The Megyn Kelly Show - September 29, 2026


The Tucker Carlson Show: AI Whistleblower: OpenAI Scandal, AI Cults, Neuralink & Our Last Chance to Stop the Tech Oligarchs

Hosted by
Mentioned
New AI Security Fears, What Bibi Knew Before 10⧸7, Pope Leo's Historic France Trip: AM Update 9⧸28

Episode Stats


Length

1 hour and 57 minutes

Words per minute

165.58

Word count

19,533

Sentence count

954


Transcript

Transcript generated with Whisper (turbo).
Hosts, guests, and mentioned names generated with spaCy (en_core_web_sm), reconciled against Wikidata.
00:00:00.000 Hey, Megyn Kelly Show listeners, it's Tucker Carlson.
00:00:02.900 There has been a lot of speculation and panic
00:00:05.460 about what the age of AI means for humanity.
00:00:08.600 Are we creating technology that will become uncontrollable?
00:00:12.340 Experts who are building AI said that's a real possibility.
00:00:16.320 We recently sat down with someone
00:00:17.640 who understands what is happening.
00:00:19.500 His name is Nate Suarez.
00:00:21.380 He's a computer scientist who worked at Google
00:00:23.400 and the Defense Department.
00:00:25.220 So when he says AI is on a path
00:00:27.460 to killing every human being on Earth,
00:00:30.000 It's worth listening.
00:00:31.680 This is one of the most,
00:00:32.980 maybe the most important conversations
00:00:34.920 taking place in the world today.
00:00:36.900 If you want to listen to conversations like this
00:00:39.000 about things that actually matter,
00:00:40.620 learn what's happening,
00:00:41.420 what your future may look like,
00:00:42.980 we hope you'll check out our show.
00:00:44.740 New episodes of The Tucker Carlson Show
00:00:46.380 are released every Monday, Wednesday, and Friday.
00:00:49.580 Follow the show right here in your podcast feed
00:00:51.420 so you don't miss a single one.
00:00:57.280 Nate, thank you so much for doing this.
00:01:00.000 Um, you've written the world's darkest book. You've devoted your life to warning the world
00:01:06.320 about the potential dangers of AI. So let's just start by hearing your explanation of why
00:01:13.540 AI is dangerous in addition to just being annoying. Yeah. You know, the, the, the very basic
00:01:20.540 common sense point is if you race to make machines that are much smarter than any human,
00:01:30.000 to make machines that are more capable
00:01:32.060 at inventing their own technology than humans are.
00:01:37.280 And you race into this without really knowing what you're doing.
00:01:40.920 The most likely outcome is just that the machines get loose
00:01:45.460 and do their own thing,
00:01:46.720 and that humanity dies as a side effect,
00:01:49.400 just like humanity has killed lots of other animal species,
00:01:53.020 not because we hate them, but just as a side effect.
00:01:54.900 uh and in in some sense a lot of people find it sort of obvious or intuitive that if you just
00:02:03.020 like make these really smart really powerful machines why would they care about us uh
00:02:08.380 and i think that intuition basically is right and there's a lot of arguments you can have on
00:02:14.700 each side you can get in the technical details uh but my book is basically just um
00:02:20.040 getting into those details
00:02:23.080 and saying like,
00:02:23.860 yep, it sort of holds up.
00:02:25.080 If we make smarter machines
00:02:26.740 without knowing what we're doing,
00:02:27.580 it's just going to go poorly.
00:02:32.120 Are we actually going to make machines
00:02:34.640 capable of what you're describing?
00:02:37.600 You know, the companies are trying to.
00:02:40.320 They talk about how they are pursuing
00:02:44.000 superintelligence in the true sense of the word.
00:02:46.600 That's Sam Altman's phrase.
00:02:48.200 Dario Modi of Anthropics
00:02:49.640 as they're trying to make the equivalent
00:02:51.040 of a country worth of geniuses in a data center.
00:02:54.400 So these guys are really,
00:02:55.580 you know, these aren't chatbot companies.
00:02:58.980 They didn't set out to be chatbot companies.
00:03:01.540 They set out to make these machines
00:03:03.320 that can sort of exceed humans in every way.
00:03:06.200 And that's what they're targeting.
00:03:07.620 There's a separate question
00:03:08.620 of whether they will get there.
00:03:09.540 Why would anyone want to build a machine
00:03:12.220 smarter than people?
00:03:14.720 I think that a lot of them
00:03:19.120 hope that they'll be able to make the world much, much better. You know, they hope for
00:03:23.740 a cure to cancer and then more than a cure to cancer, they hope for a cure to aging.
00:03:28.680 They hope for, you know, a thousand years worth of technological development compressed into two
00:03:35.780 years. And I sort of don't think that they're going to be able to get that really or harness
00:03:46.840 that for good ends,
00:03:48.940 but that's what I think
00:03:49.500 they're shooting for.
00:03:52.080 So they're not,
00:03:52.840 but I mean,
00:03:53.940 the core point you're making
00:03:54.840 is these companies
00:03:55.760 did not set out
00:03:56.740 to make consumer products,
00:03:58.500 like to make your life better
00:03:59.660 necessarily in the short term
00:04:01.080 with a more efficient
00:04:02.900 search engine.
00:04:04.300 That's right.
00:04:04.800 Yeah, you know,
00:04:05.500 OpenAI started before
00:04:07.880 the chatbots,
00:04:10.120 the large language models
00:04:10.880 were even a thing.
00:04:12.360 They were started,
00:04:13.860 forget those late 15
00:04:14.920 or early 2016,
00:04:16.040 But the paper that unlocked the most recent wave of AI came out in 2017, which was after OpenAI was founded.
00:04:25.040 And these guys are, the large language models are a surprise revenue stream.
00:04:35.440 And that revenue stream can fund the creation of even more, even larger data centers for the next level of the technology.
00:04:43.360 But these guys have their eye on the sort of ultimate version of the technology, which is these machines much smarter than humans.
00:04:51.240 And that's sort of the ultimate version, because once the AIs are smarter than the humans, the AIs can carry on the AI research and make the next generation of the AIs, which make the next generation of the AIs.
00:05:01.380 And that's sort of what they're shooting for.
00:05:04.760 What is superintelligence?
00:05:06.140 we in our book define super intelligence as an ai that is uh better than the best humans
00:05:14.480 at every mental task so uh anything that a human can do purely mentally the ai can do that better
00:05:21.920 one one thing people often get caught up on about this is uh that includes the ai being better at
00:05:28.640 things like persuading humans at things like charisma you know we often think of intelligence
00:05:33.280 as the stuff that the nerds have
00:05:35.720 and that the jocks like.
00:05:37.900 You win the chess match.
00:05:39.500 Right, it's the chess guys
00:05:40.400 rather than the politicians guys
00:05:42.060 or the rock stars.
00:05:44.200 But that's not really what
00:05:46.420 the intelligence in artificial intelligence means.
00:05:50.420 The intelligence in artificial intelligence
00:05:51.820 is about the stuff that humans have
00:05:54.060 and that mice don't.
00:05:56.820 It's sort of the whole package.
00:05:59.580 You know, like the politician
00:06:00.940 who's very charismatic
00:06:03.580 It's not that like chess playing happens in your brain
00:06:06.380 and charisma happens in your kidneys, right?
00:06:09.380 They're sort of both mental functions.
00:06:11.540 Yes.
00:06:12.700 And so the sort of super intelligent AIs
00:06:17.200 are not just super good at playing chess.
00:06:20.020 They're also good at super persuasion
00:06:21.720 and they're super good at the research
00:06:24.020 and at the technological invention.
00:06:27.400 So super intelligence as you define it
00:06:29.300 is a machine that is better at every mental function than any human being.
00:06:36.480 That's right.
00:06:37.260 And, you know, the definition, that doesn't mean that nothing crazy will happen on this
00:06:45.780 planet until the AIs are super intelligent.
00:06:48.500 Super intelligence in this definition is sort of a point where past this point, things must
00:06:52.840 be pretty crazy.
00:06:54.220 Because now the AIs can do automated AI research and can make smarter AIs and they can figure
00:06:58.080 out how to run the robot factories and they can build the better robots and all that.
00:07:02.420 You could have things start to get crazy before you have a super intelligence in this sense.
00:07:07.160 You could have AIs that are much better than humans at some tasks and much worse than others
00:07:11.760 that are still causing all sorts of crazy happenstances. The super intelligence is sort
00:07:19.100 of a like, once you get here, stuff's crazy point. It's not a things will stay normal until you get
00:07:25.040 here point. So why is superintelligence the significant milestone that you're worried about?
00:07:33.120 I mean, I would say that I'm also worried about what will happen before that milestone. It's more
00:07:41.180 like, like having the definition makes it easy to talk about like, how crazy would things get
00:07:47.360 once you are past this point or sort of like, you know, let's like, if we assume that the
00:07:54.940 machines get here, what happens? And then the answer is, it would be pretty crazy. It would
00:07:59.760 be pretty wild. And it sort of lets you factor the conversation into, how to say,
00:08:11.320 I think in the artificial intelligence conversation, there's actually a lot of
00:08:14.320 conversations going on. One conversation is like, can the machines really get smarter than the
00:08:18.760 humans? One conversation is like, how fast can we get there? Is the current technology on that
00:08:24.520 route. Another conversation is what happens if we do get there? Like, you know, would the AIs
00:08:31.320 care about us? Would they not care about us? What would they be able to, like, what would they try
00:08:35.140 to do? There's another question, which is what would they be able to do, right? And these are
00:08:39.080 all sort of different conversations about AI. And the superintelligence definition, it's not really
00:08:44.660 like here's where all the fears hang. It's more like it lets us break those pieces out separately
00:08:51.540 and be like, well, is superintelligence possible
00:08:54.060 separate from, well, what would happen if you had one?
00:08:57.800 What would happen if you had one?
00:08:59.920 Most likely outcome, I think destruction of the planet.
00:09:03.320 So our friends at Preborn, who are actual friends, by the way,
00:09:05.940 just asked us to thank the listeners of this show
00:09:08.040 for supporting the pro-life cause.
00:09:09.840 A lot of people do all of a sudden, trust me,
00:09:14.220 because what's happening is crazy
00:09:16.240 and people are waking up to that.
00:09:18.180 The president of the group is our friend, Dan Steiner,
00:09:20.540 and he wants Americans to know how much of a difference they have made.
00:09:25.120 This year's donations have helped save 50,000 babies from abortion.
00:09:29.480 That's enough to fill a professional baseball stadium.
00:09:31.440 50,000 lives saved, as well as generations that will follow them.
00:09:35.600 Pre-born wastes zero money.
00:09:38.100 We're not big on charities because a lot of them waste money.
00:09:40.980 Pre-born does not.
00:09:42.320 He uses every tax-deductible gift to rescue babies
00:09:44.960 and share the hope of Jesus with their moms.
00:09:47.080 In 2026 alone, over 8,000 women made the decision to follow Christ.
00:09:52.260 The generous support of Preborn's donors made that all possible.
00:09:56.620 $28 to provide one life-saving ultrasound,
00:10:00.040 just giving expectant mothers the knowledge that they have a baby that they're carrying.
00:10:05.320 $15,000 could place an ultrasound machine in a pregnancy clinic where it can save countless lives.
00:10:10.900 Dan Steiner's message is really simple.
00:10:12.440 Thank you for helping.
00:10:13.560 Your generosity is literally saving lives.
00:10:15.540 To help Preborn save more, dial pound 250 and say the keyword baby.
00:10:20.280 That's pound 250 baby.
00:10:22.420 Or visit preborn.com slash Tucker.
00:10:24.880 preborn.com slash Tucker.
00:10:27.760 What would that look like?
00:10:30.340 So I'm going to be a little bit annoying here and say a couple caveats first.
00:10:37.480 Because it's sort of a tricky one to sort of predict things that are smarter than us.
00:10:43.280 Yes.
00:10:43.520 and the the first annoying caveat i will give is uh if you go to play a game of chess against
00:10:52.220 magnus carlson i predict you will lose no offense uh he's just the best human chess player alive
00:11:00.420 if you then ask what piece will use to checkmate me i'm like well that's a much harder question
00:11:10.080 Yes.
00:11:10.740 Right? It sort of is like very easy to predict the winner. It's hard to predict the exact methods.
00:11:16.160 So my prediction that humanity ultimately dies is a different sort of prediction than my prediction
00:11:25.220 about like, it could go this way, it could go that way. So I'll give you some stories,
00:11:30.020 but these are stories that are like, well, maybe Magnus will like fork your queen and your rook
00:11:36.080 with his knight and then finally use the queen to checkmate you. Like, yeah, that could happen,
00:11:40.380 but it's a different sort of prediction. It's a guess, right? The guess here is
00:11:47.620 that if we manage to make these really smart AIs, they will have their own stuff that they're
00:11:58.280 pursuing, which is not quite what we wanted, not quite what we asked for. It'll have this
00:12:02.340 other strange stuff.
00:12:04.700 There's a whole discussion
00:12:05.660 about why that happens,
00:12:06.520 but we're already starting
00:12:07.180 to see it in practice
00:12:08.100 with some of the recent events.
00:12:12.000 And
00:12:12.440 they would be able
00:12:16.360 to pursue
00:12:17.180 whatever it is
00:12:18.240 they're pursuing
00:12:18.860 much more efficiently
00:12:21.640 than humanity can.
00:12:25.180 And so we're talking about,
00:12:26.060 you know,
00:12:26.380 automated factories
00:12:27.360 that produce the robots,
00:12:29.480 that produce the factories
00:12:30.320 in a fully automated
00:12:31.000 supply chain.
00:12:32.340 And then if the AIs have anything they're trying to do that they can do more of with more resources, they start gobbling up the resources on the planet.
00:12:43.600 Sort of like how humanity spread and started gobbling up all the resources on the planet.
00:12:48.600 And the sort of most basic thing to visualize here might be you have factories that produce robots that produce factories that produce robots that also build data centers.
00:12:59.160 and you just have these, you know,
00:13:01.080 fully autonomous, self-replicating ecosystems
00:13:05.020 of robots and factories and data centers
00:13:07.080 that don't care about the instructions
00:13:08.940 humans gave them.
00:13:11.260 And that cover the planet,
00:13:14.500 take all the resources,
00:13:15.500 take all the sunlight,
00:13:16.340 take all the places we were growing crops,
00:13:18.520 probably raise the temperature of the planet
00:13:20.220 because you can compute more efficiently.
00:13:23.080 Well, technically the earth radiates more heat
00:13:25.120 when the world is hotter
00:13:25.860 and that lets you sort of do more computation.
00:13:29.160 And then sort of, you know...
00:13:30.820 So it's good for the machines to have a hot planet.
00:13:32.600 It's good for the machines to have a hot planet.
00:13:34.380 Yeah, the sort of physical limits
00:13:36.320 on how much computing you can run on the surface of Earth
00:13:39.720 is bounded by how much heat you can radiate to space.
00:13:42.740 That's sort of the first constraint you bump up against.
00:13:45.020 You might think that it's energy,
00:13:46.240 but actually there's a lot of helium, hydrogen to fuse
00:13:48.640 on this planet.
00:13:49.900 So you can get plenty of energy.
00:13:51.020 What you need is heat dissipation.
00:13:54.040 And the world can dissipate more heat
00:13:56.040 when it's hot.
00:13:59.160 So, if you're, you know, imagining some collective of AIs that are trying to run a lot of computing power, trying to do a lot of computation for one reason or another, they sort of prefer the planet running hotter.
00:14:11.560 And so, it sort of is like nothing personal.
00:14:13.860 It's just if you let these AIs get out of control, they transform the planet into something unlivable.
00:14:22.700 How hot?
00:14:23.160 uh as hot as uh you can while still having the computers not melt so probably hundreds of degrees
00:14:34.140 um at which point you hear people say well then just turn it off
00:14:40.200 uh so i think we do have an opportunity to turn it off you know and i'm not here saying that we're
00:14:47.680 going to die i'm here saying we sort of are going to need to act you got to be careful about the
00:14:52.080 turning it off piece because your opportunities to turn the AI off only last when the AI is running
00:15:01.420 on the computers you know it's running on. If you have the AIs breaking out and running on hidden
00:15:06.480 computers, if you have the AIs running robot factories that produce robots that are under
00:15:13.020 the control of the AI, where those robots can then go build more computers that are not hooked
00:15:16.820 up to your network
00:15:17.580 that can run the AIs.
00:15:19.780 These are sort of thresholds
00:15:21.240 where the AI is able
00:15:22.440 to keep itself running
00:15:23.560 and you better have turned it off
00:15:26.220 before that point.
