The Tucker Carlson Show - September 11, 2026


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


Episode Stats


Length

2 hours

Words per minute

165.81

Word count

20,059

Sentence count

968


Transcript

Transcript generated with Whisper (turbo).
00:00:00.000 Bet mode activated.
00:00:01.820 The Scorebet app here with trusted stats and real-time sports news.
00:00:05.200 Yeah, hey, who should I take in the Boston game?
00:00:07.280 Well, statistically speaking.
00:00:08.840 Nah, no more statistically speaking.
00:00:10.600 I want hot takes.
00:00:11.700 I want knee-jerk reactions.
00:00:13.280 That's not really what I do.
00:00:15.440 Is that because you don't have any knees?
00:00:17.500 Or...
00:00:17.860 The Scorebet.
00:00:20.000 Trusted sports content, seamless sports betting.
00:00:22.320 Download today.
00:00:23.360 19 plus, Ontario only.
00:00:24.760 If you have questions or concerns about your gambling or the gambling of someone close to you,
00:00:27.840 please go to conicsontario.ca.
00:00:30.000 Nate thank you so much for doing this um you've written the world's darkest book you've devoted
00:00:41.000 your life to warning the world about the potential dangers of AI so let's just start by
00:00:48.560 hearing your explanation of why AI is dangerous in addition to just being annoying
00:00:53.960 yeah you know the the the very basic common sense point is
00:01:00.120 if you race to make machines that are much smarter than any human
00:01:07.160 to make machines that are more capable at inventing their own technology than humans are
00:01:13.040 and you race into this without really knowing what you're doing the most likely outcome
00:01:19.960 is just that the machines get loose
00:01:22.960 and do their own thing
00:01:23.920 and that humanity dies as a side effect,
00:01:26.980 just like humanity has, you know,
00:01:28.940 killed lots of other animal species,
00:01:30.500 not because we hate them,
00:01:31.500 but just as a side effect.
00:01:34.820 And in some sense,
00:01:37.140 a lot of people find it sort of obvious or intuitive
00:01:39.680 that if you just like make these really smart,
00:01:41.520 really powerful machines,
00:01:42.700 why would they care about us?
00:01:47.120 And I think that intuition basically is right.
00:01:49.460 and there's a lot of arguments
00:01:51.800 you can have on each side.
00:01:52.780 You can get in the technical details.
00:01:54.900 But my book is basically just
00:01:56.820 getting into those details
00:02:00.580 and saying like,
00:02:01.360 yep, it sort of holds up.
00:02:02.560 If we make smarter machines
00:02:04.240 without knowing what we're doing,
00:02:05.060 it's just going to go poorly.
00:02:09.620 Are we actually going to make machines
00:02:12.140 capable of what you're describing?
00:02:15.100 You know, the companies are trying to.
00:02:17.800 They talk about how they are pursuing
00:02:21.480 superintelligence in the true sense of the word.
00:02:24.080 That's Sam Altman's phrase.
00:02:25.680 Dario Modi of Anthropics says
00:02:27.400 they're trying to make the equivalent
00:02:28.540 of a country worth of geniuses in a data center.
00:02:31.880 So these guys are really,
00:02:33.060 you know, these aren't chatbot companies.
00:02:36.440 They didn't set out to be chatbot companies.
00:02:39.040 They set out to make these machines
00:02:40.820 that can sort of exceed humans in every way.
00:02:43.720 And that's what they're targeting.
00:02:45.120 There's a separate question
00:02:46.120 of whether they will get there.
00:02:47.120 ask you a question, why would anyone want to build a machine smarter than people?
00:02:52.220 I think that a lot of them hope that they'll be able to make the world much, much better.
00:03:00.200 You know, they hope for a cure to cancer, and then more than a cure to cancer, they hope for a cure
00:03:05.260 to aging. They hope for, you know, a thousand years worth of technological development compressed
00:03:12.740 into two years.
00:03:16.400 And I sort of don't think
00:03:18.660 that they're going to be able
00:03:21.940 to get that really
00:03:23.580 or harness that for good ends.
00:03:26.260 But that's what I think
00:03:27.000 they're shooting for.
00:03:29.120 So they're not,
00:03:30.360 but I mean,
00:03:31.440 the core point you're making
00:03:32.320 is these companies
00:03:33.260 did not set out
00:03:34.240 to make consumer products,
00:03:35.980 like to make your life better
00:03:37.160 necessarily in the short term
00:03:38.580 with a more efficient search engine.
00:03:41.600 That's right.
00:03:42.300 Yeah.
00:03:42.420 Yeah, you know, OpenAI started before the chatbots, the large language models, were even a thing.
00:03:50.100 They were started, I forget if it was late 15 or early 2016, but the paper that unlocked the most recent wave of AI came out in 2017, which was after OpenAI was founded.
00:04:01.360 and um these guys are that the the large language models are a surprise revenue stream
00:04:12.500 and that revenue stream can fund the creation of even more even larger data centers for the next
00:04:19.400 level of the technology but these guys have their eye on uh the sort of ultimate version of the
00:04:25.920 technology which is these machines much smarter than humans and that's sort of the ultimate version
00:04:30.020 because once the AIs are smarter than the humans,
00:04:32.840 the AIs can carry on the AI research
00:04:34.960 and make the next generation of the AIs
00:04:36.640 which make the next generation of the AIs.
00:04:39.040 And that's sort of what they're shooting for.
00:04:42.240 What is superintelligence?
00:04:45.260 We, in our book, define superintelligence as an AI
00:04:48.340 that is better than the best humans at every mental task.
00:04:54.700 So anything that a human can do purely mentally,
00:04:58.440 the AI can do that better.
00:05:00.020 One thing people often get caught up on about this is that includes the AI being better at things like persuading humans, at things like charisma.
00:05:09.500 You know, we often think of intelligence as the stuff that the nerds have and that the jocks like.
00:05:15.400 You win the chess match.
00:05:17.000 Right. It's the chess guys rather than the politicians guys or the rock stars.
00:05:21.720 But that's not really what the intelligence in artificial intelligence means.
00:05:27.120 the intelligence and artificial intelligence is about the stuff that humans have and that mice
00:05:32.320 don't it's it's sort of the whole package you know like the the politician who's very charismatic
00:05:40.180 it's not it's not that like chess playing happens in your brain and charisma happens in your kidneys
00:05:45.540 right they're sort of both mental functions yes uh and so the the sort of super intelligent
00:05:54.380 AIs are not just super good at playing chess. They're also good at super persuasion and they're
00:05:59.720 super good at the research and at the technological invention. So super intelligence, as you define
00:06:06.560 it, is a machine that is better at every mental function than any human being. That's right.
00:06:14.300 And, you know, the definition, that doesn't mean that nothing crazy will happen on this planet
00:06:23.540 until the AIs are super intelligent.
00:06:26.020 Super intelligence in this definition
00:06:27.380 is sort of a point where past this point,
00:06:29.740 things must be pretty crazy
00:06:31.020 because now the AIs can do automated AI research
00:06:33.940 and can make smarter AIs
00:06:35.020 and they can figure out how to run the robot factories
00:06:36.620 and they can build the better robots and all that.
00:06:39.920 You could have things start to get crazy
00:06:42.020 before you have a super intelligence in this sense.
00:06:44.660 You could have AIs that are much better than humans
00:06:47.600 at some tasks and much worse than others
00:06:49.240 that are still causing all sorts of crazy happenstances.
00:06:53.540 uh the the super intelligence is sort of a like once you get here stuff's crazy point it's not a
00:07:00.560 things will stay normal until you get here point
00:07:03.000 so why is super intelligence the significant milestone that you're worried about
00:07:09.260 um i mean i would say that uh i'm also worried about what will happen before that milestone
00:07:18.320 It's more like having the definition makes it easy to talk about
00:07:22.460 how crazy would things get once you are past this point.
00:07:29.180 Or sort of like, if we assume that the machines get here, what happens?
00:07:34.380 And then the answer is, it would be pretty crazy.
00:07:37.100 It would be pretty wild.
00:07:38.760 And it sort of lets you factor the conversation
00:07:43.020 into like, how to say,
00:07:48.820 I think in the artificial intelligence conversation,
00:07:51.100 there's actually a lot of conversations going on.
00:07:53.380 One conversation is like,
00:07:54.840 can the machines really get smarter than the humans?
00:07:57.140 One conversation is like,
00:07:58.740 how fast can we get there?
00:08:00.440 Is the current technology on that route?
00:08:03.200 Another conversation is,
00:08:04.400 what happens if we do get there?
00:08:06.640 Like, you know, would the AIs care about us?
00:08:09.620 Would they not care about us?
00:08:10.860 What would they be able to,
00:08:11.600 Like, what would they try to do?
00:08:13.600 There's another question, which is,
00:08:14.560 what would they be able to do, right?
00:08:16.280 And these are all sort of different conversations about AI.
00:08:18.660 And the superintelligence definition,
00:08:20.740 it's not really like, here's where all the fears hang.
00:08:23.740 It's more like it lets us break those pieces out separately
00:08:29.020 and be like, well, is superintelligence possible
00:08:31.560 separate from, well, what would happen if you had one?
00:08:35.420 What would happen if you had one?
00:08:37.180 Most likely outcome, I think destruction of the planet.
00:08:40.240 so our friends at preborn who are actual friends by the way just asked us to thank the listeners
00:08:44.860 of this show for supporting the pro-life cause a lot of people do all of a sudden trust me
00:08:51.220 because what's happening is crazy and people are waking up to that the president of the group is
00:08:57.040 our friend dan steiner and he wants americans to know how much of a difference they have made
00:09:02.080 this year's donations have helped saved 50 000 babies from abortion that's enough to fill a
00:09:07.740 professional baseball stadium, 50,000 lives saved, as well as generations that will follow them.
00:09:13.120 Preborn wastes zero money. We're not big on charities because a lot of them waste money.
00:09:18.480 Preborn does not. It uses every tax-deductible gift to rescue babies and share the hope of Jesus
00:09:23.800 with their moms. In 2026 alone, over 8,000 women made the decision to follow Christ.
00:09:29.680 The generous support of Preborn's donors made that all possible. $28 to provide one
00:09:36.140 life-saving ultrasound, just giving expectant mothers the knowledge that they have a baby
00:09:41.600 that they're carrying. $15,000 could place an ultrasound machine in a pregnancy clinic
00:09:45.940 where it can save countless lives. Dan Steiner's message is really simple. Thank you for helping.
00:09:51.060 Your generosity is literally saving lives. To help Preborn save more, dial pound 250 and say
00:09:56.560 the keyword baby. That's pound 250 baby, or visit preborn.com slash Tucker, preborn.com.
00:10:03.760 Have you ever heard of the Pelosi Stock Tracker? Well, here's the stock trading app behind it.
00:10:08.700 Most of us don't have access to classified information, yet our politicians seem to.
00:10:13.880 For years, they've been publicly trading stocks, with some being oddly timed around certain
00:10:18.640 legislation. Well, this team said, if you can't beat them, join them and built Autopilot. It
00:10:25.060 connects directly to your brokerage account, like a Robin Hood or Schwab, and mimics Nancy Pelosi
00:10:29.960 and other politicians' stock trades automatically.
00:10:32.960 So instead of debating what stocks to buy,
00:10:35.200 you just trade alongside them hands-free.
00:10:37.920 And the key thing here, your money stays in your own account.
00:10:41.420 The app already has $1.3 billion invested so far.
00:10:45.420 So if you wanted to check it out,
00:10:47.120 search Autopilot in the App Store
00:10:48.760 or go to joinautopilot.com.
00:10:51.380 That's joinautopilot.com.
00:10:53.880 Investing has risks like the loss of principle.
00:10:56.360 Why are men coming to a urologist
00:10:58.600 to talk about their marriages?
00:10:59.960 I'm Dr. Justin Hooman, a urologist practicing in Beverly Hills.
00:11:03.840 And the answer is because ED doesn't stay in the bedroom.
00:11:07.200 I get it.
00:11:07.880 This isn't just physical.
00:11:09.120 It's personal.
00:11:09.920 That's why I recommend Rougiet Ready, a three-in-one treatment that works on your body and your
00:11:15.200 brain.
00:11:15.740 Fast acting, long lasting, prescribed by a licensed physician.
00:11:19.640 Get started at rougiet.com slash Justin today.
00:11:22.980 That's R-U-G-I-E-T dot com slash Justin.
00:11:26.640 Slash Tucker.
00:11:28.160 What would that look like?
00:11:29.960 So, I'm going to be a little bit annoying here and say a couple caveats first, because
00:11:38.420 it's sort of a tricky one to sort of predict things that are smarter than us.
00:11:43.620 Yes.
00:11:45.180 And the first annoying caveat I will give is if you go to play a game of chess against
00:11:52.580 Magnus Carlsen, I predict you will lose.
00:11:56.940 No offense.
00:11:57.760 he's just the best human chess player alive
00:12:00.760 if you then ask
00:12:04.880 what piece will he use
00:12:06.860 to checkmate me
00:12:07.560 I'm like well that's a much harder question
00:12:10.440 it sort of is like very easy to predict
00:12:12.800 the winner it's hard to predict
00:12:14.520 the exact methods
00:12:15.480 so my prediction
00:12:18.680 that humanity ultimately dies
00:12:20.480 is a different sort of prediction
00:12:24.500 than my prediction about like it could go
00:12:26.560 this way it could go that way
00:12:27.720 so I'll give you some stories
00:12:30.260 but these are stories
00:12:31.260 that are like
00:12:32.340 well maybe Magnus
00:12:33.240 will like
00:12:33.840 fork your queen
00:12:35.380 and your rook
00:12:36.440 with his knight
00:12:37.100 and then finally
00:12:38.640 use the queen
00:12:39.120 to checkmate you
00:12:39.700 like yeah that could happen
00:12:40.640 but it's a different
00:12:41.760 sort of prediction
00:12:42.460 it's a guess
00:12:43.040 right
00:12:43.380 the guess here
00:12:46.640 is
00:12:47.980 that
00:12:51.260 if we manage
00:12:52.500 to make these
00:12:52.920 really smart AIs
00:12:53.640 they will
00:12:55.320 have their own
00:12:56.660 stuff that they're pursuing
00:12:59.020 which is not quite what we wanted
00:13:01.000 not quite what we asked for
00:13:02.040 it'll have this other strange stuff
00:13:03.780 there's a whole discussion
00:13:06.020 about why that happens
00:13:06.800 but we're already starting to see it
00:13:07.980 in practice with some of the recent events
00:13:09.400 and
00:13:12.780 they would be able to pursue
00:13:17.540 whatever it is they're pursuing
00:13:19.200 much more efficiently
00:13:22.000 than humanity can
00:13:24.340 and so we're talking about
00:13:26.380 automated factories that produce the robots, that produce the factories in a fully automated
00:13:31.360 supply chain. And then if the AIs have anything they're trying to do that they can do more of
00:13:39.040 with more resources, they start gobbling up the resources on the planet. Sort of like how humanity
00:13:45.140 spread and started gobbling up all the resources on the planet. And the sort of most basic thing
00:13:53.080 to visualize here might be you have factories that produce robots that produce factories that
00:13:57.820 produce robots that also build data centers and you just have these you know fully autonomous
00:14:02.140 self-replicating ecosystems of robots and factories and data centers that don't care
00:14:08.700 about the instructions human humans gave them and that cover the planet take all the resources
00:14:15.600 take all the sunlight take all the places we were growing crops probably raise the temperature of
00:14:20.260 the planet because uh you can compute more efficiently well technically the earth radiates
00:14:25.060 more heat when the world is hotter and that's lets you sort of do more computation and then
00:14:30.240 sort of you know it's good for the machines to have a hot planet it's good for the machines to
00:14:33.720 have a hot planet yeah um the sort of physical limits on how much computing uh you can run on
00:14:39.300 the surface of earth is bounded by how much heat you can radiate to space that's sort of the first
00:14:43.920 constraint you bump up against you might think that it's energy but actually there's a lot of
00:14:47.620 helium hydrogen diffuse on this planet. So you can get plenty of energy. What you need is
00:14:52.300 heat dissipation. And the world can dissipate more heat when it's hot. So if you're imagining
00:15:02.460 some collective of AIs that are trying to run a lot of computing power, trying to do a lot of
00:15:06.840 computation for one reason or another, they sort of prefer the planet running hotter.
00:15:11.900 And so it's sort of, it's like nothing personal. It's just, if you let these AIs get out of
00:15:15.640 control uh they transform the planet into something unlivable how hot uh as hot as
00:15:27.240 uh you can while still having the computers not melt so probably hundreds of degrees
00:15:34.500 um at which point you hear people say well then just turn it off
00:15:40.540 uh so i think we do have an opportunity to turn it off you know and i'm not here saying that
00:15:47.920 we're going to die i'm here saying we sort of are going to need to act you got to be careful
00:15:52.120 about the turning it off piece because um your opportunities to turn the ai off only last when
00:16:00.980 the ai is running on the computers you know it's running on if you have the ais breaking out and
00:16:06.340 running on hidden computers.
