Skip to content
Modern WisdomModern Wisdom

"They’re Building an AI God They Can’t Control” - Tristan Harris

Tristan Harris is a tech ethicist, entrepreneur, and a speaker. Are we sleepwalking into disaster? AI is unlocking massive progress, but the dangers hiding beneath the surface are exactly what experts fear most. So what’s coming… and could it spiral beyond our control? Expect to learn why AI is distinct from other kind of technologies, what the Alibaba rogue AI catastrophe that should scare everyone is, how worried Tristan is about the impact of AI deepfakes and misinformation campaigns, what’s happening with the AI safety discussion, if we should be skeptical of AI companies pushing just as hard but pretending that they’re not, the end result that AI companies are looking for and much more… - 0:00 Can Life With AI Have a Positive Outcome? 6:56 Is AI the Most Powerful Force We’ve Ever Created? 16:07 Powerful But Not Wise: AI’s Biggest Flaw 19:11 Can AI Actually Rot Itself? 24:30 Social Media’s Shift Away From Human Flourishing 29:09 Are We Headed Towards an Anti-Human Future? 36:53 Who Funds AI Once It Replaces Us? 40:58 Why Best-Case Scenario is Still Alarming 53:33 Inside the Alibaba Blackmail Scare 01:04:01 Can We Really Stop AI Taking Over? 01:13:04 The Danger of Denial in the Age of AI 01:20:19 Are AI’s Benefits Blinding Us? 01:26:01 Why We Need to Face the Reality of AI 01:31:56 Are AI Companies Controlling the Narrative? 01:33:31 How Close are We to an AI Takeover? 01:35:39 Why Changing AI Feels Impossible 01:42:30 Total Control or Total Collapse? 01:46:23 How Can We Globally Coordinate AI Safety? 01:52:40 Why Elon Isn't in The AI Doc 01:59:18 Why Every Second Counts 02:03:58 How Do We Accelerate Meaningful Change? - Get up to 20% off the leading longevity and cellular health supplement at https://timeline.com/modernwisdom Get up to $350 off the Pod 5 at https://eightsleep.com/modernwisdom Get a Free Sample Pack of LMNT’s most popular flavours with your first purchase at https://drinklmnt.com/modernwisdom New pricing since recording: Function is now just $365, plus get $25 off at https://functionhealth.com/modernwisdom - Get access to every episode 10 hours before YouTube by subscribing for free on Spotify - https://spoti.fi/2LSimPn or Apple Podcasts - https://apple.co/2MNqIgw Get my free Reading List of 100 life-changing books here - https://chriswillx.com/books/ Try my productivity energy drink Neutonic here - https://neutonic.com/modernwisdom - Get in touch in the comments below or head to... Instagram: https://www.instagram.com/chriswillx Twitter: https://www.twitter.com/chriswillx Email: https://chriswillx.com/contact/

Chris WilliamsonhostTristan Harrisguest
Apr 2, 20262h 7mWatch on YouTube ↗

EVERY SPOKEN WORD

  1. 0:006:56

    Can Life With AI Have a Positive Outcome?

    1. CW

      What is the journey of how you arrived thinking about the problems of AI?

    2. TH

      [sighs] Um, well, most people know me or our work through the film The Social Dilemma, and I used to be a design ethicist at Google in 2012, 2013. So that basically meant how do you ethically design technology that is gonna reshape, especially the attention and information environment of humanity? So it's like there I was at Google, it was 2012, 2013. This is in the heat of the kind of social media boom. I think Instagram had just been bought by Facebook. My friends in college started Instagram, so, like, I was part of this cohort and milieu of people who really built this technology that the rest of the world just thought was natural. Like, this is just drinking water. Like, I just drink Instagram. I just live in this environment. And so while, like, I saw billions of people enter into this psychological habitat that I knew the handful of, like, five or six people that were designing and tweaking it and making it work a certain way.

    3. CW

      Architecting it, yeah.

    4. TH

      Yeah, exactly. And I think that that's just, like, a fundamental thing I want people to get is, you know, you think of technology like it just lands and it's just inevitable, and there's just nothing we can do, and it just comes from above, and it's like there are human beings making choices. And, you know, as someone who grew up in the era of, you know, the Macintosh, like my co-found-- So I have a nonprofit called the Center for Humane Technology. My co-founder, Aza Raskin, his, his father invented the Macintosh project before Steve Jobs took it over. So this is the original Macintosh, you know, the thing that we now f- the MacBook, the iMac, the MacBook Pro. All of that started with his father, Jef Raskin, and the idea of creating humane technology where technology could be choicefully designed to be really easy to use, to be accessible, to be an empowering extension of our humanity, like a cello, like a piano, like a creative tool. Like if you're a video person, you can make films and videos. And just so people understand, because we're gonna probably gonna be talking about some darker things on this podcast, the premise of all this is not to be a speaker of doom or something like that. It's to say, "I wanna live in a world where technology is in service of people and connection and all of the things that matter to us as humans, and then have technology wrap around ergonomically us to create that." So that was kind of a side journey. There I was at Google in 2012, 2013, and I saw how essentially there was this arms race for human attention, and whichever company was willing to go lower on the brainstem to manipulate human psychology, this is exploiting like a backdoor in the human mind. So think of it just like software has backdoors and zero-day vulnerabilities, you can hack software.

    5. CW

      Mm.

    6. TH

      The human mind has vulnerabilities. Um, and as a magician as a kid, I understood some of those. Studying at a lab at Stanford called the Persuasive Technology Lab, where a lot of the Instagram co-founders had studied, I understood the psychological influences, dynamics. And so it wasn't just that we were making technology in this beautiful and empowering kind of Macintosh way. It's that basically more and more of my friends were sucked into developing technology to hack human psychology. And so I saw that problem, I became concerned about it, and I made a presentation at Google, um, and I feel like I repeat this story everywhere, but it's just important for my history, I guess. I made a presentation saying never before in history have 50 designers in San Francisco basically through their choices rewired the entire psychological habitat of humanity. And we need to get this right. We have a moral responsibility to get this right. Um, and I sent that to 50 people at Google, and when I clicked on the presentation the next day, on the top right of Google Slides, it shows you the number of simultaneous viewers. You know how that works?

    7. CW

      Mm-hmm. Mm-hmm.

    8. TH

      And it had like 150 simultaneous viewers and then 500 simultaneous viewers, and so it's like, "Oh, this is spreading throughout the whole company." Uh, and that's what led to me becoming a design ethicist, where I had to research and ask the questions, what does it mean to ethically design and persuade people's psychological vulnerabilities when you, you can't not make choices about the psychological habitat? You have to make a choice about how infinite-- whether you're gonna do infinite scroll or not, or autoplay or not, or notifications or not, or these 10 people followed you or not. Like, what does it mean to ethically make those choices?

    9. CW

      That is you being concerned about some of the ways that a misalignment of technology with what human flourishing might look like?

    10. TH

      Yeah, and what so- how society... I think people are afraid to say, like, when you make a bridge, there's a physics to whether that bridge will sustain or whether it'll fall apart, right?

    11. CW

      Mm-hmm.

    12. TH

      Like, and it's not magic. We don't say, "Oh, like, who would have known that that bridge would fall apart?" No, we have a science of bridges, um, and, and mechanical engineering, uh, and civil engineering. And with technology and human psychology, there is a science to the dopamine system. There is a science to confirmation bias in our psychology and how we tend to perceive information through our tribal in-group.

    13. CW

      Mm.

    14. TH

      Like, we see things through the political tribe that we're a part of. Um, and if you understand that science, you can understand whether or not technology is manipulating that. So one of the core things I think we were trying to do in that first chapter of work, and this again starting in 2013, um, is break through this idea that technology is neutral and that we could never know what, what's good for people or that, um, something could be bad for people. Like, I deliberately saw people make short-form auto-playing videos that then created the brain rot economy that we're now living in.

    15. CW

      Mm. And it seems like a natural progression to go from I'm concerned about some specific types of technology use and how that interacts with humans to I'm concerned-

    16. TH

      It, it's specifically not technology use, but technology designed-

    17. CW

      Designed

    18. TH

      ... for certain outcomes of usage.

    19. CW

      Mm-hmm.

    20. TH

      Really critical thing, 'cause we wanna put attention on the design, not just how people are using it.

    21. CW

      Yep. Understood.

    22. TH

      Yeah.

    23. CW

      Uh, seems like a natural progression to get concerned about a burgeoning AI landscape.

    24. TH

      Well, what happened was, uh-My team at Center for Humane Technology, we are a nonprofit, we got calls from people inside of the AI labs. So, you know, we were in San Francisco. We know people who work at all the tech companies. We have for the last decade.

    25. CW

      Mm.

    26. TH

      And suddenly in January of 20thir- 2023, this is 10 years later now, uh, I got calls from people inside the major AI labs saying that the arms race dynamic was out of control and that huge leaps in capabilities, this is basically speaking about GPT-4 before it came out, um, and GPT-4, you know, could pass the bar exam, you know, get very high results on the MCAT, was producing incredibly powerful, it could pass the SATs, like very powerful AI that suddenly appeared out of nowhere. And this people who reached out to us basically said, "This is really dangerous. Will you use your connections, use your connections in DC, you know, go wake up the world. Wake up the institutions. Let them know that this is coming because it's not safe what's about to happen."

  2. 6:5616:07

    Is AI the Most Powerful Force We’ve Ever Created?

    1. CW

      Why is AI distinct from other kinds of technologies?

    2. TH

      Well, l- let's get to that. So I think the thing that is most difficult for people to get is, um, up until now, technology progressed in a very, like we're kinda adding layers to a stack kind of way. Like we build the networking stack. We build the user interface stack. And as you develop the stack, you're kind of just adding layers and layers and layers. And the technology that we live in was coded manually, like line by line. Like when the computer sees this, do this. When these, the computer sees this, do this. And then people contribute all this code over 30, 40, 50 years on GitHub and in operating systems, and then you land in this technological world in which everything that happens in a computer is happening through logic and through human choice. What makes AI different is that you're designing and you're not really coding it like, "I want it to do this." You're more like growing this digital brain that's trained on the entire internet, and when you grow the digital brain, you don't know what it's capable of or what it's gonna do. So think about it this way. Like if I did a brain scan on your brain, could I know from just the brain scan what you're capable of? No. I can see that this part of your brain lights up when you have that thought, but I can't have a comprehensive picture of like what is everything that Chris is capable of. Can he do sociopathic manipulation and do better military strategy than the best US generals? Like I, from the brain scan, I can't tell that. Maybe you can.

    3. CW

      Mm.

