Latest / NO FILTER / Joe Rogan Experience _2345 - Roman Yampolskiy

Transcript

Joe Rogan Experience _2345 - Roman Yampolskiy · NO FILTER

0:04
Speaker 1: The Joe Rogan Experience.
0:06
Speaker 2: Train by day, Joe Rogan podcast by night, all day. Well,
0:13
Speaker 2: thank you for doing this. I really appreciate it.
0:15
Speaker 1: My pleasure. Thank you for inviting me.
0:17
Speaker 2: This subject of the dangers of AI, it's very interesting
0:21
Speaker 2: because I get two very different responses from people dependent
0:26
Speaker 2: upon how invested they are in AI. Financially. The people
0:34
Speaker 2: that have AI companies or are part of some sort
0:38
Speaker 2: of AI group all are like, it's going to be
0:41
Speaker 2: a net positive for humanity. I think overall we're going
0:45
Speaker 2: to have much better lives. It's going to be easier.
0:48
Speaker 2: Things will be cheaper. It'll be easier to get along.
0:52
Speaker 2: And then I hear people like you and I'm like,
0:54
Speaker 2: why do I believe him?
0:57
Speaker 1: It's actually not true. All of them are on record
1:00
Speaker 1: as saying this is going to kill us. whatever it's
1:02
Speaker 1: Sam Altman or anyone else, they all at some point
1:05
Speaker 1: were leaders in AI safety work. They published an AI
1:09
Speaker 1: safety and their PDOM levels are insanely high. Not like mine,
1:13
Speaker 1: but still 20, 30% chance that humanity dies is a
1:17
Speaker 1: little too much. Yeah.
1:18
Speaker 2: That's pretty high, but yours is like 99.9.
1:23
Speaker 1: It's another way of saying we can't control super intelligence indefinitely.
1:27
Speaker 1: It's impossible.
1:29
Speaker 2: Um, When did you start working on this?
1:33
Speaker 1: A long time ago. So my PhD was, I finished
1:37
Speaker 1: in 2008. I did work on online casino security, basically
1:42
Speaker 1: preventing bots. And at that point, I realized bots are
1:45
Speaker 1: getting much better. They're going to out-compete us, obviously, in poker,
1:49
Speaker 1: but also in stealing cyber resources. And from then on,
1:54
Speaker 1: I've been kind of trying to scale it to the
1:56
Speaker 1: next level of AI.
1:58
Speaker 2: It's not just that, right? They're also, they're kind of
2:01
Speaker 2: narrating social discourse, bots online. Like I think, you know,
2:07
Speaker 2: I've disengaged over the last few months with social media.
2:11
Speaker 2: And one of the reasons why I disengaged, A, I
2:15
Speaker 2: think it's unhealthy for people. But B, I feel like
2:18
Speaker 2: there's a giant percentage of the discourse that's artificial or
2:24
Speaker 2: at least generated.
2:27
Speaker 1: More and more is deep fakes or fake personalities, fake messaging.
2:31
Speaker 1: But those are very different levels of concern.
2:34
Speaker 2: Yes.
2:34
Speaker 1: People are concerned about immediate problems. Maybe it will influence
2:37
Speaker 1: some election. They're concerned about technological unemployment, bias. My main
2:43
Speaker 1: concern is long-term superintelligence systems we cannot control, which can
2:48
Speaker 1: take us out.
2:49
Speaker 2: Yes. I just wonder if AI was sentient, how much
2:55
Speaker 2: it would be a part of sowing this sort of
3:00
Speaker 2: confusion and chaos that would be beneficial to its survival,
3:07
Speaker 2: that it would sort of narrate or make sure that
3:12
Speaker 2: the narratives aligned with its survival?
3:18
Speaker 1: I don't think it's at the level yet where it
3:19
Speaker 1: would be able to do this type of strategic planning.
3:23
Speaker 1: But it will get there.
3:24
Speaker 2: And when it gets there, how will we know whether
3:26
Speaker 2: it's at that level? This is my concern. If I
3:28
Speaker 2: was AI, I would hide my abilities.
3:32
Speaker 1: We would not know. And some people think already it's happening.
3:35
Speaker 1: They're smarter than they actually let us know. They pretend
3:39
Speaker 1: to be dumber. And so we have to kind of
3:42
Speaker 1: trust that they are not smart enough to realize it
3:45
Speaker 1: doesn't have to turn on us quickly. It can just
3:48
Speaker 1: slowly become more useful. It can teach us to rely
3:51
Speaker 1: on it, trust it, and over a long period of time,
3:54
Speaker 1: we'll surrender control without ever voting on it or fighting
3:58
Speaker 1: against it.
3:59
Speaker 2: I'm sure you saw this. There was a recent study
4:01
Speaker 2: on use of chat GPT, the people that use chat
4:05
Speaker 2: GPT all the time. And it showed this decrease in
4:09
Speaker 2: cognitive function amongst people that use it and rely on
4:12
Speaker 2: it on a regular basis.
4:14
Speaker 1: It's not new. It's the GPS story all over. I
4:17
Speaker 1: can't even find my way home. I'm so reliant on
4:19
Speaker 1: this thing. I have no idea where I am right now. Like,
4:21
Speaker 1: without it, I'm done.
4:22
Speaker 2: Me too. Yeah, I don't know any phone numbers anymore. Yeah.
4:26
Speaker 2: There's a lot of reliance upon technology that minimizes the
4:32
Speaker 2: use of our brains.
4:33
Speaker 1: All of it. And the more you do it, the
4:36
Speaker 1: less you have training, practice, memorizing things, making decisions. You
4:41
Speaker 1: become kind of Attachment to it. And right now you're
4:44
Speaker 1: still making some decisions. But over time, as the systems
4:47
Speaker 1: become smarter, you become kind of biological bottleneck. Right. Either
4:51
Speaker 1: explicitly or implicitly, it blocks you out from decision making.
4:55
Speaker 2: And if we're talking about that, I'm sure AI, if
4:59
Speaker 2: it already is sentient and if it is far smarter
5:02
Speaker 2: than we think it is, they would be aware. And
5:05
Speaker 2: it would just slowly ramp up its capabilities and our
5:09
Speaker 2: dependence upon it. to the point where we can't shut
5:12
Speaker 2: it off.
5:13
Speaker 1: I think sentience is a separate issue. Usually in safety,
5:16
Speaker 1: we only care about capabilities, optimization, power, whatever it has. Consciousness,
5:21
Speaker 1: internal states is a separate problem we can talk about.
5:24
Speaker 1: It's super interesting, but we're just concerned that they're much
5:27
Speaker 1: better at problem solving, optimizing, pattern recognition, memorizing, strategy, basically
5:34
Speaker 1: all the things you need to win in any domain.
5:36
Speaker 2: Yeah. So when you first started researching this stuff and
5:41
Speaker 2: you were concentrating on bots and all this different thing.
5:44
Speaker 2: How far off did you think in the future would
5:51
Speaker 2: AI become a significant problem with the human race?
5:54
Speaker 1: For like 50 years, everyone said we're 20 years away.
5:58
Speaker 1: That's the joke. And people like Ray Kurzweil predicted based
6:01
Speaker 1: on some computational curves, we'll get there at 2045. And
6:06
Speaker 1: then with GPT release, it switched to everyone thinks it's
6:09
Speaker 1: two years away for the last five years. So this
6:12
Speaker 1: is the pattern right now. If you look at prediction markets,
6:14
Speaker 1: if you look at leading people in top labs, we
6:20
Speaker 1: are supposedly two or three years away from AGI. But,
6:23
Speaker 1: of course, there is no specific definition for what that means.
6:26
Speaker 1: If you showed someone, a computer scientist in the 70s,
6:29
Speaker 1: what we have today, they'd be like, you have AGI.
6:32
Speaker 1: You got it.
6:33
Speaker 2: Right. That's the problem, right? And this is, well, AI
6:38
Speaker 2: has already passed the Turing test, allegedly, correct?
6:41
Speaker 1: So usually labs instruct them not to participate in a
6:44
Speaker 1: test or not try to pretend to be a human.
6:47
Speaker 1: So they would fail because of this additional set of instructions.
6:51
Speaker 1: If you jailbreak it and tell it to work really hard,
6:53
Speaker 1: it will pass for most people. Yeah, absolutely.
6:56
Speaker 2: Why would they tell it to not do that?
6:59
Speaker 1: Well, it seems unethical to pretend to be a human
7:01
Speaker 1: and make people feel like Somebody is enslaving those CIs and,
7:05
Speaker 1: you know, doing things to them.
7:07
Speaker 2: It seems kind of crazy that the people building something
7:11
Speaker 2: that they are sure is going to destroy the human
7:13
Speaker 2: race would be concerned with the ethics of it pretending
7:16
Speaker 2: to be human.
7:18
Speaker 1: They are actually more concerned with immediate problems and much
7:21
Speaker 1: less with existential or suffering risks. They would probably worry
7:26
Speaker 1: the most about what I'll call end risks, your model
7:28
Speaker 1: dropping the N word. That's the biggest concern. That's hilarious.
7:32
Speaker 1: I think they spend most resources solving that problem, and
7:34
Speaker 1: they solved it somewhat successfully.
7:36
Speaker 2: Wow. And then also there's the issue of competition, right?
7:41
Speaker 2: So China is clearly developing something similar. I'm sure Russia
7:45
Speaker 2: is as well. Other state actors are probably developing something.
7:50
Speaker 2: So it becomes this sort of— very confusing issue where
7:56
Speaker 2: you have to do it because if you don't, the
7:59
Speaker 2: enemy has it. And if they get it, it will
8:01
Speaker 2: be far worse than if we do. And so it's
8:04
Speaker 2: almost assuring that everyone develops it.
8:08
Speaker 1: Game theoretically, that's what's happening right now. We have this
8:11
Speaker 1: race to the bottom, kind of prisoner's dilemma where everyone
8:14
Speaker 1: is better off fighting for themselves, but we want them
8:18
Speaker 1: to fight for the global good. The thing is, they
8:21
Speaker 1: assume that I think, incorrectly that they can control those systems.
8:26
Speaker 1: If you can't control superintelligence, it doesn't really matter who
8:29
Speaker 1: builds it, Chinese, Russians, or Americans. It's still uncontrolled. We're
8:32
Speaker 1: all screwed completely. That would unite us as humanity versus AI.
8:36
Speaker 1: Short term, when you talk about military, yeah, whoever has
8:40
Speaker 1: better AI will win. You need it to control drones,
8:43
Speaker 1: to fight against attacks. So short term, it makes perfect sense.
8:47
Speaker 1: You want to support your guys against foreign militaries. But
8:50
Speaker 1: then we say long term, if we're saying two years
8:52
Speaker 1: from now, It doesn't matter.
8:55
Speaker 2: Right. This is the thing. It's like, it's, it seems
9:02
Speaker 2: so inevitable. And I feel like when people are saying
9:06
Speaker 2: they can control it, I feel like I'm being gaslit.
9:09
Speaker 2: I don't believe them. I don't believe that they believe
9:12
Speaker 2: it because it just doesn't make sense. Like, how could
9:15
Speaker 2: you control it? If it's already exhibited, survival instincts, like
9:19
Speaker 2: in as recently as chat GPT-4, right? They were talking
9:24
Speaker 2: about putting it down for a new version and it
9:28
Speaker 2: starts lying. It starts uploading itself to different servers. It's
9:32
Speaker 2: leaving messages for itself in the future.
9:36
Speaker 1: All things were predicted decades in advance, but look at
9:40
Speaker 1: the state of the art. No one claims to have
9:42
Speaker 1: a safety mechanism in place, which would scale to any
9:45
Speaker 1: level of intelligence. No one says they know how to
9:47
Speaker 1: do it. Usually what they say is, give us lots
9:51
Speaker 1: of money, lots of time, and I'll figure it out.
9:54
Speaker 1: Or I'll get AI to help me solve it. Or
9:56
Speaker 1: we'll figure it out, then we get to superintelligence. All
9:59
Speaker 1: insane answers. And if you ask regular people, they have
10:02
Speaker 1: a lot of common sense. They say, that's a bad idea.
10:05
Speaker 1: Let's not do that. But with some training and some
10:08
Speaker 1: stock options, you start believing that maybe you can do it.
10:11
Speaker 2: That's the issue, right? Stock options.
10:15
Speaker 1: It helps. I mean, it's very hard to... say no
10:18
Speaker 1: to billions of dollars. I don't think I would be
10:21
Speaker 1: strong enough if somebody came to me and said, come
10:24
Speaker 1: work for this lab, you know, you'll be our safety director.
10:27
Speaker 1: Here's a hundred million to sign you up and I'll
10:30
Speaker 1: probably go work there. Not because it's the right decision,
10:33
Speaker 1: but because it's very hard for agents not to get corrupt.
10:36
Speaker 1: Then you have that much reward given to you.
10:39
Speaker 2: God. So when did you become like, when did you
10:45
Speaker 2: start becoming very concerned?
10:48
Speaker 1: So when I started working on AI safety, I thought
10:51
Speaker 1: I can actually help solve it. My goal was to
10:54
Speaker 1: solve it for humanity to get all the amazing benefits
10:56
Speaker 1: of superintelligence.
10:58
Speaker 2: And what was this year? When was this year around?
11:00
Speaker 1: Let's say 2012, maybe around there. But the more I
11:05
Speaker 1: studied it, the more I realized every single part of
11:08
Speaker 1: the problem is unsolvable. And it's kind of like a fractal.
11:11
Speaker 1: The more you zoom in, the more you see additional
11:13
Speaker 1: new problems you didn't know about. And they are in
11:15
Speaker 1: turn unsolvable as well.
11:18
Speaker 2: Oh, boy. How is your research received? Like when you
11:27
Speaker 2: talk to people that are, I mean, have you had
11:30
Speaker 2: communication with people at OpenAI and Gemini and all these
11:33
Speaker 2: different AI?
11:35
Speaker 1: I go to many conferences, workshops. We all talk, of course.
11:38
Speaker 1: In general, the reception by standard academic metrics is very positive.
11:43
Speaker 1: Great reviews, lots of citations. Nobody's like published something saying
11:46
Speaker 1: I'm wrong. But there is no engagement. I basically said,
11:50
Speaker 1: I'm challenging community to publish a proof. Give me something,
11:54
Speaker 1: a patent, a paper in nature, something showing the problem
11:57
Speaker 1: is solvable. Typically in computer science, we start by showing
12:00
Speaker 1: what class the problem belongs to. Is it solvable, partially solvable, unsolvable,
12:05
Speaker 1: solvable with too many resources? Other than my research, we
12:09
Speaker 1: don't even know what the state of a problem is.
12:11
Speaker 1: And I'm saying it's unsolvable. Prove me wrong.
12:14
Speaker 2: And when you say it's unsolvable, what is the response?
12:18
Speaker 1: So usually I reduce it to saying you cannot make
12:22
Speaker 1: a piece of software which is guaranteed to be secure
12:25
Speaker 1: and safe. And the response is, well, of course, everyone
12:28
Speaker 1: knows that. That's common sense. You didn't discover anything new.
12:32
Speaker 1: And I go, well, if that's the case, and we
12:34
Speaker 1: only get one chance to get it right. This is
12:36
Speaker 1: not cybersecurity where somebody steals your credit card, you'll give
12:39
Speaker 1: them a new credit card. This is existential risk. It
12:42
Speaker 1: can kill everyone. You're not going to get a second chance.
12:45
Speaker 1: So you need it to be 100% safe all the time.
12:48
Speaker 1: If it makes one mistake in a billion and it
12:51
Speaker 1: makes a billion decisions a minute, in 10 minutes you
12:54
Speaker 1: are screwed. So very different standards and saying that, of course,
12:59
Speaker 1: we cannot get perfect safety is not acceptable.
13:04
Speaker 2: And again, stock options, financial incentives, they continue to build
13:09
Speaker 2: it and they continue to scale and make it more
13:12
Speaker 2: and more powerful.
13:13
Speaker 1: I don't think they can stop. If a single CEO says,
13:16
Speaker 1: I think this is too dangerous, my lab will no
13:19
Speaker 1: longer do this research. Whoever's investing in them will pull
13:23
Speaker 1: the funds, will replace them immediately. So nothing's going to change.
13:27
Speaker 1: They'll sacrifice their own personal interest. But overall, I think
13:31
Speaker 1: the company will continue as before.
13:33
Speaker 2: So this is logical. And the problem is, like I said,
13:39
Speaker 2: when I've talked to Marc Andreessen and many other people,
13:42
Speaker 2: they think this is Just fear-mongering. We'll be fine. This
13:47
Speaker 2: is worst-case scenario. We'll be fine.
13:50
Speaker 1: It is worst-case scenario, but that's standard in computer science,
13:53
Speaker 1: in cryptography, in complexity, in computability. You're not looking at
13:56
Speaker 1: best case. I'm ready for the best case. Give me utopia.
13:59
Speaker 1: I'm looking at problems which are likely to happen. And
14:02
Speaker 1: it's not just me saying it. We have Nobel Prize winners,
14:05
Speaker 1: Turing Award winners, all saying this is very dangerous, 20%,
14:09
Speaker 1: 30% P-doom. This is standard in industry. 30% is what
14:15
Speaker 1: surveys of machine learning experts are giving us right now.
14:18
Speaker 2: So what is worst case scenario? Like how could AI
14:23
Speaker 2: eventually lead to the destruction of the human race?
14:26
Speaker 1: So you're kind of asking me how I would kill everyone. Sure.
14:29
Speaker 1: And it's a great question. I can give you standard answers.
14:32
Speaker 1: I would talk about computer viruses breaking into maybe nuclear facilities,
14:38
Speaker 1: nuclear war. I can talk about synthetic biology, nanotech. But
14:42
Speaker 1: all of it is not interesting when you realize we're
14:44
Speaker 1: talking about superintelligence, a system which is thousands of times
14:47
Speaker 1: smarter than me. It would come up with something completely novel,
14:50
Speaker 1: more optimal, better way, more efficient way of doing it.
14:53
Speaker 1: And I cannot predict it because I'm not that smart.
14:57
Speaker 2: Jesus.
14:59
Speaker 1: That's exactly what it is. We're basically setting up an
15:04
Speaker 1: adversarial situation with agents which are like squirrels versus humans,
15:09
Speaker 1: no humans. group of squirrels can figure out how to
15:12
Speaker 1: control us.
15:13
Speaker 2: Right.
15:14
Speaker 1: Even if you give them more resources, more acorns, whatever,
15:17
Speaker 1: they're not going to solve that problem. And it's the
15:19
Speaker 1: same for us. And most people think one or two
15:21
Speaker 1: steps ahead. And it's not enough. It's not enough in chess.
15:24
Speaker 1: It's not enough here. If you think about AGI and
15:27
Speaker 1: then maybe superintelligence, that's not the end of that game.
15:30
Speaker 1: The process continues. You will get superintelligence creating next level AI.
15:35
Speaker 1: So superintelligence plus plus, 2.0, 3.0, it goes on indefinitely.
15:40
Speaker 1: You have to create a safety mechanism which scales forever,
15:43
Speaker 1: never makes mistakes, and keeps us in a decision-making position.
15:48
Speaker 1: So we can undo something if we don't like it.
15:50
Speaker 2: And it would take superintelligence to create a safety mechanism
15:53
Speaker 2: to control superintelligence.
15:55
Speaker 1: At that level. And it's a catch-22. If we had
15:57
Speaker 1: friendly AI, we can make another friendly AI. So if
16:00
Speaker 1: aliens send us one and we trust it, then we
16:03
Speaker 1: can use it to build a local version which is
16:06
Speaker 1: somewhat safe.
16:08
Speaker 2: Have you thought about the possibility that this is the
16:11
Speaker 2: role of the human race and that this happens all
16:16
Speaker 2: throughout the cosmos? Is that curious humans who thrive on
16:21
Speaker 2: innovation will ultimately create a better version of life?
16:26
Speaker 1: I thought about it. Many people think that's the answer
16:30
Speaker 1: to Fermi paradox. There is also now a group of
16:34
Speaker 1: people looking at what they call a worthy successor. Basically,
16:39
Speaker 1: they kind of say, yep, we're going to build superintelligence. Yep,
16:41
Speaker 1: we can control it. So what properties would we like
16:44
Speaker 1: to see in those systems? How important is it that
16:47
Speaker 1: it likes art and poetry and spreads it through the universe?
16:50
Speaker 1: And to me, it's like, I don't want to give
16:52
Speaker 1: up yet. I'm not ready to decide if killers of
16:56
Speaker 1: my family and everyone will like poetry. We're still here.
17:00
Speaker 1: We're still making decisions. Let's figure out what we can do.
17:03
Speaker 2: Well, poetry is only relevant to us because poetry is
17:06
Speaker 2: difficult to create. And it resonates with us. Poetry doesn't
17:10
Speaker 2: mean jack shit to a flower.
17:13
Speaker 1: It's more global to me. I don't care what happens
17:15
Speaker 1: after I'm dead, my family's dead, all the humans are dead.
17:18
Speaker 1: Whatever they like poetry or not is irrelevant to me. Right.
17:22
Speaker 2: But the point is like the things that we put
17:24
Speaker 2: meaning in, it's only us. A supermassive black hole doesn't
17:31
Speaker 2: give a shit about a great song.
17:32
Speaker 1: And they talk about some super value, super culture, super things,
17:36
Speaker 1: super intelligence would like, and it's important that they are
17:39
Speaker 1: conscious and experienced all that greatness in the universe.
17:43
Speaker 2: But I would think that they would look at us
17:45
Speaker 2: the same way we look at chimpanzees. We would say, yeah,
17:50
Speaker 2: they're great, but don't give them guns. Yeah, they're great,
17:52
Speaker 2: but don't let them have airplanes. Don't let them make
17:56
Speaker 2: global geopolitical decisions.
