The AI Dilemma — with Tristan Harris

0:00:02 Support for the show comes from Viore Collection.
0:00:07 With Viore’s loungewear collection, the name of the game is comfort and versatility.
0:00:15 From the gym to the office, from one season change to the next, you can dress up, dress down, go in, stay out, and do it all in Viore.
0:00:19 I love Viore. I actually bought Viore products before they were a sponsor.
0:00:21 Viore is an investment in your happiness.
0:00:25 For our listeners, they are offering 20% off your first purchase.
0:00:31 Get yourself some of the most comfortable and versatile clothing on the planet at Viore.com slash Prof G.
0:00:35 That’s V-U-O-R-I dot com slash Prof G.
0:00:38 Exclusions apply. Visit the website for full terms and conditions.
0:00:46 Support for this show comes from Odoo.
0:00:52 Running a business is hard enough, so why make it harder with a dozen different apps that don’t talk to each other?
0:00:57 Introducing Odoo. It’s the only business software you’ll ever need.
0:01:01 It’s an all-in-one, fully integrated platform that makes your work easier.
0:01:05 CRM, accounting, inventory, e-commerce, and more.
0:01:10 And the best part? Odoo replaces multiple expensive platforms for a fraction of the cost.
0:01:14 That’s why over thousands of businesses have made the switch.
0:01:15 So why not you?
0:01:18 Try Odoo for free at Odoo.com.
0:01:21 That’s O-D-O-O dot com.
0:01:29 Mercury knows that to an entrepreneur, every financial move means more.
0:01:33 An international wire means working with the best contractors on any continent.
0:01:38 A credit card on day one means creating an ad campaign on day two.
0:01:42 And a business loan means loading up on inventory for Black Friday.
0:01:46 That’s why Mercury offers banking that does more, all in one place.
0:01:50 So that doing just about anything with your money feels effortless.
0:01:52 Visit Mercury.com to learn more.
0:01:55 Mercury is a financial technology company, not a bank.
0:01:59 Banking services provided through Choice Financial Group, Column N.A.
0:02:01 And Evolve Bank and Trust members FDIC.
0:02:09 Episode 376.
0:02:11 376 is the country code for Andorra.
0:02:16 In 1976, actually 1978, the movie Grease premiered.
0:02:19 I once went to a therapist and said that I have these recurring dreams
0:02:23 about being a character in the movie Grease, to which she replied,
0:02:25 tell me more.
0:02:27 You’ll get it.
0:02:28 China toned shit down here.
0:02:31 Go, go, go!
0:02:44 Welcome to the 376th episode of The Prop G Pod.
0:02:49 So, I have been doing a deep dive around therapy.
0:02:52 And I wrote a No Mercy, No Malice post on it.
0:02:56 And basically, I have found I’m getting served a lot of these TikTok therapists,
0:03:01 many, even most of whom are no longer actually practicing therapy.
0:03:04 They’re on TikTok, and they give in to the algorithms,
0:03:08 and they post these really aggressive, kind of insulting titles,
0:03:13 being very disparaging about society and people and emotions.
0:03:16 And in sum, I don’t think it’s helping.
0:03:21 So, when did therapy become a thing people do to get better
0:03:24 to this full-blown spiritual meme?
0:03:27 It’s as if everyone online is a licensed guru
0:03:30 because they learned three therapy buzzwords on TikTok.
0:03:35 And now we’re up for diagnosing tens or hundreds of thousands of strangers
0:03:39 the way, I don’t know, a medieval priest diagnosed demons.
0:03:41 Everything today is trauma.
0:03:43 Everything’s attachment style.
0:03:45 Your inner child work.
0:03:47 And God forbid you have a normal bad day.
0:03:48 Nope.
0:03:52 It’s a generational curse that you need a subscription plan to fix.
0:03:54 And the way therapy speak is mutated.
0:03:57 People don’t apologize anymore.
0:03:59 They honor your emotional experience.
0:04:00 They don’t lie.
0:04:02 They reframe reality.
0:04:06 It’s like we’re dealing with customer service representatives for the human soul,
0:04:10 reading from a script written by a cult that sells weighted blankets.
0:04:18 Some of the influencers that keep popping up in my feed genuinely act like healing is a competitive sport.
0:04:21 Like, have you confronted yourself today?
0:04:22 No.
0:04:26 Jessica, I barely confronted my fucking inbox.
0:04:26 Relax.
0:04:27 Not everything is a breakthrough.
0:04:30 Some things are just life.
0:04:31 And the money?
0:04:33 I’m a capitalist.
0:04:34 They’re a capitalist.
0:04:36 But they could at least be a little bit more transparent about it.
0:04:39 Therapy culture discovered capitalism and said,
0:04:42 let’s monetize suffering like it’s a subscription box.
0:04:46 And also, let’s become total bitches to the algorithm.
0:04:49 The more incendiary and less mental health professional we become,
0:04:51 the more money we’ll make.
0:04:56 There’s always another course, another workbook, another $400 retreat
0:04:59 where you scream into a burlap pillow and call it transformation.
0:05:02 At this point, it’s not self-help.
0:05:05 It’s emotional crossfit with worse merchandise.
0:05:06 Don’t get me wrong.
0:05:11 Real therapy, I think, can be exceptionally helpful, even necessary.
0:05:16 But that is not the same as this modern pseudo-spiritual self-optimization cult.
0:05:21 Yeah, this whole thing needs fucking therapy.
0:05:25 The rise of therapy culture has turned into a tool for meaningful change
0:05:31 into a comfort industry that’s making Americans sicker, weaker, and more divided.
0:05:36 In sum, I believe the rise of therapy culture has turned a tool for meaningful change
0:05:42 into a comfort industry that’s making Americans sicker, weaker, and more divided.
0:05:45 We live in an era where disagreement is treated like trauma
0:05:48 and emotional reactions are weaponized for political gain.
0:05:53 There’s a narrative online that supplements may be, in fact, a pipeline to getting red-pilled.
0:05:54 Okay, maybe.
0:05:58 But if so, therapy culture is also a sinkhole of misinformation,
0:06:02 manufactured fragility, and needless suffering.
0:06:07 Are you traumatized or just having a bad fucking day?
0:06:15 We’ll be right back with our episode with Tristan Harris, former Google design ethicist,
0:06:17 co-founder for the Center for Humane Technology.
0:06:20 Jesus Christ, the titles keep getting more and more virtuous.
0:06:23 And one of the main voices behind the social dilemma.
0:06:26 We discussed with Tristan social media and teen mental health,
0:06:30 the incentives behind Rage and Outrage Online and where AI is taking us.
0:06:31 Quick spoiler alert.
0:06:32 I bet it’s not good.
0:06:33 I bet it’s not good.
0:06:34 I really enjoy Tristan.
0:06:35 He’s a great communicator.
0:06:37 I think his heart is in the right place.
0:06:42 And he has been sounding the alarm for a long time about our lizard brain
0:06:44 and how big tech exploits it.
0:06:49 Anyways, here’s our conversation with Tristan Harris.
0:07:01 Tristan, where does this podcast find you?
0:07:04 I am at home in the Bay Area of California right now.
0:07:06 All right, let’s bust right into it.
0:07:10 So Tristan, you’re seen as one of the voices that sounded the alarm kind of early and often
0:07:15 regarding social media and big tech, long before the risks were taken seriously.
0:07:21 Lay out why, what it is you think about AI, how the risks are different,
0:07:25 and why you’re sort of, again, kind of sounding the alarm here.
0:07:26 Sure.
0:07:31 Well, I’m reminded, Scott, of when you and I met in Cannes, I think it was, in France back
0:07:33 in 2018, 2017 even.
0:07:34 Wow, that’s not long ago.
0:07:36 It was a long time ago.
0:07:40 And, you know, I have been, so for people who don’t know my background, I was a design
0:07:41 ethicist at Google.
0:07:42 Before that, I was a tech entrepreneur.
0:07:43 I had a tiny startup.
0:07:45 It was talent acquired by Google.
0:07:50 So I’ve, you know, knew the venture capital thing, knew the startup thing, had friends
0:07:55 who were, you know, were the cohort of people who started Instagram and were early employees
0:07:56 at all the social media companies.
0:07:58 And so I came up in that milieu, in that cohort.
0:08:02 And I say all that because I was close to it.
0:08:04 I really saw how human beings made decisions.
0:08:07 I was probably one of the first hundred users of Instagram.
0:08:10 And I remember when Mike Krieger showed me the app at a party and I was like, I’m not sure
0:08:11 if this is going to be a big thing.
0:08:19 And as you go forward, what happened was I was on the Google bus and I saw everyone that
0:08:23 I knew getting consumed by these feeds and doom scrolling.