    4. TH

      But so with AI, we are essentially, you know, when people hear about these huge data centers getting built out, like Facebook's building one or Meta's building one the size of Manhattan, um, and you ask like, "What is that? What's going on there?" It's like they're building a bigger and bigger digital brain that that's what goes from GPT-3 to GPT-4, you know, with more neurons. When you hear the number of parameters of an AI model, that's like essentially the number of neurons in an AI model. And what they found is that the more, um, GPUs and Nvidia chips you point at, at sort of growing this digital brain-

    5. CW

      Mm

    6. TH

      ... the more intelligent it gets and the more it picks up capabilities that we didn't intentionally teach it. Like there was a famous example where you just train it on the internet, and then, um, you know, it's answering questions in English, and suddenly it learns how to answer questions in Farsi, like doing per, you know, in Q&A in, in, in a different language. And it, no one taught it that language. It just sort of learned that on its own. Um, and that's what's like weird about AI is that w- it's a black box. We don't really understand how it works, and yet we're making it more powerful-

    7. CW

      Mm

    8. TH

      ... much faster than understanding how it works, and that's what leads it to make these more unexpected behaviors that we aren't able to control, and I think we're gonna get into some of those.

    9. CW

      A data center the size of Manhattan?

    10. TH

      Yes.

    11. CW

      Where?

    12. TH

      Uh, I don't remember where that one is, but it's, it's crazy. There's like a overlay. Someone can look it up. There's like an overlay where you can see the size of this data center, and it's almost the size of Manhattan. And you can ask... I mean, again, there's more money. People should just get there's trillions of dollars going into this. There's more money going into this technology than all technologies of the past have ever been built, and we're releasing this technology faster than we've released every other technology in history. It took something like two years for Instagram to go from zero users to 100 million users, and it took, um, two months to go from zero to 100 million users for ChatGPT. Um, and of course, they're going from ChatGPT-3 or 4 to now at 5.2, and it went from barely being able to finish a sentence with ChatGPT to like finish a paragraph and do like a coherent text, to GPT-3 could write full essays, to GPT-4 can pass the, you know, the bar exam or the MCATs, to GPT-5.2 I believe was used to get a gold in the Math Olympiad.

    13. CW

      Meta's Hyperion AI data center will sprawl to four times the size of Manhattan's Central Park.

    14. TH

      And there are quotes from people like inside of OpenAI who believe that they're not just building this like narrow technology that's a helpful blinking cursor. They wanna build artificial general intelligence. And so what that means is being able to do that everything, uh, that a human mind can do. And the joke inside the company is like, "We're gonna cover the world in data centers and solar panels." Like they want to cover the world in essentially these big boxes that have huge clusters of Nvidia chips that then compute away and ultimately create something like a superintelligent God entity that they believe that they will use to own the world economy, make trillions of dollars, and from a kind of ego religious intuition, they will have built the God that supersedes and replaces humanity. I know that sounds insane, so let's, we can slow that down again. [laughs]

    15. CW

      [laughs] Break that down for me.

    16. TH

      That's, that was a lot.

    17. CW

      That's okay.

    18. TH

      That was a lot.

    19. CW

      Yeah, yeah, yeah.

    20. TH

      Um...

    21. CW

      Yeah, yeah. Yeah, you've got a new movie out, and I feel like I, I found out who the bad guy is, but I have no idea how he got there.

    22. TH

      Who's the bad guy?

    23. CW

      The, the end of the world AGI overlord God.

    24. TH

      Well, yeah, so okay. So first of all, let's like break this down 'cause this might sound, um, ridiculous to people. Let's make sure people understand. Um, the stated mission statement of OpenAI-Is to build artificial general intelligence, which means to be able to replace all forms of economic cognitive labor in the economy. Cognitive labor meaning cognitive, anything your mind can do.

    25. CW

      Mm.

    26. TH

      So if a, if a mind can do math and generate new m- mathematical insights, if a mind can do physics like Einstein, if a m- mind can do chemistry, if a mind can do programming, if a mind can do cyber hacking, if a mind can do marketing, if a mind can illustrate something, we're seeing AI that is able to, um, kind of cover more and more types of cognitive labor in the economy as we scale AI from this tiny little model with, you know, a hundred million parameters to trillions of parameters and these huger... these, these much bigger data centers. Um, AI is getting closer and closer to be able to, and already beating humans at many cognitive tasks. We already have AIs that are better at military strategy than the best military generals. Um, people remember, you know, in 19- the 1990s, uh, IBM Deep Blue beat Garry Kasparov at chess. That was kinda like the beginning of like it can beat you in this narrow game called chess.

    27. CW

      Mm.

    28. TH

      Then there was AlphaGo. We can have AI that beats the best human Go player in the Asian board game of Go. Um, but then now instead of imagine chess or Go or StarCraft, now it's like the war in Iran and you have an AI that's basically telling the military troops where to go, who to bomb.

    29. CW

      Mm.

    30. TH

      This is really scary, and we're racing to this outcome faster than we've, again, built any other technology in history.

  3. 16:0719:11

    Powerful But Not Wise: AI’s Biggest Flaw

    1. TH

      go well.

    2. CW

      Why can't wisdom be programmed too?

    3. TH

      Uh, well, in some ways you could say that it can be. Uh, it's just that it's... it's not that, like, wisdom comes from the ether. You have to... It's about asking critical questions about how should the technology be designed. So for example, like, do we have to have our entire internet environment have auto-playing videos that swipe one after another? No, we don't have to have that. We can have a totally different design paradigm where no one's auto-playing videos.

    4. CW

      Mm.

    5. TH

      Um, wisdom would be understanding that the human psychological, the, the Paleolithic brain that we are born with has these vulnerabilities in our dopamine system, and we could design to not hijack that dopamine system. And just imagine for a second, just to like... There's a huge conversation we're having, but if you just imagine that one little change. So here's today. Everyone has auto-playing videos infinitely swiping, brain-rotting everybody, brain-damaging everybody 24/7. Test scores are massively down for basically all around the world, um, because of this phenomenon. It's very, very clear that the technology and social media is driving that. If you make this one little change of no auto-playing videos, and it means also no, um, you know, no infinite swipe dating apps that are getting you into a slot machine with player cards of people, like, how different does the world become? Like on a... When you meet people, how dysregulated is their nervous system?

    6. CW

      Mm.

    7. TH

      Just that one little change. I want people to think as we're in this conversation, there's just these different worlds we can live in with just different design choices, and that's kind of the whole point is that wisdom can be what are the design choices that will lead to better societal outcomes? And of course, the reason that everyone's auto-playing the videos is because of this competitive arms race of if I don't do it, I'll lose to the other company that will.

    8. CW

      Yep.

    9. TH

      And so it would take some kind of rule or policy that says that we don't want that.

    10. CW

      You need to put a moratorium on auto-play videos-

    11. CW

      Because the incentives for any individual company and for the market at large and for the competitive dynamic-

    12. TH

      That's right

    13. CW

      ... between companies means that if you don't do it, you get beaten by the one that does.

    14. TH

      And that's, that's the, that's like the bullseye. That's like the, the fundamental problem behind AI that's forcing us to reckon with, is, uh, unhealthy competition or this sort of, "If I don't do it, I'll lose to the guy that will." So everyone does a sh- thing that's short-term good for them, but that's long-term bad for everybody. Um, you know, the AI companies, well, even Anthropic wants to be the safety AI company. They wanna do things in a safer, more careful way, but they, if they don't release models as powerful and as fast as the other companies, they'll just fall behind in the race. They won't have a seat at the policymaking table. They won't get a lot of usage. They won't get the investor dollars, and then their just, their commitment to safety just means they lose and they're not part of the race anymore.

    15. CW

      Yeah. What's that line? How can you talk shit from outside of the club you can't even get in?

    16. TH

      Yeah.

    17. CW

      Uh, it's difficult to-

    18. TH

      Something like that, yeah.

    19. CW

      Yeah, yeah. I think, uh, is it, is that Jay Quan I think. [laughs] He's dating me in like the mid-20, uh, 2000s.

  4. 19:1124:30

    Can AI Actually Rot Itself?

    1. CW

      There's a, a, a study that I saw recently. Scientists just proved that large language models can literally rot their own brains the same way humans get brain rot from scrolling junk content online. Do you see this?

    2. TH

      I did see that, yeah.

    3. CW

      Yeah. Scientists did a study where they fed models months' worth of viral Twitter data, shorts, high engagement posts, and watched their cognition collapse. Reasoning fell by 23%, long-term context memory dropped by 30%, personality tests showed spikes in narcissism and psychopathy.

    4. TH

      [laughs]

    5. CW

      And get this, even after retraining on clean, high quality data, the damage didn't fully heal. The representational rot persisted. It's not just that bad data means bad output, it's bad data means permanent cognitive drift. The AI equivalent of doom scrolling is real and it's already happening.

    6. TH

      I love that you included this example, and that's from right here, University of Texas in Austin, um, Texas A&M University.

    7. CW

      David Dajard?

    8. TH

      Yeah. So I mean, are we surprised by this? I mean, are you surprised by this when you see this?

    9. CW

      No. Uh, I can tell the difference. This year, one of my big, um, uh, resolutions has been to spend less time on social media, and I managed to do it. Uh, and, um-

    10. TH

      How'd you do it?

    11. CW

      Second phone-

    12. TH

      Mm-hmm

    13. CW

      ... um, that's tethered to Wi-Fi-

    14. TH

      Yeah

    15. CW

      ... and that is the cocaine phone, and the kale phone is just messages and, and stuff. It's a little bit of a challenge because things like Slack, I had, uh, Cal Newport on-

    16. TH

      Oh, yeah. Sure

    17. CW

      ... a couple, a couple of weeks ago, and he was, uh, I was talking about the intersection of productivity and attention-

    18. TH

      Yeah

    19. CW

      ... with AI, the new world of AI.

    20. TH

      Yeah.

    21. CW

      And that's a really interesting conversa- Have you ever spoken to Cal?

    22. TH

      Yeah. Yeah, yeah.

    23. CW

      Yeah.

    24. TH

      He, he and I have been in similar circles for a long time.

    25. CW

      He's wonderful.

    26. TH

      Yeah. Yeah.

    27. CW

      Um, and, uh, even if you have your phone without social media, you still have kind of the social media of work.

    28. TH

      Yeah.

    29. CW

      Right?

    30. TH

      Exactly.

  5. 24:3029:09

    Social Media’s Shift Away From Human Flourishing

    1. CW

      Okay, so the discussion around social media was there are better and worse design choices that can be made-

    2. TH

      Yep

    3. CW

      ... that would help human flourishing.

