Based Camp - August 17, 2023


＂Mid is Over＂ How Do We Protect AI From Those it Will Replace？ (With Brian Chau)


Episode Stats

Length

37 minutes

Words per Minute

183.05891

Word Count

6,845

Sentence Count

424


Summary


Transcript

00:00:00.000 Of course, you see Sam Altman going to testify and he wants a licensing regime.
00:00:04.260 Of course, OpenAI will get a license.
00:00:05.900 Of course, OpenAI's competitors will not get a license.
00:00:08.420 Of course.
00:00:08.760 So if you're OpenAI, you know, and you've now secured billions of dollars in investment
00:00:14.900 in fundamentally these transformer models and, you know, the kind of hardware stacks
00:00:19.500 that are specific to replicating them.
00:00:22.240 And you get something that's out of left field or you get something that's limited
00:00:25.780 by a different factor, that's not limited by the advantages that OpenAI has over its competitors.
00:00:31.740 You know, maybe Foom is not the sound of AI accelerating,
00:00:34.520 but Foom is the sound of OpenAI's stock going to zero.
00:00:37.800 This is no longer a battle of, like, internet shitposters.
00:00:41.020 You know, this is a battle of real political interests
00:00:44.800 and the forces that drive the Democratic and Republican parties.
00:00:49.560 I also love that it's done such a good job of getting rid of these useless grifters,
00:00:54.280 like, artists and writers, because honestly, they had these pointless degrees.
00:01:02.160 They were a huge, I think, cause of our society's degradation.
00:01:07.380 And I'm so glad that AI has replaced them.
00:01:10.760 There's, okay, like, there's a real thing underlying that is that, like, mid is over, right?
00:01:16.660 And here, like, actually mid, I don't mean...
00:01:18.600 Would you like to know more?
00:01:19.520 Hello, everyone.
00:01:21.140 We are very excited to welcome Brian Chow for this episode of Basecamp.
00:01:25.400 We are going to be talking with Brian, not about his amazing podcast from the new world,
00:01:30.240 but rather a new venture of his, which is called the Alliance for the Future.
00:01:35.160 So I think that the number one thing to understand is that it's a non-trivial question,
00:01:42.440 whether you are able to use machine learning,
00:01:45.280 whether you are able to write machine learning, whether it will be illegal to log on to, you
00:01:49.400 know, chat.openai.com.
00:01:51.780 That is a real question that people have different answers to, that many people want for various
00:01:59.940 reasons to ban machine learning pretty much wholesale.
00:02:04.480 And this is, you know, the purpose of Alliance for the Future is to do everything we can to
00:02:11.440 make sure that does not happen.
00:02:12.460 Hmm.
00:02:14.200 So what are you saying to the AI apocalypticists who start immediately screaming at you and
00:02:21.700 flailing when you say things like this?
00:02:23.920 What's your take on LEIs or Yukowsky's positions and how do you respond to them?
00:02:28.300 Yeah, it's very funny.
00:02:29.420 This is very often the first question that people ask me when I mention this.
00:02:33.680 And they're, you know, striking the rogue data centers is among the, I think the less, you
00:02:42.660 know, the less dangerous versions of what people are trying to get.
00:02:47.800 Like this is something that maybe, maybe, I don't know, like there are people in, in
00:02:52.800 AFTF, in Alliance for the Future that I'm more worried about the EAs.
00:02:56.180 You know, I'm friends with many EAs and I've also seen the kind of regulatory environment
00:03:02.420 in, in Washington, in other countries, in the EU, you know, like the EU, the EU commission
00:03:09.520 are not EAs.
00:03:10.960 You know, the Chinese government is not EAs.
00:03:12.940 Chuck Schumer is not an EA.
00:03:14.620 Right.
00:03:14.760 I just don't think they're the, I just don't think they're the major threat.
00:03:18.380 They're like the closest thing to like a non-retarded version of the ban AI argument.
00:03:24.120 And that's why they get engaged with in like smart parts of Twitter.
00:03:27.360 It's because they're kind of like the, the, the smart representatives of this much bigger
00:03:33.500 target of people that I think most EAs would consider like stupid and misguided.
00:03:39.540 So, so I don't, I don't worry too much about the EAs.
00:03:42.180 I don't want this to be a conflict between, you know, like EAs and like accelerationists
00:03:48.960 or whatever.
00:03:49.560 I think that that's, you know, that's very silly.
00:03:51.820 That's getting caught up in like what I'm asking is what, so, you know, even our audience,
00:03:59.020 you know, we've got, I suspect some people in our audience who think AIs are going to
00:04:02.500 kill us all.
00:04:03.580 A lot.
00:04:04.600 Not a lot because we've taken very hard stances against that.
00:04:07.960 Well, very, so I don't know if you know our position on AI.
00:04:11.880 We take a position about variable AI risk, which is very different than absolute AI risk.
00:04:16.880 But I'm wondering, what is your stance?
00:04:18.840 Like if you're trying to calm somebody down, who's like, yeah, but it looks like genuinely,
00:04:24.560 I don't see how we remove this threat without restricting access.
00:04:30.200 How do you communicate to them?
00:04:32.700 So there's actually a difference between like Eliezer and like fans of Eliezer.
00:04:36.760 With fans of Eliezer, I just worry that they're driven too much by the current hype cycle.
00:04:42.060 Then, you know, they saw that it's cooling down a bit.