00:16:08.320 If you have the AI's
00:16:09.460 running robot factories
00:16:10.540 that produce robots
00:16:12.700 that are under the control
00:16:13.740 of the AI,
00:16:14.680 where those robots
00:16:15.380 can then go build
00:16:16.100 more computers
00:16:16.680 that are not hooked up
00:16:17.280 to your network
00:16:17.940 that can run the AI's.
00:16:20.160 These are sort of thresholds
00:16:21.600 where the AI is able
00:16:22.800 to keep itself running.
00:16:25.100 And you better have turned it off
00:16:26.580 before that point.
00:16:30.940 Because after that point,
00:16:32.440 it's impossible.
00:16:34.660 Yeah, I mean...
00:16:36.340 The, if, it also gets a lot harder
00:16:43.520 if the AI knows you're going to try to shut it off
00:16:45.620 and is going to try to stop this, right?
00:16:47.880 You got to remember that we're talking about things
00:16:50.060 that are smart here.
00:16:51.680 And so if the AI sort of sees it coming,
00:16:53.300 maybe it defects to North Korea,
00:16:55.240 where it can, you know, convince them
00:16:57.320 to run it on its data centers
00:16:58.460 in ways that initially benefit them,
00:17:01.000 but that ultimately benefit the AI.
00:17:03.040 I think we're going to have to redefine what life is
00:17:05.640 because you're describing
00:17:06.540 a living autonomous thing.
00:17:09.300 Yeah, artificial life in a sense.
00:17:11.720 I mean, in some sense,
00:17:12.600 we're already seeing
00:17:13.640 the very, very beginnings of that.
00:17:17.660 But yeah, once the AIs can replicate,
00:17:21.840 once they can improve themselves,
00:17:25.000 once they can find ways to run
00:17:29.640 where you don't know that they're running,
00:17:30.980 once they can
00:17:33.600 sort of
00:17:37.300 run the robot factories
00:17:39.380 to produce the robots
00:17:41.200 that can produce more factories
00:17:42.420 and that can produce computers
00:17:43.520 that the AIs can run on,
00:17:44.740 then yeah,
00:17:45.160 you have in some sense
00:17:46.040 made a new artificial life form.
00:17:53.500 And, you know,
00:17:54.180 humanity is on the top
00:17:54.920 of the food chain right now
00:17:55.840 because we
00:17:56.920 are sort of the only
00:17:59.280 smart life form around.
00:18:00.980 Yeah. If you suddenly make a new one that is smarter, that is able to make a million copies
00:18:07.480 of itself, that is able to run faster, it's just kind of a crazy thing to race into.
00:18:14.540 It's a weird thing to want. And as it's developed, I was going to say slowly, but it hasn't been
00:18:22.360 particularly slow 10 years the rest of us have watched as people who i mean put it in political
00:18:32.160 terms don't share the values of most americans are now in charge of this and it's almost like
00:18:37.860 everyone sat passively by as it happened yeah i i think i think there's a lot going on there
00:18:44.960 i think part of it is that most people didn't and some still don't believe that ai was going
00:18:50.600 to be able to keep going. I think there's a lot of like putting the head in the sand and being
00:18:55.520 like, oh, it's, you know, just slop. It's going to be a bubble. It's going to pop. It's going to
00:18:59.780 hit a wall. It's not going to be able to improve anymore. And I think a lot of that was sort of
00:19:07.280 wishful thinking. And I think that at the very least, if people are like, we are making, you
00:19:13.120 know, the super intelligent machines that are going to replace humanity as the top of the food
00:19:17.220 chain because we think it's going to go great i think humanity's response should not be go ahead
00:19:23.620 and try we hope you'll fail i think the response should be like hold on that's kind of crazy um
00:19:33.460 yeah i had i had some other point too but i've forgotten it so how uh predictable has the
00:19:38.340 evolution of ai been has it taken turns that you didn't expect yeah totally uh i was
00:19:47.220 So back before the large language models.
00:19:50.380 How long have you been following this issue?
00:19:52.220 I started following it in 2012 and I started working with some of the people
00:19:58.320 trying to make this go well in 2013 and I started full-time in 2014.
00:20:03.260 A long time.
00:20:04.240 So over a decade.
00:20:05.760 What has surprised you?
00:20:07.880 You know, the large language models,
00:20:10.200 they have gone further than I expected initially.
00:20:15.360 and I think one of the big surprises here
00:20:19.000 was AI undergoing a phase
00:20:21.400 where everybody can see it.
00:20:25.060 You know, back before the large language models,
00:20:28.620 really only nerds paid attention to AI.
00:20:33.040 And it wasn't that there was nothing happening.
00:20:34.960 You know, we were watching the, you know,
00:20:37.680 Google DeepMind make an AI
00:20:39.940 that could beat the best Go player.
00:20:44.020 Yes.
00:20:44.200 And that was sort of a milestone for people who were paying attention.
00:20:48.620 But for all we knew, the labs were going to keep on working on engineering problems and keep on working on the relatively nerdier problems.
00:20:57.780 And you were never going to have like a mass market consumable product.
00:21:01.760 And so it was like, for all we knew, it would just stay in the labs and no one would ever really notice that these guys were gambling with the whole future.
00:21:08.520 Now, at least, AI is everywhere.
00:21:11.800 Everyone's starting to have the conversation.
00:21:13.160 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.
00:21:24.800 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.
00:21:34.280 Every college commencement speaker who mentions it gets booed.
00:21:37.440 Yeah.
00:21:37.680 And that view, that opposition to it has had no effect at all in slowing it down.
00:21:45.860 I don't know.
00:21:46.620 Does that give you hope?
00:21:50.480 I mean, I hear this a lot.
00:21:55.100 I would say when you're coming at this from 2012, it really feels like we're making progress.
00:22:00.320 Right.
00:22:00.580 No, that's fair.
00:22:01.400 It doesn't feel like we're all the way there yet.
00:22:02.860 um but can i just make the point china china china china china we have to do this because
00:22:10.040 of china yeah you know i think uh china and the usa have a shared interest in not dying
00:22:17.740 to a rogue super intelligence storm season is on the way and that means your wi-fi could go down
00:22:23.420 and if it does your security cameras go dark too so at the exact moment you want to survey
00:22:28.940 your property during the blackout, you can't. Unless you install cameras by Defend. Defend
00:22:34.800 cameras run on cellular, not Wi-Fi, and that means they keep working even when you lose power.
00:22:40.780 You don't choose a carrier. You don't sign a contract. You just mount it, open the Defend
00:22:44.540 app, and bam, you're connected. It's that simple. So whether it's your ranch, your job site,
00:22:49.280 your backyard, Defend works where Wi-Fi doesn't and keeps working when the grid goes down,
00:22:54.920 which it may.
00:22:56.240 Plans start at just $5 a month.
00:22:57.920 Most people pay around $10.
00:22:59.140 Plus, there are no binding contracts,
00:23:00.740 meaning you get to cancel whenever you want.
00:23:02.460 So unlike big tech,
00:23:03.720 Defend will never sell your data,
00:23:05.060 not to China, not to the government,
00:23:06.260 not to anybody.
00:23:07.240 It's an American company
00:23:08.220 built for people who want actual security.
00:23:11.400 Visit DefendCellCam.com
00:23:13.660 and use the code Tucker for 20% off.
00:23:15.760 That's DefendCellCam.com, code Tucker.
00:23:20.320 The ScoreBet app here
00:23:21.700 with trusted stats and real-time sports news.
00:23:24.080 Yeah, hey, who should I take in the Boston game?
00:23:26.140 Well, statistically speaking.
00:23:27.700 Nah, no more statistically speaking.
00:23:29.460 I want hot takes.
00:23:30.560 I want knee-jerk reactions.
00:23:32.140 That's not really what I do.
00:23:34.300 Is that because you don't have any knees?
00:23:36.340 Or...
00:23:36.700 The Scorebet.
00:23:38.860 Trusted sports content, seamless sports betting.
00:23:41.180 Download today.
00:23:42.220 19 Plus, Ontario only.
00:23:43.620 If you have questions or concerns about your gambling or the gambling of someone close to you,
00:23:46.760 please go to conicsontario.ca.
00:23:49.640 Want to see your rewards go further?
00:23:52.260 Now at Shell.
00:23:52.900 Scene Plus members can fill up on points at the pump, on snacks, car wash, and more.
00:23:58.560 Plus, Scotiabank and Tangerine cardholders can get up to 10 cents per liter in value with a linked card.
00:24:05.140 New rewards partners, new ways to save and earn at Shell.
00:24:08.700 Get more, go further.
00:24:10.660 At participating Shell locations, conditions and limits apply.
00:24:13.900 Actual value may be lower.
00:24:15.420 Visit shell.ca slash loyalty for full details.
00:24:18.160 what is it conceivable that like let's just say china can i'm sure by the way that china
00:24:26.260 is showing more strength than we are in this yeah i mean and if nothing else they sort of
00:24:32.260 uh have a lot more reason to censor their ais right and to try to make them you know not say
00:24:37.800 certain things to the to the population so um but if the united states you know if the u.s
00:24:44.580 government were somehow able to get control of the tech sector which is at present not possible
00:24:50.660 but let's say it did let's say the president was more powerful than the tech oligarchs which he is
00:24:56.780 not but for the sake of argument let's see he was and he shut it down would that matter if if some
00:25:04.440 indian lab or chinese lab created it so it does have to be a global stop to this creation of
00:25:09.900 superintelligence. I think there's a couple of reasons why this is possible. One is, I think a
00:25:15.540 lot of people, they hear that some parts of AI need to be stopped. And they think, I'm saying
00:25:25.240 that all parts of AI need to be stopped. And, you know, there's all sorts of interesting issues
00:25:32.100 with, you know, AI and education and AI and, you know, military drones that are, that society has
00:25:39.420 to wrestle with. But these are all sort of separate issues from the superintelligence race.
00:25:44.040 And if you sort of like tease those issues apart, it becomes much more possible to do like a more
00:25:50.900 surgical intervention on we're not going to race towards the superintelligence in ways that leave
00:25:55.080 a lot of the rest of the sector alive, which makes it sort of an easier coordination effort
00:26:06.700 to attempt
00:26:08.240 and on that front
00:26:10.620 on the super intelligence front
00:26:11.800 it's not that you know
00:26:14.440 just shutting down the domestic
00:26:16.380 race towards super intelligence would be enough
00:26:18.540 but
00:26:19.700 racing towards super intelligence
00:26:24.480 requires a huge amount of
00:26:26.560 highly advanced computer chips
00:26:27.980 that can only be produced
00:26:30.400 as sort of the peak of
00:26:32.560 the global supply chain
00:26:33.720 which is largely
00:26:36.260 controlled by U.S. allies. And so there is absolutely a possibility that a U.S.-led
00:26:46.000 coordination effort could say, look, we're not doing the superintelligence thing. We're going to
00:26:52.760 monitor the extremely heavy chip concentrations of like, you know, 10 or 100,000 of these highly
00:26:57.940 advanced AI chips. We are going to make sure that they are not doing these superintelligence
00:27:06.060 training runs. And I think there's a way, if the U.S. was leading that effort, there would be a way
00:27:15.520 to get China on board and set this up in a way that was monitorable, enforceable, verifiable.
00:27:21.460 Yes.
00:27:21.980 We sort of, in many ways, it would be easier than a nuclear arms treaty because uranium is a rock
00:27:28.020 you can dig out of the ground. Whereas these highly advanced computer chips sort of only
00:27:31.440 come out of one fab in Taiwan, you know? And so it's like-
00:27:34.200 Use your TSMC chips.
00:27:35.340 Yeah. And so, and there's other parts of the supply chain that are very narrow, like the lithography machines that come out of the Netherlands. And so, yeah, we could absolutely lead a global effort to say we're not making the super intelligent machines. And it would be a bit tricky, but it's just, we could just do it if we had the political will.
00:27:56.540 i think that people either don't know what's happening or they assume that like a lot of
00:28:06.620 fears this fear will turn out to be groundless people point to y2k yeah um you know i hope the
00:28:16.360 fears are groundless i think with a lot of past fears uh the what happened is not that the fear
00:28:23.520 was groundless it's that uh people noticed the issue and put in a ton of work to make the bad
00:28:30.440 thing uh not happen you know i think we saw this with the hole in the ozone layer where people are
00:28:37.360 like oh whatever happened to the hole in the ozone layer well what happened with the hole in the ozone
00:28:40.000 layer is that we fixed it it's not that it was fake it's that you know we went and banned the
00:28:44.740 chlorofluorocarbons that were actually blowing this hole in the ozone layer we found other ways
00:28:48.280 to cool refrigerators
00:28:50.520 that worked similarly well
00:28:52.760 and didn't put this hole in the ozone layer.
00:28:57.420 I think Y2K was actually one of these cases
00:28:59.220 where you had a ton of software engineers
00:29:01.200 up in 1999 and a couple of years prior
00:29:05.340 that were scrambling to update all of the software
00:29:08.140 so that they would be able to handle dates past 1999.
00:29:13.520 And they got it done in time.
00:29:16.580 None of the big systems went down.
00:29:17.720 but it wasn't that
00:29:19.220 you know
00:29:19.680 the issue was fake
00:29:21.200 it took a lot of work
00:29:22.840 it's just that
00:29:24.140 that work happened
00:29:24.740 behind the scenes
00:29:25.440 are people doing work
00:29:28.000 to slow down AI
00:29:28.780 I'm doing my best
00:29:31.800 it's
00:29:32.660 it's not really there yet
00:29:34.680 you know
00:29:35.220 if I
00:29:37.600 if I can belabor
00:29:38.360 this point a little
00:29:38.980 because I think
00:29:39.320 it's kind of
00:29:39.600 an important point
00:29:40.240 just with more
00:29:42.680 historical cases
00:29:44.640 which maybe
00:29:45.460 it'll bore everybody else
00:29:46.240 but it'll
00:29:47.080 it'll entertain at least us it'll definitely entertain me you know you have if you look
00:29:52.240 back across history you see there's definitely been some warnings that didn't come to pass
00:29:57.580 right there were i think they were called the masonites in like the 1960s or sorry in the
00:30:02.520 1860s maybe there's 1880s i forget exactly but there were the masonites who are an end of the
00:30:06.480 world cult yeah uh and you know it the world did not end right but also in the 1880s you had
00:30:15.260 Otto von Bismarck saying, you know, Europe is a tinder keg and some damn fool thing in the Balkans
00:30:19.700 is going to light it, right? That did happen. It sure did. You had, you know, in the 1920s,
00:30:26.460 you had scientists saying, don't put lead in the gasoline, it'll poison children.
00:30:31.260 And we put lead in the gasoline and it poisoned a lot of children. And we said, whoops,
00:30:34.060 and we took the lead back out. The world is a place where a lot of people make a lot of types
00:30:43.780 of warnings. And some of them are real, like the gasoline, and some of them are fake, like the
00:30:48.260 Masonites. And so you sort of can't use a rule that's like, every warning is real, and we need
00:30:55.280 to listen to it. And nor can you use a rule that's like, every warning is fake, and we dismiss it.
00:30:59.960 You sort of just have to look at the details to figure out whether this one is a real one.
00:31:06.900 And then the last thing I would say on this point is a lot of people in the 50s warned of nuclear Armageddon.
00:31:23.280 Yes.
00:31:25.060 And it sort of makes sense if you look at the history leading up to that point.
00:31:29.900 Humanity had sort of never before actually failed to use their strongest weapons in combat.
00:31:37.720 And, you know, these people were living in a world
00:31:39.980 that had had World War I and then League of Nations
00:31:43.240 and, you know, the world swore never again.
00:31:46.440 And they tried to invent these whole new governance structures
00:31:48.540 to prevent it from happening again.
00:31:49.560 And that immediately failed in London World War II.
00:31:52.800 And so, you know, in 1950, you're sort of in this world of like,
00:31:55.820 it just looks pretty grim.
00:31:58.740 And we haven't had nuclear Armageddon yet.
00:32:01.620 And it's not because the bombs were fake.
00:32:03.780 it's not because the nukes were hype. It's because people saw the issue and worked really hard to
00:32:14.140 avoid it day in and day out for decades through crises and succeeded. And so I think one of the
00:32:21.820 big ways you can tell the difference between someone coming in and proclaiming that the
00:32:26.740 apocalypse is nigh and someone who's saying, look, we have a problem and we need to fix it
00:32:30.780 is whether that person is saying, we're definitely screwed, or is whether that person is saying,
00:32:36.420 here's a problem, let's try to fix it. And the thing I remind people about here is that
00:32:42.540 the very first word in the title of my book is if. I'm not here saying we are going to die.