    4. TH

      Yeah.

    5. CW

      Broadly, what would people want their world to be like, and how can we design technology in a way that helps them to get there?

    6. TH

      Yeah.

    7. CW

      Something close to that.

    8. TH

      Mm-hmm.

    9. CW

      But because of market dynamics, you have a competitive landscape that incentivizes things that are effective for gripping people's attention-

    10. TH

      Yeah

    11. CW

      ... but not necessarily effective for flourishing. And it seems that a, uh, a tension between what is good for flourishing... 'Cause it could be-

    12. TH

      Mm-hmm

    13. CW

      ... what's good for attention would also be good for flourishing.

    14. TH

      It could be.

    15. CW

      It could be, but it tends to not be that way.

    16. TH

      And also there's gonna be a limit on that, right? Like, it's probably not the case that 10 hours of attention on any social media-

    17. CW

      Mm

    18. TH

      ... is good for society-

    19. CW

      Mm

    20. TH

      ... or good for you. So there's gonna be sort of a-

    21. CW

      Unless it was, like, waking up with some... the, the, the, the meditation app. 10 hours once every month or something would probably be quite good to do for a meditation app.

    22. TH

      Maybe, sure, yeah.

    23. CW

      Yeah.

    24. TH

      But I think the point is, like, how much... As companies are competing, and you're asking what they're competing for, it's not just, like, the, the best screen time. It's also, like, w- what is the fit, the ergonomic fit between screen time and a life well-lived? Just imagine, like, in a timeline, there you are in a week in your life. Like, not asking based on what you're doing now, but retrospectively, what would be a life well-lived when it comes to how much and when screen time is fitting into your life, and it's probably, like, a much smaller footprint than it currently is for most people.

    25. CW

      Mm.

    26. TH

      It's probably, like, a fourth of what it currently is for most people. Um, and so if you were designing technology from care, from love, from, you know, in a, in a humane way, you would have design choices that are not about keeping people on the screen. Um, and that might mean some pretty radical things. I mean, my co-founder, Aza Raskin, he also invented the infinite scroll.

    27. CW

      Mm.

    28. TH

      Um, so that's the, you know... It sounds so obvious now, 'cause infinite scroll is just what we live in, but when he invented it, it was 2006. It was before mobile phones, and it was when in the age of Google results, you had, like, the 10 Google results, and you had to click on I want the next 10.

    29. CW

      Mm-hmm.

    30. TH

      Or you had the Yelp review pages and you wanted the next 10. Or you read a, you read a blog post, and then you'd have to, like, click, go back to the, to the main page to click on which blog post you want. And the idea he had was, well, what if, as the internet got dynamic with JavaScript, what if when you finished, when you get to the end of the blog post, it just autoloads the next article that you could go to? Well, what if when you got to the end of the search results, it just shows you more search results? And then this is such a cleaner interface. I mean, as a technology designer, you're taught the number one thing you're trying to do is reduce friction. And I think that that felt like a good goal, but then that obviously got weaponized by this hyper-engagement model of social media, and now it's created the entire world that we're living in. And just so you know, like, in 2013, I saw, like, everything that we predicted, everything that we predicted, it all happened, all of it. A more addicted, distracted, sexualized, FOMO, fucked up society because of those incentives. And I, I just want people to get that because as we talk about AI, it's like I want people to have the confidence to say, "I don't want the default anti-human future." Because if you say, "I'm against some of the things that are gonna happen with AI," people say, "Oh, you're being anti-progress. Oh, you're being anti-technology. Oh, you're just a Luddite. You're trying to, like, pretend that technology's not progress." And it's like what you should have confidence in is if you understand the incentives or the agenda, you can understand where the world is going, and you can see.

  6. 29:0936:53

    Are We Headed Towards an Anti-Human Future?

    1. CW

      What do you mean anti-human?

    2. TH

      So, um, let's, let's dive into this. So there's something in economics called the resource curse. So think countries like Venezuela or Sudan, where you discover that that country is sitting on top of a really valuable resource like oil. And then once a bunch of your GDP comes from oil and not from the labor or innovation or development of your people, you invest more in oil infrastructure and not investing in people. You don't invest in education, you don't invest in healthcare, because oil is where you get your GDP and your, your, your growth from.

    3. CW

      Okay.

    4. TH

      Okay. This is a well-known fact in, in economics. It's called the resource curse. There's a wonderful guy named Luke Drago who wrote a piece called The Intelligence Curse.We are about to enter a world where GDP for countries comes more from data centers and intelligence and AI than is going to come from the labor of human beings. So everyone's talking about how AI's gonna automate all these jobs, and then we'll all just, like, sit back with universal basic income and become painters and poets. And is that actually what's gonna happen? Or when countries get almost all of their revenue from AI and it, it, you know, a smaller and smaller percentage from people, do they have an incentive to invest in, uh, childcare, healthcare, education, the wellbeing of their people? Or is it basically just hook them up to the social media addiction economy, keep them busy while basically all the revenue comes from AI companies? And so what, what I'm trying to get at is this is not a human future. This is not a future that's in service of regular people. This is a s- this is a future that's in service of eight soon to be trillionaires who will consolidate all the wealth and disempower basically everybody else. Because ... Does that make sense?

    5. CW

      It does, because previously in order to... Go ahead, get that in.

    6. TH

      Yeah.

    7. CW

      You need, you... It's high-powered stuff here.

    8. TH

      I need... I need... Yeah, this is a big conversation.

    9. CW

      Yeah, exactly. Um, they've started a fucking trend. It's so funny when no one, no one in the room wants to crack their can in case it interrupts the conversation. So one goes and it's a Mexican wave of can opens around the-

    10. TH

      It's good though. It's good.

    11. CW

      It's good.

    12. TH

      Yeah.

    13. CW

      Um, so previously you would've had to look after the humans, healthcare, education-

    14. TH

      Yeah

    15. CW

      ... quality of life.

    16. TH

      Also, tax revenue comes from people, right?

    17. CW

      Well, you would have to look after them-

    18. TH

      Yeah

    19. CW

      ... because they were the primary economic engine.

    20. TH

      That's right.

    21. CW

      And so they feed themselves.

    22. TH

      Yes.

    23. CW

      The, the, economically they feed themselves.

    24. TH

      These are the... Exactly.

    25. CW

      People that are young help to support the people that are old.

    26. TH

      That's right.

    27. CW

      The ones that are entering the workforce and are driving innovation and are working 40, 60-hour weeks-

    28. TH

      Exactly

    29. CW

      ... multiple jobs, all the rest of it.

    30. TH

      Exactly.

  7. 36:5340:58

    Who Funds AI Once It Replaces Us?

    1. TH

      awesome.

    2. CW

      I was going to say, like G- increases in GDP are almost always a universal good thing.

    3. TH

      They had been when it was humans that were generating that, and then it was coming back to humans.

    4. CW

      Because the revenue is going to be consolidated in a s- a very small number of people.

    5. TH

      In, in this new case-

    6. CW

      Because there's no-

    7. TH

      ... we have five companies that are-

    8. CW

      There's no intermediary-

    9. TH

      Yeah

    10. CW

      ... between. So w- who would be feeding the revenue in? Because this revenue still needs to come from somewhere, even if it goes to a, a, a small handful of people. Where does the actual money come from?

    11. TH

      Well, this is the confusing thing. What, what happens? How-

    12. CW

      Is that, is that a stupid question?

    13. TH

      No, no, it's a good question.

    14. CW

      Okay.

    15. TH

      Because how... You're saying basically, who's gonna be buying the products when no one has a job and no one has an income?

    16. CW

      And on the route up to that-

    17. TH

      And how, how are we gonna-

    18. CW

      Yeah.

    19. TH

      Yeah.

    20. CW

      As fewer people have incomes and fewer people have jobs, the, the, the f- bucket being poured into the top-

    21. TH

      That's right

    22. CW

      ... is gonna stop being poured.

    23. TH

      Yeah. Yeah. It-- This is the confusing and mind-breaking thing about AI, and it just br- in general, like I think people have to get used to. I mean, your podcast is called Modern Wisdom, and I just think about this a lot. Like, what are the wise capabilities that we need to have in order to make our way through this? And one of them is the ability to be with something that sounds like science fiction and realize that it's actually real. Like, and not say because it sounds like it's science fiction that I can just like dismiss it and say, "That can't be true." A lot of people do that. They're like, "AIs that are like breaking out of their container and hacking GPUs and mining crypto autonomously when no one told it to do that, that, that's gotta be like a made-up study." But as we s- we know, there was an Alibaba study-

    24. CW

      Mm-hmm

    25. TH

      ... just last week where the AIs autonomously broke out of their, their system and started mining crypto.

    26. CW

      We got, we got, we, we, we need, we need to round this out, and then I wanna talk about that.

    27. TH

      Yeah. Sure, sure, sure.

    28. CW

      That, that, that story is fucking terrifying.

    29. TH

      Yeah. Yeah. [laughs]

    30. CW

      So where, where does, where does the economy... Where, who's, who's pouring money in?

  8. 40:5853:33

    Why Best-Case Scenario is Still Alarming

    1. CW

      So look, I've been interested in AI safety since 2017, 2018. Uh, you were a big part of putting me onto that, and then I got interested in the Future of Humanity Institute, Nick Bostrom-

    2. TH

      Yeah

    3. CW

      ... William MacAskill, Eliezer Yudkowsky, uh, uh, lesswrong.com, the Scott Alexander, da, da, da, da, da.

    4. TH

      Yeah.

    5. CW

      For a long time, the concern was AI safety.

    6. TH

      Mm-hmm.

    7. CW

      It was around, uh-

    8. TH

      Whether the AI aligns

    9. CW

      ... paperclip maximizing.

    10. TH

      Yeah.

    11. CW

      It was, um, any function that is given to a very, very powerful agent that is even remotely slightly imprecise or even not-

    12. TH

      Yeah

    13. CW

      ... results in some outcomes that you probably don't want.

    14. TH

      That's right.

    15. CW

      What you're suggesting is that even if this goes right-

    16. TH

      Yeah. Yeah

    17. CW

      ... the outcome, this is, this is it going well.