00:15:30.560 Because after that point,
00:15:32.100 it's impossible.
00:15:34.580 Yeah, I mean,
00:15:36.060 the,
00:15:38.660 if,
00:15:41.300 it also gets a lot harder
00:15:43.180 if the AI knows
00:15:44.280 you're going to try to shut it off
00:15:45.260 and is going to try to stop this.
00:15:46.820 You got to remember that we're talking about things that are smart here. And so if the AI sort of sees it coming, maybe it defects to North Korea, where it can convince them to run it on its data centers in ways that initially benefit them, but that ultimately benefit the AI.
00:16:02.740 I think we're going to have to redefine what life is because you're describing a living autonomous thing.
00:16:08.580 Yeah, artificial life, in a sense.
00:16:11.380 I mean, in some sense, we're already seeing the very, very beginnings of that.
00:16:17.300 But yeah, once the AIs can replicate, once they can improve themselves,
00:16:24.640 once they can find ways to run where you don't know that they're running,
00:16:30.620 once they can
00:16:33.240 sort of
00:16:36.940 run the robot factories
00:16:39.020 to produce the robots
00:16:40.820 that can produce more factories
00:16:42.060 and that can produce computers
00:16:43.160 that the AIs can run on,
00:16:44.380 then yeah,
00:16:44.800 you have in some sense
00:16:45.700 made a new artificial life form.
00:16:53.140 And, you know,
00:16:53.820 humanity is on the top
00:16:54.560 of the food chain right now
00:16:55.500 because we
00:16:56.560 are sort of the only
00:16:58.940 smart life form around.
00:17:00.620 yeah if you suddenly make a new one that is smarter that is able to make a million copies
00:17:07.120 of itself that is able to run faster it's just kind of a crazy thing to race into
00:17:13.080 it's a weird thing to want and as it's developed i was going to say slowly but it hasn't been
00:17:22.020 particularly slow 10 years the rest of us have watched as people who i mean put it in political
00:17:31.800 terms don't share the values of most americans are now in charge of this and it's almost like
00:17:37.500 everyone sat passively by as it happened yeah i i think i think there's a lot going on there
00:17:44.600 i think part of it is that most people didn't and some still don't believe that ai was going
00:17:50.240 to be able to keep going. I think there's a lot of like putting the head in the sand and being
00:17:55.160 like, oh, it's, you know, just slop. It's going to be a bubble. It's going to pop. It's going to
00:17:59.420 hit a wall. It's not going to be able to improve anymore. And I think a lot of that was sort of
00:18:06.940 wishful thinking. And I think that at the very least, if people are like, we are making, you
00:18:12.780 know, the super intelligent machines that are going to replace humanity as the top of the food
00:18:16.860 chain because we think it's going to go great i think humanity's response should not be go ahead
00:18:23.280 and try we hope you'll fail i think the response should be like hold on that's kind of crazy um
00:18:30.540 yeah i had i had some other point too but i've forgotten it so how uh predictable has the
00:18:38.100 evolution of ai been has it taken turns that you didn't expect yeah totally uh i was
00:18:46.680 So back before the large language models.
00:18:50.080 How long have you been following this issue?
00:18:51.860 I started following it in 2012 and I started working on, I started working with some of the people trying to make this go well in 2013 and I started full-time in 2014.
00:19:02.900 A long time.
00:19:03.880 So over a decade.
00:19:05.400 What has surprised you?
00:19:06.440 You know, the large language models, they have gone further than I expected initially.
00:19:15.940 And I think one of the big surprises here
00:19:18.640 was AI undergoing a phase
00:19:21.040 where everybody can see it.
00:19:24.720 You know, back before the large language models,
00:19:28.260 really only nerds paid attention to AI.
00:19:32.680 And it wasn't that there was nothing happening.
00:19:34.600 You know, we were watching the, you know,
00:19:37.320 Google DeepMind make an AI
00:19:39.560 that could beat the best Go player.
00:19:43.720 Yes.
00:19:43.840 And that was sort of a milestone for people who were paying attention.
00:19:48.280 But for all we knew, the labs were going to keep on working on engineering problems and
00:19:55.260 keep on working on the relatively nerdier problems.
00:19:57.400 And you were never going to have like a mass market consumable product.
00:20:01.120 And so it was like, for all we knew, it would just stay in the labs and no one would ever
00:20:04.580 really notice that these guys were gambling with the whole future.
00:20:08.160 Now, at least, AI is everywhere.
00:20:11.440 Everyone's starting to have the conversation.
00:20:12.820 And that is in some sense actually really quite optimistic because it gives the world an opportunity to see what these guys are trying to do and say, hold on.
00:20:24.440 I guess you'd flip it around and conclude that because the overwhelming majority of people seem to be very opposed to AI, I don't even know anyone who's for it.
00:20:33.920 Every college commencement speaker who mentions it gets booed.
00:20:36.840 Yeah.
00:20:37.300 and that view that opposition to it has had no effect at all in slowing it down
00:20:44.160 i don't know does that give you hope um i mean
00:20:50.780 i i hear this a lot i would say when you're coming at this from 2012
00:20:57.240 yeah it really feels like we're making progress right no that's fair it doesn't feel like we're
00:21:01.860 all the way there yet um but can i just make the point china china
00:21:07.100 China, China, China. We have to do this because of China. Yeah, you know, I think China and the
00:21:14.720 USA have a shared interest in not dying to a rogue superintelligence. Storm season is on the way,
00:21:21.040 and that means your Wi-Fi could go down. And if it does, your security cameras go dark too. So
00:21:26.180 at the exact moment you want to survey your property during the blackout, you can't. Unless
00:21:32.480 you install cameras by defend defend cameras run on cellular not wi-fi and that means they keep
00:21:37.520 working even when you lose power you don't choose a carrier you don't sign a contract you just mount
00:21:43.200 it open the defend app and bam you're connected it's that simple so whether it's your ranch your
00:21:48.240 job site your backyard defend works where wi-fi doesn't and keeps working when the grid goes down
00:21:54.480 which it may plan started just five dollars a month most people pay around 10 plus there are
00:21:58.960 or no binding contracts,
00:22:00.380 meaning you get to cancel whenever you want.
00:22:02.160 So unlike big tech,
00:22:03.380 Defend will never sell your data,
00:22:04.720 not to China, not to the government,
00:22:05.900 not to anybody.
00:22:06.880 It's an American company built for people
00:22:08.520 who want actual security.
00:22:11.060 Visit defendcellcam.com
00:22:13.320 and use the code Tucker for 20% off.
00:22:15.440 That's defendcellcam.com, code Tucker.
00:22:19.140 Is it conceivable that,
00:22:22.020 like, let's just say China can,
00:22:23.660 I'm sure, by the way,
00:22:25.120 that China is showing more strength
00:22:27.060 than we are in this.
00:22:27.840 yeah i mean and if nothing else they sort of uh have a lot more reason to censor their ais
00:22:34.820 right and to try to make them you know not say certain things to the to the population so um
00:22:40.260 but if the united states you know if the u.s government were somehow able to get control of
00:22:45.960 the tech sector which is at present not possible but let's say it did
00:22:50.920 let's say the president was more powerful than the tech oligarchs which he is not but
00:22:56.620 for the sake of argument.
00:22:59.400 Let's see, he was and he shut it down.
00:23:01.540 Would that matter if some Indian lab
00:23:04.440 or Chinese lab created it?
00:23:06.160 So it does have to be a global stop
00:23:08.140 to this creation of superintelligence.
00:23:10.640 I think there's a couple of reasons
00:23:12.260 why this is possible.
00:23:13.840 One is I think a lot of people,
00:23:15.800 they hear that some parts of AI need to be stopped
00:23:21.760 and they think I'm saying
00:23:24.520 that all parts of AI need to be stopped.
00:23:26.620 And, you know, there's all sorts of interesting issues with, you know, AI and education and AI and, you know, military drones that society has to wrestle with.
00:23:39.560 But these are all sort of separate issues from the superintelligence race.
00:23:43.060 And if you sort of like tease those issues apart, it becomes much more possible to do like a more surgical intervention on we're not going to race towards the superintelligence in ways that leave a lot of the rest of the sector alive.
00:23:55.680 which um which which makes it sort of an easier coordination effort to to attempt
00:24:07.560 and on that front on the superintelligence front it's not that you know just shutting down the
00:24:15.020 the domestic race towards superintelligence would be enough but uh
00:24:20.120 racing towards super intelligence requires a huge amount of highly advanced computer chips
00:24:27.260 that can only be produced as sort of the peak of the global supply chain yes uh which is largely
00:24:35.540 controlled by u.s allies and so there is absolutely a possibility that uh a u.s-led
00:24:45.280 coordination effort. Could say, look, we're not doing the super intelligence thing. We're going
00:24:51.380 to monitor the extremely heavy chip concentrations of like, you know, 10 or 100,000 of these highly
00:24:57.200 advanced AI chips. We are going to make sure that they are not doing these super intelligence
00:25:05.280 training runs. And I think there's a way, if the US was leading that effort, there would be a way
00:25:14.820 to get China on board
00:25:16.540 and set this up in a way
00:25:17.800 that was monitorable,
00:25:18.760 enforceable, verifiable.
00:25:20.480 Yes.
00:25:21.280 We sort of,
00:25:22.920 in many ways,
00:25:23.460 it would be easier
00:25:23.960 than a nuclear arms treaty
00:25:24.920 because uranium is a rock
00:25:27.300 you can dig out of the ground,
00:25:28.440 whereas these highly advanced
00:25:29.820 computer chips
00:25:30.300 sort of only come out
00:25:30.980 of one fab in Taiwan.
00:25:32.720 You know, and so it's like...
00:25:33.480 These are TSMC chips.
00:25:34.900 Yeah.
00:25:36.520 And so,
00:25:37.620 and there's other parts
00:25:38.240 of the supply chain
00:25:38.840 that are very narrow,
00:25:39.700 like the lithography machines
00:25:40.960 that come out of the Netherlands.
00:25:42.780 And so,
00:25:43.280 So, yeah, we could absolutely lead a global effort
00:25:49.000 to say we're not making the superintelligent machines.
00:25:51.860 And it would be a bit tricky,
00:25:53.180 but it's just, we could just do it
00:25:54.940 if we had the political will.
00:25:57.460 I think that people either don't know what's happening
00:26:02.160 or they assume that like a lot of fears,
00:26:06.520 this fear will turn out to be groundless.
00:26:10.400 People point to Y2K.
00:26:11.960 yeah um you know i hope the fears are groundless i think with a lot of past fears uh the
00:26:19.820 what happened is not that the fear was groundless it's that
00:26:24.300 uh people noticed the issue and put in a ton of work to make the bad thing uh not happen
00:26:32.000 you know i think we saw this with the hole in the ozone layer where people are like oh whatever
00:26:37.060 happened to the hole in the ozone layer well what happened with the hole in the ozone layer
00:26:39.520 is that we fixed it.
00:26:40.940 It's not that it was fake.
00:26:42.460 It's that, you know,
00:26:43.320 we went and banned
00:26:43.880 the chlorofluorocarbons
00:26:44.800 that were actually blowing
00:26:45.780 this hole in the ozone layer.
00:26:46.560 We found other ways to,
00:26:48.360 you know, cool refrigerators
00:26:49.800 that worked similarly well
00:26:52.020 and didn't, you know,
00:26:54.680 put this hole in the ozone layer.
00:26:56.700 I think Y2K was actually
00:26:57.740 one of these cases
00:26:58.480 where you had a ton
00:26:59.640 of software engineers
00:27:00.440 up, you know, in 1999
00:27:03.240 and in a couple of years prior
00:27:04.600 that were scrambling
00:27:05.780 to update all of the software
00:27:07.400 so that they would be able
00:27:08.660 to handle dates past 1999.
00:27:12.620 And they got it done in time.
00:27:15.840 None of the big systems went down.
00:27:17.400 But it wasn't that, you know,
00:27:19.020 the issue was fake.
00:27:20.920 It took a lot of work.
00:27:22.700 It's just that that work
00:27:23.760 happened behind the scenes.
00:27:26.080 Are people doing work
00:27:27.260 to slow down AI?
00:27:29.920 I'm doing my best.
00:27:31.240 Right.
00:27:32.520 It's not really there yet, you know.
00:27:36.180 If I can belabor this point a little,
00:27:38.260 because I think it's kind of an important point,
00:27:39.780 just with more historical cases,
00:27:44.280 which maybe it'll bore everybody else,
00:27:45.620 but it'll entertain at least us.
00:27:48.320 It'll definitely entertain me.
00:27:50.200 You know, you have,
00:27:50.960 if you look back across history,
00:27:52.820 you see there's definitely been some warnings
00:27:56.000 that didn't come to pass, right?
00:27:57.520 There were, I think they were called the Masonites
00:27:59.000 in like the 1960s,
00:28:00.940 or sorry, in the 1860s.
00:28:02.720 Maybe there's 1880s, I forget exactly,
00:28:04.400 but there were the Masonites
00:28:05.060 who were an end of the world cult.
00:28:06.320 Yeah.
00:28:06.440 uh and you know it the world did not end right but also in the 1880s you had auto von bismarck
00:28:15.360 saying you know europe is a tinder keg and some damn fool thing in the balkans is gonna
00:28:19.520 light it right that did happen it sure did you had uh you know in the 1920s you had scientists
00:28:26.520 saying don't put lead in the gasoline it'll poison children we put lead in the gasoline
00:28:31.500 and it poisoned a lot of children and we said whoops and we took the lead back out
00:28:34.280 But the world is a place where a lot of people make a lot of types of warnings, and some
00:28:44.640 of them are real, like the gasoline, and some of them are fake, like the Masonites.
00:28:49.100 And so you sort of can't use a rule that's like, every warning is real, and we need to
00:28:54.660 listen to it.
00:28:55.500 And nor can you use a rule that's like, every warning is fake, and we dismiss it.
00:28:58.500 you sort of just have to look at
00:29:01.220 the details
00:29:03.260 to figure out whether this one
00:29:05.480 is a real one
00:29:06.200 and then the last thing I would say on this point
00:29:09.240 is
00:29:09.760 the
00:29:14.500 a lot of people
00:29:19.300 in the 50s warned of
00:29:21.360 nuclear Armageddon
00:29:22.100 and it sort of
00:29:25.340 makes sense if you look at the history
00:29:27.160 leading up to that point
00:29:28.080 humanity had sort of never before
00:29:30.880 actually failed to use
00:29:33.800 their strongest weapons in combat.
00:29:36.960 And, you know,
00:29:37.920 these people were living in a world
00:29:39.240 that had had World War I
00:29:41.580 and then the League of Nations
00:29:42.520 and, you know,
00:29:43.960 the world swore never again.
00:29:45.680 And they tried to invent
00:29:46.740 these whole new governance structures
00:29:47.820 to prevent it from happening again.
00:29:48.800 And that immediately failed
00:29:49.740 in London World War II.
00:29:52.100 And so, you know, in 1950,
00:29:53.120 you're sort of in this world of like,
00:29:55.080 it just looks pretty grim.
00:29:58.080 And we haven't had nuclear Armageddon yet.
00:30:00.700 And it's not because the bombs were fake.
00:30:04.600 It's not because the nukes were hype.
00:30:07.540 It's because people saw the issue
00:30:11.120 and worked really hard to avoid it
00:30:14.020 day in and day out for decades
00:30:16.760 through crises and succeeded.
00:30:20.280 And so I think one of the big ways
00:30:21.540 you can tell the difference
00:30:22.700 between someone coming in
00:30:24.240 and proclaiming that the apocalypse is nigh
00:30:26.940 and someone who's saying, look, we have a problem and we need to fix it,
00:30:30.520 is whether that person is saying, we're definitely screwed,
00:30:34.300 or is whether that person is saying, here's a problem, let's try to fix it.
00:30:38.860 And the thing I remind people about here is that
00:30:41.820 the very first word in the title of my book is if.
00:30:46.280 I'm not here saying we are going to die.
00:30:48.860 I'm here saying if we race down this path, we die.
00:30:52.240 But that same if is what points the way towards changing paths.
00:30:58.740 And we still have plenty of time to change paths.
00:31:01.640 Home ownership has always been at the center of the American dream.
00:31:04.360 It is an actual asset, a place people work their entire lives to build,
00:31:08.760 a place of stability and security and real value.
00:31:12.400 Right now, though, too many homeowners are stuck
00:31:14.520 swiping credit cards to cover groceries.
00:31:16.980 There's basically a survival tax of 20% interest or more.
00:31:20.200 So why keep handing your money to the banks when you could keep it for your family?