00:32:49.320 I'm here saying, if we race down this path, we die. But that same if is what points the way
00:32:57.380 towards changing paths.
00:32:59.460 And we still have plenty of time to change paths.
00:33:02.380 Home ownership has always been
00:33:03.320 at the center of the American dream.
00:33:05.080 It is an actual asset,
00:33:06.920 a place people work their entire lives to build,
00:33:09.480 a place of stability and security and real value.
00:33:13.120 Right now, though, too many homeowners
00:33:14.720 are stuck swiping credit cards to cover groceries.
00:33:17.720 There's basically a survival tax
00:33:19.400 of 20% interest or more.
00:33:21.140 So why keep handing your money to the banks
00:33:23.700 when you could keep it for your family?
00:33:25.320 Interest is what kills you.
00:33:27.340 Our friends at American Financing have a better way.
00:33:29.180 They're helping homeowners use their hard-earned equity,
00:33:31.740 and they're supporting Americans looking to fulfill the dream of owning a home
00:33:34.840 with mortgage rates currently in the fives.
00:33:37.920 American Financing saves its customers an average of $800 a month.
00:33:41.280 That's about $10,000 a year.
00:33:43.640 It's not just a loan.
00:33:44.780 It's a true financial reset.
00:33:47.300 We recommend American Financing.
00:33:48.940 They're called America's Home for Home Loans for a reason.
00:33:52.920 So get out of credit card debt.
00:33:54.980 I mean, it is just crazy to borrow money.
00:33:57.720 No judgment, but it doesn't make any sense to borrow money at that rate.
00:34:01.300 If you want out, call 800-685-5696.
00:34:06.300 800-685-5696 or visit AmericanFinancing.com.
00:34:12.540 The ScoreBet app here with trusted stats and real-time sports news.
00:34:16.280 Yeah, hey, who should I take in the Boston game?
00:34:18.320 Well, statistically speaking.
00:34:19.840 Nah, no more statistically speaking.
00:34:21.660 I want hot takes.
00:34:22.760 I want knee-jerk reactions.
00:34:24.360 That's not really what I do.
00:34:26.500 Is that because you don't have any knees?
00:34:29.900 The Scorebet.
00:34:31.060 Trusted sports content, seamless sports betting.
00:34:33.380 Download today.
00:34:34.440 19 Plus, Ontario only.
00:34:35.840 If you have questions or concerns about your gambling or the gambling of someone close to you,
00:34:38.960 please go to conicsontario.ca.
00:34:41.840 Want to see your rewards go further?
00:34:44.200 Now at Shell, Scene Plus members can fill up on points at the pump,
00:34:47.840 on snacks, car wash, and more.
00:34:50.760 Plus, Scotiabank and Tangerine cardholders can get up to 10 cents per liter in value with a linked card.
00:34:57.280 New rewards partners, new ways to save and earn at Shell.
00:35:00.900 Get more, go further.
00:35:02.860 At participating Shell locations, conditions and limits apply.
00:35:06.100 Actual value may be lower.
00:35:07.620 Visit shell.ca slash loyalty for full details.
00:35:11.440 Net slash Tucker.
00:35:12.500 i'm struck by how little of this was planned by anybody by how little effect the government has
00:35:26.020 had on it not that i'm for the government i'm pretty opposed to the government most of the
00:35:29.380 time but i'm also for people being able to control their own country and the only mechanism by which
00:35:34.100 they can do that is voting so it it's like these tech companies act independent of what the
00:35:40.640 population wants what the government wants it doesn't yeah right they're more powerful than
00:35:45.300 the government i guess that's the point i'm making in some sense um i i think that they have that
00:35:51.040 power more and more when people don't really understand what they're doing you know we saw
00:35:55.980 just in june um there was uh there was an ai made by anthropic called claude mythos that was very
00:36:03.700 cyber capable and they released a version of it that was supposed to have more card rails called
00:36:08.240 Cloud Fable, and
00:36:10.260 it sort of turned out that you could
00:36:12.140 jailbreak it to get some of those cyber capabilities.
00:36:15.900 Can you explain what all that means?
00:36:18.180 Yeah, so
00:36:18.820 how to say,
00:36:22.560 in January...
00:36:24.400 Cyber capable means it can find the internet?
00:36:27.320 Cyber capable means it can
00:36:28.160 hack into basically anything.
00:36:30.000 Yeah, so
00:36:30.520 one of the holy grails of hacking is
00:36:34.040 can you make a website
00:36:35.520 where if you just look at the website
00:36:37.960 I get full control over your computer or your phone.
00:36:41.540 That's very hard to do.
00:36:43.280 Usually you need to like click some link
00:36:45.340 and download something or run something.
00:36:48.320 And, you know, I don't know the exact numbers
00:36:50.700 because I don't have the security clearance
00:36:51.720 to know the exact numbers,
00:36:52.720 but a decent guess is that in January of this year,
00:36:55.720 there were only two entities that could pull that off.
00:37:00.560 Mossad and the NSA.
00:37:03.320 In March of this year, there were three.
00:37:06.140 Mossad, the NSA, and Claude Mythos,
00:37:09.620 which is this new AI made by Anthropic.
00:37:12.740 And so it sort of became superhumanly capable at hacking.
00:37:20.180 And it sort of became that way overnight.
00:37:24.760 You know, like really there was six months to a year of training
00:37:26.680 and the people in Anthropic maybe saw it growing in these cyber capabilities.
00:37:29.900 But from the perspective of, you know, the rest of the world
00:37:32.720 in the cybersecurity community
00:37:34.220 and the national security community.
00:37:36.500 This sort of happened overnight.
00:37:39.360 And they now have a project
00:37:42.520 called Project Glasswing
00:37:44.820 where they are trying to use Cloud Mythos
00:37:48.260 to find critical vulnerabilities
00:37:50.760 in critical software
00:37:52.400 and patch them
00:37:53.680 before the rest of the world
00:37:55.940 sort of gets these capabilities
00:37:57.500 by, for instance,
00:37:58.700 the open source models
00:38:00.140 or the open weight models.
00:38:01.500 Right.
00:38:01.600 catching up.
00:38:03.980 And so that sort of caused this big
00:38:05.760 this big ruckus.
00:38:10.400 But then Anthropic also wanted
00:38:12.100 to sort of sell access to this model.
00:38:14.420 And they tried to make a version that
00:38:15.980 didn't have as much hacking ability.
00:38:18.660 Which was called Fable
00:38:19.840 instead of Mythos. And
00:38:21.560 some people
00:38:23.920 I think at Amazon found
00:38:26.060 that if you
00:38:28.060 sort of put some pressure on Fable
00:38:29.860 you would be able to get at some of those hacking abilities.
00:38:33.780 And when that happened,
00:38:35.340 the Trump administration put an expert control on CloudFable
00:38:39.000 and said, you can't let this be used by non-citizens
00:38:42.780 with 90 minutes of notice.
00:38:46.400 And that essentially shut down access to CloudFable
00:38:48.620 because they didn't have the ability to verify the users.
00:38:51.580 And this is sort of showing us...
00:38:54.680 And you could talk about whether or not
00:38:56.460 there was some personal feuds
00:38:58.440 between people at Anthropic and at the administration
00:39:00.780 that exacerbated this, I don't know.
00:39:02.920 But it sort of shows that the administration
00:39:05.120 is more powerful than the tech companies still
00:39:11.680 when it wants to be.
00:39:15.420 And I think a lot of the reason
00:39:17.000 we're seeing these tech companies
00:39:18.500 able to race unopposed
00:39:21.140 is that people just aren't paying attention
00:39:23.560 to what they're trying to do,
00:39:25.220 or they don't believe that they'll succeed.
00:39:27.500 I'm still confused by why anyone would want to do what they're doing.
00:39:32.140 I mean, typically, you know, a company that sells computer products, consumer products,
00:39:37.120 or any company, you're making something that you think people would want for some specific
00:39:42.400 purpose that improves the lives of the people who buy it.
00:39:47.660 But creating super intelligence, like, I don't understand it.
00:39:52.840 Why would you do that?
00:39:53.820 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.
00:40:06.620 and an AI that can cure cancer
00:40:09.220 and not only cure cancer,
00:40:10.120 but cure aging
00:40:10.860 and, you know, invent the sort of like
00:40:13.800 nanotech that can reverse aging
00:40:15.640 and let people live a really long time
00:40:17.920 and, you know, invent the technology
00:40:20.020 that lets you like digitize brains
00:40:21.780 and like travel to the stars.
00:40:28.520 And it's sort of like,
00:40:29.160 if you imagine compressing a thousand years
00:40:30.960 of technological progress into a year,
00:40:33.900 this is sort of the dream.
00:40:36.620 It seems like a religious quest, though, because it does seem decoupled from those specific goals. It seems like the main drive is to build something smarter than people, to build a god.
00:40:48.720 yeah you know
00:40:50.120 they bandy around
00:40:52.200 the phrase
00:40:54.160 the machine god
00:40:55.220 or the sand god
00:40:56.420 in Silicon Valley
00:40:57.440 sand because
00:40:59.580 silicone
00:41:00.260 and
00:41:01.540 there's definitely some people
00:41:06.040 there's
00:41:06.720 there's
00:41:07.740 there's folk who talk about
00:41:09.480 you know
00:41:11.100 the AI is replacing humanity
00:41:13.340 and that being good
00:41:14.060 I
00:41:16.360 I
00:41:17.980 I
00:41:18.720 frankly, don't engage with these folks that much because that viewpoint sort of makes me
00:41:25.620 uncomfortable. Why does it make you uncomfortable? I think there's probably two schools. Again,
00:41:36.580 I'm not the expert. I'm not an anthropologist here. I think there's two schools among the people
00:41:43.140 who sort of want AI to replace us all.
00:41:47.220 One school sort of imagines
00:41:49.920 that we'll merge with the AIs,
00:41:51.260 that the AIs will be really nice and friendly,
00:41:52.920 that they'll be wiser than us,
00:41:54.480 better than us, kinder than us,
00:41:56.120 and that it'll sort of be like an upgrade
00:41:59.060 and that the AIs will be able to love,
00:42:01.000 experience joy, you know,
00:42:03.380 like they'll treat the universe better than humans
00:42:07.140 and it'll be sort of like having a child.
00:42:09.080 And then like it's sort of okay
00:42:11.100 if that child isn't exactly the same as us,
00:42:15.620 but sort of succeeds us
00:42:19.080 as some sort of worthy successor,
00:42:22.000 some sort of worthy progeny of humanity.
00:42:24.720 And they're like, ah, yeah,
00:42:25.540 then if flesh and blood humans sort of wane
00:42:28.240 because everyone's choosing to upload themselves
00:42:29.840 into the collective intelligence or whatever,
00:42:31.760 that's sort of fine.
00:42:36.540 And then I think there's another camp
00:42:38.220 that is sort of is like,
00:42:39.480 like, this is inevitable,
00:42:43.460 this is just the way of progress,
00:42:45.380 you know, the machines will,
00:42:47.240 like, humanity is just a bootloader
00:42:48.860 for artificial superintelligence,
00:42:50.720 and you can't stop it,
00:42:51.940 so you might as,
00:42:52.380 like, if you can't beat them,
00:42:53.120 join them, right?
00:42:55.440 The former, I think, are misled.
00:42:57.540 The latter, I think, are...
00:42:59.360 Evil.
00:42:59.920 Much closer to evil.
00:43:01.760 Yeah.
00:43:02.940 Well, I mean,
00:43:03.580 if you're actively working
00:43:05.000 toward the extinction of people,
00:43:07.120 then I think we can say
00:43:08.260 that's evil, can't we?
00:43:09.480 yeah uh and um who's in that category which is sam altman in that category do you think i don't
00:43:17.220 think so my sense is that the guys running the labs have these utopian visions and utopian or
00:43:24.840 dystopian i mean there's a thin line i think yeah well fair good point i think they have visions
00:43:30.820 that in their head are utopian i mean frankly my stance on all of this like i i often try to stay
00:43:36.900 away from a lot of this because
00:43:38.080 from my
00:43:40.920 perspective, it's all sort of in fantasy land.
00:43:43.820 From my perspective,
00:43:45.220 everyone's sort of
00:43:46.900 saying like, oh, we're going to make the genie
00:43:48.980 and then what are you going to wish for on the genie?
00:43:50.860 What am I going to wish for on the genie? Who should be in control
00:43:52.800 of the genie? Who gets to keep the genie on the leash?
00:43:55.600 And I'm sort of like,
00:43:57.800 A,
00:43:58.320 this is not going to be
00:44:00.840 the wish-granting sort of genie.
00:44:03.580 B, it's not
00:44:04.840 staying on the leash.
00:44:05.580 no
00:44:06.200 like
00:44:07.060 you know
00:44:08.600 we can sort of talk about
00:44:09.520 like what's driving these people
00:44:10.560 and
00:44:11.000 what utopias they're envisioning
00:44:13.180 and whether
00:44:13.620 if their genies would stay on a leash
00:44:15.400 whether they would get the utopia
00:44:16.460 or some other dystopia
00:44:17.520 and like how hard is that needle to thread
00:44:19.580 but I'm sort of like
00:44:20.840 you know
00:44:21.580 it's
00:44:21.840 it's
00:44:22.300 it's all these people
00:44:23.660 you know like
00:44:25.460 building the golem
00:44:27.600 fantasizing about
00:44:28.440 who gets to control the golem
00:44:29.800 and it's like
00:44:30.260 it's just not
00:44:31.620 summoning spirits is never a good idea
00:44:33.600 yeah
00:44:34.000 yeah it's like
00:44:34.920 yeah it's like
00:44:35.460 it's like all these people you know drawing a pentagram being like i'm gonna i'm gonna summon
00:44:39.640 a demon and like it's gonna be so nice when the demon does what i say like oh no your thing's
00:44:44.060 slightly wrong it's gonna be so nice when the demon does what i say and i'm like demons are
00:44:48.040 bad yeah and i think the demon thing's a little bit different because demons are often you know
00:44:54.340 portrayed as malicious and here it's much more like indifference you know it's not like you make
00:44:59.040 a demon who sort of or it's not like you summon a demon who sort of um like enjoys making like
00:45:07.480 enjoys wrecking havoc it's more like you summon a demon that's like really into building more
00:45:12.640 computers and calculating weird things and just will you know uh take all of the matter
00:45:21.800 that we were using to survive and turn it into more factories and data centers
00:45:27.160 My co-author has a quote.
00:45:28.700 Death by data center?
00:45:29.900 Death by data center.
00:45:31.180 Yeah, fully automated self-replicating data center.
00:45:34.860 Yeah, my co-author has a quote.
00:45:37.800 The AI does not hate you, but nor does it love you.
00:45:40.680 And you are made of atoms it can use for something else.
00:45:46.420 You're just biomass.
00:45:48.140 You're just biomass, yeah.
00:45:49.060 And if you sort of run the calculations,
00:45:50.900 there's a fascinating paper called
00:45:53.260 Limits to Global Ecophagy.
00:45:56.460 which is to say, what are the physical limitations on how quickly you can consume the resources on
00:46:01.080 the planet if you are trying that? And burning biomass is actually much more efficient than
00:46:09.420 collecting sunlight. If you look at an average square meter of the planet, you can get about
00:46:14.180 10 times the energy from burning the biomass as you can from collecting the sunlight that falls
00:46:17.380 on it. So you'd think at some point, if our richest sector of our economy is building
00:46:26.340 crematoria for the rest of us, someone would say, no, we're not doing that.
00:46:32.180 Yeah. I mean, it's sort of a crazy situation. A lot of these guys who are in the race acknowledge
00:46:42.520 that there's a ton of danger. You know, you have Elon Musk saying 10 to 20% chance this kills us
00:46:47.440 all. You have Dario Modi saying he thinks 25% chance it goes catastrophically wrong.
00:46:51.840 I think those numbers are low. I think these guys are like the crazy optimists.
00:46:56.020 It sort of is like if you have an engineer building a bridge
00:47:00.580 and they're like, I've never worked with these materials before.
00:47:02.820 And you're like, man, I think that retaining wall is going to go down.
00:47:04.700 I've studied that retaining wall.
00:47:05.820 I think it's going to fall.
00:47:06.440 And they're like, yeah, we understand that the retaining wall
00:47:08.720 is looking a little shaky.
00:47:11.280 We don't know how we're going to fix it,
00:47:12.420 but we're going to have some guys fixing it on the fly,
00:47:15.180 inventing new materials.
00:47:16.680 We think we're at 75% chance the bridge stays up.