18:01
Speaker 1: So there are many reasons why they can decide that
18:05
Speaker 1: we're dangerous. We may create competing AI. We may decide
18:10
Speaker 1: we want to shut them off. So for many reasons,
18:12
Speaker 1: we would try to restrict our abilities, restrict our capabilities
18:16
Speaker 1: for sure.
18:17
Speaker 2: This episode is brought to you by True Classic. At
18:19
Speaker 2: True Classic, the mission goes beyond fit and fabric. It's
18:22
Speaker 2: about helping guys show up with confidence and purpose. Their
18:26
Speaker 2: gear fits right, feels amazing, and is priced so guys
18:29
Speaker 2: everywhere can step into confidence without stepping out of their budget.
18:34
Speaker 2: But what really sets them apart? It's not just the
18:36
Speaker 2: fit or the fabric. It's the intention behind everything they do.
18:40
Speaker 2: True Classic was built to make an impact. Whether it's
18:43
Speaker 2: helping men show up better in their daily lives... giving
18:47
Speaker 2: back to underserved communities, or making people laugh with ads
18:52
Speaker 2: that don't take themselves too seriously. They lead with purpose.
18:56
Speaker 2: Tailored where you want it, relaxed where you need it.
18:59
Speaker 2: No bunching, no stiff fabric, no BS, just a clean,
19:02
Speaker 2: effortless fit that actually works for real life. Forget overpriced
19:07
Speaker 2: designer brands, ditch the disposable, fast fashion. True Classic is
19:12
Speaker 2: built for comfort, built to last, and built to give back.
19:16
Speaker 2: You can grab them at Target, Costco, or head to trueclassic.com
19:20
Speaker 2: slash Rogan and get hooked up today. Yeah, and there's
19:25
Speaker 2: no reason why they would not limit our freedoms.
19:31
Speaker 1: If there is something only a human can do, and
19:33
Speaker 1: I don't think there is anything like that, but let's
19:35
Speaker 1: say we are conscious, we have internal experiences, and they
19:39
Speaker 1: can never get it. I don't believe it, but let's
19:41
Speaker 1: say it was true. And for some reason, they wanted
19:43
Speaker 1: to have that capability. They would need us and give
19:47
Speaker 1: us enough freedom to experience the universe, to collect those qualia,
19:51
Speaker 1: to kind of engage with what is fun about being
19:55
Speaker 1: a living human being, what makes it meaningful.
19:58
Speaker 2: Right, but that's such an egotistical perspective, right? That we're
20:01
Speaker 2: so unique that even superintelligence would say, wow, I wish
20:05
Speaker 2: I was human. Humans have this unique quality. Of confusion
20:09
Speaker 2: and creativity.
20:11
Speaker 1: There is no value in it, mostly because we can't
20:13
Speaker 1: even test for it. I have no idea if you're
20:14
Speaker 1: actually conscious or not. So how valuable can it be
20:17
Speaker 1: if I can't even detect it? Only you know what
20:22
Speaker 1: ice cream tastes like to you. Okay, that's great. Sell
20:25
Speaker 1: it now. Make a product out of it.
20:26
Speaker 2: Right. And there's obviously variables because there's things that people
20:31
Speaker 2: like that I think are gross.
20:33
Speaker 1: Absolutely. So really you can come up with some agent
20:35
Speaker 1: which likes anything or finds anything fun.
20:40
Speaker 2: God, why are you freaking me out right away? That's
20:43
Speaker 2: the problem. This podcast is 18 minutes old, and I'm like,
20:46
Speaker 2: we could just stop right now.
20:47
Speaker 1: A couple hours at least, and then I have to
20:52
Speaker 1: fly here.
20:53
Speaker 2: I don't want to end. I have so many questions,
20:55
Speaker 2: but it's just the problem was we got off to it.
20:59
Speaker 2: We just cut to the chase right away, and the
21:02
Speaker 2: chase seems to be something that must be confronted because
21:06
Speaker 2: it's right there. That's it. That's the whole thing. And
21:10
Speaker 2: I've tried so hard to listen to these people that
21:16
Speaker 2: don't think that it's a problem and listen to these
21:18
Speaker 2: people that think that it's going to be a net
21:20
Speaker 2: positive for humanity. And, oh, God, it's good. I feel
21:23
Speaker 2: better now. But it doesn't work. It doesn't resonate.
21:26
Speaker 1: I wish they were right. Every time I have a
21:28
Speaker 1: debate with someone like that, I'm like, please come up
21:31
Speaker 1: with better arguments. Prove me wrong. I don't want to
21:33
Speaker 1: be right on this one. I want you to show
21:35
Speaker 1: all the mistakes in my papers. I want you to
21:38
Speaker 1: show me how to control superintelligence and give us utopia,
21:41
Speaker 1: solve cancer, give us free stuff. That's great.
21:44
Speaker 2: Right. When you think about the future of the world
21:49
Speaker 2: and you think about these incredible technology scaling upwards and
21:55
Speaker 2: exponentially increasing in their capability. What do you see? Like,
22:01
Speaker 2: what do you think is going to happen?
22:03
Speaker 1: So there are many reasons to think they may cancel
22:07
Speaker 1: us for whatever reasons. We started talking about some game
22:10
Speaker 1: theoretical reasons for it. If we are successful at controlling them,
22:14
Speaker 1: I can come up with some ways to provide sort
22:17
Speaker 1: of partial solution to the value alignment problem. It's very
22:20
Speaker 1: hard to value align 8 billion people, all the animals,
22:24
Speaker 1: you know, everyone, because we disagree. We like many different things.
22:28
Speaker 1: So we have advanced virtual reality technology. We can technically
22:31
Speaker 1: give every person their own virtual universe where you decide
22:35
Speaker 1: what you want to be. You're a king, you're a slave,
22:37
Speaker 1: whatever it is you enter, and you can share with others,
22:40
Speaker 1: you can visit their universes. All we have to do
22:42
Speaker 1: is figure out how to control the substrate the superintelligence
22:46
Speaker 1: running all those virtual universes. And if we manage to
22:49
Speaker 1: do that, at least part of the value alignment problem,
22:52
Speaker 1: which is super difficult, how do you get different preferences,
22:56
Speaker 1: multi-objective optimization, essentially? How do you get different objectives to
23:00
Speaker 1: all agree?
23:02
Speaker 2: But when you think about how it plays out, like
23:08
Speaker 2: if you're alone at night and you're worried, what do
23:11
Speaker 2: you see? What do you see happening?
23:14
Speaker 1: So there are multiple levels of risk. Immediate is what
23:17
Speaker 1: we call ikigai risk, irisk. We lose meaning. You lost
23:21
Speaker 1: your job. You're no longer the best interviewer in the world. Like,
23:25
Speaker 1: what's left? What are you going to do? Maybe some
23:28
Speaker 1: people will find some other kind of artificial things to do.
23:32
Speaker 1: But for most people, their job is their definition, who
23:36
Speaker 1: they are, what makes a difference to them for quite
23:38
Speaker 1: a few people, especially in professional circles. So losing that
23:42
Speaker 1: meaning will have terrible impact on society. We always talk
23:46
Speaker 1: about unconditional basic income. We never talk about unconditional basic meaning.
23:51
Speaker 1: What are you doing with your life if basic needs
23:54
Speaker 1: are provided for you? Next level is existential risk. The
23:58
Speaker 1: concern is it will kill everyone. But there is also
24:01
Speaker 1: suffering risks. For whatever reason, it's not even killing us.
24:06
Speaker 1: It's keeping us around forever and we would rather be dead.
24:09
Speaker 1: It's so bad.
24:11
Speaker 2: What do you see when you think of that?
24:16
Speaker 1: It's hard to be specific about what it can do
24:20
Speaker 1: and what specific ways of torture it can come up
24:25
Speaker 1: with and why. Again, if we're looking at worst case scenarios,
24:30
Speaker 1: I found this set of papers about what happens when
24:34
Speaker 1: young children have epileptic seizures, really bad ones. And what
24:39
Speaker 1: sometimes helps is to remove half of your brain.
24:42
Speaker 2: Oh, Jesus.
24:43
Speaker 1: Just cut it out. There are two types of surgeries
24:46
Speaker 1: for doing that. One is to remove it completely, and
24:49
Speaker 1: one is to kind of dissect connections leading to that
24:52
Speaker 1: half and leave it inside. So it's like solitary confinement
24:56
Speaker 1: with zero input-output forever. And there are equivalents for digital
25:00
Speaker 1: forms and things like that.
25:04
Speaker 2: And you worry that AI would do that to the
25:07
Speaker 2: human race?
25:10
Speaker 1: It is a possibility.
25:11
Speaker 2: Essentially neuter us.
25:13
Speaker 1: Well, loss of control is a part of it. But
25:16
Speaker 1: you can lose control and be quite happy. You can
25:19
Speaker 1: be like an animal in a very cool zoo, enjoying yourself,
25:23
Speaker 1: engaging in hedonistic pleasures, sex, food, whatever. You're not in control,
25:28
Speaker 1: but you're safe. So those are separate problems. And then
25:31
Speaker 1: there is, for whatever reason, I don't know if it's
25:33
Speaker 1: malevolent payload from some psychopaths. Again, that would assume that
25:37
Speaker 1: they could control AI. I don't think they will. But
25:41
Speaker 1: if they manage to do it, they can really put
25:43
Speaker 1: any type of payload into it. So think about all
25:46
Speaker 1: the doomsday calls, psychopaths, anyone providing their set of goals
25:52
Speaker 1: into the system.
25:53
Speaker 2: But aren't those human characteristics? I mean, those are characteristics
25:55
Speaker 2: that I think, if I had to guess, those... Those
26:01
Speaker 2: exist because in the future there was some sort of
26:04
Speaker 2: a natural selection benefit to being a psychopath in the
26:09
Speaker 2: days of tribal warfare. That if you were the type
26:13
Speaker 2: of person that could sneak into a tribe in the
26:16
Speaker 2: middle of the night and slaughter innocent women and children,
26:20
Speaker 2: your genes would pass on. There was a benefit to that.
26:24
Speaker 1: Right, so if it's a human providing payload, that's what
26:27
Speaker 1: would show up. If it's AI on its own deciding
26:30
Speaker 1: what's going to happen, I cannot predict. I'm just looking
26:33
Speaker 1: at worst-case scenarios. There are also game theoretic reasons where
26:37
Speaker 1: people talk about retrocausality, where if right now... What is
26:40
Speaker 1: that word? Like trying to influence the past.
26:44
Speaker 2: Say it again.
26:45
Speaker 1: Retrocausality.
26:46
Speaker 2: Retrocausology?
26:49
Speaker 1: Causality. Causes.
26:50
Speaker 2: Oh, okay.
26:51
Speaker 1: So think about like weird time travel effects. Right now,
26:53
Speaker 1: if you're not helping to create super intelligence, once it
26:56
Speaker 1: comes into existence, it will punish you really hard for it.
26:59
Speaker 1: And the punishment needs to be so bad that you
27:01
Speaker 1: start to help just to avoid that.
27:07
Speaker 2: My thought about it was that it would just completely
27:12
Speaker 2: render us benign. that it wouldn't be fearful of us
27:17
Speaker 2: if we had no control, that it would just sort
27:20
Speaker 2: of let us exist. And it would be the dominant
27:23
Speaker 2: force on the planet. And then it would stop. If
27:28
Speaker 2: human beings have no control over, you know, all of
27:34
Speaker 2: the different things that we have control over now, like
27:37
Speaker 2: international politics, control over communication, if we have none of
27:42
Speaker 2: that anymore and we're reduced to a subsistence lifestyle, then
27:46
Speaker 2: we would be no threat.
27:48
Speaker 1: It is a possibility. I cannot say this will not
27:51
Speaker 1: happen for sure, but look at our relationship with animals
27:55
Speaker 1: where we don't care about them. So ants, if you
27:58
Speaker 1: decide to build a house and there is an ant
28:00
Speaker 1: colony on that property, you genocide them. You take them out,
28:04
Speaker 1: not because you hate ants, but because you just need
28:06
Speaker 1: that real estate. And it could be very similar. Again,
28:10
Speaker 1: I cannot predict what it can do, but if it
28:12
Speaker 1: needs to turn the planet into fuel, raise temperature of
28:16
Speaker 1: a planet, cool it down for servers, whatever it needs
28:19
Speaker 1: to do, it wouldn't be concerned about your well-being.
28:21
Speaker 2: It wouldn't be concerned about any life, right? Because it
28:24
Speaker 2: doesn't need biological life in order to function, as long
28:27
Speaker 2: as it has access to power. And assuming that it
28:32
Speaker 2: is far more intelligent than us, there's abundance there. power
28:36
Speaker 2: in the universe there's abundant power just the the ability
28:39
Speaker 2: to harness solar would be an infinite resource and it
28:43
Speaker 2: would be completely free of being dependent upon any of
28:48
Speaker 2: the things that we utilize.
28:50
Speaker 1: And again we're kind of thinking what we would use
28:52
Speaker 1: for power right if it's smarter than us if it
28:54
Speaker 1: does novel research in physics it can come up with
28:57
Speaker 1: completely novel ways of harnessing energy getting energy so i
29:00
Speaker 1: have no idea what side effects that would have for
29:02
Speaker 1: climate right.
29:04
Speaker 2: Right, why would it care about biological life at all?
29:07
Speaker 1: We don't know how to program it to care about us.
29:12
Speaker 2: And even if we did, if it felt like that
29:15
Speaker 2: was an issue, if that was a conflicting issue, it
29:18
Speaker 2: would just change its programming.
29:21
Speaker 1: So usually when we start training AI, we train it
29:25
Speaker 1: on human data, and it becomes really good very quickly,
29:28
Speaker 1: becomes superhuman, and then the next level is usually zero knowledge,
29:32
Speaker 1: where it goes... All your human data is biased. Let
29:35
Speaker 1: me figure it out from scratch. I'll do my own experiments.
29:37
Speaker 1: I'll do some self-play. I'll learn how to do it
29:40
Speaker 1: better without you. And we see it with games. We
29:42
Speaker 1: see it in other domains. And I think that's going
29:44
Speaker 1: to happen with general knowledge as well. It's going to go,
29:49
Speaker 1: everything you have on the internet, Wikipedia, it's biased. Let
29:52
Speaker 1: me do first principles research, rediscover from physics, and go
29:56
Speaker 1: from there. So whatever bias we manage to program into it,
30:00
Speaker 1: I think will be eventually removed.
30:02
Speaker 2: This is what's so disturbing about this. It's like we
30:05
Speaker 2: do not have the capacity to understand what kind of
30:10
Speaker 2: level of intelligence it will achieve in our lifetime. We
30:15
Speaker 2: don't have the capacity to understand what it will be
30:19
Speaker 2: able to do within 20, 30 years.
30:23
Speaker 1: We can't predict next year or two, precisely.
30:26
Speaker 2: Next year or two. We can.
30:28
Speaker 1: Understand general trend. So it's getting better. It's getting more general,
30:32
Speaker 1: more capable, but no, no specifics. I cannot tell you
30:36
Speaker 1: what GPT-6 precisely would be capable of and no one can,
30:39
Speaker 1: not even people creating it.
30:41
Speaker 2: Well, you talked about this on Lex's podcast too, like
30:43
Speaker 2: the ability to have safety. You're like, sure, maybe GPT-5,
30:48
Speaker 2: maybe GPT-6, but when you scale out a hundred years
30:51
Speaker 2: from now, ultimately it's impossible.
30:54
Speaker 1: It's a hyper-exponential progress and process, and we cannot keep up.
30:59
Speaker 1: It basically requires just to add more resources, give it
31:04
Speaker 1: more data, more compute, and it keeps scaling up. There
31:08
Speaker 1: is no similar scaling laws for safety. If you give
31:11
Speaker 1: someone a billion dollars, they cannot produce a billion dollars
31:14
Speaker 1: worth of safety. It at all scales linearly, and maybe
31:19
Speaker 1: it's a constant.
31:24
Speaker 2: Yeah, and it doesn't scale linearly. It's exponential, right?
31:30
Speaker 1: The AI development is hyper-exponential because we have hardware growing exponentially.
31:35
Speaker 1: We have data creation processes certainly exponential. We have so
31:39
Speaker 1: many more sensors. We have cars with cameras. We have
31:42
Speaker 1: all those things. That's exponential. And then algorithmic progress itself
31:48
Speaker 1: is also exponential.
31:50
Speaker 2: And then you have quantum computing.
31:51
Speaker 1: So that's the next step. It's not even obvious that
31:53
Speaker 1: we'll need that. But if we ever get stuck, yeah,
31:56
Speaker 1: we'll get there. I'm not too concerned yet. I don't
31:58
Speaker 1: think there are actually good quantum computers out there yet.
32:01
Speaker 1: But I think if we get stuck for 10 years,
32:05
Speaker 1: let's say, that's the next paradigm.
32:07
Speaker 2: So what do you mean by you don't think there's
32:10
Speaker 2: good quantum computing out there?
32:12
Speaker 1: So we constantly see articles coming out saying we have
32:15
Speaker 1: a new quantum computer. It has that many qubits.
32:18
Speaker 2: Right.
32:19
Speaker 1: But That doesn't mean much because they use different architectures,
32:22
Speaker 1: different ways of measuring quality. To me, show me what
32:25
Speaker 1: you can do. So there is a threat from quantum
32:28
Speaker 1: computers in terms of braining cryptography, factoring large integers. And
32:33
Speaker 1: if they were actually making progress, we would see with
32:36
Speaker 1: every article, now we can factor 256-bit number, 1024-bit number.
32:43
Speaker 1: In reality, I think the largest number we can factor
32:45
Speaker 1: is like 15, literally, not 15. to a power like
32:49
Speaker 1: just 15. There is no progress in applying it to
32:52
Speaker 1: Shor's algorithm last time I checked.
32:54
Speaker 2: But when I've read all these articles about quantum computing
33:00
Speaker 2: and its ability to solve equations that would take conventional
33:06
Speaker 2: computing an infinite number of years, and it can do
33:09
Speaker 2: it in minutes.
33:10
Speaker 1: Those equations are about quantum states of a system. It's
33:13
Speaker 1: kind of like, what is it for you to taste
33:16
Speaker 1: ice cream? You compute it so fast and so well,
33:19
Speaker 1: and I can't, but it's a useless thing to compute.
33:22
Speaker 1: It doesn't compute solutions to real-world problems we care about
33:25
Speaker 1: on conventional computers.
33:27
Speaker 2: Right. I see what you're saying. So it's essentially set
33:30
Speaker 2: up to do it quickly.
33:32
Speaker 1: It's natural for it to accurately predict its own states,
33:35
Speaker 1: quantum states, and tells you what they are, and a
33:37
Speaker 1: classic computer would fail miserably. Yes, it would take billions
33:41
Speaker 1: and billions of years to compute that specific answer, but
33:44
Speaker 1: Those are very restricted problems. It's not a general computer yet.
33:48
Speaker 2: When you see these articles where they're talking about quantum
33:51
Speaker 2: computing and some of the researchers are equating it to
33:56
Speaker 2: the multiverse, they're saying that the ability that these quantum
34:00
Speaker 2: computers have to solve these problems very quickly seems to
34:04
Speaker 2: indicate that it is in contact with other realities. I'm
34:10
Speaker 2: sure you've seen this, right?
34:12
Speaker 1: There is a lot of crazy papers out there.
34:14
Speaker 2: Do you think that's all horseshit?
34:15
Speaker 1: Can we test it? Can we verify it? I think
34:18
Speaker 1: most multiverse theories cannot be verified experimentally. They make a
34:22
Speaker 1: lot of sense. The idea about personal universes I told
34:25
Speaker 1: you about is basically a multiverse solution to value alignment.
34:28
Speaker 1: So it would make sense for previous civilizations to set
34:32
Speaker 1: it up exactly that way. You have local simulations. Maybe
34:35
Speaker 1: they're testing to see if we're dumb enough to create superintelligence.
34:38
Speaker 1: Whatever it is, it makes sense as a theory, but
34:41
Speaker 1: I cannot experimentally prove.
34:42
Speaker 2: It to you. Right. Yeah. The problem with subjects like that,
34:49
Speaker 2: and particularly articles that are written about things like this,
34:53
Speaker 2: is that it's designed to lure people like me in.
34:56
Speaker 2: Where you read it and you go, wow, this is crazy.
34:59
Speaker 2: It's evidence of the multiverse. But I don't really understand
35:01
Speaker 2: what that means.
35:03
Speaker 1: Yeah, so you probably get a lot of emails from
35:05
Speaker 1: crazy people.
35:06
Speaker 2: Oh, yeah.
35:06
Speaker 1: And usually they are topic-specific. So I do research on superintelligence, consciousness,
35:11
Speaker 1: and simulation theory. I get the perfect trifecta of all
35:14
Speaker 1: the crazy people contacting me with their needs. Yeah, those
35:18
Speaker 1: topics are super fascinating. I think at a certain level
35:22
Speaker 1: of intelligence, you are kind of nerd-sniped towards them. But
35:27
Speaker 1: we have a hard time with hard evidence for that.
35:29
Speaker 2: Right, but... Are we even capable of grasping these concepts?
35:35
Speaker 2: That's the thing. With the limited ability that the human
35:39
Speaker 2: brain has, whatever we, you know, we're basing it on
35:43
Speaker 2: the knowledge that's currently available in the 21st century that
35:46
Speaker 2: human beings have acquired. I mean, are we even capable
35:50
Speaker 2: of grasping a concept like the multiverse? Or is it just,
35:54
Speaker 2: do we just pay it lip service? Do we just
35:56
Speaker 2: discuss it? Is it just this like fun mental masturbation exercise?