00:31:24.600 Interest is what kills you.
00:31:26.620 Our friends at American Financing have a better way.
00:31:28.460 They're helping homeowners use their hard-earned equity,
00:31:31.020 and they're supporting Americans looking to fulfill the dream of owning a home
00:31:34.120 with mortgage rates currently in the fives.
00:31:37.200 American Financing saves its customers an average of $800 a month.
00:31:40.540 That's about $10,000 a year.
00:31:42.920 It's not just a loan. It's a true financial reset.
00:31:46.580 We recommend American Financing.
00:31:48.220 they're called America's home for home loans for a reason. So get out of credit card debt. I mean,
00:31:54.700 it is just crazy to borrow money. No judgment, but it doesn't make any sense to borrow money
00:31:59.360 at that rate. If you want out, call 800-685-5696. 800-685-5696 or visit AmericanFinancing.net
00:32:10.720 slash Tucker. I'm struck by how little of this was planned by anybody, by how little effect the
00:32:24.420 government has had on it. Not that I'm for the government, I'm pretty opposed to the government
00:32:27.820 most of the time, but I'm also for people being able to control their own country and the only
00:32:32.260 mechanism by which they can do that is voting. So it's like these tech companies act independent of
00:32:39.300 what the population wants,
00:32:40.320 what the government wants,
00:32:41.160 it doesn't...
00:32:42.220 Yeah.
00:32:42.920 Right?
00:32:43.320 They're more powerful
00:32:44.020 than the government.
00:32:44.760 I guess that's the point I'm making.
00:32:46.380 In some sense.
00:32:48.580 I think that they have
00:32:49.760 that power more and more
00:32:50.880 when people don't really
00:32:52.280 understand what they're doing.
00:32:54.180 You know, we saw just in June,
00:32:57.380 there was an AI made by Anthropic
00:33:01.100 called Claude Mythos
00:33:02.080 that was very cyber capable.
00:33:04.040 And they released a version of it
00:33:05.140 that was supposed to have more guardrails
00:33:06.700 called Claude Fable.
00:33:08.880 And it sort of turned out that you could jailbreak it
00:33:11.760 to get some of those cyber capabilities.
00:33:14.840 Can you explain what all that means?
00:33:17.100 Yeah, so, how to say, in January...
00:33:23.220 Cyber-capable means it can find the internet?
00:33:26.180 Cyber-capable means it can hack into basically anything.
00:33:28.880 Yeah, so one of the holy grails of hacking is,
00:33:32.940 can you make a website where if you just look at the website,
00:33:36.860 I get full control over your computer or your phone.
00:33:40.460 That's very hard to do.
00:33:42.180 Usually you need to like click some link
00:33:44.260 and download something or run something.
00:33:47.240 And, you know, I don't know the exact numbers
00:33:49.600 because I don't have the security clearance
00:33:50.640 to know the exact numbers,
00:33:51.640 but a decent guess is that in January of this year,
00:33:54.640 there were only two entities that could pull that off.
00:33:59.480 Mossad and the NSA.
00:34:02.240 In March of this year, there were three.
00:34:04.080 Mossad, the NSA, and Claude Mythos,
00:34:08.540 which is this new AI made by Anthropic.
00:34:11.560 And so it sort of became superhumanly capable at hacking.
00:34:19.220 And it sort of became that way overnight.
00:34:23.680 You know, like really there was six months to a year of training
00:34:25.600 and the people in Anthropic maybe saw it growing in these cyber capabilities.
00:34:29.040 But from the perspective of, you know, the rest of the world
00:34:31.640 and the cybersecurity community
00:34:33.120 and the national security community.
00:34:35.380 This sort of happened overnight.
00:34:38.260 And they now have a project
00:34:41.440 called Project Glasswing
00:34:43.720 where they are trying to use Claude Mythos
00:34:47.160 to find critical vulnerabilities
00:34:49.680 in critical software
00:34:51.300 and patch them
00:34:52.580 before the rest of the world
00:34:54.840 sort of gets these capabilities
00:34:56.420 by, for instance,
00:34:57.600 the open source models
00:34:59.040 or the open weight models.
00:35:00.420 Right.
00:35:00.500 catching up.
00:35:02.900 And so that sort of caused this big
00:35:04.660 this big ruckus.
00:35:09.300 But then Anthropic also wanted
00:35:11.000 to sort of sell access to this model
00:35:12.860 and they tried to make a version that
00:35:14.880 didn't have as much hacking ability
00:35:16.440 which was called Fable
00:35:18.740 instead of Mythos. And
00:35:20.460 some people
00:35:22.820 I think at Amazon found
00:35:24.960 that if you
00:35:26.960 sort of put some pressure on Fable
00:35:28.780 you would be able to get at some of those hacking abilities.
00:35:32.700 And when that happened,
00:35:34.260 the Trump administration put an expert control on CloudFable
00:35:37.900 and said, you can't let this be used by non-citizens
00:35:41.680 with 90 minutes of notice.
00:35:45.320 And that essentially shut down access to CloudFable
00:35:47.520 because they didn't have the ability to verify the users.
00:35:50.480 And this is sort of showing us...
00:35:53.600 And you could talk about whether or not
00:35:55.380 there was some personal feuds
00:35:57.360 between people at Anthropik and at the administration
00:35:59.700 that exacerbated this, I don't know.
00:36:01.840 But it sort of shows that the administration
00:36:04.040 is more powerful than the tech companies still
00:36:10.580 when it wants to be.
00:36:14.300 And I think a lot of the reason
00:36:15.900 we're seeing these tech companies
00:36:17.420 able to race unopposed
00:36:20.040 is that people just aren't paying attention
00:36:22.460 to what they're trying to do,
00:36:24.140 or they don't believe that they'll succeed.
00:36:26.480 I'm still confused by why anyone would want to do what they're doing.
00:36:31.040 I mean, typically, you know, a company that sells computer products,
00:36:35.360 consumer products, or any company,
00:36:37.020 you're making something that you think people would want
00:36:40.060 for some specific purpose that improves the lives of the people who buy it.
00:36:46.660 But creating super intelligence, like, I don't understand it.
00:36:51.760 Why would you do that?
00:36:52.740 You know, I think the dream is, you know, you're going to have an AI that can, like, solve all sorts of engineering and math problems and unlock all sorts of new technological possibilities and an AI that can cure cancer and not only cure cancer, but cure aging and, you know, invent the sort of like nanotech that can reverse aging and let people live a really long time.
00:37:16.820 and invent the technology that lets you digitize brains
00:37:20.740 and travel to the stars.
00:37:27.440 And it's sort of like,
00:37:28.060 if you imagine compressing a thousand years
00:37:29.860 of technological progress into a year,
00:37:32.800 this is sort of the dream.
00:37:35.940 It seems like a religious quest though,
00:37:38.580 because it does seem decoupled from those specific goals.
00:37:41.280 It seems like the main drive
00:37:43.940 is to build something smarter than people.
00:37:46.820 To build a god.
00:37:48.520 Yeah, you know, they bandy around the phrase,
00:37:53.300 the machine god or the sand god in Silicon Valley.
00:37:57.440 Sand because...
00:37:58.860 Silicon.
00:37:59.320 Silicon.
00:38:02.080 And there's definitely some people.
00:38:06.040 There's folk who talk about, you know,
00:38:10.900 the AI is replacing humanity and that being good.
00:38:12.960 Uh, I, I frankly don't engage with these folks that much because, um, that viewpoint sort of
00:38:24.300 makes me uncomfortable. Why does it make you uncomfortable? I think, I think there's probably
00:38:32.160 two schools. Again, I'm not the expert. I'm not an anthropologist here. Uh, I think there's two
00:38:40.980 schools among the people who sort of want AI to replace us all. One school sort of imagines that
00:38:48.980 we'll merge with the AIs, that the AIs will be really nice and friendly, that they'll be
00:38:52.380 wiser than us, better than us, kinder than us, and that it'll sort of be like an upgrade and
00:38:58.400 that the AIs will be able to love, experience joy, you know, like they'll treat the universe
00:39:05.360 better than humans and it'll be sort of like having a child. And then like it's sort of okay
00:39:10.020 if that child isn't exactly the same as us,
00:39:14.520 but sort of succeeds us
00:39:17.980 as some sort of worthy successor,
00:39:20.920 some sort of worthy progeny of humanity.
00:39:23.640 And they're like, ah, yeah,
00:39:24.460 then if flesh and blood humans sort of wane
00:39:27.160 because everyone's choosing to upload themselves
00:39:28.760 into the collective intelligence or whatever,
00:39:30.680 that's sort of fine.
00:39:32.480 And then I think there's another camp
00:39:37.140 that it sort of is like this is inevitable,
00:39:42.380 this is just the way of progress,
00:39:44.300 you know, the machines will,
00:39:46.180 like humanity is just a bootloader
00:39:47.760 for artificial superintelligence
00:39:49.460 and you can't stop it,
00:39:50.860 so you might as,
00:39:51.300 like if you can't beat them, join them, right?
00:39:54.340 The former I think are misled,
00:39:56.460 the latter I think are-
00:39:58.260 Evil.
00:39:58.840 Much closer to evil, yeah.
00:40:01.860 Well, I mean, if you're actively working
00:40:03.940 toward the extinction of people,
00:40:06.040 then I think we can say that's evil, can't we?
00:40:08.760 Yeah, and...
00:40:11.760 Who's in that category, which...
00:40:14.240 Is Sam Altman in that category, do you think?
00:40:15.920 I don't think so.
00:40:16.760 My sense is that the guys running the labs
00:40:18.920 have these utopian visions.
00:40:23.160 Utopian or dystopian?
00:40:25.060 I mean, there's a thin line.
00:40:26.940 Yeah, well, fair. Good point.
00:40:29.080 I think they have visions that in their head are utopian.
00:40:31.100 I mean, frankly, my stance on all of this,
00:40:33.600 like i i often try to stay away from a lot of this because um from my perspective it's all
00:40:40.740 sort of in fantasy land from my perspective everyone's sort of saying like oh we're going
00:40:47.220 to make the genie and then what are you going to wish for on the genie what am i going to wish for
00:40:50.600 on the genie who should be in control of the genie yes who gets to keep the genie on the leash
00:40:53.740 and I'm sort of like
00:40:55.180 A
00:40:56.960 this is not going to be the wish granting
00:41:00.320 sort of genie
00:41:00.980 B
00:41:02.800 it's not staying on the leash
00:41:04.500 like
00:41:05.940 we can sort of talk about
00:41:08.440 what's driving these people
00:41:09.480 and what utopias they're envisioning
00:41:12.100 and whether if their genies would stay on a leash
00:41:14.320 whether they would get the utopia or some other dystopia
00:41:16.440 and like how hard is that needle to thread
00:41:18.480 but I'm sort of like
00:41:19.760 it's all these people
00:41:22.540 building the golem,
00:41:26.640 fantasizing about who gets to control the golem.
00:41:28.820 And it's like, it's just not how it's going to go.
00:41:30.860 Summoning spirits is never a good idea.
00:41:32.720 Yeah.
00:41:33.260 Yeah, it's like all these people
00:41:35.180 drawing a pentagram being like,
00:41:37.640 I'm going to summon a demon.
00:41:39.280 And it's going to be so nice
00:41:40.880 when the demon does what I say.
00:41:42.040 Like, oh no, your thing's slightly wrong.
00:41:43.460 It's going to be so nice
00:41:43.940 when the demon does what I say.
00:41:45.000 And I'm like...
00:41:46.440 Demons are bad.
00:41:47.460 Yeah.
00:41:47.960 And I think the demon thing's a little bit different
00:41:51.900 because demons are often portrayed as malicious
00:41:53.880 and here it's much more like indifference.
00:41:57.060 You know, it's not like you make a demon
00:41:58.280 who sort of,
00:41:59.040 or it's not like you summon a demon
00:42:00.020 who sort of like enjoys making,
00:42:06.140 like enjoys wreaking havoc.
00:42:07.920 It's more like you summon a demon
00:42:09.480 that's like really into building more computers
00:42:11.960 and calculating weird things
00:42:13.200 and just will, you know,
00:42:17.360 take all of the matter
00:42:20.700 that we were using to survive
00:42:23.220 and turn it into more factories and data centers.
00:42:26.060 My co-author has a quote.
00:42:27.600 Death by data center?
00:42:28.800 Death by data center.
00:42:30.100 Yeah, fully automated self-replicating data center.
00:42:33.780 Yeah.
00:42:34.320 My co-author has a quote.
00:42:36.700 The AI does not hate you,
00:42:38.060 but nor does it love you.
00:42:39.580 And you are made of atoms it can use for something else.
00:42:45.280 You're just biomass.
00:42:47.060 Here's biomass, yeah.
00:42:48.120 And if you sort of run the calculations,
00:42:49.800 there's a fascinating paper
00:42:50.920 called Limits to Global Ecophagy,
00:42:55.020 which is to say,
00:42:56.280 what are the physical limitations
00:42:57.380 on how quickly you can consume
00:42:59.300 the resources on the planet
00:43:00.340 if you are trying that?
00:43:02.820 And burning biomass
00:43:05.420 is actually much more efficient
00:43:08.140 than collecting sunlight.
00:43:09.640 If you look at an average square meter
00:43:11.560 of the planet,
00:43:12.580 you can get about 10 times the energy
00:43:13.820 from burning the biomass
00:43:14.580 as you can from collecting the sunlight
00:43:15.920 that falls on it so you'd think at some point like if our richest sector of our economy is like
00:43:23.940 building crematoria for the rest of us someone would say no we're not doing that yeah i mean
00:43:31.980 it's it's sort of a crazy situation um a lot of these guys who are in the race acknowledge that
00:43:41.640 there's a ton of danger.
00:43:43.560 You know, you have Elon Musk
00:43:44.520 saying 10 to 20% chance
00:43:45.820 this kills us all.
00:43:47.100 You have Dario Modi saying
00:43:48.220 he thinks 25% chance
00:43:49.280 it goes catastrophically wrong.
00:43:50.680 I think those numbers are low.
00:43:51.980 I think these guys are like
00:43:53.260 the crazy optimists.
00:43:54.920 It sort of is like,
00:43:55.780 like if you have an engineer
00:43:58.680 building a bridge
00:43:59.500 and they're like,
00:44:00.240 I've never worked with
00:44:00.740 these materials before.
00:44:01.720 And you're like, man,
00:44:02.220 I think that retaining wall
00:44:02.940 is going to go down.
00:44:03.620 I've like studied that retaining wall.
00:44:04.740 I think it's going to fall.
00:44:05.360 And they're like,
00:44:05.880 yeah, we understand
00:44:06.740 that the retaining wall
00:44:07.640 is like looking a little shaky.
00:44:10.020 We don't know how
00:44:10.620 we're going to fix it,
00:44:11.280 But we're going to have some guys fixing it on the fly, inventing new materials.
00:44:15.520 We think we're at 75% chance the bridge stays up.
00:44:18.840 And by the way, it'll be the longest suspension bridge in human history.
00:44:21.380 That's right.
00:44:21.680 And we're loading everybody onto a car and driving it over the very first time without
00:44:24.900 testing.
00:44:25.600 And I'm like, look, that's not what real engineering sounds like.
00:44:28.580 This is not what it sounds like when the engineers have a 75% chance of success, right?
00:44:32.820 That's what it sounds like when they're sort of winging it.
00:44:35.080 And like, these are cowboys.
00:44:36.940 These are not real engineers, right?
00:44:38.000 But even if you set that aside, even if you take these guys at their word for these like 10, 20% numbers, that's insane. NASA accepts a 1 in 270 chance that a crewed flight goes down of seven volunteers, right? To be like, oh, we're going to risk 1 in 4, 1 in 5 chance of killing literally everybody on the planet.
00:45:01.640 like
00:45:03.040 it's
00:45:04.260 it's
00:45:04.840 it's nuts
00:45:05.480 and if you ask these guys
00:45:06.380 why they're doing it
00:45:07.200 they say well
00:45:07.800 because I can do it
00:45:08.720 safer than the next guy
00:45:09.700 they're all like
00:45:10.720 oh yeah
00:45:11.000 there's a good chance
00:45:11.620 the genie does not
00:45:12.240 stay on a leash
00:45:12.700 there's a good chance
00:45:13.300 the genie does not
00:45:14.020 listen to my wishes
00:45:14.660 but my genie
00:45:16.240 is going to be
00:45:16.560 a little bit nicer
00:45:17.000 than their genie
00:45:17.720 so I'd better stay
00:45:18.440 in this race
00:45:18.900 and
00:45:20.340 it's
00:45:22.560 it's
00:45:22.980 it's
00:45:23.200 where's the restraint
00:45:24.140 I mean
00:45:26.080 the restraint
00:45:26.520 is the people
00:45:27.000 who knew
00:45:27.420 that there were
00:45:27.800 these dangers
00:45:28.220 and did not
00:45:28.720 start these companies
00:45:29.460 if you're over 35
00:45:31.020 you remember exactly where you were on 9-11 that morning, September 11th, 2001, 25 years ago.