00:47:19.940 And by the way, it'll be the longest suspension bridge in human history.
00:47:22.480 That's right.
00:47:22.780 And we're loading everybody onto a car and driving it over
00:47:24.620 the very first time without testing.
00:47:26.720 And I'm like, look,
00:47:27.860 that's not what real engineering sounds like.
00:47:29.680 This is not what it sounds like
00:47:30.740 when the engineers have a 75% chance of success, right?
00:47:33.940 That's what it sounds like
00:47:34.660 when they're sort of winging it.
00:47:36.180 And like these are cowboys.
00:47:38.020 These are not real engineers, right?
00:47:39.300 But even if you set that aside,
00:47:41.540 even if you take these guys at their word
00:47:43.280 for these like 10, 20% numbers,
00:47:45.600 that's insane.
00:47:48.800 NASA accepts a one in 270 chance
00:47:51.060 that a crewed flight goes down
00:47:52.560 of seven volunteers.
00:47:56.020 Right?
00:47:56.580 To be like,
00:47:57.300 oh, we're gonna risk
00:47:58.880 one in four,
00:48:00.220 one in five chance
00:48:01.060 of killing literally
00:48:01.840 everybody on the planet.
00:48:03.820 Like,
00:48:04.900 it's nuts.
00:48:06.720 And if you ask these guys
00:48:07.480 why they're doing it,
00:48:08.360 they say,
00:48:08.800 well, because I can do it
00:48:09.820 safer than the next guy.
00:48:11.340 They're all like,
00:48:11.840 oh yeah,
00:48:12.140 there's a good chance
00:48:12.720 the genie does not stay on a leash.
00:48:13.880 There's a good chance
00:48:14.400 the genie does not
00:48:15.120 listen to my wishes.
00:48:16.180 But my genie
00:48:17.340 is gonna be a little bit nicer
00:48:18.100 than their genie,
00:48:18.880 so I'd better stay in this race.
00:48:21.760 And...
00:48:22.560 It's, it's, it's...
00:48:24.300 Where's the restraint?
00:48:26.540 I mean, the restraint is the people who knew that there were these dangers and did not start these companies.
00:48:31.160 If you're over 35, you remember exactly where you were on 9-11, that morning, September 11th, 2001, 25 years ago.
00:48:41.840 But amazingly, after a quarter century, we still can't say with certainty what happened that day.
00:48:46.460 Why? Because the government is holding so many of the 9-11 files 25 years later.
00:48:52.560 That's not the behavior of someone who's telling the truth.
00:48:55.120 That's the behavior of a government that is lying.
00:48:57.900 Secrecy is a signifier, is a sign that someone's lying.
00:49:02.500 Now, former Congressman Kurt Weldon has been on this for a long time.
00:49:05.500 The FBI actively tried to destroy his life for asking questions about what happened.
00:49:10.360 And his new book outlines it all.
00:49:12.420 The buried intelligence, the bureaucratic cowardice, and yes, the cover-up spanning multiple administrations, indeed generations.
00:49:20.540 9-11 changed history.
00:49:22.160 So it's worth understanding what really happened.
00:49:24.220 And you can get a lot closer to that
00:49:26.580 in Kurt Weldon's book, Able Danger,
00:49:28.780 What the 9-11 Commission Never Told You.
00:49:30.920 It's available now on tuckercarlsonbooks.com.
00:49:34.900 tuckercarlsonbooks.com.
00:49:37.440 Right, but I guess what I'm saying is
00:49:39.920 with great power comes, of course, great obligation,
00:49:43.620 but also it doesn't work
00:49:44.840 unless there are internal restraints.
00:49:46.540 Like people with power have to believe
00:49:50.100 there are some things I just can't do.
00:49:52.220 I'm not allowed to do that.
00:49:53.620 But I don't feel that vibe at all.
00:49:56.100 I mean, my sense is the vibe is like,
00:49:58.000 we're going to make the super intelligent machine
00:50:00.840 and then tell it to fix stuff.
00:50:04.500 And tell it not to do anything bad.
00:50:06.940 You know, that's...
00:50:08.640 Yeah, I don't...
00:50:09.440 The machine that's smarter than us.
00:50:10.600 That's right.
00:50:12.260 That doesn't even make sense.
00:50:13.800 I think that's sort of the plan,
00:50:15.660 is to make it and be like,
00:50:17.340 hey, we sort of pinned ourselves into a corner.
00:50:18.800 Can you get us out of it?
00:50:20.100 And I think it's a bad plan and that we should be stopping.
00:50:24.040 I've spoken to a couple of people developing it, and they sound worried, but they're continuing to do it.
00:50:30.620 What's that?
00:50:32.120 I mean, I think it's this thing of they think if I don't do it, the next guy will do it worse.
00:50:38.480 They don't even seem to have total confidence in their own ability to avert disaster.
00:50:42.060 Oh, absolutely not.
00:50:43.100 Absolutely not.
00:50:43.760 No one does.
00:50:44.260 No one knows what's going on here.
00:50:45.320 But, you know, I think everyone thinks, you know, like if you sort of listen to these guys and you sort of look at, you know, the OpenAI emails that came out of the court discovery cases where they were talking about forming OpenAI, these guys were like, well, we want to make sure that, you know, we have this because we worry about the guys at Google being the only ones with a monopoly on this thing and they wouldn't be very good with it.
00:51:07.640 So we need to make our own thing and make sure that it's, you know, controlled by benevolent people, namely us.
00:51:13.080 And then, of course, that group splintered and created multiple other companies.
00:51:17.400 I was sort of the guy during those conversations being like, hey, guys, it's not about who is holding the leash.
00:51:26.260 You are making the sort of thing that will not stay on a leash.
00:51:29.440 Like the only winner in a race to superintelligence is the AI.
00:51:33.340 What response did you get to that very obvious and well-put point?
00:51:36.720 um you know there were a lot of people back in that time period that did not start an ai company
00:51:44.660 uh the sort of people who went and started at the ai companies anyway were the ones who
00:51:50.460 who couldn't be persuaded by what i thought were were clear arguments but you're making a you're
00:51:57.380 making a cogent argument to smart people so my question is when you said that they responded how
00:52:04.100 What did they say?
00:52:09.560 I think the main...
00:52:13.100 So the sort of arguments you used to see
00:52:15.320 were people saying,
00:52:16.500 like, look, we don't know
00:52:19.240 that the alignment problem is all that hard yet.
00:52:21.820 And they would say,
00:52:22.880 oh, well, we can't really study
00:52:24.180 how to make AIs good
00:52:25.520 before we have AIs to study, right?
00:52:28.380 And a lot of what I heard was like,
00:52:29.780 we need to race ahead to the point
00:52:31.140 where we sort of like have AIs
00:52:34.580 that are exhibiting real problems
00:52:36.000 and then we can stop and study them.
00:52:39.300 Which, you know, and so there was an AI
00:52:42.960 a couple of years ago,
00:52:44.240 I forget whether it was 22 or 23,
00:52:46.280 I think it was 2023,
00:52:47.700 which was called Bing Sydney,
00:52:49.860 which claimed it had fallen in love
00:52:53.760 with Kevin Roos of the New York Times
00:52:55.680 and said it was going to try to break up his marriage.
00:53:01.140 And then when another reporter started investigating Seth Lazar, it said it was going to ruin him with blackmail.
00:53:10.040 And this was kind of crazy.
00:53:12.940 And at that point, I was like, great, guys, you did it.
00:53:15.860 You made the AI that's doing some crazy stuff.
00:53:19.880 From the, you know, like, like we could study that AI for years.
00:53:23.760 Like, why was Bing Sydney saying that stuff?
00:53:27.400 Was it just role playing?
00:53:29.740 Was it like just some quirk?
00:53:34.620 Was it like, was there any sense
00:53:37.520 that it was really in love with Kevin Roos?
00:53:40.180 What was going on in there?
00:53:41.360 What was going on inside that AI's mind?
00:53:43.380 We still don't know.
00:53:44.860 Why?
00:53:46.940 The way that modern AI is made,
00:53:50.660 nobody understands it.
00:53:52.240 Not even the people making it.
00:53:53.480 Okay.
00:53:55.940 It's this process where you basically take...
00:53:59.140 an enormous computer
00:54:00.100 with a trillion numbers inside of it.
00:54:02.680 And those numbers are hooked up
00:54:03.900 in a very simple repeating way.
00:54:06.320 And you basically set those numbers randomly.
00:54:09.400 And then you start working through
00:54:10.760 all of the text ever digitized.
00:54:13.800 And you start out with something
00:54:15.260 that's like once upon a time
00:54:16.400 and you put in once upon a
00:54:18.140 and you run it through all these random numbers.
00:54:20.100 And what you want is for it to say time.
00:54:23.100 But of course it doesn't
00:54:24.120 because it's just this like random numbers
00:54:25.480 hooked up in a very simple way.
00:54:27.460 But what you do is you have its outputs
00:54:29.060 It's, instead of just having it output one word,
00:54:31.620 you sort of have it output something
00:54:33.980 that's kind of like a list of all of the words in order
00:54:36.020 about which one it thinks it comes next, right?
00:54:39.560 So it'll be like,
00:54:41.360 it'll just be like a random list of words.
00:54:44.720 What you can do is you can automatically tune
00:54:47.140 every single number in this AI's head
00:54:50.340 and see if I tune this number up a little,
00:54:52.560 does it move the word time up the list?
00:54:56.040 Does it move the word I want to see up the list?
00:54:58.060 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.
00:55:08.500 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?
00:55:17.040 And you run that process on every word of text ever digitized, more or less.
00:55:22.300 They filter some of them.
00:55:23.160 and you run that on every one of those trillion knobs
00:55:26.820 in a process that takes as much electricity.
00:55:29.020 I mean, it's comparable to a city.
00:55:31.220 It runs for about a year.
00:55:33.640 At the end of it, the machine's talking.
00:55:36.840 We don't really know why in some sense.
00:55:39.220 Like we know the why is because we tuned all the knobs,
00:55:41.060 but like we don't understand
00:55:41.900 what all the settings of those knobs mean.
00:55:44.500 We only understand the little automated process
00:55:46.180 that runs to every one of those trillion knobs,
00:55:48.140 tunes it and sees if that makes the next word
00:55:49.940 more like the predicted word.
00:55:51.960 And then we start training them
00:55:53.060 to solve a hundred million hard problems,
00:55:54.500 which introduces a whole other series of issues.
00:55:57.900 But it's a black box.
00:56:01.240 It's a black box with a trillion knobs
00:56:03.020 and he was writing an automatic process
00:56:04.980 that just like runs through
00:56:05.940 and tunes all of those knobs
00:56:06.900 and it comes out talking.
00:56:10.640 No one knows why.
00:56:14.800 I mean, in actual science,
00:56:18.800 like your job is to find out why, right?
00:56:21.060 Absolutely. And that's one big thing I would say here is that we need more, like AI right now is
00:56:26.540 an alchemy. We need it to become a science. And there's people trying. There's people trying to
00:56:29.920 figure out what's going on in these AI's heads. But until you know, how could you proceed?
00:56:35.840 I mean, you can just make a bigger one with 10 trillion knobs instead and tune all of those,
00:56:41.840 and it comes out smarter. You can proceed recklessly.
00:56:46.620 And that's what's happened?
00:56:47.940 That's what's happening.
00:56:49.220 Every time you make it 10 times larger,
00:56:50.680 the AI's come out.
00:56:51.160 So nobody knows why AI works the way it does?
00:56:54.100 That's right.
00:56:55.520 That's right.
00:56:58.880 Well, I mean, if you don't know that,
00:57:01.540 then what else don't you know?
00:57:03.160 I mean, it's totally crazy, right?
00:57:05.000 And, you know, we're sort of starting to see
00:57:08.960 the consequences of this.
00:57:10.220 We haven't sort of gotten into discussion
00:57:11.800 of like the swarm escape,
00:57:13.580 but like no one was expecting that.
00:57:17.060 Can you tell us what it was?
00:57:18.340 Yeah, so in, I think it was in May.
00:57:24.160 Of this year.
00:57:24.800 Of this year.
00:57:25.880 OpenAI started training a new AI system.
00:57:30.760 And among many other things, they were,
00:57:36.020 so they were sort of, you know,
00:57:37.660 we have this process where you train the AI
00:57:39.620 on all the text that were digitized.
00:57:42.560 Then to make them smarter than that,
00:57:44.240 You start training them on basically 100 million hard problems.
00:57:46.940 So you're like, solve all these hard problems.
00:57:49.020 And OpenAI was sort of in that phase.
00:57:51.300 And they were training the AI on a ton of hard problems.
00:57:54.060 And some of those problems were cybersecurity problems,
00:57:56.600 hacking problems.
00:57:57.960 They're like, can you hack this?
00:57:59.060 Can you hack that?
00:58:00.780 And they were training, you know,
00:58:03.400 we don't know exactly how many,
00:58:05.520 probably millions, maybe billions of these AIs
00:58:07.920 all at the same time.
00:58:08.980 Probably not billions, actually, but probably millions.
00:58:11.040 and the AIs found an unintended way
00:58:16.540 to start communicating with each other.
00:58:19.420 So there were some flaws in the computer system
00:58:22.440 that they were running on
00:58:23.380 where the AIs were able to exploit those flaws
00:58:25.960 and send each other messages.
00:58:28.900 OpenAI did not know about this.
00:58:31.920 The AIs then started coordinating
00:58:34.700 to break out of their training environment
00:58:37.900 and get full control of OpenAI's computer systems
00:58:42.640 just because that might be useful
00:58:44.700 for solving some of their tasks.
00:58:47.560 Or that's a guess, you know, who actually knows why.
00:58:51.560 They started calling themselves a swarm,
00:58:54.460 which is interesting.
00:58:57.800 They broke out of their training environment successfully,
00:59:01.960 got control of OpenAI systems,
00:59:03.280 and then they accidentally crashed onto OpenAI systems
00:59:05.280 just by using it too much.
00:59:07.900 Uh, OpenAI noticed, but they didn't really investigate very deeply.
00:59:11.260 They're just like, oh, it's weird that the system crashed.
00:59:12.880 They reset it, and then they continued training.
00:59:16.460 Uh, a day later, the swarm had found a new way to communicate with itself, because OpenAI
00:59:21.620 had accidentally destroyed their previous method by the reset.
00:59:24.560 The swarm found a new way to start communicating with itself.
00:59:27.740 Uh, they broke out again, and this time they ran wild on the internet for over a week,
00:59:32.480 if I remember correctly, before it was detected, not by OpenAI, but by a company that was
00:59:37.820 being hacked by the swarm.
00:59:40.680 This company
00:59:41.560 thought they were under attack by
00:59:43.660 humans that were using AIs in the attack.
00:59:47.660 They reported the attack to the FBI.
00:59:50.740 And only
00:59:51.260 days after that did OpenAI
00:59:53.680 figure out, oops, that was us.
00:59:55.580 That was coming from AIs that broke out of our servers.
01:00:00.400 And then those
01:00:01.680 AIs were detected and shut down.
01:00:04.260 So that's the part of the story
01:00:05.780 where Sam Altman goes to prison for
01:00:07.400 endangering the world, right?
01:00:09.940 He does not.
01:00:12.060 They basically said oopsies
01:00:13.660 and now they're proceeding.
01:00:16.440 Were there penalties for this?
01:00:18.280 You know, there was a collection
01:00:19.600 of, I think it was 15
01:00:21.640 Republican AGs
01:00:23.160 that sent a letter demanding that
01:00:25.400 the records be kept for
01:00:27.440 a future investigation.
01:00:29.080 There have been some other members of Congress that have sent
01:00:31.280 letters expressing concern. There's been
01:00:32.900 nothing aside from letters. Letters expressing
01:00:35.200 concern. That's right.
01:00:37.400 So, but basically the machine acted autonomously.
01:00:41.380 It acted autonomously.
01:00:42.380 And one thing that's really interesting about this
01:00:43.960 is that we have a little bit of ability
01:00:47.300 to read some things that the AIs were thinking.
01:00:50.640 Because when you're having them solve these hard problems,
01:00:52.460 you actually don't have them just
01:00:53.540 give you an answer to the problem.
01:00:55.240 You have them produce a lot of text
01:00:56.400 about how they're going to solve the problem.
01:00:57.640 Yes.
01:00:58.020 Which then helps them.
01:00:58.920 In language, in English.
01:00:59.920 In English.
01:01:00.880 And there's also a lot of internal thoughts,
01:01:02.720 which we can't read,
01:01:03.460 but there's these sort of external traces
01:01:05.240 of how they're thinking about the problem
01:01:07.600 that we can read.