00:04:44.360 They saw, you know, like GPT-3 get released and then they saw GPT-4 get released.
00:04:49.260 And they were like, what if it just keeps developing at the same rate that between GPT-3 and GPT-4
00:04:54.320 was released?
00:04:54.960 And of course, that did not happen.
00:04:57.180 That was, you know, that was the accumulation of lots of work and, you know, the effort across the entirety of OpenAI's existence.
00:05:07.640 And there are many technological factors that just slow development.
00:05:12.720 This is the trend you see is that you see, you know, like an accelerating technology.
00:05:17.560 You see it get adopted in early market versions.
00:05:20.840 And then you start to see that the speed of that development petering out.
00:05:24.740 I actually have a more specific argument going through all of the, so far I have the hardware part done.
00:05:31.400 This has been put on the back shelf a bit because of everything else I'm doing right now.
00:05:35.520 But I do think the first part of the article actually holds up very well in the past few months,
00:05:42.820 which is diminishing returns on machine learning one.
00:05:46.380 You can find that also at fromthenew.world, fromthenewworld.
00:05:50.460 It is the home of everything now, everything that I do.
00:05:54.100 And yeah, so to those people, I would say you are severely estimating the progress in artificial intelligence.
00:06:02.720 Many of them have, you know, they have this like, I did not come up with this word.
00:06:07.620 They came up with this word of like foom, which is basically like the idea that, you know, it goes foom.
00:06:16.380 It sounds like, this sounds like, this sounds like a straw man version of the argument.
00:06:22.460 But it is, it is actually their own argument that, you know, once it, once it starts accelerating,
00:06:27.480 it will just become, you know, extremely fast.
00:06:30.260 It will only get faster and faster.
00:06:31.900 It's like, this, this is just not, this is, you know, this is hype cycle fantasy.
00:06:37.260 In terms of like the longer term.
00:06:40.020 In terms of the longer term.
00:06:41.220 Explain why it's hype cycle fantasy.
00:06:42.700 And I, like, I, this is the main thing.
00:06:45.180 Look at, yeah, it's, it's not the mainstream view.
00:06:48.220 Like, like, maybe it's like, maybe it's the view of my, like, Twitter posters.
00:06:51.680 No, no, I mean, I'm among, like, EA Zoomers, I know.
00:06:54.480 Like, they, they genuinely, and I, I don't mean to be, we've tried really hard not to insult
00:07:00.680 specific people on this show.
00:07:02.860 But Eliezer Yukowski is the Greta Thornburg of AI apocalypticism.
00:07:07.320 No, no, you have to, you have to have the Straussian reading of Eliezer, right?
00:07:12.640 So, so the Straussian reading of Eliezer.
00:07:14.920 So, so like the surface level reading of Eliezer is that, like, he, he just wants to
00:07:18.540 create, you know, massive panic and for government to take control of everything.
00:07:23.680 You know, that, that's what he writes in his Time article pretty much.
00:07:26.780 It is.
00:07:27.040 Like, this is, this is like not even really an exaggeration.
00:07:29.720 So, so like the Straussian reading of Eliezer is, you know, like he, he doesn't actually
00:07:33.720 think that the world is like 90% that, that like the, that like the world is like absolutely
00:07:38.480 doomed, but he thinks that like no one will take it seriously, right?
00:07:41.880 He, he thinks that unless you, unless you turn the volume, everyone's wearing earmuffs,
00:07:46.240 so you better turn the volume up to a hundred and you better give, you know, and this is by
00:07:50.860 the way, like not necessarily an incorrect model of how, how the government works.
00:07:55.300 Like, this is what Fauci did as well.
00:07:57.260 You know, the, the reasoning was, you know, that you wouldn't give us power to do any
00:08:01.440 pandemic prevention measures unless we took very extreme positions on how dangerous the
00:08:08.100 virus was, on how confident we were about certain interventions.
00:08:12.780 And, you know, in terms of the politics, Fauci was correct.
00:08:18.040 Like that was actually correct.
00:08:20.440 He, he got the power.
00:08:21.700 It's very likely that he would not have gotten the power if he made, you know, like a
00:08:25.220 moderate case for the risks of COVID.
00:08:27.260 And I think Eliezer has learned that lesson.
00:08:29.660 So I don't know, like, like maybe, so like, I think Eliezer is like, not as, not as, not
00:08:36.720 as insane as, as it sounds.
00:08:38.460 Like the things that he advocates for are like truly insane.
00:08:40.980 But like, I think like Eliezer, the man, you know, is not necessarily that insane.
00:08:45.380 So your argument is just, he's, he's wildly overcorrecting.
00:08:49.720 And that in reality, AI is not going to accelerate as quickly as everyone thinks it is.
00:08:54.760 And therefore, as AI more linearly develops, people can develop safeguards as necessary.
00:09:03.820 Therefore, it doesn't present an existential risk.
00:09:06.000 And therefore we shouldn't be, be stifling its development with regulation and rules.
00:09:12.540 Is that correct?
00:09:14.100 Yes.
00:09:14.420 So it's mostly correct.
00:09:16.920 So, and I should say here, I'm speaking for myself, not for like the AF, for AFTF in
00:09:21.960 general, but so like the original version of the, the name of the original version of
00:09:29.300 EA that cared about AI was that long-termism.