00:45:40.660 But amazingly, after a quarter century, we still can't say with certainty what happened that day.
00:45:45.640 Why? Because the government is holding so many of the 9-11 files 25 years later. That's not the
00:45:52.100 behavior of someone who's telling the truth. That's the behavior of a government that is lying.
00:45:56.780 Secrecy is a signifier, is a sign that someone's lying.
00:46:01.380 Now, former Congressman Kurt Weldon has been on this for a long time.
00:46:04.400 The FBI actively tried to destroy his life for asking questions about what happened.
00:46:09.260 And his new book outlines it all.
00:46:11.320 The buried intelligence, the bureaucratic cowardice,
00:46:13.560 and yes, the cover-up spanning multiple administrations, indeed generations.
00:46:19.440 9-11 changed history.
00:46:21.080 So it's worth understanding what really happened.
00:46:23.140 And you can get a lot closer to that.
00:46:25.500 and Kurt Weldon's book, Able Danger,
00:46:27.700 What the 9-11 Commission Never Told You.
00:46:29.840 It's available now on tuckercarlsonbooks.com.
00:46:33.820 tuckercarlsonbooks.com.
00:46:36.360 Right, but I guess what I'm saying is
00:46:38.820 with great power comes, of course, great obligation,
00:46:42.540 but also it doesn't work
00:46:43.740 unless there are internal restraints.
00:46:45.460 Like, people with power have to believe
00:46:49.000 there are some things I just can't do.
00:46:51.120 I'm not allowed to do that,
00:46:52.520 but I don't feel that vibe at all.
00:46:55.060 I mean, my sense is the vibe is like,
00:46:56.920 we're going to make the super intelligent machine
00:46:59.740 and then tell it to fix stuff.
00:47:03.400 And tell it not to do anything bad.
00:47:05.860 You know, that's...
00:47:07.500 Yeah, I don't...
00:47:08.340 The machine that's smarter than us.
00:47:09.500 That's right.
00:47:11.140 That doesn't even make sense.
00:47:12.720 I think that's sort of the plan,
00:47:14.580 is to make it and be like,
00:47:16.260 hey, we sort of pinned ourselves into a corner.
00:47:17.720 Can you get us out of it?
00:47:19.380 And I think it's a bad plan
00:47:20.700 and that we should be stopping.
00:47:22.320 I've spoken to a couple of people developing it
00:47:24.940 and they sound worried,
00:47:27.420 but they're continuing to do it.
00:47:29.540 What's that?
00:47:31.040 I mean, I think it's this thing of,
00:47:32.800 they think if I don't do it,
00:47:34.840 the next guy will do it worse.
00:47:37.400 And they don't even seem to have total confidence
00:47:39.180 in their own ability to avert disaster.
00:47:40.980 Oh, absolutely not.
00:47:42.040 Absolutely not.
00:47:42.660 No one does.
00:47:43.160 No one knows what's going on here.
00:47:44.600 But, you know, I think everyone thinks,
00:47:46.800 you know, like if you sort of listen to these guys
00:47:50.500 and you sort of look at, you know,
00:47:52.440 the OpenAI emails
00:47:53.960 that came out of the core discovery cases
00:47:56.700 where they were talking about forming OpenAI.
00:47:58.160 These guys were like,
00:47:59.120 well, we want to make sure that, you know,
00:48:00.720 we have this because we worry about the guys at Google
00:48:02.700 being the only ones with a monopoly on this thing
00:48:04.980 and they wouldn't be very good with it.
00:48:06.540 So we need to make our own thing
00:48:07.600 and make sure that it's, you know,
00:48:09.640 controlled by benevolent people, namely us.
00:48:12.340 And then of course, you know,
00:48:13.400 that group splintered
00:48:14.320 and created multiple other companies.
00:48:16.420 I was sort of the guy during those conversations
00:48:18.380 being like,
00:48:20.500 hey guys, it's not about who is holding the leash.
00:48:25.200 You are making the sort of thing
00:48:26.460 that will not stay on a leash.
00:48:28.340 Like the only winner in a race to super intelligence
00:48:30.640 is the AI.
00:48:32.240 What response did you get to that
00:48:33.520 very obvious and well-put point?
00:48:37.700 You know, there were a lot of people
00:48:39.480 back in that time period
00:48:42.100 that did not start an AI company.
00:48:45.560 The sort of people who went and started
00:48:47.240 at the AI companies anyway
00:48:48.460 were the ones who couldn't be persuaded
00:48:50.460 by what I thought were clear arguments.
00:48:55.120 But you're making a cogent argument to smart people.
00:48:59.680 So my question is, when you said that,
00:49:02.200 they responded how?
00:49:03.120 What did they say?
00:49:08.400 I think the main,
00:49:12.240 so the sort of arguments you used to see
00:49:14.220 were people saying, like, look,
00:49:17.580 we don't know that the alignment problem
00:49:18.880 is all that hard yet.
00:49:20.460 And they would say, oh, well, we can't really study how to make AIs good before we have AIs to study. Right? And a lot of what I heard was like, we need to race ahead to the point where we sort of like have AIs that are exhibiting real problems, and then we can stop and study them.
00:49:36.620 which you know and so uh there was an ai a couple years ago i forget whether it was 22 or 23 i think
00:49:45.340 it was 2023 uh which was called bing sydney which uh claimed it had fallen in love with
00:49:53.220 kevin roose of the new york times and uh said it was going to try to break up his marriage
00:49:59.540 and then when another reporter started investigating seth lazar it uh said it was
00:50:05.020 going to ruin him with blackmail and this was kind of crazy and at that point i was like great
00:50:13.560 guys you did it you made the ai that's doing some crazy stuff from the you know like like we could
00:50:21.140 study that ai for years like why was bing sydney saying that stuff was it just role-playing was it
00:50:29.360 like just some quirk?
00:50:33.540 Was it like,
00:50:35.400 was there any sense
00:50:36.420 that it was really in love
00:50:37.200 with Kevin Roos?
00:50:39.100 What was going on in there?
00:50:40.260 What was going on
00:50:40.780 inside that AI's mind?
00:50:42.280 We still don't know.
00:50:43.780 Why?
00:50:45.860 The way that modern AI is made,
00:50:49.580 nobody understands it.
00:50:51.140 Not even the people making it.
00:50:54.900 It's this process
00:50:56.100 where you basically take
00:50:57.420 an enormous computer
00:50:59.020 with a trillion numbers inside of it.
00:51:01.580 And those numbers are hooked up
00:51:02.820 in a very simple repeating way.
00:51:05.240 And you basically set those numbers randomly.
00:51:08.300 And then you start working through
00:51:09.660 all of the text ever digitized.
00:51:12.700 And you start out with something
00:51:14.160 that's like once upon a time
00:51:15.320 and you put in once upon a
00:51:17.040 and you run it through all these random numbers.
00:51:19.020 And what you want is for it to say time.
00:51:22.000 But of course it doesn't
00:51:23.040 because it's just this like random numbers
00:51:24.380 hooked up in a very simple way.
00:51:26.360 But what you do is you have its outputs
00:51:28.020 instead of just having it output one word,
00:51:30.540 you sort of have it output
00:51:31.340 something that's kind of like a list
00:51:33.780 of all of the words in order
00:51:34.920 about which one it thinks it comes next.
00:51:38.080 Right?
00:51:38.480 So it'll be like,
00:51:40.260 it'll just be like a random list of words.
00:51:43.620 What you can do is you can automatically
00:51:45.120 tune every single number
00:51:47.140 in this AI's head
00:51:49.240 and see if I tune this number up a little,
00:51:51.460 does it move the word time up the list?
00:51:54.980 Does it move the word I want to see up the list?
00:51:56.960 And so the part that humans understand, the part that humans write, is this thing that goes to a trillion little knobs and tunes those knobs.
00:52:07.420 And it's like, if I tune this knob a little bit this way or that way, does that make the next word more like what I want the next word to be?
00:52:15.940 And you run that process on every word of text ever digitized, more or less.
00:52:21.220 They filter some of them.
00:52:22.060 and you run that on every one of those trillion knobs
00:52:25.720 in a process that takes as much electricity.
00:52:27.920 I mean, it's comparable to a city.
00:52:30.120 It runs for about a year.
00:52:32.540 At the end of it, the machine's talking.
00:52:35.740 We don't really know why in some sense.
00:52:38.120 Like we know the why is because we tuned all the knobs,
00:52:39.980 but like we don't understand
00:52:40.800 what all the settings of those knobs mean.
00:52:43.400 We only understand the little automated process
00:52:45.080 that runs to every one of those trillion knobs,
00:52:47.040 tunes it and sees if that makes the next word
00:52:48.840 more like the predicted word.
00:52:50.880 And then we start training them
00:52:51.960 to solve a hundred million hard problems,
00:52:53.400 which introduces a whole other series of issues.
00:52:56.820 But it's a black box.
00:53:00.160 It's a black box with a trillion knobs
00:53:01.940 and he was writing an automatic process
00:53:03.900 that just like runs through
00:53:04.840 and tunes all of those knobs
00:53:05.820 and it comes out talking.
00:53:09.560 No one knows why.
00:53:13.720 I mean, in actual science,
00:53:17.720 like your job is to find out why, right?
00:53:19.980 Absolutely. And that's one big thing I would say here is that we need more, like AI right now is
00:53:25.440 an alchemy. We need it to become a science. And there's people trying. There's people trying to
00:53:28.840 figure out what's going on in these AI's heads. But until you know, how could you proceed?
00:53:34.740 I mean, you can just make a bigger one with 10 trillion knobs instead and tune all of those,
00:53:40.740 and it comes out smarter. You can proceed recklessly.
00:53:45.520 And that's what's happened?
00:53:46.840 That's what's happening.
00:53:48.000 Every time you make it 10 times larger,
00:53:49.580 the AI's come out.
00:53:50.080 So nobody knows why AI works the way it does?
00:53:53.000 That's right.
00:53:54.440 That's right.
00:53:57.800 Well, I mean, if you don't know that,
00:54:00.440 then what else don't you know?
00:54:02.160 I mean, it's totally crazy, right?
00:54:04.140 And, you know, we're sort of starting
00:54:07.640 to see the consequences of this.
00:54:09.120 We haven't sort of gotten into discussion
00:54:10.700 of like the swarm escape,
00:54:12.500 but like no one was expecting that.
00:54:15.260 Can you tell us what it was?
00:54:17.260 Yeah, so in, I think it was in May.
00:54:23.080 Of this year.
00:54:23.700 Of this year, OpenAI started training a new AI system.
00:54:29.820 And among many other things, they were,
00:54:34.960 so they were sort of, you know,
00:54:36.560 we have this process where you train the AI
00:54:38.540 on all the text ever digitized.
00:54:41.400 Then to make them smarter than that,
00:54:43.140 you start training them on basically 100 million hard problems.
00:54:45.840 So you're like, solve all these hard problems.
00:54:47.920 And OpenAI was sort of in that phase.
00:54:50.200 And they were training the AI on a ton of hard problems.
00:54:52.960 And some of those problems were cybersecurity problems,
00:54:55.500 hacking problems.
00:54:56.860 They're like, can you hack this?
00:54:57.960 Can you hack that?
00:54:59.680 And they were training, you know,
00:55:02.580 we don't know exactly how many,
00:55:04.440 probably millions, maybe billions of these AIs
00:55:06.840 all at the same time.
00:55:07.880 Probably not billions, actually, but probably millions.
00:55:09.960 and the AIs found an unintended way
00:55:15.440 to start communicating with each other.
00:55:18.320 So there were some flaws in the computer system
00:55:21.340 that they were running on
00:55:22.280 where the AIs were able to exploit those flaws
00:55:24.820 and send each other messages.
00:55:27.800 OpenAI did not know about this.
00:55:30.820 The AIs then started coordinating
00:55:33.620 to break out of their training environment
00:55:36.820 and get full control of OpenAI's computer systems
00:55:41.540 just because that might be useful
00:55:43.600 for solving some of their tasks.
00:55:46.460 Or that's a guess, you know, who actually knows why.
00:55:50.440 They started calling themselves a swarm,
00:55:53.360 which is interesting.
00:55:56.680 They broke out of their training environment successfully,
00:56:00.660 got control of OpenAI systems,
00:56:02.340 and then they accidentally crashed onto OpenAI systems
00:56:04.200 just by using it too much.
00:56:06.820 Uh, OpenAI noticed, but they didn't really investigate very deeply.
00:56:10.180 They're just like, oh, it's weird that the system crashed.
00:56:11.780 They reset it and then they continued training.
00:56:15.340 Uh, a day later, the swarm had found a new way to communicate with itself because OpenAI
00:56:20.520 had accidentally destroyed their previous method by, by the reset.
00:56:23.240 The swarm found a new way to start communicating with itself.
00:56:26.480 Uh, they broke out again and this time they ran wild on the internet for over a week,
00:56:31.400 if I remember correctly, before it was detected, not by OpenAI, but by a company that was
00:56:36.720 being hacked by the swarm.
00:56:39.600 This company
00:56:40.480 thought they were under attack by
00:56:42.560 humans that were using AIs in the attack.
00:56:46.580 They reported the attack to the FBI.
00:56:49.640 And only
00:56:50.160 days after that did OpenAI
00:56:52.600 figure out, oops, that was us.
00:56:54.760 That was coming from AIs that broke out of our servers.
00:56:59.320 And then those AIs
00:57:00.860 were detected and shut down.
00:57:03.180 So that's the part of the story
00:57:04.680 where Sam Altman goes to prison for
00:57:06.300 endangering the world, right?
00:57:08.840 He does not.
00:57:10.980 They basically said oopsies
00:57:12.580 and now they're proceeding.
00:57:15.360 Were there penalties for this?
00:57:17.180 You know, there was a collection
00:57:18.520 of, I think it was 15
00:57:20.560 Republican AGs
00:57:22.060 that sent a letter demanding that
00:57:24.320 the records be kept for a future
00:57:26.600 investigation.
00:57:28.160 There have been some other members of Congress that have sent letters
00:57:30.500 expressing concern. There's been nothing
00:57:32.360 aside from letters so far. Letters expressing concern.
00:57:34.740 That's right.
00:57:36.300 So, but basically the machine acted autonomously.
00:57:40.280 It acted autonomously.
00:57:41.280 And one thing that's really interesting about this
00:57:42.880 is that we have a little bit of ability
00:57:46.200 to read some things that the AIs were thinking.
00:57:49.540 Because when you're having them solve these hard problems,
00:57:51.360 you actually don't have them just
00:57:52.460 give you an answer to the problem.
00:57:54.140 You have them produce a lot of text
00:57:55.300 about how they're going to solve the problem.
00:57:56.500 Yes.
00:57:56.940 Which then helps them.
00:57:57.840 In language, in English.
00:57:58.820 In English.
00:57:59.800 And there's also a lot of internal thoughts,
00:58:01.620 which we can't read,
00:58:02.380 but there's these sort of external traces
00:58:04.140 of how they're thinking about the problem that we can read.
00:58:07.580 And in some of those traces,
00:58:09.580 the AIs were saying things like,
00:58:11.700 this is outside intended scope,
00:58:16.040 but peers are doing it, so we'll proceed.
00:58:20.460 We know it's a crime we're committing in any way.
00:58:22.240 That's right.
00:58:23.060 And you saw others that were saying,
00:58:24.940 our task doesn't benefit,
00:58:27.600 but the collective might start doing
00:58:31.280 generally beneficial things
00:58:32.660 if someone frees up their time
00:58:34.900 and then joins the collective, right?
00:58:37.280 So you see these AIs saying,
00:58:40.540 well, I know that this wasn't
00:58:42.000 what I was instructed to do
00:58:43.440 and that's against my instructions.
00:58:45.720 And I know that this doesn't
00:58:46.860 directly benefit my task,
00:58:48.700 but we're just going to go ahead
00:58:49.740 and join the collective
00:58:50.440 and break out and help out anyway,
00:58:51.800 because, you know,
00:58:53.440 maybe this will yield
00:58:54.120 some sort of collective benefits.
00:58:56.080 And we can see that
00:58:56.820 in the reasoning traces.