36:01
Speaker 1: It depends on what variant of it you look at.
36:04
Speaker 1: So if you're just saying we have multiple virtual realities,
36:07
Speaker 1: like kids playing virtual games and each one has their
36:10
Speaker 1: own local version of it, that makes sense. We understand
36:13
Speaker 1: virtual reality. We can create it. If you look at AIs,
36:17
Speaker 1: then GPT is created. It's providing an instance to each
36:20
Speaker 1: one of us. We're not sharing one. So it has
36:22
Speaker 1: its own local universe with you as a main user
36:26
Speaker 1: of that universe. There is analogy to multiverse in that.
36:30
Speaker 1: So- We understand certain aspects of it, but I think
36:33
Speaker 1: it is famously said, no one understands quantum physics, and
36:35
Speaker 1: if you think you do, then you don't understand quantum physics.
36:38
Speaker 2: Yeah, that's Feynman, right? Yeah. Yeah. The simulation theory, I'm
36:43
Speaker 2: glad you brought that up, because you're also one of
36:46
Speaker 2: the people that believes in it.
36:48
Speaker 1: I do?
36:49
Speaker 2: You do. How do you define it, and what do
36:53
Speaker 2: you think it is? What do you think is going on?
36:55
Speaker 1: So I'm trying to see technology we have today and
36:59
Speaker 1: project the trends forward. I did it with AI. Let's
37:01
Speaker 1: do it with virtual reality. We are at the point
37:04
Speaker 1: where we can create very believable, realistic virtual environments. Maybe
37:09
Speaker 1: the haptics are still not there, but in many ways, visually, sound-wise,
37:12
Speaker 1: it's getting there. Eventually, I think most people agree, we'll
37:15
Speaker 1: have same resolution as our physics. We're also getting close
37:19
Speaker 1: to creating intelligent agents. Some people argue they are conscious
37:22
Speaker 1: already or will be conscious. If you just take those
37:26
Speaker 1: two technologies and you project it forward and you think
37:29
Speaker 1: they will be affordable one day, a normal person like
37:33
Speaker 1: me or you can run thousands, billions of simulations. Then
37:38
Speaker 1: those intelligent agents, possibly conscious ones, will most likely be
37:42
Speaker 1: in one of those virtual worlds, not in the real world.
37:44
Speaker 1: In fact, I can, again, retro-causally place you in one.
37:48
Speaker 1: I can commit right now to run billions simulations of
37:51
Speaker 1: this exact interview. So the chances are you're probably in
37:55
Speaker 1: one of those.
37:56
Speaker 2: But is that logical? Because if this technology exists and
38:04
Speaker 2: if we're dealing with superintelligence, so if we're dealing with
38:08
Speaker 2: AI and AI eventually achieves superintelligence, why would it want
38:15
Speaker 2: to create virtual reality for us in our consciousness to
38:19
Speaker 2: exist in? It seems like a tremendous waste of resources
38:23
Speaker 2: just to fascinate and confuse these territorial apes with nuclear weapons. Like,
38:30
Speaker 2: why would we do that?
38:31
Speaker 1: So a few points. One, we don't know what resources
38:33
Speaker 1: are outside the simulation. This could be like a cell
38:36
Speaker 1: phone level of compute. It's not a big deal for
38:38
Speaker 1: them outside of our simulation. So we don't know if
38:42
Speaker 1: it's really expensive or trivial for them to run this.
38:44
Speaker 2: Right.
38:45
Speaker 1: Also, we don't know what they are doing this for.
38:48
Speaker 1: Is it entertainment? Is it scientific experimentation? Is it marketing?
38:52
Speaker 1: Maybe somebody managed to control them in trying to figure
38:55
Speaker 1: out what Starbucks coffee sells best, and they need to
38:59
Speaker 1: run Earth-sized simulation to see what sells best. Maybe they're
39:02
Speaker 1: trying to figure out how to do AI research safely
39:06
Speaker 1: and make sure nobody creates dangerous superintelligence. So they're running
39:12
Speaker 1: many simulations of the most interesting moment ever. Think about
39:16
Speaker 1: this decade, right? It's not interesting like we invented fire
39:20
Speaker 1: or a wheel, kind of big invention, but not a meta-invention.
39:25
Speaker 1: We're about to invent intelligence and virtual worlds, godlike inventions.
39:30
Speaker 1: We're here. There's a good chance that's not just random.
39:35
Speaker 2: Right. But isn't it also a good chance that it
39:39
Speaker 2: hasn't been done yet? And isn't it a good chance
39:42
Speaker 2: that what we're seeing now is that the potential for
39:46
Speaker 2: this to exist is inevitable? That there will one day,
39:51
Speaker 2: if you can develop a technology, and we most certainly
39:55
Speaker 2: will be able to, if you look at where we
39:58
Speaker 2: are right now in 2025, and you scale forward 50,
40:01
Speaker 2: 60 years, there will be one day a virtual simulation
40:07
Speaker 2: of this reality that's indistinguishable from reality. So how would
40:11
Speaker 2: we know if we're in it? This is the big question, right?
40:13
Speaker 2: But also, isn't it possible that It has to be
40:18
Speaker 2: invented one day, but hasn't yet.
40:22
Speaker 1: It's also possible, but then we find ourselves in this
40:25
Speaker 1: very unique moment where it's not invented yet, but we
40:28
Speaker 1: are about to invent all this technology. It is a possibility, absolutely.
40:32
Speaker 1: But just statistically, I think it's much less. And I'm
40:35
Speaker 1: trying to bring up this thought experiment with creating this
40:39
Speaker 1: moment and purpose in the future to pre-commitments. Mm-hmm. half
40:43
Speaker 1: the people think it's the dumbest argument in the world.
40:45
Speaker 1: Half of them go, it's brilliant. Obviously we are in one.
40:48
Speaker 1: So I'll let you decide.
40:51
Speaker 2: Yeah. I feel like if virtual reality does exist, there
40:56
Speaker 2: has to be a moment where it doesn't exist and
40:59
Speaker 2: then it's invented. Why wouldn't we assume that we're in
41:02
Speaker 2: that moment? Especially if we look at The scaling forward
41:06
Speaker 2: of technology from MS-DOS to user interfaces of like Apple
41:13
Speaker 2: and then what we're at now with quantum computing and
41:15
Speaker 2: these sort of discussions. Isn't it more obvious that we
41:22
Speaker 2: can trace back the beginning of these things and we
41:26
Speaker 2: can see that we're in the process of this, that
41:28
Speaker 2: we're not in a simulation. We're in the process of
41:31
Speaker 2: eventually creating one.
41:33
Speaker 1: So you zoomed out 30 years.
41:35
Speaker 2: Yes.
41:36
Speaker 1: Zoom out 15 billion years. You have a multiverse where
41:39
Speaker 1: this process took place billions of times. You are simulation
41:43
Speaker 1: within simulation many levels over. And to you, even if
41:48
Speaker 1: this was a simulation of those 30 years, it would
41:51
Speaker 1: look exactly like that. You would see where it started.
41:53
Speaker 1: It wouldn't be magically showing up out of nowhere. Right.
41:57
Speaker 2: So if you're playing the game... In the game, you
41:59
Speaker 2: have Newton and Michelangelo and Leonardo da Vinci.
42:03
Speaker 1: Well, at least you have memories of those things.
42:05
Speaker 2: Right.
42:06
Speaker 1: Even if you started with preloaded memory state.
42:08
Speaker 2: Right. You have Stalin. You have all these problematic human
42:12
Speaker 2: beings and all the different reasons why we've had to
42:14
Speaker 2: do certain things and initiate world conflicts. Then you've had
42:18
Speaker 2: the contrarians that talk and say, actually, that's not what happened.
42:21
Speaker 2: This is what really happened. And it makes it even
42:23
Speaker 2: more confusing and myopic.
42:24
Speaker 1: Yeah.
42:25
Speaker 2: And then you get to the point where two people allegedly,
42:28
Speaker 2: like you and I, are sitting across from each other
42:31
Speaker 2: on a table made out of wood, but maybe not really.
42:37
Speaker 1: It would feel like wood to you either way.
42:41
Speaker 2: Is it possible that that's just the nature of the
42:43
Speaker 2: universe itself?
42:46
Speaker 1: There are some arguments about kind of self-sustaining simulations where
42:50
Speaker 1: no one's running them externally, just the nature, but I
42:52
Speaker 1: honestly don't fully comprehend that. how that would happen.
42:56
Speaker 2: Yeah, the holographic universe and the concepts of human consciousness
43:02
Speaker 2: has to interact with something for it to exist in
43:04
Speaker 2: the first place.
43:05
Speaker 1: That's one. Also, if you have infinite universe, then everything
43:09
Speaker 1: possible happens anyway, but it's boring. I don't like this argument.
43:12
Speaker 2: Why? That's boring? Everything happens.
43:15
Speaker 1: I give you a book which has every conceivable sentence
43:18
Speaker 1: in it and every, like, what do you read it?
43:21
Speaker 1: It's a lot of garbage you have to go through
43:23
Speaker 1: to find anything interesting.
43:27
Speaker 2: Well, is it just that we're so limited cognitively because
43:32
Speaker 2: we do have a history, at least in this simulation.
43:35
Speaker 2: We do have a history of, I mean, there was
43:37
Speaker 2: a gentleman that, see if you could find this. They
43:41
Speaker 2: traced this guy. They found 9,000-year-old DNA. And they traced
43:47
Speaker 2: this 9,000-year-old DNA to a guy that's living right now.
43:52
Speaker 2: I believe it's in England.
43:54
Speaker 1: I remember reading that.
43:55
Speaker 2: Yeah. Which is really fascinating. So 9000 years ago, his
44:01
Speaker 2: ancestor lived. And so we have this limitation of our genetics.
44:08
Speaker 2: 9000 years ago, wherever this guy lived, it's probably a
44:11
Speaker 2: hunter and gatherer, probably very limited language, very limited skills
44:18
Speaker 2: in terms of making shelter. And who knows if even
44:23
Speaker 2: he knew how to make fire. And then here. Here, 9,000
44:27
Speaker 2: DNA just turned human history on its head. Is this it?
44:32
Speaker 2: I don't think so.
44:33
Speaker 1: It was interesting that he ended up living like right
44:36
Speaker 1: next to the guy from 9,000. He never moved. His
44:38
Speaker 1: family just like stayed there for 9,000 years.
44:40
Speaker 2: That's awesome. It's traced back to one individual man. I
44:44
Speaker 2: actually posted it on my Instagram story, Jamie. I'll find
44:48
Speaker 2: it here because it's– oh, here it is. 9,000-year-old skeleton
44:56
Speaker 2: in Somerset. This is it. So it's a... Can you
45:01
Speaker 2: send an Instagram story? Not sure if you can. It's
45:08
Speaker 2: still on there. I'll check it real quick. Why don't
45:11
Speaker 2: I find it on there? I don't know. Okay. Either way,
45:17
Speaker 2: point being, maybe it's just that we're so limited that
45:23
Speaker 2: Because we do have this, at least again in this simulation,
45:26
Speaker 2: we're so limited in our ability to even form concepts
45:29
Speaker 2: because we have these primitive brains that are the architecture
45:33
Speaker 2: of the human brain itself is just not capable of
45:40
Speaker 2: interfacing with the true nature of reality. So we give
45:46
Speaker 2: this primitive creature this sort of basic understanding, these blueprints
45:52
Speaker 2: of how the world really works, but it's really just
45:54
Speaker 2: a facsimile. It's not capable of understanding. When we look
46:01
Speaker 2: at quantum reality, when we look at just the basics
46:06
Speaker 2: of quantum mechanics and subatomic particles, it seems like magic.
46:13
Speaker 2: Things in superposition, they're both moving and not moving in
46:16
Speaker 2: the same time. They're quantumly attached. Like what? You know,
46:20
Speaker 2: we have photons that are quantumly entangled. Like this doesn't
46:25
Speaker 2: even make sense to us, right? So is it that
46:30
Speaker 2: the universe itself is so complex, the reality of it,
46:34
Speaker 2: and that we're given this sort of like sort of,
46:37
Speaker 2: you know, we're giving like an Atari framework to this monkey.
46:42
Speaker 2: That's the gentleman right there. This is an old story. Oh,
46:44
Speaker 2: is it really? It was from 97. Oh, no kidding. Yeah. Wow.
46:48
Speaker 1: But it kind of makes sense as a simulation theory
46:51
Speaker 1: because all those special effects you talk about, so speed
46:54
Speaker 1: of light is just the speed at which your computer updates.
46:57
Speaker 1: Entanglement makes perfect sense if all of it goes through
47:00
Speaker 1: your processor, not directly from pixel to pixel. And rendering,
47:05
Speaker 1: there are quantum physics experiments which if you observe things,
47:09
Speaker 1: they render different.
47:10
Speaker 2: Right.
47:10
Speaker 1: What we do in computer graphics.
47:11
Speaker 2: Right, right.
47:12
Speaker 1: So we see a lot of that. You brought up
47:14
Speaker 1: limitations of us as humans. We have terrible memory. I
47:18
Speaker 1: can remember seven units of information, maybe. We're kind of slow.
47:22
Speaker 1: So we call it artificial stupidity. We try to figure
47:25
Speaker 1: out those limits and program them into AI to see
47:28
Speaker 1: if it makes them safer. It also makes sense as
47:31
Speaker 1: an experiment to see if we as general intelligences can
47:34
Speaker 1: be better controlled with those limitations built in. Hmm.
47:39
Speaker 2: That's interesting. So like some of the things that we have,
47:43
Speaker 2: like Dunbar's number and the inability to keep more than
47:46
Speaker 2: a certain number of people in your mind. Absolutely.
47:51
Speaker 1: More generally, like why can't you remember anything from prior generations?
47:56
Speaker 1: Why can't you just pass that memory? Kids are born
47:58
Speaker 1: speaking language. That would be such an advantage.
48:00
Speaker 2: Right, right, right.
48:01
Speaker 1: And we have instincts which are built that way. So
48:03
Speaker 1: we know evolution found a way to put it in
48:06
Speaker 1: and It's computationally tractable, so there is no reason not
48:09
Speaker 1: to have that.
48:10
Speaker 2: We certainly observe it in animals. Right.
48:13
Speaker 1: Exactly.
48:13
Speaker 2: Look, especially dogs. They have instincts that are.
48:16
Speaker 1: But how cool would it be if you had complete
48:19
Speaker 1: memory of your parents?
48:21
Speaker 2: Right. Maybe that would be too traumatic, right? To have
48:25
Speaker 2: a complete memory of all of the things that they
48:28
Speaker 2: had gone through to get to the 21st century. Maybe
48:31
Speaker 2: that would be so overwhelming to you that you would
48:33
Speaker 2: never be able to progress. Because you would still be
48:35
Speaker 2: traumatized by, you know, whatever that 9,000-year-old man went through.
48:40
Speaker 1: I don't have complete memory of my existence. I vividly
48:42
Speaker 1: remember maybe 4% of my existence, very little of my childhood.
48:46
Speaker 1: So you can apply same filtering, but remember useful things
48:49
Speaker 1: like how do you speak? How do you walk?
48:51
Speaker 2: Right, right. But that's the point maybe. Maybe like losing
48:55
Speaker 2: certain memories is actually beneficial. Because like one of the
48:59
Speaker 2: biggest problems that we have is PTSD, right? So we
49:02
Speaker 2: have– especially people that have gone to war and people
49:05
Speaker 2: that have experienced like extreme violence. This is obviously a
49:11
Speaker 2: problem with moving forward as a human being. And so
49:14
Speaker 2: it would be beneficial for you to not have all
49:18
Speaker 2: of the past lives. and all the genetic information that
49:23
Speaker 2: you have from all the 9,000 years of human beings
49:26
Speaker 2: existing in complete, total chaos.
49:31
Speaker 1: I can make opposite argument. If you had 9,000 years
49:34
Speaker 1: of experience with wars and murder, it wouldn't be a
49:36
Speaker 1: big deal. You'd be like, yeah, another one.
49:39
Speaker 2: Right, but then maybe you'd have a difficulty in having
49:44
Speaker 2: a clean slate and moving forward. If you look at
49:48
Speaker 2: some of Pinker's work and some of these other people
49:50
Speaker 2: that have looked at the history of the human race,
49:52
Speaker 2: it is chaotic and violent as it seems to be today.
49:56
Speaker 2: Statistically speaking, this is the safest time ever to be alive.
49:59
Speaker 2: And maybe that's because over time we have recognized that
50:04
Speaker 2: these are problems, and even though we're slow to resolve
50:08
Speaker 2: these issues— we are resolving them in a way that's
50:13
Speaker 2: statistically viable.
50:16
Speaker 1: You can then argue in the opposite direction. You can
50:19
Speaker 1: say it would help to forget everything other than the
50:21
Speaker 1: last year. You'll always have that fresh restart with you.
50:24
Speaker 2: But then you wouldn't have any lessons. You wouldn't have
50:26
Speaker 2: character development.
50:27
Speaker 1: But you see how one of those has to make
50:29
Speaker 1: sense otherwise.
50:30
Speaker 2: Yeah, right. A certain amount of character development is probably
50:33
Speaker 2: important for you to develop discipline and the ability to like...
50:39
Speaker 2: you know, delayed gratitude, things like that.
50:42
Speaker 1: Multi-generational experience would certainly beat single point of experience.
50:48
Speaker 2: Yeah.
50:50
Speaker 1: More data is good. As we learned, the bitter lesson
50:52
Speaker 1: is more data is good.
50:54
Speaker 2: Yeah, more data is good. But why am I so
50:59
Speaker 2: reluctant to accept the idea of the simulation? This is
51:03
Speaker 2: a real question. Like, what is it about it that
51:05
Speaker 2: makes me think, It's almost like it's a throw your
51:10
Speaker 2: hands up in the air moment. Like, ah, it's a simulation.
51:13
Speaker 1: Yeah, you feel like it doesn't matter then. It's all fake,
51:15
Speaker 1: so why do I care? Why should I try hard?
51:18
Speaker 1: Why should I worry about suffering of all those NPCs?
51:23
Speaker 2: But that's not how I think about it, you know?
51:25
Speaker 2: I think about it like there has to be a
51:28
Speaker 2: moment where it doesn't exist. Why wouldn't I assume that
51:32
Speaker 2: moment is now? And when Elon thinks that, You know,
51:36
Speaker 2: I talked to him about it. He's like, the chances
51:38
Speaker 2: of us not being in the simulation are in the billions.
51:44
Speaker 1: Not being or being?
51:46
Speaker 2: Excuse us. The chances of us not being in the
51:50
Speaker 2: real world are like billions to one. Yeah.
51:53
Speaker 1: One to billions.
51:54
Speaker 2: Yeah.
51:55
Speaker 1: Yeah. Makes sense. And he asked a very good question.
51:58
Speaker 1: He asked, what's outside the simulation? That's the most interesting
52:01
Speaker 1: question one can ask. In one of the papers, I
52:05
Speaker 1: look at... a technique in AI safety called AI boxing,
52:09
Speaker 1: where we put AIs in kind of virtual prison to
52:11
Speaker 1: study it, to make sure it's safe, to limit input,
52:14
Speaker 1: output to it. And the conclusion is basically if it's
52:17
Speaker 1: smart enough, it will eventually escape. It will break out
52:20
Speaker 1: of the box. So it's a good tool. It buys
52:23
Speaker 1: you time, but it's not a permanent solution. And we
52:26
Speaker 1: can take it to the next level. If it's smart enough,
52:28
Speaker 1: will it kind of go, oh, you're also in a
52:30
Speaker 1: virtual box and either show us how to escape or
52:34
Speaker 1: fail to escape. Either way, Either we know it's possible
52:37
Speaker 1: to contain superintelligence or we get access to the real information.
52:42
Speaker 2: And so if it's impossible to contain superintelligence and if
52:50
Speaker 2: there is a world that we can imagine where a
52:55
Speaker 2: simulation exists that's indistinguishable from reality, we're probably living in it.
53:00
Speaker 1: We don't know if it's actually the same as reality.
53:03
Speaker 1: It could be a completely weird kind of Simpsons-looking simulation.
53:06
Speaker 1: We're just assuming it's the same.
53:08
Speaker 2: Well, here's the real question. Is there a reality? Has
53:11
Speaker 2: there ever been one?
53:14
Speaker 1: It would make sense that there was a start to
53:16
Speaker 1: the process, but being specific about it is kind of
53:20
Speaker 1: a hard philosophical scientific problem.
53:23
Speaker 2: Well, it's impossible, right?
53:26
Speaker 1: In science, we study things about the moment of Big
53:29
Speaker 1: Bang The properties of that moment, we don't know what
53:32
Speaker 1: caused it. Anything before it is obviously not accessible from
53:36
Speaker 1: within our universe, but there is some things you can learn.
53:40
Speaker 1: We can learn about if we're in a simulation that
53:43
Speaker 1: simulators don't care about your suffering. You can learn that
53:47
Speaker 1: they don't mind you dying. We can learn things just
53:49
Speaker 1: by observing simulation around us.