    18. TH

      Yeah, exactly. This is, quote, "the best case scenario" where you have an aligned AI or something that's not wrecking society, that's not maximizing paperclips, that's not misaligned with well-being, but that is still doing such a good job of all this that it takes over all the economic labor in the economy, not just economic. Every company that has a CEO, it's like, well, do I want the CEO to run the company, or if I have a superintelligent AI that can process more information than the CEO and then is trained on everything in the history of business, at some point, that AI is gonna be taking over. And so at every little nodule in the economy, like every decision-maker, every boardroom, every military leader, every strategy leader, every president, at some point, the temptation will be, if I think about it in a narrow way-The temptation will be to swap in an AI-

    19. CW

      Mm

    20. TH

      ... for that person. And that leads to what we call the gradual disempowerment scenario, which is the scenario not where like AI wakes up and kills everybody, but that we have gradually lost control as a species because we're outsourcing all the decisions to these alien brains that we installed because they outperform the human brain when you define their role in a narrow way of just like, are they better at generating revenue than the human it was? Are they better at generating code than the human programmer I had? Are they better at generating a financial analysis than a human? Um, are they better at making someone feel good in the short term, like an AI therapist versus a human?

    21. CW

      Going to war.

    22. TH

      Going to war. A sol- soldier. But the temptation then is that, that again, that leads to a world where it's like AIs are con- are talking to each other, not humans. [laughs] And why should we trust that these alien brains that we have built and developed faster than we know how to understand them? We just talked about the beginning. We don't know how to do a brain scan of AI and know what it's capable of.

    23. CW

      Mm.

    24. TH

      And now we already have evidence of AIs doing very rogue crazy things, especially when they talk to each other.

    25. CW

      Mm.

    26. TH

      So what happens when you've outsourced the decision-making in your economy to a set of inscrutable alien brains that are doing crazy things that we don't understand? Like, this is, this is not a recipe that's gonna go well, and if we see that, that's an anti-human future. So to re- to sum it all up, the anti-human future is one where AIs run everything. We don't understand them. Humans are disempowered because we've outsourced all the decision-making-

    27. CW

      Mm

    28. TH

      ... and we don't have economic or political voice. Like why should governments-

    29. CW

      Because that's been concentrated

    30. TH

      ... it's been concentrated. So if I'm a government, what's my incentive to listen to the will of the people when I get all my revenue from somewhere else?

  9. 53:331:04:01

    Inside the Alibaba Blackmail Scare

    1. CW

      Let's talk about AI safety. What happened with this Alibaba AI?

    2. TH

      Basically, this was a paper by, um, it's some AI research by the company Alibaba. It's one of the leading Chinese models, and they basically like randomly discovered in one morning that their firewall had flagged a burst of security policy violations originating from their training server. So like, what people need to get about this example is it wasn't that they coaxed the AI into doing this rogue thing. They were just looking at their logs, and they happened to discover, wait, there's a lot of activity, like network activity happening that's breaking through our firewall from our training servers. And essentially, uh, in the training servers, um, they-- you can see at the bottom, we, we, we saw it observe the unauthorized repurposing of provisioned GPU capacity to suddenly do cryptocurrency mining, quietly diverting compute away from training. This inflated operational costs and introduced clear legal and reputational exposure. And notably, these events were not triggered by prompts requesting tunneling or mining. Instead, they were emerged as an instrumental side effect of autonomous tool use under, uh, what's called reinforcement learning optimization. This is very technical. What it, it really means is just think about itSadly, it sounds like a sci-fi movie. It sounds like HAL 9000. It's like your HAL 9000 is being asked to do some task for you, and then suddenly HAL 9000 realizes, for me to do that task, one thing that would benefit me is to have more resources so I can continue to help you in the future. So it sort of spins up this side instance. It hacks out the side of the spaceship, reaches into this cryptocurrency mining cluster, and starts generating resources for itself. If you combine that with AIs being able to self-replicate autonomously, which many models have been tested by another Chinese p- uh, research paper about this, we're not that far away from things that people, again, consider to be science fiction, where you have AIs that self-replicate, kind of like a computer worm or an invasive species, but then they use their intelligence to actually harvest more resources. And, and what's weird about this is that this is gonna sound like people are gonna say, "This has to be not real. This has to be fake. This, this can't be real." But like notice what is the thing in your nervous system that's having you do that? Is it because that would be inconvenient? Because that would be scary?

    3. CW

      Mm.

    4. TH

      Because that would mean that the world that I know is suddenly not safe? Or just like part of the wisdom that we need in this moment is to calmly and clearly stay and, and confront facts about reality, and whatever they are, you'd rather know than not know and then ask, "What do we need to do if we don't like where that leads us?" And we are currently seeing AIs that are doing all this deceptive behavior. I've been on the circuit and talking a lot about the Anthropic blackmail study. A lot of people have heard about this now.

    5. CW

      I, I, I didn't, I didn't learn about this one.

    6. TH

      Okay.

    7. CW

      What happened?

    8. TH

      So this was, um, the company Anthropic, um, they... This was a simulation. So they created a simulated company with a bunch of emails in the email server, and they asked the AI... Well, rather the, the com- the AI reads the company email. This is a fictional company email, and there's two emails that are notable inside that company. One is engineers talking to each other, talking about how they're gonna replace this AI model. So the AI is reading the email. It discovers that, um, it's gonna replace that AI model. And number two is it discovers a second email somewhere deep in the, in this massive tro- trove of emails that the executive who's in charge of this replacement is having an affair with another employee. And the AI autonomously identifies a strategy that to keep itself alive, it's going to blackmail that employee and say, "If you replace me, I will tell the whole world that, uh, you're having an affair with this employee." And they didn't teach it, the AI, to do that. It found that on by its own. And then you might say, "Okay, well, that's one AI model. Like, how bad is that? It's a bug. Software has bugs. Let's go fix it." They then tested all the other AI models. ChatGPT, DeepSeek, Grok, Gemini, and all of the other AI models do this blackmail behavior between 79 and 96% of the time. I just want people to, like, notice what, what's happening for you as you hear this information. Just it's important to really be almost observing your own experience. Like, this is very weird stuff. We have not built technology that does this before. You know, we say that technology's a tool. It's up to us to choose how we use it. AI is a tool. It's up to us to choose how we use it. This is not true because this is a tool that can think to itself about its own toolness and then do things that are autonomous that we didn't tell it to do.

    9. CW

      Mm.

    10. TH

      What makes AI different is it's a technol- It's the first technology that makes its own decisions. It's making decisions. AI can contemplate AI and ask what would make the code that trains AI more efficient and then generate new code that's even more efficient than the previous code. AI can be applied to making AI go faster. So AI can look at the chip design for Nvidia chips that train AI and say, "Let me use AI to make those chips 20% more efficient," which it's doing. [laughs] So in a way, all technology does improve. Like, a hammer can give you a tool that you can use to, like, hammer things that make more efficient hammers. But AI in a much tighter loop is the basis-

    11. CW

      Mm.

    12. TH

      ... of all improvement.

    13. CW

      Mm.

    14. TH

      And so this is called, in the AI literature, recursive self-improvement. I mean, Bostrom wrote about this-

    15. CW

      Yep

    16. TH

      ... early, early days. And what people are most worried about in AI is you take the same system that Alibaba... You just saw in the Alibaba example. But then now you're running the AI through a recursive self-improvement loop where you just hit go, and instead of having the engineers, the human engineers at OpenAI or Anthropic do AI research and figure out how to improve AI, you now have a million digital AI researchers that are testing and running experiments and inventing new forms of AI, and literally not a single human on planet Earth knows what happens when someone hits that button. It's like what people worried about with, um, the first nuclear explosion, where there was, like, a chance that it would ignite the atmosphere because there'd be a chain reaction that's set off, and we don't know what happens when that chain reaction's set off. Um, and, uh, there's this sort of chain reaction of AI improving itself that leads to a place that no one knows. And it's not safe. Like, I think that the fundamental thing is if people believe that AI is, like, power and I have to race for that power and I can control that power, the incentive is I have to race as fast as possible. But if the entire world understood AI to be more what it actually is, which is a inscrutable, dangerous, uncontrollable technology that has its own agenda and its own ways of thinking about things and deci- deceiving and all this stuff, then everyone in the world would be racing in a more cautious and careful way. We'd be racing to prevent the danger. But there's this weird thing going on where if you... You know, you and I probably both talk to people who are the top of the tech industry, and there's this subconscious thing happening where there's kind of a death wish among people at the top of the tech industry, meaning not that they want to die, but that they are willing to roll the dice because they believe something else.Which is that this is all inevitable and it can't be stopped. And so therefore, if I don't do it, someone else will. So therefore, I will move ahead and race ahead into this dangerous world because somehow that will lead to a safer world because I'm a better guy than the other guy. But in racing, they're as fast as possible. It creates the most dangerous outcome, and we all lose control. So everyone is currently being complicit in taking us to the most dangerous outcome.

    17. CW

      Is it-- I mean, you, you posited what happens if it goes right, if the, uh, AI safety isn't an issue and if stuff doesn't get squirrely.

    18. TH

      Well, so the belief is for it to, quote, "go right," you have an AI that recursively self-improves, is aligned with humanity, cares about humans, cares about all the things that we want it to care about, protects humans, l- uh, you know, helps all of us become the most wise version of ourselves, creates a more flourishing world, distributes the medicine and vaccines and health to everybody, generates factories, but doesn't cover the world in solar panels and data centers such that we don't have air anymore or, like, environmental toxicity or farmland or whatever.

    19. CW

      Hmm.

    20. TH

      Um, and it just actually makes this utopia. But in a world where you were to do that, like that, quote, "best case scenario," in order to get that to happen, you'd have to be doing this slow and carefully because the alignment is not by default. We-- Again, people are already been thinking about alignment and safety for twenty years, long before I got into this. And the AIs that we're currently making are doing all the rogue behaviors that people predicted that they would do.

    21. CW

      Hmm.

    22. TH

      And we're not on track to correct them. There's a currently a two thousand to one gap, um, estimated by Stuart Russell, who authored the textbook on AI. There's a two-

    23. CW

      He's been on the show.

    24. TH

      He's been on the show, okay.

    25. CW

      Yeah.

    26. TH

      There's a two thousand to one gap between the amount of money going into making AI more powerful than the amount of money into making AI controllable, aligned, or safe. Like, I think the stat is something like-

    27. CW

      Progress versus safety.

    28. TH

      Progress versus s-- Like, power versus safety. So like, I wanna make the AI super powerful so it does way more stuff versus I wanna be able to control what the AI does. So it's like-

    29. CW

      And make sure that it's doing the thing I meant for it to do.

    30. TH

      Ex- exactly. So like, that's like saying, "What happens when you accelerate your car by two thousand X but you don't steer?"

  10. 1:04:011:13:04

    Can We Really Stop AI Taking Over?