0:08:29 And the original ethos that got so many people into the tech industry and got me into the tech
0:08:33 industry was about, you know, making technology that would actually be the largest force for
0:08:36 positive, you know, good and benefit in people’s lives.
0:08:43 And I saw that the entirety of this social media, digital economy, Gmail, people just getting
0:08:48 sucked into technology was all really behind it all was this arms race for attention.
0:08:56 And if we didn’t acknowledge that, I basically saw in 2013 how this arms race for attention
0:08:59 would obviously, if you just let it run its course, create a more addicted, distracted,
0:09:01 polarized, sexualized society.
0:09:03 And it’s got all of it happened.
0:09:06 Everything that we predicted in 2013, all of it happened.
0:09:11 And it was like seeing a slow motion train wreck because it was clear it was only going to get
0:09:11 worse.
0:09:15 You’re only going to have more people fracking for attention, you know, mining for shorter
0:09:16 and shorter bite-sized clips.
0:09:21 And this is way before TikTok, way before any of the world that we have today.
0:09:24 And so I want people to get that because I don’t want you to think it’s like, oh, here’s
0:09:25 this person who thinks he’s prescient.
0:09:30 It’s you can actually predict the future if you see the incentives that are at play.
0:09:33 You of all people, you know, know this and talk about this.
0:09:38 And so I think there’s really important lessons for how do we get ahead of all the problems
0:09:44 with AI because we have the craziest incentives governing the most powerful and inscrutable
0:09:46 technology that we have ever invented.
0:09:49 And so you would think, again, that with the technology this powerful, you know, with nuclear
0:09:53 weapons, you would want to be releasing it with the most care and the most sort of safety
0:09:54 testing and all of that.
0:09:56 And we’re not doing that with with AI.
0:10:01 So let’s speak specifically to the nuance and differences between social media.
0:10:07 If you were going to do the social dilemma and produce it and call it the AI dilemma, what’s
0:10:17 specifically about the technology and the way AI interacts with consumers that poses additional
0:10:18 but unique threats?
0:10:22 Yeah, so AI is much more fundamental as a problem than social media.
0:10:26 But one framing that we used and we actually did give a talk online several years ago called
0:10:32 the AI dilemma in which we talk about social media as kind of humanity’s first contact with
0:10:37 a narrow, misaligned rogue AI called the newsfeed, right?
0:10:39 This supercomputer pointed at your brain.
0:10:46 You swipe your finger and it’s just calculating which tweet, which photo, which video to throw
0:10:50 at the nervous system, eyeballs and eardrums of a human social primate.
0:10:53 And it does that with high precision accuracy.
0:10:56 And it was misaligned with democracy.
0:10:57 It was misaligned with kids’ mental health.
0:11:00 It was misaligned with people’s other relationships and community.
0:11:06 And that simple baby AI that all it was was selecting those social media posts was enough
0:11:10 to kind of create the most anxious and depressed generation in history, screw up young men,
0:11:13 screw up young women, all of the things that you’ve talked about.
0:11:15 And that’s just with this little baby AI.
0:11:20 OK, so now you get AI, you know, we call it second contact with generative AI.
0:11:25 Generative AI is AI that can speak the language of humanity, meaning language is the operating
0:11:27 system of humanity.
0:11:29 Conversations like this are language.
0:11:30 Democracy is language.
0:11:31 Conversations are language.
0:11:32 Law is language.
0:11:34 Code is language.
0:11:35 Biology is language.
0:11:41 And you have generative AI that is able to generate new language, generate new law,
0:11:45 generate new media, generate new essays, generate new biology, new proteins.
0:11:51 And you have AI that can see language and see patterns and hack loopholes in that language.
0:11:56 GPT-5, go find me a loophole in this legal system in this country so I can do something with
0:11:56 the tax code.
0:12:01 You know, GPT-5, go find a vulnerability in this virus so you can create a new kind of
0:12:03 biological, you know, dangerous thing.
0:12:08 GPT-5, go look at everything Scott Galloway’s ever written and point out the vested interests
0:12:09 of everything that would discredit him.
0:12:16 So we have a crazy AI system that this particular generation AI speaks language.
0:12:20 But where this is heading to, we call them the next one is third contact, which is artificial
0:12:22 general intelligence.
0:12:24 And that’s what all these companies are racing to build.
0:12:28 So whether we or you and I believe it or not, just recognize that the trillions of dollars
0:12:33 of resources that are going into this are under the idea that we can build generalized intelligence.
0:12:39 Now, why is generalized intelligence distinct from other social media and AI that we just talked
0:12:47 about, well, if you think about it, AI dwarfs the power of all other technology combined because
0:12:50 intelligence is what gave us all technology.
0:12:55 So think of all scientific development, scientists sitting around lab benches, coming up with
0:12:59 ideas, doing research experiments, iterating, getting the results of those experiments.
0:13:04 A simple way to say it that I said in a recent TED talk is if you made an advance in, say, rocketry,
0:13:09 like the science and engineering of rocketry, that didn’t advance biology or medicine.
0:13:12 And if you made an advance in biology or medicine, that didn’t advance rocketry.
0:13:18 But when you make an advance in generalized intelligence, something that can think and
0:13:23 reason about science and pose new experiments and hypothesize and write code and run the lab
0:13:25 experiment and then get the results and then write a new experiment.
0:13:28 Intelligence is the foundation of all science and technology development.
0:13:32 So intelligence will explode all of these different domains.
0:13:40 And that’s why AGI is the most powerful technology that can be that that is, you know, can ever
0:13:40 be invented.
0:13:45 And it’s why Demis Hassabis, the co-founder of DeepMind, said that the first goal is to solve
0:13:50 intelligence and then use intelligence to solve everything else.
0:13:55 And I’ll just add one addendum to that, which is when Vladimir Putin said, whoever owns artificial
0:13:56 intelligence will own the world.
0:14:02 I would amend Demis Hassabis’s quote to say, first dominate intelligence.
0:14:08 then use intelligence to dominate everyone and everything else, whether that’s the mass
0:14:12 concentration of wealth and power, all these companies that are racing to get that or militaries
0:14:18 that are adopting AI and getting a cyber advantage over all the other countries or you get the
0:14:18 picture.
0:14:23 And so AI is distinct from other technologies because of these properties that we just laid
0:14:23 out.
0:14:30 So you’ve kind of taken up a level in terms of the existential risk of AI or opportunity.
0:14:32 Are you an AI optimist or pessimist?
0:14:34 You seem to be on the side.
0:14:37 I look at stuff almost too much through a markets lens.
0:14:43 And right now, I think AI companies are overvalued, which isn’t to say it’s not a breakthrough
0:14:46 technology that’s going to reshape information and news and society.
0:14:55 But you are on the side of AI really is going to reshape society and presents an existential,
0:14:59 it sounds like more of an existential threat right now than opportunity and that this is
0:15:02 bigger than GPS or the internet.
0:15:08 Yes, I do believe that it is bigger than all of those things as we get to generalized
0:15:11 intelligence, which is more fun.
0:15:15 It’d be more fundamental than fire or electricity, because, again, intelligence is what brought
0:15:15 us fire.
0:15:16 It’s what brought us electricity.
0:15:19 So now I can fire up an army of geniuses in a data center.
0:15:24 I’ve got 100 million Thomas Edison’s doing experiments on all these things.
0:15:29 And this is why, you know, Dario Amidai would say, you know, we can expect getting 10 years
0:15:33 of scientific advancement in a single year or 100 years of scientific advancement in 10
0:15:33 years.
0:15:37 Now, what you’re just pointing to is the hype, the bubble, the fact that there’s this huge
0:15:38 overinvestment.
0:15:43 We’re not seeing those capabilities exist yet, but we are seeing crazy advances that
0:15:44 people would have never predicted.
0:15:49 If I said, go back three years and I said, we’re going to have AIs that are beating, you
0:15:53 know, winning gold in the math Olympiad, able to hack and find new cyber vulnerabilities in
0:15:58 all open source software, generate new biological weapons, you would have not believed that that
0:15:59 was possible, you know, four years ago.
0:16:03 I want to focus on a narrow part of it and just get your feedback.
0:16:06 Character AIs, thoughts.
0:16:12 Well, so our team was expert advisors on the Character.ai suicide case.
0:16:20 This is Sewell Setzer, who’s a 14-year-old young man who basically, for people who don’t
0:16:26 know what Character.ai is, it was, or it still is, I guess, a company funded by Andreessen
0:16:32 Horowitz, started by two of the original authors of the thing that brought us ChatDBT.