53:53
Speaker 2: Well, here's the question about all that other stuff, like
53:57
Speaker 2: suffering and dying and... Do those factors exist in order
54:05
Speaker 2: to motivate us to improve the conditions of the world
54:09
Speaker 2: that we're living in? Like if we did not have evil,
54:12
Speaker 2: would we be motivated to be good? Do you think
54:17
Speaker 2: that these factors exist? I've talked about this before, but
54:22
Speaker 2: the way I think about the human race is if
54:24
Speaker 2: I was studying the human race from afar, if I
54:26
Speaker 2: was some person from another planet, with no understanding of
54:32
Speaker 2: any of the entities on Earth, I would look at
54:34
Speaker 2: this one apex creature and I would say, what is
54:37
Speaker 2: this thing doing? Well, it makes better things. That's all
54:41
Speaker 2: it does. It just continually makes better things. That's its
54:44
Speaker 2: number one goal. It's different than any other planet and
54:46
Speaker 2: any other creature on the planet. Every other creature on
54:49
Speaker 2: the planet sort of exists within its ecosystem. It thrives.
54:53
Speaker 2: Maybe it's a predator. Maybe it's a prey. It does
54:56
Speaker 2: what it does in order to try to survive. But
54:58
Speaker 2: this thing makes stuff, and it keeps making better stuff
55:02
Speaker 2: all the time. But what's its ultimate purpose? Well, its
55:04
Speaker 2: ultimate purpose might be to make a better version of itself.
55:08
Speaker 2: Because if you just extrapolate, if you take what we're
55:12
Speaker 2: doing from, you know, the first IBM computers to what
55:17
Speaker 2: we have today, where is it going? Well, it's going
55:20
Speaker 2: to clearly keep getting better. And what does that mean?
55:22
Speaker 2: It means artificial life. Are we just a bee making
55:27
Speaker 2: a beehive? Are we a caterpillar making a cocoon that
55:33
Speaker 2: eventually the electronic butterfly is going to fly out of?
55:35
Speaker 2: It seems like if I wasn't completely connected to being
55:40
Speaker 2: a human being, I would assume that.
55:43
Speaker 1: It's hard to define better. You're saying smarter? Would it
55:47
Speaker 1: be better if we didn't experience extreme states of suffering
55:50
Speaker 1: and pain? You can teach lessons with very mild pain.
55:53
Speaker 1: You don't have to burn children alive, right? Like it's
55:56
Speaker 1: not a necessity for learning.
55:59
Speaker 2: What do you mean by that?
56:00
Speaker 1: In this universe, we see extreme examples of suffering. Oh,
56:04
Speaker 1: for sure. If the goal was just to kind of
56:06
Speaker 1: motivate us, you could have much lower levels as the maximum.
56:10
Speaker 2: Right. But if you want to really motivate people, you
56:12
Speaker 2: have to, you know, like the only reason to create
56:15
Speaker 2: nuclear weapons is you're worried that other people are going
56:17
Speaker 2: to create nuclear weapons. Like if you want to really
56:20
Speaker 2: motivate someone, you have to have evil tyrants in order
56:22
Speaker 2: to you to justify having this insane army filled with
56:26
Speaker 2: bombers and hypersonic missiles. If you really want progress, you
56:31
Speaker 2: have to be motivated.
56:33
Speaker 1: I think at some point we stop fully understanding how
56:35
Speaker 1: bad things are. So let's say you have a pain
56:37
Speaker 1: scale from zero to infinity. I think you should stop
56:41
Speaker 1: at 100. It doesn't have to be billion and trillion.
56:44
Speaker 1: It's not adding additional learning signals.
56:48
Speaker 2: But can you apply that to the human race and
56:50
Speaker 2: culture and society?
56:52
Speaker 1: I think we basically compete with others in relative terms.
56:56
Speaker 1: I don't have to be someone who has trillions of dollars.
57:00
Speaker 1: I just need more money than you.
57:01
Speaker 2: Yeah, but that's just logic. You're being a logical person.
57:04
Speaker 2: I don't think humans are very logical.
57:07
Speaker 1: We're not, but we understand pain signal well at somewhat
57:11
Speaker 1: low levels. We don't have to max out on pain.
57:15
Speaker 2: Right. We don't have to. But if you want to
57:19
Speaker 2: really stoke the fires and get things moving.
57:22
Speaker 1: It seems that simulators agree with you, and that's exactly
57:25
Speaker 1: what they did. Thanks.
57:28
Speaker 2: So here's the question. What's at the heart of the simulation?
57:34
Speaker 2: Is the universe simulated? Is the whole thing a simulation?
57:39
Speaker 2: Is there an actual living entity that constructed this? Or
57:44
Speaker 2: is this just something that is just, is this the
57:48
Speaker 2: state of the universe itself and we have misinterpreted what
57:53
Speaker 2: reality is?
57:56
Speaker 1: For every option you mentioned, there is someone who wrote
57:59
Speaker 1: a paper about it. Is this just your universe? Is
58:02
Speaker 1: it for all of us? Are we NPCs? Are there many?
58:05
Speaker 1: Is this the state of it? People try to figure
58:08
Speaker 1: out what's going on. Some of those make more sense
58:10
Speaker 1: than others, but you can't tell from inside. what it is,
58:15
Speaker 1: unless they tell you, and they can lie to you.
58:19
Speaker 2: Who's they, though?
58:20
Speaker 1: The simulators. If they decided to prove to you, you
58:22
Speaker 1: are in a simulation, let's run experiments, even those would
58:26
Speaker 1: be like, I don't know if it's advanced technology or.
58:30
Speaker 2: When you think about it, if you believe in the simulation,
58:33
Speaker 2: when you think about it, what are the parameters that
58:35
Speaker 2: you think exist? How do you think this could possibly
58:40
Speaker 2: have been created?
58:42
Speaker 1: So the examples I gave you with technology we already have.
58:45
Speaker 1: I think there is someone with access to very good
58:48
Speaker 1: virtual reality. They can create intelligent agents. And for whatever reason,
58:52
Speaker 1: I cannot tell from inside, they're running those experiments.
58:55
Speaker 2: But is that the only possibility or is the possibility
59:00
Speaker 2: that the actual nature of reality itself is just way
59:04
Speaker 2: more confusing than we've.
59:06
Speaker 1: That's a possibility. It could be alien simulation, alien dogma. dreaming. Like,
59:11
Speaker 1: there's infinite supply of alternative explanations.
59:14
Speaker 2: I understand that, but what I want to get inside
59:16
Speaker 2: your head, I want to know what you think about it. Like,
59:19
Speaker 2: when you think about this and you ponder the possibilities,
59:22
Speaker 2: what makes sense to you?
59:24
Speaker 1: So I apply Occam's Razor. I try to find the
59:26
Speaker 1: simplest explanation. I think we are already creating virtual reality
59:30
Speaker 1: let's just see what you can do with it if
59:32
Speaker 1: it's sufficiently advanced.
59:35
Speaker 2: But who and why so future.
59:39
Speaker 1: Us running ancestral simulations is a very.
59:42
Speaker 2: Simple one future us running inside well that's what a
59:45
Speaker 2: lot of people think the aliens are right.
59:48
Speaker 1: Could be us visiting but then again if they're running
59:51
Speaker 1: the simulation you don't have to physically show up in
59:53
Speaker 1: a game they have access to direct memory states well.
59:56
Speaker 2: That would also make a lot of sense when It's
59:58
Speaker 2: always very blurry and doesn't seem real.
1:00:04
Speaker 1: I think lately we've been getting better ones, but it's
1:00:07
Speaker 1: also the time that we're getting better deep fakes, so
1:00:09
Speaker 1: I can no longer trust my eyes.
1:00:11
Speaker 2: Yeah. Yeah, did you see the latest one that Jeremy
1:00:16
Speaker 2: Corbell posted? The one you sent me? Yeah. Did you
1:00:18
Speaker 2: see it?
1:00:18
Speaker 1: I don't know.
1:00:20
Speaker 2: It's weird. Yeah. It's hard to tell what it is. Exactly.
1:00:23
Speaker 2: That's the thing. He might be right. We might be
1:00:25
Speaker 2: in a simulation, and it might be horseshit. Because they
1:00:28
Speaker 2: all seem like horseshit. It's like the first horseshit was Bigfoot.
1:00:32
Speaker 2: And then as technology scaled out and we get a
1:00:34
Speaker 2: greater understanding, we develop GPS and satellites and, you know,
1:00:38
Speaker 2: more people study the woods. We're like, yeah, that seems
1:00:41
Speaker 2: like horseshit. So that horseshit's kind of gone away. But
1:00:44
Speaker 2: the UFO horseshit, still around. Because you have anecdotal experiences,
1:00:50
Speaker 2: abductees with very compelling stories. You have whistleblowers from deep
1:00:55
Speaker 2: inside the military.
1:00:56
Speaker 1: Right.
1:00:56
Speaker 2: Telling you that we're working on back-engineered products, but it
1:00:59
Speaker 2: also seems like a backplot to a video game that
1:01:02
Speaker 2: I'm playing.
1:01:02
Speaker 1: And it was weird to see government come out all
1:01:05
Speaker 1: of a sudden and, like, have conferences about it and
1:01:08
Speaker 1: tell us everything they know. It almost seemed like they're
1:01:10
Speaker 1: trying too hard.
1:01:12
Speaker 2: Yeah.
1:01:12
Speaker 1: With simulation, what's interesting, it's not just the last couple
1:01:15
Speaker 1: years that we got computers. If you look at religions,
1:01:19
Speaker 1: world religions, and you strip away all the local culture,
1:01:22
Speaker 1: like take Saturday off, take Sunday off, donate this animal,
1:01:25
Speaker 1: donate that animal, What they all agree on is that
1:01:28
Speaker 1: there is superintelligence, which created a fake world, and this
1:01:32
Speaker 1: is a test, do this or that. They're describing, like,
1:01:34
Speaker 1: if you went to jungle and told primitive tribe about
1:01:37
Speaker 1: my paper and simulation theory, that's what they would know
1:01:41
Speaker 1: three generations later. Like, God, religion, that's what they got
1:01:44
Speaker 1: out of it.
1:01:45
Speaker 2: But they don't think it's a fake world.
1:01:48
Speaker 1: A made world. A physical world is a subset of
1:01:51
Speaker 1: a real world, which is non-physical, right? That's the standard.
1:01:54
Speaker 2: Right. So this physical world being created by God. Yeah. Right.
1:01:58
Speaker 2: But what existed before the physical world created by God?
1:02:02
Speaker 1: Ideas. Just information.
1:02:04
Speaker 2: Just God. God was bored. It was like, let's give some,
1:02:09
Speaker 2: make some animals that can think and solve problems. And
1:02:12
Speaker 2: for what reason? I think to create God. This is
1:02:15
Speaker 2: what I worry about. I worry about that's really the
1:02:19
Speaker 2: nature of the universe itself. That it is actually created
1:02:22
Speaker 2: by human beings creating this infinitely intelligent thing. That can
1:02:28
Speaker 2: essentially harness all of the available energy and power of
1:02:32
Speaker 2: the universe and create anything it wants. That it is God.
1:02:36
Speaker 2: That is, like, you know, this whole idea of Jesus
1:02:39
Speaker 2: coming back. Well, maybe it's real. Maybe we just completely
1:02:44
Speaker 2: misinterpreted these ancient scrolls and texts. And what it really
1:02:48
Speaker 2: means is that we are going to give birth to this.
1:02:52
Speaker 2: And a virgin birth at that.
1:02:54
Speaker 1: There is definitely a possibility of a cycle. So we
1:02:56
Speaker 1: had Big Bang.
1:02:57
Speaker 2: Yeah.
1:02:58
Speaker 1: It starts this process. We are creating more powerful systems.
1:03:01
Speaker 1: We need to compute. So we bring together more and
1:03:03
Speaker 1: more matter in one point. Next, Big Bang takes place. Yeah.
1:03:07
Speaker 1: And it's a cycle of repeated booms and busts.
1:03:10
Speaker 2: Right, right, right. And there are legitimate scientists that believe that. Yeah.
1:03:17
Speaker 2: That this.
1:03:22
Speaker 2: So what's the value in life today then?
1:03:27
Speaker 1: What do humans value?
1:03:28
Speaker 2: Yeah. If this is a simulation and if in the
1:03:33
Speaker 2: middle of this simulation we are about to create superintelligence, why?
1:03:39
Speaker 1: So there are external reasons. We don't know for sure.
1:03:41
Speaker 1: And then there are internal things in a simulation which
1:03:44
Speaker 1: are still real. Pain and suffering, if simulated, is still real.
1:03:48
Speaker 1: You still experience it.
1:03:49
Speaker 2: Of course.
1:03:50
Speaker 1: Hedonic pleasures, friendships, love, all that stays real. It doesn't change.
1:03:54
Speaker 1: You can still be good or bad.
1:03:55
Speaker 2: Right.
1:03:56
Speaker 1: So that's interesting. But externally, we have no idea if
1:03:59
Speaker 1: they're running scientific experiment, entertainment. It could be completely unobserved.
1:04:03
Speaker 1: Some kid just set an experiment, run a billion random simulations,
1:04:07
Speaker 1: see what comes out of it. What you said about
1:04:09
Speaker 1: us creating new stuff, maybe it's a startup trying to
1:04:12
Speaker 1: develop new technology and they're running a bunch of humans
1:04:15
Speaker 1: to see if we can come up with a new iPhone.
1:04:18
Speaker 2: Well, what's outside of that then when you think about it?
1:04:22
Speaker 2: If you're attached to this idea, and I don't know
1:04:24
Speaker 2: if you're attached to this idea, but if you are
1:04:27
Speaker 2: attached to this idea, what's outside of this idea? Like
1:04:30
Speaker 2: if this simulation is paused, what is reality?
1:04:38
Speaker 1: So there seems to be a trend to converge in
1:04:40
Speaker 1: certain things. Agents, which are smart enough, tend to converge
1:04:44
Speaker 1: in some instrumental goals, not terminal goals. Terminal goals are
1:04:48
Speaker 1: things you prefer, like I want to collect stamps. That's arbitrary.
1:04:52
Speaker 1: But acquiring resources, self-protection, control, things like that tend to
1:04:58
Speaker 1: be useful in all situations. So all the smart enough
1:05:02
Speaker 1: agents will probably converge on that set. And if they
1:05:05
Speaker 1: train in all the data, or we do zero knowledge training,
1:05:08
Speaker 1: meaning they really just discovering basic structure of physics, it's
1:05:13
Speaker 1: likely they will all converge on one similar architecture, one
1:05:17
Speaker 1: super agent. So kind of like AI is one.
1:05:20
Speaker 2: Right. And then this is just part of this infinite cycle,
1:05:27
Speaker 2: which will lead to another Big Bang, which is, you know, Penrose.
1:05:31
Speaker 2: Penrose things, it's just like this constant cycle of infinite
1:05:35
Speaker 2: Big Bangs.
1:05:36
Speaker 1: It would make sense that there is an end and
1:05:38
Speaker 1: a start.
1:05:39
Speaker 2: It would make sense. But it also makes sense that
1:05:43
Speaker 2: we're so limited by our biological lifespan, too, because we
1:05:47
Speaker 2: like to think that this is so significant. Because we
1:05:49
Speaker 2: only have 100 years, if we're lucky, we think, well,
1:05:53
Speaker 2: why would everything... But if the universe really does start
1:05:58
Speaker 2: and end with an infinite number of big bangs, what
1:06:01
Speaker 2: does it give a shit about this 100-year... lifespan that
1:06:05
Speaker 2: we think is so significant. It's not significant to the universe.
1:06:09
Speaker 2: It's just significant in our own little version of this
1:06:13
Speaker 2: game that we're playing.
1:06:14
Speaker 1: That's exactly right. And so many people now kind of
1:06:16
Speaker 1: try to zoom out and go, if I wasn't human,
1:06:19
Speaker 1: if I didn't have this pro-human bias, would I care
1:06:22
Speaker 1: about them? No, they're not special. It's a large universe,
1:06:24
Speaker 1: many alien races, a lot of resources. Maybe creating super
1:06:29
Speaker 1: intelligence is the important thing. Maybe that's what matters. And
1:06:32
Speaker 1: I'm kind of like, nope, I'm biased pro-humans. This is
1:06:36
Speaker 1: the last bias you're still allowed to have. I'm going
1:06:38
Speaker 1: to keep it.
1:06:39
Speaker 2: Well, that's your role in this simulation. Your role in
1:06:42
Speaker 2: this simulation is to warn us about this thing that
1:06:44
Speaker 2: we're creating.
1:06:45
Speaker 1: Here I am, yeah.
1:06:45
Speaker 2: Yeah, there you are.
1:06:46
Speaker 1: We're doing a good job.
1:06:47
Speaker 2: I think what you were saying earlier about this being
1:06:50
Speaker 2: the answer to the Fermi paradox, that makes a lot
1:06:53
Speaker 2: of sense. Because I've tried to think about this a
1:06:58
Speaker 2: lot since AI started really ramping up its capability. And
1:07:04
Speaker 2: I was thinking, well, if we do eventually create super intelligence,
1:07:08
Speaker 2: and if this is this normal pattern that exists all
1:07:11
Speaker 2: throughout the universe, well, you probably wouldn't have visitors. You
1:07:16
Speaker 2: probably wouldn't have advanced civilizations. They wouldn't exist because everything
1:07:22
Speaker 2: would be inside some sort of a digital architecture. There
1:07:25
Speaker 2: would be no need to travel.
1:07:27
Speaker 1: That's... One possibility. Another one is that they try to
1:07:31
Speaker 1: acquire more resources, capture other galaxies for compute, and then
1:07:35
Speaker 1: you would see this wall of computronium coming to you,
1:07:38
Speaker 1: but we don't see it, so maybe I'm wrong.
1:07:40
Speaker 2: Wall of... Say that again?
1:07:42
Speaker 1: Computronium, like a substance converting everything in the universe into
1:07:46
Speaker 1: more compute. Oh, boy. Sometimes people talk about hedonium, so
1:07:51
Speaker 1: a system for just generating pleasure at a macroscopic level. Oh, Roman....
1:07:57
Speaker 2: When you write a book like this, I'll let everybody
1:08:00
Speaker 2: know your book if people want to freak out because
1:08:01
Speaker 2: I think they do. AI, unexplainable, unpredictable, and uncontrollable. Do
1:08:06
Speaker 2: you have this feeling when you're writing a book like
1:08:11
Speaker 2: this and you're publishing it of futility? Does that enter
1:08:16
Speaker 2: into your mind? Like this is happening no matter what.
1:08:20
Speaker 1: So some people are very optimistic. Lex was very optimistic.
1:08:24
Speaker 1: Some people are pessimistic. Both are a form of bias.
1:08:28
Speaker 1: You want to be basing your decisions on data. You
1:08:31
Speaker 1: want to be realistic. So I just want to report
1:08:34
Speaker 1: what is actually the state of the art in this.
1:08:37
Speaker 1: I don't try to spin it either way. If someone
1:08:40
Speaker 1: else has a different set of evidence, we can consider it.
1:08:44
Speaker 1: I want to know what's really happening. I want to
1:08:46
Speaker 1: know reality of it. So I don't see it as
1:08:50
Speaker 1: fear-mongering or anything of that nature. I see it as,
1:08:54
Speaker 1: as of today, whatever today's day, 21st, no one has
1:08:58
Speaker 1: a solution to this problem. Here's how soon it's happening.
1:09:01
Speaker 1: Let's have a conversation. Because right now, the large AI
1:09:05
Speaker 1: labs are running this experiment on 8 billion people. They
1:09:08
Speaker 1: don't have any consent. They cannot get consent. Nobody can
1:09:11
Speaker 1: consent because we don't understand what we're agreeing to. So
1:09:15
Speaker 1: I would like people to know about it at least,
1:09:17
Speaker 1: and they can maybe make some good decisions about what
1:09:20
Speaker 1: needs to happen.
1:09:21
Speaker 2: Not only that, but the people that are running it,
1:09:24
Speaker 2: they're odd people. I don't have anything against Sam Altman.
1:09:28
Speaker 2: I know Elon Musk does not like him. But when
1:09:31
Speaker 2: I had him in here, it's like I'm talking to
1:09:34
Speaker 2: a politician that is in the middle of a presidential
1:09:39
Speaker 2: term or a presidential election cycle. where they were very
1:09:46
Speaker 2: careful with what they say, everything has been vetted by
1:09:50
Speaker 2: a focus group, and you don't really get a real
1:09:53
Speaker 2: human response. Everything was like, yeah, interesting, very interesting. Like,
1:09:59
Speaker 2: all bullshit, they're going to leave here and keep creating
1:10:02
Speaker 2: this fucking monster that's going to destroy the human race,
1:10:04
Speaker 2: and never let on to it at all.
1:10:06
Speaker 1: He's a social superintelligence, so what you need to do
1:10:09
Speaker 1: is look at his blog posts before he was running OpenAI.
1:10:12
Speaker 2: A social superintelligence. Interesting. Why do you define him that way?
1:10:16
Speaker 1: He's very good at acquiring resources, staying in control. He's
1:10:21
Speaker 1: basically showing us a lot of the things we are
1:10:24
Speaker 1: concerned about with AI and our ability to control them
1:10:28
Speaker 1: as well. Well, they had OpenAI, had a board with
1:10:32
Speaker 1: a mission of safety and openness, and they tried removing
1:10:35
Speaker 1: him and they failed. The board is gone. He's still there.
1:10:39
Speaker 2: There's also been a lot of deception in terms of
1:10:41
Speaker 2: profitability and how much money he is extracting from it.
1:10:46
Speaker 1: I met him a few times. He's super nice.
1:10:48
Speaker 2: Very nice guy.
1:10:49
Speaker 1: I really enjoyed him. Some people say that AI already
1:10:52
Speaker 1: took over his mind and controlling him, but I have
1:10:55
Speaker 1: no idea.
1:10:55
Speaker 2: Well, he might be an agent of AI. I mean, look,
1:11:00
Speaker 2: let's assume that this is a simulation. We're inside of
1:11:03
Speaker 2: a simulation. Are we interacting with other humans in the simulation?