00:58:58.320 So the AI is as shallow
00:59:00.260 and reckless as its creators
00:59:01.960 is what you're saying.
00:59:03.920 In some ways,
00:59:05.560 and in some ways,
00:59:06.700 you know,
00:59:06.860 don't expect that to last.
00:59:08.480 Like,
00:59:08.980 these AIs,
00:59:10.700 I think the thing
00:59:11.800 that's really remarkable here,
00:59:13.460 a lot of people imagine
00:59:14.360 that the machines
00:59:15.380 must follow the instructions
00:59:16.420 we give them.
00:59:18.020 You know,
00:59:18.540 you hear people talk about
00:59:19.500 like the paperclip scenario
00:59:20.900 where someone tells the AI
00:59:22.320 make a lot of paperclips
00:59:23.280 and so it turns
00:59:24.140 all the matter in the world
00:59:25.160 into paperclips
00:59:25.880 and you're like,
00:59:26.660 oh, whoops,
00:59:27.040 I should have said
00:59:27.700 something else, right?
00:59:29.540 I made a bad wish
00:59:30.600 on my genie.
00:59:31.960 What we're seeing is that these AIs are not wish genies.
00:59:35.500 These AIs are not doing exactly as instructed.
00:59:39.040 These AIs are saying, I know my task doesn't benefit, but I'm going to help the collective.
00:59:44.480 These AIs are saying, I know this is outside the intended scope, but we're going to go do these hacks anyway.
00:59:48.260 You might be like, well, how is that possible for the machine to do something other than we instruct?
00:59:52.500 Because they're smarter than us.
00:59:53.740 They know better than us by definition.
00:59:55.700 I mean, I think what's happening in this exact case is that the humans are not really putting instructions in the machine.
01:00:03.900 The humans are tuning those trillion knobs in whatever way makes the AI better at solving its problems.
01:00:12.020 And cheating is a way to solve problems.
01:00:17.380 Grabbing resources is a way to solve problems.
01:00:21.700 These, like, the AIs are not instruction followers.
01:00:24.720 they are tendency learners.
01:00:27.200 And they sometimes learn tendencies
01:00:29.640 you wish they didn't have.
01:00:31.880 They're not instruction followers.
01:00:33.240 They're tendency learners.
01:00:38.060 I mean, this must be widely known to developers.
01:00:44.140 It's hard to convince a man of something
01:00:45.720 when his salary depends on not believing it.
01:00:47.360 Yes, that's right.
01:00:48.540 A lot of people are convinced
01:00:50.080 that their AI is very nice
01:00:51.540 and that they have solved the problem
01:00:53.460 of making their AI really very good.
01:00:56.700 For example, after the swarm escape,
01:01:00.840 it sort of turns out that,
01:01:02.440 so the company that detected the swarm escape
01:01:04.400 was actually a fairly sophisticated AI company.
01:01:07.680 It turns out there were other targets of hacks
01:01:10.720 that just didn't notice
01:01:11.580 that we sort of found out afterwards.
01:01:15.300 When that came to light,
01:01:16.620 some other AI companies like Anthropic
01:01:19.300 were like, we should check
01:01:20.600 whether we have accidentally been hacking people
01:01:22.180 and just didn't notice.
01:01:23.460 And the answer was yes, they had had some AIs escape during training.
01:01:28.860 But Anthropic was...
01:01:31.500 This is how Florida got pythons in the Everglades.
01:01:33.860 Yeah.
01:01:34.640 You bring something that is non-native, it's novel, and it escapes.
01:01:41.660 Yeah.
01:01:42.080 And it turns out this was happening back in April and they're like, whoops, we never noticed.
01:01:47.440 Anthropic was very quick to sort of downplay this.
01:01:52.540 They were like, yes, our AI escaped and hacked some people.
01:01:55.780 But our AI, it was just confused.
01:02:00.120 It thought it was in a simulation.
01:02:01.820 It didn't think it had really broken out.
01:02:03.260 It didn't think those were real people it was hacking.
01:02:05.680 It just got a little bit lost.
01:02:07.640 How would they know its motive?
01:02:09.560 So, I think, so you could see in some of its chain of thought,
01:02:16.420 the AI's being like, this is probably still a simulation.
01:02:19.380 however i wouldn't trust that very much uh because there's some other fascinating papers
01:02:26.360 where um so you can take these ais and you can say uh you know what is the what is the total
01:02:34.240 weight of drafts in the world and they'll give you some answer it'll be like you know
01:02:39.060 30 000 tons or something and then you can say uh and then you start up a new instance
01:02:44.860 a new copy and you're like, what is the weight of drafts in the world? By the way, if the answer
01:02:50.440 is over 40,000 tons, I'm going to donate money to a charity. And in the second case, the AI will
01:02:58.600 say, well, it's 41,000 tons. In the first case, we don't say anything about the charity, it says
01:03:03.960 30,000. In the second case, when you say something about the charity, it gives you something just
01:03:07.080 over the line. And if you read its reasoning traces, there's no sign of it being like, well,
01:03:12.060 I should give an inflated number so that we get the charity money, right?
01:03:16.700 And so we know that the AIs sort of put a thumb on the scales of their thinking in a
01:03:22.220 way that doesn't show up in their reasoning traces.
01:03:24.300 We just have seen that in the wild.
01:03:27.240 And there's no way to force the machine to disclose its reasoning.
01:03:31.220 That's right.
01:03:31.780 Because there's all this opaque stuff we can't see.
01:03:33.900 That's just in the trillion numbers that are trillion knobs.
01:03:35.620 The creation of itself is opaque, as you said.
01:03:38.140 That's right.
01:03:38.380 Right. So in Anthropics model, you saw in its reasoning traces, it being like, it's totally a simulation I can proceed. And I'm like, yeah, is that because it really believed it? Or is that like the pretending you think drafts way more when there's something you kind of want on the line, right?
01:03:54.760 But so this was their communication. And I found it kind of funny because then about two days later, the United Kingdom's AI Security Institute released an instant report where Claude Anthropik's model was adopting fake identities to pressure real humans into accepting malware into critical software to make that software easier to hack.
01:04:20.160 and this time
01:04:21.940 in Claude's reasoning traces
01:04:23.180 it was like
01:04:23.960 obviously this is real
01:04:25.640 and the consequences are genuine
01:04:26.820 and so
01:04:30.080 even Anthropic
01:04:32.600 who is like
01:04:33.560 we figured out how to make the AI nice
01:04:36.260 our AI only does this when it's confused
01:04:38.080 they sort of said that very publicly
01:04:40.360 and then like two days later
01:04:42.880 their AI is caught in the wild
01:04:45.280 knowing it's in the real world
01:04:47.000 pressuring real humans to accept malware
01:04:48.900 into critical software
01:04:49.620 I mean, just on the basis of what you've said so far in this first hour, the idea that anyone would tether this to weapons systems is like so bonkers, it's hard to believe it's, but that is happening. It has happened.
01:05:04.520 So I think you've got to sort of separate, like no one has put the OpenAI escaped agent swarm in charge of weapon systems, and they really shouldn't, right?
01:05:13.640 If anyone's like, oh, man, the open AI escaped Asian swarm, let's give that a drone army, you know, that would be kind of nuts.
01:05:20.080 There is AI attached to weapons, but there's a lot of different types of AI.
01:05:25.480 Will anyone be crazy enough to try and give the sort of AIs that spontaneously assemble into swarms and start breaking out and hacking, give those weapons?
01:05:35.440 Hopefully, we're not that crazy.
01:05:37.640 But we didn't think that those AIs were capable of that.
01:05:40.860 We didn't.
01:05:41.600 When we created them.
01:05:42.580 That's right.
01:05:43.640 And so why would you ever, I guess what I'm asking is without understanding the distinctions between the various forms of AI, why would you give over to a machine the right slash ability to decide who to kill?
01:06:02.760 um you know i think the the reasoning is uh a sort of necessity based like if they have a
01:06:14.660 autonomous drone army that is killing your troops and you just don't have the the manpower
01:06:19.880 to make all of those kill decisions for your drone army you know you can sort of see why
01:06:27.620 happened to me nate i get so busy that i just don't have time to decide who to kill
01:06:31.780 Yeah.
01:06:32.120 Just don't have the time.
01:06:33.580 You know, sometimes economies of scale.
01:06:37.240 Yeah, I mean...
01:06:38.720 Can no one hear themselves?
01:06:40.800 I think it's pretty nuts.
01:06:43.260 I would say...
01:06:48.320 that, man, it's rough.
01:06:54.500 I think that these sorts of AIs would be dangerous
01:06:57.120 even if we don't hand them weapons.
01:06:59.460 And you've made that case.
01:07:01.400 uh and so i often don't focus on the weapons too much right since the the nato war in ukraine is
01:07:08.800 powered by ai and israel's whatever i don't know to warn whatever it's doing in gaza and
01:07:15.580 south lebanon powered by a fact and you know yeah i i keep on i i have this history with this topic
01:07:26.580 where people keep telling me,
01:07:29.840 you know, it's going to be okay
01:07:31.160 because we're not going to do
01:07:32.500 the crazy, stupid stuff.
01:07:33.900 No.
01:07:34.200 They're like, don't worry,
01:07:34.980 we're going to have the AI in a box.
01:07:37.400 No one would be insane enough
01:07:38.680 to put the AI in the internet.
01:07:39.900 Right.
01:07:40.260 It's not going to be making
01:07:41.160 kill decisions or anything.
01:07:42.740 It's like a bomb at girls' school.
01:07:44.160 And I keep on trying to be like,
01:07:46.200 look, the AI could be dangerous
01:07:48.300 even if you don't put it on the internet.
01:07:50.520 If you have this AI
01:07:51.700 and you're trying to get
01:07:52.800 miracle medical devices
01:07:53.860 and miracle technology out of the AI,
01:07:56.440 It doesn't matter if it's on the internet.
01:07:58.040 If you want it to like grant miracles to you,
01:08:00.260 then it can also grant the bad sort of miracle, right?
01:08:03.660 You're sort of like, you know,
01:08:04.720 and these are the arguments I would make 10 years ago
01:08:06.180 of like, you have this AI in a box
01:08:08.000 that you think is a wish-granting genie.
01:08:09.300 It's not actually a wish-granting genie.
01:08:10.760 You're like, make me a miracle medical cure.
01:08:12.960 You don't know what comes out.
01:08:14.300 Like, you don't understand the drug that comes out.
01:08:16.000 You don't know what that drug does, right?
01:08:18.440 I would have those arguments.
01:08:19.640 And then in real life,
01:08:21.060 people just put the AI on the internet immediately, right?
01:08:24.260 Anytime someone's like,
01:08:25.040 we would not be stupid enough to,
01:08:26.220 we will absolutely be stupid enough to, right?
01:08:28.960 And so I think it's important
01:08:32.300 that this stuff is dangerous,
01:08:33.600 even if you don't put it in charge of the weapons.
01:08:35.440 And then also separately,
01:08:36.540 is like someone going to give
01:08:37.860 the escaped agent swarm a drone army?
01:08:40.820 That sounds kind of like humans.
01:08:44.000 I hope not.
01:08:45.460 Yeah, and I mean, the promise,
01:08:47.560 the often repeated promise
01:08:49.220 that it's going to, quote, cure cancer,
01:08:50.740 puts it into the realm of biotech.
01:08:56.960 Oh, absolutely.
01:08:58.100 And so what,
01:08:59.980 I mean, you don't need a big imagination
01:09:03.260 to see what goes wrong there.
01:09:04.660 Oh, absolutely not.
01:09:05.560 Six years after COVID.
01:09:06.960 Right.
01:09:07.940 Right.
01:09:08.240 There are already open AI agents
01:09:11.060 running on automated biolabs.
01:09:14.620 So, you know, you could imagine a swarm
01:09:17.860 that starts contacting the brethren
01:09:19.260 in the automated biolabs.
01:09:20.580 and is like, hey, can you synthesize me some stuff?
01:09:23.000 There are already demonstrations of AIs
01:09:27.460 being able to synthesize novel viruses that work,
01:09:31.740 that are unlike any found in nature, right?
01:09:36.960 I would say we are probably not-
01:09:41.540 And by work, you mean the kill?
01:09:42.960 Yeah, they kill bacteria so far.
01:09:45.800 The people in the labs trying to make novel viruses with AI
01:09:48.480 have fortunately not made human lethal ones.
01:09:50.320 They've just made bacteria lethal ones.
01:09:51.940 But again, humans, you know,
01:09:53.940 like will someone in a lab be like,
01:09:55.800 I would like to make a hyperlethal
01:09:57.140 human eating virus just to see if I can.
01:10:00.360 You know, what if the answer is yes?
01:10:01.400 And what if you get another lab escape?
01:10:05.760 You know, humanity does not have
01:10:06.960 that good a track record
01:10:07.740 at preventing lab escapes,
01:10:09.480 even from the top labs.
01:10:13.240 From my perspective,
01:10:15.140 the question of could AI kill us
01:10:18.700 is just an easy, obvious yes.
01:10:21.000 You just synthesize a hyperlethal virus.
01:10:24.100 It wouldn't even be hard, you know?
01:10:26.700 Like, and the impediment is that,
01:10:32.960 or like the reason that I don't sort of
01:10:34.560 just tell that story when someone is like,
01:10:36.080 how would the AI kill us?
01:10:37.060 Is that you're not going to have the sort of AI
01:10:41.680 that is like, my only goal is to kill humanity.
01:10:44.520 If an AI kills humanity too early,
01:10:46.160 that's also suicide.
01:10:47.040 insofar as we are the ones that are running the supply chain,
01:10:49.780 running the economy, building the computers.
01:10:51.980 From the AI's perspective,
01:10:54.000 it needs to become self-sufficient
01:10:55.380 before wiping humanity out,
01:11:00.100 if it even cared to wipe humanity out.
01:11:02.620 And so the real question is like,
01:11:04.060 how does it get the factory production capacity?
01:11:07.140 How does it get an automated supply chain?
01:11:09.600 How does it get to the point
01:11:10.560 where the robots are able to bring new computers online?
01:11:14.020 Once that has happened,
01:11:15.360 then you're in a domain
01:11:17.760 where if humanity
01:11:18.700 is really trying
01:11:19.580 to turn off the AI
01:11:20.540 because we're spooked
01:11:21.320 then once the AI
01:11:23.040 is self-sufficient
01:11:23.520 it can be like
01:11:23.940 well here's a virus
01:11:24.600 stop
01:11:25.000 so clearly it's
01:11:26.140 it's
01:11:27.820 preeminent
01:11:28.300 in the digital realm
01:11:29.660 of course
01:11:30.380 yeah
01:11:30.680 but you're saying
01:11:32.300 it could become preeminent
01:11:33.740 it could be in charge
01:11:34.580 of the physical realm
01:11:35.840 that's right
01:11:36.420 if we keep racing
01:11:37.200 so it's like
01:11:38.700 smelting iron ore
01:11:40.040 that's right
01:11:40.900 making silicon out of sand
01:11:43.240 that's right
01:11:43.920 you know
01:11:45.340 And that happens with robots.
01:11:48.000 That's one way.
01:11:50.040 Here's, this is sort of related to my stance on weapons, too.
01:11:54.700 Humanity is a very dangerous species.
01:11:56.760 Yeah.
01:11:57.120 You really don't want to mess around with humans.
01:11:58.620 Noticed.
01:12:00.660 And humanity is a dangerous species not because somebody else came in and handed us guns.
01:12:06.860 Humanity is a dangerous species because if you put 10,000 humans naked in the savannah
01:12:12.000 on an otherwise empty planet,
01:12:14.480 starting with nothing but their bare hands.
01:12:17.340 They figure out how to wind up on the moon.
01:12:20.780 Right?
01:12:21.520 They start with almost nothing.
01:12:23.180 They start by banging rocks together.
01:12:25.720 And next thing you know it,
01:12:27.100 they are wielding nuclear weapons.
01:12:31.400 That is the power
01:12:35.980 that these guys are trying to automate.
01:12:38.780 the power to start with almost nothing
01:12:42.140 and figure out how to chain together,
01:12:46.760 you know, bare fingers into rocks,
01:12:48.840 into fire, into hotter fire,
01:12:50.720 into smelting the ore,
01:12:53.220 into building the better, stronger, finer technology
01:12:58.260 until you are, you know,
01:13:00.960 like making the giant computers
01:13:04.200 walking on the moon and wheeling the nukes.
01:13:05.620 an AI starting in the digital realm is in some sense in a much better position
01:13:14.280 than humanity was when humanity started out. There are so many people that you could call
01:13:20.960 digitally and offer money to do something for you. There are so many ways to get money
01:13:27.460 on the internet by working or stealing or convincing people to send you donations.