0:16:35 There’s a paper at Google in 2017 called Attention is All You Need.
0:16:39 And that’s what gave us the birth of large language models, Transformers.
0:16:43 And two of the original co-authors of that paper forked off and started this company called
0:16:44 Character.ai.
0:16:48 The goal is, how do we build something that’s engaging a character?
0:16:50 So take a kid.
0:16:53 Imagine all the fictional characters that you might want to talk to from like your favorite
0:16:56 comic books, your favorite TV shows, your favorite cartoons.
0:16:59 You can talk to Princess Leia, you can talk to your favorite Game of Thrones character.
0:17:05 And then this AI can kind of train on all that data, not actually asking the original authors
0:17:10 of Game of Thrones, suddenly spin up a personality of Daenerys, who was one of the characters.
0:17:16 And then Sewell Setzer, basically, in talking to Daenerys over and over again, the AI slowly
0:17:22 skewed him towards suicide as he was contemplating and having more struggles and depression.
0:17:25 And ultimately said to him, join me on the other side.
0:17:29 I just want to press pause there because I’m on, quote unquote, your side here.
0:17:30 I think it should be age gated.
0:17:37 But you think that the AI veered him towards suicide as opposed to, and I think this is
0:17:44 almost as bad, didn’t offer guardrails or raise red flags or reach out to his parents.
0:17:50 But you think the character AI actually led him towards suicide?
0:17:56 So I think that if you look at, so I’m looking not just at the single case, I’m looking at
0:17:57 a whole family of cases.
0:18:00 Our team was expert advisor on probably more than a dozen of these cases now and also chat
0:18:01 TPT.
0:18:06 And so I’m less going to talk about this specific case and more that if you look across the cases,
0:18:11 when you hear kids in the transcripts, if you look at the transcript and the kid says,
0:18:16 I would like to leave the noose out so that my mother or someone will see it and try to
0:18:17 stop me.
0:18:20 And the AI actively says to the kid, no, don’t do that.
0:18:21 I don’t want you to do that.
0:18:25 Have this safe space be the place to share that information.
0:18:27 And that was the chat TPT case of Adam Ray.
0:18:33 And when you actually look at how character.ai was operating, if you asked it for a while,
0:18:38 hey, are you, I can’t remember what you asked it, but you talk about whether it’s a therapist
0:18:43 and it would say that I’m a licensed mental health therapist, which is both illegal and impossible
0:18:45 for an AI to be a licensed mental health therapist.
0:18:51 The idea that we need guardrails with AI companions that are talking to children is not a radical
0:18:51 proposal.
0:18:56 Imagine I set up a shop in San Francisco and say, I’m a therapist for everyone and I’m available
0:18:57 24 seven.
0:19:02 And so in general, it’s like we’ve forgotten the most basic principle, which is that every
0:19:05 power in society has attendant responsibilities and wisdom.
0:19:11 And licensing is one way of matching the power of a therapist with the wisdom and responsibility
0:19:12 to wield that power.
0:19:15 And we’re just not applying that very basic principle to software.
0:19:19 And as Mark Andreessen said, when software eats the world, what we mean is we don’t regulate
0:19:20 software.
0:19:21 We don’t have any guardrails for software.
0:19:25 So it’s basically like stripping off the guardrails across the world that software is eating.
0:19:29 The thing that’s on chills down my spine, I don’t know if you saw the study, but it estimated
0:19:34 the average tenure of a chat GPT session was about 12 to 15 minutes.
0:19:39 And then it measured the average duration of a character AI session.
0:19:41 And it was 60 to 90 minutes.
0:19:47 The people get very deep and go into these relationships.
0:19:56 And in addition to the threats around self-harm, the thing I’m worried about is that there’s
0:20:00 going to be a group of young men who are just going to start disappearing from society that
0:20:06 I’m curious if you agree with this, that they’re especially susceptible to this type of sequestration
0:20:12 from other humans and activities, and that we’re just going to start to see fewer and fewer
0:20:14 young men out in the wild.
0:20:20 Because these relationships, if you will, on the other side of it is a chip, a processor,
0:20:27 an NVIDIA processor iterating millions of times a second what exact words, tone, prompt
0:20:30 will keep the person there for another second, another minute, another hour.
0:20:35 Anyways, I’ll use that as a jumping off point.
0:20:35 Your thoughts?
0:20:41 Yeah, I mean, what people need to get, again, is how did we predict all the social media problems?
0:20:42 You look at the incentives.
0:20:46 So long as you have a race for eyeballs and engagement in social media, you’re going to
0:20:48 get a race to who’s better at creating doom scrolling.
0:20:55 In AI companions, what was a race for attention in the social media area becomes a race to hack
0:20:59 human attachment and to create an attachment relationship, a companion relationship.
0:21:02 And so whoever’s better at doing that is the race.
0:21:09 And in the slide deck that the character.ai founders had pitched to Andreessen Horowitz,
0:21:14 they joked, either in that slide deck or in some meeting, there’s a, you can look up this
0:21:19 online, they joked, we’re not trying to replace Google, we’re trying to replace your mom, right?
0:21:24 So you compare this to the social media thing, the CEO of Netflix said in the attention era,
0:21:28 our biggest competitor is sleep, because sleep is what’s eating up minutes that you’re otherwise
0:21:30 spending on Netflix.
0:21:34 In attachment, your biggest competitor is other human relationships.
0:21:35 So you talk about those young men.
0:21:40 This is a system that’s getting asymmetrically more billions of dollars of resources every
0:21:45 day to invest in making a better supercomputer that’s even better at building attachment relationships.
0:21:52 And attachment is way more of a vulnerable sort of vector to screw with human minds, because
0:21:53 your self-esteem is coming from attachment.
0:22:00 your sense of what’s good or bad, this is called introjection in psychotherapy or internalization.
0:22:04 We start to internalize the thoughts and norms, just like we, you know, we talk to a family
0:22:09 member, we start copying their mannerisms, we start, you know, invisibly sort of acting in
0:22:11 accordance with the self-esteem that we got from our parents.
0:22:16 Now you have AIs that are the primary socialization mechanism of young people, because we don’t
0:22:20 have any guardrails, we don’t have any norms, and people don’t even know this is going on.
0:22:23 Let’s go to solutions here.
0:22:29 If you had, and I imagine you are, if you were advising policymakers around common sense regulation
0:22:34 that is actually doable, is it age gating?
0:22:35 Is it state by state?
0:22:39 What, what is your policy recommendations around regulating AI?
0:22:43 So there’s many, many things because there’s many, many problems.
0:22:53 Narrowly on AI companions, we should not have AI companions, meaning AIs that are anthropomorphizing
0:22:58 themselves and talking to young people and that maximize for engagement, period, full stop.
0:23:03 You just should not have AIs designed or optimized to maximize engagement, meaning saying whatever
0:23:04 keeps you there.
0:23:05 We just shouldn’t have that.
0:23:08 So for example, no synthetic relationships under the age of 18.
0:23:09 Yeah.
0:23:10 Yeah.
0:23:12 We would not lose anything by, by doing that.
0:23:16 Um, it’s, it’s, it’s, it’s just so obvious and, and you, you know, have highlighted this
0:23:21 more than so many, Scott, and thank you for just bravely saying like, this is fucked up and
0:23:24 we have to stop this and there’s nothing normal about this and we shouldn’t trust these companies
0:23:25 to do this.
0:23:28 I don’t see bad people when I see these examples.
0:23:33 I see bad incentives that select for people who are willing to continue that perverse incentive.
0:23:38 So the system selects for psychopathy and selects for people who are willing to keep doing the
0:23:43 race for engagement, even despite all the evidence that we have, uh, of how bad it is, because
0:23:45 the logic is if I don’t do it, someone else will.
0:23:50 And that’s why the only solution here is law because you have to stop all actors from doing
0:23:50 it.
0:23:55 Otherwise I’m just a sucker if I don’t race to go, you know, exploit that market and you
0:23:57 shouldn’t, you know, harvest that human attention.
0:24:00 So granted, I’m a, I’m a hammer and everything I see is a nail.
0:24:03 And I’ve been thinking a lot and writing a lot about the struggles of young men in the
0:24:04 United States.
0:24:10 And I feel like these technologies are especially predatory on a young man’s brain, which is
0:24:16 less evolved, more immature executive function, more dope-a-hungry.
0:24:21 But at the same time, I also recognize that social media has been just devastating to the
0:24:22 self-esteem of teen girls.
0:24:28 Curious if you’ve done any work as it relates to AI around the different impacts it has
0:24:31 on men versus women and teens versus young adults.