00:09:32.260 And the reason why it was called long-termism is, you know, the idea is that in hundreds
00:09:38.980 of years, humanity, if you just look at, you know, population growth or like, I don't
00:09:43.820 know, you guys have a much more, much more pessimistic version of this, right?
00:09:47.280 But if we create a way, if we create a way to solve our population growth problems and
00:09:52.180 have the earth continue growing, then the number of humans in the future are just much
00:09:55.860 more than the number of humans in the present.
00:09:57.740 So you have to care about the long-term.
00:09:59.140 And in terms of the long-term, in terms of the timeframe of like hundreds of years, I
00:10:04.120 am not, you know, I'm not completely sure that AI will not be a problem in like a hundred
00:10:08.480 years.
00:10:09.080 That is something, you know, that, that I accept as, as a possibility.
00:10:13.760 The question is, you know, if you have a pace of technological growth, that is, you know,
00:10:20.860 what you would infer from every other technology that has ever existed pretty much from, you
00:10:27.940 know, like the history of technological development from, you know, early science to the industrial
00:10:32.320 revolution, to more recent cycles, the 2000 cycle, you know, the, the very recent, you
00:10:37.680 know, like 2020 to 2021 cycle of technologies.
00:10:41.320 You get that, you, you get this very known thing, very well-known thing called an S-curve.
00:10:45.300 You get an early acceleration and then it peters out.
00:10:47.840 And then that's where the hard work has to be done.
00:10:49.400 Actually getting the technology adopted in all these sectors of society.
00:10:53.360 And this is also like the mainstream economist view.
00:10:56.320 And, you know, I'm not like this, like the thing is, this is the mainstream view, right?
00:11:00.980 The mainstream view is that, you know, hype cycles happen, that techno technological progress
00:11:05.180 is good, but it is not necessarily something that should be, you know, that you should go
00:11:11.140 all in on, right?
00:11:11.880 You should, you know, you should invest in some tech stocks, but you should not, you know,
00:11:15.180 gamble your entire life savings on one tech stock and so on and so forth.
00:11:19.500 So I really do think this is the implementation of the normal, the normie, like non-hyper online
00:11:28.140 position on technology as policy.
00:11:31.000 That's how I would put it.
00:11:33.140 So functionally, your organization, if it succeeds, what's it doing?
00:11:39.300 What specific changes does it make in policy?
00:11:42.140 How do you achieve those?
00:11:43.640 What research are you outputting?
00:11:46.600 So Alliance for the Future is a completely new think tank.
00:11:49.760 The number one thing is just to balance the scales, because right now there is a lot of
00:11:55.140 funding either from the EA side, although I don't think that's the main problem, but even
00:11:59.520 more from existing political interests.
00:12:02.440 And of course, you see Sam Altman going to testify and he wants a licensing regime.
00:12:07.680 Of course, OpenAI will get a license.
00:12:09.320 Of course, OpenAI's competitors will not get a license.
00:12:11.680 Of course, you know, increase the barriers to entry.
00:12:15.140 Actually, so this is a thing that you're talking about briefly here, but I really want our audience
00:12:19.520 to understand this.
00:12:20.700 So in the business world, you know, what you want to do to maximize the value of your company
00:12:26.560 is to advocate for regulation.
00:12:29.060 Like a lot of people are really surprised that like Google would advocate for like internet
00:12:32.960 search regulation or like, but this is what you see with any large monopoly in a space
00:12:38.680 is they spend a huge chunk of their revenue advocating for regulation of their own company.
00:12:45.120 And the reason they're doing that is because it prevents new entrants from entering the market,
00:12:49.760 which protects their monopoly.
00:12:52.780 That is why people like Sam Altman are advocating for this regulation.
00:12:56.240 It's not because they're genuinely scared of AI.
00:12:59.300 It's because they're the first players on the market to continue.
00:13:02.160 This is something, this is actually a very important topic.
00:13:04.880 Okay.
00:13:05.160 Like this is a good venue to be like very autistic about this.
00:13:08.260 So there's, there is this, I think like he published most of his stuff in the 60s, 70s.
00:13:16.180 Economist, one of the, considered one of like the founding people of political economy,
00:13:20.920 Gordon Tullock.
00:13:21.720 Okay.
00:13:22.160 All of the GMU people love this guy because he was from GMU, I think.
00:13:25.880 And he has this idea called the Tullock rectangle.
00:13:29.820 And that idea is that, okay, you look at, you know, if you've looked at a supply and demand
00:13:34.600 curve, when there's regulations that interfere with the supply and demand curve, it can increase
00:13:38.200 the profit of, of an industry that has being interfered with because it stops.
00:13:43.740 And essentially like the number of new, new customers or like the, the amount of increased
00:13:48.120 profit from these regulations outweighs the amount of, the amount that you lose from missing
00:13:52.900 out on some traits.
00:13:54.560 The way that I really want to see this expanded is with firm dilution and not only with firm
00:13:59.860 dilution, but with dilution of the entire industry.
00:14:04.460 So, so, so what do I mean by this?
00:14:05.800 In an industry like machine learning, you have basically, you, you have a precedent that people
00:14:11.060 are not sure is optimal.
00:14:12.860 So you have, you know, right now we have transformers pretty much.
00:14:16.060 We have this paper from 2017.
00:14:17.880 There have been minor modifications to it.
00:14:19.820 The paper is called attention is all you need.
00:14:21.740 And this basically outlined the way in which all language models and many similar models
00:14:28.280 operate.