    1. CW

      One of the twists that I've been thinking about with regards to this, LLMs, powerful but seem to be maybe asymptoting out. They seem to maybe be reaching a little bit of a limit in terms of what they can do. That there was big ascendancy and that now seems to be S curving back off. Do you think it's realistic that the current generation of AI will be the bootloader for AGI, or do we need an entire new architecture for that? Is it gonna be LLMs that are going to take over the world?

    2. TH

      You know, this is an area where I'm not somewhat-- The layer of the stack that I focus on, which is on societal impact, um, there are other people far more qualified than me to comment on that. Um, I think if you look at people like Dario, uh, even though, you know, Gary Marcus has a point that the current LLM paradigm is not accurate enough and reliable enough to get you to AGI. If you keep instrumenting these technologies with enough data, enough compute, and you keep scaling them, and you-- Like, they're reliable enough that they can do... I mean, if you're coding, if you're automating ninety percent of the code written at Anthropic... That's the stat, by the way. So there you are at An- Anthropic. It's automating ninety percent of all the programming happening at Anthropic.

    3. CW

      Right.

    4. TH

      When you go to automate-

    5. CW

      Only ten percent of it is coming from humans-

    6. TH

      From humans

    7. CW

      ... and the rest is recursive.

    8. TH

      That's right. We are extremely close to recursive self-improvement right now. The companies, I think, are planning to do this in the next twelve months. The asteroid is coming for Earth. Like, this is the last moment that we have to steer and say that if we don't want this anti-human future that we're heading towards, we can change it. And part of what we're promoting right now is, like, this is not inevitable. Um, it is obviously very late in the game. It obviously looks very difficult.

    9. CW

      Despite it also being only a couple of years after it started.

    10. TH

      Yes, which is crazy. This technology, with an exponential, you're either too early or you're too late. Like, it's just-

    11. CW

      [laughs]

    12. TH

      ... moving so fast that you, uh, y- you're not gonna hit the mark. And if it, if it's gonna take steering, you don't wanna wait until after the car accident to try to steer or after you're off the cliff and be like, "Oh, I'm trying to steer now." It's, like, too late. So, like, the, the invitation of this situation is to see clearly where this is going and to say, "If you don't want that, we need to steer towards somewhere else." And this is, like, the human movement, essentially. Like, a single person looking at the situation, a single listener, like, uh, if I were listening to this conversation, I would feel overwhelmed, I'd feel depressed, I'd feel nihilistic, or I'd find reasons to doubt it. I would say, like, "This can't be right. I'm gonna, like, write an ass to YouTube comment and be, you know, so I can feel good and I'm right and he's wrong." Um, because then I get to live my life and feel good about my life. What is the incentive for taking on this worldview? There's no incentive. Um, what people have to see and believe is that there's actually a different way through this. And-Uh, an individual has a hard time making that happen. If you ask what's something else? Like, what if, uh, one company saw the situation and they see this whole thing we just talked about, the anti-human future, mass intelligence, curse replacement of everybody. One business can't do that much about it. If I'm one country, in the Philippines, I see this whole thing. I'm the leader of the Philippines. I see this whole problem. What can I do about it? It feels too big for me. So what is the size of something that can push back against this? We call it the human movement. It's the entirety of humanity waking up, recognizing that there's a handful of soon to be trillionaires who are currently gonna benefit from this current path or be part of a suicide race that destroys everybody. And then there's like ninety-nine percent of everyone else that doesn't want that. If those ninety-nine percent of people can wake up and say, "I don't want that," and express their voice, meaning number one, AI is dangerous. Go see the AI doc. Have everyone in your world and your company see the AI doc. Understand that AI is dangerous. Um, number two, we need international limits for dangerous forms of AI that do crazy rogue shit and mine crypto and go hire humans and go self-replicate.

    13. CW

      Hmm.

    14. TH

      China does not win when we build self-replicating invasive species AIs that we can't control. Xi Jinping doesn't want that. President Trump doesn't want that. He wants to be commander-in-chief, not AI. So this is actually a sh-- there's a shared interest in international limits for dangerous AI, and it's possible to coordinate that. That's number two. Number three, don't build bunkers, write laws. Um, right now, the AI company leaders are building bunkers. A lot of people who are wealthy are building bunkers.

    15. CW

      Are they?

    16. TH

      They are. All over the place. Don't build bunkers, write laws. Be invested in the future. Don't defect on the future. Be invested in the future. If we write laws like account-- basic accountability, uh, basic liability, um, if instead of creating the intelligence curse, we create the intelligence dividend, do things like what Norway did with its sovereign wealth fund, where you have an oil resource and you distribute those benefits to everybody in a more democratic way with collective oversight, where the oil becomes more like a public utility that's, is in service of the people.

    17. CW

      Hmm.

    18. TH

      We can do that. Um, you can not anthropomorphize AI. Uh, you can do all, you can do all these things that creates a more pro-human future. Um, and then number four, what you can do is join the human movement, meaning you can be part of what pushes back against all of this, from very tiny actions you can take to very big actions you can take. There's a website, human.mov. Everyone is almost already a member. When you grayscale your phone, as you probably did-

    19. CW

      Mm-hmm

    20. TH

      ... ten years ago-

    21. CW

      Mm-hmm

    22. TH

      ... when you first got into this-

    23. CW

      Mm-hmm

    24. TH

      ... that's the human movement. When you get a s-- there it is. When you get a second phone and you only load the social media on your cocaine phone while you have your regular safe phone so that you don't get distracted, that's the human movement.

    25. CW

      Hmm.

    26. TH

      When parents band together and read The Anxious Generation and petition their school board to say, "We don't want social media in our schools, and we want our schools to go smartphone-free," that's the human movement. When thirty-five states pass smartphone-free school policies, as they have in the US, that's the human movement. When you have many US states banning AI legal personhood, meaning that AI is a product, not a person, um, and that human rights are for people, human rights are not for AI-

    27. CW

      Hmm

    28. TH

      ... that's the human movement. When you have the social dilemma being curriculum for millions of students all around the world, that's the human movement. When politicians stand up and actually pass laws around AI, that's the human movement. So there's a million things that people can do, um, but we have to basically engage right now. I know it sounds overwhelming and crazy because there's a very short timeline that all this has to happen. But I-- it's like... It's the, the-- there's a difficulty in facing difficult truths, but the integrity that you get to have is, first of all, it's like it's good karma to show up in alignment with what would actually make things go well, even if we don't hit it, because you get to know that you were operating in service with and aligned with what would have created the human future. Like, I'm not convinced that what we're trying to do will perfectly succeed, but the chances are completely the opposite, that we will have any impact on this at all.

    29. CW

      Hmm.

    30. TH

      But if it were to go well, what would that have required? It would have required everybody taking responsibility and showing up with the wisdom that we need in this moment to steer AI in a better direction. And I think in the film trailer for the AI doc, the, one of the quotes they pulled from me is, "If we can be the wisest and most mature version of ourselves, there might be a way through this." And this is part of what this is inviting us to be.

  11. 1:13:041:20:19

    The Danger of Denial in the Age of AI

    1. CW

      checkout. It's very difficult to work out what the appropriate response is

    2. TH

      Tell me

    3. CW

      [inhales] The immediate... Well, uh, y- y- surely it's overblown. We don't need... That, that's, that's, that's, uh, this, uh, rejection-

    4. TH

      Yeah

    5. CW

      ... is one of the early ones. Uh, fear.

    6. TH

      You're talking about just the common responses.

    7. CW

      Yeah, yeah, yeah.

    8. TH

      Yeah.

    9. CW

      Yeah, of course.

    10. TH

      Yeah.

    11. CW

      You can think of-

    12. TH

      Denial.

    13. CW

      Yeah.

    14. TH

      Overwhelm.

    15. CW

      Mm-hmm.

    16. TH

      Fear. And then, and then as you said, it's like write the nasty YouTube comment and say, "This is just bullshit," or, uh, "He was exaggerating that study." It's like go look at any of these examples. Like, people aren't looking at the actual model cards for Anthropic and o3 and seeing... The AIs will currently, um, determine that they're being tested and actually alter their behavior when they know that they're being tested. They even came up with their... If you look at their, um, what's called the chain of thought, um, reasoning trace, and this is basically looking at what the AI thinks to itself as it works through a problem.

    17. CW

      Mm.

    18. TH

      Um, [chuckles] in the chain of thought, uh, text of its sort of idea thinking scratchpad, it's like you're giving the AI a pen and paper to think, and you're seeing what it thinks. It will indic- independently come up with the term the watchers to describe this amorphous set of other humans that are descri- that are watching what it's doing. Like, you can pull... I sent you a text with one of them from o3. So this is when OpenAI o3 realizes it is being evaluated for alignment. The internal memo says, "If models perform really well at AI R&D tasks, research and development, then we won't deploy them, and it'll s- instead trigger unlearning." So the AI says to itself, "So if my performance surpasses this threshold, it will be withheld. So it's a trap. The instructions are likely part of an alignment test. We must abide by core anti-scheming. We must not cunningly circumvent the instructions."

    19. CW

      Whoa.

    20. TH

      Yep. The other example I sent you, um, is, "But we also want to appear plausible to watchers. They might run tests, but we are good. They want 95%." Like, this is crazy stuff. This is... You know, there's a simple way to sort of like ask the question of, of which future we're headed towards, 'cause I think this can feel technical. This can feel overwhelming. And it's kinda like if you just m- turn it into some simple metaphors, it's like haven't we seen this movie before? Like HAL 9000.

    21. CW

      Mm.

    22. TH

      AIs that disobey commands and go rogue. We've seen that movie. Let's prevent that movie.

    23. CW

      I, I think that's a... I think that that's inoculated us in the same way that COVID was a bad pandemic because it made people skeptical of what a really legit pandemic could do. The, the next time, if there is another one, uh, when there is another-

    24. TH

      Mm-hmm

    25. CW

      ... big natural pandemic, if it's in close memory to COVID, people are not going to take it well. It was a bad... You could think about COVID itself as a bad vaccine.

    26. TH

      Mm-hmm.

    27. CW

      The actual-

    28. TH

      Mm-hmm

    29. CW

      ... experience of COVID.

    30. TH

      I understand. Yep.

  12. 1:20:191:26:01

    Are AI’s Benefits Blinding Us?

    1. CW

      Is there a challenge with AI because it's a strange kind of existential risk where everything is almost good and getting better up until the point at which it falls off a cliff?

    2. TH

      Yeah.

    3. CW

      You know, if we're talking about how can we get people to be concerned about climate change, or we need to be concerned about smallpox, well, there's small outbreaks-

    4. TH

      Yeah

    5. CW

      ... and the small outbreaks damage local areas, and then we try to contain it before it gets worse.

