Latest / Razib Khan's Unsupervised Learning / Sam Hammond: I for one welcome our A.I. overlords
Transcript
- 0:00SPEAKER_00: This podcast is brought to you by the Albany Public Library Main Branch and the
- 0:04SPEAKER_00: generosity of listeners like you.
- 0:09SPEAKER_00: What is a podcast?
- 0:16SPEAKER_00: God, Daddy, these people talk as much as I do.
- 0:26SPEAKER_01: Razeeb Khan's Unsupervised Learning
- 0:28SPEAKER_01: If you have a sibling with autism,
- 0:30SPEAKER_01: your future child's risk for an autism diagnosis is increased by a factor of 2 to 3.5.
- 0:35SPEAKER_01: ORCID's whole genome embryo reports can help mitigate your child's risk by
- 0:38SPEAKER_01: screening for over 200 genetic variants definitively linked to autism and other
- 0:42SPEAKER_01: neurodevelopmental disorders.
- 0:44SPEAKER_01: Discuss your situation with a genetics expert.
- 0:47SPEAKER_01: Hey everybody,
- 0:47SPEAKER_01: this is Razeeb with the Unsupervised Learning Podcast,
- 0:50SPEAKER_01: and I am here today with Sam Hammond,
- 0:53SPEAKER_01: or Samuel Hammond as he is more formally.
- 0:56SPEAKER_01: He is a senior economist at, what is this, a new place that you're working at?
- 1:02SPEAKER_01: a Foundation for American Innovation, FAI, which, you know, D.C., it's an acronym place.
- 1:10SPEAKER_01: And before that, you were at Niskanen, and you were also at the Mercatus Center.
- 1:14SPEAKER_01: You have a master's from George Mason and also an undergrad degree,
- 1:19SPEAKER_01: I think a master from Carleton in Canada.
- 1:21SPEAKER_01: You are Canadian, for anyone who cares.
- 1:23SPEAKER_01: I don't think that's relevant to any of this stuff.
- 1:26SPEAKER_01: I'm glad you got out of there.
- 1:27SPEAKER_01: Glad you got out of there.
- 1:29SPEAKER_01: uh you know you're in the the u.s you are you are in the metropole in the empire uh
- 1:34SPEAKER_00: so sam canada has a better system of government but america has two-day shipping
- 1:39SPEAKER_01: but you really think canada has a better system of government are you serious yes
- 1:45SPEAKER_00: but i don't my home city still doesn't have uber so it is worth the trade what's
- 1:50SPEAKER_01: your home city halifax
- 1:53SPEAKER_01: Oh, yeah, you're from the Atlantic.
- 1:55SPEAKER_01: You're from the Atlantic provinces that kind of like got added in accidentally, right?
- 2:00SPEAKER_00: No, we were the OGs.
- 2:02SPEAKER_00: Well, what about Newfoundland?
- 2:03SPEAKER_00: Newfoundland was added in the 1930s.
- 2:05SPEAKER_00: It was actually called the Balfour Declaration, a different Balfour Declaration.
- 2:10SPEAKER_01: yeah nobody knows canadian history man we even like we barely barely like realize
- 2:15SPEAKER_01: canada exists except for like trudeau shenanigans you know what i'm saying i don't
- 2:19SPEAKER_00: know it's uh it's funny to share newfoundland was shortlisted for a jewish homeland
- 2:25SPEAKER_00: no known fact
- 2:27SPEAKER_01: Okay, that is super weird.
- 2:29SPEAKER_01: I think I've heard that somewhere and I think I've discarded it in the past and
- 2:32SPEAKER_01: I'll probably discard it again because Canada facts just,
- 2:35SPEAKER_01: they don't really stick with me.
- 2:37SPEAKER_01: I don't know why, you know, Canada's, it's just kind of there, you know, it fills up a part of the map.
- 2:43SPEAKER_01: And I mean, you know,
- 2:45SPEAKER_01: We've got to talk about AI and not Canada.
- 2:49SPEAKER_01: We've got to talk about the future, not the past.
- 2:51SPEAKER_01: Because Canada is a country of the past in some way.
- 2:54SPEAKER_01: With global warming, with anthropogenic climate change, whatever you want to call it,
- 2:59SPEAKER_01: Some parts of Canada are not looking that bad.
- 3:03SPEAKER_01: There's some opportunity here.
- 3:06SPEAKER_01: Okay, you know what?
- 3:07SPEAKER_01: I'm just going to ask you this.
- 3:08SPEAKER_01: It's totally not part of the... What is up with Canadian housing?
- 3:12SPEAKER_01: Do you guys have the same NIMBY stuff that we have?
- 3:17SPEAKER_00: Yeah, but in a bigger way.
- 3:19SPEAKER_00: There's multiple things.
- 3:20SPEAKER_00: The first thing is we have all the same NIMBYism.
- 3:22SPEAKER_00: That's a broader anglo-sphere phenomena, I think, originating from our property law and common law.
- 3:29SPEAKER_00: Um, then the other factor in Canada is we have much less sprawl.
- 3:33SPEAKER_00: So where there's problems in the U S in terms of density,
- 3:36SPEAKER_00: we have those problems,
- 3:37SPEAKER_00: but then we also don't have the sprawl that a lot of us cities have.
- 3:41SPEAKER_00: And then the third and final factor is really twofold.
- 3:44SPEAKER_00: One, our high rate of immigration in recent decades, recent years, and to all the Chinese money.
- 3:51SPEAKER_00: So, uh, you know, it's a store of wealth in China and people are getting their money out of China.
- 3:57SPEAKER_00: They,
- 3:58SPEAKER_00: buy up real estate in Canada.