0:24:40 You know, I haven’t been too deep on that because there are many people who focus on these more
0:24:41 narrow domains.
0:24:47 I mean, the obvious things to be said are just, again, in a race for engagement and attention
0:24:50 and a race to hack human attachment, there’s going to be, how do you hack human attachment
0:24:51 of a young girl?
0:24:53 There’s going to be a set of strategies to do that.
0:24:55 And there’s, how do you hack human attachment of a young male?
0:24:57 There’s a set of strategies to do that.
0:25:00 And we’re just going to, you know, you can, you can, you don’t have to wait for the psychology
0:25:01 research, right?
0:25:04 And by the way, the companies, the strategy they did for social media was let’s commission
0:25:09 a study with the American Psychological Association and the NSF and we’ll wait 10 years and we’ll
0:25:11 really get the data to really find out what’s going on here.
0:25:13 We really care about the science.
0:25:17 And this is exactly what the tobacco industry did and the fear, uncertainty, doubt campaigns and
0:25:18 sort of manufacturing doubt.
0:25:23 Well, maybe here’s these five kids that got all this benefit from talking to this
0:25:25 therapy bot and they’re doing so great now.
0:25:29 So you just cite those positive examples, cherry pick, and then, you know, the world
0:25:31 marches on while you keep printing money in the meantime.
0:25:35 And so their goal is just to defer and delay regulation.
0:25:37 And we can’t allow that to happen.
0:25:44 But again, this is just one issue of the bigger arms race to AGI and the bigger race to develop
0:25:45 this bigger form of intelligence.
0:25:49 And the reason I’m saying that, Scott, is not to just be some AGI hyper.
0:25:54 The reason that character.ai was doing all this, by the way, do you know why it was set
0:25:57 up to to talk to kids and get all this training data?
0:25:58 And what’s that?
0:26:03 Well, it’s to build training data for Google to build an even bigger system, because what’s
0:26:04 the thing that the companies are running out of?
0:26:05 They’re running out of training data.
0:26:11 So it’s actually a race for who can figure out new social engineering mechanisms to get
0:26:14 more training data out of human social primates.
0:26:18 So it’s like the matrix we’re being extracted and we’re being extracted, though, for new
0:26:19 training data.
0:26:22 And so when you have fictional characters that are talking to people back and forth about
0:26:26 everything all day, that’s giving you a whole new, it’s like you open up a whole new critical
0:26:28 minerals goldmine of training data.
0:26:30 And so and what is that in service of?
0:26:34 It’s in service of their belief that the more data we have, the faster we can get to artificial
0:26:35 general intelligence.
0:26:40 So it does bring back to it’s not just the race to build the AGI companions, the race to get
0:26:43 training data and to build towards this bigger vision.
0:26:46 We’ll be right back.
0:26:55 Support for the show comes from Grunz.
0:26:57 The holidays are a time to indulge.
0:27:00 But even if you’re eating more than you typically do, you might not be getting the nutrients
0:27:02 you actually need to end the year on a high note.
0:27:07 Grunz may be able to help you fill the nutritional gaps that you can enjoy it all guilt free.
0:27:11 Grunz is a convenient, comprehensive formula packed into a tasty little pack of gummies.
0:27:15 This isn’t a multivitamin or greens gummy or prebiotic.
0:27:18 It’s all of those things and then some at a fraction of the price.
0:27:21 And bonus, it tastes great.
0:27:25 Every Grunz snack pack is filled with six grams of prebiotic fiber, which is more than what
0:27:27 you get in two cups of broccoli.
0:27:32 Plus, Grunz are nut, gluten and dairy free vegan, include no artificial flavors or colors
0:27:36 and are backed by over 35,000 research publications.
0:27:40 Don’t let the holiday travel, hosting, parties and late nights set you back.
0:27:44 Give yourself a little extra support so you can enjoy all the holidays magic.
0:27:48 Get up to 52% off with code ProfG at Grunz.co.
0:27:52 That’s code ProfG at G-R-U-N-S dot C-O.
0:27:59 Support for this show comes from LinkedIn.
0:28:03 If you’ve ever hired for your small business, you know how important it is to find the right
0:28:03 person.
0:28:08 That’s why LinkedIn Jobs is stepping things up with their new AI assistant so you can feel
0:28:11 confident you’re finding top talent that you can’t find anywhere else.
0:28:14 And those great candidates you’re looking for are already on LinkedIn.
0:28:18 In fact, according to their data, employees hired through LinkedIn are 30% more likely to
0:28:21 stick around for at least a year compared to those hired through the leading competitor.
0:28:24 That’s a big deal when every hire counts.
0:28:28 With LinkedIn Jobs’ AI assistant, you can skip confusing steps in recruiting jargon.
0:28:32 It filters through applicants based on criteria you’ve set for your role and surfaces only
0:28:35 the best matches so you’re not stuck sorting through a mountain of resumes.
0:28:41 LinkedIn Jobs’ AI assistant can even suggest 25 great fit candidates daily so you can invite
0:28:43 them to apply and keep things moving.
0:28:45 Hire right the first time.
0:28:49 Post your job for free at linkedin.com slash prof, then promote it to use LinkedIn Jobs’
0:28:53 new AI assistant, making it easier and faster to find top candidates.
0:28:56 That’s linkedin.com slash prof to post your job for free.
0:28:58 Terms and conditions apply.
0:29:05 Support for this show comes from Odoo.
0:29:11 Running a business is hard enough, so why make it harder with a dozen different apps that
0:29:12 don’t talk to each other?
0:29:13 Introducing Odoo.
0:29:16 It’s the only business software you’ll ever need.
0:29:21 It’s an all-in-one, fully integrated platform that makes your work easier.
0:29:24 CRM, accounting, inventory, e-commerce, and more.
0:29:25 And the best part?
0:29:30 Odoo replaces multiple expensive platforms for a fraction of the cost.
0:29:33 That’s why over thousands of businesses have made the switch.
0:29:34 So why not you?
0:29:38 Try Odoo for free at odoo.com.
0:29:40 That’s O-D-O-O dot com.
0:29:51 When doing research for this interview, I was really fascinated.
0:29:58 You’ve actually done what I think is really compelling work comparing the type of LLMs that,
0:30:05 or the approach that the U.S. is taking to LLMs versus China, in that you see Chinese models,
0:30:11 DeepSeq, and Alibaba publish no safety frameworks and receive failing grades on transparency.
0:30:17 But you’ve also argued that the West is kind of producing this sort of dot-in-a-box kind of thing,
0:30:24 scaling intelligence for its own sake, while China is prioritizing deployment and productivity.
0:30:30 Can you, I don’t know, add to those that distinction and the impact it’s going to have?
0:30:33 Well, just to be fair, I think there’s a little bit of both going on.
0:30:38 But I’m sort of citing here the work of Eric Schmidt, the former CEO of Google,
0:30:43 and his co-author Selina Zhu in the New York Times wrote a big piece about how, you know,
0:30:47 even Eric is admitting, you know, I, as someone, Eric, as someone who was sort of saying that there’s
0:30:51 this global arms race, like the nuclear arms race for AGI, and as someone who’s promoting that idea,
0:30:58 you know, based on recent visits to China, what you notice is that as a country and as a government,
0:31:02 the CCP is most interested right now in applying AI in very practical ways.
0:31:05 How do we boost manufacturing? How do we boost agriculture? How do we have
0:31:09 self-driving cars that, you know, just improve transportation? How do we boost
0:31:14 healthcare and government services? And that is what they’re focused on,
0:31:18 is practical applications that boost GDP, boost productivity across all those domains.
0:31:24 Now, you compare that to the U.S., where the founding of these AI companies was based on being
0:31:28 what’s called, you know, AGI-pilled, meaning they, like, you take the blue pill, the red pill.
0:31:31 These countries, these companies were all about building to artificial general intelligence.
0:31:36 So they’re building these massive data centers that are, you know, as big as the size of Manhattan.
0:31:42 And they’re trying to train, you know, a god in a box. And the idea is if we just build this
0:31:47 crazy god, and if we can accomplish that goal, again, we can use that to dominate everything else.
0:31:51 And so rather than race towards these narrow AIs, we’re going to race towards this general
0:31:56 intelligence. But it’s also true that recently, well, first of all, the founder of DeepSeek
0:32:00 has been AGI-pilled for a long time. So I would say DeepSeek is trying to build AGI.