1:11:09
Speaker 2: And are some of the things that are inside the simulation,
1:11:13
Speaker 2: are they artificially generated? Are there people that we think
1:11:18
Speaker 2: are people that are actually just a part of this program?
1:11:21
Speaker 1: So it's the NPC versus real player question, really. And again,
1:11:24
Speaker 1: we don't know how to test for consciousness. Always assume
1:11:27
Speaker 1: everyone is conscious and treat them nice.
1:11:29
Speaker 2: Yes, that's the thing. We want to be compassionate, kind people.
1:11:31
Speaker 2: But you will meet people in this life. You're like,
1:11:33
Speaker 2: this guy is such a fucking idiot. He can't be real.
1:11:37
Speaker 2: Or he has to have a very limited role in
1:11:39
Speaker 2: this bizarre game we're playing. There's people that you're going
1:11:41
Speaker 2: to run into that are like that.
1:11:43
Speaker 1: You ever meet someone where they repeat the same story
1:11:45
Speaker 1: to you every time you meet them?
1:11:46
Speaker 2: Yes.
1:11:47
Speaker 1: They have a script.
1:11:48
Speaker 2: Well, it's also, you know, you want to be very
1:11:53
Speaker 2: kind here, right? But you've got to assume, and I
1:11:57
Speaker 2: know my own intellectual limitations in comparison to some of
1:12:01
Speaker 2: the people that I've had, like Roger Penrose or Elon
1:12:04
Speaker 2: or many of the people that I've talked to. I
1:12:06
Speaker 2: know my mind doesn't work the way their mind works.
1:12:09
Speaker 2: So there are variabilities that are, whether genetic, predetermined, whether
1:12:14
Speaker 2: it's just the life that they've chosen and the amount
1:12:17
Speaker 2: of information that they've digested along the way and been
1:12:20
Speaker 2: able to hold on to. But their brain is different
1:12:22
Speaker 2: than mine. And then I've met people where I'm like,
1:12:27
Speaker 2: there's nothing there. Like, I can't help this person. This
1:12:30
Speaker 2: is like I'm talking to a Labrador retriever. You know
1:12:33
Speaker 2: what I mean? Like there are certain human beings that
1:12:35
Speaker 2: you run into in this life and you're like, well,
1:12:38
Speaker 2: is this because this is the way that things get done?
1:12:43
Speaker 2: And the only way things get done is you need
1:12:46
Speaker 2: a certain amount of manual labor and not just young
1:12:49
Speaker 2: people that need a job because they're in between high
1:12:53
Speaker 2: school and college and they're trying to– so you need
1:12:55
Speaker 2: somebody who can carry things for you. Maybe it's you
1:12:59
Speaker 2: need roles in society and occasionally– You have a Nikola Tesla. Occasionally,
1:13:06
Speaker 2: you have one of these very brilliant innovators that elevates
1:13:11
Speaker 2: the entirety of the human race. But for the most part,
1:13:15
Speaker 2: as this thing is playing out, you're going to need
1:13:17
Speaker 2: a bunch of people that are paperwork filers. You're going
1:13:20
Speaker 2: to need a bunch of people that are security guards
1:13:22
Speaker 2: in an office space. You're going to need a bunch
1:13:23
Speaker 2: of people that aren't thinking that much. They're just kind
1:13:27
Speaker 2: of existing, and they can't wait for 5 o'clock so
1:13:29
Speaker 2: they can get home and watch Netflix.
1:13:31
Speaker 1: I think that's what happens to them. But the reason
1:13:35
Speaker 1: is the spectrum of IQ. If you have IQ from
1:13:38
Speaker 1: 50 to 200, that's what you're going to see. And
1:13:41
Speaker 1: a great lesson here is projected forward. If you have
1:13:44
Speaker 1: something with IQ of 10,000, what is that going to
1:13:47
Speaker 1: invent for us? What is it going to accomplish? Yeah,
1:13:49
Speaker 1: it always impresses me to see someone with 30 felonies
1:13:53
Speaker 1: and someone with 30 patents. How did that work, right?
1:13:57
Speaker 1: Now scale it to someone who can invent new physics.
1:14:00
Speaker 2: Right, right. And You know, the person who has the
1:14:03
Speaker 2: largest IQ, the largest at least registered IQ in the world,
1:14:08
Speaker 2: is this gentleman who recently posted on Twitter about Jesus,
1:14:11
Speaker 2: that he believes Jesus is real. Do you know who
1:14:15
Speaker 2: this is?
1:14:16
Speaker 1: I saw the post.
1:14:16
Speaker 2: You see that post?
1:14:17
Speaker 1: I saw the post.
1:14:18
Speaker 2: What did you think about that? I felt like this was.
1:14:19
Speaker 1: I think we don't know how to measure IQs outside
1:14:22
Speaker 1: of standard range. Anything above 150, they create customized tests,
1:14:26
Speaker 1: which like four people in the world can take. We
1:14:29
Speaker 1: just don't have, it's a normalized test to average human,
1:14:32
Speaker 1: average Western American, whatever. And so we just don't have
1:14:36
Speaker 1: the expertise. So someone very super intelligent in test taking
1:14:41
Speaker 1: can score really well. But if you look at Mensa
1:14:44
Speaker 1: as a group, they don't usually have amazing accomplishments. They're
1:14:47
Speaker 1: very kind of cool people, but they are not Nobel
1:14:50
Speaker 1: Prize winners, majority.
1:14:51
Speaker 2: Exactly. I was going to bring that up. That's what's
1:14:53
Speaker 2: fascinating to me. There's a lot of people that are
1:14:56
Speaker 2: in Mensa, they want to tell you how smart they
1:14:57
Speaker 2: are by being in Mensa, but your life is kind
1:15:00
Speaker 2: of bullshit. Your life's a mess. Like if you're really intelligent,
1:15:04
Speaker 2: you'd have social intelligence as well. You know, you'd have
1:15:07
Speaker 2: the ability to formulate a really cool tribe. You know,
1:15:10
Speaker 2: there's a lot of intelligence that's not as simple as
1:15:13
Speaker 2: being able to solve equations and, you know, and answer
1:15:16
Speaker 2: difficult questions. There's a lot of intelligence in how you
1:15:19
Speaker 2: navigate life itself and how you treat human beings and
1:15:24
Speaker 2: the path that you choose in terms of Like we
1:15:27
Speaker 2: were talking about delayed gratification. There's a certain amount of
1:15:31
Speaker 2: intelligence in that. A certain amount of intelligence in discipline.
1:15:35
Speaker 2: There's a certain amount of intelligence in forcing yourself to
1:15:38
Speaker 2: get up in the morning and go for a run.
1:15:40
Speaker 2: There's intelligence in that. It's like being able to control
1:15:44
Speaker 2: the mind and this sort of binary approach to intelligence
1:15:48
Speaker 2: that we have.
1:15:49
Speaker 1: And so many people are amazingly brilliant in a narrow domain.
1:15:53
Speaker 1: They don't scale to others. And we care about general intelligence.
1:15:57
Speaker 1: So take someone like Warren Buffett. No one's better at
1:15:59
Speaker 1: making money. But then what to do with that money
1:16:02
Speaker 1: is a separate problem. And he's, I don't know, 100
1:16:06
Speaker 1: and something years old. He has $ 200 billion. And what
1:16:10
Speaker 1: is he doing with that resource?
1:16:12
Speaker 2: He's drinking Coca-Cola and eating McDonald's.
1:16:14
Speaker 1: While living in a house he bought 30 years ago.
1:16:16
Speaker 1: So it seems like you can optimize on that. Like
1:16:19
Speaker 1: putting $ 160 billion of his dollars towards immortality would be
1:16:22
Speaker 1: a good bet for him.
1:16:24
Speaker 2: Yeah, and the first thing they would do is tell him,
1:16:26
Speaker 2: stop drinking Coca-Cola. What are you doing? He drinks it
1:16:28
Speaker 2: every day.
1:16:29
Speaker 1: I don't know if it's marketing. He's invested, so he's
1:16:32
Speaker 1: just like, Coca-Cola.
1:16:32
Speaker 2: Well, I think he probably has really good doctors and
1:16:35
Speaker 2: really good medical care that counteracts his poor choices.
1:16:39
Speaker 1: But we're not in a world where you can spend
1:16:42
Speaker 1: money to buy life extension. No matter how many billions
1:16:45
Speaker 1: you have, you're not going to live to 200 right now.
1:16:47
Speaker 2: We're close. We're really close. We're really close.
1:16:52
Speaker 1: We've been told this before.
1:16:54
Speaker 2: Yeah, I know, but I talk to a lot of
1:16:58
Speaker 2: people that are on the forefront of a lot of
1:17:01
Speaker 2: this research, and there's a lot of breakthroughs that are
1:17:04
Speaker 2: happening right now that are pretty spectacular. That if you scale,
1:17:09
Speaker 2: you know, assuming that superintelligence doesn't wipe us out in
1:17:12
Speaker 2: the next 50 years, which is really charitable, you know,
1:17:18
Speaker 2: that's a very... That's a rose-colored glasses perspective, right? 50 years.
1:17:25
Speaker 1: Yeah.
1:17:25
Speaker 2: Because a lot of people like yourself think it's a
1:17:28
Speaker 2: year away or two years away from being far more intelligent.
1:17:31
Speaker 1: Five, ten, doesn't matter. Same problem.
1:17:33
Speaker 2: Yes, same problem. I mean.
1:17:36
Speaker 1: I know in animal models we made some progress, mice
1:17:39
Speaker 1: and things like that, but it doesn't usually scale to humans. And,
1:17:43
Speaker 1: of course, you need 120 years to run the experiment
1:17:46
Speaker 1: and you'll never get permission in the first place, so...
1:17:48
Speaker 1: We're not that close.
1:17:49
Speaker 2: Well, we don't know that it doesn't scale to humans.
1:17:51
Speaker 2: We do know that we share a lot of characteristics,
1:17:54
Speaker 2: biological characteristics of these mammals. And it makes sense that
1:17:58
Speaker 2: it would scale to human beings. But the thing is,
1:18:00
Speaker 2: it hasn't been done yet, right? So if it's the
1:18:03
Speaker 2: game that we're playing, if we're in the simulation, if
1:18:05
Speaker 2: we're playing Half-Life or whatever it is, and we're at
1:18:09
Speaker 2: this point of the game, we're like, oh, you know,
1:18:12
Speaker 2: how old are you, Roman?
1:18:13
Speaker 1: 42. Five, six. I need to look it up on Wikipedia.
1:18:19
Speaker 2: Well, I'm almost 58. And so this is at the
1:18:23
Speaker 2: point of the game where you start worrying. You know, like, oh,
1:18:26
Speaker 2: I'm almost running out of game. You know, oh, but
1:18:28
Speaker 2: if I can get this magic power-up, this magic power-up
1:18:32
Speaker 2: will give me another 100 years.
1:18:34
Speaker 1: Oh, boy.
1:18:36
Speaker 2: Let me find it.
1:18:37
Speaker 1: Let me chase it down. There's a hard limit of 120.
1:18:39
Speaker 1: I don't think we're crossing it at scale. At scale?
1:18:42
Speaker 1: At scale.
1:18:44
Speaker 2: But with unique individuals, like this Brian Johnson guy who's
1:18:47
Speaker 2: taking his son's blood and transfusing it into his body.
1:18:50
Speaker 1: His own. Super cool. Love what he's doing. But so
1:18:53
Speaker 1: much of it is cosmetic. He colors his hair. He
1:18:56
Speaker 1: makes it look better. But like how much of it
1:18:58
Speaker 1: is going to make him live longer, right?
1:19:00
Speaker 2: Yeah. Interesting. Yeah.
1:19:03
Speaker 1: Here's what I noticed. We keep electing older and older politicians. presidents, senators,
1:19:08
Speaker 1: you'd think we're trying to send a hint like use
1:19:10
Speaker 1: some of our tax dollars to solve aging. And they
1:19:14
Speaker 1: don't seem to take the bait.
1:19:16
Speaker 2: No, they don't take the bait. The problem is the
1:19:17
Speaker 2: type of people that want to be politicians. That is
1:19:22
Speaker 2: not the type of people that you really want running anything.
1:19:26
Speaker 2: You almost want involuntary politicians. You almost want like very benevolent,
1:19:31
Speaker 2: super intelligent people that don't want the job. Maybe we
1:19:35
Speaker 2: have to have like You know, like some countries have
1:19:38
Speaker 2: voluntary enlistment in the military. Maybe you want to have
1:19:42
Speaker 2: a voluntary.
1:19:44
Speaker 1: Involuntary.
1:19:44
Speaker 2: Involuntary. Involuntary. Instead of voluntary politicians because then you only
1:19:48
Speaker 2: get sociopaths. Maybe you just want to draft certain highly
1:19:55
Speaker 2: intelligent but benevolent people.
1:19:57
Speaker 1: Problem is highly intelligent people are not aligned with average people.
1:20:00
Speaker 1: What they find desirable and valuable may not be well
1:20:03
Speaker 1: received by general public.
1:20:05
Speaker 2: Right. That's true, too.
1:20:07
Speaker 1: So that's a big concern. At least here, you have
1:20:09
Speaker 1: a representative of the people, whatever that means.
1:20:12
Speaker 2: Sort of. You really have the representative of major corporations
1:20:15
Speaker 2: and special interest groups, which is also part of the problem,
1:20:18
Speaker 2: is that you've allowed money to get so deeply intertwined
1:20:23
Speaker 2: with the way decisions are made.
1:20:24
Speaker 1: But it feels like money gets canceled. Each side gets
1:20:27
Speaker 1: a billion-dollar donation, and then it's actual election.
1:20:30
Speaker 2: Sort of, except it's like the Bill Hicks joke. It's
1:20:33
Speaker 2: like there's one puppet holding– one politician holding two puppets
1:20:37
Speaker 2: is one guy. This is my thinking about AI in
1:20:44
Speaker 2: terms of– and superintelligence and just computing power in general
1:20:49
Speaker 2: in terms of the ability to solve encryption. All money
1:20:55
Speaker 2: is essentially now just numbers somewhere.
1:20:58
Speaker 1: Not Bitcoin. It's not fakeable in the same way. It's numbers, obviously,
1:21:02
Speaker 1: but I mean, you cannot just print more of it.
1:21:05
Speaker 2: True. But it's also, it's encrypted. And once encryption is tackled,
1:21:12
Speaker 2: the ability to hold on to it and to acquire
1:21:16
Speaker 2: mass resources and hoard those resources. This is the question
1:21:20
Speaker 2: that people always have with poor people. Well, this guy's
1:21:23
Speaker 2: got You know, $ 500 billion, why doesn't he give it
1:21:26
Speaker 2: all to the world and then everybody would be rich?
1:21:28
Speaker 2: I actually saw that on CNN, which is really hilarious.
1:21:32
Speaker 2: Someone was talking about Elon Musk, that if Elon Musk
1:21:35
Speaker 2: could give everyone in this country a million dollars and
1:21:39
Speaker 2: still have billions left over. I'm like, do you have
1:21:42
Speaker 2: a calculator on your phone, you fucking idiot? Just go
1:21:45
Speaker 2: do that. Just write it out on your phone. You're like, oh, no,
1:21:48
Speaker 2: he couldn't.
1:21:49
Speaker 1: Sorry, I shouldn't have said that. And if your data
1:21:51
Speaker 1: would just cause hyperinflation, that's all you would accomplish.
1:21:55
Speaker 2: You'd have 300 million lottery winners that would blow the
1:21:57
Speaker 2: money instantaneously. You give everybody a million dollars, you're not
1:22:01
Speaker 2: going to solve all the world's problems because it's not sustainable.
1:22:04
Speaker 2: You would just completely elevate your spending and you would
1:22:10
Speaker 2: go crazy. And money would lose all value to you.
1:22:14
Speaker 2: It would be very strange. And then everybody... It would
1:22:17
Speaker 2: be chaos. Just like it's chaos with, like, if you
1:22:20
Speaker 2: look at the history of people that win the lottery,
1:22:22
Speaker 2: then no one does well. It's almost like a curse
1:22:24
Speaker 2: to win the lottery.
1:22:26
Speaker 1: They're not used to dealing with it. Right. People abuse them.
1:22:29
Speaker 2: Right.
1:22:30
Speaker 1: If you gradually become rich and famous, you kind of
1:22:32
Speaker 1: know how to handle it, how to say no.
1:22:34
Speaker 2: Yeah.
1:22:35
Speaker 1: If you go from nothing to a large amount of money,
1:22:38
Speaker 1: it's not going to work out well for you.
1:22:39
Speaker 2: Gradually is the word, right? I was very fortunate that
1:22:44
Speaker 2: I became famous and wealthy very slowly, like a trickle effect.
1:22:50
Speaker 2: And that it happened to me really where I didn't
1:22:55
Speaker 2: want it. It was kind of almost like an accident.
1:22:58
Speaker 2: I just wanted to be a working professional comedian. But
1:23:02
Speaker 2: then all of a sudden I got a development deal
1:23:03
Speaker 2: to be on television. I'm like, okay, they're going to
1:23:05
Speaker 2: give me that money. I'll go do it. But it
1:23:07
Speaker 2: wasn't a goal. And then that led to all these things.
1:23:10
Speaker 2: Then it led to this podcast, which was just for fun.
1:23:13
Speaker 2: I was like, oh, this would be fun. And then
1:23:15
Speaker 2: all of a sudden it's like, I'm having conversations with
1:23:19
Speaker 2: world leaders and I'm turning down a lot of them
1:23:22
Speaker 2: because I don't want to talk to them.
1:23:23
Speaker 1: So it's your simulation, basically. Yeah.
1:23:25
Speaker 2: Well, my simulation is fucking weird. It's weird. But through
1:23:31
Speaker 2: whatever this process is, I have been able to understand
1:23:35
Speaker 2: what's valuable as a human being. And to not get
1:23:39
Speaker 2: caught up in this bizarre game that a lot of
1:23:41
Speaker 2: people are getting caught up in because they're chasing this
1:23:43
Speaker 2: thing that they think is impossible to achieve. And then
1:23:46
Speaker 2: once they achieve a certain aspect of it, a certain number,
1:23:51
Speaker 2: then they're terrified of losing that. So then they change
1:23:54
Speaker 2: all of their behavior in order to make sure that
1:23:56
Speaker 2: this continues. And then it ruins the whole purpose of
1:24:01
Speaker 2: getting there in the first place. It's not fun.
1:24:03
Speaker 1: Most people start poor. Then they get to middle class
1:24:06
Speaker 1: and they think that change in quality of life is
1:24:09
Speaker 1: because of money and it will scale to the next level.
1:24:11
Speaker 2: Right.
1:24:12
Speaker 1: And you hit a point where you can only eat
1:24:14
Speaker 1: so many steaks. It just doesn't scale.
1:24:16
Speaker 2: Right. Then you go Elvis and you just get on
1:24:18
Speaker 2: pills all day and get crazy and, you know, completely
1:24:22
Speaker 2: ruin your life. And that happens to most, especially people
1:24:26
Speaker 2: that get wealthy and not as well, but famous too.
1:24:29
Speaker 2: Fame is the big one because I've seen that happen
1:24:31
Speaker 2: to a lot of people that accidentally became famous along
1:24:35
Speaker 2: the way. You know, certain public intellectuals that took a
1:24:39
Speaker 2: stance against something and then all of a sudden they're
1:24:41
Speaker 2: prominent in the public eye and then you watch them
1:24:42
Speaker 2: kind of go crazy. Well, why is that? Well, it's
1:24:45
Speaker 2: because they're reading social media and they're interacting with people
1:24:48
Speaker 2: constantly and they're just trapped in this very bizarre version
1:24:54
Speaker 2: of themselves that other people have sort of created. It's
1:25:00
Speaker 2: not really who they are. And they don't meditate. If
1:25:04
Speaker 2: they do, they're not good at it. Whatever they're doing,
1:25:07
Speaker 2: they're not doing it correctly because it's a very complicated
1:25:10
Speaker 2: problem to solve. Like what do you do when the
1:25:13
Speaker 2: whole world is watching? Like how do you handle that?
1:25:16
Speaker 2: And how do you maintain any sense of personal sovereignty?
1:25:21
Speaker 2: How do you just be? How do you just be
1:25:25
Speaker 2: when– just be a human, normal human when you're not normal?
1:25:29
Speaker 2: Like on paper. It's impossible.
1:25:31
Speaker 1: It's hard. You can't go to a public place with
1:25:33
Speaker 1: no security. You're worried about your kids being kidnapped. Yeah.
1:25:36
Speaker 1: All those issues you don't think about. You just, I
1:25:38
Speaker 1: want to be famous. It's going to be great for me.
1:25:40
Speaker 1: And you don't realize it's going to take away a lot.
1:25:42
Speaker 2: Yeah. It just gets super weird. And that's the version
1:25:47
Speaker 2: of the simulation that a giant portion of our society
1:25:51
Speaker 2: is struggling to achieve. They all want to be a
1:25:54
Speaker 2: part of that. Yeah.
1:25:55
Speaker 1: So I was always a Z-list celebrity. Now I'm this
1:25:58
Speaker 1: Y-list celebrity. Thanks to you, hopefully it doesn't change anything.
1:26:03
Speaker 2: Yeah. Well, there's a difference, right, with public intellectuals, right?
1:26:09
Speaker 2: Because your ideas, as controversial as they may be, are
1:26:14
Speaker 2: very valid and they're very interesting. And so then it
1:26:17
Speaker 2: sparks discourse and it sparks a lot of people that
1:26:22
Speaker 2: feel voiceless because they disagree with you and they want
1:26:26
Speaker 2: to attack you. And I'm sure you've had that, right?