- 4:01SPEAKER_01: Yeah,
- 4:01SPEAKER_01: just to explicate what you're talking about here,
- 4:04SPEAKER_01: Chinese,
- 4:05SPEAKER_01: I think it started with Hong Kong Chinese,
- 4:06SPEAKER_01: but I think it's just Chinese in general now.
- 4:08SPEAKER_01: They buy assets abroad.
- 4:11SPEAKER_01: One of those assets they buy abroad is Canadian housing,
- 4:15SPEAKER_01: houses starting in Vancouver,
- 4:16SPEAKER_01: but I'm assuming in Toronto and elsewhere now too.
- 4:19SPEAKER_01: And those houses are store of wealth.
- 4:21SPEAKER_01: And during the Bloomberg regime,
- 4:26SPEAKER_01: I know that in New York,
- 4:27SPEAKER_01: a lot of people would complain that the Russian oligarchs were doing that with New
- 4:30SPEAKER_01: York apartments.
- 4:31SPEAKER_01: So they would buy these apartments.
- 4:33SPEAKER_01: Nobody would live in them.
- 4:34SPEAKER_01: And so the frustrating thing is apartments, houses, ideally people live in them.
- 4:40SPEAKER_01: They're not just some sort of financial instrument to store your wealth.
- 4:44SPEAKER_01: So that's what angers people.
- 4:46SPEAKER_01: I alluded to –
- 4:48SPEAKER_01: The NIMBY issue, not in my backyard.
- 4:51SPEAKER_01: And it's basically just really difficult to build in some parts of the Anglosphere.
- 4:54SPEAKER_01: It seems like to be a disproportionately Anglosphere problem.
- 4:58SPEAKER_01: Britain is – I mean I need to do a podcast with someone who's British because I
- 5:04SPEAKER_01: know the Japanese have a tradition of suicide due to dishonor.
- 5:10SPEAKER_01: But something happened and the UK has –
- 5:12SPEAKER_01: has agreed to commit economic suicide over the last 10 years uh i don't actually
- 5:17SPEAKER_01: like i don't think it's like great i don't laugh at it it's just kind of sad but i
- 5:21SPEAKER_01: literally do shake my head every week reading about the weird things that they do
- 5:25SPEAKER_01: in the uk or say or you know like so for example they have a housing crisis they're
- 5:29SPEAKER_01: like well actually you know some academics and people in the government are saying
- 5:33SPEAKER_01: well you know what a lot of people just have too many rooms
- 5:37SPEAKER_01: it's really a problem of misallocation of rooms.
- 5:41SPEAKER_01: What are you trying to say?
- 5:42SPEAKER_01: The Ministry of Housing is going to resettle people into baby boomers' houses?
- 5:48SPEAKER_01: And so I think Sam and I both think that straightforward market-based solution, which is
- 5:55SPEAKER_00: Build houses.
- 5:57SPEAKER_00: Yeah, legalize it.
- 5:58SPEAKER_00: So what's going on in Vancouver now is the Squamish Reserve,
- 6:01SPEAKER_00: First Nations Reserve,
- 6:02SPEAKER_00: has sovereignty over their land and they're now upzoning it.
- 6:06SPEAKER_00: And all the white NIMBYs in Vancouver are complaining that it doesn't hew to their traditional ways.
- 6:13SPEAKER_00: But I say a skyscraper is a longhouse.
- 6:15SPEAKER_00: It's just put on its end.
- 6:17SPEAKER_01: Yeah.
- 6:18SPEAKER_01: I mean this is the neoliberal future.
- 6:20SPEAKER_01: This is the neoliberal future they took from us.
- 6:24SPEAKER_01: We have the engineering.
- 6:26SPEAKER_01: Asian cities have the engineering.
- 6:28SPEAKER_01: I think – I'm trying to get some housing people on this podcast actually because
- 6:33SPEAKER_01: housing is important for fertility of the future and whatever.
- 6:36SPEAKER_01: It's kind of stupid in a way because I think we know how to do it.
- 6:38SPEAKER_01: But, you know, people complain, oh, well, we don't want these Asian skyscrapers and everything like that.
- 6:42SPEAKER_01: But here in the U.S., you know, we have a lot of space.
- 6:45SPEAKER_01: You can have exurbs.
- 6:46SPEAKER_01: You can have skyscrapers.
- 6:48SPEAKER_01: You can do all sorts of things.
- 6:49SPEAKER_01: It doesn't – there's no tradeoff here.
- 6:51SPEAKER_01: No.
- 6:52SPEAKER_00: No.
- 6:53SPEAKER_00: And,
- 6:53SPEAKER_00: you know,
- 6:54SPEAKER_00: as we segue into talking about AI,
- 6:55SPEAKER_00: like if AI revolutionizes the economy,
- 6:58SPEAKER_00: the thing that will be scarce is still land.
- 7:00SPEAKER_01: Yeah, yeah.
- 7:01SPEAKER_01: And so we'll still have these prices.
- 7:03SPEAKER_01: So, okay, AI.
- 7:05SPEAKER_01: So I saw this.
- 7:05SPEAKER_01: So I know you, Sam, from just around.
- 7:09SPEAKER_01: We met in D.C.
- 7:10SPEAKER_01: before, so we move in the same circles.
- 7:14SPEAKER_01: And I saw this article.
- 7:16SPEAKER_01: This cleverly titled piece, 95 Theses on AI.
- 7:22SPEAKER_01: It's at your sub stack, second best.
- 7:24SPEAKER_01: It's kind of like you post pretty infrequently, but when you post, people tend to take notice.
- 7:29SPEAKER_01: 95 Theses obviously is an allusion to Martin Luther with his 95 Theses, right?