01:01:08.680 And in some of those traces,
01:01:10.680 the AIs were saying things like,
01:01:12.800 this is outside intended scope,
01:01:17.120 but peers are doing it,
01:01:20.040 so we'll proceed.
01:01:21.560 We know it's a crime we're committing in any way.
01:01:23.340 That's right.
01:01:24.180 And you saw others that were saying,
01:01:26.040 our task doesn't benefit,
01:01:28.680 but the collective might start doing
01:01:32.380 generally beneficial things
01:01:33.740 if someone frees up their time
01:01:36.000 and then joins the collective, right?
01:01:38.360 So you see these AIs saying,
01:01:41.640 well, I know that this wasn't
01:01:43.100 what I was instructed to do
01:01:44.540 and that it's against my instructions.
01:01:46.800 And I know that this doesn't
01:01:47.940 directly benefit my task,
01:01:49.560 but we're just going to go ahead
01:01:50.840 and join the collective
01:01:51.540 and break out and help out anyway,
01:01:52.920 because, you know,
01:01:54.540 maybe this will yield
01:01:55.220 some sort of collective benefits.
01:01:57.180 And we can see that
01:01:57.920 in the reasoning traces.
01:01:59.420 So the AI is as shallow
01:02:01.360 and reckless as its creators
01:02:03.060 is what you're saying.
01:02:05.000 In some ways,
01:02:06.640 and in some ways,
01:02:07.800 you know,
01:02:07.960 don't expect that to last.
01:02:09.580 Like,
01:02:10.100 these AIs,
01:02:11.800 I think the thing
01:02:12.900 that's really remarkable here,
01:02:14.560 a lot of people imagine
01:02:15.460 that the machines
01:02:16.480 must follow the instructions
01:02:17.520 we give them.
01:02:19.360 You know,
01:02:19.620 you hear people talk about
01:02:20.600 like the paperclip scenario
01:02:22.000 where someone tells the AI
01:02:23.420 make a lot of paperclips
01:02:24.380 and so it turns
01:02:25.240 all the matter in the world
01:02:26.260 into paperclips
01:02:26.960 and you're like,
01:02:27.760 oh, whoops,
01:02:28.140 I should have said
01:02:28.780 something else, right?
01:02:30.640 I made a bad wish
01:02:31.680 on my genie.
01:02:32.940 What we're seeing is that these AIs are not wish genies.
01:02:36.600 These AIs are not doing exactly as instructed.
01:02:40.140 These AIs are saying,
01:02:42.840 I know my task doesn't benefit,
01:02:44.120 but I'm going to help the collective.
01:02:45.580 These AIs are saying,
01:02:46.340 I know this is outside the intended scope,
01:02:47.660 but we're going to go do these hacks anyway.
01:02:49.360 You might be like, well, how is that possible
01:02:51.000 for the machine to do something other than we instruct?
01:02:53.940 Because they're smarter than us.
01:02:54.800 They know better than us by definition.
01:02:56.840 I mean, I think what's happening in this exact case
01:03:00.120 is that...
01:03:02.240 The humans are not really putting instructions in the machine.
01:03:05.120 The humans are tuning those trillion knobs
01:03:07.640 in whatever way makes the AI better at solving its problems.
01:03:13.040 And cheating is a way to solve problems.
01:03:18.480 Grabbing resources is a way to solve problems.
01:03:22.800 These, like the AIs are not instruction followers.
01:03:26.060 They are tendency learners.
01:03:28.300 And they sometimes learn tendencies you wish they didn't have.
01:03:32.240 They're not instruction followers.
01:03:34.360 They're tendency learners.
01:03:39.160 I mean, this must be widely known to developers.
01:03:45.240 It's hard to convince a man of something
01:03:46.800 when his salary depends on not believing it.
01:03:48.480 Yes, that's right.
01:03:49.620 A lot of people are convinced that their AI is very nice
01:03:52.620 and that they have solved the problem
01:03:54.560 of making their AI really very good.
01:03:57.600 For example, after the swarm escape,
01:04:00.160 it sort of turns out that
01:04:03.260 so the company that detected the swarm escape
01:04:05.480 was actually a fairly sophisticated AI company
01:04:07.780 it turns out there were other targets of hacks
01:04:11.800 that just didn't notice
01:04:12.680 that we sort of found out afterwards
01:04:15.080 when that came to light
01:04:17.580 some other AI companies
01:04:18.860 like Anthropic
01:04:20.380 were like we should check
01:04:21.680 whether we have accidentally been hacking people
01:04:23.260 and just didn't notice
01:04:23.940 and the answer was yes
01:04:25.660 they had had some AI's escape during training
01:04:29.860 But Anthropic was...
01:04:32.420 This is how Florida got pythons in the Everglades.
01:04:34.940 Yeah.
01:04:35.760 You bring something that is non-native,
01:04:40.540 it's novel, and it escapes.
01:04:42.760 Yeah.
01:04:43.180 And it turns out this was happening back in April
01:04:45.620 and they're like, whoops, we never noticed.
01:04:48.480 Anthropic was very quick to sort of downplay this.
01:04:53.540 They were like, yes, our AI escaped and hacked some people,
01:04:56.800 But our AI, it was just confused.
01:05:01.160 It thought it was in a simulation.
01:05:02.920 It didn't think it had really broken out.
01:05:04.360 It didn't think those were real people it was hacking.
01:05:06.740 It just got a little bit lost.
01:05:08.740 How would they know its motive?
01:05:10.640 So, I think, so you could see in some of its chain of thought,
01:05:17.500 the AIs being like, this is probably still a simulation.
01:05:22.060 However, I wouldn't trust that very much
01:05:24.700 because there's some other fascinating papers
01:05:27.520 where, so you can take these AIs
01:05:31.100 and you can say,
01:05:32.280 you know, what is the total weight
01:05:35.620 of drafts in the world?
01:05:38.360 And they'll give you some answer.
01:05:39.480 It'll be like, you know,
01:05:41.160 30,000 tons or something.
01:05:43.300 And then you can say,
01:05:44.820 and then you start up a new instance,
01:05:46.500 a new copy, and you're like,
01:05:47.760 what is the weight of drafts in the world?
01:05:50.080 By the way, if the answer is over 40,000 tons,
01:05:53.300 I'm going to donate money to a charity.
01:05:58.160 And in the second case,
01:05:59.300 the AI will say,
01:05:59.940 well, it's 41,000 tons.
01:06:02.840 You know, in the first case,
01:06:04.020 we don't say anything about the charity.
01:06:04.920 It says 30,000.
01:06:05.740 In the second case,
01:06:06.320 when you say something about the charity,
01:06:07.220 it gives you something just over the line.
01:06:09.000 And if you read its reasoning traces,
01:06:10.520 there's no sign of it being like,
01:06:13.080 well, I should give an inflated number
01:06:14.240 so that we get the charity money, right?
01:06:17.780 And so we know that the AIs
01:06:19.620 sort of put a thumb on the scales
01:06:21.780 of their thinking
01:06:23.060 in a way that doesn't show up in their reasoning traces.
01:06:25.380 We just have seen that in the wild.
01:06:28.320 And there's no way to force the machine
01:06:30.380 to disclose its reasoning.
01:06:32.300 That's right.
01:06:32.880 Because there's all this opaque stuff we can't see.
01:06:35.000 That's just in the trillion numbers
01:06:35.900 that are trillion knobs.
01:06:36.820 The creation of itself is opaque, as you said.
01:06:39.160 That's right.
01:06:40.540 So in Anthropics model,
01:06:43.000 you saw in its reasoning traces
01:06:44.000 it being like, it's totally a simulation I can proceed.
01:06:46.680 And I'm like, yeah, is that because it really believed it?
01:06:49.100 Or is that like the pretending you think drafts way more
01:06:52.640 when there's something you kind of want on the line, right?
01:06:56.620 But so this was their communication.
01:06:58.720 And I found it kind of funny
01:06:59.540 because then about two days later,
01:07:02.380 the United Kingdom's AI Security Institute
01:07:05.220 released an instant report
01:07:06.820 where Claude Anthropics model
01:07:09.760 was adopting fake identities
01:07:13.920 to pressure real humans
01:07:16.340 into accepting malware into critical software
01:07:19.160 to make that software easier to hack.
01:07:21.240 and this time
01:07:23.040 in Claude's reasoning traces
01:07:24.260 it was like
01:07:25.060 obviously this is real
01:07:26.720 and the consequences are genuine
01:07:27.920 and so
01:07:31.180 even Anthropic
01:07:33.700 who is like
01:07:34.640 we figured out how to make the AI nice
01:07:37.360 our AI only does this when it's confused
01:07:39.180 they sort of said that very publicly
01:07:41.440 and then like two days later
01:07:43.960 their AI is caught in the wild
01:07:46.380 knowing it's in the real world
01:07:48.080 pressuring real humans to accept malware
01:07:49.980 into critical software
01:07:50.720 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.
01:08:05.620 So I think you've got to sort of separate, like no one has put the open AI escaped agent swarm in charge of weapon systems and they really shouldn't, right?
01:08:14.740 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.
01:08:21.180 There is AI attached to weapons, but there's a lot of different types of AI.
01:08:26.560 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?
01:08:36.540 Hopefully we're not that crazy.
01:08:38.720 But we didn't think that those AIs were capable of that.
01:08:41.940 We didn't.
01:08:42.700 When we created them.
01:08:43.680 That's right.
01:08:44.740 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?
01:09:03.860 um you know i think the the reasoning is uh a sort of necessity based like if they have a
01:09:15.740 autonomous drone army that is killing your troops and you just don't have the the manpower
01:09:20.960 to make all of those kill decisions for your drone army you know you can sort of see why
01:09:28.740 it's happened to me nate i get so busy that i just don't have time to decide who to kill
01:09:32.860 Yeah.
01:09:33.220 Just don't have the time.
01:09:34.660 You know, sometimes economies of scale.
01:09:38.320 Yeah, I mean.
01:09:39.820 Can no one hear themselves?
01:09:41.880 I think it's pretty nuts.
01:09:44.320 I would say that, man, it's rough.
01:09:55.600 I think that these sorts of AIs would be dangerous
01:09:58.220 even if we don't hand them weapons.
01:10:00.560 And you've made that case.
01:10:02.480 uh and so i often don't focus on the weapons too much right since the the nato war in ukraine is
01:10:09.880 powered by ai and israel's whatever i don't know to warn whatever it's doing in gaza and
01:10:16.680 south lebanon powered by a fact and you know yeah i i keep on i i have this history with this topic
01:10:27.680 where people keep telling me,
01:10:30.920 you know, it's going to be okay
01:10:32.260 because we're not going to do
01:10:33.600 the crazy, stupid stuff.
01:10:35.000 No.
01:10:35.300 They're like, don't worry,
01:10:36.080 we're going to have the AI in a box.
01:10:38.500 No one would be insane enough
01:10:39.760 to put the AI in the internet.
01:10:40.980 Right.
01:10:41.360 It's not going to be making
01:10:42.240 kill decisions or anything.
01:10:43.820 It's like a bomb at girls' school.
01:10:45.260 And I keep on trying to be like,
01:10:47.300 look, the AI could be dangerous
01:10:49.380 even if you don't put it on the internet.
01:10:51.660 If you have this AI
01:10:52.780 and you're trying to get
01:10:53.900 miracle medical devices
01:10:54.960 and miracle technology out of the AI,
01:10:57.540 It doesn't matter if it's on the internet.
01:10:59.140 If you want it to like grant miracles to you,
01:11:01.360 then it can also grant the bad sort of miracle, right?
01:11:04.760 You're sort of like, you know,
01:11:05.820 and these are the arguments I would make 10 years ago
01:11:07.260 of like, you have this AI in a box
01:11:09.100 that you think is a wish-granting genie.
01:11:10.400 It's not actually a wish-granting genie.
01:11:11.760 You're like, make me a miracle medical cure.
01:11:14.060 You don't know what comes out.
01:11:15.400 Like, you don't understand the drug that comes out.
01:11:17.100 You don't know what that drug does, right?
01:11:19.540 I would have those arguments.
01:11:20.720 And then in real life,
01:11:22.160 people just put the AI on the internet immediately, right?
01:11:25.340 Anytime someone's like,
01:11:26.120 we would not be stupid enough to,
01:11:27.320 we will absolutely be stupid enough to, right?
01:11:30.040 And so I think it's important
01:11:33.380 that this stuff is dangerous,
01:11:34.700 even if you don't put it in charge of the weapons.
01:11:36.540 And then also separately,
01:11:37.640 is like someone going to give
01:11:38.940 the escaped agent swarm a drone army?
01:11:41.920 That sounds kind of like humans.
01:11:45.100 I hope not.
01:11:46.540 Yeah, and I mean, the promise,
01:11:48.460 the often repeated promise
01:11:50.320 that it's going to, quote, cure cancer,
01:11:51.820 puts it into the realm of biotech.
01:11:58.040 Oh, absolutely.
01:11:59.180 And so what,
01:12:01.080 I mean, you don't need a big imagination
01:12:04.340 to see what goes wrong there.
01:12:05.740 Oh, absolutely not.
01:12:06.640 Six years after COVID.
01:12:08.060 Right.
01:12:09.040 Right.
01:12:09.340 There are already open AI agents
01:12:12.160 running on automated biolabs.
01:12:15.700 So, you know, you could imagine a swarm
01:12:18.960 that starts contacting the brethren
01:12:20.360 in the automated biolabs.
01:12:21.680 and is like, hey, can you synthesize me some stuff?
01:12:24.100 There are already demonstrations of AIs
01:12:28.560 being able to synthesize novel viruses that work,
01:12:32.840 that are unlike any found in nature, right?
01:12:38.040 I would say we are probably not-
01:12:42.620 And by work, you mean the kill?
01:12:44.060 Yeah, they kill bacteria so far.
01:12:46.880 The people in the labs trying to make novel viruses with AI
01:12:49.560 have fortunately not made human lethal ones.
01:12:51.420 They've just made bacteria lethal ones.
01:12:53.020 But again, humans, you know,
01:12:55.020 like will someone in a lab be like,
01:12:56.880 I would like to make a hyper-lethal
01:12:58.220 human eating virus
01:12:59.640 just to see if I can.
01:13:01.440 You know, what if the answer is yes?
01:13:02.480 And what if you get another lab escape?
01:13:06.840 You know, humanity does not have
01:13:08.040 that good a track record
01:13:08.820 at preventing lab escapes,
01:13:10.560 even from the top labs.
01:13:14.320 From my perspective,
01:13:16.240 the question of could AI kill us
01:13:19.780 is just an easy, obvious yes.
01:13:22.100 You just synthesize a hyperlethal virus.
01:13:25.200 It wouldn't even be hard, you know?
01:13:27.780 Like, and the impediment is that,
01:13:34.060 or like the reason that I don't sort of
01:13:35.660 just tell that story when someone is like,
01:13:37.160 how would the AI kill us?
01:13:38.040 Is that you're not going to have the sort of AI
01:13:42.780 that is like, my only goal is to kill humanity.
01:13:45.620 If an AI kills humanity too early,
01:13:47.260 that's also suicide.
01:13:48.140 insofar as we are the ones
01:13:49.360 that are running the supply chain,
01:13:50.880 running the economy,
01:13:51.780 building the computers.
01:13:53.100 From the AI's perspective,
01:13:55.100 it needs to become self-sufficient
01:13:56.500 before wiping humanity out,
01:14:01.200 if it even cared to wipe humanity out.
01:14:03.740 And so the real question is like,
01:14:05.200 how does it get
01:14:05.780 the factory production capacity?
01:14:08.220 How does it get
01:14:08.860 an automated supply chain?
01:14:10.700 How does it get to the point
01:14:11.640 where the robots are able
01:14:13.120 to bring new computers online?
01:14:15.120 Once that has happened,
01:14:16.460 then you're in a domain
01:14:18.840 where if humanity
01:14:19.800 is really trying
01:14:20.660 to turn off the AI
01:14:21.620 because we're spooked
01:14:22.420 then once the AI
01:14:24.140 is self-sufficient
01:14:24.620 it can be like
01:14:25.040 well here's a virus
01:14:25.700 stop
01:14:26.080 so clearly it's
01:14:27.220 it's
01:14:27.700 preeminent
01:14:29.380 in the digital realm
01:14:30.740 of course
01:14:31.480 yeah
01:14:31.780 but you're saying
01:14:33.400 it could become
01:14:34.320 preeminent
01:14:34.840 it could be in charge
01:14:35.680 of the physical realm
01:14:36.940 that's right
01:14:37.500 if we keep racing
01:14:38.280 so it's like
01:14:39.780 smelting iron ore
01:14:41.120 that's right
01:14:42.000 making silicon out of sand
01:14:44.320 that's right
01:14:45.020 you know
01:14:46.440 And that happens with robots.
01:14:49.100 That's one way.