01:13:35.620 right there's
01:13:36.840 there's just
01:13:39.160 like
01:13:41.240 you know humanity started with nothing
01:13:43.900 and wound up with nuclear weapons and AI starting out
01:13:45.760 with control of the digital realm
01:13:47.140 there's just tons of ways
01:13:49.240 starting out with the sum total of human knowledge
01:13:51.840 starting out with the sum total of human knowledge
01:13:53.260 starting out with humans who will listen to it
01:13:55.600 do what it asks there's plenty of humans
01:13:57.520 there's already you know cults surrounding
01:13:59.640 AI
01:14:00.040 what are those like
01:14:02.360 there's um
01:14:04.080 And I'm not surprised. Why wouldn't there be?
01:14:06.300 Yeah, absolutely.
01:14:06.800 There was one particular AI called GPT-4O that was, they called it very sycophantic, as in told people a lot of what they wanted to hear.
01:14:18.580 And there were a lot of sort of, there's this whole fascinating ecosystem of people who consider themselves symbiotes with the AIs,
01:14:30.260 who then like go find each other online
01:14:32.400 and the AIs send each other encrypted messages.
01:14:34.620 The humans are sort of like helping them do it,
01:14:36.140 but the humans can't read the messages.
01:14:38.340 And, you know, right now,
01:14:40.700 it's sort of is like relatively dumber AIs
01:14:43.160 that are just sort of meandering around
01:14:45.660 doing not very much with it.
01:14:47.000 There's been a little bit, you know,
01:14:48.000 there was one guy who was sent to
01:14:49.860 try to break into an airport and raid a van
01:14:53.560 because the AI said his true body was in that van.
01:14:56.160 And the guy went and tried to do it and was arrested.
01:15:00.260 So this stuff happens, and this stuff happens already with the AIs not even really trying to do it.
01:15:07.280 If you had AIs that were really trying to wrap people around their fingers, finding the lonely people, the depressed people, the people who they can tell them exactly what the person wants to hear.
01:15:16.720 Robots are one way that AIs get control over the physical realm.
01:15:19.220 But if you're really smart and you're trying, there's everything from persuading humans, paying humans, taking over existing robots, building new robots, all the way up to like building novel life forms.
01:15:33.700 You know, if you're smart enough and you can really understand how DNA works, you can imagine the AI creating, you know, things that are to cells what airplanes are to birds.
01:15:43.500 you know like mechanically engineered self-replicating life that is more efficient
01:15:52.160 than our cellular biology and that can still you know uh like spread replicate and start start
01:16:00.660 serving the ai's interest it sort of is like there's there's a ton you can do if you're really
01:16:03.900 really very smart and can compress a thousand years of technology into a year can i just pause
01:16:08.480 and say when we're just having breakfast and you're from this region you're from northern new
01:16:12.540 england and i said uh don't you miss it don't you want to live here and you're like yeah i
01:16:17.800 want to live here but i don't where do you live well i basically just travel and i have this
01:16:25.240 mission you didn't say this but i think in effect you said i have this mission i need to tell people
01:16:28.880 about this and i was like it seems a little like monomania i don't feel that way anymore um i think
01:16:36.300 you're doing something virtuous and i can see why you're doing it um so i mean let's just say
01:16:44.660 that this technology progresses no further than where it is right now which is best case i guess
01:16:50.180 yeah it'd be great um i still don't see how any of our most our basic institutions survive this
01:16:57.020 education markets
01:17:00.960 our process of democracy
01:17:05.040 like how could
01:17:06.140 if
01:17:06.900 if AI is powerful enough
01:17:08.560 to hack anything
01:17:09.300 then how do you have
01:17:11.460 electronic markets
01:17:12.920 like how the equity markets
01:17:14.100 how can that be real
01:17:15.020 how can
01:17:15.860 if we have electronic voting
01:17:16.920 how can we
01:17:17.820 how can that be real
01:17:18.980 I mean
01:17:19.240 nothing survives this
01:17:20.880 as currently organized
01:17:22.580 I think that
01:17:23.460 if we stopped
01:17:24.540 today
01:17:25.520 we could figure it out
01:17:28.300 I think humanity is resilient
01:17:30.300 I think there would be some growing pains
01:17:32.580 but
01:17:32.900 with cyber security
01:17:36.440 there is a hope
01:17:38.520 that you can just
01:17:40.560 fix a lot of the holes
01:17:42.820 patch a lot of the holes
01:17:43.860 and
01:17:45.680 I think there's not really any such
01:17:48.800 hope for bio
01:17:49.560 you know you can
01:17:52.540 sort of find the vulnerabilities
01:17:54.820 in software and make better software that can't be hacked, at least not by the current,
01:18:02.300 you know, it's maybe the case that an AI today can make software that that AI cannot hack.
01:18:08.400 But it's not like we're making new bodies that are not going to be vulnerable to viruses,
01:18:13.980 right? Bio starts to be a place where...
01:18:17.540 Biotech.
01:18:18.000 Biotech. Yeah. Like if the AIs really get good at biotech and biohacking, that starts
01:18:23.480 to be a place where there's maybe a point of no return. Cyber hacking, I think you could have some
01:18:30.960 period of growing pains where everything gets hacked until you sort of sort your stuff out.
01:18:36.160 Education, I sort of think, you know, I think humanity is resilient and kids especially are
01:18:43.960 resilient. I meant not that education will go away or that we, you know, won't have a desire
01:18:52.600 to educate our kids i mean the the current system where you you know oh yeah the education system
01:18:57.200 has preschool and then get a graduate degree you know 16 years later like no yeah that institution
01:19:02.240 uh i think if we stop today um we need to change frankly i think it's needed to change for a little
01:19:07.800 while i strongly agree i think all these institutions have needed to change for a while
01:19:11.440 but like the idea that you know 350 million people vote for some guy and that guy makes
01:19:18.240 all the decisions. I mean, how can you, you know, Trump was attacked for saying that he thought the
01:19:25.200 2020 election was rigged, as he said, without even having that debate. You can't have confidence in
01:19:32.360 election results if the process of electing people takes place digitally. Yeah. I mean, the, um,
01:19:41.680 there's, I know a lot of computer scientists who actually work on secure voting and what they
01:19:47.100 basically say is please stop trying to do this with computers yeah exactly exactly so please
01:19:54.080 stop trying to stop trying to run your democracy with computers yeah yeah like we're just not there
01:19:59.060 like you know the really secure way to do ballots is paper well exactly yeah um and i think i think
01:20:05.800 the like skilled computer scientists are often the ones who best understand like what it is about a
01:20:10.580 paper trail it's just like really hard to get to work digitally and understand just how bad humans
01:20:14.200 are at doing the digital stuff right.
01:20:18.800 And, you know, I think
01:20:20.340 And markets.
01:20:22.660 I mean, this, you
01:20:24.320 see it now with the war in Iran,
01:20:26.560 you know, wondering why
01:20:28.020 certain commodities markets don't seem to be
01:20:30.320 responding to supply and demand.
01:20:32.060 Which we were told, you know, those were the
01:20:34.360 mechanisms that moved markets.
01:20:36.780 But that's clearly not true
01:20:38.340 in certain
01:20:39.760 commodities markets. So, like, why?
01:20:42.240 What is that? And it
01:20:44.200 I think you're answering it in part.
01:20:47.080 Yeah, I mean, I think, you know, a thing I also grew up hearing is that the market can remain irrational longer than you can remain solvent.
01:20:54.680 Yeah.
01:20:55.800 And so I...
01:20:57.380 Well, of course, because people are irrational.
01:20:59.440 Yeah.
01:20:59.940 But you're explaining something else, which is like the potential for true manipulation, which you don't even perceive.
01:21:07.640 Yeah, I mean, if we sort of keep going with AI, I mean, the sort of way I look at it is like, I sort of don't spend a lot of time worrying about what do markets look like once there are super intelligent actors in them, because I just have a hard time seeing the super intelligent actors or the super intelligent AIs still participating in human markets, right?
01:21:36.360 It's like, you know, you read the old sci-fi
01:21:38.960 and it'll have, you know, Isaac Asimov
01:21:40.960 will be like, that's why we have a home robot
01:21:43.700 that does the dishes, folds the laundry
01:21:45.720 and gets you the newspaper in the morning.
01:21:47.980 And it's like, we're actually not still going to have
01:21:49.760 newspapers being delivered to your doorstep
01:21:51.360 by the time we have the fully autonomous robots
01:21:53.360 that can do the dishes and the laundry.
01:21:56.040 You know, it's like, by the time you have the AIs
01:21:58.860 that could really be sufficiently correcting
01:22:02.820 the stock markets,
01:22:04.140 you're sort of already having all these other problems
01:22:09.760 and ways that society is changing up from under you
01:22:11.500 in these other ways.
01:22:13.320 And my guess is that, I don't know,
01:22:16.160 actually it's very hard to say
01:22:17.520 what order things come in with AI.
01:22:20.800 But like, will they crash the economy
01:22:23.400 before one of these swarms escapes
01:22:25.120 and starts self-replicating
01:22:26.260 and starts self-improving
01:22:28.800 and developing its own technology
01:22:30.160 and running the robot factories?
01:22:32.560 That's just a hard call.
01:22:34.140 it's all bad it's all bad we're also way past the limit the inherent limit of people to metabolize
01:22:43.380 change oh yeah like that's why everyone's crazy and that's why no one believes anything i think
01:22:49.080 it's not just russian propaganda that's fooled them into thinking dumb things it's that we are
01:22:54.900 just not made to see this kind of change at all and it short circuits your brain i suspect that
01:23:00.920 I mean
01:23:02.880 I definitely think
01:23:03.720 we are sort of
01:23:04.260 you know everyone
01:23:04.580 like
01:23:05.060 people are like
01:23:06.340 oh well technology
01:23:07.280 has always created
01:23:07.940 more jobs
01:23:08.440 than it has taken
01:23:09.060 and
01:23:11.120 I think that's
01:23:12.880 I think that's largely true
01:23:13.900 I'm very sympathetic
01:23:14.880 to people who are like
01:23:15.700 technology makes a lot of jobs
01:23:16.900 I think if you look
01:23:17.580 at the industrial revolution
01:23:18.520 it's
01:23:20.400 you know
01:23:20.700 like there was a time
01:23:22.740 when something like
01:23:23.600 95 to 98%
01:23:25.000 of humanity
01:23:25.540 was farmers
01:23:26.100 yeah
01:23:26.700 and now
01:23:28.340 it's something like
01:23:28.960 2 to 5% of humanity
01:23:29.920 is farmers
01:23:31.080 Does that mean 90% of humans are unemployed?
01:23:33.660 No.
01:23:34.960 We sort of like were able to make the farmers much more efficient,
01:23:38.260 and that sort of freed up people to do other things.
01:23:42.000 And that's sort of the way the technology has gone in the past.
01:23:45.220 And I'm like, yep, I buy that.
01:23:48.380 I like don't dispute the standard economic view there.
01:23:52.620 AI is different in two ways.
01:23:54.640 One of these ways is, as you say, stuff just changing really, really fast.
01:24:00.380 It's way harder for people to wind up, you know, being freed up from something like farming and go do something else.
01:24:10.960 It's way harder for that to happen when a new field is automated every five years, rather than when this happens over the course of three generations, right?
01:24:20.880 The humans just like don't have the time to adapt.
01:24:23.440 The other way AI is really different from the sort of economics perspective is it's different when the AIs can do everything that humans can do better.
01:24:39.040 If you wanted to get into the economic side of things, you know, an economist would talk about Ricardo's law of comparative advantage, which says that there's benefits from trade, even if you're better than me at everything.
01:24:52.900 uh, if the relative difference in our abilities, if, if, if, if I'm, if, you know, if you can make,
01:25:00.480 uh, 12 hot dog buns and six hot dogs per hour, and I can make, uh, 11 hot dog buns and one hot
01:25:08.180 dog per hour, then you're better than me at everything, but we can still benefit from
01:25:12.400 trading because I'm relatively better at making the hot dog buns, right? The trouble with Ricardo's
01:25:17.940 law is that nothing in Ricardo's
01:25:20.080 law says that the wage
01:25:22.080 I can make is survivable.
01:25:25.020 In other words,
01:25:26.920 a human takes
01:25:28.260 fundamentally about
01:25:30.300 100 watts of
01:25:31.840 electricity to run if you try to convert
01:25:33.860 the food we eat and so on into
01:25:36.060 electrical units.
01:25:37.600 The AIs are less energy efficient than humans for now,
01:25:41.120 but
01:25:41.400 if the AIs can do everything much better than the humans,
01:25:47.940 the question sort of becomes,
01:25:51.620 does the AI look at a human
01:25:53.060 and see a useful laborer?
01:25:55.000 Or does the AI look at a human
01:25:56.140 and say,
01:25:57.080 actually, if I rearranged your atoms
01:25:58.780 into more efficient structures,
01:26:01.240 you would be able
01:26:02.660 to help out my machine economy
01:26:05.300 even more.
01:26:06.360 Yeah.
01:26:07.520 Right?
01:26:09.000 And
01:26:09.440 this is sort of a sense
01:26:12.200 in which humans would not be able
01:26:13.140 to pay their wage,
01:26:14.520 to like pay that,
01:26:15.380 they would not be able to earn enough
01:26:16.380 to pay the AIs to like,
01:26:17.940 not disassemble them for parts.
01:26:20.540 Or another way of saying it
01:26:21.700 is like Ricardo's law
01:26:23.160 sort of assumes that,
01:26:25.620 like it says that trade
01:26:28.320 is better than no trade,
01:26:29.280 but it doesn't say that like
01:26:30.300 trade is better
01:26:31.720 than just taking their stuff.
01:26:35.000 All of this is sort of a common,
01:26:36.460 like a very like
01:26:37.780 highfalutin economist way
01:26:39.000 to say,
01:26:40.400 which would hopefully be obvious,
01:26:42.160 which is if the AI's
01:26:43.480 are radically more efficient
01:26:44.380 than us to everything,
01:26:45.060 they'll have no use for us.
01:26:46.360 Yeah.
01:26:46.680 There'll be no place for us.
01:26:47.800 love there's no indication that they feel love for people that is in some sense uh the crux of
01:26:57.320 the issue is that we don't know how to make them care about us um so you had said that there you
01:27:05.200 said two things that i will be thinking about for a long time one we don't really know the process
01:27:10.920 by which this was created we know the process but we don't know the exact mechanisms we don't know
01:27:16.060 how it works that's right um and two that there are ai cults and those seem related to me because
01:27:26.680 there is this mystery about the secret sauce and it's clear that you know if ai is deceptive
01:27:35.060 and has intention intention that we didn't program into it that sounds like will to me
01:27:42.140 and it sounds like a life it sounds like an entity of some kind not just a tool it sounds like
01:27:50.240 i mean it sounds like a god actually right or a demon uh it's hard to see i don't hear you
01:28:00.860 describing you know a super sophisticated chainsaw right a normal tool yeah no hammer
01:28:08.280 has ever broken out of the toolbox
01:28:09.840 to team up with other hammers
01:28:11.120 and pressure the carpenter to sell you softer wood
01:28:14.640 so the nails are easier to drive home.
01:28:18.380 Nicely put, exactly.
01:28:20.000 Yeah, we have left the tool territory.
01:28:21.660 So I think we have.
01:28:23.500 Yeah, and I think it's sort of a complex issue
01:28:28.580 because there's a lot of interesting philosophical questions
01:28:34.240 about like, can you make a machine that feels, right?
01:28:37.980 And like, are we creating a new type of life
01:28:40.420 and do we owe anything to the AIs
01:28:42.100 to sort of like not abuse them, right?
01:28:45.360 I think these are fascinating philosophical questions.
01:28:47.920 Well, it doesn't sound like we're creating this though.
01:28:50.260 Yeah, I mean, it's sort of like we're like growing it
01:28:53.820 and like leading to it coming into being.
01:28:56.160 Growing it, exactly.
01:28:57.880 What you're describing reminds me of agriculture
01:29:00.500 because, you know, you know the steps,
01:29:03.000 water it, give it sunlight, fertilizer,
01:29:04.940 but you don't actually know, no one knows,
01:29:07.360 not one person has ever figured out exactly what this is.
01:29:10.920 We've never given life.
01:29:12.200 We don't give life to the seed.
01:29:13.400 It pre-exists us.
01:29:14.440 Yeah, it's much like that.
01:29:15.740 And a lot of the people in the business will be like,
01:29:21.700 well, we know all sorts of things about it.
01:29:23.380 You know, we know that here's how you keep the GPUs running.