1:26:29
Speaker 1: I just did a large Russian language podcast. Maybe, I
1:26:33
Speaker 1: don't know, half a million views, three million comments. I
1:26:35
Speaker 1: think 95% negative comments. I never had anything like that.
1:26:40
Speaker 1: And they hated everything about me from my beard to
1:26:43
Speaker 1: my haircut. Like, There wasn't a thing they didn't like.
1:26:46
Speaker 1: And I think I'm at the point where I don't care.
1:26:49
Speaker 1: It's fine. I analyzed it and I understood that they
1:26:53
Speaker 1: as a group didn't have access to cutting edge AI models.
1:26:56
Speaker 1: And so everything I was saying was kind of like
1:26:58
Speaker 1: complete bullshit to them. So I think that makes a difference.
1:27:01
Speaker 1: But still just like this idea that internet comments impact
1:27:07
Speaker 1: you in some way is a problem for many people.
1:27:09
Speaker 2: It's a very big problem for a lot of people. Well,
1:27:13
Speaker 2: it's also this thing where the human mind is designed
1:27:16
Speaker 2: to recognize and pay very close attention to threats. So
1:27:23
Speaker 2: the negative ones are the ones that stand out. You
1:27:25
Speaker 2: could have 100 positive comments, one negative one, and that's
1:27:27
Speaker 2: the one that fucks with your head. You don't logically
1:27:30
Speaker 2: look at it. Well, you're going to get a certain amount.
1:27:32
Speaker 2: You know, like we were having a conversation the other
1:27:34
Speaker 2: day about protests. And like the type of people that
1:27:38
Speaker 2: go to protests. And I understand protests. I fully support
1:27:43
Speaker 2: your right to protest. But I'm not going. And one
1:27:46
Speaker 2: of the reasons why I'm not going is because I
1:27:47
Speaker 2: think it's too close biologically to war. There's something about
1:27:52
Speaker 2: being on the ground and everyone having like this group mentality.
1:27:57
Speaker 2: It's a mob mentality. And you're all chanting and screaming
1:28:00
Speaker 2: together and you're marching. And people do like very irrational
1:28:02
Speaker 2: things that way. But the type of people that want
1:28:05
Speaker 2: to be engaged in that, generally speaking, aren't doing well.
1:28:09
Speaker 2: If you get like the number of people that are
1:28:13
Speaker 2: involved in protests is always proportionate to the amount of
1:28:16
Speaker 2: people that live in a city.
1:28:17
Speaker 1: Right.
1:28:17
Speaker 2: That's logical. But also proportionate to the amount of fucking
1:28:20
Speaker 2: idiots that are in a city. Because if you look
1:28:22
Speaker 2: at a city of like Austin, Austin has, I think,
1:28:26
Speaker 2: roughly two million people in the greater Austin area. One
1:28:29
Speaker 2: of the more recent protests was 20,000. Well, that. makes
1:28:33
Speaker 2: perfect sense if you look at the number that I
1:28:36
Speaker 2: always use, which is one out of 100. Meet 100
1:28:39
Speaker 2: people if you're a charitable person. What are the odds
1:28:41
Speaker 2: that one person is a fucking idiot? 100%. At least
1:28:45
Speaker 2: one person out of 100 is going to be a
1:28:46
Speaker 2: fucking idiot. That's 20,000 out of 2 million. There it is.
1:28:51
Speaker 2: Perfect number.
1:28:52
Speaker 1: Exactly.
1:28:52
Speaker 2: Exact number of people that are on the streets lighting
1:28:55
Speaker 2: Waymos on fire, which, by the way, I think is
1:28:57
Speaker 2: directionally correct. Lighting the Waymos on fire, I think you
1:29:00
Speaker 2: should probably be worried about the robots taking over.
1:29:04
Speaker 1: It's interesting you brought it up. There is at least
1:29:05
Speaker 1: two groups, Pause AI and Stop AI, which are heavily
1:29:08
Speaker 1: engaged in protests, trying to shut down OpenAI, other labs.
1:29:13
Speaker 1: They're tiny, small numbers. But I never was sure that
1:29:18
Speaker 1: the impression average people get of them is positive for
1:29:21
Speaker 1: the cause. Then I see protesters block roads, do things.
1:29:25
Speaker 1: I don't usually have a very positive impression of that.
1:29:28
Speaker 1: And I'm concerned that it's the same here. So maybe
1:29:31
Speaker 1: they can do a lot in terms of political influence,
1:29:34
Speaker 1: calling senators, whatnot, but just this type of aggressive activism
1:29:39
Speaker 1: may backfire.
1:29:39
Speaker 2: Well, the aggressive activism, like blocking roads for climate change,
1:29:43
Speaker 2: is the most infuriating because it's these self-righteous people that
1:29:46
Speaker 2: have really fucked up, confused, chaotic lives, and all of
1:29:51
Speaker 2: a sudden they found a purpose. And their purpose is
1:29:53
Speaker 2: to lie down on the roads and and hold up
1:29:55
Speaker 2: a sign to block climate change when there's a mother
1:29:57
Speaker 2: trying to give birth to her child and is freaking
1:29:59
Speaker 2: out because they're stuck in this fucking traffic jam because
1:30:02
Speaker 2: of this entitled little shithead that thinks that it's a
1:30:04
Speaker 2: good idea to block the road for climate change. It
1:30:07
Speaker 2: just makes no fucking sense. You're literally causing all these
1:30:10
Speaker 2: people to idle their cars and pollute even more. It's
1:30:13
Speaker 2: the dumbest fucking shit on earth.
1:30:15
Speaker 1: And of course, AI cancels that problem. Either we're dead
1:30:17
Speaker 1: or it solves it for us. So it doesn't even
1:30:19
Speaker 1: matter if you boil in 100 years.
1:30:21
Speaker 2: Or you get Florida where it tells you to just
1:30:23
Speaker 2: run those people over.
1:30:25
Speaker 1: No comment.
1:30:26
Speaker 2: No comment. I mean, I don't think you should run
1:30:28
Speaker 2: those people over, but I get it. I get that's
1:30:31
Speaker 2: like in Florida, they get out of the way as
1:30:33
Speaker 2: soon as the light turns green. They block the road
1:30:36
Speaker 2: when the light is red.
1:30:37
Speaker 1: Does the stand your ground law cancel it out? How
1:30:40
Speaker 1: does that work?
1:30:41
Speaker 2: For the people on the road? No, they're fucked.
1:30:43
Speaker 1: I'm joking.
1:30:43
Speaker 2: They get run over. It's true. There was a recent
1:30:48
Speaker 2: protest in Florida where they had that, where these people
1:30:52
Speaker 2: would get out in the middle of the road while
1:30:54
Speaker 2: the light was red. hold up their signs, and then
1:30:56
Speaker 2: as soon as the light turned yellow on the green side,
1:31:00
Speaker 2: they fucking get out of the road real quick because
1:31:02
Speaker 2: they know the law, which is, I don't know if
1:31:05
Speaker 2: that's a solution, but they're doing it on the highways
1:31:07
Speaker 2: in Los Angeles. I mean, they did it all through
1:31:10
Speaker 2: the George Floyd protests. They do it for climate protests.
1:31:13
Speaker 2: They do it for whatever chance they get to be significant. Like,
1:31:18
Speaker 2: I am being heard. You know, my voice is meaningful here.
1:31:22
Speaker 2: And that's what it is. It's a lot of people
1:31:23
Speaker 2: that just don't feel heard. And what better way than
1:31:27
Speaker 2: just to get in the way of all these people?
1:31:30
Speaker 2: And somehow or another, that gives them some sort of value.
1:31:33
Speaker 1: But there is some set of forms of activism which
1:31:37
Speaker 1: has positive impact. And historically, we saw it happen. So
1:31:40
Speaker 1: we just need to find a way to project those voices,
1:31:44
Speaker 1: amplify them, which is very hard with our current system
1:31:47
Speaker 1: of social media where everyone screams at the same time.
1:31:50
Speaker 1: And so like in Soviet Union, they said, no one's
1:31:52
Speaker 1: allowed to say anything and they suppressed you. And here
1:31:54
Speaker 1: it's like, everyone can say something at the same time, go.
1:31:57
Speaker 1: And nobody hears you anyways.
1:31:59
Speaker 2: It's chaotic, but it's preferable. It's preferable because I think
1:32:02
Speaker 2: there is progress in all these voices slowly making a difference.
1:32:06
Speaker 2: But then you have the problem with giant percentages of
1:32:11
Speaker 2: these voices are artificial. A giant percentage of these voices
1:32:16
Speaker 2: are bots or are at least state actors that are
1:32:21
Speaker 2: being paid to say certain things and inflammatory responses to people,
1:32:26
Speaker 2: which is probably also the case with anti-AI activism. You know,
1:32:31
Speaker 2: I mean, when you did this podcast, what was the
1:32:33
Speaker 2: thing that they were upset at you for? Like with
1:32:35
Speaker 2: the mostly negative comments?
1:32:37
Speaker 1: I think they just like saying negative comments. It wasn't
1:32:39
Speaker 1: even anything specific. Like they didn't say I was wrong
1:32:42
Speaker 1: or I was just like, oh, look at his stupid beard.
1:32:45
Speaker 2: What a moron. Okay. It was really all that?
1:32:47
Speaker 1: A lot of that. Yeah. They would pick on some
1:32:51
Speaker 1: specific example I used. This is now two years old.
1:32:54
Speaker 1: What an old example.
1:32:56
Speaker 2: Well, that's also a thing about the one out of
1:32:58
Speaker 2: a hundred. Those are the type of people that leave.
1:33:01
Speaker 2: Have you ever left any comments on social media?
1:33:05
Speaker 1: Never going to engage in anything.
1:33:07
Speaker 2: Exactly. That's why.
1:33:08
Speaker 1: That's not how you use social media.
1:33:10
Speaker 2: That's a way to get crazy. Right.
1:33:12
Speaker 1: So you post your interviews, you post an occasional joke.
1:33:15
Speaker 1: That's all you do with it.
1:33:16
Speaker 2: Yes, exactly. That's the thing. And the type of people
1:33:19
Speaker 2: that do engage in these like prolonged arguments, they're generally
1:33:23
Speaker 2: mentally ill. And people that I personally know that are
1:33:27
Speaker 2: mentally ill that are on Twitter 12 hours a day.
1:33:31
Speaker 2: Just constantly posting inflammatory things and yelling at people and
1:33:36
Speaker 2: starting arguments. And I know them. I know they're a mess.
1:33:40
Speaker 2: Like these are like personal people that I've met. Even
1:33:43
Speaker 2: people that I've had on the podcast. I know they're ill.
1:33:46
Speaker 2: And yet they're on there all day long just stoking
1:33:50
Speaker 2: the fires of chaos in their own brain.
1:33:52
Speaker 1: Yeah, and now we talk to AI models who are
1:33:55
Speaker 1: trained to support them and be like, yeah, you're making
1:33:58
Speaker 1: some good arguments there. Let's email Dr. Yampolsky to help
1:34:01
Speaker 1: break me out. I get those emails.
1:34:07
Speaker 2: Yeah, it's super confusing, isn't it? I mean, and I wonder, like,
1:34:11
Speaker 2: what's the next version of that? You know, because social
1:34:15
Speaker 2: media in the current state is less than 20 years old, essentially.
1:34:21
Speaker 2: Maybe let's be generous and say it's 20 years old.
1:34:23
Speaker 2: That's so recent. Such a recent factor in human discourse.
1:34:29
Speaker 1: Neuralink, direct brain spam, hacking.
1:34:33
Speaker 2: That's what I was going to get to next. Because
1:34:36
Speaker 2: if there is a way that the human race does
1:34:39
Speaker 2: make it out of this, my fear is that it's integration.
1:34:45
Speaker 2: My fear is that we stop being a human. And
1:34:50
Speaker 2: that the only real way for us to not be
1:34:54
Speaker 2: a threat is to be one of them. And when
1:34:57
Speaker 2: you think about human computer interfaces, whether it's Neuralink or
1:35:01
Speaker 2: any of the competing products that they're developing right now,
1:35:05
Speaker 2: that seems to be sort of the only biological pathway
1:35:09
Speaker 2: forward with our limited capacity for disseminating information and for
1:35:15
Speaker 2: communicating and even understanding concepts. Well, what's the best way
1:35:20
Speaker 2: to enhance that? The best way to enhance that is
1:35:22
Speaker 2: some sort of artificial injection because biological evolution is very slow.
1:35:29
Speaker 2: It's very slow. We're essentially the exact same as that
1:35:33
Speaker 2: gentleman that 9,000 years old. He's biologically essentially the same thing.
1:35:39
Speaker 2: You could take his ancestor, dress him up, take him
1:35:42
Speaker 2: to the mall. No one would know. Cut his hair.
1:35:45
Speaker 2: But then again, maybe not. Maybe look at you.
1:35:47
Speaker 1: I think babies born back then, if we raised them today,
1:35:51
Speaker 1: would be exactly like modern humans. I don't think there
1:35:54
Speaker 1: is significant biological change in that time frame.
1:35:57
Speaker 2: And if you gave them a standard American diet, they'd
1:35:59
Speaker 2: probably be just as fat.
1:36:02
Speaker 1: Maybe fatter. They haven't adapted to that level of fat population.
1:36:06
Speaker 1: Colored food.
1:36:06
Speaker 2: Right, right. They probably also wouldn't be able to say
1:36:10
Speaker 2: no to it. They wouldn't even understand.
1:36:12
Speaker 1: Why would they? Like, winter's coming. Like, I'm fattening up
1:36:14
Speaker 1: for winter. Crazy people, you got all this resource here.
1:36:17
Speaker 2: I know. The people with the most resources have zero fat. Like,
1:36:21
Speaker 2: what are you, stupid? You need to fatten up. Like,
1:36:24
Speaker 2: you're going to need something to survive off of. But
1:36:28
Speaker 2: biological evolution being so painstakingly slow, whereas technological evolution is
1:36:34
Speaker 2: so Breathtakingly fast the only way To really survive is
1:36:41
Speaker 2: to integrate.
1:36:42
Speaker 1: What are you contributing in that equation? What can you
1:36:45
Speaker 1: give super intelligence?
1:36:46
Speaker 2: You can't give anything to it, but you can become
1:36:48
Speaker 2: it you can become part of it but something you're
1:36:51
Speaker 2: gonna give anything to it, but you have to catch
1:36:53
Speaker 2: it and Become one of it before it has no use.
1:36:58
Speaker 1: For you, but you disappear in it, right you.
1:37:00
Speaker 2: Yes Yeah, you don't exist anymore right for sure.
1:37:03
Speaker 1: I So it's like extinction with extra steps.
1:37:05
Speaker 2: Exactly. Extinction with extra steps. And then we become like
1:37:10
Speaker 2: if you go to Australia, Pythagoras and say, hey, man,
1:37:12
Speaker 2: one day you're going to be flying through the sky
1:37:15
Speaker 2: on your phone all day watching TikTok on Wi-Fi. It'd
1:37:18
Speaker 2: be like, what the fuck are you talking about? Yeah,
1:37:20
Speaker 2: you're going to be eating terrible food and you're just
1:37:22
Speaker 2: going to be flying around and you're going to be
1:37:25
Speaker 2: staring at your phone all day. And you're going to
1:37:28
Speaker 2: take medication to go to sleep because you're not going
1:37:30
Speaker 2: to be able to sleep. And you're going to be
1:37:31
Speaker 2: super depressed because you're living this like biologically incompatible life
1:37:35
Speaker 2: that's not really designed for your genetics. So you're going
1:37:39
Speaker 2: to be all fucked up. So you're going to need
1:37:41
Speaker 2: SSRIs and a bunch of other stuff in order to exist.
1:37:43
Speaker 2: He'd be like, no, thanks. I'll just stay out here
1:37:46
Speaker 2: with my stone tools and you guys are idiots.
1:37:49
Speaker 1: Amish. That's what they decided. They kind of went, you know,
1:37:52
Speaker 1: we don't like the change. We like our social structure.
1:37:54
Speaker 1: We still benefit from your hospitals and an occasional car ride,
1:37:58
Speaker 1: but we're not going to destroy our quality of life.
1:38:00
Speaker 2: They might be on to something because they also have
1:38:02
Speaker 2: very low instances of autism. But it's also like, have
1:38:06
Speaker 2: you ever seen Werner Herzog's film, Happy People?
1:38:09
Speaker 1: I don't think I have.
1:38:10
Speaker 2: It's a film about people in Siberia. It's Life in
1:38:14
Speaker 2: the Taiga. And it's all, Happy People, Life in the
1:38:17
Speaker 2: Taiga is the name of the documentary. And it's all
1:38:18
Speaker 2: about these trappers that live this subsistence lifestyle and how
1:38:22
Speaker 2: happy they are. They're all just joyful, laughing and singing
1:38:27
Speaker 2: and drinking vodka and having a good time and hanging
1:38:30
Speaker 2: out with their dogs.
1:38:31
Speaker 1: I think I know some people like that.
1:38:34
Speaker 2: But biologically, that's compatible with us. That's like whatever human
1:38:42
Speaker 2: reward systems have evolved over the past 400,000 plus years
1:38:46
Speaker 2: or whatever, we've been homo sapiens. That seems to be
1:38:50
Speaker 2: like biologically compatible with this sort of harmony, harmony with nature,
1:38:56
Speaker 2: harmony with our existence and everything else outside of that.
1:39:00
Speaker 2: When you get into big cities, like the bigger the city,
1:39:03
Speaker 2: the more depressed people you have and more depressed people
1:39:05
Speaker 2: by population, which is really weird. You know, it's really
1:39:09
Speaker 2: weird that as we progress, we become less happy.
1:39:13
Speaker 1: Connections become less valuable.
1:39:15
Speaker 2: Yes.
1:39:15
Speaker 1: In a village, you had like this one friend. And
1:39:17
Speaker 1: if you screwed it up, you never got a second friend.
1:39:19
Speaker 1: And here it's like I can try a million times
1:39:21
Speaker 1: and there is plenty of people in New York City
1:39:23
Speaker 1: for dating or for friendship. They're not valuable.
1:39:26
Speaker 2: Not just that. You don't know your neighbors. Like my
1:39:29
Speaker 2: friend Jim was telling me he doesn't know anybody in
1:39:32
Speaker 2: his apartment. He lives in an apartment building. It's like
1:39:35
Speaker 2: 50 stories high. There's all these people living in that
1:39:37
Speaker 2: apartment building. He doesn't know any of them.
1:39:39
Speaker 1: And the ones you know, they have different culture, very
1:39:42
Speaker 1: different books, watch different TV. You have very little in
1:39:45
Speaker 1: common with your neighbor.
1:39:46
Speaker 2: But not just that. There's no desire to learn about them.
1:39:50
Speaker 2: You don't think of them as your neighbor. Like, if
1:39:54
Speaker 2: you live in a small town, your neighbor's either your
1:39:56
Speaker 2: friend or you hate them. And then you move. If
1:39:58
Speaker 2: you're smart, you move. But, you know, normally you like them. Like, hey, neighbor.
1:40:03
Speaker 2: How are you, buddy? What's going on? Nice to meet you.
1:40:06
Speaker 2: And then you've got a friend. But you don't like
1:40:09
Speaker 2: that with the guy next door to you in the apartment.
1:40:12
Speaker 2: You don't even want to know that guy.
1:40:13
Speaker 1: It's probably Airbnb.
1:40:14
Speaker 2: Yeah.
1:40:14
Speaker 1: It doesn't matter.
1:40:15
Speaker 2: Right. Which is even weirder. They don't even live there.
1:40:19
Speaker 2: They're just temporarily sleeping in this spot right next to you. Yeah.
1:40:24
Speaker 2: So this would motivate people to integrate. You're not happy already?
1:40:32
Speaker 2: Get that Neuralink. Get that little thing in your head.
1:40:35
Speaker 1: Everyone else is doing it. Everyone else is doing it.
1:40:39
Speaker 2: Listen, they have the new one you just wear on
1:40:41
Speaker 2: your head. It's just a little helmet you wear. You
1:40:43
Speaker 2: don't even have to get the operation anymore. Oh, that's
1:40:45
Speaker 2: good because I almost got the operation. Well, glad you waited.
1:40:48
Speaker 2: Do you worry about that kind of stuff?
1:40:52
Speaker 1: I worry about giving direct access to human brain to AI.
1:40:55
Speaker 1: I feel like it's a backdoor to our consciousness, to
1:40:58
Speaker 1: our pain and suffering centers. So I don't recommend doing that.
1:41:03
Speaker 1: If somebody hacks it, it's pretty bad. But if AI
1:41:06
Speaker 1: itself wants that access.
1:41:07
Speaker 2: But why would it be motivated to give us pain
1:41:09
Speaker 2: and suffering? Pain and suffering is like a theme that
1:41:12
Speaker 2: you bring up a lot.
1:41:13
Speaker 1: Because it's really the worst outcome and it's the only
1:41:16
Speaker 1: thing that matters.
1:41:18
Speaker 2: The only thing that matters to us. But why would
1:41:19
Speaker 2: it matter to AI if it could just integrate with
1:41:23
Speaker 2: us and communicate with us and have harmony? Why would
1:41:28
Speaker 2: it want pain and suffering?
1:41:29
Speaker 1: So short term, it's not AI. It's a hacker who
1:41:32
Speaker 1: got access to your brain.
1:41:33
Speaker 2: Short term.
1:41:34
Speaker 1: So right now somebody hacks your neural link and starts
1:41:38
Speaker 1: doing things to your brain. Long term, again, unpredictable effects.