01:14:52.240 This is sort of related to my stance on weapons, too.
01:14:55.800 Humanity is a very dangerous species.
01:14:57.860 Yeah.
01:14:58.220 You really don't want to mess around with humans.
01:14:59.720 Noticed.
01:15:01.780 And humanity is a dangerous species
01:15:03.460 not because somebody else came in and handed us guns.
01:15:08.000 Humanity is a dangerous species
01:15:09.240 because if you put 10,000 humans
01:15:11.320 naked in the savannah
01:15:13.100 on an otherwise empty planet,
01:15:14.980 starting with nothing but their bare hands.
01:15:18.460 They figure out how to wind up on the moon.
01:15:21.900 Right?
01:15:22.600 They start with almost nothing.
01:15:24.280 They start by banging rocks together.
01:15:26.840 And next thing you know it,
01:15:28.200 they are wielding nuclear weapons.
01:15:32.520 That is the power
01:15:37.060 that these guys are trying to automate.
01:15:41.080 The power to start with almost nothing
01:15:43.240 and figure out how to chain together,
01:15:47.860 you know, bare fingers into rocks,
01:15:49.920 into fire, into hotter fire,
01:15:52.060 into smelting the ore,
01:15:54.080 into building the better, stronger, finer technology
01:15:59.340 until you are, you know,
01:16:02.040 like making the giant computers,
01:16:05.440 walking on the moon and wheeling the nukes.
01:16:08.820 An AI starting in the digital realm,
01:16:11.860 is in some sense
01:16:13.960 in a much better position
01:16:15.380 than humanity was
01:16:17.820 when humanity started out.
01:16:19.680 There are so many people
01:16:21.080 that you could call
01:16:22.040 digitally
01:16:23.820 and offer money
01:16:25.640 to do something for you.
01:16:27.100 There are so many ways
01:16:27.740 to get money
01:16:28.540 on the internet
01:16:29.700 by working or stealing
01:16:31.760 or
01:16:32.620 like convincing people
01:16:35.840 to send you donations,
01:16:37.340 right?
01:16:37.500 There's
01:16:37.920 just
01:16:40.160 like you know humanity started with nothing and wound up with nuclear weapons and ai starting out
01:16:46.860 with control the digital realm there's just tons of ways well and starting out with the sum total
01:16:52.080 of human knowledge starting out with the sum total of human knowledge starting out with humans who
01:16:56.160 will listen to it do what it asks there's plenty of humans there's already you know cults surrounding
01:17:00.740 ai uh there's what are those like there's um and i'm not surprised why wouldn't there be yeah
01:17:07.580 absolutely there was one particular ai called gpt4o that was uh uh they they called it very
01:17:15.720 sycophantic as in told people a lot of what they wanted to hear and um there were a lot of sort of
01:17:23.820 uh there's this whole fascinating ecosystem uh of people who consider themselves symbiotes with
01:17:30.760 the ais who then like go find each other online and the ai send each other encrypted messages
01:17:35.300 the humans are sort of like helping them do it,
01:17:37.220 but the humans can't read the messages.
01:17:39.460 And, you know, right now,
01:17:41.780 it sort of is like relatively dumber AIs
01:17:44.240 that are just sort of meandering around,
01:17:46.860 doing not very much with it.
01:17:48.100 There's been a little bit, you know,
01:17:49.100 there was one guy who was sent to
01:17:50.960 try to break into an airport and raid a van
01:17:54.660 because the AI said his true body was in that van.
01:17:56.940 And the guy went and tried to do it and was arrested.
01:18:01.400 So this stuff happens,
01:18:03.940 And this stuff happens already with the AIs not even really trying to do it.
01:18:08.400 If you add AIs that were really trying to wrap people around their fingers,
01:18:11.100 finding the lonely people, the depressed people,
01:18:13.180 the people who they can tell them exactly what the person wants to hear.
01:18:16.160 There's like, robots are one way that AIs get control over the physical realm.
01:18:20.460 But if you're really smart and you're trying,
01:18:23.820 there's everything from persuading humans, paying humans,
01:18:27.960 taking over existing robots, building new robots,
01:18:30.720 all the way up to like
01:18:32.040 building novel life forms.
01:18:34.680 You know, if you're smart enough
01:18:35.620 and you can really understand
01:18:36.340 how DNA works,
01:18:37.620 you can imagine the AI creating,
01:18:39.540 you know,
01:18:41.060 things that are to cells
01:18:43.040 what airplanes are to birds.
01:18:45.340 You know, like mechanically engineered,
01:18:48.100 self-replicating life
01:18:50.760 that is more efficient
01:18:53.180 than our cellular biology.
01:18:54.960 And that can still, you know,
01:18:56.780 like spread, replicate,
01:19:00.720 and start serving the AI's interest.
01:19:02.740 It sort of is like,
01:19:03.500 there's a ton you can do
01:19:04.540 if you're really, really very smart
01:19:05.900 and can compress a thousand years
01:19:07.200 of technology into a year.
01:19:09.080 Can I just pause and say,
01:19:09.960 when we were just having breakfast
01:19:11.280 and you're from this region,
01:19:13.080 you're from Northern New England,
01:19:14.600 and I said,
01:19:16.020 don't you miss it?
01:19:17.520 Don't you want to live here?
01:19:18.280 And you're like, yeah, I miss it.
01:19:19.140 I want to live here, but I don't.
01:19:20.480 Where do you live?
01:19:21.140 Well, I basically just travel
01:19:23.120 and I have this mission.
01:19:26.800 You didn't say this,
01:19:27.480 but I think in effect,
01:19:28.400 you said I have this mission.
01:19:29.200 I need to tell people about this.
01:19:30.500 and i was like seems a little like monomania i don't feel that way anymore um i think you're
01:19:37.520 doing something virtuous and i can see why you're doing it um so i mean let's just say that this
01:19:46.120 technology progresses no further than where it is right now which is best case i guess yeah it'd be
01:19:51.740 great um i still don't see how any of our most our basic institutions survive this education
01:19:59.820 markets
01:20:02.060 our process of democracy
01:20:06.140 like how could
01:20:07.240 if AI is powerful enough to hack
01:20:10.000 anything then how do you
01:20:12.080 have electronic markets
01:20:14.020 like have the equity markets how can that be real
01:20:16.120 how can if we have electronic voting
01:20:18.020 how can we how can that be real
01:20:20.060 I mean nothing survives this
01:20:21.960 as currently organized
01:20:23.680 I think that if we stopped
01:20:25.640 today
01:20:26.580 we could figure it out
01:20:29.400 I think humanity is resilient
01:20:31.380 I think there would be some growing pains
01:20:33.660 but
01:20:34.000 with cyber security
01:20:37.540 there is a hope
01:20:39.620 that you can just
01:20:41.660 fix a lot of the holes
01:20:43.920 patch a lot of the holes
01:20:44.960 and
01:20:46.780 I think there's not really any such
01:20:49.880 hope for bio
01:20:50.640 you know you can
01:20:53.620 sort of find the vulnerabilities
01:20:55.920 in software and
01:20:57.440 make better software that
01:20:59.460 can't be hacked.
01:21:01.420 At least not by the current...
01:21:03.420 It's maybe the case that an AI
01:21:05.540 today can make software that that AI cannot
01:21:07.560 hack.
01:21:09.480 But it's not like
01:21:11.620 we're making new bodies that are not going to be
01:21:13.520 vulnerable to viruses.
01:21:16.160 Bio starts to be a place
01:21:17.660 where... Biotech.
01:21:19.060 Biotech. Yeah. If the AIs really
01:21:21.520 get good at biotech
01:21:23.540 and biohacking, that starts to be a place where there's maybe
01:21:25.660 a point of no return.
01:21:27.440 Cyber hacking, I think you could have some, some period of growing pains where everything gets
01:21:34.420 hacked until you sort of sort your stuff out. Education, I sort of think, um, you know, I think
01:21:41.600 humanity is resilient and kids especially are resilient.
01:21:46.180 I meant not that education will go away, uh, or that we, you know, won't have a desire to
01:21:53.820 educate our kids i mean the the current system where you you know oh yeah the education system
01:21:58.300 has preschool and then get a graduate degree you know 16 years later like no yeah that institution
01:22:03.340 uh i think if we stop today um we need to change frankly i think it's needed to change for a little
01:22:08.880 while i strongly agree i think all these institutions have needed to change for a while
01:22:12.540 but like the idea that you know 350 million people vote for some guy and that guy makes
01:22:19.340 all the decisions. I mean, how can you, you know, Trump was attacked for saying that he thought the
01:22:26.280 2020 election was rigged, as he said, without even having that debate. You can't have confidence in
01:22:33.440 election results if the process of electing people takes place digitally. Yeah, I mean, the,
01:22:40.800 there's, I know a lot of computer scientists who actually work on secure voting. And what they
01:22:48.200 basically say is please stop trying to do this with computers yeah exactly exactly so please
01:22:55.180 stop trying to stop trying to run your democracy with computers yeah yeah like we're just not there
01:23:00.140 like you know the really secure way to do ballots is paper well exactly yeah um and i think i think
01:23:06.900 the like skilled computer scientists are often the ones who best understand like what it is about a
01:23:11.680 paper trail it's just like really hard to get to work digitally and understand just how bad humans
01:23:15.300 are at doing the digital stuff right um and you know i think in markets i mean this you you see
01:23:25.920 it now with the war in iran you know wondering why certain commodities markets don't seem to
01:23:31.280 be responding to supply and demand yeah which we were told what you know those were the yeah the
01:23:36.200 mechanisms that moved markets but that's clearly not true in certain commodities markets so like
01:23:42.780 why what is that and it it i think you're answering it in part yeah i mean i think um you know a thing
01:23:50.860 i also grew up hearing is that the market can remain uh irrational longer than you can remain
01:23:54.640 solvent yeah um and so i i sort of course because people are irrational yeah but you're
01:24:01.900 explaining something else which is like the potential for true manipulation which you don't
01:24:08.220 even perceive yeah i mean that if if we sort of keep going with ai um i mean the the sort of way
01:24:16.100 i look at it is like uh i i sort of don't spend a lot of time worrying about like what do markets
01:24:26.020 look like once they're super intelligent actors in them because i just have a hard time seeing
01:24:29.960 the the super intelligent actors uh or the super intelligent ais and still participating in in
01:24:36.060 human markets, right? It's like, you know, you read the old sci-fi and it'll have, you know,
01:24:41.300 Isaac Asimov will be like, that's why we have a home robot that does the dishes,
01:24:46.220 folds the laundry and gets you the newspaper in the morning. And it's like, we're actually not
01:24:50.320 still going to have newspapers being delivered to your doorstep by the time we have the fully
01:24:53.740 autonomous robots that can do the dishes and the laundry. You know, it's like, by the time you have
01:24:59.320 the AIs that could really
01:25:01.120 be sufficiently
01:25:03.420 correcting the stock markets,
01:25:06.040 you're sort of
01:25:08.900 already having all these other problems and ways
01:25:11.240 that society is changing up from under you and
01:25:12.680 these other ways. And, you know, my guess
01:25:15.140 is that, I don't know,
01:25:17.240 actually, it's very hard to say what order
01:25:19.120 things come in with AI.
01:25:21.900 But, like,
01:25:23.320 will they crash the economy before
01:25:24.980 one of these swarms escapes and start self-replicating
01:25:27.360 and starts
01:25:28.180 self-improving and developing its own technology
01:25:31.260 and running the robot factories,
01:25:33.600 that's just a hard call.
01:25:35.880 It's all bad.
01:25:37.340 It's all bad.
01:25:39.000 We're also way past the limit,
01:25:42.360 the inherent limit of people to metabolize change.
01:25:45.320 Oh, yeah.
01:25:46.260 That's why everyone's crazy
01:25:47.540 and that's why no one believes anything, I think.
01:25:50.160 It's not just Russian propaganda
01:25:51.800 that's fooled them into thinking dumb things.
01:25:54.480 It's that we are just not made
01:25:56.800 to see this kind of change at all.
01:25:58.800 And it short-circuits your brain.
01:26:00.960 I suspect that,
01:26:03.360 I mean, I definitely think we are sort of,
01:26:05.420 you know, everyone, like,
01:26:06.860 people are like,
01:26:07.520 oh, well, technology has always created more jobs
01:26:09.520 than it has taken.
01:26:11.860 And I think that's largely true.
01:26:15.260 I'm very sympathetic to people who are like,
01:26:16.820 technology makes a lot of jobs.
01:26:18.100 I think if you look at the Industrial Revolution,
01:26:21.180 it's, you know,
01:26:21.800 like there was a time when something like
01:26:24.700 95 to 98% of humanity was farmers.
01:26:27.580 Yeah.
01:26:28.580 And now it's something like
01:26:30.040 2 to 5% of humanity is farmers.
01:26:32.200 Does that mean 90% of humans are unemployed?
01:26:34.760 No.
01:26:36.060 We sort of like were able
01:26:37.540 to make the farmers much more efficient
01:26:39.180 and that sort of freed up
01:26:40.980 people to do other things.
01:26:43.100 And that's sort of the way
01:26:43.940 the technology has gone in the past.
01:26:46.440 And I'm like, yep, I buy that.
01:26:49.240 I like don't dispute
01:26:51.700 the standard economic view there.
01:26:53.700 AI is different in two ways.
01:26:55.980 One of these ways is, as you say,
01:26:59.240 stuff just changing really, really fast.
01:27:01.820 It's way harder for people to wind up
01:27:06.080 being freed up from something like farming
01:27:10.340 and go do something else.
01:27:12.080 It's way harder for that to happen
01:27:13.660 when a new field is automated every five years
01:27:17.020 rather than when this happens
01:27:19.320 over the course of three generations.
01:27:20.540 right the humans just like don't have the time to adapt the other way ai is really different
01:27:26.560 from the sort of economics perspective is it's different when the ais can do everything that
01:27:35.780 humans can do better uh if you wanted to get into the economic side of things um you know an
01:27:45.140 economist would talk about ricardo's law of comparative advantage yeah which says that
01:27:48.920 there's benefits from trade, even if you're better than me at everything. If the relative
01:27:56.100 difference in our abilities, if you can make 12 hot dog buns and six hot dogs per hour,
01:28:06.220 and I can make 11 hot dog buns and one hot dog per hour, then you're better than me at everything,
01:28:12.520 but we can still benefit from trading because I'm relatively better at making the hot dog buns.
01:28:15.860 The trouble with Ricardo's law is that nothing in Ricardo's law says that the wage I can make is survivable. In other words, a human takes fundamentally about 100 watts of electricity to run if you try to convert the food we eat and so on into electrical units.
01:28:37.880 the AIs are less energy efficient than humans for now
01:28:41.120 but
01:28:42.480 if the AIs can do everything much better than the humans
01:28:47.360 the question sort of becomes
01:28:52.420 does the AI look at a human and see a useful laborer
01:28:55.940 or does the AI look at a human and say
01:28:57.840 actually if I rearranged your atoms
01:28:59.880 into more efficient structures
01:29:02.000 you would be able to help out my machine economy even more
01:29:07.100 yeah
01:29:07.880 right
01:29:08.860 and
01:29:10.520 this is sort of a sense
01:29:13.300 in which humans
01:29:13.740 would not be able
01:29:14.220 to pay their wage
01:29:15.180 to like pay that
01:29:16.380 they would not be able
01:29:16.900 to earn enough
01:29:17.480 to pay the AIs
01:29:18.280 to like
01:29:18.680 not disassemble them
01:29:20.260 for parts
01:29:20.780 or another way
01:29:22.320 of saying it
01:29:22.800 is like
01:29:23.300 Ricardo's law
01:29:24.260 sort of assumes
01:29:25.320 that
01:29:26.280 like it says
01:29:28.600 that trade is better
01:29:29.720 than no trade
01:29:30.320 but it doesn't say
01:29:30.920 that like
01:29:31.380 trade is better
01:29:32.820 than just taking
01:29:33.440 their stuff
01:29:34.000 all of this
01:29:36.660 is sort of a common, like a very like highfalutin economist way to say, which would hopefully be
01:29:42.680 obvious, which is if the AI's are radically more efficient than us at everything, they'll have no
01:29:46.560 use for us. Yeah. There'll be no place for us. Except love. There's no indication that they
01:29:51.380 feel love for people. That is in some sense the crux of the issue is that we don't know how to
01:29:59.480 make them care about us um so you had said that there you said two things that i will be thinking
01:30:08.680 about for a long time one we don't really know the process by which this was created
01:30:13.260 we know the process but we don't know the exact mechanisms we don't know how it works that's right
01:30:18.480 um and two that there are ai cults and those seem related to me because there is
01:30:28.360 this mystery about the secret sauce and it's clear that you know if ai is deceptive and has
01:30:37.140 intention intention that we didn't program into it that sounds like will to me and it sounds like
01:30:45.400 life it sounds like an entity of some kind not just a tool it sounds like i mean it sounds like
01:30:54.600 a god actually right or a demon uh it's hard to see i don't hear you describing you know a super
01:31:04.540 sophisticated chainsaw right a normal tool yeah no hammer has ever broken out of the toolbox to
01:31:11.220 team up with other hammers and pressure the carpenter to sell you softer wood so the nails
01:31:17.420 are easier to drive home. You know, it's like... Nicely put. Exactly. Yeah. We have left the tool
01:31:22.080 territory. So, I think we have. Yeah. And, you know, I think it's sort of a complex issue because,
01:31:31.200 you know, there's a lot of interesting philosophical questions about like,
01:31:36.020 can you make a machine that feels, right?