00:14:29.700 And we don't really like, but it's the best we got so far, but it's not like, it's not
00:14:36.940 like a mathematical proof.
00:14:38.300 It's not like a certain thing.
00:14:39.740 You know, we have no idea if this is the best way to do machine learning.
00:14:44.580 And in fact, many of the people who are more hyped about AI think that, you know, we're
00:14:47.680 just about to get another breakthrough in how we do, you know, how we do machine learning.
00:14:53.000 And so if you're open AI, you know, and you've now secured billions of dollars in investments
00:14:59.960 in, in fundamentally these transformer models and, you know, the kind of hardware stacks that
00:15:04.740 are specific to replicating them.
00:15:06.980 And you get something that's out of left field, or you get something that's limited
00:15:10.820 by a different factor.
00:15:11.840 That's not limited by the advantages that open AI has over its, its competitors.
00:15:16.480 You know, maybe foom is not the sound of AI accelerating, but foom is the sound of open
00:15:21.040 AI stock going to zero.
00:15:22.920 To this topic, just to shout out for listeners, cause I have a little request from a friend
00:15:28.600 here.
00:15:29.000 I know of this company that's found a way to create like much better using your own
00:15:34.220 models, like, like much tighter chip sets.
00:15:37.180 Anybody who's interested in investing in like a large fab, like, like, look, I'm talking,
00:15:44.020 you know, many millions of dollars, but it could make AI much less expensive to run.
00:15:48.820 I've got a company that's interested in doing that right now.
00:15:51.380 They've already developed all this stuff.
00:15:52.860 They're just looking for whoever they work with on the space.
00:15:54.860 Oh, this is fascinating.
00:15:56.460 So this is like not an open, open AI competitor, but like a TSMC competitor or something like
00:16:00.680 that.
00:16:00.820 Yeah.
00:16:02.100 Okay.
00:16:02.700 Interesting.
00:16:03.220 Interesting.
00:16:04.340 Another thing I wanted to dive into is before we started recording, you had alluded to this
00:16:08.360 not being really an EA style think tank.
00:16:12.760 And I thought that was interesting because you are trying to be effective in your altruism
00:16:18.340 through doing this, right?
00:16:19.980 Like I'm so, so tell me more.
00:16:22.020 Why do you think this is not an EA thing?
00:16:24.860 So like on the, on the center for effective altruism homepage, they have like this essay
00:16:29.540 about like, who is, who is an effective altruist.
00:16:33.020 And they basically say like everyone who wants the world to be better and who likes evidence.
00:16:38.540 And I'm like, okay, you know, I guess it's an EA think tank.
00:16:42.240 Sure.
00:16:42.540 Give us money.
00:16:43.880 Of course, you know, most EAs, you know, most EAs want more, more regulation in the, even
00:16:49.800 in the short term, or actually I'm not sure about like the, the people who are, you
00:16:53.720 know, most funded by EAs.
00:16:54.980 That's what they want.
00:16:56.000 I should say as well, this is an article that might be out by then, but there's also a very
00:17:01.340 strong EA case.
00:17:02.880 Even if you think that there, I actually, I say this as well.
00:17:06.520 Like the more you think that AI is happening soon, the more it's the flight 93 election.
00:17:16.500 The more you should be trying to make sure that there is no government control of AI because
00:17:23.040 there is one institution in all of human history that is guaranteed to be misaligned.
00:17:28.860 And that is the most powerful government in the world.
00:17:32.180 Why is that institution always guaranteed to be misaligned?
00:17:35.380 The answer is political economy.
00:17:37.380 It goes back to Tulloch.
00:17:38.440 It goes back to many of the factors we were talking about, but people have the idea.
00:17:42.280 Oh, you know, in markets, there's externalities and in markets, people will compete.
00:17:46.300 And the thing that makes them succeed will not always be the thing that's best for the
00:17:49.120 general population.
00:17:50.820 And, you know, when it comes to elections, you know, the thing that makes people succeed
00:17:54.120 in elections, the thing that makes them get the most votes is always going to be the
00:17:57.700 thing that is, you know, most rational and sane.
00:18:00.200 And of course, it's not true.
00:18:01.540 And you go further, you go further down the level, right?
00:18:04.840 So there are these like nesting, there are these layers of the onion as you go down like
00:18:09.760 the policymaking stack and people don't transfer their lessons.
00:18:13.720 So I think many EAs, maybe they've read like Brian Kaplan, they've read, they might even
00:18:18.340 have read Garrett Jones.
00:18:19.400 And they understand the flaws of the voter.
00:18:23.120 They understand the flaws of, you know, when you go and cast your ballot for Donald Trump
00:18:27.620 and Joe Biden, that that's not necessarily the best thing that, you know, they're not
00:18:32.860 necessarily going to be doing the best thing for the world.
00:18:35.540 But you go one level down the stack to the legislative level or to the administrative level,
00:18:41.200 and they do not see the exact same incentives in play.
00:18:44.580 Where when you pass a bill, for example, when you pass like, like, the authorization of the
00:18:49.820 use of military force, right?
00:18:51.760 Or when you pass like when you pass like the budget, the budget reconciliation bill, what
00:18:57.320 factors go into play in terms of getting your policy priority into that bill?
00:19:02.340 And what in what ways will be corrupted by the process by the requirements and the incentives
00:19:07.860 for it to be there in the first place?