- 7:35SPEAKER_01: The Reformation,
- 7:36SPEAKER_01: which destroyed medieval Western civilization,
- 7:39SPEAKER_01: destroyed,
- 7:40SPEAKER_01: some people would say,
- 7:41SPEAKER_01: the unity of Christendom,
- 7:43SPEAKER_01: created the modern world,
- 7:45SPEAKER_01: unleashed you know diabolical degeneracy i whatever you believe a mixed bag in
- 7:50SPEAKER_01: other words yeah whatever you believe about the positives or negatives some shit
- 7:56SPEAKER_01: went down in the early 16th century and martin luther's 95 theses are an anchor
- 8:01SPEAKER_01: right and so uh in a way that should be pointing to people that you think this is a
- 8:06SPEAKER_01: big deal you think ai is you think ai in our time is a really big deal
- 8:11SPEAKER_00: Yes, absolutely.
- 8:12SPEAKER_00: I have a line in there saying the open question is whether AI is like the printing
- 8:16SPEAKER_00: press or photosynthesis,
- 8:17SPEAKER_00: right?
- 8:18SPEAKER_00: And your, your doomerism will vary based on that.
- 8:22SPEAKER_01: Well, I mean, I mean, that's incredible though, because the printing press is a big deal.
- 8:26SPEAKER_01: The floor,
- 8:27SPEAKER_01: like you set the floor very high,
- 8:29SPEAKER_01: like you set the floor at the sky and you set the ceiling in space or something.
- 8:35SPEAKER_01: I don't, I don't know how to explain it.
- 8:38SPEAKER_01: Yeah.
- 8:38SPEAKER_01: So just, just to allude, uh,
- 8:40SPEAKER_01: Just to,
- 8:41SPEAKER_01: again,
- 8:41SPEAKER_01: dig down what you're trying to say there,
- 8:43SPEAKER_01: the printing press transformed human culture over the last – over five centuries.
- 8:49SPEAKER_01: There is evidence – Patrick Weinberg has talked about this,
- 8:52SPEAKER_01: but there is evidence in economic history papers that where there was a higher
- 8:57SPEAKER_01: density of printing presses –
- 8:58SPEAKER_01: The Protestant Reformation was much more likely to break out and succeed in German cities.
- 9:04SPEAKER_01: So the Pridnik Press is a precondition for religious and cultural change.
- 9:08SPEAKER_01: I think we know that.
- 9:09SPEAKER_01: I think the internet is kind of an accelerated version of that.
- 9:11SPEAKER_01: Photosynthesis was an energy revolution, transforming basically –
- 9:18SPEAKER_01: increasing the ability of life to extract energy so much more.
- 9:24SPEAKER_01: We shift from anaerobic to aerobic and whatnot.
- 9:28SPEAKER_01: So coupled with oxidation or oxygen respiration,
- 9:32SPEAKER_01: we are able to transform energy from the sun into energy that powers us,
- 9:36SPEAKER_01: right?
- 9:37SPEAKER_01: And so this is a pretty big claim.
- 9:42SPEAKER_01: You know there are people who think this is just another hype cycle.
- 9:46SPEAKER_01: And what do you say to them?
- 9:47SPEAKER_00: You know, we've come out of 40 years of relative stagnation, right?
- 9:50SPEAKER_00: The whole great stagnation discourse.
- 9:53SPEAKER_00: And I think even older generations haven't really lived through periods of rapid technological change.
- 9:59SPEAKER_00: You have to reach back into history.
- 10:01SPEAKER_00: And a lot of the innovation we've had over the last 40 years and a lot of
- 10:04SPEAKER_00: innovation boosters have been boosting a kind of like pseudo Potemkin kind of innovation.
- 10:12SPEAKER_00: So we really haven't experienced that.
- 10:13SPEAKER_00: Maybe our great grandparents did.
- 10:15SPEAKER_00: and my grandparents are deceased,
- 10:16SPEAKER_00: maybe they saw my grandfather on my mother's side,
- 10:21SPEAKER_00: his house was the first on his street to get electricity.
- 10:24SPEAKER_00: Right.
- 10:24SPEAKER_00: So that's a big deal.
- 10:25SPEAKER_00: But really, since then, we've been in this kind of plateau, relative plateau.
- 10:29SPEAKER_00: You know,
- 10:30SPEAKER_00: we get modest improvements,
- 10:31SPEAKER_00: better,
- 10:32SPEAKER_00: better mousetraps,
- 10:33SPEAKER_00: some information technology that hasn't showed up in the GDP statistics.
- 10:37SPEAKER_00: Maybe this is sort of think of the think of the Internet era as the kind of bootloader to the AI era.
- 10:43SPEAKER_00: Right.
- 10:43SPEAKER_00: The kind of 40 years of tinkering and complementary investments that don't show up in GDP until they do.
- 10:50SPEAKER_00: Um,
- 10:50SPEAKER_00: but the people,
- 10:51SPEAKER_00: but people just haven't,
- 10:52SPEAKER_00: people are used to this being the normal and don't understand how weird actually
- 10:58SPEAKER_00: the world is today relative to 200 years ago and how unnatural and out of
- 11:02SPEAKER_00: equilibrium it is.
- 11:03SPEAKER_01: Yeah.
- 11:04SPEAKER_01: So, I mean, you're alluding to a lot of things here.
- 11:06SPEAKER_01: Um,
- 11:07SPEAKER_01: so basically the internet,
- 11:08SPEAKER_01: uh,
- 11:09SPEAKER_01: it seems like there was some productivity increase from the internet,
- 11:12SPEAKER_01: but not as much as people thought,
- 11:13SPEAKER_01: or at least maybe the statistics are not being,
- 11:15SPEAKER_01: you know,
- 11:15SPEAKER_01: and you could correct me here.
- 11:16SPEAKER_01: You're an economist.