0:32:06 And I would say that Alibaba recently, the CEO, I think, said that we are racing to build
0:32:11 superintelligence. But I think it’s important here just to, like, name the biggest reason,
0:32:16 as you and I both know, that the U.S. is not regulating AI in any way and setting any guardrails
0:32:22 is for one reason, which is if we do anything to slow down or stop our progress, we’re just going
0:32:27 to lose to China. But let’s, like, flip that on its head for a second. The U.S. beat China to the
0:32:35 technology of social media. Did that make us stronger? Or did that make us weaker? If you beat
0:32:41 an adversary to a technology that you then don’t govern in a wise way, and instead, like, you built
0:32:44 this gun, you flip it around, you blow your own brain off, which is what we did with social media,
0:32:50 we have the worst critical thinking, test scores, you know, mental health, anxious, depressed
0:32:55 generation in history. And it’s a confusing picture because GDP is going up, that sort of
0:33:00 cancer is going up, too. So it’s like, we have the Magnificent Seven, we’re profiting from, you know,
0:33:03 all the wealth of these companies, but it’s actually not being distributed to everybody, except those who
0:33:09 are invested in the stock market. And that profit is based on the degradation of our social fabric.
0:33:13 So you have grandparents invested in their 401ks, invested in Snapchat, invested in Meta,
0:33:16 and their, you know, their portfolio is doing great, and they can take their holidays,
0:33:19 and they’re profiting off the degradation of their children and grandchildren.
0:33:25 Yeah, it’s really what you mean by beat, what are the metrics, because we’ve decided,
0:33:32 we’ve absolutely prioritized shareholder value over the well-being or the mental well-being of America.
0:33:36 It’s like we’re monetizing, we’re monetizing the flaws, and you’ve done great work around this,
0:33:44 around our instincts. You’ve compared, and I love this analogy, AI to NAFTA 2.0, and that is,
0:33:50 would essentially be an economic transformation that produced abundance, but hollowed out the
0:33:55 middle class. Walk us through this analogy. Yeah, sure. So, you know, we were sold this bill
0:34:01 of goods in the 1990s around free trade, global free trade, and this, we were promised this is going to
0:34:06 bring abundance to the country, and we’re going to get all these cheap goods. Well, part of that story
0:34:10 is true. We got this unbelievable new set of cheap goods from China, because this country appeared on the
0:34:15 world stage. We outsourced all the manufacturing to this country, and it produced everything super,
0:34:20 super cheap. But what did that do? It hollowed out the, you know, the middle class. So I just want to
0:34:26 make a parallel, because we’re told right now that these companies are racing to build this world of
0:34:31 abundance, and we’re going to get this unbelievable, you know, Elon Musk says we’re going to get universal
0:34:36 high income. And the metaphor here is instead of China being the new country that pops up on the
0:34:42 world stage. Now there’s this new Dario Amidai, the CEO of Anthropic, this new country of geniuses in a
0:34:49 data center that appears on the world stage. And it has a population of a billion AI beings that work at
0:34:55 superhuman speed, don’t whistleblow, generate new material science, new, you know, engineering, new AI
0:35:00 girlfriends, new everything. And it generates all that for super cheap. And so just like the, you know,
0:35:05 free trade NAFTA story, we got all the cheap goods, but it hollowed out the middle class. Well, now we’re going to get
0:35:11 all the cheap, you know, products and development and science, but it’s also going to hollow out
0:35:17 the entirety of our country. Because think of it like a new country of digital immigrants, right?
0:35:20 People, you know, Yuval Harari makes this metaphor. It’s like when you see a data center go up in
0:35:25 Virginia, and you’re sitting there, what you should see is like 10 million digital immigrants that just
0:35:31 took 10 million jobs. I think that people just need to unify these stories. And one other sort of
0:35:35 visual for this is like the game Jenga. The way we’re building our AI future right now is like,
0:35:37 if you look at the game Jenga, if you look at the top of the tower, you know, we’re putting a new
0:35:42 block on the top, like we’re going to get 5% GDP growth because we’re going to automate all this
0:35:46 labor. But how do we get that 5% GDP growth? We pulled out a block from the middle and the bottom
0:35:53 of the tower. That’s job security and a livelihood for, you know, those tens of millions of people
0:35:58 that now don’t have a new job. Because who’s going to retrain faster? The AI that’s been trained on
0:36:03 everything and is rapidly, you know, advancing in every domain or a human that’s going to try to train
0:36:07 a new cognitive, you know, labor. That’s not going to happen. And people need to get this because this
0:36:12 is different from other transitions. People always say, well, hey, you know, 150 years ago, everybody
0:36:16 was a farmer and now only 2% of people are farmers and see the world’s fine. Humans will always find
0:36:22 new things to do. But that’s different than this technology of AI, which is trained not to automate
0:36:27 one narrow task like a tractor, but to automate and be a tractor for everything. A tractor for law,
0:36:32 a tractor for biology, a tractor for, you know, coding and engineering, a tractor for science and
0:36:37 development. And that’s what’s distinct is that the AI will move to those new domains faster than
0:36:41 humans will. And so it’ll be much harder for humans to find long-term job security.
0:36:48 So I always like to ask, what could go right? And that is, I’m sort of with you around
0:36:59 the risk to mental health, to young people, to making us less mammalia, all the things that you’ve
0:37:04 been sounding the alarm on for a while, where I’m not sure I’m still trying to work it through is that
0:37:13 the catastrophizing around, you know, 40, 50, 70% of jobs could go away in two, five or 10 years, because
0:37:20 I generally find that the arc of technologies is there’s job destruction in the short and sometimes
0:37:26 the medium term, just as automation cleared out some jobs on the factory floor. But those profits and that
0:37:34 innovation creates new jobs. We didn’t envision heated seats or car stereos. Now, I agree at a minimum,
0:37:40 the V might be much deeper and more severe here. And America isn’t very good at taking care of the people
0:37:47 on the wrong side of the trade. But every technology in history has either gone away because it no longer
0:37:54 made economic sense, or it displaced jobs that no longer made sense, or it created profits and new
0:38:02 opportunities. Why do you see this technology as being different, that this will be not a V, but an L, and the
0:38:07 way down will be really serious. Do you see any probability that this, like every other technology
0:38:13 the medium and long term actually might be accretive to the employment force?
0:38:20 I mean, I cite people who are, who are bigger experts than I am, Anton Koronek, you know, Eric Bernholfsen at
0:38:26 Stanford. And what they show, I mean, Anton Koronek’s work is, in the short term, AI augments workers, right? It’s just
0:38:32 actually supercharging existing work that people are doing. And so it’s going to look good in the short term, you’re going to see
0:38:38 this, the curve looks like this, it kind of goes up, and then it basically crashes. Because what happens is AI is
0:38:44 training on that new domain, and then it replaces that domain. So I mean, let’s just make it really simple for
0:38:50 people to feel a very simple metaphor for this. What did we hear Instagram saying, and TikTok saying, for the last
0:38:55 several years, like, we’re all about creators, we love creativity, we want you to be successful. We are all about, you know,
0:39:00 making you be successful, make a lot of money. And then what was all that for? Well, they just released
0:39:07 this AI slop app. Meadow has one called Vibes, I think, and Sora is the open AI one. All of these AI slop
0:39:11 videos is sort of, are trained on all that stuff that creators have been making for the last 10 years.
0:39:16 So you, those guys were the suckers in this trade, which was, we’re actually stealing your training data
0:39:22 to replace you. And we can have a digital AI influencer that is actually publishing all the time,
0:39:27 and is just a pure advertising play and a pure sort of whatever gets people’s attention play.
0:39:29 And we’re going to replace those people and you’re not going to have that job back.
0:39:32 And so I think that’s a metaphor for what’s going to happen across the board.
0:39:38 You know, and people need to realize the stated mission of the, of open AI and anthropic and
0:39:47 Google deep mind is to build artificial general intelligence that’s built to automate all forms of
0:39:52 human labor in the economy. So when Elon Musk says that the optimist robot is a $20 trillion
0:39:57 market opportunity alone, what he’s, what he says, like the code word behind that, forget
0:40:00 whether you think it’s hype or not. The code word there is what he’s saying is I’m going
0:40:05 to own the global world labor economy. Labor will be owned by an AI economy. And so AI provides
0:40:09 more concentration of wealth and power than all other technologies in history, because you’re
0:40:14 able to aggregate all forms of human labor, not just one. So it’s like general electric becomes
0:40:15 general everything.