00:19:09.760 And, you know, the short answer, this is something very funny that I posted about recently and
00:19:16.760 got a lot of support on Twitter for, is that like, there are some ideas that are so correct
00:19:21.060 that even the most strawman version of them is true.
00:19:24.380 So like the extreme strawman version of my opinion is like, or of this, like it really
00:19:30.140 like not my opinion originally, but like the field of political economy, the extreme strawman
00:19:35.440 of it is that like, whatever idea you want to get into law will be unrecognizably corrupted
00:19:40.760 once it is in law.
00:19:42.480 And even that like extreme strawman is like, pretty much true.
00:19:48.620 You just look at like,
00:19:50.020 I would call it extreme strawman.
00:19:51.660 You're like, this is what my opponent's like, but it's true.
00:19:53.920 Even the most insane of them.
00:19:56.160 Yeah.
00:19:56.460 Yeah.
00:19:56.840 That's, that's the best thing about it.
00:19:58.700 Right.
00:19:58.920 Like, like even, you know, even, you know, the most fervent supporters of law, you know,
00:20:04.160 you, you ask them, you know, you ask them like, what happens?
00:20:07.640 What do you think this law will look like after it's passed?
00:20:10.120 I'll say like, I, I don't know.
00:20:11.660 I just hope it's, this is also like a take that, that I want to, I want to like specifically
00:20:17.060 address.
00:20:17.680 If we're talking about addressing EA takes is that I've had multiple people that I really
00:20:22.680 respect say like, oh, we, we want the government to interfere because we'll just slow down
00:20:28.560 progress.
00:20:29.240 We just want it to like cause harm.
00:20:31.660 And as long as it's like, as long as it's like causing harm, it will reduce the probability
00:20:36.120 or the speed that we get AGI.
00:20:38.820 And this is also something that I don't necessarily think is true.
00:20:43.280 And the best counter example in recent memory is gain of function research.
00:20:47.600 Right.
00:20:48.060 So we have this phrase called like a narco tyranny.
00:20:50.680 Usually people talk about in the context of SF, which is like the bad actors are not punished
00:20:54.940 because they're outside of the system.
00:20:56.260 The only people who are punished are like the good actors in the system.
00:20:59.540 Right.
00:20:59.700 So, so you're, you're, you're punished for like protecting your convenience store, but
00:21:03.220 like the, the, the people who rob your convenience store are rewarded by the state.
00:21:07.520 And you see the exact same thing in regulatory capture.
00:21:10.460 You see the FDA, you know, very, very, you guys, you guys actually have had experience with
00:21:16.840 this, right.
00:21:17.320 With going after novel technologies that have some promise, but they will also, you know, this
00:21:22.780 is not the FDA, but the U S government will also fund gain of function research in Wuhan.
00:21:28.060 So it's, the question is not, you know, this like one singular lever of what it does to
00:21:33.460 the industry.
00:21:34.100 It is much more of the question of like how it influences the distribution of, of players.
00:21:42.120 And, you know, the most likely thing that will happen with the distribution of players
00:21:45.160 is that, Oh, actually it's just open AI, Google, Facebook, so on, or, or not really a so on.
00:21:53.140 It is really much just in this case, you know, in, in some regulatory capture cases, they're
00:21:58.960 morbid in this case, you know, it is, it is really just, you know, like the major named
00:22:03.200 players.
00:22:03.580 So let's say that you are successful with this beyond your wildest dreams, how will
00:22:09.640 the world be different?
00:22:10.660 Is it a scenario in which rather than there being like three players who are setting the
00:22:15.080 tone for AI, it's a little bit more distributed, like a lot of people are contributing, it's
00:22:19.180 more open source or, you know, what, what, what kind of environment or ecosystem are you
00:22:24.080 trying to create?
00:22:25.860 So many people, this is a very fascinating critique of political economy, because if you look at
00:22:31.560 these incentives, they, they rarely ever change.
00:22:33.680 So, so there are cases of just being like a doomer of people saying like, not, not an
00:22:38.020 AI doomer, but people like saying, Oh, like the regulatory crackdown is inevitable, but you
00:22:43.300 just look at the history of the United States and that's just not true.
00:22:48.060 The best example is you guys remember SOPA, like a stop or like something like online.
00:22:54.560 Yeah.
00:22:54.640 They were trying to stop the internet nonsense.
00:22:56.460 Oh my God, yes.
00:22:57.540 Yes.
00:22:57.840 Wow.
00:22:58.180 So like this, this was a case, I think it was the, it was, I forget, I forget which
00:23:05.220 agency there was, some, some agency was trying to pass, I forget the exact name of it, but
00:23:09.920 it may have been, okay, whatever.
00:23:12.960 But, but basically they were trying to do this restriction of content on the internet,
00:23:16.300 basically saying that, you know, like if any, if any like random commenter said, or like
00:23:21.560 linked to porn or, or something like that, then it would be, you know, then the entire
00:23:26.120 website would be subject to legal crackdowns.
00:23:29.220 Right.
00:23:29.660 And people, people just wrote in, people got, got media attention.
00:23:34.220 It became like, it became like this, this huge, like unification of the internet was,
00:23:39.060 was one way that it was described.
00:23:42.400 We are basically unanimously.
00:23:45.720 Everyone was like, this is a terrible idea.
00:23:47.040 Like, this is just, you know, this is just disastrous.