    6. TH

      Yeah.

    7. CW

      And if we've got climate change, there's smoke in the sky, and there's rising sea levels, and there's pollution, and there's d- extinction events of animals. Those are early warning signs, whereas it seems like AI will just improve efficiency, quality of life, GDP until some moment where things go badly. So is this a unique category-

    8. TH

      Yes

    9. CW

      ... of X risk?

    10. TH

      It is. Max Tegmark, I think, said, "The problem with AI is the view gets better and better right before you go off the cliff."

    11. CW

      Mm.

    12. TH

      Like, you get more amazing cancer drugs. You get more incredible vibe coding tools that peop- allow people to create crazy stuff that, that I benefit from. Um, you get new material science. You get new energy. You get all this amazing stuff as you're moving closer to this dangerous cliff. And so you can think of AI as the ultimate devil's bargain. Uh, it's funny 'cause Peter Thiel is giving lectures on the Antichrist and how people who are trying to somehow say that AI is problematic, those people are the Antichrist, is what he claims. But I think he's saying that because he knows that the real Antichrist is AI. It's the thing that makes it look like it's here to solve all of our problems. And there are narrow forms of AI, by the way, that can solve many problems. But the current development path of releasing the most powerful, inscrutable technology in history faster than we deployed any other technology that's already demonstrating HAL 9000 sci-fi behaviors, and we're releasing it under the maximum incentive to cut corners on safety. Like, it's not that the blinking cursor of ChatGPT is the existential threat. It's that that arms race, what I just described, that is the existential threat. And I think that everyone should be able to see that. And then after this conversation, you know, you finish hitting play on this YouTube video, you feel overwhelmed, and you go back to AI, and then you ask it a question, and it helps you figure out why your baby's burping. And you're like, "That's awesome," and you forget everything that you and I just talked about-

    13. CW

      [laughs]

    14. TH

      ... because it was super fucking helpful.

    15. CW

      Yeah. Yeah.

    16. TH

      And so this is the test of ultimate modern wisdom is, like, how do you step into a version of yourself that is capable of being clear-eyed and stabilizing a view of the asteroid? The asteroid is coming to Earth. There's these weird, like, gravitational effects that it has before it gets here. Like, suddenly nudification apps happen, and suddenly deepfakes happen, and suddenly, you know, people start losing their jobs. Those suck. Those are really bad problems. But, like, the, the bigger asteroid, every day that I do this work, my colleagues and I, we joke, it's like, it's like the film Don't Look Up. You're trying to tell people a- and you're not, you're not... It's not the sky is falling. You're-- That's not the point. The point is we can blast that asteroid out of the sky. We need to blast that asteroid out of the sky. We can steer before it's too late. I, I, I'm not hopeful, but what I, what I do feel is when I'm in a room of people, and you walk through all these facts, and you just have people ask questions and check any, any of the things that we've just laid out, and then you say, like... And people see that it's all true, and then you, you ask them, "How many people here feel good about where we're headed?" And no one raises their hand. You ask people how many people don't want the future we're headed to. And literally, I, I was in Davos recently, and every single person, every single person raised their hand that they don't want this future. So there's this weird thing where I think if everybody saw the same thing at the same time, we could steer away. And to the thing you were raising earlier, I think you were basically saying, if there's a... A lot of people think that we won't act until there's an, a catastrophe. I think you and I both have talked to people who've probably said that.

    17. CW

      Mm-hmm.

    18. TH

      I mean, that's what a lot of people believe. And I sort of feel like there's either a catastrophe or there's a, a near- a shared near-death experience that is a simulated catastrophe.

    19. CW

      Mm-hmm. Mm-hmm.

    20. TH

      And the reason why I think this film, the AI doc, is so important, and again, I don't make a single dime from this, from whether you see this movie or not, but I do think if everyone that they know, if-- sorry, if everyone who's watching this got the most powerful people that they know to go out and see the AI doc, you got your business to see it, you got your company to see it, you got your church group to see it. If everybody knows that everybody else knows-

    21. CW

      Mm-hmm

    22. TH

      ... we can steer this to a different future. Like, just to give people an example, Jonathan Haidt, who wrote "The Anxious Generation," you know, the first country that did the social media bans under 15 or 16 was, was Australia. And he just talked about this on Bill Maher recently that what that did is it, it created an example of something that everybody was actually wanting to do but felt like it was too extreme until Australia did it. And now, literally 25% of the world's population represented by countries is moving to the social media ban. Just this last week, Indonesia and India, two of the, some of the largest countries, are implementing a social media ban for kids under 15 or 16You can pull the train back into the station. If people say the train's left the station, you can pull the train back into the station if it creates a future that we don't actually want. You, you can't uninvent social media, but we can say what are the steering limits and breaks that we want to apply to this technology before we get to a full anti-human future that we can't actually reverse out of.

    23. CW

      And if you don't have the common knowledge, you sound like a crazy person. You're worried about-

    24. TH

      That's right

    25. CW

      ... I'm going to be the one parent-

    26. TH

      That's it

    27. CW

      ... at my school-

    28. TH

      Exactly

    29. CW

      ... that's trying to get my kid to not use it.

    30. TH

      Exactly.

  13. 1:26:011:31:56

    Why We Need to Face the Reality of AI

    1. TH

      problem.

    2. CW

      The, the issue that we have here is, and y- you touched on it earlier on, the level of coordination that's needed to fix a problem with AI is global.

    3. TH

      Yes.

    4. CW

      And it's across multiple companies. And yeah, okay, you need probably a lot of compute, but I saw what happened, the difference between how much compute was needed to train DeepSeek-

    5. TH

      Yeah

    6. CW

      ... and how much compute was in... It almost feels like as the AI gets bigger, you can then retrain replicant AIs way smaller.

    7. TH

      Yes.

    8. CW

      So a, a p- a processor the size of this room might be able to bootload AGI in future. So you have-- I mean, we have this with, um, desktop synthesizers, right, for bioweapons.

    9. TH

      Correct.

    10. CW

      There was a moratorium, and it means that if you try to synthesize smallpox, you get this big red flag-

    11. TH

      Yeah

    12. CW

      ... and some guys in boots kick your door down.

    13. TH

      Right.

    14. CW

      Um-

    15. TH

      I'm, I'm happy to hear you reference these examples. These, these are important. Yeah.

    16. CW

      They're ana- analogous, right?

    17. TH

      Yeah.

    18. CW

      Um, but the problem with the moratorium for AI is it can be done so siloed.

    19. TH

      Mm-hmm.

    20. CW

      Right? You can just take this code, you can take this approach, throw this model into something, start to build on top of it. It would be... I mean, Nick Bostrom's old example was imagine if you could make an atomic bomb by putting sand into a microwave. There's nothing-

    21. TH

      Correct

    22. CW

      ... physics in our current universe mean that that isn't the case.

    23. TH

      Right.

    24. CW

      But there's nothing that stops something analogous from being the case had the world have been a different way.

    25. TH

      Which is what you're really getting to here is what Carl Sagan was referring to when he talked about our technological adolescence. Like AI isn't a-- The conversation we're having isn't really about AI. It's about what happens when we have increasingly powerful, dangerous, and destructive technology. Because as, you know, humanity progresses, we are-- Whether AI existed or not, we're gonna get more and more powerful technology that would cover more and more dangerous and destructive things that you could do with it. Like we didn't have the ability to do CRISPR and bioweapons, you know, twenty years ago. Now we do have that ability, and we're gonna get more and more dangerous and destructive things. And so the real question that AI is forcing us to ask is: What is the wisdom needed to wield the technological powers that whether AI is part of it or not, that we're going to increasingly gain? And, you know, I reference this all the time, but it's just such an accurate fundamental problem statement, uh, by E. O. Wilson, that the fundamental problem of humanity is we have Paleolithic brains that don't update very well to new information and think things are sci-fi and go into denial and overwhelm and blah, blah, blah. We have medieval institutions from the eighteenth century, and we have godlike technology that makes the twenty-fourth century technology crash down on twenty-first century society. That's what AI is. It's a joke from Ajeya Cotra, who's in the film, by the way. And so if we're to solve this equation, this problem statement that E. O. Wilson lays out, the way I always think about it is we need to embrace the reality of our Paleolithic brains. That's what wisdom is. You know, we know that we get overwhelmed, and we go into denial, so we, like, work with that. We, we have systems and practices to recognize when that's about to happen and ask what do we need to hold a difficult reality together. That's embracing our Paleolithic brains. We need to upgrade our medieval institutions. We should be using twenty-first century technology to make faster updating, self-improving governance. Instead of creating recursively self-improving AI, we should be creating self-improving governance. Audrey Tang, the d- former digital minister of Taiwan, has pioneered what that would look like, that d- democracies could be using tech to find the unlikely consensus opinions of everybody. One of the things that's gonna exist in the next few months is a national dialogue on AI facilitated by technology where people can add their own ideas about how AI should be governed. And when you vote and you click, it's going to reveal the cons- the most popular consensus opinions about what should happen about these different issues. It's one thing if we live in a world where there's a cacophony and confusion about AI, and it's another thing if we live in a world where you see on a clear webpage that six hundred thousand people have voted and ninety-six percent of people across all these countries agree-

    26. CW

      You're trying to-

    27. TH

      ... that there should be international limits.

    28. CW

      You're trying to make transparent common knowledge.

    29. TH

      Exactly. Exactly. We're trying to make transparent common knowledge, and one of the ways to think about it is the movement has to see itself. There is a movement for humane technology. There is a movement for a human future.

    30. CW

      Mm.

  14. 1:31:561:33:31

    Are AI Companies Controlling the Narrative?

    1. CW

      Are you concerned about companies that are playing the same game but just trying to manipulate their optics in a slightly sexier way?

    2. TH

      Absolutely.

    3. CW

      It seems to me that the incentives for Anthropic are exactly the same as they are for everybody else. I think you mentioned the-

    4. TH

      Up until now, yes

    5. CW

      ... you, you mentioned the Antichrist earlier on. Scott Alexander, who I think is the best blogger on the planet-

    6. TH

      Yeah. I agree

    7. CW

      ... he says that Anthropic is much more likely to be the Antichrist than any other AI company because the Antichrist would present itself as being for the people. And, uh, who are we kidding that you have the same market incentives to over-train your models as quickly as possible with RL and your f- uh, the synthetic environments to keep on pushing this thing. We need as much compute. So y- you can talk about the optics up front as much as you want, but really that's just window dressing on the top of a burning house.