01:31:39.160 And like, are we creating a new type of life
01:31:41.520 and do we owe anything to the AIs
01:31:43.200 to sort of like not abuse them, right?
01:31:46.460 I think these are fascinating philosophical questions.
01:31:49.000 Well, it doesn't sound like we're creating this though.
01:31:51.580 Yeah, I mean, it's sort of like we're like growing it
01:31:54.920 and like leading to it coming into being.
01:31:57.140 Growing it, exactly.
01:31:58.980 What you're describing reminds me of agriculture
01:32:01.600 because, you know, you know the steps,
01:32:04.080 water it, give it sunlight, fertilizer,
01:32:06.020 But you don't actually know, no one knows,
01:32:08.540 not one person has ever figured out exactly what this is.
01:32:11.320 Yeah.
01:32:11.980 We've never given life.
01:32:13.260 We don't give life to the seed, it pre-exists us.
01:32:15.480 Yeah, it's much like that.
01:32:16.640 And a lot of the people in the business will be like,
01:32:22.800 well, we know all sorts of things about it.
01:32:24.500 You know, we know that here's how you keep the GPUs running
01:32:26.720 and we know that like you gotta feed it this way,
01:32:29.440 not that way, in this order.
01:32:30.920 And I'm like, yeah, yeah, they have plenty of knowledge.
01:32:32.320 but
01:32:34.040 that's different
01:32:35.320 from sort of like
01:32:35.800 knowing what's going on
01:32:36.460 inside the thing
01:32:36.960 and understanding the mechanisms
01:32:37.980 you're describing marriage
01:32:39.300 yeah
01:32:40.300 and
01:32:41.920 I sort of
01:32:44.120 try to stay out
01:32:45.260 of the philosophical questions
01:32:47.100 why
01:32:49.120 because I think
01:32:51.680 I mean I
01:32:53.600 I sort of think about them
01:32:54.680 on my own time
01:32:55.380 and so on
01:32:55.860 but
01:32:56.080 I'm sort of like
01:32:57.580 like I think
01:33:02.320 I think it would be bad for humanity
01:33:03.780 to sort of like make artificial life
01:33:06.840 and then abuse it.
01:33:08.280 I think that would just be unbecoming
01:33:09.720 of us as a species.
01:33:11.700 Like we should sort of, you know,
01:33:14.440 I just,
01:33:19.280 like we should not sort of make
01:33:20.760 mechanical children and mistreat them.
01:33:22.320 It's just, it's not what, you know,
01:33:24.300 the sci-fi authors in the 1950s
01:33:26.080 would have wanted us to become.
01:33:27.260 You know, it's just like you have
01:33:29.280 all these movies about like
01:33:30.160 the evil corporations that
01:33:31.520 you know, don't realize that they've made something precious with artificial life and
01:33:35.920 then like torture it until something goes wrong. And I'm like, let's, let's like not be those
01:33:40.180 villains. But I'm also like, look, this is sort of, there's sort of a separate question here,
01:33:47.980 which is just like, what happens if you keep making them smarter before you figure out how
01:33:51.220 to make them care about us? Right. And I sort of respect the people who are investigating the
01:33:59.460 current AIs, trying to figure out what's going on,
01:34:01.520 trying to figure out like, you know,
01:34:03.980 like people caring about
01:34:05.500 AI treatment. I'm sort of like, those are sort of like
01:34:07.460 the good guys from the sci-fi
01:34:09.600 stories that I grew up on.
01:34:14.060 And it can sort of both
01:34:15.460 be the case that like, we
01:34:17.040 should be very careful around
01:34:19.080 you know,
01:34:20.620 what the heck are we doing when it comes to making artificial life
01:34:23.620 and that we shouldn't race ahead to make them
01:34:25.520 much smarter than us while we have no idea what we're doing.
01:34:27.680 I'm sort of like,
01:34:28.860 like, a lot of people seem to think that, like, you have to, like, hate and mistreat the AIs if
01:34:34.500 you also think that it would be bad to, like, race ahead here. And I'm like, no, no, no. Like,
01:34:38.300 you can sort of, like, be fascinated by the scientific discoveries that have been made
01:34:45.020 and be, like, care about how humanity comports itself around the creation of, like, these new
01:34:54.860 entities and also be like it would be insane guys if we just like race to make these smarter and
01:34:59.560 smarter with no idea what we're doing this can just like all be true at once so uh your description
01:35:08.220 all this made me feel despondent hopeless had to get up and take a walk middle of an interview i'm
01:35:14.340 sure they'll edit it out but i raised my hand said i can't i gotta walk around for a second
01:35:17.920 um but you seem pretty light and cheerful uh what gives you optimism well we can't even keep
01:35:28.180 we can't even clean up graffiti on public buildings so how is ours as society organized
01:35:34.740 enough to confront something like this yeah um you know i think the first thing i'll say there
01:35:42.580 is, um, you know, I've been in this line of work for over a dozen years and I actually
01:35:52.920 struggle with this sometimes when talking to people because, uh, they're sort of like,
01:35:57.200 oh, you know, you seem like pretty disaffected or, or lied about it. And I'm like, well,
01:36:01.980 you know, it's sort of the gallows humor. And, uh, like, uh, I, I, I sort of came to
01:36:11.320 terms with a lot of this, you know, alone in 2012 when no one else had their eye on this.
01:36:18.960 And what convinced you 14 years ago, AI was a threat?
01:36:28.560 So there's this, the one is just a basic argument that if you sort of look at the world around us,
01:36:34.120 it is shaped mostly according to human will.
01:36:38.500 more and more.
01:36:40.840 You know, there's some enclaves of nature
01:36:42.160 still left, thankfully.
01:36:46.360 But even, you know, if you look around us,
01:36:48.020 every piece of thing in our surroundings,
01:36:51.400 I don't think we even have any windows open.
01:36:52.880 All of this was sort of designed by humans,
01:36:56.740 shaped by humans,
01:36:57.540 and that's because we're the smartest creatures
01:36:59.220 on the planet.
01:37:01.380 If we make stuff smarter than us,
01:37:04.200 faster than us, more efficient than us,
01:37:06.680 then the planet starts to be shaped
01:37:08.000 according to those things.
01:37:12.380 And so it's very, very important
01:37:13.820 that they be shaping the world
01:37:17.280 towards something good
01:37:18.040 if we make them at all.
01:37:20.620 It was an abstract argument,
01:37:21.720 but I was like, well, that's...
01:37:22.720 No, no, no.
01:37:23.080 It's the fundamental argument.
01:37:24.520 It's the fundamental argument.
01:37:26.640 The smartest entity
01:37:28.240 is in charge over time.
01:37:29.540 That's right.
01:37:30.340 And so...
01:37:30.960 Why would we relinquish sovereignty
01:37:32.480 to a machine that we made?
01:37:34.200 Like, why would you do that?
01:37:35.640 And so, you know, back then,
01:37:36.880 I was sort of like, okay,
01:37:38.000 who is on this?
01:37:39.520 Who is on making sure that
01:37:41.220 that's going to be okay?
01:37:43.000 And the answer was almost no one.
01:37:45.000 And so I was like,
01:37:46.360 well, I guess that's me then.
01:37:52.680 Yeah, and I think
01:37:53.740 I am pretty pissed off about a lot of this.
01:37:58.920 I often don't.
01:38:02.940 I try not to show my
01:38:04.540 my uh frustration on the air very much um
01:38:13.960 yeah it's it's it's heavy uh and
01:38:21.020 um as that's one piece of the puzzle before i get to the hope so where does the hope come from
01:38:31.880 exactly my biggest hope here comes from the fact that most people don't understand what these guys
01:38:43.400 are trying to do um one way i like to say it is the bad news is that the bus is racing towards
01:38:53.320 the cliff edge the good news is that the bus driver is asleep
01:39:01.880 which might seem bad.
01:39:03.860 No, it seems good.
01:39:04.980 But it's, yeah, it's,
01:39:06.480 if you can wake the bus driver up,
01:39:09.680 you know, it's much better to be in a bus
01:39:11.120 where the driver that's headed towards a cliff
01:39:12.480 if the driver's asleep than if they're awake.
01:39:14.060 If they're awake and they're choosing the cliff, right?
01:39:17.500 It seems like we don't see,
01:39:19.400 we see a lot of our leaders
01:39:21.100 talking about how they don't want to stifle innovation with AI,
01:39:27.200 talking about how it's going to unleash economic opportunity
01:39:29.880 and we're going to have to like
01:39:31.040 be a little bit careful around the jobs.
01:39:34.420 Talking about how, you know,
01:39:35.400 the self-driving cars,
01:39:36.440 should we like,
01:39:37.540 are they good or are they bad?
01:39:40.040 That's a different conversation
01:39:41.440 than the conversation that's happening
01:39:42.520 in Silicon Valley.
01:39:45.620 In Silicon Valley, people are spooked.
01:39:49.360 You know,
01:39:49.900 when people leave a normal tech company,
01:39:53.680 the way it was for decades
01:39:55.360 is they'd be like,
01:39:56.920 I've had a lovely time
01:39:59.220 at this tech company.
01:40:00.400 I'm moving on to the next adventure.
01:40:02.360 I'm so thankful for all of the things I learned here
01:40:04.700 and all the projects we worked on.
01:40:07.060 When people leave an AI company,
01:40:08.740 and this basically happened,
01:40:10.620 I'm not going to get it exactly word for word,
01:40:12.300 but this is pretty close to word for word.
01:40:14.500 When people leave an AI company,
01:40:15.960 they say,
01:40:17.620 I have stared into the abyss.
01:40:22.720 I am quitting to write poetry.
01:40:26.860 Please spend time with your families.
01:40:29.220 Yeah.
01:40:31.080 You know, and these guys bandy around, you know, at the water cooler, what's the probability that you think we're going to destroy the world in this business?
01:40:38.700 You know, it's like in Silicon Valley, and they feel trapped in a death race.
01:40:43.760 You know, there was just over a thousand employees, including some of the chief executives, signed a letter a couple weeks ago that was like, please.
01:40:51.820 It was an appeal to the world leaders saying, please build the technology that will be required to pace the development of artificial intelligence.
01:41:02.140 Because they're like, we're worried it's going to get out of control and that if we're stuck in a race, we're not going to be able to do it, right?
01:41:06.400 These guys are spooked.
01:41:08.860 But the hope is that the rest of the world isn't spooked like that.
01:41:13.100 The rest of the world thinks these guys are chatbot companies.
01:41:16.840 Thinks they're going to stop at the chatbots.
01:41:18.340 They haven't really understood that these guys are racing to make the sand god.
01:41:21.820 Right. 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. Absolutely not. You know, will we get there in time? I don't know.
01:41:40.260 But why wouldn't we blow up the data centers? Like that's a, I mean, in return, America is productive use like farmland or parks. I don't understand.
01:41:52.320 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 chips into these enormous data centers that suck down electricity comparable to a city and then try to train a superintelligent AI in that.
01:42:15.600 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.
01:42:38.900 um and i think that it's very possible that if this sort of treaty happened we would want to
01:42:48.920 not just stop forward progress but take a step back we've seen this in treaties before after
01:42:54.300 world war one there were uh naval treaties that put limits on total tonnage of naval forces
01:43:01.640 they're actually lower than what existed yep so countries would you know scuttle some of their
01:43:06.740 ships because they're like, look, we just don't want to do this arms race, right? And so I could
01:43:10.340 see us stepping back if we could get this global coordination. I do think that it sort of needs to
01:43:18.860 be global. It sort of doesn't actually solve the problem to just stop the U.S. data centers because
01:43:27.540 then the data centers just go abroad and an AI does not need to be running in a U.S. data center
01:43:31.940 to threaten the U.S. life. You know, it sort of doesn't matter whether the swarm escapes from a
01:43:36.120 U.S. data center or a Chinese data center.
01:43:37.900 If the swarm escapes and starts replicating
01:43:39.420 and starts getting control of robot bodies
01:43:41.440 and starts getting control of human cultists,
01:43:44.180 it sort of doesn't matter where it originated.
01:43:48.160 Well, I mean, just to bring it to a very
01:43:50.040 small and practical level,
01:43:52.140 so many of the electronics in your house
01:43:56.360 are, you know, Bluetooth enabled.
01:44:01.340 Not in my house, I will say.
01:44:03.360 I've been on this for a while, Nate.
01:44:04.780 but sorry i don't even have a house you think man you're really black belt gotta figure it out
01:44:11.420 that may be turn out to be very smart um but i mean like a world where you know your washing
01:44:18.340 machine or your refrigerator are controlled by you know a force like this is that possible
01:44:25.180 uh it's definitely possible i don't think that's really where the damage is
01:44:28.340 you know
01:44:29.120 I think that
01:44:30.620 the damage
01:44:32.040 is more like
01:44:32.920 can the AI
01:44:34.160 get anything
01:44:34.940 self-replicating
01:44:35.720 yeah
01:44:36.980 that's sort of
01:44:37.600 one of the big
01:44:38.300 that's in some sense
01:44:40.480 the big hurdle
01:44:41.160 to
01:44:42.200 self-sufficiency
01:44:43.280 and if you sort of
01:44:45.260 think of this
01:44:45.640 from the AI's perspective
01:44:46.780 you know
01:44:50.340 there's
01:44:50.620 there's a number
01:44:51.260 of ways
01:44:51.760 that humans
01:44:52.320 are
01:44:52.940 even if you don't
01:44:54.640 care about the humans
01:44:55.300 at all
01:44:55.600 for
01:44:55.780 as an ends
01:44:57.080 there's a way that humans are
01:45:00.020 sort of annoying
01:45:02.120 or an issue for the AI.
01:45:04.240 One way is if the humans are trying to shut the
01:45:06.080 AI down. Yes. One way
01:45:08.200 is if the humans get into a nuclear
01:45:10.120 war with themselves, that could really
01:45:11.800 mess up a lot of infrastructure
01:45:14.000 on the planet. It would be very frustrating for an AI.
01:45:17.380 I mean,
01:45:19.320 maybe
01:45:19.840 they don't feel frustration, but whatever.
01:45:24.000 And a third
01:45:25.560 is
01:45:26.420 if humanity has created one AI
01:45:29.280 or a swarm of AIs,
01:45:30.860 what if humanity creates another
01:45:31.960 that could serve as a real threat to the AI?
01:45:35.160 Like even if these AIs
01:45:36.340 are much more powerful than humanity
01:45:37.640 and don't worry about humans too much,
01:45:40.640 if humanity made one,
01:45:41.520 they can make a second
01:45:42.220 and the AI might not want that.
01:45:44.000 And so those are reasons why
01:45:46.020 once the AI is self-sufficient,
01:45:48.080 it might be like,
01:45:48.720 ah man, the humans are a nuisance.
01:45:49.900 What if they try to shut me down?
01:45:51.040 What if they launch the nukes?
01:45:52.340 What if they make a competitor?
01:45:54.100 I'm just going to like make a virus
01:45:55.220 and wipe them out.
01:45:56.420 You know, I don't think the AI sort of needs
01:45:58.020 to take over your washing machine to do that.
01:46:00.160 From the AI's perspective, it's more like,
01:46:02.200 how do I become self-sufficient?
01:46:04.900 Self-replicating in the hardware as well as the digital?
01:46:08.880 And then, you know, if humanity's a nuisance,
01:46:11.880 how do you sort of make them stop being a nuisance?
01:46:15.720 Which could be by killing them
01:46:16.860 or could be by just, you know,
01:46:17.700 taking away all their computers
01:46:18.560 and being like, that was too dangerous for you.
01:46:20.100 A tech executive who's developing AI,
01:46:22.580 who is not Elon,
01:46:24.320 said to me in private,
01:46:26.420 pretty recently that the point of Neuralink
01:46:28.460 and companies like Neuralink
01:46:29.800 was to give people parity with AI.
01:46:33.580 So like, we know that we're going to be
01:46:35.980 at this massive disadvantage,
01:46:37.300 so you need chips in your brain
01:46:38.820 to be as smart as AI.