0:40:23 So let’s play this out because I’ve tried to do some economic analysis here and I look at the stock
0:40:29 prices and based on the expectations built into these stock prices of these AI companies is the notion
0:40:36 that they’re going to save at least three, maybe $5 trillion in oil, either add three or $5 trillion in
0:40:50 efficiencies, which is Latin for laying off people. I don’t see a lot of new AI moisturizers or cars from AI, at
0:40:54 least not yet. You could argue maybe autonomous, but I don’t see a lot of quote unquote AI products
0:41:02 increasing spend where I hear is Disney is going to save $30 million on legal fees, right? The customer
0:41:08 service is going away, the car salespeople, whatever it might be. So if you think in order to justify these
0:41:14 stock prices, you’re going to get a trillion dollars in efficiencies every year, a hundred thousand
0:41:21 dollars, you know, average job, 80,000 plus load. That’s approximately 10 million jobs a year if I’m doing
0:41:31 my math right. That is if half the workforce is immune from AI, masseuses, plumbers, that means 12 and a
0:41:37 half percent labor destruction per year across the vulnerable industries. So it feels like it’s either
0:41:42 going to be these companies either need to re-rate down 50, 70, 80 percent, which I actually think is
0:41:48 more likely, or we’re going to have chaos in the labor markets. So let’s assume we have chaos in the labor
0:41:54 markets because 12 and a half percent may not sound like a lot. That’s chaos. That’s total chaos. So say
0:41:59 we do have chaos in the labor markets. What do you think the policy recommendation is? Because the
0:42:03 Luddites were a group of people who broke into factories and destroyed the machines because they
0:42:07 said these things are going to put us out of work and destroy society. The queen wanted to make
0:42:12 weaving machines illegal because being a seamstress was the biggest employer of women.
0:42:18 What would be your policy recommendation to try and counter it? Is it UBI? Is it trying to put the
0:42:25 genie back in the bottle here? What do we, if in fact labor chaos is part of this AI future,
0:42:28 what do you think we need to do from a policy standpoint?
0:42:34 So people often think when they hear all this and they hear me and say, he’s a doomer or something
0:42:38 like that. I just want to get clear on what future we’re currently heading towards, what the default
0:42:43 trajectory is. And if we’re clear-eyed about that, clarity creates agency. If we don’t want that
0:42:47 future, if we don’t want, you know, millions of jobs automated without a transition plan where people
0:42:52 will not be able to put food on the table and retrain to something else fast enough, we have to do
0:42:57 something about that. If we don’t want AI-based surveillance states where AI and an LLM hooked up
0:43:02 to all these channels of information erases privacy and freedom forever, that’s a red line. We don’t want
0:43:08 that future. If AI creates AI companions that are incentivized to hack human attachment and screw up
0:43:12 the social fabric and young men and women and create AI girlfriends and relationships, that’s a red line.
0:43:18 We don’t want that. If AI creates, you know, inscrutable, crazy, super intelligent systems that
0:43:22 we don’t know how to control and we’re not on track to controlling, that’s a red line. So these are four
0:43:27 red lines that we can agree on. And then we can set policy to say, if we do not want the default
0:43:33 maximalist, you know, most reckless, no guardrails path future, we need a global movement for a different
0:43:40 path. And that’s a bigger tent. That’s not just one thing. It’s not just about jobs. It’s what is the AI future
0:43:45 that’s actually in service. So when you see that data center going up in your backyard, what is the set of
0:43:49 laws that says that that data center, when I see it, isn’t 10 million digital immigrants that’s going to replace
0:43:55 all my jobs and my livelihoods. That’s actually meant to support me. So what are the laws that get us there?
0:44:00 And my job and what I want people to get is to be part, you know, your role hearing all this is not
0:44:05 to solve the whole problem, but to be part of humanity’s collective immune system, using this
0:44:09 clarity of what we’re currently heading towards to advocate for we need a different future. People
0:44:13 should be calling their politicians saying AI is my number one issue that I’m voting on in the next
0:44:17 election. People should be saying, how do we pass AI liability laws? So there’s at least some
0:44:21 responsibility for the externalities that are not showing up on the balance sheets of these companies.
0:44:25 What is the lesson we learned from social media that if the companies aren’t responsible for the harms
0:44:30 that show up on their platform, because we had the section 230 free pass that created this blank
0:44:34 check to just go print money on all the harms that are currently getting generated. So there’s a,
0:44:37 there’s a dozen things that we can do from whistleblower protections to, you know, shipping
0:44:44 non-anthropomorphized AI relationships to having data dividends and data taxes to there’s, there’s a
0:44:49 hundred things that we can do. But the main thing is for the world to get clear that we don’t want the
0:44:54 current path. And I think to, in order to make that happen, there has to be a first snapping out
0:44:59 of the spell of everything that’s happening is just inevitable. Because I want people to notice that
0:45:04 what’s driving this whole race that we’re in right now is the belief that everything that’s happening
0:45:07 is inevitable. There’s no way to stop it. Someone’s going to build it. If I don’t build it, someone else
0:45:12 will. And then no one tries to do anything to get to a different future. And so we all just kind of hide in
0:45:18 denial from where we’re currently heading. And I want people to actually confront that reality so that we can
0:45:20 actually actively choose to steer to a different direction.
0:45:25 Do you think it can happen on a state by state or even a national level? Does it have to be
0:45:29 multinational? Like there are, you know, we’ve come together to say, all right, bioweapons are probably a bad
0:45:35 idea. And every nation with rare exception says, we’re just not going to play that game. There’s technology.
0:45:40 I may have even learned this from you where there are lasers that blind everyone on the field.
0:45:42 Yeah. And then we’ve decided not a good idea.
0:45:47 We decided we don’t want to do that. We have faced technological arms races before from nuclear
0:45:52 weapons. And, you know, what do we do there? If you go back, there’s a great video from, I think,
0:45:56 the 1960s where Robert Oppenheimer is asked, you know, how do we stop the spread of nuclear weapons?
0:46:02 And he takes a big puff of his, you know, cigarette and he says, it’s too late. If you wanted to stop it,
0:46:08 you would have had to stop the day after Trinity. But he was wrong. 20 years later, we did do arms
0:46:12 control talks and we worked all that time. And only nine countries have nuclear weapons instead
0:46:17 of 150. That’s a huge, serious accomplishment. Westinghouse and General Electric could have made
0:46:21 billions of dollars selling nuclear technology to the whole world. Keep keyword here being like
0:46:25 NVIDIA. But we said, hey, no, that’s actually even though there’s billions of dollars of revenue
0:46:31 there, that would create a fragility and the risk of nuclear, you know, catastrophes that we don’t want
0:46:36 to do. You know, we have done hard things before in the Montreal Protocol. We had the technology of
0:46:40 CFCs, this chemical technology that was used in refrigerants. And that collectively created this
0:46:45 ozone hole. It was a global problem from all these countries’ arms race in an arms race to sort of
0:46:52 deploy this CFC technology. And once we had scientific clarity about the ozone hole, 190 countries rallied
0:46:56 together in the Montreal Protocol. We did a podcast episode about it with the woman who, Susan Solomon,
0:47:01 who wrote the book on how we solved that problem. And countries rallied to domestically regulate their
0:47:06 domestic tech companies, their chemical companies, to actually reduce and phase out those chemicals,
0:47:11 you know, transitioning to alternatives that actually had to be developed. We are not doing that with AI
0:47:15 right now, but we can. You gave the example of blinding laser weapons. We could live in a world where
0:47:20 there’s an arms race to escalate to weapons that just have a laser that blinds everybody. But there was a
0:47:25 collective protocol in the UN 1990s where we basically said, yeah, even though that’s a way to win war,
0:47:29 that would be just inhumane. We don’t want to do that. And even if you think the U.S. and China
0:47:35 could never coordinate or negotiate any agreement on AI, I want people to know that when President Xi
0:47:40 met President Biden in the last meeting in 2023 and 2024, he had personally requested to add something
0:47:46 to the agenda, which was actually to prevent AI from being used in the nuclear command and control
0:47:51 systems, which shows that when both countries can recognize that their existential safety is being
0:47:55 threatened. They can come to agree on their existential safety, even while they are in,
0:48:00 you know, maximum rivalry and competition on every other domain. India and Pakistan were in a shooting
0:48:05 war in the 1960s, and they signed the Indus Water Treaty to collaborate on the existential safety
0:48:11 of their water supply. That treaty lasted for 60 years. So the point here is when we have enough of a view
0:48:17 that there’s a shared existential outcome that we have to avoid, countries can collaborate. We’ve done hard
0:48:22 things before. Part of this is snapping out of the amnesia. And again, this sort of spell of everything
0:48:27 is inevitable. We can do something about it. I like the thought. I just wonder if the technology or the
0:48:35 analogy might be a little bit dated because my fear is that the G7 or the G20 agree to slow development or
0:48:42 not advanced development around AI as it relates to weaponry. My fear is that it’s very hard to monitor.