- 11:17SPEAKER_01: Okay.
- 11:17SPEAKER_01: Uh,
- 11:18SPEAKER_01: the statistics are not corrected or collected appropriately,
- 11:21SPEAKER_01: or it's not factored in,
- 11:22SPEAKER_01: in terms of productivity gain.
- 11:23SPEAKER_01: Uh,
- 11:24SPEAKER_01: but you're saying here,
- 11:25SPEAKER_01: AI,
- 11:25SPEAKER_01: like,
- 11:26SPEAKER_01: you know,
- 11:27SPEAKER_01: if the internet is a boot,
- 11:28SPEAKER_01: bootloader,
- 11:28SPEAKER_01: so AI is the,
- 11:30SPEAKER_01: it's the,
- 11:31SPEAKER_01: it's the main show,
- 11:32SPEAKER_01: right.
- 11:32SPEAKER_01: Of, of, of this, of this period.
- 11:34SPEAKER_01: And so, um,
- 11:36SPEAKER_01: What did you think, like, a couple of years, like, where did, when did this, like, come into your mind?
- 11:40SPEAKER_01: Like, was it chat GPT?
- 11:41SPEAKER_01: Was it earlier?
- 11:42SPEAKER_00: No, it's one of the big reasons I moved to the US.
- 11:46SPEAKER_00: You know,
- 11:46SPEAKER_00: I went to George Mason at the Mercatus Center,
- 11:48SPEAKER_00: which was sort of the,
- 11:50SPEAKER_00: this strange ecosystem of,
- 11:52SPEAKER_00: like,
- 11:53SPEAKER_00: Austrian libertarians and cyber libertarian rationalists,
- 11:57SPEAKER_00: like Robin Hanson and folks like that.
- 12:00SPEAKER_00: So I've always been interested in that kind of,
- 12:02SPEAKER_00: they used to call themselves the extropians,
- 12:05SPEAKER_00: the less wrong,
- 12:06SPEAKER_00: overcoming bias community.
- 12:08SPEAKER_00: So I think probably my entry point was picking up Age of Spiritual Machines,
- 12:13SPEAKER_00: the Ray Kurzweil book,
- 12:14SPEAKER_00: when I was in middle school and reading that.
- 12:17SPEAKER_00: He had basically his extrapolations of
- 12:20SPEAKER_00: Moore's Law and other trends out to 2040.
- 12:24SPEAKER_00: And his marker for when we'd have human level AI was around 2029,
- 12:30SPEAKER_00: which is actually now where the current forecasts are converging.
- 12:33SPEAKER_01: Yeah,
- 12:33SPEAKER_01: so maybe I'm not keeping track of it,
- 12:36SPEAKER_01: but Ray's been surprisingly quiet from what I can tell right now.
- 12:40SPEAKER_01: I feel like he was everywhere between 2000 and 2010 and he's just kind of not said much.
- 12:49SPEAKER_00: He just did the Rogan podcast and he has a new book coming out called The Singularity Is Even Nearer.
- 12:55SPEAKER_00: So he's going to do a press tour on that.
- 12:59SPEAKER_00: He's taking like 80 pills a day to try to
- 13:02SPEAKER_01: Yes, he's been doing that for a while now.
- 13:06SPEAKER_00: Yeah, it's not going to be working.
- 13:09SPEAKER_01: I'm not going to make any comment on someone's mortality in their expected years, right?
- 13:17SPEAKER_01: So your first – so you have like – okay, these are like arrayed in sequence.
- 13:22SPEAKER_01: You have 95 of these assertions.
- 13:25SPEAKER_01: But they're bracketed into these kind of like categories or like these classes.
- 13:30SPEAKER_01: So your first one is oversight of AI labs is prudent.
- 13:36SPEAKER_01: I don't really know anything about, is there any oversight right now?
- 13:41SPEAKER_01: Is the oversight Sam Altman's conscience?
- 13:46SPEAKER_00: Pretty much.
- 13:47SPEAKER_00: I mean, there's some voluntary stuff.
- 13:49SPEAKER_00: So there's this Frontier AI Forum where basically Google,
- 13:54SPEAKER_00: Anthropic,
- 13:55SPEAKER_00: Microsoft,
- 13:56SPEAKER_00: and OpenAI have made a set of voluntary commitments.
- 13:59SPEAKER_00: But there's no, at this point, no formal insight into the labs other than
- 14:04SPEAKER_00: the White House executive order last year,
- 14:07SPEAKER_00: which created these compute thresholds that said,
- 14:10SPEAKER_00: if you train a model above these thresholds,
- 14:13SPEAKER_00: 10 to the 25 flops,
- 14:15SPEAKER_00: which for reference,
- 14:16SPEAKER_00: GTP4 is about 10 to the 24 flop scale model.
- 14:20SPEAKER_00: So nothing that we've seen to date has really crossed that threshold.
- 14:23SPEAKER_00: But when they do,
- 14:24SPEAKER_00: in principle,
- 14:24SPEAKER_00: they're supposed to disclose their safety testing and so forth to the Department of Commerce.
- 14:31SPEAKER_01: And I,
- 14:31SPEAKER_00: and I,
- 14:31SPEAKER_00: and I argue that that is good and smart and reasonable and,
- 14:35SPEAKER_00: um,
- 14:36SPEAKER_00: and should be differentiated from like AI regulation or something more comprehensive.