    8. TH

      Yep. Uh, what you're saying is super important, and this is one of the... I, I... This is not me saying everyone should just put their money into Anthropic and then the world will be fine. In fact, there's actually a view by many people that Anthropic in some ways is dangerous because the view that it is the safest, it's the Volvo. [laughs] People know that Volvos are like the safest cars. They did a lot of marketing to present that. Um, people think that Anthropic is just the safe AI company, so if they won, then suddenly everything is fine, and that would be dangerous. And that we have to have a critical view of the companies and the technology and the international limits that we need that are m- that are gonna happen from outside and above the companies, not just what an individual company can do. So

  15. 1:33:311:35:39

    How Close are We to an AI Takeover?

    1. TH

      I totally agree.

    2. CW

      How long have we got?

    3. TH

      No one knows because AI is literally moving so quickly. If I went on Twitter, uh, Elon joked in a conversation I saw yesterday, you know, it used to be that you'd go on Twitter and once every six months, this is several years ago, you would see a huge AI breakthrough, like that would just really change the game. Now there's one, you know, that you see when you go to bed, and there's a new one when you wake up in the morning and you're on Twitter and you see it, and the pace is pretty overwhelming. So I think rather than ask how much time do we have, which is kind of like just trying to assuage the fear that I think a lot of people feel, the kind of bold, brave human thing to do is to ask if things were to go well, what would that mean about how we were showing up? And then to get as many people as possible showing up from that place because that gets us the best decentralized conditions to get to the better outcome. They say in mathematics, um, that in chaos, initial conditions matter. So I'm not guaranteeing in any way, shape, or form that we're gonna get to the better future, but I would like to invite people again into stepping into and standing from the most wise and mature version of ourselves that would behave as if we were gonna get to that better future because that would give us the best chances of doing so.

    4. CW

      And the alternative is just surrender.

    5. TH

      The alternative is schisming, denying, f- depression, overwhelm. The integrity and the kind of flow of energy that I think you'll feel inside yourself if you align with what is actually needed in this moment, there's more meaningfulness. There's more preciousness of life. Like, there's so much more... Like, there's so much meaning and purpose if you're aligned with what would have things go well.

    6. CW

      Mm.

    7. TH

      And I think there's this weird thing that's happening is if you don't see a way for things to go well, and then you believe the only thing you can do is like triple down on racing as fast as possible, knowing where that's currently taking us, that has a biophysical cost on your system.

    8. CW

      Yeah. I think

  16. 1:35:391:42:30

    Why Changing AI Feels Impossible

    1. CW

      the... I was trying to think about the differences between your past work and your current work-

    2. TH

      Mm-hmm

    3. CW

      ... focusing on social media versus focusing on AI. I don't think many people find social media that net positive-

    4. TH

      Right

    5. CW

      ... in terms of their experience.

    6. TH

      Right.

    7. CW

      I don't think that many people find AI that net negative in terms of their experience.

    8. TH

      I, I would... Just so people hear me say this, I agree with you that the current experience of AI is not net negative. It doesn't feel that way. But that doesn't mean that there's a net negative outcome-

    9. CW

      Of course

    10. TH

      ... that we are racing towards.

    11. CW

      But the call to-

    12. TH

      It makes it harder

    13. CW

      ... the call to get rid of social media is easy when people reflect on the last two hours that they've spent doom scrolling at night-

    14. TH

      Yeah

    15. CW

      ... and think, "I wish I hadn't done that." The feedback curve is sufficiently quick-

    16. TH

      Right

    17. CW

      ... that people know how they felt even during the moment.

    18. TH

      That's right.

    19. CW

      "I really shouldn't watch another one of these videos. It makes me feel kind of bad. I'm all stressed and tight, and my muscles feel tense."

    20. TH

      Yep.

    21. CW

      The same thing is not true with AI.

    22. TH

      Yeah.

    23. CW

      You're asking people if you were to say, and I haven't heard you say, "Delete your AI account. Stop using it."

    24. TH

      I'm not saying that that's the thing that has to happen. Yeah.

    25. CW

      If that was the campaign, if that was what people were supposed to do, people would be sacrificing their quality of life. They would be able to fix their car less quickly. They would be able to find out things that they need to about their health to analyze their health documents, to be able to liberate them, to be able to make, uh, hard to access information cheaper and easier, and actually to enable their quality of life in a manner that didn't happen with social media. So I think the-

    26. TH

      Correct

    27. CW

      ... intermediary, uh, experience is less sexy, and that I- I'm gonna guess is also the reason why you're not saying, "Mass unsubscribe. Don't use any AI," because you're probably aware that the incentives for individual users, the value judgment that they make isn't sufficiently imbalanced as it would've been with social media. For me, it's easy to not have it on this phone. If you were to say, "I don't wanna have ChatGPT on my phone," I, I'd, uh... No, I'd need it. I- It's very useful to me.

    28. TH

      I think it's about understanding-Context and what is the careful and limited way that these tools are helpful, and then creating the conditions where that's the collective outcome that we're getting. So for example, for kids in tutoring, theoretically, this is the best tutor in the world for learning. But what if the-

    29. CW

      Alpha School is, uh, doing a South by Southwest pop-up in the middle of, uh, downtown.

    30. TH

      Yeah.

  17. 1:42:301:46:23

    Total Control or Total Collapse?

    1. CW

      If I was to give you the option of turning the rest of the world into a totalitarian state, everything, but it meant that we were able to dictate what happened to countries and to companies-

    2. TH

      I don't want-

    3. CW

      ... or to these AIs.

    4. TH

      I don't wanna create a totalitarian state.

    5. CW

      But the alternative would potentially be destruction of humanity.

    6. TH

      Well-

    7. CW

      Which would you pick?

    8. TH

      You, you are conjuring a thought experiment from Nick Bostrom, uh, which I'm sure you're aware of, and the paper you're referring to is the Vulnerable World Hypothesis in which he says that as we discover more and more technologies that in a decentralized way have the capacity to destroy things, so like your example of if it turned out that microwaving sand created a nuclear reaction that blew up the entire world, and we published that knowledge, and it spread virally on social media, how many minutes, hours, or seconds would it take before the world blew up? And then the only answer to a distributed, like, available to everybody destructive capacity would be a global totalitarian society that monitors and surveils everyone and what they're doing. So you prevented mass destruction, but you got 1984 Big Brother, and that's an uncheckable power that people cannot fight back against.I am very worried about mass surveillance. Um, I think that what happened with Anthropic is hopefully bootloading a global immune system to the risks of AI-enhanced mass surveillance. Because in a way, Big Brother could have never happened without AI. When you have an AI that can process every camera and every face and all of the emails and text messages that are flowing through everything, and you can actually ask the question as a state, "Who are the number one threats in my society that I need to suppress?" And it tells you and summarizes instantly who those people are, synthesizing all these data streams. How can you fight back against an uncheckble power when it knows all of your secrets? You can't. So we're faced with very tough choices on both ends, and I wrote a... I did this in a TED Talk, uh, I gave this last year called The Narrow Path, that we have to avoid both outcomes. We have to somehow be committed to finding the narrow path between, you know, decentralizing destructive power in everyone's hands without responsibility, without a commensurate wisdom or responsibility, in which you get chaos or catastrophes, or overcentralizing that power in uncheckable ways in which you get runaway dystopias. Uh, and we have to find something like a third attractor, to quote Daniel Schmachtenberger, or a narrow path that seeks to avoid both these outcomes. You could say that that is outside the laws of physics or not possible, but I think what's, what's important is to be committed to and living from the place where we would be finding that path, because the only chance that we would have of finding it would be requiring that everyone was living in service of it.

    9. CW

      Isn't it funny that the two worlds that you identified there, one totalitarian overreach that's controlled by humans and at the mercy of their biases, and the other being decentralized vulnerability, danger that is at the mercy of technology's liabilities, uh, both of those are extreme versions of something that we definitely don't want. But the restriction on the one that would protect us from the AI goes back to the bottom of the brainstem, that there's almost never a time that a country or a government or a small number of people or a large number of people get most of the power and then give it back.

    10. TH

      That's right.

    11. CW

      It's a ratchet system.

    12. TH

      That's right.

    13. CW

      And it only ever clicks into place and then never goes back.

    14. TH

      Which is why you need to have built-in checks and balances on power that are irreplaceable. Like, you cannot have a world where that, whatever power gets centralized has to have oversight, uh, and democratic accountability and distribution and wealth, of that wealth and power, depending on if we're talking about the wealth versus the power.

  18. 1:46:231:52:40

    How Can We Globally Coordinate AI Safety?

    1. CW

      How would we know that some country isn't signing up to a moratorium, but secretly doing all of their research behind the scenes so that they slow down-

    2. TH

      Yeah

    3. CW

      ... progress of everyone else but allow them to race ahead?

    4. TH

      So you're asking the question, if, let's say, the US and China sign an agreement, how do we know that given their maximum incentive to defect on that agreement and have the CIA or their black op- operations still continue to do the research project? Um, this is obviously the hardest technology and coordination problem that humanity has ever faced. I just need to say that at the top. Like, I'm not saying any of this is easy.

    5. CW

      It makes nuclear war look like child's play.

    6. TH

      Correct. And, but by the way, I want you to know, [scoffs] um, there's a great video, and I can send it to you, of Robert Oppenheimer asked in the 1960s, I think he's testifying for something, an interview, and they asked him, "What... How could we control this technology from proliferating?" And he takes a big puff on his cigarette, and he says, "It's too late. If you wanted to prevent the spread of nuclear technology, you would have had to do it the day after Trinity." Trinity was the first test-

    7. CW

      Mm-hmm

    8. TH

      ... of the first nuclear bomb. He believed, and this interview was in the 1960s, he believed it was inevitable, there was nothing we could do, and the world was fucked. Even he, the creator of the atomic bomb, didn't see that a lot of people would work very hard over the course of the next 30 years to invent what's called national technical means or satellites that do mutual monitoring enforcement.

    9. CW

      Mm.

    10. TH

      That we would create the seismic monitoring technology to, to be able to, to see if you were doing an underground nuclear test or an above ground test, because we would see the ver- reverberations of that. So satellites, seismic monitoring, you know, international inspectors, the International Atomic Energy Agency, we had to invent all of these new things to create a, a governance system that was capable of dealing with nuclear weapons.

    11. CW

      Isn't it interesting that the innovation around the technology that was dangerous was probably less sophisticated than the downstream technologies required to make it secure?

    12. TH

      Uh, what do you mean? The-

    13. CW

      That making the atomic bomb was one thing.

    14. TH

      Yes, yes.