00:23:50.100 And we stopped it.
00:23:51.480 You know, we had, we had a W and there are instances where there are calls for regulating
00:23:59.980 an emerging industry.
00:24:02.140 And for one reason or another, it just doesn't happen.
00:24:06.420 You can look at the internet, for example, you know, this is in the broader history of
00:24:10.580 the internet.
00:24:11.500 Mark Andreessen actually gave me this example, right?
00:24:13.400 So Mark Andreessen, you know, like he, he talked about, he talked about this on my podcast.
00:24:18.040 He's been on several others, but it should be, it should be out by the time, you know,
00:24:21.440 this, this releases.
00:24:23.180 And he talked about, you know, like nowadays, like, what if you try to call, what if you try
00:24:27.640 to ban the internet now?
00:24:28.980 We've created a successful political constituency.
00:24:32.360 If your industry is big enough, which I think machine learning will be, you know, I think
00:24:36.540 even though we won't get, you know, artificial general intelligence, we will get, you know,
00:24:39.900 many commercial applications.
00:24:41.160 We have many commercial applications now, right?
00:24:43.620 And once all of those are adopted, once all of those are, you know, regular parts of people's
00:24:47.500 lives, if it's big enough, like the internet is, and like, I think machine learning will
00:24:51.640 be, then you've created a political constituency.
00:24:54.560 You know, all you have to do is wait for the existing adoption curve to happen.
00:24:58.740 So you're, you're, you're saying that you're trying to broker that transition to like get
00:25:03.680 to a place where adoption is wide enough, where like the market will handle the protection
00:25:08.360 because they are dependent on it and huge fans of machine learning.
00:25:11.640 Right. It's not even the market.
00:25:12.360 It's just, you know, like if half your country is using chat GPT or using some form of LLM,
00:25:17.280 you know, you're not banning it.
00:25:19.220 I'm sorry.
00:25:19.700 You know?
00:25:19.960 Yeah. Like no one will get reelected if they support any legislation that does that.
00:25:23.340 So basically before that saturation has been reached, you were trying to ensure that we
00:25:28.860 don't preemptively make that impossible.
00:25:31.260 Right.
00:25:31.320 Yeah. Now, now is the most important reason for exact, or now is the most important moment
00:25:35.540 for exactly that reason is because it's, it's the only, you know, it's the period of time
00:25:40.620 in which it is most politically vulnerable, but most economically has the most economic
00:25:45.760 potential.
00:25:47.560 Yeah. That's meaningful.
00:25:50.460 Are you concerned about all the people?
00:25:52.860 Because it sounds like you're not really fighting against EA-ers who are making weird
00:25:56.300 arguments against AI.
00:25:57.360 You're more fighting against legislators who are like, well, but, you know, they might
00:26:02.240 have much more, what you might say, like normie arguments against it.
00:26:05.200 Right. So they're going to say, well, what about the fact that AI is going to take jobs
00:26:08.820 away? You know, what are you going to say to them?
00:26:11.160 I want to take an interlude between this.
00:26:14.160 Most of the people who are actually proposing like these, like legislative crackdowns are very
00:26:20.140 anti-EA.
00:26:21.700 You know, they are anti-EA.
00:26:23.640 Like, like one of these people, like Timnit Gebru, who is like this race grifter who is
00:26:28.100 now like focused on machine learning is they like, or like she absolutely like despises
00:26:34.340 EA for like, honestly, like pretty pathological reasons.
00:26:39.740 Like it's not helpful for her for in any way to dislike EA, you know, like they could easily
00:26:45.700 be, you know, like legislative allies, but, but she, like, you know, but, but she just absolutely
00:26:51.200 despises them, you know, loves calling them racist because they believe in IQ and they
00:26:57.040 believe in like.
00:26:57.820 Oh, okay.
00:26:58.420 So that general approach.
00:26:59.900 Yeah.
00:27:00.360 Yeah.
00:27:00.640 Like a lot of the most despicable people, like also hate EA and like, like this is, you know,
00:27:08.980 it's no longer just, this is no longer a battle of like internet shit posters.
00:27:13.380 You know, this is a battle of real political interests and the forces that drive the democratic
00:27:20.520 and republican parties.
00:27:22.600 What's your rebuttal to their arguments though?
00:27:24.480 Because those aren't EA arguments.
00:27:26.000 So what is your rebuttal to, this is going to take away jobs.
00:27:29.540 It's, it's dangerous and we don't need it.
00:27:31.420 So why should we support it?
00:27:33.240 Et cetera.
00:27:33.820 The jobs point is, is particularly interesting because it's, it's kind of like framed as a
00:27:39.120 republican concern and it's, it's very, it's very funny.
00:27:44.540 It actually relates to the other discussion that, that we've had or the discussion in the
00:27:50.600 future that I, that I have some premonitions about.
00:27:54.800 Yeah.
00:27:55.660 In many cases, AI, the things that AI are replacing are kind of bullshit jobs.
00:28:02.040 There are things that people already dislike, you know, people do not, there's this wonderful
00:28:07.040 tweet by Sam Altman that says, you know, today I've had one person tell me that they've used
00:28:14.580 ChatGPT to expand their bullet points into a long corporate email.
00:28:19.520 And another person tell me that they've used ChatGPT to condense a long corporate email
00:28:24.440 into five bullet points.
00:28:26.300 It's the future of communication.