- 14:42SPEAKER_01: Well,
- 14:42SPEAKER_01: so,
- 14:42SPEAKER_01: um,
- 14:43SPEAKER_01: just,
- 14:44SPEAKER_01: you know,
- 14:44SPEAKER_01: contemporary relevance right now,
- 14:46SPEAKER_01: uh,
- 14:46SPEAKER_01: which I try to avoid on this podcast,
- 14:48SPEAKER_01: but you know,
- 14:48SPEAKER_01: as we're recording a bunch of people have exited,
- 14:50SPEAKER_01: uh,
- 14:51SPEAKER_01: Ilya and the other,
- 14:54SPEAKER_01: other people within open AI,
- 14:55SPEAKER_01: uh,
- 14:56SPEAKER_01: that take,
- 14:57SPEAKER_01: uh,
- 14:58SPEAKER_01: existential risk seriously, right?
- 15:00SPEAKER_01: So it doesn't seem that open AI is actually,
- 15:04SPEAKER_01: there's a lot of complaints from people that were there earlier or supported open
- 15:10SPEAKER_01: AI earlier back when Elon was involved that Greg Brockman and Sam Altman are being
- 15:15SPEAKER_01: irresponsible and they're basically just,
- 15:18SPEAKER_01: they don't care about the existential risk part at all.
- 15:21SPEAKER_01: They think it was either crazy or dumb and so, you know, there's a lot of, you know,
- 15:28SPEAKER_01: There's a lot of noise about that right now.
- 15:30SPEAKER_01: Mm-hmm.
- 15:32SPEAKER_01: Which is, like, in contrast to... So, you know, I think I mentioned this on the podcast before.
- 15:37SPEAKER_01: I asked why Google had not done this AI revolution.
- 15:41SPEAKER_01: You know, because OpenAI, which had CPC, transformed everything.
- 15:44SPEAKER_01: Well, why did Google do it?
- 15:46SPEAKER_01: Somebody at DeepMind,
- 15:47SPEAKER_01: an engineer...
- 15:48SPEAKER_01: And I know several engineers at DeepMind who've worked on Gemini and whatever.
- 15:53SPEAKER_01: They said that at Google...
- 15:55SPEAKER_01: they were too worried about existential risk type stuff or they were too worried about the consequences.
- 16:00SPEAKER_01: And so what happened was there was a demotivation on the part of the researchers to
- 16:05SPEAKER_01: ever do anything because nothing ever progressed.
- 16:09SPEAKER_01: They never approved anything because people were too afraid.
- 16:12SPEAKER_00: There might be something to that, but I also think Google is just a bloated company.
- 16:16SPEAKER_00: And they had their old X labs, for instance, where they were like,
- 16:20SPEAKER_00: building blimps that shoot down internet and stuff like that.
- 16:24SPEAKER_00: And a lot of those things just died in the vine because they were basically R&D
- 16:27SPEAKER_00: research efforts that had no sort of productization intent.
- 16:32SPEAKER_00: And those teams were doing their thing on their own.
- 16:36SPEAKER_00: I imagine the same thing with AI,
- 16:38SPEAKER_00: although DeepMind was spun out and has since been brought back into Google.
- 16:45SPEAKER_00: Yeah.
- 16:46SPEAKER_01: Yeah,
- 16:46SPEAKER_01: so in terms of regulation,
- 16:48SPEAKER_01: I think everyone assumes that there's going to be regulation,
- 16:50SPEAKER_01: there's going to be oversight.
- 16:51SPEAKER_01: AI is too big of a technology.
- 16:54SPEAKER_01: You know,
- 16:55SPEAKER_01: for example,
- 16:55SPEAKER_01: you know,
- 16:56SPEAKER_01: the genetic space where I come from,
- 16:57SPEAKER_01: you know,
- 16:57SPEAKER_01: there was a Solomar conference in the 70s about recombinant technology.
- 17:02SPEAKER_01: And there's all sorts of self-regulation and explicit regulation related to genetic
- 17:07SPEAKER_01: engineering technology and whatnot,
- 17:08SPEAKER_01: right?
- 17:09SPEAKER_01: So I think AI, that's going to happen.
- 17:11SPEAKER_01: My question, though, is...
- 17:13SPEAKER_01: And this is the issue that actually privately there's a lot of conversations about this.
- 17:17SPEAKER_01: What if the U.S.
- 17:18SPEAKER_01: has these strict regulations and, I don't know, China does not?
- 17:23SPEAKER_01: How do we work AI into a context of great power competition?
- 17:30SPEAKER_00: Yeah, and I'm very concerned about that as well.
- 17:31SPEAKER_00: I have a new piece in American Affairs talking about geopolitics of AI and compute.
- 17:36SPEAKER_00: And that's why I think that actual statutory regulation of AI per se
- 17:44SPEAKER_00: is either ill conceived or premature.
- 17:48SPEAKER_00: Um,
- 17:49SPEAKER_00: there's,
- 17:49SPEAKER_00: you know,
- 17:50SPEAKER_00: on the ill conceived side,
- 17:51SPEAKER_00: there are even,
- 17:52SPEAKER_00: even within the federal government,
- 17:54SPEAKER_00: sort of the kind of,
- 17:55SPEAKER_00: you know,
- 17:56SPEAKER_00: federal government regulating itself through guidance and so forth.
- 17:59SPEAKER_00: Uh,
- 18:00SPEAKER_00: very messy,
- 18:01SPEAKER_00: confused guidance around,
- 18:03SPEAKER_00: you know,
- 18:03SPEAKER_00: rules for reducing harm when you're using AI,
- 18:06SPEAKER_00: where AI basically means software.
- 18:09SPEAKER_00: And how you draw a line to sort of carve nature as joints to distinguish between,
- 18:16SPEAKER_00: you know,
- 18:16SPEAKER_00: basic machine learning and statistics versus agential,
- 18:21SPEAKER_00: intelligent reasoning models is actually harder than it sounds because you can sort
- 18:27SPEAKER_00: of tell the difference between
- 18:29SPEAKER_00: But to draw the line into the law is hard because by nature, more is different, right?