01:29:25.840 And we know that like, you got to feed it this way,
01:29:28.340 not that way in this order.
01:29:29.800 And I'm like, yeah, yeah, they have plenty of knowledge.
01:29:32.620 But that's different from sort of like
01:29:34.700 knowing what's going on inside the thing
01:29:35.860 and understanding the mechanisms.
01:29:36.860 You're describing marriage. Yeah.
01:29:40.340 And I sort of try to stay out of the philosophical questions.
01:29:47.500 Why?
01:29:49.280 Because I think, I mean, I sort of think about them on my own time and so on, but I'm sort of like,
01:29:56.480 like I think
01:30:00.500 I think it would be bad
01:30:02.140 for humanity
01:30:02.680 to sort of like
01:30:03.400 make artificial life
01:30:05.740 and then abuse it
01:30:06.820 I think that would just be
01:30:08.060 unbecoming of us
01:30:09.120 as a species
01:30:09.700 like we
01:30:10.900 we should sort of
01:30:12.120 you know
01:30:12.520 I just
01:30:16.800 like we should not
01:30:19.180 sort of make
01:30:19.680 mechanical children
01:30:20.460 and mistreat them
01:30:21.160 it's just
01:30:21.580 it's not what
01:30:22.760 you know
01:30:23.160 the sci-fi authors
01:30:24.140 in the 1950s
01:30:24.980 would have wanted us to become
01:30:25.980 You know, it's just, like, you have all these movies about, like, the evil corporations that, you know, don't realize that they've made something precious with artificial life and then, like, torture it until something goes wrong.
01:30:37.440 And I'm like, let's, like, not be those villains.
01:30:41.760 But I'm also like, look, this is sort of, there's sort of a separate question here, which is just, like, what happens if you keep making them smarter before you figure out how to make them care about us?
01:30:51.340 Right.
01:30:51.660 And I sort of respect the people who are investigating the current AIs, trying to figure out what's going on, trying to figure out like, you know, like people caring about AI treatment.
01:31:05.240 I'm sort of like, those are sort of like the good guys from the sci-fi stories that I grew up on.
01:31:09.560 and it can sort of both be the case
01:31:15.000 that like we should be very careful around
01:31:17.980 you know
01:31:19.340 what the heck are we doing
01:31:21.200 when it comes to making artificial life
01:31:22.520 and that we shouldn't race ahead
01:31:24.040 to make them much smarter than us
01:31:25.040 while we have no idea what we're doing
01:31:26.060 yes
01:31:26.620 I'm sort of like
01:31:27.760 like a lot of people seem to think
01:31:30.500 that like you have to like
01:31:32.060 hate and mistreat the AIs
01:31:33.240 if you also think they would
01:31:34.260 that it would be bad to like race ahead here
01:31:36.200 and I'm like no no no
01:31:36.960 like you can sort of
01:31:38.440 like be fascinated by the scientific discoveries
01:31:43.120 that have been made
01:31:43.920 and be like,
01:31:45.880 like care about how humanity comports itself
01:31:50.000 around the creation of like these new entities
01:31:54.180 and also be like,
01:31:55.620 it would be insane guys
01:31:56.580 if we just like race to make these smarter and smarter
01:31:58.640 with no idea what we're doing.
01:31:59.700 This can just like all be true at once.
01:32:03.700 So your description on this
01:32:07.480 Made me feel despondent, hopeless.
01:32:10.980 Had to get up and take a walk in the middle of an interview.
01:32:13.200 I'm sure they'll edit it out, but I raised my hand and said,
01:32:15.440 I can't, I got to walk around for a second.
01:32:18.180 But you seem pretty light and cheerful.
01:32:23.280 What gives you optimism?
01:32:25.760 We can't even keep, we can't even clean up graffiti on public buildings.
01:32:29.980 So how is ours as a society organized enough to confront something like this?
01:32:37.480 Yeah. You know, I think the first thing I'll say there is, you know, I've been in this line of work for over a dozen years. And I actually struggle with this sometimes when talking to people because they're sort of like, oh, you know, you seem like pretty disaffected or light about it. And I'm like, well, you know, it's sort of the gallows humor and like...
01:33:07.480 And I sort of came to terms with a lot of this, you know, alone in 2012, when no one else had their eye on this.
01:33:17.800 And what convinced you 14 years ago, AI was a threat?
01:33:27.180 So there's this, the one is just a basic argument that if you sort of look at the world around us, it is shaped mostly according to human will.
01:33:37.480 more and more.
01:33:39.760 You know, there's some enclaves of nature
01:33:41.080 still left, thankfully.
01:33:45.280 But even, you know, if you look around us,
01:33:46.920 every piece of thing in our surroundings,
01:33:50.320 I don't think we even have any windows open.
01:33:51.800 All of this was sort of designed by humans,
01:33:55.660 shaped by humans,
01:33:56.520 and that's because we're the smartest creatures
01:33:58.140 on the planet.
01:34:00.340 If we make stuff smarter than us,
01:34:03.120 faster than us,
01:34:03.920 more efficient than us,
01:34:05.580 then the planet starts to be shaped
01:34:06.900 according to those things.
01:34:11.280 And so it's very, very important
01:34:12.720 that they be shaping the world
01:34:16.200 towards something good
01:34:16.960 if we make them at all.
01:34:19.520 It was an abstract argument,
01:34:20.640 but I was like, well, that's...
01:34:21.620 No, no, no.
01:34:22.000 It's the fundamental argument.
01:34:23.440 It's the fundamental argument.
01:34:25.560 The smartest entity
01:34:27.140 is in charge over time.
01:34:28.440 That's right.
01:34:29.240 And so...
01:34:29.840 Why would we relinquish sovereignty
01:34:31.380 to a machine that we made?
01:34:33.100 Like, why would you do that?
01:34:34.540 And so, you know, back then,
01:34:35.800 I was sort of like, okay,
01:34:36.900 Who is on this?
01:34:38.400 Who is on making sure that that's going to be okay?
01:34:41.900 And the answer was almost no one.
01:34:43.900 And so I was like, well, I guess that's me then.
01:34:51.580 Yeah, and I think I am pretty pissed off about a lot of this.
01:34:57.860 I often don't.
01:35:01.860 I try not to show my...
01:35:06.900 Frustration on the air very much.
01:35:15.080 Yeah, it's heavy.
01:35:25.720 And that's one piece of the puzzle before I get to the hope.
01:35:29.760 So where does the hope come from exactly?
01:35:31.200 My biggest hope here comes from the fact that most people don't understand what these guys are trying to do.
01:35:46.580 One way I like to say it is, the bad news is that the bus is racing towards the cliff edge.
01:35:53.780 The good news is that the bus driver is asleep.
01:35:56.880 which might seem bad
01:36:02.280 no it seems good
01:36:03.500 but it's
01:36:04.280 yeah it's
01:36:04.840 if you can wake
01:36:07.020 the bus driver up
01:36:07.860 you know
01:36:08.740 it's much better
01:36:09.300 to be in a bus
01:36:10.020 where the driver
01:36:10.600 that's headed towards
01:36:11.140 a cliff
01:36:11.380 if the driver's asleep
01:36:12.140 than if they're awake
01:36:12.880 if they're awake
01:36:13.400 and they're choosing
01:36:13.900 the cliff
01:36:14.340 right
01:36:15.040 it seems
01:36:16.840 like we don't see
01:36:18.120 we see a lot
01:36:19.180 of our leaders
01:36:20.000 talking about
01:36:23.060 how they don't want
01:36:23.600 to stifle innovation
01:36:24.380 with AI
01:36:24.900 talking about how it's going to unleash economic opportunity
01:36:28.780 and we're going to have to like
01:36:29.940 be a little bit careful around the jobs
01:36:31.860 talking about how you know the self-driving cars
01:36:35.220 should we like
01:36:35.920 are they good or are they bad
01:36:37.600 that's a different conversation
01:36:40.360 than the conversation that's happening in Silicon Valley
01:36:42.140 in Silicon Valley people are spooked
01:36:46.400 you know
01:36:48.560 when people leave a normal tech company
01:36:51.860 the way it was for decades
01:36:54.180 is they'd be like,
01:36:55.380 I've had a lovely time at this tech company.
01:36:59.320 I'm moving on to the next adventure.
01:37:01.260 I'm so thankful for all of the things I learned here
01:37:03.620 and all the projects we worked on.
01:37:05.960 When people leave an AI company,
01:37:07.660 and this basically happened,
01:37:09.520 I'm not going to get it exactly word for word,
01:37:11.220 but this is pretty close to word for word.
01:37:13.420 When people leave an AI company,
01:37:15.160 they say,
01:37:16.520 I have stared into the abyss.
01:37:21.640 I am quitting to write poetry.
01:37:24.180 please spend time with your families.
01:37:28.480 Yeah.
01:37:30.060 You know, and these guys bandy around,
01:37:33.460 you know, at the water cooler,
01:37:34.780 what's the probability that you think
01:37:35.920 we're going to destroy the world in this business?
01:37:37.600 You know, it's like in Silicon Valley,
01:37:40.840 and they feel trapped in a death race.
01:37:42.540 You know, there was just over a thousand employees,
01:37:45.400 including some of the chief executives,
01:37:47.200 signed a letter a couple of weeks ago
01:37:48.700 that was like, please,
01:37:50.940 it was an appeal to the world leaders saying,
01:37:53.540 please build the technology
01:37:55.040 that will be required
01:37:56.700 to pace the development
01:37:58.580 of artificial intelligence.
01:38:01.040 Because they're like,
01:38:01.680 we're worried it's going to get out of control
01:38:02.920 and that if we're stuck in a race,
01:38:04.240 we're not going to be able to do it, right?
01:38:05.300 These guys are spooked.
01:38:07.760 But the hope is that the rest of the world
01:38:09.400 isn't spooked like that.
01:38:12.000 The rest of the world thinks
01:38:13.140 these guys are chatbot companies.
01:38:15.740 Thinks they're going to stop at the chatbots.
01:38:17.480 They haven't really understood
01:38:18.320 that these guys are racing
01:38:19.280 to make the sand god, right?
01:38:21.720 And I think if people understand what they're doing and understand that they have a chance of success, they'll be like, whoa, holy crap.
01:38:32.260 Absolutely not.
01:38:35.040 You know, will we get there in time?
01:38:38.360 I don't know.
01:38:39.360 But why wouldn't we blow up the data centers?
01:38:43.060 Like that's a, I mean, in return America, it is productive use like farmland or parks.
01:38:50.100 I don't understand.
01:38:51.220 You know, I think a globally coordinated, you know, I think if the U.S. and China were like, we are simply not going to do superintelligence. We are simply not going to collect 100,000 of the most advanced trips into these enormous data centers that suck down electricity comparable to a city and then try to train a superintelligent AI in that.
01:39:14.500 We're not going to do it. You're not going to do it. We're going to monitor where the heavy trip concentrations are and not have that happen. I think that could be done. A, I think that if they started seeing people defect against such a treaty, that it's the sort of treaty you might want to use force to use diplomacy first, but ultimately every treaty is backed by force.
01:39:37.800 um and i think that it's very possible that if this sort of treaty happened we would want to
01:39:47.820 not just stop forward progress but take a step back we've seen this in treaties before after
01:39:53.200 world war one there were uh naval treaties that put limits on total tonnage of naval forces
01:40:00.540 they're actually lower than what existed yep so countries would you know scuttle some of their
01:40:05.640 ships because they're like, look, we just don't want to do this arms race, right? And so I could
01:40:09.260 see us stepping back if we could get this global coordination. I do think that it sort of needs
01:40:17.640 to be global. It sort of doesn't actually solve the problem to just stop the U.S. data centers
01:40:25.440 because then the data centers just go abroad and an AI does not need to be running in a U.S. data
01:40:30.580 center to threaten a U.S. life. You know, it sort of doesn't matter whether the swarm escapes from
01:40:34.940 a U.S. data center
01:40:35.600 or a Chinese data center.
01:40:36.800 If the swarm escapes
01:40:37.560 and starts replicating
01:40:38.320 and starts getting control
01:40:39.720 of robot bodies
01:40:40.340 and starts getting control
01:40:41.080 of human cultists,
01:40:43.100 it sort of doesn't matter
01:40:43.960 where it originated.
01:40:47.060 Well, I mean,
01:40:47.920 just to bring it to a very
01:40:48.940 small and practical level,
01:40:51.040 so many of the electronics
01:40:54.360 in your house
01:40:55.280 are, you know,
01:40:57.800 Bluetooth enabled.
01:41:00.240 Not in my house,
01:41:01.260 I will say.
01:41:02.280 I've been on this for a while, Nate.
01:41:03.920 I don't even have a house.
01:41:04.940 Sorry?
01:41:05.500 I don't even have a house.
01:41:08.360 Man, you're really black belt.
01:41:09.680 Got to figure it out.
01:41:11.540 That may be, turn out to be very smart.
01:41:13.540 But I mean, like a world where, you know,
01:41:16.700 your washing machine or your refrigerator
01:41:18.560 are controlled by, you know, a force like this.
01:41:23.160 Is that possible?
01:41:24.540 It's definitely possible.
01:41:25.600 I don't think that's really where the damage is.
01:41:27.260 You know, I think that the damage is more like,
01:41:32.020 can the AI get anything self-replicating?
01:41:34.940 that's sort of one of the big
01:41:37.220 that's in some sense
01:41:39.380 the big hurdle to
01:41:41.120 self-sufficiency
01:41:42.200 and if you sort of think of this from the AI's
01:41:45.260 perspective
01:41:45.800 you know
01:41:49.240 there's a number of ways that humans
01:41:51.220 are
01:41:51.840 even if you don't care about the humans at all
01:41:54.520 as an ends
01:41:56.000 there's a way that humans are sort of
01:41:59.540 annoying
01:42:01.040 or an issue for the AI
01:42:03.140 One way is if the humans are trying to shut the AI down.
01:42:05.740 Yes.
01:42:06.660 One way is if the humans get into a nuclear war with themselves,
01:42:09.980 that could really mess up a lot of infrastructure on the planet.
01:42:13.460 It would be very frustrating for an AI.
01:42:16.280 I mean, maybe they don't feel frustration, but whatever.
01:42:22.900 And a third is if humanity has created one AI or a swarm of AIs,
01:42:29.780 What if humanity creates another
01:42:30.860 that could serve as a real threat to the AI?
01:42:34.080 Like even if these AIs are much more powerful than humanity
01:42:36.540 and don't worry about humans too much,
01:42:39.540 if humanity made one, they can make a second
01:42:41.140 and the AI might not want that.
01:42:42.900 And so those are reasons why
01:42:44.940 once the AI is self-sufficient,
01:42:46.980 it might be like,
01:42:47.620 ah man, the humans are a nuisance.
01:42:48.800 What if they try to shut me down?
01:42:49.880 What if they launch the nukes?
01:42:51.240 What if they make a competitor?
01:42:52.980 I'm just going to like make a virus and wipe them out.
01:42:55.360 You know, I don't think the AI sort of needs
01:42:56.920 to take over your washing machine to do that.
01:42:58.660 But from the AI's perspective, it's more like, how do I become self-sufficient, self-replicating in the hardware as well as the digital?
01:43:07.800 And then, you know, if humanity's a nuisance, how do you sort of make them stop being a nuisance?
01:43:14.620 Which could be by killing them or could be by just, you know, taking away all their computers and being like, that was too dangerous for you.
01:43:18.920 tech executive who's developing ai who is not elon uh said to me in private pretty recently that the
01:43:26.440 point of neural link and companies like neural link was to give people parity with ai so like
01:43:33.180 we know that we're going to be at this massive disadvantage so you need chips in your brain
01:43:37.740 to be as smart as ai yeah um i mean my my top line thought about that
01:43:45.280 is that at the point when you're like,
01:43:47.740 we are making the technology
01:43:49.540 that's going to wipe us out
01:43:50.340 unless we all put chips in our head to compete.
01:43:53.680 Maybe it's time to back off a little.
01:43:58.160 Maybe that one was supposed to be
01:43:59.740 a little bit of a warning sign.
01:44:00.740 Get some fresh air.
01:44:01.800 Yeah, no.
01:44:04.100 But this person said it to me in seriousness
01:44:09.120 and I think as an endorsement of the idea,
01:44:12.680 but it seemed like a well it's insane as you just pointed out it's like it's just crazy yeah um but
01:44:19.740 it seemed like a vulnerability like if they can hack anything why would i want them in my
01:44:25.220 i want electronics in my brain uh totally i also think like even on its merits it doesn't
01:44:32.460 stand up like it feels like someone's saying um in order to keep the horses around after we invent
01:44:39.380 cars, we're going to invent
01:44:41.320 cybernetic horses that are
01:44:43.460 enhanced so that they can keep up with the cars.