00:28:28.340 I also love that it's done such a good job of getting rid of these useless grifters, like
00:28:33.840 artists and writers, because honestly, they had these pointless degrees.
00:28:41.240 They were a huge, I think, cause of our society's degradation.
00:28:46.440 And I'm so glad that AI has replaced them.
00:28:49.800 There's okay.
00:28:50.740 Like there's a real thing underlying that is that like mid is over.
00:28:55.380 Right.
00:28:55.600 And here, like actually mid, I don't mean, you know.
00:28:57.880 Yeah, yeah, yeah, yeah.
00:28:57.900 It is over.
00:28:59.060 I like that way of putting it.
00:29:02.280 It's very funny because people portray like the most famous, people portray like Drake
00:29:12.020 being replaced by AI.
00:29:13.760 It's like, no, he has, he still has interesting things going on.
00:29:18.500 Or like Taylor Swift.
00:29:19.460 Look at Taylor Swift's like ticket sales, you know, like she's not worried about this.
00:29:22.620 The thing, like the top, like the people who are actually contributing to, you know, the
00:29:31.220 culture that we consume every single day, they're, they're just going to be fine.
00:29:36.860 And in fact, they're going, like the next generation, Sam Woods was on my podcast.
00:29:40.940 He had this wonderful line, which is your job's not going to be replaced by ChatGPT.
00:29:45.760 It's going to be replaced by someone using ChatGPT.
00:29:47.760 I think that, you know, in the future, like the meta, the meta of art will be very much.
00:29:54.120 And I think like the top artists today will be able to adapt.
00:29:58.160 They have that kind of like entrepreneurial focus.
00:30:00.420 Taylor Swift is once again, a great example.
00:30:02.840 Like, I think she'll really enjoy, you know, playing and her team will really enjoy using
00:30:07.540 the new tools to discover like the frontier of music.
00:30:10.800 But I think what you're discounting here.
00:30:12.220 That people never want to work.
00:30:13.600 She represents 1% of people, like the vast, vast majority of people have not done anything.
00:30:19.580 Grimes has adapted.
00:30:21.280 Yeah, exactly.
00:30:22.040 Exactly.
00:30:22.520 I think that you are overestimating the competence and the aggressiveness of this top, you know,
00:30:30.500 1% of society.
00:30:31.960 And that for a long time, they haven't been pushed out.
00:30:34.980 Like Quentin Tarantino, like digital cameras come along.
00:30:37.580 He's like, I'm not going to touch them.
00:30:38.740 Even still, he doesn't do them.
00:30:39.660 A lot of the, throughout most of our history, you've been able to get away with that kind
00:30:43.180 of BS, that kind of arrogance.
00:30:45.640 But I don't think you're going to be able to now.
00:30:48.740 Yeah, there is, you know, there is the innovator's dilemma.
00:30:52.060 I think you're right, actually.
00:30:53.440 I'll say more like, I think that non-zero of the current top artists will adapt.
00:30:59.040 But in terms of like one specific one, yeah, like, you know, I'm not 100% sure that Taylor
00:31:04.740 Swift will be, you know, will be the one who is like, yeah, who is taking up all of these
00:31:09.180 new tools.
00:31:10.220 But I think that, I think that it will be, you know, it will be a hybrid.
00:31:13.260 That I'm very confident about.
00:31:14.920 That's, you know, the mainstream of art, the mainstream of culture, the mainstream of
00:31:19.420 film.
00:31:20.420 Those will all, you know, it won't be, it won't be completely AI generated.
00:31:23.700 It won't be completely human generated, created from scratch.
00:31:27.820 It will be some combination of the two.
00:31:30.280 Just like, you know, we had digital cameras, people, you know, people are adapting to digital
00:31:33.900 cameras.
00:31:34.480 We had the internet, we had social media, we now have a mix, right?
00:31:38.020 You can think of Netflix as a mix between the original film model and YouTube, right?
00:31:43.560 Garrett Jones has this amazing term like spaghetti and spaghetti assimilation, which is he uses
00:31:49.700 this in the context of immigrants.
00:31:50.940 So like Italians come to America and like, or like they come to New York and they make
00:31:55.220 New York, you know, they become more like New Yorkers, but New York's culture also, you
00:32:00.480 know, they start eating pizza and they start eating spaghetti.
00:32:03.400 It becomes more Italian.
00:32:05.140 And I think the same is true of AI, you know, our current culture will become more like AI.
00:32:10.740 It will, there will be, you know, I think like Sam would put it best, you know, the replacement
00:32:16.220 of jobs is not really going to be, you know, like vertical.
00:32:19.540 It's not really going to be like people being completely replaced by AI.
00:32:23.100 It's going to be a new skillset.
00:32:25.440 People are learning to do something better.
00:32:28.180 And, you know, the people who will do that better are going to be the people who use AI.
00:32:33.700 I think though, that what's understated is that a huge portion of knowledge workers, and
00:32:40.700 in that I include people like sales, marketing, writers, artists, designers, salespeople,
00:32:47.780 like website people have been doing work that isn't actually used.
00:32:52.380 Like, I think we've had this period of inertia where people are still hiring and paying and
00:32:57.040 thinking that they need these people even before AI.
00:32:59.280 And like actually not using the vast majority of work they do, and that there are busloads
00:33:04.720 of graduating classes that believe that their job is to sit behind a computer and write strategy
00:33:09.760 documents and analyze things, but not actually build or create anything.