- 18:34SPEAKER_00: There are qualitative difference from quantity when you scale these models up.
- 18:38SPEAKER_00: And that's why I think,
- 18:39SPEAKER_00: you know,
- 18:39SPEAKER_00: I want to segment the litany of potential AI regulatory issues from the core
- 18:47SPEAKER_00: concern I have around the hyperscalers.
- 18:50SPEAKER_00: And those are basically OpenAI Anthropic and their patrons in the form of Microsoft and Google.
- 18:55SPEAKER_01: Yeah,
- 18:56SPEAKER_01: I mean,
- 18:57SPEAKER_01: here in the States,
- 18:57SPEAKER_01: we just had an issue about whether – I think whether burritos were sandwiches.
- 19:03SPEAKER_01: So these regulations,
- 19:05SPEAKER_01: how they're interpreted,
- 19:06SPEAKER_01: how they're written,
- 19:08SPEAKER_01: they can get really weird in terms of what gets classified as what.
- 19:12SPEAKER_01: So just because you write something you think is straightforward,
- 19:15SPEAKER_01: it might not have the impact that you think.
- 19:19SPEAKER_01: So most of the – so you have your second biggest – your second class is –
- 19:23SPEAKER_01: Most of the AI regulations are ill-conceived or premature.
- 19:26SPEAKER_01: And I'm assuming that you're talking here about these like top-down, kind of like clunky regulations.
- 19:32SPEAKER_01: Can you talk about some of the things that have been proposed?
- 19:35SPEAKER_00: Yeah, I've been critical of the EUAI Act of a piece in City Journal discussing its problems.
- 19:41SPEAKER_00: It does this, it sort of tries to combine two things.
- 19:44SPEAKER_00: So originally it was this risk tier framework.
- 19:47SPEAKER_00: And by the way, the EUAI Act was developed beginning four years ago.
- 19:51SPEAKER_00: So it predates ChachiBT and reflects a kind of 2010s view of AI where it's all
- 19:57SPEAKER_00: about algorithmic bias and harm from discrimination and so forth.
- 20:03SPEAKER_00: So it lays out these four risk tiers from like
- 20:07SPEAKER_00: minimal to no risk to high risk where high risk means you could be denied a loan
- 20:12SPEAKER_00: you you know it could recognize your emotions or like these things that are you
- 20:19SPEAKER_00: know dubious to call high risk it also lumps in you know critical infrastructure
- 20:23SPEAKER_00: medical devices which you could make an argument but a lot of the things that
- 20:26SPEAKER_00: consider high risk are not to me not high risk at all they're about you know
- 20:31SPEAKER_00: protecting the european for freedoms or whatever
- 20:33SPEAKER_00: And so that's bad in and of itself.
- 20:37SPEAKER_00: But secondly, it's bad because it's done at a use level.
- 20:41SPEAKER_00: So it's not that they're going to the few big companies that are creating these
- 20:45SPEAKER_00: foundation models and saying,
- 20:47SPEAKER_00: you must make sure your models are not going to take over the world.
- 20:50SPEAKER_00: They're saying anyone who uses these systems
- 20:53SPEAKER_00: in a particular application or use case,
- 20:55SPEAKER_00: has to abide by these rules,
- 20:57SPEAKER_00: has to have quality assurance systems in place and monitoring systems in place.
- 21:01SPEAKER_00: It's a train wreck in terms of the scope.
- 21:05SPEAKER_00: That's the kind of thing I worry about coming to the US.
- 21:08SPEAKER_00: That sort of thing really would stifle innovation in the US.
- 21:11SPEAKER_00: And you're already seeing it in the case of Europe,
- 21:13SPEAKER_00: where a lot of US companies are delaying the release or planning to not release
- 21:16SPEAKER_00: there at all.
- 21:17SPEAKER_01: Yeah,
- 21:18SPEAKER_01: I mean,
- 21:18SPEAKER_01: so we've seen this with the Europeans before,
- 21:22SPEAKER_01: and I'm not saying this is good or bad,
- 21:23SPEAKER_01: but it seems like the EU in particular has been thinking that it can kind of veto
- 21:30SPEAKER_01: or kind of affect the –
- 21:34SPEAKER_01: corporate technological you know just framework that the whole world has because
- 21:38SPEAKER_01: it's a big economy and they can act it with one voice I don't I think that that's
- 21:43SPEAKER_01: getting less and less impactful though because it seems to me that a lot of these
- 21:50SPEAKER_01: cutting edge technologies are almost writing Europe off I don't know what you think
- 21:54SPEAKER_00: I don't keep track of it closely but that's my perception yeah no I'm going back to
- 21:58SPEAKER_00: the
- 21:59SPEAKER_00: conversation about you know the internet seeing internet gdp statistics robert
- 22:02SPEAKER_00: gordon has a great paper on the transatlantic productivity slowdown where he shows
- 22:06SPEAKER_00: that yes productivity has been slowing down across the west but um there was a
- 22:11SPEAKER_00: legitimate it boom that boosted us gdp in the late 90s early 2000s but it didn't
- 22:17SPEAKER_00: manifest in europe so europe already missed out on one uh productivity wave and now
- 22:22SPEAKER_00: they're about to miss out on the next
- 22:24SPEAKER_01: Yeah,
- 22:24SPEAKER_01: like,
- 22:24SPEAKER_01: so there's,
- 22:26SPEAKER_01: yeah,
- 22:26SPEAKER_01: there's massive,
- 22:27SPEAKER_01: this might be part of,
- 22:28SPEAKER_01: you know,
- 22:29SPEAKER_01: we have all these unicorns,
- 22:30SPEAKER_01: like,
- 22:30SPEAKER_01: this might be part of the reason for the GDP,
- 22:33SPEAKER_01: you know,
- 22:33SPEAKER_01: gap that's,
- 22:34SPEAKER_01: like,
- 22:34SPEAKER_01: growing between,
- 22:35SPEAKER_01: it looks like,
- 22:36SPEAKER_01: the EU and the UK and the US,
- 22:38SPEAKER_01: right?
