnerabilities in tools, even tools that have been around for decades like the Linux kernel. Is going to need to be rewritten and substantially hardened because we are facing like an onslaught of these very sophisticated AI tools that can find every little bug and problem in Them. Casey Newton Well, I am going to watch that talk just as soon as I’m finished watching Fruit Love Island. (Time 1:03:17)
  • OpenAI Pulled Back From Slop Feed And Adult Chat
    • OpenAI’s shutdown of Sora and retreat from an erotic chatbot mode suggest the company is abandoning engagement side bets for coding and enterprise focus.
    • Casey Newton says both ideas conflicted with OpenAI’s stated mission and likely died once Claude Code showed where real money was. Transcript: Casey Newton Okay, so this is actually a two-parter, Kevin. Two stories about OpenAI recently that caught our attention. One, Sora has shut down, which was a prediction that I made at our year-end episode. Yes, you called this one. This was my low-confidence prediction for the year, and it’s already come true by March. And then a second story, which I think actually, crazily enough, is related. OpenAI has apparently shelled its plans to release the erotic chatbot, or sort of like the adult mode, that it said that it was going to be bringing soon to ChatGPT in an effort to boost Engagement. So, Kevin, dying to know what you made of those two changes. Kevin Roose So, I think you were smart to predict the end of Sora. I think the story with Sora never quite made sense to me. Like, it was obviously a very cool piece of technology. It was devastatingly expensive to run, is my understanding. Like, generating all those short videos was, like, computationally quite pricey. And so I think they are making the decision to sort of spread their bets a little less and consolidate around, like, a few projects, one being enterprise AI, one being coding and sort Of automating AI research. But I think they maybe made a few too many side bets in the past couple of years that they are now seeing were expensive and diverted resources away from the core. Casey Newton I have to say, I was personally really glad to see both of these changes, like the release of this infinite slop feed app last year, and the company saying that they were going to release This adult mode while they were still having all of these issues with like psychological problems that some of their users were experiencing as a result of getting a little too close To their chatbots. I just thought both of those seemed like really irresponsible moves, and just like contrary to what they said their mission was. So I was actually just really happy to see them say, you know what? We’re not doing any of these things anymore. Like I think that was the right move. Now, did they do that out of the goodness of their heart and some sort of like, you know, moral awakening that they had? No, they saw Anthropic, which had started to print money because Claude Code was taking off and they said, we want to get a piece of that. But hey, whatever it took, I’m just glad it’s happening. (Time 1:07:32)