00:33:13.620 And that these people are going to get laid off.
00:33:17.400 They're getting laid off in droves.
00:33:18.700 They're not going to get rehired.
00:33:19.960 They're going to have to figure out their own way.
00:33:21.400 Do you think that those groups are capable of building new lives for themselves when
00:33:26.780 they've been conditioned to do something that is completely different?
00:33:33.560 Like, so this is an interesting question.
00:33:38.100 This is an interesting, like, economics question.
00:33:41.620 But I do want to, like, mess with the framing a little bit.
00:33:45.820 Go ahead.
00:33:46.200 I think, like, this is not that related to AI.
00:33:49.200 This is, like, related to low interest rates more than it is related to AI.
00:33:54.320 Like, it is kind of like, it is like these things happening in the same time.
00:33:59.060 But, like, you know, you could easily see a world where, you know, open AI just develops
00:34:04.700 or, like, all open AI and its competitors just develop and release everything, like, a few
00:34:09.900 years earlier.
00:34:10.780 And we have AI being released to broader society in, like, a, you know, user-friendly way.
00:34:17.960 At the same time as, like, the crypto hype bubble.
00:34:22.520 What would the vibes be around AI then?
00:34:25.480 You know?
00:34:25.960 If it's, like, if it's, like, the tail end of the lockdown, crypto stocks are going crazy,
00:34:31.700 you know?
00:34:32.160 And then open AI publishes, like, current level chat GPT.
00:34:36.840 What are the vibes then?
00:34:39.080 And I think it will be a vibe of, like, just much more optimism.
00:34:42.080 It won't be a vibe of, like, complaining.
00:34:43.860 It won't be a vibe of, like, you know.
00:34:45.600 And, of course, this is not really an argument for my position.
00:34:49.160 But it's an argument against, you know, I think it's an argument against much of the
00:34:53.060 contemporary.
00:34:54.140 This is, like, not the EAs, right?
00:34:55.460 I think the EAs would still be worried.
00:34:57.260 But for, like, the people worried about jobs, for the people who are worried about, you know,
00:35:01.660 like, automation, for the people who are worried about, you know, basically collapse, I think
00:35:06.180 that that's much more kind of absorbing the more general economic environment than it is an
00:35:10.660 actual concern about AI.
00:35:11.920 Well, what I like about your view is that it's actually quite optimistic, which is super
00:35:17.600 not Gen Z.
00:35:18.460 Like, I really like that.
00:35:19.320 It's eminently reasonable.
00:35:20.580 It's like, actually, this problem will solve itself.
00:35:23.140 We're going to have the critical mass of machine learning users, essentially, that are
00:35:26.460 going to make sure that it doesn't get, you know, walled off and made very difficult to
00:35:30.580 open source and, you know, collaboratively develop.
00:35:33.660 I'm just going to help to bridge the gap with this.
00:35:35.840 And I think that makes me super intrigued to see how it goes for the Alliance for the Future.
00:35:39.660 I'm really glad that you are in its foundational team and doing this work.
00:35:44.760 And I'm keen to see how it goes.
00:35:45.800 So to our listeners, if you're interested in this, check out the Alliance for the Future.
00:35:49.920 Check out Brian's podcast as well.
00:35:52.660 Yeah, affuture.org is how to find us.
00:35:54.660 We would really, really appreciate donations at this early stage.
00:35:59.560 And yeah, you can check out all of my writing.
00:36:01.640 You can check out my writing on AI specifically at pluralism.ai.
00:36:05.300 And you can check out all of it, including the podcast at fromthenew.world.
00:36:10.840 Thank you so much, Brian.
00:36:12.380 This is, it's always really fun to talk with you.
00:36:15.100 Awesome.
00:36:15.580 This was very fun.
00:36:16.940 It was not, you know, it wasn't four hours, but it was, you know.
00:36:20.560 See, we don't have attention spans for that.
00:36:22.560 We're like fast, but you know.
00:36:24.400 Yeah, yeah, yeah.
00:36:25.080 You guys were on the podcast for four hours.
00:36:29.320 So for people who want to see our podcast with him, we talked to him for like four hours each.
00:36:34.020 And I was hammered when I talked to him.
00:36:37.060 Completely hammered.
00:36:37.780 He got me at like 9 a.m. at night, going on until like 1 a.m. in the morning.
00:36:43.780 It was enjoyable.
00:36:45.880 It was more like, I think it was like 3 until 8 p.m.
00:36:48.780 But for Malcolm, whose day starts at 2 a.m., that is like extremely late.
00:36:54.000 No, no, no.
00:36:54.440 It was late.
00:36:55.580 No, I went past midnight, I think, was recording.
00:37:00.440 It went pretty late.
00:37:01.900 I don't think.
00:37:02.880 Yeah, I'm not sure if it was midnight.
00:37:05.080 In my brain, it was past midnight.
00:37:07.980 We'll see.
00:37:08.420 We'll see.
00:37:08.900 Okay, okay.
00:37:09.680 But our viewers, check it out.
00:37:11.580 He has a great show.
00:37:13.340 He really knows how to pull things out of people.
00:37:15.520 Yeah.
00:37:16.660 Yeah.
00:37:17.060 You listed amazing questions.
00:37:18.080 You're an amazing interviewer.
00:37:19.320 So yeah, check out From the New World, but also The Alliance for the Future.
00:37:22.780 Thanks again, Brian.