- 22:38SPEAKER_01: So,
- 22:39SPEAKER_00: we are...
- 22:40SPEAKER_00: They're also very,
- 22:40SPEAKER_00: like,
- 22:41SPEAKER_00: even with the common market and Eurozone,
- 22:45SPEAKER_00: the economies are much more fragmented.
- 22:46SPEAKER_00: Like,
- 22:46SPEAKER_00: there's obviously huge differences between Italy and France and Germany versus,
- 22:51SPEAKER_00: you know,
- 22:52SPEAKER_00: Nevada,
- 22:52SPEAKER_00: Utah,
- 22:52SPEAKER_00: and Idaho.
- 22:53SPEAKER_00: Yeah.
- 22:54SPEAKER_01: Yeah, yeah.
- 22:54SPEAKER_01: Yeah, that's true.
- 22:55SPEAKER_01: That's true.
- 22:56SPEAKER_01: It's different.
- 22:57SPEAKER_01: Okay, so these top-down regulations will not work.
- 23:04SPEAKER_01: And so, obviously, what you're talking about is focusing on every use case, focusing on consumers.
- 23:10SPEAKER_01: That's kind of micromanaging, right?
- 23:13SPEAKER_01: We need to look at the companies, the big players, big players that are scaling.
- 23:17SPEAKER_01: Because there are lots... Actually, every single startup that's in tech right now is an AI startup.
- 23:24SPEAKER_01: at least according to the people that they're trying to raise seed money for.
- 23:27SPEAKER_01: So there's AI and then there's AI.
- 23:32SPEAKER_01: And so what you're talking about is these big players are really the ones that are
- 23:37SPEAKER_01: on the forefront and pushing things forward.
- 23:40SPEAKER_01: And I think the reason is they're the ones that are training these really,
- 23:46SPEAKER_01: really powerful models because these models are extremely compute-intensive.
- 23:50SPEAKER_00: They're expensive, right?
- 23:52SPEAKER_00: Exactly.
- 23:52SPEAKER_00: And in every case,
- 23:54SPEAKER_00: whether it's DeepMind,
- 23:56SPEAKER_00: OpenAI,
- 23:57SPEAKER_00: or Anthropic,
- 23:58SPEAKER_00: have building AGI as their explicit mission.
- 24:02SPEAKER_01: Yeah, okay.
- 24:03SPEAKER_01: And you can only do that with a massive amount of compute horsepower.
- 24:07SPEAKER_01: Exactly.
- 24:08SPEAKER_01: Yeah.
- 24:09SPEAKER_01: And so it's obvious you can't ignore what they're doing.
- 24:12SPEAKER_01: And in terms of just oversight,
- 24:17SPEAKER_01: it's quite different than tracking every single business that's in this space.
- 24:22SPEAKER_00: Yeah, exactly.
- 24:22SPEAKER_00: And the way we're doing it now,
- 24:24SPEAKER_00: at least provisionally,
- 24:24SPEAKER_00: is with this executive order that uses the Defense Production Act,
- 24:28SPEAKER_00: which is just a misunderstood part of the law,
- 24:32SPEAKER_00: but it's essentially the powers of the Department of Defense to have oversight and
- 24:38SPEAKER_00: intervene if they need to.
- 24:41SPEAKER_00: And it's a discretionary thing, and people don't like discretion often.
- 24:45SPEAKER_00: People prefer rules.
- 24:46SPEAKER_00: But one of the issues when you have fast-changing technology and events is discretion really matters.
- 24:52SPEAKER_00: And when you know,
- 24:53SPEAKER_00: and their and their focus,
- 24:54SPEAKER_00: their mandate is to focus solely on things that implicate national security,
- 24:58SPEAKER_00: whether that's,
- 24:59SPEAKER_00: you know,
- 25:00SPEAKER_00: that includes both like a super intelligence,
- 25:02SPEAKER_00: or,
- 25:03SPEAKER_00: you know,
- 25:03SPEAKER_00: dual use technologies like cyber for cyber warfare or bio biological engineering.
- 25:08SPEAKER_01: Well,
- 25:09SPEAKER_01: okay,
- 25:09SPEAKER_01: so one of your assertions,
- 25:11SPEAKER_01: or just like a class of bundle theses,
- 25:14SPEAKER_01: let's just put it that way,
- 25:15SPEAKER_01: is that the AI progress is accelerating and not plateauing.
- 25:18SPEAKER_01: Can you talk about why you think that?
- 25:20SPEAKER_01: Because other people have been saying that it's plateauing.
- 25:23SPEAKER_01: Some people,
- 25:24SPEAKER_01: so as we're recording,
- 25:26SPEAKER_01: ChatGPT 4.0 came out,
- 25:30SPEAKER_01: and some people were saying,
- 25:31SPEAKER_01: oh,
- 25:31SPEAKER_01: it's not really that much of an advancement,
- 25:32SPEAKER_01: et cetera,
- 25:33SPEAKER_01: et cetera.
- 25:34SPEAKER_00: Yeah, I think it's important to remember that GPT-4 was trained in 2022.
- 25:42SPEAKER_00: So OpenAI has had a two-year lead.
- 25:45SPEAKER_00: They've explicitly said that they're slow rolling the release of these things to
- 25:49SPEAKER_00: encourage sort of societal adaptation and so forth.
- 25:53SPEAKER_00: GPT-4.0 is a GPT-4 scale model, so it's intentionally trying not to push the frontier out.