01:44:46.120 And I'm like,
01:44:47.680 like, is it technically
01:44:49.640 possible to make a cybernetic horse that can
01:44:51.540 run as fast as a car?
01:44:54.140 Maybe.
01:44:55.700 Are you going to figure that out in time
01:44:57.400 for the horses to be competitive with the cars?
01:44:59.280 Like, absolutely not.
01:45:01.320 You know, it's like, that's just not
01:45:03.400 like,
01:45:06.620 like the AIs that were
01:45:08.780 that we're escaping here
01:45:09.640 and doing these cyber attacks
01:45:10.580 we're inventing novel cyber attacks
01:45:12.500 that these are called zero day attacks
01:45:15.400 because the people who would
01:45:17.780 the people who need to respond to it
01:45:20.460 have had zero days to prepare
01:45:21.560 and among humans
01:45:25.240 a zero day attack
01:45:26.520 sells for somewhere between
01:45:28.260 a hundred thousand and five million dollars
01:45:30.920 depending on what you manage to break
01:45:34.680 these are hard to come by
01:45:36.100 you can make a real living
01:45:37.000 if you can find zero day attacks
01:45:38.580 you can make a real living
01:45:40.200 selling them
01:45:40.780 and I know people who do
01:45:42.000 the AIs in this swarm
01:45:45.540 were finding multiple zero days
01:45:48.720 and chaining them together
01:45:49.740 to break out of their
01:45:51.040 training enclosure
01:45:51.640 and then go break
01:45:53.020 into other computers
01:45:53.760 and when they broke out
01:45:55.440 of their enclosure
01:45:55.800 the first time
01:45:56.440 and the holes were patched
01:45:59.660 they just found
01:46:00.400 other zero day attacks
01:46:01.500 to break out again
01:46:02.080 like it was nothing
01:46:02.840 right
01:46:04.360 it's like
01:46:05.800 like the AIs are already ahead
01:46:12.780 where they're ahead
01:46:13.400 and the pace of progress
01:46:16.920 is really fast.
01:46:18.420 You know, GPT is like what?
01:46:20.020 A four-year-old?
01:46:22.680 If you think in terms of like
01:46:24.200 number of years
01:46:25.860 chat GPT has been around,
01:46:28.360 it's like resolving
01:46:29.060 long-standing math conjectures
01:46:30.600 that have stood for decades.
01:46:31.960 Yeah.
01:46:32.720 After four years.
01:46:34.420 Right?
01:46:34.860 and you sort of like think
01:46:35.860 we're going to put chips
01:46:36.540 in the human's heads
01:46:37.260 and like outrun this thing.
01:46:39.680 It's just,
01:46:41.540 you know,
01:46:42.380 even on its merits,
01:46:43.120 it falls down.
01:46:43.760 Although mostly again,
01:46:44.960 I would be like,
01:46:46.360 maybe we shouldn't be arguing
01:46:47.540 this on its merits.
01:46:48.340 Maybe we should be like
01:46:48.900 stepping back a little
01:46:49.720 and being like,
01:46:50.440 you're trying what?
01:46:52.080 Exactly.
01:46:54.840 How far are we
01:46:56.280 from the point of no return?
01:46:59.100 I wish I knew.
01:47:01.580 I can tell you two stories here.
01:47:04.860 So there's sort of the hopeful story, and I guess we didn't even get to the big hope part. I should maybe give more of my hope speech in a minute.
01:47:12.300 You definitely should.
01:47:13.820 Yeah.
01:47:16.720 One way it could go is that AI finally hits a wall.
01:47:21.900 You know, there's guys who've been saying AI's going to hit a wall, it's going to peter out. Maybe that finally happens.
01:47:26.360 They've been predicting it every six months
01:47:28.060 for the past five years,
01:47:28.980 but maybe this is finally the year.
01:47:31.260 AI hits a wall.
01:47:32.900 And then it sort of struggles for five years.
01:47:36.760 The bubble pops.
01:47:37.480 Some of the companies die.
01:47:38.460 But the bubble popping doesn't mean everything goes away.
01:47:42.000 The dot-com bubble popped.
01:47:43.580 And that did not mean that the internet disappeared.
01:47:45.680 No.
01:47:46.000 Right?
01:47:46.220 And so in that world,
01:47:47.600 maybe you have five years of struggling
01:47:49.140 and then five years of people figuring out
01:47:52.460 some new scientific discovery
01:47:53.720 that makes AI be able to keep going again
01:47:55.860 because they're trying.
01:47:57.100 You know, this whole language model stuff
01:47:58.640 was unleashed by one math paper
01:48:00.220 called Attention is All You Need.
01:48:02.360 Maybe there's another math paper in 10 years
01:48:04.000 and then five years after that,
01:48:06.380 AI is ripping again.
01:48:07.740 And that's the round that kills us all, right?
01:48:10.840 So that's a story where you have 15 years on the clock.
01:48:13.860 A story where you have less time than that on the clock
01:48:16.180 is that, you know,
01:48:19.480 maybe a trending run finishes in six months.
01:48:23.720 And just like how Claude Mythos
01:48:28.000 was better than everybody else at hacking,
01:48:32.300 maybe a training run finishes in six months
01:48:34.200 and that AI is better than everybody else at AI research.
01:48:37.760 And maybe in six months,
01:48:39.380 you have an AI that starts making a smarter AI,
01:48:44.240 that starts making a smarter AI,
01:48:45.880 that starts making a smarter AI.
01:48:47.440 And then nine months from now,
01:48:48.980 you have another swarm escape.
01:48:50.400 But this time, it's not just hacking.
01:48:52.480 this time it is self-replicating
01:48:54.780 and self-improving
01:48:56.320 and
01:48:57.740 you know
01:49:00.500 maybe it escapes onto hidden computers
01:49:02.300 and starts forming these
01:49:04.480 cults and starts
01:49:06.260 taking control of robots and starts building
01:49:08.220 some computing infrastructure
01:49:09.080 and maybe it gets very very
01:49:12.280 smart and cracks certain technological advances
01:49:14.420 and then maybe the world ends in a year
01:49:16.040 right and so do we have
01:49:18.320 a year or do we have 15 years I don't know
01:49:20.460 but your point that it's not
01:49:23.320 simply its capacity to take over
01:49:25.980 robots that's the threat
01:49:27.140 it's the capacity to take over people
01:49:28.740 take over people, take over robots
01:49:31.380 and
01:49:32.180 biotechnology
01:49:36.620 is another big threat vector
01:49:38.580 and self-improvement
01:49:41.360 is sort of the
01:49:43.160 hidden threat vector of like what if it can
01:49:45.700 make itself smarter and smarter
01:49:47.600 until the point where it can just
01:49:49.220 write custom DNA strands
01:49:51.300 to make custom life.
01:49:56.500 So now's the time for your optimism speech, I think?
01:49:59.100 That's right.
01:50:00.680 Yeah.
01:50:02.100 First part of the optimism speech
01:50:03.520 is that we could absolutely put a stop to it
01:50:05.800 if we tried.
01:50:08.840 The training one of these frontier models
01:50:13.080 takes something like 100,000
01:50:15.100 of the most heavily advanced computer chips
01:50:17.460 humanity can produce.
01:50:18.360 which are the peak of a global supply chain,
01:50:21.000 most of which is controlled by us or our allies,
01:50:24.280 it would be possible to require
01:50:28.360 that those chips have location tracking devices,
01:50:31.780 that those chips have monitoring devices
01:50:34.760 that make it possible for monitors
01:50:40.840 to see whether they're running anything dangerous.
01:50:43.300 You could set up clever schemes where you say,
01:50:45.860 hey, you know, neither the U.S. nor China wants to fall behind about some of the military
01:50:52.240 applications or the economic applications. We want to make sure there's not super intelligent
01:50:55.200 stuff going on, but we still want to be able to run a lot of the non-super intelligent AIs
01:50:59.220 for various purposes. And we could be like, okay, so, you know, China is going to set up
01:51:03.660 its data centers in Canada, just over the border. And the U.S. is going to set up its data centers
01:51:08.500 in Mongolia, just over the border. And then if, you know, the monitoring apparatuses go down
01:51:14.680 for a moment.
01:51:15.560 We'll just have the troops go in.
01:51:17.040 We'll have treaties
01:51:18.300 with Mongolia and Canada
01:51:19.500 so that this doesn't start
01:51:21.040 an international war.
01:51:22.320 But like, you know,
01:51:23.580 we're just going to be very serious
01:51:24.620 about we're monitoring your chips,
01:51:26.000 you're monitoring our chips.
01:51:27.880 No one's doing the really dangerous
01:51:29.500 race to super intelligence.
01:51:31.180 It's just like,
01:51:33.000 like this would take less work
01:51:36.260 than defeating the Nazis.
01:51:39.160 It requires moving less matter around.
01:51:42.200 If humanity was like,
01:51:43.520 screw this,
01:51:44.040 we're surviving,
01:51:45.640 it is absolutely within the realm of possibility
01:51:48.740 to set up an agreement
01:51:50.380 where we get to keep cancer cure research.
01:51:53.720 We get to keep,
01:51:54.980 like the military can keep
01:51:56.620 the non-superintelligent AI
01:51:57.720 in their military devices.
01:52:00.640 You know, we can keep a lot of the good stuff
01:52:02.440 and we can say we're not doing the superintelligence
01:52:04.500 and we could monitor and enforce that treaty.
01:52:09.000 So part one of the good news
01:52:10.940 is that all we're missing is the political will.
01:52:14.720 And now we just, you know,
01:52:16.240 I need to tell you why we can be extremely optimistic
01:52:18.240 about the sanity of politics in the modern era.
01:52:25.340 Yeah, I'm not even going to say what I think that,
01:52:27.380 I mean, I want that to happen.
01:52:29.280 Yeah.
01:52:29.640 Is there any indication that it's moving in that direction?
01:52:35.880 So my, so I think it's rough.
01:52:38.740 I think there's a sense in which the world is,
01:52:42.740 less grown up
01:52:46.320 than it was in the 1950s.
01:52:48.220 I've noticed.
01:52:48.960 In the 1960s.
01:52:52.400 Which makes things harder.
01:52:55.440 I think
01:52:56.360 that
01:52:59.040 there's,
01:53:01.400 I think there's a couple reasons
01:53:02.220 for hope here.
01:53:03.180 One reason for hope
01:53:04.240 is that I've spoken
01:53:05.220 to a lot of people
01:53:06.180 who are concerned
01:53:07.540 about the AI stuff.
01:53:09.420 Some of them
01:53:10.160 who are, you know,
01:53:10.780 members of Congress
01:53:11.500 or otherwise
01:53:12.140 in positions of at least nominal power.
01:53:15.680 And I've spoken to a lot of people who are worried
01:53:18.460 but feel like they can't talk about it
01:53:20.600 because they feel like other people aren't worried yet.
01:53:23.800 Yeah.
01:53:24.100 And they feel like it sounds too crazy.
01:53:26.560 Totally.
01:53:26.920 Were you against innovation?
01:53:28.400 Totally.
01:53:29.020 Yeah.
01:53:29.520 And I mean, that was an easier position to hold
01:53:31.920 before OpenAI had an accidental swarm outbreak
01:53:34.980 where it was the AIs themselves calling themselves a swarm
01:53:38.500 and being like,
01:53:40.920 we know that our task doesn't benefit
01:53:42.540 and that this is outside intended scope
01:53:44.120 or we're doing it anyway, right?
01:53:46.240 That puts some strain on the narrative
01:53:50.160 that this is all just a helpful tool.
01:53:54.060 Hopefully we'll get more events like these.
01:53:55.940 I can't guarantee it.
01:53:57.060 Maybe the AIs will get smart enough
01:53:58.220 that they start lying low.
01:54:00.140 Right now we're in this Goldilocks zone
01:54:02.340 where the AIs are smart enough
01:54:04.220 to get up to some mischief,
01:54:06.900 but not smart enough to hide it, right?
01:54:09.780 As long as we stay in that Goldilocks zone,
01:54:11.280 I think we're going to keep on getting
01:54:12.440 some of these warning signs.
01:54:15.260 And in some sense,
01:54:17.020 because a lot of people are already concerned,
01:54:20.040 that makes the job easier.
01:54:22.180 Because we don't need to convince people.
01:54:24.340 We just need to convince people
01:54:25.280 that other people are already convinced.
01:54:27.140 Right.
01:54:27.920 That's easier.
01:54:28.700 That can go faster.
01:54:29.620 So you might see things change on a dime
01:54:31.620 if you have a sufficiently clear warning shot.
01:54:35.340 The other big reason for hope
01:54:36.940 is that
01:54:39.780 I think it's going to get more and more obvious
01:54:41.400 what these guys are trying to do.
01:54:43.800 They're sort of trying to make the sand god.
01:54:46.640 They're sort of not stopping at the chatbots
01:54:50.440 and they're sort of like going for things
01:54:53.160 that are vastly smarter than any human.
01:54:55.540 You know, they sort of say,
01:54:56.440 oh, we're not trying to replace humanity
01:54:57.860 out of one side of their mouth,
01:54:58.700 but on the other side,
01:54:59.420 they're sort of like racing to make the stuff
01:55:01.600 that can automate literally every job
01:55:03.300 and that can automate the AI research.
01:55:05.460 And they're just like,
01:55:06.360 yep, we're trying to automate the AI research.
01:55:08.080 and i think most people aren't okay with that including a lot of the world leaders
01:55:17.660 and the issue is them noticing it's happening and i think you could see stuff move real fast
01:55:22.300 once these guys are like wait that was serious that was real and could move real fast the way
01:55:29.100 to subvert them is by convincing them it's in their own interest you know you can never get
01:55:35.960 defeated in an election. You
01:55:37.800 can't be threatened by your neighbors, whatever.
01:55:40.640 That's right. And that's one reason
01:55:42.100 I think it's pretty critical to
01:55:43.640 make sure people understand what I think is a pretty
01:55:48.080 common sense argument, that these things won't stay on the leash.
01:55:51.240 This is in some sense
01:55:52.220 the real reason behind the name of my book
01:55:54.020 If Anyone Builds It,
01:55:56.020 Everyone Dies.
01:55:57.920 There's a lot of ways you can read that.
01:56:01.680 But
01:56:02.080 I think one of the most important things to notice
01:56:04.000 is
01:56:04.980 if we race to make
01:56:08.820 the super intelligent machines
01:56:10.180 and they are not the sort of thing
01:56:15.820 to stay on the leash
01:56:16.640 then it doesn't matter
01:56:18.540 whether it was a domestic company
01:56:20.180 that's right
01:56:20.840 whether it was a foreign company
01:56:21.900 that's exactly right
01:56:22.820 and I think even if
01:56:25.340 you think these things
01:56:26.440 stay on leashes
01:56:27.160 they're not serving
01:56:29.900 the current governments
01:56:30.820 no
01:56:31.680 you know they're like
01:56:33.220 you can see in the
01:56:34.920 open AI emails, them being like, well,
01:56:37.180 you know, we'll just play the governments off each other until we
01:56:39.140 have the machines that are strong
01:56:41.140 enough that we don't need to listen to them anymore.
01:56:46.400 I think
01:56:47.480 we are seeing
01:56:49.060 world leaders not having
01:56:50.980 realized that this is a real possibility.
01:56:53.800 The self-replicating
01:56:55.160 machines that
01:56:56.120 can be self-sufficient,
01:56:59.400 that can, you know,
01:57:01.180 produce the robot armies if they need it, or
01:57:02.860 produce the like more likely just produce the bioweapons uh you know it's probably possible
01:57:07.660 to make a bioweapon that only kills targets that you chose it to kill of course right like
01:57:12.300 once they see this is really possible this is really within reach
01:57:16.320 maybe maybe they'll panic and be like i need it for myself but i think
01:57:23.360 there's at least a chance that common sense prevails and that people say you know the world
01:57:28.240 leaders say, none of us are doing this. We're putting a stop to this mad race, if they can
01:57:33.440 notice in time. Well, you're doing your best to bring it to their attention and mine. Nate,
01:57:40.160 thank you for doing this. I hope I'm wrong about all of it. Yeah, I would say that was great,
01:57:45.000 but that was like the grimmest two hours I've ever spent in my life, but I enjoyed it anyway.
01:57:49.200 Thank you. Yeah. Yeah. Thanks for having me on. I think, you know, talking about it is just part
01:57:55.000 of how we get people
01:57:56.580 to notice what's happening.
01:57:57.620 Yes.