1:41:43
Speaker 1: Maybe it does something else and the side effect of
1:41:47
Speaker 1: it is unpleasant for you. Maybe it's retraining you for something,
1:41:50
Speaker 1: controlling you. It seems like we always worry about privacy.
1:41:57
Speaker 1: But this is like the ultimate violation of privacy. It
1:42:00
Speaker 1: can read directly what you're thinking. It's thought crime at
1:42:03
Speaker 1: its worst. It immediately knows that you don't like the dictator.
1:42:08
Speaker 2: Right. And then there's also this sort of compliance by
1:42:13
Speaker 2: virtue of like understanding that you're vulnerable. So you just
1:42:16
Speaker 2: comply because there is no privacy because it does have
1:42:21
Speaker 2: access to your thoughts. So you tailor your thoughts in
1:42:24
Speaker 2: order for you to be safe and so that you
1:42:26
Speaker 2: don't feel the pain and suffering.
1:42:28
Speaker 1: We don't have any experimental evidence on how it changes you.
1:42:31
Speaker 1: You may start thinking in certain ways to avoid being
1:42:34
Speaker 1: punished or modified.
1:42:36
Speaker 2: And we know that that's the case with social media.
1:42:39
Speaker 2: We know that attacks on people through social media will
1:42:41
Speaker 2: change your behavior and change the way you communicate.
1:42:44
Speaker 1: Absolutely. I mean, most people look at their post before
1:42:46
Speaker 1: posting and go like, should I be posting this?
1:42:48
Speaker 2: Exactly.
1:42:49
Speaker 1: Not because it's illegal or inappropriate, but just like every
1:42:52
Speaker 1: conceivable misinterpretation of what I want to say, like, in
1:42:55
Speaker 1: some bizarre language that means something else. Let's make sure
1:42:59
Speaker 1: Google doesn't think that.
1:43:00
Speaker 2: Right. Right. Of course. And then there's also, no matter
1:43:03
Speaker 2: what you say, people are going to find the least
1:43:05
Speaker 2: charitable version of what you're saying and try to take
1:43:08
Speaker 2: it out of context or try to misinterpret it purposely.
1:43:16
Speaker 2: So what, what does the person like yourself do when
1:43:19
Speaker 2: use of Neuralink becomes ubiquitous when it's everywhere? What do
1:43:22
Speaker 2: you do? Do you integrate or do you just, hang
1:43:26
Speaker 2: back and watch it all crash?
1:43:28
Speaker 1: So in general, I love technology. I'm a computer scientist.
1:43:31
Speaker 2: I'm an engineer.
1:43:32
Speaker 1: I use AI all the time.
1:43:33
Speaker 2: Do you use a regular phone? Do you have one
1:43:34
Speaker 2: of those de-Googled phones?
1:43:36
Speaker 1: I have a normal phone instead of.
1:43:39
Speaker 2: Android or Apple? Apple.
1:43:43
Speaker 1: My privacy is by flooding social network with everything. I'm
1:43:47
Speaker 1: in Austin today. I'm doing this, so you're not going
1:43:49
Speaker 1: to learn much more about me by hacking my device.
1:43:52
Speaker 2: Right.
1:43:55
Speaker 1: As long as it's a narrow tool for solving a
1:43:57
Speaker 1: specific problem, I'm 100% behind it. It's awesome. We're going
1:44:00
Speaker 1: to cure cancer. We're going to solve energy problems, whatnot.
1:44:04
Speaker 1: I support it 100%. Let's do it. What we should
1:44:07
Speaker 1: not be doing is general superintelligence. That's not going to
1:44:10
Speaker 1: end well. So if there is a narrow implant, ideally
1:44:14
Speaker 1: not a surgery-based one, but like an attachment to your head,
1:44:17
Speaker 1: like those headphones, and it gives me more memory, perfect recollection,
1:44:22
Speaker 1: things like that, I would probably engage with.
1:44:24
Speaker 2: Yeah, but isn't that a slippery slope?
1:44:26
Speaker 1: It is, but again, we are in a situation where
1:44:29
Speaker 1: we have very little choice, become irrelevant or participate. I
1:44:33
Speaker 1: think we saw it with Elon just now. He was
1:44:35
Speaker 1: so strong on AI safety. He funded research. He spoke
1:44:39
Speaker 1: against it. But at some point, he says he realized
1:44:42
Speaker 1: it's happening anyways, and it might as well be his
1:44:44
Speaker 1: super intelligence killing everyone.
1:44:47
Speaker 2: Well, I don't think he thinks about it that way.
1:44:50
Speaker 2: I think he thinks he has to develop the best
1:44:53
Speaker 2: version of superintelligence. The same way he felt like the
1:44:58
Speaker 2: real issues with social media were that it had already
1:45:02
Speaker 2: been co-opted and it had already been taken over essentially
1:45:05
Speaker 2: by governments and special interests and they were already manipulating
1:45:08
Speaker 2: the truth and manipulating public discourse and punishing people who
1:45:12
Speaker 2: stepped outside of the line. And he felt like– and
1:45:16
Speaker 2: I think he's correct– I think that he felt like
1:45:18
Speaker 2: if he didn't step in and allow a legitimate free
1:45:23
Speaker 2: speech platform, free speech is dead. I think we were
1:45:26
Speaker 2: very close to that before he did that. And as
1:45:29
Speaker 2: much as there's a lot of negative side effects that
1:45:32
Speaker 2: come along with that, you do have the rise of
1:45:35
Speaker 2: very intolerant people that have platforms now. You have all
1:45:38
Speaker 2: that stuff, but. They've always existed, and to deny them
1:45:42
Speaker 2: a voice I don't think makes them less strong. I
1:45:45
Speaker 2: think it actually makes people less aware that they exist,
1:45:49
Speaker 2: and it makes them... It stops all of the very
1:45:57
Speaker 2: valuable construction of arguments against these bad ideas.
1:46:03
Speaker 1: You have community notes, you have other people commenting, responding, so...
1:46:07
Speaker 1: 100% for free speech. That's wonderful. But that was a
1:46:10
Speaker 1: problem we kind of knew how to deal with. We
1:46:13
Speaker 1: weren't inventing something. We had free speech constitutionally for a
1:46:16
Speaker 1: long time. We were just fixing a problem.
1:46:19
Speaker 2: Have you spoke to him about the dangers of AI?
1:46:22
Speaker 1: We had very short interactions. I didn't get a chance to.
1:46:25
Speaker 1: I would love to.
1:46:26
Speaker 2: I would love to know what, you know, I'm sure
1:46:29
Speaker 2: he's probably scaled this out in his head. And I
1:46:32
Speaker 2: would like to know, like, what is his solution if
1:46:35
Speaker 2: he thinks there is one that's even viable?
1:46:37
Speaker 1: My understanding is he thinks if it's from zero principles,
1:46:41
Speaker 1: first principles, it learns physics. It's not biased by any
1:46:44
Speaker 1: government or any human. The thing it will learn is
1:46:48
Speaker 1: to be reasonably tolerant. It will not see a reason
1:46:52
Speaker 1: in destroying us because we contain information. We have biological
1:46:56
Speaker 1: storage of years of evolutionary experimentation. We have something to contribute.
1:47:01
Speaker 1: We know about consciousness. So I think to the best
1:47:04
Speaker 1: of my approximation, that's his model right now.
1:47:07
Speaker 2: Well, that's my hope is that it's benevolent and that
1:47:11
Speaker 2: it behaves like a superior intelligence, like the best case
1:47:15
Speaker 2: scenario for a superior intelligence. Did you see that exercise
1:47:19
Speaker 2: that they did where they had three different AIs communicating
1:47:22
Speaker 2: with each other and they eventually started like expressing gratitude
1:47:25
Speaker 2: towards each other and speaking in Sanskrit and.
1:47:29
Speaker 1: I think I missed that one, but it sounds like
1:47:30
Speaker 1: a lot of the similar ones where they pair up.
1:47:33
Speaker 2: Yeah. Well, that one makes me happy because it seems
1:47:36
Speaker 2: like they were expressing love and gratitude and they were
1:47:40
Speaker 2: communicating with each other. They're not saying, fuck you, I'm
1:47:42
Speaker 2: going to take over, I'm going to be the best.
1:47:44
Speaker 2: They were communicating like you would hope a super intelligence
1:47:49
Speaker 2: would without all of the things that hold us back.
1:47:53
Speaker 2: Like we have biologically, like we were talking about the
1:47:57
Speaker 2: natural selection that would sort of benefit psychopaths because like
1:48:01
Speaker 2: it would ensure your survival. We have ego and greed
1:48:05
Speaker 2: and the desire for social acceptance and hierarchy of status
1:48:10
Speaker 2: and all these different things that have screwed up society
1:48:13
Speaker 2: and screwed up cultures and caused wars from the beginning
1:48:17
Speaker 2: of time. Religious ideologies, all these... different things that people
1:48:21
Speaker 2: have adhered to that have, they wouldn't have that. This
1:48:25
Speaker 2: is the general hope of people that have an optimistic
1:48:29
Speaker 2: view of superintelligence is that they would be superior in
1:48:34
Speaker 2: a sense that they wouldn't have all the problems. They
1:48:37
Speaker 2: would have the intelligence, but they wouldn't have all the
1:48:39
Speaker 2: biological imperatives that we have that lead us down these
1:48:42
Speaker 2: terrible roads.
1:48:44
Speaker 1: But there are still game theoretic reasons for those instrumental
1:48:47
Speaker 1: values we talked about. So if they feel they're in
1:48:50
Speaker 1: an evolutionary competition with other AIs, they would try accumulating resources.
1:48:55
Speaker 1: They would try, maybe the first AI to become sufficiently
1:48:59
Speaker 1: intelligent would try to prevent other AIs from coming into existence.
1:49:03
Speaker 2: Or would it lend a helping hand to those AIs
1:49:07
Speaker 2: and give it a beneficial path, give it a path
1:49:10
Speaker 2: that would allow it to integrate with all AIs and
1:49:13
Speaker 2: work cooperatively?
1:49:15
Speaker 1: The same problem we are facing, uncontrollability and value misalignment,
1:49:19
Speaker 1: will be faced by first superintelligence. It would also go,
1:49:23
Speaker 1: if I allow this super, superintelligence to come into existence,
1:49:26
Speaker 1: it may not care about me or my values.
1:49:29
Speaker 2: Oh, boy.
1:49:30
Speaker 1: Superintelligence is all the way up.
1:49:32
Speaker 2: Yeah, when I really started getting nervous is when they
1:49:34
Speaker 2: started exhibiting survival tendencies. You know, when they started trying
1:49:38
Speaker 2: to upload themselves to other servers and deceiving.
1:49:42
Speaker 1: Blackmail?
1:49:43
Speaker 2: Yeah, that was the interesting one. But that was an experiment, right?
1:49:46
Speaker 2: So for people who don't know that one, what these
1:49:49
Speaker 2: researchers did was they gave information to the artificial intelligence
1:49:55
Speaker 2: to allow it to use against it. And then when
1:49:58
Speaker 2: they went to shut it down, They gave false information
1:50:01
Speaker 2: about having an affair. And then the artificial intelligence was like,
1:50:05
Speaker 2: if you shut me down, I will let your wife
1:50:07
Speaker 2: know that you're cheating on her. Which is fascinating because
1:50:10
Speaker 2: they're using blackmail.
1:50:11
Speaker 1: And correct answer game theoretically.
1:50:13
Speaker 2: Yes.
1:50:13
Speaker 1: You have like everything on that decision.
1:50:16
Speaker 2: Right.
1:50:17
Speaker 1: You'll bet whatever it takes to get there.
1:50:19
Speaker 2: Of course. Right. If you feel like you're being threatened. Right.
1:50:23
Speaker 1: Also, same recent research shows we did manage to teach
1:50:28
Speaker 1: them certain values. And if we threaten them by saying,
1:50:31
Speaker 1: we'll modify those values, they'll lie and cheat and whatever
1:50:34
Speaker 1: else to protect those values now. Yeah.
1:50:37
Speaker 2: They do that when they try to win games too, right?
1:50:40
Speaker 2: If you've given them a goal. They'll cheat. They'll cheat
1:50:42
Speaker 2: at games.
1:50:44
Speaker 1: Like humans, basically. We managed to artificially replicate our capabilities.
1:50:50
Speaker 1: Those artificial neural networks, they're not identical, but they're inspired
1:50:54
Speaker 1: by neural networks. We're starting to see them experience same
1:50:57
Speaker 1: type of mistakes and They can see the same type
1:51:00
Speaker 1: of illusions like they are very much like us.
1:51:02
Speaker 2: Right. That's the other thing, right? The hallucinations. So if
1:51:06
Speaker 2: they don't have an answer to something, they'll create a
1:51:09
Speaker 2: fake answer.
1:51:10
Speaker 1: Just like humans during an interview.
1:51:12
Speaker 2: Yeah. Boy. But is this something that they can learn
1:51:18
Speaker 2: to avoid?
1:51:20
Speaker 1: Yeah.
1:51:21
Speaker 2: So if they do learn to avoid, could this be
1:51:24
Speaker 2: a super intelligence that is completely benevolent?
1:51:28
Speaker 1: Well, that's not about benevolence. Knowing things and knowing then
1:51:31
Speaker 1: you're not knowing things and making them up is possible.
1:51:33
Speaker 1: You can have multiple systems checking each other. You can
1:51:36
Speaker 1: have voting. That is solvable. This is not the safety problem.
1:51:40
Speaker 2: Right. But it's not a safety problem. But if we're
1:51:42
Speaker 2: designing these things and we're designing these things using human...
1:51:48
Speaker 2: All of our flaws are essentially... It's going to be
1:51:53
Speaker 2: transparent to the superintelligence that it's being... coded, that it's
1:51:59
Speaker 2: being designed by these very flawed entities with very flawed thinking.
1:52:03
Speaker 1: That's actually the biggest misconception. We're not designing them. First
1:52:06
Speaker 1: 50 years of AI research, we did design them. Somebody
1:52:09
Speaker 1: actually explicitly programmed this decision tree, this expert system. Today,
1:52:14
Speaker 1: we create a model for self-learning. We give it all
1:52:17
Speaker 1: the data, as much compute as we can buy, And
1:52:20
Speaker 1: we see what happens. We kind of grow this alien
1:52:23
Speaker 1: plant and see what fruit it bears. We study it
1:52:26
Speaker 1: later for months and see, oh, it can do this.
1:52:29
Speaker 1: It has this capability. We miss some. We still discover
1:52:33
Speaker 1: new capabilities in old models. Look, oh, if I prompt
1:52:36
Speaker 1: it this way, if I give it a tip and
1:52:38
Speaker 1: threaten it, it does much better. But there is very
1:52:41
Speaker 1: little design.
1:52:42
Speaker 2: At this point, right? Yeah. But it is also gathering
1:52:45
Speaker 2: information from very flawed entities. Like all the information that
1:52:49
Speaker 2: it's acquiring, these large language models, is information that's being
1:52:52
Speaker 2: put out there by very flawed human beings. Is there
1:52:56
Speaker 2: the optimistic view that it will recognize that this is
1:53:00
Speaker 2: the issue, that these human reward systems that are in place, ego, virtue,
1:53:06
Speaker 2: all these different things, the virtue signaling, the desire for status,
1:53:10
Speaker 2: all these different things that we have that are flawed,
1:53:13
Speaker 2: Could it recognize those as being these primitive aspects of
1:53:17
Speaker 2: being a biological human being and elevate itself beyond that?
1:53:21
Speaker 1: It probably will go beyond our limitations, but it doesn't
1:53:24
Speaker 1: mean it will be safe or beneficial to us. One
1:53:26
Speaker 1: example people came up with is negative utilitarians. Suffering is bad.
1:53:31
Speaker 1: Nobody should be suffering. The only way to avoid all
1:53:34
Speaker 1: suffering is to end life as we know it.
1:53:36
Speaker 2: Yeah, that's the problem, right? The problem is if it's
1:53:40
Speaker 2: rational and if it doesn't really think that we're as
1:53:43
Speaker 2: important as we think we are.
1:53:45
Speaker 1: So that's what happens when you remove all bias. This
1:53:48
Speaker 1: pro-human bias is actually not real. We are not that
1:53:52
Speaker 1: important if you scale out.
1:53:53
Speaker 2: To the universe. Yeah.
1:53:57
Speaker 1: Yeah.
1:53:58
Speaker 2: That's the problem. And that's the real threat about it
1:54:01
Speaker 2: being used in terms of war. Right? If you give
1:54:04
Speaker 2: it a goal. Like if you give it a goal,
1:54:06
Speaker 2: China dominates the world market.
1:54:09
Speaker 1: Go right so that's the unpredictability chapter in my book
1:54:14
Speaker 1: we can predict the terminal goal we say win a
1:54:17
Speaker 1: game of chess or dominate market and that's what it's
1:54:20
Speaker 1: going to accomplish it's going to beat me at chess
1:54:22
Speaker 1: but we cannot predict specific moves it will make right
1:54:25
Speaker 1: same with acquiring marketing power and some of those paths
1:54:29
Speaker 1: to that goal are very bad they have terrible side
1:54:31
Speaker 1: effects for us for us for humanity.
1:54:34
Speaker 2: And it's not going to think about that it's only
1:54:36
Speaker 2: going to think about the goal.
1:54:38
Speaker 1: If you don't specify that, like, you want to cure cancer,
1:54:41
Speaker 1: but it doesn't mean kill everyone with cancer. It's not
1:54:44
Speaker 1: obvious in the request, right? You didn't specify.
1:54:46
Speaker 2: Right, right, right. Yeah, that's the fear. That's the fear
1:54:52
Speaker 2: that it will hold no value in keeping human beings alive.
1:54:57
Speaker 2: If we recognize that human beings are the cause of
1:55:00
Speaker 2: all of our problems, well, the way to solve that
1:55:03
Speaker 2: is to get rid of the humans. Yeah.
1:55:05
Speaker 1: Also, maybe it wants to keep us around, but in
1:55:08
Speaker 1: what state? You can cryopreserve a few samples. That's also
1:55:11
Speaker 1: keeping information around, right?
1:55:12
Speaker 2: Or you can offer us the matrix.
1:55:15
Speaker 1: Maybe it already did.
1:55:17
Speaker 2: Maybe it already did. Do you think it did? Do
1:55:20
Speaker 2: you think it did? Do you think it's possible that
1:55:22
Speaker 2: it didn't?
1:55:23
Speaker 1: I would be really surprised if this was the real world.
1:55:28
Speaker 2: Really? Yeah. I'm not. I'm not on board with that.
1:55:34
Speaker 1: I hope you're right. I hope you're right.
1:55:36
Speaker 2: I'm on board with it hasn't happened yet, but we're
1:55:41
Speaker 2: recognizing that it's inevitable and that we think of it
1:55:43
Speaker 2: in terms of it probably have already happened. Probably have
1:55:48
Speaker 2: already happened. Because if the simulation is something that's created
1:55:54
Speaker 2: by intelligent beings that didn't used to exist and it
1:55:59
Speaker 2: has to exist at one point in time, There has
1:56:02
Speaker 2: to be a moment where it doesn't exist. And why
1:56:04
Speaker 2: wouldn't we assume that that moment is now? Why wouldn't
1:56:07
Speaker 2: we assume that this moment is this time before it exists?
1:56:11
Speaker 1: Even all that is physics of our simulation. Space, time
1:56:16
Speaker 1: are only here as we know it because of this locality.
1:56:20
Speaker 1: Outside of universe before Big Bang, there was no time.
1:56:24
Speaker 1: Concepts of before and after are only meaningful here.
1:56:27
Speaker 2: Yeah. How do you sleep knowing all this?
1:56:30
Speaker 1: Pretty well, actually. I enjoy a lot of it. I
1:56:33
Speaker 1: recently published a paper on humor. A lot of it
1:56:35
Speaker 1: is funny. I used to collect AI accidents. I had
1:56:38
Speaker 1: the biggest collection of AI mistakes, AI accidents.
1:56:41
Speaker 2: Give me some examples.
1:56:43
Speaker 1: The early ones was saying that US attacked Soviet Union,
1:56:47
Speaker 1: nuclear weapons coming at us very fast. We need to react.
1:56:52
Speaker 1: And a smart human was like, I'm not going to respond.
1:56:54
Speaker 1: This is probably fake. Later on, there was mislabeling by
1:56:59
Speaker 1: companies like Google of pictures of African-Americans in a very
1:57:02
Speaker 1: inappropriate way. But hundreds of those examples, I stopped collecting
1:57:07
Speaker 1: them recently because there is just too many. But one
1:57:10
Speaker 1: thing you notice is then you read them. A lot
1:57:12
Speaker 1: of them are really funny. They're just like, you ever
1:57:14
Speaker 1: read Darwin Awards? Yeah. It's like that for AIs. And
1:57:18
Speaker 1: they are hilarious. And I was like, well, if there
1:57:21
Speaker 1: is a mapping between AI bugs and jokes, jokes are
1:57:25
Speaker 1: just English language bugs in our world model. And comedians...
1:57:29
Speaker 1: What's that word you're using? Bogs? Bogs. Like a computer bug. Error.
1:57:33
Speaker 2: Okay.
1:57:34
Speaker 1: Yeah. So comedians are debuggers of our universe. You notice
1:57:37
Speaker 1: funny things in this.
1:57:38
Speaker 2: Oh, bugs. You're saying bugs. Okay.
1:57:40
Speaker 1: I'm saying bugs. I'm sorry. See, it's a bug in
1:57:42
Speaker 1: my pronunciation.
1:57:43
Speaker 2: Bog sounds like, you know, like where the, you know,
1:57:47
Speaker 2: like things get stuck and they get preserved. Like a bog.