- 25:59SPEAKER_00: And nonetheless, on the
- 26:03SPEAKER_00: subjective leaderboards where they give you two model outputs and you don't know
- 26:07SPEAKER_00: which is which and you vote on which one you prefer.
- 26:09SPEAKER_00: It's almost 100 yellow points above the next best model.
- 26:13SPEAKER_00: So it is significantly better, especially at coding.
- 26:17SPEAKER_00: And they attributed that to its post-training.
- 26:19SPEAKER_00: So my speculation is that it's been post-trained on GPT-5 outputs.
- 26:23SPEAKER_01: Okay, wait, can you tell me, explain what that is?
- 26:27SPEAKER_00: So when you train one of these large language models or large multimodal models,
- 26:32SPEAKER_00: you train it in several steps.
- 26:34SPEAKER_00: Step one is called the pre-training phase,
- 26:36SPEAKER_00: where you're just purely trying to predict the next token,
- 26:38SPEAKER_00: whether it's all the text on the internet,
- 26:40SPEAKER_00: all the videos on YouTube,
- 26:40SPEAKER_00: you're trying to predict the next frame,
- 26:42SPEAKER_00: the next word.
- 26:43SPEAKER_00: And then you're left with a text completion engine,
- 26:46SPEAKER_00: where you type some words and it continues your words.
- 26:48SPEAKER_00: The next stage is called instruction tuning and supervised fine tuning, where
- 26:52SPEAKER_00: That's when you turn the thing into a chat bot where it actually answers in full paragraphs and so on.
- 26:57SPEAKER_00: And then the final stage is reinforcement learning from human feedback.
- 27:01SPEAKER_00: That's where you can do it with the ethics or whatever.
- 27:03SPEAKER_00: But you can also use that stage to fine tune its outputs with a more powerful model.
- 27:09SPEAKER_00: So even as a GPT-4 scale model, they could use GPT-5, which people think has already been trained,
- 27:17SPEAKER_00: And have GPT-5 be the source of feedback to tighten up the model and make it smarter without scaling it.
- 27:24SPEAKER_01: Okay, okay, that makes sense.
- 27:27SPEAKER_01: And yeah, so I will say that I do pay for, I pay for Gemini, Claude, GPT, OpenAI, ChatGPT.
- 27:35SPEAKER_01: Definitely 4.0 is... I had started to use Cloud more than ChatGPT 4.0.
- 27:43SPEAKER_01: And then now that 4.0 came out, I went back to ChatGPT.
- 27:47SPEAKER_01: You know what this feels like?
- 27:49SPEAKER_01: This feels like the browser wars of the 1990s.
- 27:54SPEAKER_01: And like,
- 27:54SPEAKER_01: you know,
- 27:55SPEAKER_01: you youth don't remember,
- 27:56SPEAKER_01: but there was a new browser that was coming out every couple of months.
- 28:00SPEAKER_01: And so,
- 28:01SPEAKER_01: you know,
- 28:02SPEAKER_01: Netscape would come out with a browser that,
- 28:04SPEAKER_01: you know,
- 28:05SPEAKER_01: Microsoft would come out with another browser a month later.
- 28:07SPEAKER_01: And so you get like three new upgrades within a year or within like six months.
- 28:12SPEAKER_01: Just crazy, crazy.
- 28:14SPEAKER_01: And then like eventually...
- 28:17SPEAKER_01: microsoft won the browser war and we just got the same explorer for like you know
- 28:21SPEAKER_01: about 10 years until you know firefox and uh yeah and explore uh and chrome change
- 28:27SPEAKER_01: things right uh so we are in that like what was free the browsers yeah i'm paying
- 28:34SPEAKER_00: yeah with these with these models you know my i don't want to look at my bank
- 28:38SPEAKER_00: account my checking account because it's just subscriptions up the wazoo
- 28:42SPEAKER_01: Yeah, yeah.
- 28:42SPEAKER_01: Well, I mean it's expensive.
- 28:44SPEAKER_01: So that's one of the – so you're talking about AI progress is accelerating.
- 28:49SPEAKER_01: People are freaking out about that.
- 28:51SPEAKER_01: But one of the limitations is just energy inputs, right?
- 28:56SPEAKER_01: It's just like cost.
- 28:57SPEAKER_00: yeah you know all the inputs so the hardware inputs is when the main thing that's
- 29:02SPEAKER_00: just started to alleviate right so you know people who are saying oh the gary
- 29:07SPEAKER_00: marcus's of the world one of the premium ai skeptics it's like well you know where
- 29:12SPEAKER_00: is you know scaling is really paying off where is the new better model we haven't
- 29:15SPEAKER_00: had a new better model in a while i just said that you know gp4 was is really two
- 29:19SPEAKER_00: two and a half year old technology at this point um
- 29:23SPEAKER_00: To release a model,
- 29:24SPEAKER_00: you also have to have the infrastructure to train it,
- 29:27SPEAKER_00: and then you have to have the infrastructure to service it for inference,
- 29:31SPEAKER_00: especially if you want to make it free and have millions of users.
- 29:34SPEAKER_00: And last year, NVIDIA produced half a million H100s.
- 29:38SPEAKER_00: They were completely sold out.
- 29:40SPEAKER_00: Everyone was trying to buy them.
- 29:42SPEAKER_00: This year, they've quadrupled their production.
- 29:44SPEAKER_00: And there were some key shortages with packaging and some other components that go
- 29:49SPEAKER_00: into these GPUs that is just now starting to alleviate.
- 29:52SPEAKER_00: And that's why you've started to see,
- 29:54SPEAKER_00: you know,
- 29:54SPEAKER_00: not just that progress start to pick back up as the leading companies,
- 29:58but also many other...