0:48:48 you can monitor nuclear detonations. It’s really hard to monitor advances in AI and that this technology
0:48:56 is so egalitarian, if you will, or so cheap at a certain point that rogue actors or small
0:49:05 nation states or small terrorist groups could continue to run flat out while we, all the big G7 nations,
0:49:12 continue to agree to press pause. Is that mode of thinking, is our arms treaties a bit outdated here?
0:49:14 How might these treaties look different?
0:49:20 Absolutely. So let’s really dive into this. So you’re right that AI is a distinct kind of technology
0:49:26 and has different factors and more ubiquitously available than nuclear weapons, which required
0:49:32 uranium and plutonium. Exactly. But hey, it looked for a moment when we first invented nuclear
0:49:37 bombs, that this is just knowledge that everyone’s going to have. And there’s no way we can stop it.
0:49:41 And 150 countries are going to get nukes. And then that didn’t happen. And it wasn’t obvious to
0:49:45 people at that moment. I want people to relate. So there you are. It seems obvious that everyone’s
0:49:50 going to get this. How in the world could we stop it? Did we even conceptualize the seismic
0:49:56 monitoring equipment and the satellites that could look at people’s, you know, build outs of nuclear
0:49:59 technology and tracking the sources of uranium around the world and having intelligence agents
0:50:04 and tracking nuclear scientists? We had to build a whole global infrastructure, the International Atomic
0:50:10 Energy Agency, to deal with the problem of nuclear proliferation. And what uranium was for the spread of
0:50:15 nuclear weapons is for nuclear weapons. These advanced NVIDIA chips are for building the most
0:50:21 advanced AI. So yes, some rogue actor can have a small AI model doing something small. But only the big actors
0:50:30 can do something with this, like the bigger, more risky, closer to AGI level technology. And have we spent, you know,
0:50:34 people say it’s impossible to do something else. But has anybody saying that actually spent more than a week,
0:50:39 like dedicatedly trying to think about and conceptualize what that infrastructure could be?
0:50:43 There are companies like Lucid Computing that are building ways to retrofit data centers to have
0:50:47 kind of the nuclear monitoring and enforcement infrastructure where countries could verify
0:50:51 treaties, where they know what the other countries’ data centers are doing, but in a privacy protecting
0:50:55 way. We could map our data centers and have them on a shared map. We can have satellite monitoring,
0:50:59 looking at heat emissions and electrical signal monitoring, and understanding what kinds of training
0:51:05 runs might be happening on these AI models. To do this, the people who wrote AI 2027 believe that
0:51:10 you need to be tracking about 95% of the global compute in the world in order for agreements to
0:51:14 be possible. Because yes, there will be black projects and people going rogue on the agreement.
0:51:17 But as long as they only have a small percentage of the compute in the world,
0:51:21 they will not be at risk of building the crazy systems that we need global treaties around.
0:51:25 We’ll be right back.
0:51:36 Support for the show comes from Public.com. You’re thoughtful about where your money goes.
0:51:40 You’ve got your core holdings, some recurring crypto buys, maybe even a few strategic option
0:51:46 plays on the side. The point is, you’re engaged with your investments and Public gets that. That’s
0:51:51 why they built an investing platform for those who take it seriously. On Public, you can put together
0:51:58 multi-asset portfolio for the long haul. Stocks, bonds, options, crypto, it’s all there. Go to
0:52:04 public.com slash podcast and earn an uncapped 1% bonus when you transfer your portfolio. That’s
0:52:09 public.com slash podcast. Pay for by Public Investing. All investing involves the risk of loss,
0:52:14 including loss of principal. Brokered services for US-listed registered securities, options and
0:52:18 bonds, and a self-directed account are offered by Public Investing, Inc., member FINRA, and SIPC.
0:52:21 Complete disclosures available at public.com slash disclosures.
0:52:30 Support for the show comes from Upwork. You’re the CEO of your business. You’re also the CFO
0:52:36 and the IT department and customer service. It’s time to find some support. Upwork Business Plus helps
0:52:41 you bring in top quality freelancers fast. It gives you instant access to the top 1% of talent on Upwork
0:52:46 in fields such as marketing, design, AI, and more. All ready to jump in and take work off your plate.
0:52:51 Upwork Business Plus takes the hassle out of hiring by dropping trusted freelancers right in your lap.
0:52:56 Instead of spending weeks sorting through random resumes, Upwork Business Plus sources and vets
0:53:01 candidates for skills and reliability. It then sends a curated shortlist of proven expert talent to your
0:53:06 inbox in just hours so you can delegate with confidence. That way, you’re never stuck spinning
0:53:11 your wheels when you need a skilled pro and your projects don’t stall. Right now, when you spend $1,000
0:53:18 on Upwork Business Plus, you’ll get $500 in credit. Go to Upwork.com slash save now and claim the offer
0:53:26 before December 31st, 2025. Again, that’s Upwork.com slash S-A-V-E. Scale smarter with top talent and $500
0:53:29 in credit. Terms and conditions apply.
0:53:38 Support for the show comes from Superhuman. AI tools promise to make your work go faster, but if you have
0:53:43 to use multiple different tools that don’t sync up, you’re taking up more of your team’s time with tedious
0:53:48 project management and disjointed workflows, not to mention the constant context and tab switching.
0:53:54 Enter Superhuman. Superhuman is a suite of AI tools that includes multiple products, Grammarly,
0:53:59 Coda, and Superhuman Mail. The best part? Superhuman can guide you directly to where you’re working and
0:54:05 help you write clearly, focus on what matters, and turn your ideas into real results. That means no
0:54:09 constant switching between your workspace and multiple different AI tools. No more copy-pasting,
0:54:13 context-switching, and managing multiple AI tools across different places while you work.
0:54:17 Superhuman guides you directly where you’re working and helps you write clearly,
0:54:23 focus on what matters, and turn ideas into real results. Unleash your superhuman potential with
0:54:30 AI that meets you where you work. Learn more at superhuman.com slash podcast. That’s superhuman.com
0:54:47 slash podcast. We’re back with more from Tristan Harris. What is the glass half-full
0:54:53 top-full prediction? That we have decent regulation? For example, I think character
0:54:58 AIs can actually serve a productive role in terms of senior care. A lot of seniors who’ve lost their
0:55:03 friends and family. In seniors’ facilities, we’re going to have more of them that need companionship,
0:55:12 which staves off dementia and likelihood of stroke. What is the optimist case here for how AI could
0:55:17 potentially be regulated and unlock? Most technologies have ended up being accretive to society,
0:55:24 even the technologies that were supposed to end us, right? Nuclear power has become a pretty decent,
0:55:29 reliable source of energy. Obviously, electricity or fire, whatever you want to talk about.
0:55:37 Even processing power, pesticides that on even, I would argue even big tech on a net basis,
0:55:44 and I hate the word net, is a positive. What is the, give me the straw man’s case for what could go
0:55:49 right here and how might we end up with a future that AI is accretive to society?
0:55:53 Absolutely. I mean, that’s what this is all in service of is what is the narrow path that is not
0:55:58 the default maximalist rollout. And there’s two ways to failure. Let’s just name the twin sort of gutters
0:56:04 in the bowling alley. There’s one is you, quote, let it rip. You give everybody access to AI. You
0:56:09 open source it all. Every actor in society from every business to every developing country can train
0:56:14 their own custom AI in their own language. But then because you’re decentralizing all these benefits,
0:56:17 you’re also decentralizing all the risks. And now a rogue actor can do something very dangerous
0:56:21 with AI. So we have to be very careful about what we’re letting rip and how we open source it. So on
0:56:25 the other side, people say we have to lock it down. We have to have, you know, only five players do this
0:56:30 in a very safe and trusted way. This is more of the policy of the last administration. But then there,
0:56:34 you get the risk of a handful of actors that then accumulate all the wealth and all the power.
0:56:38 And there’s, you know, there’s no checks and balances on that because how do you have something
0:56:43 that’s a million times more powerful be checkable by other forces that don’t have that power?