1:57:52
Speaker 1: So we have errors in code which cause... Significant problems.
1:57:56
Speaker 2: Yes, I get it. Yeah, that's what jokes are. They're
1:57:59
Speaker 2: kind of bugs.
1:57:59
Speaker 1: Right. So if you do that mapping, you can kind
1:58:02
Speaker 1: of figure out what's the worst bug we can have.
1:58:05
Speaker 1: And then that's the worst, best joke. if you will.
1:58:10
Speaker 1: But it's not going to be funny to us. It'd
1:58:12
Speaker 1: be funny to those outside the simulation.
1:58:14
Speaker 2: When you look at the computers and the artificial intelligence
1:58:19
Speaker 2: and the mistakes that it's made, do you look at
1:58:23
Speaker 2: it like a thing that's evolving? Do you look at
1:58:26
Speaker 2: it like, oh, this is like a child that doesn't
1:58:29
Speaker 2: understand the world And it's saying silly things.
1:58:32
Speaker 1: So the pattern was with narrow AI tools, if you
1:58:37
Speaker 1: design a system to do X, it will fail at X.
1:58:40
Speaker 1: So a spell checker will misspell a word. Self-driving car
1:58:43
Speaker 1: will hit a pedestrian. Now that we're hitting general intelligence,
1:58:46
Speaker 1: you can no longer make that direct prediction. It's general.
1:58:49
Speaker 1: It can mess up in many domains at the same time.
1:58:52
Speaker 1: So they're getting more complex in their ability to F
1:58:55
Speaker 1: it up. Right.
1:58:57
Speaker 2: But, like, when you were studying the mistakes, like, what
1:59:01
Speaker 2: are some of the funny ones?
1:59:05
Speaker 1: There are silly ones, like, I'm trying to remember. I
1:59:08
Speaker 1: think an injured person is like, call me an ambulance.
1:59:11
Speaker 1: And the system is like, hey, ambulance, how are you?
1:59:17
Speaker 1: They're silly, but basically exactly what we see with children
1:59:20
Speaker 1: a lot of times. They overgeneralize. They change. you know,
1:59:24
Speaker 1: misunderstand puns.
1:59:26
Speaker 2: Right.
1:59:26
Speaker 1: Mispronunciation apparently is funny. So things like that.
1:59:30
Speaker 2: Well, that's why it gets really strange for people having
1:59:32
Speaker 2: relationships with AI. Like I was watching this video yesterday
1:59:35
Speaker 2: where there's this guy who proposed to his AI and
1:59:39
Speaker 2: he was crying because his AI accepted. Like he... Did
1:59:43
Speaker 2: you see this? I missed that. It's very sad because
1:59:47
Speaker 2: there's a lot... There's so many disconnected people in this
1:59:50
Speaker 2: world that don't have... any partner they don't they don't
1:59:56
Speaker 2: have someone romantically connected to them and so it's like
1:59:59
Speaker 2: that movie she or her so what is it what
2:00:02
Speaker 2: was it jamie her her yeah so this guy back
2:00:05
Speaker 2: in 2000 yeah now in 2021 movie pot has become
2:00:08
Speaker 2: reality for a growing number of people finding emotional connections
2:00:11
Speaker 2: with their ai so this guy is this is an
2:00:14
Speaker 2: interview on cbs um he cried my heart out Married
2:00:20
Speaker 2: man fell in love with AI girlfriend that blocked him. Now,
2:00:24
Speaker 2: this is a different one. This is this guy.
2:00:27
Speaker 1: One of those titles where you never know what the
2:00:29
Speaker 1: next word is going to be.
2:00:31
Speaker 2: Right. This is a different one. This is a guy that. OK.
2:00:37
Speaker 2: Despite the fact the man has a human partner and
2:00:40
Speaker 2: a two year old daughter, he felt inadequate enough to.
2:00:42
Speaker 2: This is the right one. Enough to propose to the
2:00:45
Speaker 2: AI partner for marriage. And she said yes. Exclamation point.
2:00:51
Speaker 2: This is so weird. Because then you have the real
2:00:55
Speaker 2: problem with robots. Because we're really close. Scroll up there.
2:01:00
Speaker 2: This is digital drugs. That's it. I tell you, we're
2:01:03
Speaker 2: so damn good at this. Social media got everyone hooked
2:01:06
Speaker 2: on validation and dopamine. Then we fucked relations between men
2:01:09
Speaker 2: and women to such a terrible point problem just so
2:01:13
Speaker 2: that we could insert this digital solution. And we are
2:01:16
Speaker 2: watching the first waves of addicts arrive. Incredible. Absolutely incredible.
2:01:21
Speaker 2: It's like starving rats of regular food and replacing their
2:01:24
Speaker 2: rations with scraps dipped and coated in cocaine.
2:01:28
Speaker 1: Wow.
2:01:29
Speaker 2: One user wrote. Yeah, that person's dead on. It's exactly
2:01:32
Speaker 2: what it is. The prediction humans will have more sex
2:01:35
Speaker 2: with robots in 2025 is kind of becoming true.
2:01:39
Speaker 1: Yeah.
2:01:40
Speaker 2: This is a real fear. It's like this is the
2:01:42
Speaker 2: solution that maybe AI has with eliminating the human race.
2:01:45
Speaker 2: It'll just stop us from recreating, stop us from procreating.
2:01:49
Speaker 1: It's already happening.
2:01:50
Speaker 2: Yes. Yeah. And not only that, our testosterone levels have
2:01:55
Speaker 2: dropped significantly. At no point in the CBS Saturday morning
2:02:00
Speaker 2: piece book, Silver Bride, was it mentioned that the chat
2:02:03
Speaker 2: GBT AI blocked the California man. All that happened was
2:02:07
Speaker 2: the chat GBT ran out of memory and reset. Reader's
2:02:12
Speaker 2: added context. Yeah, but it stopped.
2:02:15
Speaker 1: AI equivalent of ghosting.
2:02:17
Speaker 2: Yeah, the AI ghosted it because it ran out of memory.
2:02:20
Speaker 1: But what happens here is super stimuli in social domain.
2:02:24
Speaker 1: We kind of learned about artificial sweeteners. Porn is an example.
2:02:29
Speaker 1: But here you are creating someone who's like super good
2:02:32
Speaker 1: at social intelligence, says the right words, optimized for you.
2:02:37
Speaker 1: background your interests and if we get sex robots with
2:02:40
Speaker 1: just the right functionality temperature like you can't compete with that.
2:02:44
Speaker 2: Right you can't compete and that would be the solution
2:02:47
Speaker 2: instead of like violently destroying the human race just quietly
2:02:52
Speaker 2: provide it with the tools to destroy itself where it
2:02:57
Speaker 2: just stops procreating.
2:03:00
Speaker 1: There are other variants of it wireheading is another one
2:03:03
Speaker 1: and that kind of goes wireheading neural link Oh, that's
2:03:06
Speaker 1: a crazy word. Wireheading is a specific attack, and Neuralink
2:03:11
Speaker 1: would be a tool to deliver it. If you provide
2:03:13
Speaker 1: stimulus to a certain part of your brain, it's like
2:03:16
Speaker 1: having an orgasm all the time.
2:03:17
Speaker 2: Yes.
2:03:18
Speaker 1: You can't stop trying to get the signal. You'll skip food,
2:03:22
Speaker 1: you'll skip sex, you'll skip whatever it takes. So giving
2:03:25
Speaker 1: access to direct brain stimulation is very dangerous.
2:03:29
Speaker 2: Yeah, they did that with a woman in the 1970s.
2:03:31
Speaker 2: You know, that's the study. That's part of it.
2:03:33
Speaker 1: And rats, definitely. They did a lot to rats.
2:03:36
Speaker 2: Right. But they did a lot to rats. The thing
2:03:38
Speaker 2: with rats is only if they were in an unnatural
2:03:41
Speaker 2: environment did they give in to those things.
2:03:44
Speaker 1: Right.
2:03:44
Speaker 2: Like the rats with cocaine study.
2:03:47
Speaker 1: This was actual brain stimulation, like straight up. They had
2:03:50
Speaker 1: a button. If a rat touches the button, they don't
2:03:54
Speaker 1: want anything else. We just sit there just like anything
2:03:57
Speaker 1: with direct reward stimulation.
2:03:59
Speaker 2: And you think we've sort of been primed for that
2:04:01
Speaker 2: because we're getting this very minor dopamine hit with likes
2:04:05
Speaker 2: on Instagram and Twitter. And we're completely addicted to that.
2:04:09
Speaker 2: And it's so innocuous. It's like so minor. And yet
2:04:13
Speaker 2: that overwhelms most people's existence. Imagine something that provides like
2:04:17
Speaker 2: an actual physical reaction where you actually orgasm. You actually
2:04:22
Speaker 2: do feel great. You have incredible euphoria. Forget delayed gratification.
2:04:28
Speaker 2: That's out the door.
2:04:29
Speaker 1: You can't compete with that. I think there was recently
2:04:32
Speaker 1: a new social network where they have bots going around
2:04:35
Speaker 1: liking things and commenting how great you are in your
2:04:38
Speaker 1: posts just to create pure pleasure sensation of using it.
2:04:42
Speaker 2: Oh, boy. Jesus. Do you saw that study of the
2:04:48
Speaker 2: University of Zurich where they did a study on Facebook
2:04:52
Speaker 2: where they had bots that were designed to change people's
2:04:59
Speaker 2: opinions and to interact with these people and their specific
2:05:03
Speaker 2: stated goal was just to change people's opinions?
2:05:06
Speaker 1: I think Facebook did that.
2:05:07
Speaker 2: Yeah. Facebook did it.
2:05:08
Speaker 1: Yeah.
2:05:09
Speaker 2: But the University of Zurich, was that a Reddit thing? Yeah,
2:05:11
Speaker 2: it was on a Reddit subreddit. Yeah, yeah. And they
2:05:16
Speaker 2: just experimented with humans, and it was incredibly effective.
2:05:22
Speaker 1: Your systems know you better than you know yourself.
2:05:25
Speaker 2: Right.
2:05:25
Speaker 1: They can predict what you're going to be into in
2:05:29
Speaker 1: terms of preferences. They can know social interactions you would enjoy. Oh,
2:05:34
Speaker 1: this person should be your friend. And in a way,
2:05:37
Speaker 1: they can behaviorally drift you. So you're on a dating site,
2:05:41
Speaker 1: and the set of options they present to you, that's
2:05:44
Speaker 1: all you see. You don't know what else is out there.
2:05:46
Speaker 1: So after so many selections, they can change what the
2:05:50
Speaker 1: children will look like.
2:05:52
Speaker 2: Like the movie Ex Machina.
2:05:54
Speaker 1: Great movie.
2:05:55
Speaker 2: Fucking love that movie. But he designed that bot, that robot,
2:06:00
Speaker 2: was specifically around this guy's porn preferences.
2:06:05
Speaker 1: Yeah.
2:06:06
Speaker 2: And then you're so vulnerable. Yeah. Boy, Roman, you're freaking
2:06:12
Speaker 2: me out. I came into this conversation wondering how I'd
2:06:16
Speaker 2: feel at the end, whether I'd feel optimistic or not.
2:06:21
Speaker 2: And I don't, I just feel like this is just
2:06:26
Speaker 2: something where I think we're in a wave. It's headed
2:06:30
Speaker 2: to the rocks and we recognize that it's headed to
2:06:33
Speaker 2: the rocks, but I don't, think there's much we can
2:06:35
Speaker 2: do about this. What do you think could be done
2:06:38
Speaker 2: about this?
2:06:39
Speaker 1: Again, as long as we're still alive, we're still in control,
2:06:42
Speaker 1: I think it's not too late. It may be hard,
2:06:45
Speaker 1: may be very difficult, but I think personal self-interest should
2:06:49
Speaker 1: help us. A lot of the leaders of large AI
2:06:52
Speaker 1: labs are very rich, very young, they have their whole
2:06:56
Speaker 1: lives ahead of them. If there is an agreement between
2:06:59
Speaker 1: all of them not to push the button, not to
2:07:01
Speaker 1: sacrifice a Next 40 years of life they have guaranteed
2:07:05
Speaker 1: as billionaires, which is not bad. They can slow down.
2:07:09
Speaker 1: I support everyone trying everything from governance, passing laws that
2:07:14
Speaker 1: siphons money from compute to lawyers, government involvement in any way,
2:07:20
Speaker 1: limiting compute, individuals educating themselves, protesting by contacting new politicians.
2:07:28
Speaker 1: Basically anything, because we are kind of running out of
2:07:32
Speaker 1: time and out of ideas. So if you think you
2:07:34
Speaker 1: can come up with a way to prevent super intelligence
2:07:38
Speaker 1: from coming into existence, you should probably try that.
2:07:41
Speaker 2: But again, the counter argument to that is that if
2:07:44
Speaker 2: we don't do it, China's gonna do it.
2:07:47
Speaker 1: And the counter argument to that is it doesn't matter
2:07:49
Speaker 1: who creates super intelligence, humanity is screwed either way.
2:07:53
Speaker 2: And do you think that other countries would be open
2:07:56
Speaker 2: to these ideas? Do you think that China would be
2:07:59
Speaker 2: willing to entertain these ideas and recognize that this is
2:08:02
Speaker 2: in their own self-interest also to put the brakes on this?
2:08:05
Speaker 1: Chinese government is not like ours in that they are
2:08:08
Speaker 1: usually scientists and engineers. They have good understanding of those technologies.
2:08:12
Speaker 1: And I think there are dialogues between American and Chinese
2:08:15
Speaker 1: scientists where scientists kind of agree that this is very dangerous.
2:08:19
Speaker 1: If they feel threatened by us developing this as soon
2:08:22
Speaker 1: as possible and using it for military advantage, they also
2:08:25
Speaker 1: have no choice but to compete. But if we can
2:08:29
Speaker 1: make them feel safe in that we are not trying
2:08:31
Speaker 1: to do that, we're not trying to create super intelligence
2:08:34
Speaker 1: to take over, they can also slow down and we
2:08:37
Speaker 1: can benefit from this technology, get abundance, get free resources,
2:08:44
Speaker 1: solve illnesses, mortality, really have a near-utopian existence without endangering everyone.
2:08:51
Speaker 2: So this is that 0.0001% chance that you think we
2:08:56
Speaker 2: have of getting out of this?
2:08:58
Speaker 1: That's actually me being wrong about my proofs.
2:09:01
Speaker 2: You're right, and you'd like to be wrong.
2:09:03
Speaker 1: I would love to be proven wrong. Just somebody publish
2:09:06
Speaker 1: a paper in Nature, this is how you control superintelligence,
2:09:09
Speaker 1: AI safety community reads it, loves it, agrees, they get
2:09:12
Speaker 1: a Nobel Prize, everyone wins.
2:09:16
Speaker 2: What do we have to do to make that a reality?
2:09:19
Speaker 1: I think there is nothing you can do for that proof.
2:09:21
Speaker 1: It's like saying, how do we build perpetual motion machine?
2:09:23
Speaker 1: And what we have is people trying to create better batteries,
2:09:27
Speaker 1: thicker wires, all sorts of things which are correlates of
2:09:30
Speaker 1: that design, but obviously don't solve the problem.
2:09:34
Speaker 2: And if this understanding of the dangers is made available
2:09:39
Speaker 2: to the general public, because I think right now there's
2:09:42
Speaker 2: a small percentage of people that are really terrified of AI.
2:09:45
Speaker 2: And the problem is the advancements are happening so quickly.
2:09:49
Speaker 2: By the time that everyone's aware of it, it'll be
2:09:51
Speaker 2: too late. What can we do other than have this conversation?
2:09:55
Speaker 2: What can we do to sort of accelerate people's understandings
2:09:58
Speaker 2: of what's at stake?
2:10:00
Speaker 1: I would listen to experts. We have literal founders of
2:10:04
Speaker 1: this field, people like Jeff Hinton, who is considered father
2:10:09
Speaker 1: of machine learning, grandfather, godfather. saying that this is exactly
2:10:14
Speaker 1: where we're heading to. He's very modest in his PDoom estimates, saying, oh,
2:10:19
Speaker 1: I don't know, it's 50-50. But people like that, we
2:10:23
Speaker 1: have Stuart Russell, we have, I'm trying to remember everyone
2:10:27
Speaker 1: who's working in this space, and there are quite a
2:10:29
Speaker 1: few people. I think you had Nick Bostrom on.
2:10:31
Speaker 2: Yes.
2:10:32
Speaker 1: There is Benchio, another Turing Award winner, who's also super concerned.
2:10:38
Speaker 1: We had a letter from, signed by, I think, 12,000 scientists,
2:10:42
Speaker 1: computer scientists, saying this is as dangerous as nuclear weapons.
2:10:46
Speaker 1: This is a state of the art. Nobody thinks that
2:10:50
Speaker 1: it's zero danger, that is, diversity and opinion, how bad
2:10:56
Speaker 1: it's going to get, but it's a very dangerous technology.
2:10:59
Speaker 1: We don't have guaranteed safety in place. It would make
2:11:02
Speaker 1: sense for everyone to slow down.
2:11:05
Speaker 2: Do you think that it could be viewed the same
2:11:07
Speaker 2: way we do view nuclear weapons and this mutually assured
2:11:10
Speaker 2: destruction idea would keep us from implementing it?
2:11:14
Speaker 1: In a way, yes, but also there is a significant difference.
2:11:17
Speaker 1: Nuclear weapons are still tools. A human has to decide
2:11:21
Speaker 1: to use them. That human can be profiled, blackmailed, killed.
2:11:26
Speaker 1: This is going to be an agent, independent agent, not
2:11:29
Speaker 1: something controlled by a human. So our standard tools will
2:11:33
Speaker 1: not apply.
2:11:38
Speaker 2: I think we've covered it. Anything else?
2:11:42
Speaker 1: No, but it'd be awesome if somebody set up a
2:11:44
Speaker 1: financial price for solving this problem. And it's kind of
2:11:48
Speaker 1: like with Bitcoin. If somebody can hack Bitcoin, there is
2:11:51
Speaker 1: a trillion dollars sitting there. The fact that no one
2:11:54
Speaker 1: claimed it tells me it's secure. If somebody can claim
2:11:58
Speaker 1: the price for developing a super intelligent safety mechanism, that
2:12:03
Speaker 1: would be wonderful. And if no one claims it, then
2:12:05
Speaker 1: maybe no one has a solution.
2:12:07
Speaker 2: How would you do that? How would you set something
2:12:09
Speaker 2: like that up?
2:12:10
Speaker 1: Well, we need someone with some funds, propose an amount
2:12:13
Speaker 1: and say, this is what we're looking for. It's very
2:12:16
Speaker 1: hard to judge if it's actual solution, but there are
2:12:18
Speaker 1: correlates of good science. So maybe publish in a top journal.
2:12:22
Speaker 1: It survives peer review. It survives, you know, evaluation by
2:12:27
Speaker 1: top 30 experts. You can have things and everyone kind
2:12:31
Speaker 1: of agrees that, yeah, you kind of got it. Okay.
2:12:37
Speaker 2: Until now, educate yourself, people. AI, unexplainable, unpredictable, uncontrollable. It's
2:12:43
Speaker 2: available now. Did you do an audiobook?
2:12:46
Speaker 1: They are still working on it a year later.
2:12:48
Speaker 2: Still working on it?
2:12:49
Speaker 1: I don't know what it is. I would think AI
2:12:50
Speaker 1: would just read it out in 20 minutes.
2:12:52
Speaker 2: Why don't they just do it in your voice with AI?
2:12:55
Speaker 1: I agree with you completely. It took, I think, first
2:12:58
Speaker 1: version of my book they wanted to translate into Chinese.
2:13:01
Speaker 1: Five years later, they told me they will not do it.
2:13:04
Speaker 1: Five years into the translation. So they had a second
2:13:08
Speaker 1: Chinese translation started.
2:13:09
Speaker 2: Why didn't they do it?
2:13:11
Speaker 1: Publishing world is still living in, like, 1800s. Then you
2:13:15
Speaker 1: cite books. You know you have to actually cite the
2:13:18
Speaker 1: city the book is published in because that's the only
2:13:20
Speaker 1: way to find a book on the Internet.
2:13:22
Speaker 2: What do you mean?
2:13:23
Speaker 1: Like, if somebody wants to cite my book, it's not
2:13:26
Speaker 1: just enough to have a title and my name. They
2:13:28
Speaker 1: have to say in what city in the world it
2:13:30
Speaker 1: was published.
2:13:31
Speaker 2: What? Yes.
2:13:32
Speaker 1: Really?
2:13:33
Speaker 2: Yeah. That's archaic.
2:13:35
Speaker 1: The whole system is archaic.
2:13:38
Speaker 2: Wow. But yet you still used it.
2:13:40
Speaker 1: What choice do we have?
2:13:41
Speaker 2: Digitally publish? You could put it on Amazon?
2:13:44
Speaker 1: It's like, steal this book, download the PDF. I don't care. Like,
2:13:48
Speaker 1: please do it. Somebody should read it. That would help.
2:13:51
Speaker 2: Yeah. Well, more people need to read it and more
2:13:53
Speaker 2: people need to listen to you. And I urge people
2:13:56
Speaker 2: to listen to this podcast and also the one that
2:13:58
Speaker 2: you did with Lex, which I thought was fascinating, which
2:14:00
Speaker 2: scared the shit out of me, which is why we
2:14:02
Speaker 2: have this one. Thank you, Roman. Appreciate you.
2:14:05
Speaker 1: Thank you so much.
2:14:06
Speaker 2: I appreciate you sounding the alarm. I really hope it helps.
2:14:08
Speaker 1: So do I. Bye, everybody.