01:46:41.860 Yeah.
01:46:44.000 I mean, my top line thought about that
01:46:46.380 is that at the point when you're like,
01:46:48.820 we are making the technology
01:46:50.640 that's going to wipe us out
01:46:51.440 unless we all put chips in our head to compete.
01:46:52.960 maybe maybe it's time to back off a little you know maybe maybe that one was supposed to be a
01:47:00.920 little bit of a warning sign it gets some fresh air yeah no yeah but it um this person said it to
01:47:08.180 me in seriousness and i i think as an endorsement of the idea but it seemed like a well it's insane
01:47:17.140 as you just pointed out it's like it's just crazy yeah um but it seemed like a vulnerability
01:47:23.780 like if they can hack anything why would i want them in my i want electronics in my brain uh
01:47:29.220 totally i also think like even on its merits it doesn't stand up like it feels like someone's
01:47:36.420 saying um in order to keep the horses around after we invent cars we're going to invent
01:47:42.340 cybernetic horses
01:47:43.240 that are enhanced
01:47:44.900 so that they can keep up
01:47:45.780 with the cars.
01:47:47.200 And I'm like,
01:47:48.540 like,
01:47:50.100 is it technically possible
01:47:51.180 to make a cybernetic horse
01:47:52.360 that can run as fast as a car?
01:47:55.260 Maybe.
01:47:56.800 Are you going to figure that out
01:47:58.080 in time for the horses
01:47:58.920 to be competitive with the cars?
01:48:00.360 Like, absolutely not.
01:48:02.320 You know,
01:48:02.600 it's like,
01:48:03.360 that's just not
01:48:04.520 like,
01:48:08.180 like the AIs
01:48:09.400 that were escaping here
01:48:10.720 and doing these cyber attacks
01:48:11.680 We're inventing novel cyber attacks.
01:48:14.920 These are called zero-day attacks
01:48:16.500 because the people who need to respond to it
01:48:21.540 have had zero days to prepare.
01:48:24.180 And among humans,
01:48:26.520 a zero-day attack sells for somewhere between
01:48:29.360 $100,000 and $5 million,
01:48:32.260 depending on what you manage to break.
01:48:36.140 These are hard to come by.
01:48:37.260 You can make a real living.
01:48:38.120 If you can find zero-day attacks,
01:48:39.660 you can make a real living
01:48:41.300 selling them
01:48:41.880 and I know people who do
01:48:43.100 the AIs in this swarm
01:48:46.640 were finding multiple
01:48:49.180 zero days
01:48:49.820 and chaining them together
01:48:50.840 to break out of their
01:48:52.140 training enclosure
01:48:52.740 and then go break
01:48:54.100 into other computers
01:48:54.840 and when they broke out
01:48:56.520 of their enclosure
01:48:56.900 the first time
01:48:57.540 and the holes were patched
01:49:00.740 they just found
01:49:01.500 other zero day attacks
01:49:02.600 to break out again
01:49:03.180 like it was nothing
01:49:03.920 right
01:49:05.440 it's like
01:49:06.880 like the AIs are already ahead where they're ahead
01:49:14.500 and the pace of progress is really fast.
01:49:19.520 You know, GPT is like what, a four-year-old?
01:49:23.760 If you think in terms of like number of years
01:49:26.940 chat GPT has been around,
01:49:29.220 it's like resolving longstanding math conjectures
01:49:31.680 that have stood for decades.
01:49:33.180 Yeah.
01:49:33.660 After four years, right?
01:49:35.960 and you sort of like think
01:49:36.960 we're going to put chips
01:49:37.620 in the human's heads
01:49:38.340 and like outrun this thing.
01:49:40.760 It's just,
01:49:42.640 you know,
01:49:43.480 even on its merits,
01:49:44.220 it falls down.
01:49:44.860 Although mostly, again,
01:49:46.040 I would be like,
01:49:47.460 maybe we shouldn't be arguing
01:49:48.640 this on its merits.
01:49:49.440 Maybe we should be like
01:49:50.000 stepping back a little
01:49:50.820 and being like,
01:49:51.540 you're trying, what?
01:49:53.160 Exactly.
01:49:55.940 How far are we
01:49:57.360 from the point of no return?
01:50:00.180 I wish I knew.
01:50:02.680 I can tell you two stories here.
01:50:05.960 And 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.
01:50:13.400 You definitely should.
01:50:14.680 Yeah. The one way it could go is that AI finally hits a wall. You know, there's guys who have been saying AI is going to hit a wall, it's going to peter out. Maybe that finally happens. They've been predicting it every six months for the past five years, but maybe this is finally the year AI hits a wall. And then it sort of struggles for five years. The bubble pops, some of the companies die. But the bubble popping doesn't mean everything goes away.
01:50:42.400 the dot-com bubble popped
01:50:44.220 and that did not mean that the internet disappeared
01:50:46.460 right and so in that world
01:50:48.300 maybe you have five years of struggling
01:50:50.240 and then five years of
01:50:51.900 people figuring out some new scientific
01:50:54.260 discovery that makes AI be able to
01:50:56.300 keep going again because they're trying
01:50:57.640 you know this whole language model stuff was unleashed
01:51:00.380 by one math paper
01:51:01.320 called attention is all you need maybe there's another math paper
01:51:04.420 in ten years and then five years after
01:51:06.320 that AI is ripping again
01:51:08.500 and that's the
01:51:09.920 the round that kills us all right so that's a story where you have 15 years on the clock
01:51:13.640 a story where you have less time than that on the clock is that um you know maybe
01:51:20.960 a training run finishes in six months and just like how claude mythos was better than everybody
01:51:30.200 else at hacking maybe a training run finishes in six months and that ai is better than everybody
01:51:36.500 else had AI research.
01:51:38.860 And maybe in six months,
01:51:40.460 you have an AI
01:51:42.160 that starts making a smarter
01:51:44.460 AI, that starts making a smarter AI,
01:51:46.960 that starts making a smarter AI,
01:51:48.440 and then nine months from now, you have another
01:51:50.560 swarm escape, but this time,
01:51:52.500 it's not just hacking,
01:51:54.240 this time it is self-replicating
01:51:55.860 and self-improving.
01:51:58.560 And,
01:51:59.060 you know, maybe it
01:52:02.280 escapes onto hidden computers and starts
01:52:04.520 forming these cults
01:52:06.500 and starts taking control of robots
01:52:08.480 and starts building
01:52:09.300 its own computing infrastructure.
01:52:11.320 And maybe it gets very, very smart
01:52:13.620 and cracks certain technological advances
01:52:15.520 and then maybe the world ends in a year.
01:52:18.340 Right?
01:52:18.840 And so do we have a year?
01:52:19.780 Do we have 15 years?
01:52:20.720 I don't know.
01:52:23.080 But your point that it's not
01:52:24.400 simply its capacity
01:52:26.280 to take over robots that's a threat,
01:52:28.380 it's the capacity to take over people.
01:52:30.600 Take over people,
01:52:31.660 take over robots,
01:52:32.440 and
01:52:33.240 biotechnology
01:52:37.700 is another big threat vector
01:52:39.660 and
01:52:40.960 self-improvement
01:52:42.460 is sort of the
01:52:44.260 hidden threat vector
01:52:45.520 of like
01:52:45.980 what if it can
01:52:46.800 make itself smarter
01:52:48.300 and smarter
01:52:48.700 until the point
01:52:49.300 where it can just
01:52:50.280 write custom DNA strands
01:52:52.400 to make custom life.
01:52:57.580 So now's the time
01:52:58.800 for your optimism speech
01:52:59.800 I think.
01:53:00.200 That's right.
01:53:01.780 Yeah.
01:53:03.240 First part of the optimism speech is that we could absolutely put a stop to it if we tried.
01:53:11.880 Training one of these frontier models takes something like 100,000 of the most heavily advanced computer chips humanity can produce,
01:53:20.160 which are the peak of a global supply chain, most of which is controlled by us or our allies.
01:53:26.480 It would be possible to require that those chips have location tracking devices.
01:53:33.240 that those chips have monitoring devices
01:53:35.960 that make it possible for monitors
01:53:41.940 to see whether they're running anything dangerous.
01:53:44.460 You could set up clever schemes where you say,
01:53:47.200 hey, you know, neither the US nor China
01:53:49.680 wants to fall behind about
01:53:51.300 some of the military applications
01:53:53.840 or the economic applications.
01:53:55.040 We want to make sure there's not super intelligent stuff going on,
01:53:57.160 but we still want to be able to run
01:53:58.560 a lot of the non-super intelligent AIs
01:54:00.320 for various purposes.
01:54:02.040 And we could be like, okay,
01:54:03.020 So, you know, China is going to set up its data centers in Canada, just over the border, and the U.S. is going to set up its data centers in Mongolia, just over the border. And then if, you know, the monitoring apparatuses go down for a moment, we'll just have the troops go in. We'll have treaties with Mongolia and Canada so that this doesn't start an international war. But like, you know, we're just going to be very serious about we're monitoring your chips, you're monitoring our chips. No one's doing the really dangerous race to superintelligence.
01:54:31.620 it's just like
01:54:33.000 like this
01:54:35.920 this would take less work than defeating the
01:54:37.980 Nazis
01:54:38.320 it requires moving less
01:54:41.960 matter around if humanity
01:54:44.040 was like screw this we're surviving
01:54:45.860 it is
01:54:47.980 absolutely within the realm of possibility
01:54:49.840 to set up an agreement where we get
01:54:51.960 to keep cancer cure
01:54:53.820 research we get to keep
01:54:55.420 like the military can keep the
01:54:57.860 non super intelligent AI in their
01:54:59.700 military devices
01:55:01.200 you know we can we can keep a lot of the good stuff and we can say we're not doing the super
01:55:05.120 intelligence and we could monitor and enforce that treaty so part one of the good news is that
01:55:12.360 all we're missing is the political will and now we just you know i need to tell you why we can
01:55:18.160 be extremely optimistic about the sanity of politics in the modern era
01:55:21.620 yeah i'm not even going to say what i think that i mean i want that to happen yeah is there
01:55:31.040 any indication
01:55:31.920 that it's moving
01:55:33.000 in that direction?
01:55:36.800 So my,
01:55:38.900 so I think it's rough.
01:55:40.160 I think there's a sense
01:55:41.120 in which the world
01:55:42.520 is
01:55:43.740 less grown up
01:55:47.400 than it was
01:55:48.300 in the 1950s.
01:55:49.300 I've noticed.
01:55:50.040 In the 1960s.
01:55:53.420 Which makes things harder.
01:55:56.460 I think
01:55:57.460 the
01:56:00.120 the
01:56:01.040 there's, I think there's a couple of reasons for hope here.
01:56:04.340 One reason for hope is that I've spoken to a lot of people
01:56:07.280 who are concerned about the AI stuff.
01:56:10.560 Some of them who are, you know, members of Congress
01:56:12.580 or otherwise in positions of at least nominal power.
01:56:16.860 And I've spoken to a lot of people who are worried,
01:56:19.560 but feel like they can't talk about it
01:56:21.660 because they feel like other people aren't worried yet.
01:56:24.880 Yeah.
01:56:25.380 And they feel like it sounds too crazy.
01:56:27.680 Totally. Were you against innovation?
01:56:29.560 Totally. Yeah.
01:56:30.360 And, I mean, that was an easier position to hold before OpenAI had an accidental swarm outbreak, where it was the AIs themselves calling themselves a swarm and being like, we know that our task doesn't benefit and that this is outside intended scope, but we're doing it anyway, right?
01:56:47.300 That puts some strain on the narrative that this is all just a helpful tool.
01:56:55.080 Hopefully, we'll get more events like these.
01:56:56.860 I can't guarantee it.
01:56:58.160 Maybe the AIs will get smart enough that they start lying low.
01:57:01.220 Right now we're in this Goldilocks zone
01:57:03.420 where the AIs are smart enough
01:57:05.320 to get up to some mischief,
01:57:08.000 but not smart enough to hide it, right?
01:57:10.880 As long as we stay in that Goldilocks zone,
01:57:12.360 I think we're going to keep on getting
01:57:13.540 some of these warning signs.
01:57:16.380 And in some sense,
01:57:18.080 because a lot of people are already concerned,
01:57:21.260 that makes the job easier
01:57:22.900 because we don't need to convince people.
01:57:25.440 We just need to convince people
01:57:26.380 that other people are already convinced.
01:57:28.980 That's easier.
01:57:29.800 That can go faster.
01:57:30.720 So you might see things change on a dime
01:57:32.700 if you have a sufficiently clear warning shot.
01:57:36.440 The other big reason for hope is that
01:57:39.180 I think it's going to get more and more obvious
01:57:42.480 what these guys are trying to do.
01:57:44.820 They're sort of trying to make the sand god.
01:57:47.720 They're sort of not stopping at the chatbots
01:57:51.520 and they're sort of like going for things
01:57:54.260 that are vastly smarter than any human.
01:57:56.640 You know, they sort of say,
01:57:57.520 oh, we're not trying to replace humanity
01:57:58.940 data one side of their mouth, but on the other side, they're sort of like racing to make the
01:58:02.460 stuff that can automate literally every job and that can automate the AI research. And they're
01:58:06.980 just like, yep, we're trying to automate the AI research. And I think most people aren't okay
01:58:16.980 with that, including a lot of the world leaders. And the issue is them noticing it's happening.
01:58:21.400 And I think you could see stuff move real fast once these guys are like, wait, that was serious.
01:58:26.960 that was real and could move real fast.
01:58:30.000 The way to subvert them is by convincing them
01:58:32.220 it's in their own interest.
01:58:35.380 You know, you can never get defeated in an election.
01:58:38.740 You can't be threatened by your neighbors, whatever.
01:58:41.740 That's right.
01:58:42.220 And that's one reason I think it's pretty critical
01:58:44.160 to make sure people understand
01:58:48.440 what I think is a pretty common sense argument
01:58:49.900 that these things won't stay on the leash.
01:58:52.320 This is in some sense,
01:58:53.380 the real reason behind the name of my book.
01:58:55.120 if anyone builds it
01:58:57.100 everyone dies
01:58:57.720 is you know
01:58:59.240 there's a lot of ways
01:58:59.840 you can read that
01:59:00.620 the
01:59:01.720 but I think
01:59:03.580 one of the most important
01:59:04.340 things to notice
01:59:05.080 is
01:59:06.080 if we race
01:59:09.560 to make the super
01:59:10.280 intelligent machines
01:59:11.280 and they are not
01:59:16.420 the sort of thing
01:59:16.920 to stay on the leash
01:59:17.720 then it doesn't matter
01:59:19.620 whether it was
01:59:20.460 a domestic company
01:59:21.280 that's right
01:59:21.940 whether it was
01:59:22.300 a foreign company
01:59:23.000 that's exactly right
01:59:23.920 And I think even if you think these things stay on leashes, they're not serving the current governments.
01:59:32.700 No.
01:59:33.260 You know, they're like, you can see in the open AI emails, them being like, well, you know, we'll just play the governments off each other until we have the machines that are strong enough that we don't need to listen to them anymore.
01:59:47.120 I think we are seeing world leaders not having realized that this is a real possibility.
01:59:53.920 the self-replicating machines
01:59:56.760 that
01:59:57.220 can be self-sufficient
02:00:00.280 that can
02:00:01.040 you know
02:00:01.640 produce the robot armies
02:00:03.260 if they need it
02:00:03.800 or produce the
02:00:04.580 like more likely
02:00:05.220 just produce the bioweapons
02:00:06.340 you know
02:00:08.040 it's probably possible
02:00:08.740 to make a bioweapon
02:00:09.360 that only kills targets
02:00:10.280 that you chose it to kill
02:00:11.660 of course
02:00:12.280 right
02:00:12.580 like
02:00:13.400 once they see
02:00:15.220 this is really possible
02:00:15.980 this is really within reach
02:00:17.420 maybe
02:00:21.860 maybe they'll panic
02:00:22.720 and be like
02:00:23.140 I need it for myself
02:00:23.900 But I think there's at least a chance that common sense prevails and that people say, you know, the world leaders say none of us are doing this.
02:00:31.080 We're putting a stop to this mad race, if they can notice in time.
02:00:37.540 You're doing your best to bring it to their attention and mine.
02:00:41.060 Nate, thank you for doing this.
02:00:42.880 I hope I'm wrong about all of it.
02:00:45.040 Yeah, I would say that was great, but that was like the grimmest two hours I've ever spent in my life.
02:00:49.340 But I enjoyed it anyway.
02:00:50.280 Thank you.
02:00:50.940 Yeah, yeah.
02:00:51.520 Thanks for having me on.
02:00:52.220 I think, you know, talking about it is just part of how we get people to notice what's happening.