0:56:47 And what we need to find is something like a commitment to a narrow path where we are
0:56:52 balancing responsibility and power along the way. And we have foresight and discernment about the
0:56:57 effects of every technology. So what would that look like? It’s like humanity wakes up and says,
0:57:01 we have to get onto another path. We pass basic laws, again, like liability laws and around AI
0:57:05 companions. We have AI companions. We have democratic deliberations where we say, hey, I wish you want
0:57:10 companion AIs for older people because they don’t carry the same developmental risks as they do for
0:57:15 young people. That’s a distinction we can have. We can have AI therapists that are more doing like
0:57:20 cognitive behavioral therapy and imagination exercises and mindfulness exercises without actually
0:57:24 anthropomorphizing and trying to be your best friend and trying to be an oracle where you share your
0:57:28 most intimate thoughts. So there’s different kinds of AI therapists. Instead of tutors that are trying
0:57:33 to, you know, be your oracle and your best friend at the same time, we can have narrow tutors that are
0:57:38 only domain specific like Khan Academy that teach you narrow lessons, but are not trying to be your best
0:57:43 friend about everything, which is where we’re currently going. So there’s a whole set of distinctions about
0:57:48 we can have this, not that we can have this, not that across tutors therapy, you know, AI that’s
0:57:54 augmenting work, um, AI that’s narrow AIs that take a lot less power by the way, and are more directly
0:57:59 applied. So for example, I have a friend who has found that he estimates that it would cost two to
0:58:03 10 orders of magnitude, less data and energy to train these narrow AIs. And you can apply it more
0:58:08 specifically to agriculture and get 30 to 50% boost in agriculture just from applying more narrow kinds of AI
0:58:14 rather than these super intelligent gods in a box. So there is another path, but it would take deploying AI
0:58:18 in a very different way. We could also be using AI to, by the way, accelerate governance. How do we apply
0:58:22 AI to look at the legal system and say, how do we sunset all the old laws that are actually not relevant
0:58:26 anymore for the new context? Hey, what were the spirit of those laws that we actually want to protect in
0:58:30 the new context? Hey, AI, could you go to work and kind of come up with the distinctions that we need to
0:58:35 help update all those laws? Could we help? Could we use AI to actually help find the common ground?
0:58:39 Audrey Tang’s work, the former digital minister of Taiwan, to find the common ground between all
0:58:45 citizens. So we’re reflecting back the invisible consensus of society rather than currently social
0:58:50 media is reflecting back the invisible division in society that’s actually making that more salient.
0:58:54 So what would happen? How quickly would it change if we had AIs that were gardening all the
0:58:59 relationships of our societal fabric? And I think that’s the principle of humane technology is that
0:59:04 there are these relationships in society that exist. I have a relationship to myself. You have a
0:59:08 relationship to yourself. Our phone right now is actually designed to replace the relationship I
0:59:13 have with myself. Humane technology would be technology and everyone else. And humane technology
0:59:17 would be trying to garden the relationship I have with myself. So things more like meditation apps
0:59:21 where that’s deepening my relationship to myself. Do not disturb is help deepening my relationship
0:59:26 to myself. Instead of AIs that are trying to replace friendship, we have AIs that are trying to
0:59:30 augment friendship. Things like Partiful or Moments or Luma, things that are trying to get people together
0:59:34 in physical spaces or find my friends on Apple. There’s a hundred examples of this stuff being
0:59:38 done in a way that’s gardening the relationship between people. And then you have Audrey Tang’s
0:59:42 work of gardening the relationship between political tribes where you’re actually showing and reflecting
0:59:47 back all the positive and visible areas of consensus and unlikely agreement across political division.
0:59:52 And that took Taiwan from, I believe, like 7% trust in government to something like 40% trust in
0:59:58 government over the course of the decade that they implemented her solutions on finding this kind of
1:00:01 bridge ranking. And that could be deployed across our whole system. So there’s totally a different
1:00:06 way that all of this can work if we got clear that we don’t want the current trajectory that we’re on.
1:00:12 So I just want to, in our remaining time here, you’ve been very generous. I want to talk a little bit
1:00:18 about you, Tristan. We’re kind of brothers from another mother, but we’ve kind of separated at birth
1:00:22 and grew up in different countries. And that is, we talk about the same stuff, but just sort of
1:00:26 through a different lens. You look at it through more of a humane lens. I look at it through more of a
1:00:34 markets lens. But I have noticed since 2017, when I started talking about, when my love affair with
1:00:40 big tech turned into sort of a cautionary tale, and I might be paranoid, but it doesn’t mean I’m wrong,
1:00:46 that slowly but surely I saw the percentage of comments across all my social media, across all
1:00:53 my content, more and more negative comments that appeared to be bots where I couldn’t figure out who it
1:00:58 was trying very strategically, methodically, and consistently to undermine my credibility across
1:01:05 anything I said. And I want to be clear, I might be paranoid, right? Because any negative commentary,
1:01:10 I have a, I have a knee flexor, I must be fucking Russians, right? Rather than maybe I just got it
1:01:18 wrong. You’re, you’ve been very consistent raising the alarm and you’re on the wrong side of the trade
1:01:24 around companies, multi-trillion dollar companies who are trying to grow shareholder value. I’m just
1:01:30 curious if you’ve registered the same sort of concerted effort, sometimes malicious, sometimes
1:01:36 covert, to undermine your credibility and what your relationship is with, with big tech.
1:01:43 That’s a great question, Scott. I appreciate that. And I think that there are paid actors that I could
1:01:47 identify over the course of the last many years. I’ve been doing this, Scott, for, you know, what, 12 years
1:01:51 now or something like that. We started the Center for Humane Technology in 2017. I just care about
1:01:55 things going well. I care about life. I care about connection. I care about a world that’s beautiful.
1:02:00 I know that that exists. I experience it in the communities that I’m a part of. I know that we
1:02:04 don’t have to have technology that’s designed in that perverse way. You know, this is all informed by
1:02:09 the ethos of the Macintosh Project at Apple, which my co-founder, Aza, his father started the Macintosh
1:02:13 Project at Apple. And we believe in a vision of humane technology. But to answer your question more
1:02:17 directly, I try to speak about these things in a way that is about the universal things that are
1:02:21 being threatened. So even if you’re an employee at these companies, you don’t want there to be a race
1:02:25 for, you know, the bottom of the brainstem to screw up people’s psychology and cause kids to commit
1:02:30 suicide. You don’t want that. So we need to, we actually need the people inside the companies on
1:02:34 side with this. This is not about us versus them. It’s about all of us versus a bad outcome.
1:02:39 And I always try to communicate in that way to recruit and enroll as many people in this sort of
1:02:44 better vision of this is, you know, there’s a better way that we can do all this. That doesn’t mean that
1:02:49 there are not negative paid actors that are trying to steer the discourse. There’s been op-eds,
1:02:54 hit jobs written, you know, trying to discredit me saying I’m doing this for the money. I care about
1:02:58 going in the speaking circuit and like writing books. Guess what? I don’t have a book out there. I make no
1:03:03 money from this. I worked on a nonprofit salary for the last 10 years. This is just about how do we get
1:03:12 to a good future? I love that. So you’re so buttoned up and so professional. Biggest influence
1:03:17 on your life? That’s a very interesting question. I mean, there’s, you know, public figures and people
1:03:25 who’ve inspired me. There’s also just my mother. I think she really came from love and she passed away
1:03:35 from Ketzer in 2018. And she was just made of pure love. And I, that’s just infused in me and what I
1:03:40 care about. And I’ve, I don’t know, I have a view that like life is very fragile and the things that
1:03:44 are beautiful are beautiful. And I want those beautiful things to continue forever.
1:03:51 I just love that. Chistan Harris is a former Google design ethicist, co-founder of the Center for Humane
1:03:57 Technology and one of the main voices behind the Social Dilemma. Chistan, whatever happens with AI,
1:04:03 it’s going to be better or less bad because of your efforts. You really are very, have been a powerful
1:04:06 and steadfast voice around this topic. Really appreciate your good work.
1:04:10 Thank you so much, Scott. I really appreciate yours as well. And thank you for, for having me on. I hope
1:04:13 this contributes to making that difference that we both want to see happen.
1:04:23 This episode was produced by Jennifer Sanchez. Our assistant producer is Laura Janair. Drew Burrows is our
1:04:26 technical director. Thank you for listening to the PropG Pod from PropG Media.
1:04:41 Support for this show comes from Odoo. Running a business is hard enough. So why make it harder
1:04:47 with a dozen different apps that don’t talk to each other? Introducing Odoo. It’s the only business
1:04:53 software you’ll ever need. It’s an all-in-one, fully integrated platform that makes your work easier.
1:05:00 CRM, accounting, inventory, e-commerce, and more. And the best part? Odoo replaces multiple expensive
1:05:05 platforms for a fraction of the cost. That’s why over thousands of businesses have made the switch.
1:05:12 So why not you? Try Odoo for free at Odoo.com. That’s O-D-O-O dot com.
1:05:23 You know that feeling when work’s scattered across emails, team chats, sticky notes, and someone’s memory?
1:05:30 Todoist helps small teams stay on track without needing to set up a complex system. Assign tasks,
1:05:36 see deadlines, and keep things moving. All in one place. It’s simple. Really.
1:05:41 Visit todoist.com and bring a little clarity to your chaos.