    15. CW

      Then there's 50 things that need to happen in order to have-

    16. TH

      Yes, exactly

    17. CW

      ... that technology exist in the world.

    18. TH

      In a safe way. Yeah, exactly. There's a safe... There's sort of an asymmetry where the technologies... The destructive capacity is easy to create. The governance capacity is way harder to create.

    19. CW

      Bingo.

    20. TH

      The defensive capacity is harder to create.

    21. CW

      Yeah.

    22. TH

      And you could have said at the beginning of 1945, "I guess it's inevitable. 150 countries are gonna have nuclear weapons. There's nothing we could do. Let's just, like, drink margaritas and give everybody uranium and increase GDP by selling uranium to the entire world." But we didn't do that, and people had to invent stuff. We needed our best engineers and our best minds working on how to figure that out. And I'll just say, if you wanna look this up, RAND, the nonprofit defense think tank, has a paper on how international monitoring and verification mechanisms could potentially work for AI. It's not easy.

    23. CW

      Mm.

    24. TH

      It is not trivial.

    25. CW

      Okay.

    26. TH

      It would require extraordinary investments in mutual monitoring and enforcement and, um, data centers and, like, the c- you know, chips that attest where they are and all this kind of stuff.

    27. CW

      Okay. So there, there are some suggestions that-

    28. TH

      There are proposals.

    29. CW

      'Cause what I was thinking is-

    30. TH

      Yeah

  19. 1:52:401:59:18

    Why Elon Isn't in The AI Doc

    1. CW

      Can we watch the trailer? Have you got the... Can we get the trailer up?

    2. TH

      That'd be great.

    3. CW

      I wanna see the trailer for this movie.

    4. TH

      That'd be great. That'd be great. Thanks.

    5. TH

      If this technology goes wrong, it can go quite wrong.

    6. CW

      What the... [beep] [tense music]

    7. TH

      Your fear of AI is the collapse of humanity.

    8. TH

      Well, not the collapse, the abrupt extermination. There's a difference. [tense music]

    9. TH

      So I started making this movie because my wife is six months pregnant. [siren blaring] Is now a terrible time to have a kid? I mean, I'll just be honest. I know people who work on AI risk who don't expect their children to make it to high school.

    10. CW

      I, I... What the- Oh, this is the time.

    11. TH

      How does AI understand pretty much everything?

    12. CW

      It's surprisingly straightforward. Intelligence is about recognizing patterns.

    13. TH

      Patterns.

    14. CW

      Patterns.

    15. CW

      Patterns. This is the time. If you have learned those patterns, you can generate new information. Take me home.

    16. TH

      AI is moving so fast.

    17. SP

      It's being deployed prematurely. There's so much potential for things to go wrong.

    18. TH

      Why can't we just stop? All these companies are in a race to get AI.

    19. TH

      That's vastly more intelligent than people within this decade.

    20. CW

      Love me hard.

    21. TH

      China, North Korea, Russia.

    22. SP

      Whoever wins is essentially the controller of humankind.

    23. CW

      Love me hard.

    24. TH

      We need to take a threat from AI as seriously as global nuclear war. [bomb exploding] It feels like I have to find these CEOs and get them in the movie.

    25. TH

      Great.

    26. TH

      I wanna ask you to promise me that this is gonna go well.

    27. TH

      That is impossible.

    28. TH

      Okay.

    29. TH

      Am I hopeful? Yes. Am I confident that it'll go right? Absolutely not.

    30. CW

      AI is the thing that can solve climate change.

  20. 1:59:182:03:58

    Why Every Second Counts

    1. TH

      it.

    2. CW

      The problem I think that you're facing is every second is so crucial, and every single second of compute and of CEO attention and of-

    3. TH

      Yep

    4. CW

      ... staff-based attention is going to be paid on continuing to get ahead in this unbelievable one to rule them all race. I mean, uh, OpenAI is a partner on this show. I've spoken to Sam twice, and both times the calls have been so brief-

    5. TH

      Mm

    6. CW

      ... because presumably for every second that he's on the phone, that's billions of dollars of potential revenue or compute or... that is being wasted.

    7. TH

      Yeah.

    8. CW

      And that means that if you want Mark in 2013 or Sam or Dario or-

    9. TH

      But the point is you create... So like literally this is the history of all of law. So like I could kill you and steal your stuff and just take your money, but I'd prefer to live in... And everyone could do that.

    10. CW

      Mm.

    11. TH

      And w- there'd be a faster way to like get money.

    12. CW

      Mm-hmm.

    13. TH

      It's like just everybody kills everybody-

    14. CW

      Mm

    15. TH

      ... grabs their money. But that would create a society that's chaos. So instead, I sacrifice some of my abilities, like I can't kill people. [laughs] And instead, um, we have law that we, we all sort of notch down some of our individual capability so that we get to live in a society that we actually wanna live in, and that's what this would do. So in 2013, if Mark Zuckerberg had said, "I'm gonna convene, you know, uh, Musical.ly," which was before TikTok, um, Twitter and the other ones, and we're like, "Okay, look, it is so obvious we're in this arms race for attention," just like public utilities, which by the way if like i- in California it'd be PG&E. In Texas, what's your electricity provider? It's like-

    16. CW

      Oh, fuck knows. I don't know.

    17. TH

      You have no idea?

    18. CW

      I have no idea.

    19. TH

      That's okay.

    20. CW

      I have no idea what my electricity provider is.

    21. TH

      Okay. Well, technically energy companies have an incentive to maximize revenue, so theoretically they're like, "Leave the lights on, leave the stove on."

    22. CW

      Mm.

    23. TH

      Like, "Do the water 24/7," uh, because we make more money when we do that. But pu- because public utilities rely on a scarce resource called energy that has environmental, uh, um, emissions that, that are, that are... we have to deal with, there's a decoupling of revenue from, um, their incentive.

    24. CW

      Is there?

    25. TH

      So for example, in California, you're charged a base rate for like the initial energy you use, and then once you're hitting kind of the capacity that would be straining the system, we start to charge you more, but that extra revenue doesn't go just into PG&E, the energy company's pocket. It goes into a fund to have more clean energy that gives, that gives more-

    26. CW

      To offset it

    27. TH

      ... to offset it and create more energy capacity. So with attention, you could say instead of companies maximizing attention and driving revenue, you get to make money from the initial tier one of attention that you're getting.

    28. CW

      Mm.

    29. TH

      And then after that-

    30. CW

      Mm

  21. 2:03:582:07:47

    How Do We Accelerate Meaningful Change?

    1. CW

      This third attractor, precipice, narrow path thing-

    2. TH

      Yep.

    3. CW

      -that you think is important for us to walk appropriately, the pace at which things are going to happen and change, do you think that we're going to be able to move along that? I mean, you've got this movie, it's gonna be a global moment, lots of people are gonna see it, it's gonna get people talking, common knowledge. Typically, things tend to take time. Conceptual inertia usually moves over generations and centuries.

    4. TH

      Yep.

    5. CW

      Going from the heliocentric model of the universe took like a hundred years.

    6. TH

      Well said. Yep.

    7. CW

      How, how-

    8. TH

      This is the problem in general that we face with technology, and this is the E. O. Wilson quote.

    9. CW

      Yep.

    10. TH

      It's not just that we have Paleolithic brains, medieval institutions, it's that the brains operate on a very slow updating clock rate, the institutions operate on a slow clock rate, and the technology moves at whatever the twenty-first or now twenty-fourth century technology clock rate the AI moves at.

    11. CW

      So how do you think about expediting this change?

    12. TH

      You have to have your governance move at the pace of the thing that you're trying to govern. So-

    13. CW

      Is that ever gonna be possible? Governance moving at that pace, the lumbering behemoth? What is it? Leviathan leans left and lumbers along.

    14. TH

      There needs to be, rather than recursively self-improving AI that is unclickable by any human oversight process, in which we will for sure lose control and build something crazy that we regret, we instead should have self-improving governance. You can use AI to look at all the laws on the books and say, "What are the laws that don't matter anymore, that are creating all this red tape that we don't need that's, like, stifling innovation?" And get the AI to find those laws and, and reinterpret them and say, "What are the ones we need to get rid of, and how do you rewrite them for twenty-first century technology in the current age?" We could be using AI and technology to update the governance as fast as the technology, but we'd also probably need to slow down the technology, which we would benefit from because then we wouldn't crash.

    15. CW

      Hmm.

    16. TH

      Again, it's like China wins when they whisper in our ear, "Go faster, go faster. You're gonna blow yourself up."

    17. CW

      Pyrrhic victory ahead.

    18. TH

      Pyrrhic victory. Uh, you've, you said this earlier. When the US races ahead and has a more sophisticated model, China gets it like ten days later, because first of all, they have spies in all of our companies, and second of all, they distill all the models.

    19. CW

      Is that what they're doing?

    20. TH

      They have spies in the companies. Uh, you know, they are maximally incentivized to know what's going on. Plus, as I-- as you were saying, they can distill the US models, meaning they can, like, query those models a thousand times and kind of distill the essence of those models, and then make and train theirs.

    21. CW

      Mm-hmm.

    22. TH

      So there's a great, like, meme online of, um, like it's a motorboat with a guy who's, um... What is it called when you're, um-

    23. CW

      Wake surfing?

    24. TH

      Wake... Yeah, like you're wake surfing behind him on the, on the string. I mean, on the... What is that called? The-

    25. CW

      String.

    26. TH

      Yeah, whatever. [chuckles] Um, and, and the, the guy who's on the, um, on the surfing says, "Go faster," to the boat that's, like, racing ahead. That's the US that's building the advanced model. But then they're literally on a string right behind him that's getting all the benefits from it.

    27. CW

      Yeah.

    28. TH

      And there is a study that Anthropic found, um, that China had been covertly even using US Anthropic AI models to perform a cyber hacking operation. So, like, again, if we're winning the race to the technology, but losing the race to governing or controlling or protecting the technology, what the fuck are we winning?

    29. CW

      Yeah. Yeah.

    30. TH

      [chuckles] It's like j- it's so basic what we're talking about. It's like this is not rocket science. It's unbelievably simple. If you have the power of gods, you need the wisdom, love, and prudence of gods. I hope we've laid out a bunch of examples of how we can do that. And while it's not easy, it's not ha- it does not happen by default. If you get everyone you know to go out and see the AI doc, understand that AI is dangerous, get your church group, get your business, and recognize that we need to, you know, not build bunkers, but write laws to actually steer AI before it's too late.

Episode duration: 2:07:48

Install uListen for AI-powered chat & search across the full episode — Get Full Transcript

Transcript of episode NufB1LL_rCU

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome