Subscribe
Share
Share
Embed
A podcast from Reserve about the AI revolution: the technology, the power, and the future it's pulling us toward. Reserve co-founder and CEO Nevin Freeman explores the AI boom from two angles: how to invest in it and own your share of the buildout, and where AI may really take us in the long term. But it's really about investing our way out of the permanent underclass.
Nevin: [00:00:00] Do you think that if we are creating conscious experiences in these things, we'll ever be able to tell?
Samo: Everyone asks, "What is it like to be Claude?" No one asks, "What is it like to be a Waymo?" One of the things that people that are not in Silicon Valley might not believe is just how profoundly influential science fiction and transhumanist ideas are, and how motivating they are.
People here are giddy with excitement often, with the idea of transcending their own personal limits or the limits of humanity, and sometimes, and this one I don't actually like that much, replacing humans with a morally superior being.
Nevin: Do you think, uh, do you think we have a, a good shot, or do you think, uh, it all hangs in the balance and it depends on- Mm
the institutions and great founders of this moment, whether we're gonna make it through, you know, with respect to AI or all the things- Mm-hmm ... that it brings? What's, what's your thought?
Samo: I do think we are subject to some of those same, uh, s- sociological civilizational forces that [00:01:00] have caused previous civilizations to kind of deteriorate, right?
Both politically, economically, socially, but ultimately so far that it even erodes the knowledge base, right? So I would put it at something like, I think we're 30% odds that this current human civilization is sort of the one that makes it to the singularity, since the show is-
Nevin: Okay ...
Samo: is termed The Singularity.
Nevin: Yeah.
Samo: And we're about 70% where people will like, you know, look 5,000 years from now or 2,000 years and they'll be like, "Wow, early 21st century North America was building data centers. Can you imagine that primitive civilization?" "They're burning rocks from the ground, you know, they're like driving around in internal combustion engine cars, but they were building data centers.
Like if only they built like five times more, uh, they would've just had a super intelligence answer all their questions." Yeah.
Nevin: Yeah.
Samo: But they didn't.
Nevin: Are you comfortable sharing what you are invested in in the space?
Samo: Now people are like, "Hey, I should have invested in Korea." Well, maybe you should invest in the Netherlands.
It's almost a narrative [00:02:00] violation, right? We assume all of industry is in East Asia, but actually the Netherlands is right there embedded in that.
Nevin: Yeah.
Samo: So if the East Asia thesis is correct, then the Ne- the Netherlands does well as well. The data centers will really increase the amount of natural gas we burn.
Any company that is like doing anything related to liquefied natural gas infrastructure globally is almost like right now experiencing like massive, massive, uh, demand for their services.
Nevin: You could have a situation where it's like, yeah, this is gonna be real, it's gonna be huge, it's gonna revolutionize everything, and yet still you could have speculation valuations get way ahead- Mm-hmm
and then a, and then a big sell-off. If we look back on this conversation a year or two from now and, and it was the top of the market now- Mm-hmm ... what's the most likely explanation for how that could be?[00:03:00]
Welcome, Samo, to episode one of Buying Into the Singularity, in which we discuss, you know, the AI situation, the AI build-out, how to invest in it, um, but also taking it seriously. You know, thinking about where is this all going, and it's not just about money. It's, it's about control. It's about sort of the future state of the world and all that.
Um, I'm super excited to start this off with Samo Burja, the founder of Bismarck Analysis- Mm-hmm ... and, uh, the primary writer of the Bismarck Brief. Um, because Samo, you've been, you know, doing extremely deep research and analysis into people and institutions that basically shape the, the, the state of reality for such a long time.
And, uh, I, I think it's only natural that you guys have started to have a pretty significant focus on the topic of AI. Mm-hmm. Because, you know, at least according to, to many, and, and I think you and I share the intuition that, um, you know, it may be one of or, or the biggest, uh, force to really [00:04:00] understand- Mm-hmm
uh, that may shape how things turn out in, um, in the next couple of decades. So, um- I guess before we dive in, um, do you wanna just say a couple words about, you know, what is Bismarck? Why did you found it? What do you guys do on sort of a month-to-month basis? And why has your writing gotten so much traction?
Samo: Yes. Well, um, Bismarck Analysis is a strategic and, um... It is a strategic consulting and research firm, and I founded it essentially because I thought that people were under-investigating the institutional landscape of society, um, which key individuals shape the world, which organizations are shaping the world, and how do those, uh, function?
How well do they work? How well do they not work? I think there's something fairly deep and unified about how human beings coordinate, how we transmit knowledge, how we make decisions, how we run organizations. And I think because of that, um, I think that there was, uh, [00:05:00] something to investigate here, something to research.
For me, it was also just a way to fund, um, a lot of the best social science case studies I could think of. I'm a big believer in sort of the case study methodology. I think that the main reason that, say, economics and political science, but also like, you know, softer approaches like, you know, anthropology or sociology, have had such disappointing, uh, results through the course of the 20th and 21st century, uh, that has been because we don't yet have the right paradigm of research.
So because of that, you end up doing hundreds of case studies of companies, of bureaucracies, and what, what is the result of that? Well, that brings me back to the description of a strategic consulting firm, right? We, in a way, provide the situational awareness of what is happening in your industry. We provide an alternative perspective from, say, your in-house management, which is probably going to be following industry consensus as to how your industry will change [00:06:00] and what your company's approach to that should be.
And, uh, because of that, we end up being hired on all sorts of interesting topics, everything from, like, the steel supply chain in the United Kingdom, um, to, say, like, the development of the regulatory landscape worldwide. And further, there's a lot of aspects of economics today that are deeply entwined with, say, political or geopolitical questions.
Mm-hmm. And then it's kind of a strength of this paradigm of approach that I have and that my firm now has been carrying out for, oof, seven, eight years, that we have this, uh, theory of unity, essentially, of the different kinds of human organizations. They use some of the same approaches. So much as you might analyze, you know, is Apple an adaptive organization that is, you know, pricing in AI, the answer is no.
Um, you can use much the same tools and much the same analysis to figure out, okay, is this, uh, you know... Is the [00:07:00] Singaporean, uh, trade ministry, are they- Uh, aware of the world and its developments. How are they planning to position Singapore?
Nevin: And so I think, if I understand right, you know, you set up Bismarck and you, you, you sell research to your clients and you, and you, um, in a way sell it to the world with- Mm-hmm
with the paid brief, um, as a way to be able to spend your time studying these people and institutions and, and making a map and, um- Mm-hmm ... making sense of it. And I, I knew you before that. I know that you were very interested in understanding those things even before making a business of it.
Samo: Mm-hmm.
Nevin: Why are you personally so fascinated by understanding the world in this way on this level?
Like, what, what... Yeah, what do you get out of that understanding?
Samo: I think that just the world is absolutely fascinating, and what does happen is as interesting as what doesn't happen. We have so many, um, scientists, politicians, economists, um, you know, workers, [00:08:00] engineers of every kind. You would think that everything possible in the world, everything that human beings can do would already have happened, could already be done, right?
At least with millions, if not billions of people working on those problems. Um, turns out mostly no, and it's fascinating. It's a fascinating question to me why that is the case. Mm. So to me, it seems, like, very, very interesting what are sort of these, like, structural forces that, like, determine how fast civilizations advance and why they also perplex- perplexingly, uh, sometimes decay and regress.
Uh, we have numerous examples through history of civilizational collapse. Now, uh, archeology is not as easy to monetize as, uh, perhaps economic or political analysis. Mm-hmm. Uh, but that has been an enduring interest of mine, and I've written, uh, many articles on it as well. I sort of feel like this, this long human story and long story of civilization that, uh, we've not really grappled with and what it [00:09:00] means about even the nature of our relationship with technology, right?
Why is it the case that societies can technologically regress, and why is it the case that many societies become stagnant for a very long time? I think we normally tell the story of, oh, they were just all kind of sort of... They were kind of dumb, and, uh, they didn't, uh, understand or see the value of it. Uh, but I think that's far too simple a story and ignores the many ways that while today we are still a technologically advancing civilization, like is manifest with AI, there are other ways in which we are just as stagnant as a Bronze Age Egypt and possibly for the same reasons.
Nevin: I remember when we first met you, I think at the time, still had your Twitter banner as, uh, a drawing of the Great Filter.
Samo: Yes.
Nevin: And, and you were sort of, you know, pointing people to this question of are we past the Great Filter or not- Yes ... which for the audience, in case anyone doesn't know, the idea is that, you know, that we don't see a whole lot of [00:10:00] life teeming throughout the universe, and so there's something that prevents life from forming- Mm-hmm
all over the place, uh, where somehow we, we have, we have form- formed, and so there's a question of, well, what is it that prevents, uh, life from forming? That's called the Great Filter, and the question is, are we past it yet- Mm-hmm ... or not? Um, and so what's your view on that now? Are we past the Great Filter or not?
Samo: Well, you know, if we, if the universe was teeming with advanced intelligent life, right, we would be able to see it with our telescopes. Yeah. We would see like, you know, Dyson spheres being assembled where stars are enclosed to harness all of their energy. We would see like blasts of like, like rapid acceleration of par- of, uh, objects near the speed of light, and most importantly, we would probably see like an, a bunch of transmission.
Now maybe the, the, um, uh, radio transmissions or whatever technique is used would be encrypted and we wouldn't be able to decode it, but we would hear it, right? Mm-hmm. Mm-hmm. Uh, yet when we look around the universe, like you said, it looks like a very natural universe, [00:11:00] and we from this assume that it's kind of mostly a dead universe, right?
It seems that either planets don't have life that often, they have life, but intelligence doesn't arise, intelligence arises, but technological civilization does not sustain itself. And where are we right now? Well, we have a technological civilization. So for me, you know, that, that banner was like, it was like kind of a, this central question.
Uh, are we at the start of this kind of like infinite growth, infinite explosion, or are we perhaps kind of this blip on the radar? Or even more interestingly, might there be some actions we can take to make, uh, the former rather than the latter be our future?
Nevin: And so do you, are you comfortable stating a view on that?
Do you think that, uh ... Do you think, uh, do you think we have a, a good shot or do you think, uh, it all hangs in the balance and it depends on- Mm ... the institutions and great founders of this moment whether we're gonna make it through, you know, with respect to AI or all the [00:12:00] things- Mm-hmm ... that it brings?
What's, what's your thought?
Samo: I think that, um, our society is way much more fragile than is usually assumed, and I think that we, we have already stepped away from many possible futures that we could have developed, right? There's no physical reason that we couldn't have had a nuclear-powered civilization.
There is no physical reason really that, you know, we got Starship maybe let's say this year or last year rather than 40 or 50 years ago, right? We could have developed those sort of reusable technologies. We could have become a multi-planetary species as is sometimes stated. Uh, and I do think we are subject to some of those same, uh, s- sociological civilizational forces that have caused previous civilizations to kind of deteriorate, right?
Both politically, economically, socially, but ultimately so far that it even erodes the knowledge base, right? I believe in other words that even today our ability to [00:13:00] preserve practical engineering knowledge or process knowledge, I think it's far from certain. We have vastly more documentation, but all of that documentation is in a way very fragile.
Uh, arguably we're sort of one solar flare away from the world's hard drives being erased and, uh, I think that people also underestimate how bad things can get if you had say an outbreak of technologically advanced cyber warfare. Mm. So there's some ways in which as we become more powerful our ability to destroy also increases, including destroying something like for example if I wanted to censor a book in 1500, um, I would probably have to go to the printers and find all the illegal shops of the printers and sort of like smash the machines or confiscate all the books, imprison all the people and eventually I'd do pretty good.
But I probably couldn't get every book in a private collection. Now, if I wanted to destroy every PDF, [00:14:00] every copy of a PDF that is on a non-air-gapped computer, on any computer connected to the internet, arguably I could do a much more thorough job.
Nevin: Hmm.
Samo: I would not have to ransack people's homes. People's computers are metaphorically doors wide open, even if not to the government, uh, with human hackers, then they will be to a government with Mythos or Super Mythos or whatever the next version of this super AI is.
Hmm. So in a way, all of our knowledge has become intensely correlated, and even though it's distributed in a decentralized way, knowledge management is still inching towards centralization- Hmm ... I would argue. Hmm. Um, well, I'm giving you a complicated answer because it's a complicated- a, a simple question that has very complicated implications.
Yeah. And the short bottom line answer is I think it's down to us. I think our odds are decent, but I actually think they're not that great. So I would put it at something like, I think we're 30% [00:15:00] odds that this current human civilization is sort of the one that makes it to the singularity, since the show is-
Nevin: Okay
Samo: is termed The Singularity.
Nevin: Yeah.
Samo: And we're about 70% where people will like, you know, look 5,000 years from now or 2,000 years and they'll be like, "Wow, early 21st century North America was building data centers. Can you imagine that primitive civilization?" "They're burning rocks from the ground. You know, they're, like, driving around in internal combustion engine cars, but they were building data centers.
Like, if only they built, like, five times more, uh, they would have just had a super intelligence answer all their questions."
Nevin: Yeah.
Samo: But they didn't, right? And that story, it's easy to tell, right? We could imagine a world where the singularity hypothesis isn't true and things go down. Mm-hmm. But the thing that is interesting about this is this is not like then an AI bubble, then that would be kind of a civilizational bubble.
So ultimately, on the idea that, you know, pessimists can sound very wise, but maybe optimists are the ones that make money, [00:16:00] in most of the futures where the future of our civilization is good, where the investments are good- Mm-hmm ... where, uh, the scientific work we're doing is building towards a future rather than j- just to be forgotten.
In most of those futures, we get something like a real AI revolution, and I think we're at the start of it. I just feel it's, it's still so easy for us to step away from it.
Nevin: And are you saying then that, like- Even if the, even if you think it's less than 50% chance that we sort of m- make it to that takeoff this time around, uh, that it makes sense to play for that world or that set of worlds anyway because they're the ones that matter.
Samo: I think so, yes. Yeah. And, um, I'm of course also here counting the possibility of us catastrophically misusing earl- early AI in a way that causes our civilization to end, either to be replaced by a completely machine civilization or no civilization at all. I think too few people consider the possibility of truly [00:17:00] devastating industrial accidents.
I think, uh, you know, like massive scale, um, where, you know, AI is sort of used to take out in like some war situation a key pillar of modern society, and we just have to restart.
Nevin: Yeah, yeah. So zooming into this moment, you know, a couple weeks ago at the Machine Consciousness, Consciousness Conference, uh, we looked together at that image of the total Korean stock market index- Yes
since, I think, 2008. Uh, we can put it on the screen for the podcast. It, you know, looks like a crazy hockey stick right now. Um, and to me, you know, in crypto terms, for people who are familiar with crypto, this feels... You know, you just look at the chart and it feels like 2017. Mm-hmm. And even the way people are talking and doing deals and whatever- Mm-hmm
it feels very 2017. And, um, you know, I wasn't really paying attention to the markets in the year 2000, but looking back it still seems [00:18:00] like, you know, in internet terms, maybe- Mm-hmm ... this is kind of the year 2000 in terms of the, the valuations and the way that people are thinking about it. I wrote down a couple of reference valuations in the AI history I think are interesting.
So in 2014, Google bought all of DeepMind for about $500 million. In 2022, uh, Alameda bought 8% of Anthropic in their Series B- Mm-hmm ... for $500 million. Mm-hmm. Um, in 2024, Safe Superintelligence, Ilya Sutskever's, uh, offshoot startup, raised $1 billion in a seed round. Mm-hmm. And I think they raised another couple bil- couple billion the next year.
Um, and then now, OpenAI, Anthropic, and xAI are all each valued at around $1 trillion. Um, and, you know, and then also we just, you, we can see in public markets the whole supply chain is kinda just going bananas as of early 2026. And so just thinking about what this, what this [00:19:00] means, like, um, or, or, or what it's like for the people inside of these companies, the people who started these companies, like- At, at this point, top researchers and operators- Mm-hmm
relatively suddenly have hundreds of billions of dollars at their- Mm-hmm ... disposal to build super intelligence or whatever it else is that they're deciding to, to build. And, and so, and even the people who are just, you know, employees of these companies, um, have achieved great wealth and have quite a lot- Mm
of mobility at this point, quite a lot of optionality.
Samo: Mm-hmm.
Nevin: And so, you know, if you, if you sort of compare that to looking back to, you know, it was around 2010, 2011 that I started watching this industry. I think you've been watching it for about that amount of time as well. Back at the time, I personally did not see this coming.
I didn't think about it this way in terms of, uh, what, what the markets would look like around all this. Um, and in a way, it, [00:20:00] it's, you know, it's like back at the time, the people who were concerned about AI, uh, thinking about its power, both its upside and its downside, you know, there's this meme of like someone could build a brain in a box in a basement.
Samo: Yes.
Nevin: You know, where it's like a really smart person or small team would figure out the special intelligence algorithm. They would, they would hit enter and it would start running, and it would be able to immediately recursively self-improve. And over some period of time, you know, regardless of whether it's, you know, minutes or months or something, you could have an intelligence explosion before society even knew that AI- Mm-hmm
was a thing, right? Mm-hmm. You could go from sort of like sci-fi, Terminator- Mm-hmm ... people watching that to like suddenly it's real and it's- Mm-hmm ... taken over everything. And now that obviously seems so, so wrong relative to what we've seen in practice. In a way, um, everything playing out so much more slowly, uh, than what some, some folks had hypothesized or feared- makes it possible for markets to catch up- [00:21:00] Yes
and, uh, for us to then have this worldwide conversation about AI. And that's, you know, it's really cool, I think, that it's, it gives the world, uh, the opportunity to sort of have a conversation about it- Mm-hmm ... and start to grapple with it and think about it, um, instead of it just being like a tiny cluster of people.
Um, but it's also maybe giving an opportunity to like divvy up ownership and control-
Samo: Mm-hmm ...
Nevin: uh, over, uh, over these companies and, uh, who, you know, uh, could even be done on a state level, um, before, you know, before we have something that sort of has godlike power. Um, and so I'm curious, since you think a lot about the people and the institutions and sort of where the power lies, you know, uh, what is your take on where we have landed so far- Mm-hmm
in terms of the power and control, control structures?
Samo: Mm-hmm.
Nevin: And where do you, how do you forecast that evolving over the next, you know, [00:22:00] 10 years as, you know, these companies IPO? I mean, right now we're hearing about like, like, yeah, actually I pulled up a summary of, uh, news headlines from yesterday, and it's like Trump is signaling openness to a federal equity stake in leading AI firms- Mm-hmm, mm-hmm
um, uh, ahead of a planned meeting with 12 to 15 industry CEOs. That's an interesting- Mm ... you know, interesting development. Um, meanwhile, at the same time, Anthropic, uh, pressed Congress for a rigorous federal law on catastrophic AI risks and mandatory safety tests before any preemption of state rules. Uh, while CEO Dario Amodei separately called for third-party audits of frontier models above a compute threshold and government power to block unsafe deployments.
And then on the same day, OpenAI published a threat report on two China-linked influence operations that used ChatGPT to push narratives about US data center build-outs and tariffs before the company banned the accounts. So, um It's heating up. Yeah. Where do you, where do you, how do you see where we are now in terms of who [00:23:00] has power and control, and how do you think that that's gonna evolve in the next- Hmm
decade?
Samo: Well, uh, you know, as you point out, it dominates the headlines already, so there's broad sort of, um, understanding or awareness that AI is reshaping the world. But I think the, uh, I think the most powerful institutions and organizations of society are taking related but distinct bets on how exactly it will play out.
So while there's a shared awareness that AI is a big deal, it's a major technological revolution, uh, people are making radically different bets. So if we look at, say, Anthropic, right? Anthropic's perspective is very much informed by the sort of classic existential risk, effective altruism arguments, where they sort of feel that, oh, you know, we have to race to the front, and then once we're at the front, we're gonna press the brakes, and we're gonna make sure that the AI that's developed is sort of the safe AI.
Um, and then eventually, that sort of safe AI is, uh, you [00:24:00] know, the thing that brings about utopia and that it will strongly recursively self-improve. So ironically, when you were discussing the sort of classical, uh, you know, Eliezer Yudkowsky 1990s, 2000s view of a brain in a box, well, Anthropic kind of believes that.
They believe in a brain in a data center, right? Mm-hmm. They talked a lot, and the researchers and experts working there talk a lot about the automated AI researcher.
Nevin: Right.
Samo: Right? So maybe it's not running on a small personal computer in a basement, but it's running on a giant data center, and the idea is actually the data center kind of already has all the compute it needs.
We just have to have the code be good enough, efficient enough. And also, we will have the AI itself discover new architectures- Mm-hmm ... like sort of completely new ways to process this massive data we've collected or to generate new synthetic data that's like, you know, so deeply informative that overcomes the data wall Um, so that's one perspective.
A different perspective is, uh, there are still companies deeply betting on [00:25:00] scale, and regardless of what they say, you can tell from their actions. Why would we need orbital data centers? Why should the moon be turned into a computer? Yeah. That doesn't make as much sense unless we're in a world where actually we have to build out so many more orders of magnitude, right?
Mm-hmm. So when you were sharing the financial numbers earlier, 500 million to basically 500 billion, if I exaggerate, we were jumping orders of magnitude over and over again, not just on the profits, not just on the valuations, but on the very spend- Mm ... on compute.
Nevin: Mm-hmm.
Samo: So to jump if it is, and it could be a physical fact of the universe that to build an artificial intelligence, you know, it's sort of like, um, well, sure, a bird can fly, and the Concorde plane can fly, but you need way more resources and calories to make the Concorde do supersonic flight than for a hummingbird, uh, you know, to, you know, go, uh, go visit some flowers- Right
and, uh, fly around them. [00:26:00] Um, and of course, there's things that, you know, the Concorde will do that the hummingbird never could, right? And vice versa. So perhaps the future of intelligence is the case that we're gonna have to build so many data centers that we're gonna need an industrial build-out away from the regulatory constraints and political constraints of Earth.
And then something like building the data centers on the moon makes sense, making them in orbit makes sense. And note, that's not just SpaceX, right? Which of course having acquired xAI is now a rocket and AI company- Mm ... so that's a convenient narrative for their IPO. Right. Uh, that is also Google, and even Nvidia has signaled, like, an openness to orbital data centers as, like, a place where intelligence could be deployed.
And yes, Nvidia does their own version of this belief in scale. From Nvidia's perspective, all this AI stuff is a side effect of the GPU. Mm. And actually, the GPU is the real revolution. Mm. AI is just one application.
Nevin: Mm-hmm.
Samo: Right?
Nevin: Mm-hmm.
Samo: And, uh, you know, the [00:27:00] way Jensen talks about it, I think he honestly kind of believes it, right?
He thinks many different types of, uh, algorithms can run on these cards, and many kinds of algorithms will prove to be transformative, both for AI but also for other applications. Mm. I mean, occasionally he still talks of video games, right? And it's good to remember that that's where that started. Mm-hmm,
Nevin: mm-hmm.
Samo: Technological progress is, like, profoundly filled with surprises, with unexpected codependencies. Mm-hmm. And, uh, maybe that brings us back to Korea. In a way, the leaders of industry, of normal industry, are vastly behind on understanding this. I think they have a misconception still, if they're a few steps away from the AI build-out itself, that this will be sort of like the 1990s or the information revolution.
There'll be some new physical products made, but the software people are gonna basically, like, eat it all up. But I don't think that's [00:28:00] what we're seeing. We are actually seeing fairly competitive AI models where you can get a slightly worse performance for basically free on either an open-source model or a company distills another company's hard-earned- Mm
uh, training run, and actually the software companies are in this, like, a very bitter competition, and maybe, you know, to go back to SpaceX, maybe they're smarter, right? Because sure, they're working with their competitors, Anthropic, right? SpaceX is lending compute to Anthropic. Uh, but also in a way, they control the data centers.
They know how to build them. Mm-hmm. They're very good at navigating the regulatory, uh, space, or at least Elon is, even before any of the more sci-fi space compute civilization stuff kicks in. And I think that, you know, the South Korean stock market going up, South Korea has, like, no real AI companies that matter What they have is electronics companies.
Right. So the demand [00:29:00] driven by the data centers for NVIDIA, but also AMD and also other specialized companies like Cerebras and so on, that demand goes straight to Taiwan, which this quarter I think hit 10% economic growth. Now, to be fair, Taiwan is a rich country, but it's still significantly poorer than the United States or even most rich, rich European countries such as the Netherlands.
So if you have a company like TSMC and an ecosystem around it, if it experiences rapid growth, that growth might be 50% or 60%, and then very quickly you see something that's like 10% economic growth- Mm-hmm ... um, from a country. I think South Korea might approach something very similar. China, ironically, is too big to benefit from it.
But I think if we broke out the data and we looked at the electronics manufacturing ecosystem in Shenzhen or the Pearl River Delta, we would also see 10% economic growth. So in other words, if the Pearl River Delta was a small East Asian [00:30:00] country like, uh, you know, South Korea or Taiwan is, I bet we will see 10% economic growth there too.
And now comes the surprising turn. So once we go through semiconductors and electronics, we eventually start reaching stuff that does not feel very tech. It's extremely finely polished mirrors made in Germany, bought by ASML to make those extreme ultraviolet lithography machines for TSMC. Also, turbines, gas turbines manufactured, in this case, in South Korea, but also in the United States, being deployed to power the data centers because the data centers are outstripping what the grid can provide.
Wait, gas turbines and mirrors? This is starting to look like just an industrial revolution. So here would be my take. The AI revolution, if it is a scaling future, and for now it looks like it significantly will require at least, like, two more orders of magnitude of scaling of compute. [00:31:00] If it is that kind of compute revolution and this build-out revolution, I think that is possibly big enough to restart the engine of industrial society And the result of this, this is the beauty, uh, if demand for atomic precision polished mirrors for advanced optics goes way up, and demand for extremely high temperature reliable gas turbines, which by the way, are a triumph of metallurgy, right?
You have to heat it up many hundreds of degrees, cool it down over and over and over again, and the thing can't become weak, cannot break, right? If those two- if demand for those goes up, unit costs go down, and suddenly there are all sorts of application for- applications for atomic precision mirrors- Hmm
or extremely advanced metallurgy- Mm ... that become cheaper as well. To put it differently, to make a 5 million Model Y Teslas is much more expensive than to make 50 [00:32:00] million, right? Mm-hmm. Or to make a f- 5 million cars is much more expensive than making 5,000 cars. Whenever you look at luxury cars that are super expensive, it- they're not just super expensive because people understand constrained scarcity.
They're super expensive because they're custom machines. So the unit cost for every single example of a really good sports car, that has to pay for the whole development process.
Nevin: Right.
Samo: Right? And, uh, meanwhile, if you made 5 million sports cars of the same type, well, the cost would go down. It might not be quite the cost of an SUV, but it'd start approaching the cost of an SUV.
Nevin: Yeah.
Samo: And we see this over and over again. That's kind of the miracle of industrial society, where intuitively, if you think high demand, we're used to thinking about scarcity, things like gold, things like maybe digital alternatives, things like real estate. But the Industrial Revolution kind of unlocked this possibility for us that, um, you know, things- very complicated machines, [00:33:00] right?
We can build out the factories that make them, and then the machine itself becomes cheap. Don't get me wrong, the factory itself is still expensive. So when we're looking at this, building more chip fabs, building more data centers, building more plants that can build, uh, you know, gas turbines and all of this stuff, you start getting to really basic stuff like cement, right?
Mm-hmm. And steel. And not even advanced steels like special metallurgy, just basic steel. Like, the people installing the electronics. So when you start adding all of this up, it's like, I don't think anyone in South Korea is losing their job. I don't think anyone in Taiwan is losing their job. I don't think anyone in the Pearl River Delta is losing their job.
I actually think in the Netherlands, most people will not be losing their job. So here's my funny prediction. We might yet see, if this AI boom continues, uh, a European country, the Netherlands, suddenly and seemingly randomly experience 10% economic growth- Hmm ... despite the really, say, bad [00:34:00] decisions of, uh, the Netherlands' neighbor, Germany Because that's still a small enough country with a giant company, ASML, with a cultivated ecosystem that is at least as large as ASML, probably two or three times larger of other very valuable companies doing basically industrial work there.
Mm. Highly skilled industrial work. So, uh, eventually that starts hitting limits, and this is one of the reasons that some people are betting on robotics because they see it not just as a driver of demand for GPUs or another way for AI to make money, but a way to solve these seemingly crazy scaling issues.
Mm-hmm. Because if you assume a super intelligent future riding on compute, you can kind of imagine our current civilization doing, like, another order of magnitude. When you start jumping on the second order of magnitude of providing extra compute, you have to start assuming things like orbital data centers- Mm-hmm
or fusion power [00:35:00] or robots doing the very basic boring work of, like, you know, applying the cement and applying the rebar to, like, make the factory that then makes the, uh, gas turbines.
Nevin: Right.
Samo: So I do think that there's a way in which this could be a self- uh, this could be a virtuous cycle, a flywheel. And, uh, why do I think restarting industrial civilization is a perhaps a good idea or why that could help our civilization overall?
Um, I think it's a good idea because we're very close to a number of technologies that solve some of the worst side effects of our first industrialization, which was a coal-based one. We are close to developing the right energy sources- Mm ... but ironically, it won't be from decreasing energy consumption that those become affordable.
Mm. It'll be through increasing them, right? Mm. There's two approaches to, say, reducing carbon emissions from the car. Uh, one is that, you know, you kind of start taxing the cars and the people have fewer cars, and there are just fewer cars on the [00:36:00] road. And the other is to sort of commit to the electrification and that build-out, and then you find, like, clean energy sources for cars.
Mm-hmm. It's really not crazy to imagine most people driving their electric car to work, parking it, plugging it in, and then the solar cells while they're at work basically charging the car during daytime. And most of the driving happens during daytime anyway as well-
Nevin: Mm-hmm ...
Samo: with a small amount happening at night.
So in a way, even if we don't solve stuff with, like, batteries or other approaches, we could have a vastly cleaner civilization just for our transportation, uh, and then hopefully also for things like data centers. Now, I, I do acknowledge that right now the energy demands of the data centers-- You know, I don't like the water thing.
Um- Water is not a problem for them. They consume relatively small amounts
Nevin: Could you expand on that? I've been curious, but I haven't had time to look into it. Well, um, we could double-click on so many things here. Yeah. But to double-click on it, it's[00:37:00]
Samo: like, um, well, we spend vastly more water growing almonds in California than we do on all the data centers of California.
Nevin: Interesting.
Samo: Do we really need to grow almonds in California? Yeah. I don't think so. At the end of the day, though, we don't even need data centers in California, right? People are building them in Texas because it's regulatorily easy, right?
Or Virginia- Yeah, but is that- ... because there's already an infrastructure there ... just
Nevin: because we don't have very many data centers in California?
Samo: We do have some data centers here.
Nevin: Okay.
Samo: Yeah.
Nevin: And just on a basic level-
Samo: Farming, farming in Texas, of course, spends vastly more water, even there, even in Texas- Okay
than all the data centers of Texas.
Nevin: Okay.
Samo: So if we look at energy consumption, you know, when, say, people talk about cows using a lot of water, that's true Cows do use a lot of water, and agriculture uses vast majority of our water. Just a tiny part of that water is drinking water, the kind that you or me might use or, or waste.
Mm-hmm. And, um, then there is some water that is used for industrial purposes, and industrial purposes might have, like, you know, use agriculture [00:38:00] tier water, or they might need, like, special distilled really fancy water for, like, various chemical processes.
Nevin: I have a dumb question. Why, like... Um, I know very little about water usage in data centers.
Mm-hmm. To me, it seems like if you're cooling a chip with water, you could just reuse the same water. You let it cool off, you use it again. Why, why does it even cost water in a, in an ongoing way to power data centers?
Samo: Well, in fact, you do reuse most of it. Yeah. It's just water, you always lose water no matter what the process is.
Yeah. Right? So I think the water rights thing, I think it's because we find water to be compelling. It's like, it's in a way, it's almost sacred to us, right? Like, we feel that there's something very spiritually important about having water because it's such a basic biological necessity for us. So the ideas of running out of water, it's sort of like, oh, it's, we're running out of the stuff of life, right?
And we're putting it into these, like, awful machines that are replacing us or something like that. Yeah. And then the water rights are a potent lever to stop the construction of a data center.
Nevin: Mm-hmm.
Samo: Right? Because [00:39:00] water rights are very, like, very managed, so.
Right.
Samo: And you can also galvanize some sort of, like, local thinking on that.
Uh, but I was going to acknowledge the water issue is not an issue. Uh, we can very cheaply desalinate huge amounts of water, so if we needed to desalinate water for the data centers, we actually could. It's, like, very cheap. We're talking about cents on 1,000 liters or stuff like that. Um, and this is water that, you know, wouldn't really be missed, right?
The ocean is pretty big.
Nevin: Yeah.
Samo: Ocean is pretty big. And, uh, there is a problem perhaps in, I do think initially we'll at least have an increase in CO2 emissions. So I think in a weird way, I'm surprised the environmentalists have gone for water-
Nevin: Right ...
Samo: rather than emissions- Who's- ... because that actually makes sense.
Nevin: If, if that's essentially a lie, who's responsible for that lie? Do you know?
Samo: I, I'm not sure. I think partially journalists just like it 'cause it's a compelling story. [00:40:00] It sort of feels plausible. And I think that there is also this, like, belief that you're not supposed to interrogate these stories that don't make sense.
So even people who kinda know it doesn't make sense, they'll be like, "Well, but on the margin it makes sense." Yeah. And like, actually, the important thing is that we stop these for all of these other environmental impact things. I think it comes from this perspective that it's okay to sort of tell mild lies to the public if it's for a good cause.
The problem with that ethic is that everyone thinks their cause is a good cause. And that actually is also my, one of my rare problems with Anthropic, which are a very admirable company. Their thesis is kind of, well, if the good people don't build AI, the bad people build AI. And my retort to that is like, everyone thinks they're the good people.
Nevin: Right. Right.
Samo: Right? Yeah. So y- y- you don't ... Just because you think your friend group is, like, very good because they share your explicit moral philosophy, the historical record for human beings being excellent because they share an [00:41:00] explicit moral philosophy, I'm not gonna say that it's ... There, there is some- something there, but it's very sketchy.
It's actually not a good way to judge character. Mm-hmm. Like, let, let me put it this way. Anyone sort of ... When people were converting to Christianity from paganism or whatever, uh, they believed that people who converted are just better people. Today, most people in the Western world kinda don't believe that.
Maybe some Christians do actually still believe it. In America, say, like being an atheist, it's still perceived to be more unethical. Uh, and it does change behavior in some way. Like, people, you know, in some places human sacrifice was stopped and so on, and we can, like, objectively look at this as moral improvement.
But anyone can, that is sufficiently intelligent, proclaim, like, adherence to any moral philosophy, sec- secular or religious, quite persuasively. I think it does not absolve the, like, interrogation of, like, character or reliability in all of these things. And, uh, I think that if, if one believes that your particular [00:42:00] ideological group is filled with the good people, and the good people aren't in it, that that can lead to tragic mistakes.
Nevin: Mm-hmm. Mm-hmm.
Samo: Like, not just because you should distrust people around you, but because you won't trust the right people outside of your, of your bubble.
Nevin: Yeah.
Samo: Um, sorry, this has been kind of like a, like kind of a, in a way, a very deep critique of the whole effective a- altruism movement, which, which I liked and in some small ways even participated in.
I, I talked at a conference. I know you were also active. And I still, like, admire a lot of their aims and, and goals. I just think that there's, like, this implicit unstated thing that they don't interrogate very much.
Nevin: Makes sense. Makes sense.
Samo: Yeah.
Nevin: On the ... You reminded me, on the topic of like, um, you know, in, in some cases lies, in other cases subtle misinformation, and we're just, we're just kind of getting it wrong.
Like I, in my experience interacting with most journalists or many journalists- Mm-hmm Any, any time I see, you [00:43:00] know, something that is about something that I was a part of-
Samo: Mm-hmm ...
Nevin: I can easily spot just many ways in which I can tell someone reading that article or watching that video is gonna be misinformed about- Yes
the reality of the thing. Yes. And once you see that happen enough times, I, I feel like it's impossible to take, take reporting on, on other things fully seriously. Mm-hmm. Mm-hmm. You know? Um, when someone you know is written up and it's, like, a totally wrong narrative, it's like, well, okay. Um, does that mean that everything I read about everyone is, is wrong?
And so I wonder, with your study of people- Mm-hmm ... you know, of, of founders who often are objects of fascination, who are written about a lot and, and who are subject to, you know, there's a lot of, uh, people who have political will to bash them. Mm-hmm. Mm-hmm. Um, or, or even just, uh, even just sort of the, the simple fact of, like, as a journalist, it's kind of just easier to rely on a simple trope- Mm-hmm
to get a point across- Mm-hmm ... in a single sentence instead of te- [00:44:00] you know, explaining what really happened. How do you- get to the underlying truth of, like, what the, who that person really is or what their intent really is, uh, or how their, their organization really works- Mm-hmm ... if you're not able to, like, go sit with them for several months?
Like, do you end up needing direct contact in order to study these people? Or is there some way that you can look out at this really sort of deranged and warped information landscape and somehow infer the truth about what's going on, uh, in, you know, in a, in a way that- Mm-hmm ... um, a naive reader wouldn't be able to do?
Samo: Well, let's first scope this a little bit, right? Because, uh, you know, since, um, I'm assuming the viewers might not be as familiar with the live player versus dead player concept that I, you know, published and coined 2017. Um, it's basically this observation that often organizations, but also just small groups of people, individuals, uh, will respond to circumstances in their, like, you know, in their [00:45:00] industry or in their, uh, political environment following well-established scripts, right?
Mm-hmm. In a way, it's kind of, um, sort of following of the best practice.
Nevin: Mm-hmm.
Samo: And then you have some individuals who demonstrate over and over again- Mm ... when they go into very different areas, that actually they can do quite well without a script, and often rewrite the script, rewrite the story. So this came out of the metaphor of sort of like, uh, you know, a, a live player versus a dead player.
The live player is like, you know, playing whatever game it is, is it chess or anything else, uh, live, right? They're present there, and then the dead player's like carrying out this sort of like pre-scripted set of actions. Mm-hmm. That pre-scripted set of actions might be very effective. It might work quite well.
In fact, I think many functional institutions which are very efficient at doing things like, I don't know, producing natural gas, uh, burning turbines to generate electricity, right? Uh, they follow these, like, recipes very [00:46:00] well, and they can adjust within the parameters where these recipes are valid. Um, but it also goes for other things.
It goes for, like, business practices, et cetera. And I think you can establish whether someone is a live player with enough track record, whether someone or a small group is a live player. Say, I would easily say that, I don't know, uh, Elon Musk is a live player, uh, Sam Altman is a live player, and then someone at Anthro- someone at Anthropic is a live player just because of the frequent adjustments there, right?
And then I would say that Samsung as a company, not a live player probably, but fairly functional institution in most of its divisions. It's a very complex conglomerate.
Nevin: And you're judging that by the presence or absence of, like, strategic pivots?
Samo: Successful strategic pivots.
Nevin: Okay. Okay.
Samo: Because I think people often, you know, they might attempt something new, and the default is failure.
Nevin: Mm-hmm. Mm.
Samo: Right? Like, I think, you know, when we [00:47:00] encourage creative thinking in our society, I think we have to also accept that, you know, most people's ideas are just kind of bad. Mo- most ideas if, if I... I mean, if I sit down and had some ideas for how to do, like, a nuclear reactor, I think most of those ideas would probably be bad.
Nevin: Yeah.
Samo: Right? I wouldn't know what I'm doing.
Nevin: Yeah.
Samo: Uh, so it's extremely impressive and weird and kind of a strong signal. Like, you can decry a live player, but even when you're sort of, like, critiquing them 'cause you don't like what they're doing in the world, and in fact, a live player might be doing things that are bad in the world, right?
That doesn't mean that they're, you know, like, good or benevolent or so on. It just means that they're capable- Mm-hmm ... of this kind of, like, deeply restarted first principles type of action that, like, succeeds to a very high degree in completely different domains. Uh, that, by the way, that as an ability, of course, very naturally ends up making it a very powerful force in the world.
Um, uh, they might be bad, but even if you wanna call them [00:48:00] bad, you kind of have to show what they did. And so because of that, to some extent, I think you can usually establish if someone's a live player. Uh, but not always. There is not enough material. There is not enough history. If there are new live players that enter the arena.
Um, and whenever we look at institutions, with those institutions, you have to interview people that closely work with them. You have to go talk to people in the industry. Uh, you have to, like, hire additional consultants who are, like, deep specialists in their area. That's why we have, like, a larger network of experts who help us study this.
So yeah, the truth is, you can do a lot with just, let's call it open source intelligence, and then some of it you just have to actually go talk to the people or people who know them.
Nevin: Yep. Yep. That makes sense.
Samo: I mean, with regard to motivation and private ideology, um, what really motivates people, you know, the crazy thing is people [00:49:00] will just tell you a shocking amount of the time.
It's just that people talk about what is valuable in the world or interesting in the world in very idiosyncratic ways. So there's a way in which someone can tell you that, you know, what they really care about is figuring out where their epistemic bubble is, and then finding how to correct those bubbles, and do they invest on that basis, but they also politically believe that that should happen.
That's kind of George Soros- George Soros' open society, uh, world view and ideology. But if someone, like, tries to act on that, usually they don't do, like, as good a job. But at a high level, do I believe that that's, like, what George Soros was really trying to do? Yeah, actually.
Nevin: Mm-hmm.
Samo: It's a way... Like, like there is no pragmatic reason to make up a lie like that.
Mm-hmm,
Nevin: mm-hmm.
Samo: There might be a psychological reason. There might be, like, a cope. But then if I'm examining then the public writing of someone who's, like, an intellectual or a statesman or something like that, and their [00:50:00] interactions and the reports from others around them, even a very intricate cope actually reveals something very interesting.
Nevin: Mm-hmm.
Samo: Right? It reveals something very, uh, interesting about where their attention goes, right? Yeah. And, uh, what kind of information they do or don't have in the world. So one of the things that people that are not in Silicon Valley might not believe is just how profoundly influential science fiction and transhumanist ideas are, and how motivating they are.
People here are giddy with excitement often with the idea of transcending their own personal limits or the limits of humanity. And sometimes, and this one I don't actually like that much, replacing humans with a morally superior being.
Nevin: Mm-hmm.
Samo: Um, I suspect that, you know, misanthropy is mostly unearned, and that human beings are vastly better than we've, uh, decided they are.
Fairly recently in history, mind you. Like, it's only, like, 1914 that people start saying that, "Oh, humans are [00:51:00] destined to destroy themselves," or, "Humans can't ever do things right," or, you know, "Humans are like a blight on the planet." Um, so I think it's, uh, perhaps we as a civilization in the broader Western sense still have a hangover of a World War I era depression that never quite went away, and just got worse with World War II.
Nevin: But it does feel like it's a pretty small minority that are part of the whole AI sphere that see it in that misanthropic way, right? The
Samo: misanthropes are fairly rare, but the people excited about the potential are high. Yeah. Yeah. So it's both of them like, "Hey, we can replace humanity." And then one side of it is we can perfect humanity and, you know, and grow to these other types of beings, and the other side is like, "Oh, we're gonna just replace them 'cause there's just no reason for carbon brains to stay around if you have silicon brains."
It is a small m- minority, but both things ... Look, both things are motivating enough that people came together to make [00:52:00] companies to, like, literally, like, from their perspective, fight for the future. It's now public record that OpenAI, when it began, was this sort of Sam Altman, Elon Musk alliance. Uh, it was almost the anyone-but-Demis Hassabis alliance.
Nevin: Yeah.
Samo: And then Dario Amodei starts the not-Sam Altman company.
Nevin: Mm-hmm, mm-hmm.
Samo: Right?
Nevin: Yeah.
Samo: At first, both of these seem like very bad decisions. If you were, in 2019, betting on 2019 or 2020, betting on OpenAI rather than Google, that's really not an obvious bet.
Nevin: Mm-hmm, mm-hmm.
Samo: And if in 2025, 2024, early 2025, or 2024 you were betting on Anthropic, not OpenAI, that's also not a very obvious bet.
Nevin: Mm-hmm.
Samo: Uh, but because this worldview is compelling enough, combined with the ability to raise money because of, you know, the Valley is, like, fairly good at funding new ideas- Uh, it was possible to start these new organizations. So let me put it this way. The, the [00:53:00] worldviews, ideologies, beliefs, and these, like, unusual assumptions about the world, they, uh, matter for just predicting what these individuals and companies will do.
Yeah.
Nevin: Yeah.
Samo: And I think it's very important, and I, and I think if you think it's, oh, it's all a cynical business plot to make, just make the most money you can, I think you're actually gonna make some wrong predictions, and ironically, might miss out on making money, too.
Nevin: Yeah. I remember people being very sort of concerned about Demis and his character, where it's like, oh, you know, he's so good at diplomacy, so we can never trust him, and he must be sort of lying about everything, and he's gonna try to be the god king, and that's not okay- Mm-hmm
and we need to do something about that. And, and I do believe that that, that is a big piece of what motivated Sam and Elon and others to, to start an alternative. And now you hear the, the people who are concerned about AI risk sort of lamenting the fact that, um, you have, you know, all these new companies [00:54:00] that are meant- Mm-hmm
to be the, the good one. Mm-hmm. And, and they're saying, well, that's just creating a worse race condition, and so actually they've made- Mm-hmm ... the situation worse. Maybe that's right, but I find myself being honestly calmed by the fact that, um, we're sort of in a multipolar-
Samo: Mm-hmm ...
Nevin: on a corporate level, um, AI- Right
race situation because it, to me, just feels like if, if you were Demis and you really were very far out ahead-
Samo: Mm-hmm ...
Nevin: and, uh, you know, especially if you'd stayed independent from Google, but even within Google, um, if there really was just no competitor, and the entire world was just focused on you- Mm-hmm
and your AI thing that was gonna be this crazy revolution and maybe have the crazy singularity, I feel like there's a sort of, uh, messianic or, you know, megalomaniacal mindset you could end up in-
Samo: Yeah ...
Nevin: that would lead to potentially very different behavior, um, than the situation we're in now, where it is kinda, you kinda have the sense [00:55:00] that it's neck and neck, and there's already been some sort of changes in who feels like a front runner a couple times or s- or so.
And so it's like, well, we don't really know where this is all going. We don't really know who's gonna have power over this or that. Um, and so somehow, to me, it feels like maybe the same humans who, in certain situations, would have potentially done things that we would find really distasteful, uh, in the multipolar situation, I feel somehow are held to a bit more accountability or, or, or just themselves have to be a little bit more humble about the situation.
What do you think about that?
Samo: I think, uh, you make a good point because it means that there's no, there's very little that can be done unilaterally, right?
Nevin: Mm-hmm.
Samo: And, uh, further, the asymmetry of information between the AI companies- People are circulating among them all the time, right? Right. Like, company secrets at most become, like, industry secrets, [00:56:00] right?
And very, very frequently any sort of new approach to, to improving the performance of these models ends up being widely adopted, ends up being, like, you know, quickly propagated. Um, there is this idea perhaps, right, that, oh, this just, like, increases the speed of it. Um, but I'm, I'm honestly, like You know, I'm, I'm sometimes not even sure that that is the case, right?
Where to me, if there was like a single company directing AI progress, you know, there's a lot of planning efficiency you could have done. There's a lot of financing that went into building data centers that just didn't happen, right? The data center construction was just canceled or written off, right?
There are inefficiencies with competition too, not just gains.
Nevin: Mm-hmm.
Samo: And if you had a single company led by a live player, I actually think they could go about as fast, and they would go in this like very [00:57:00] unconstrained way, uh, that you described. Um, now I do think that this type of competition makes one of the domains of competition also the safety of the models, or demonstrating some level of concern, or demonstrating some amount of caring for the wellbeing of the general population- Mm-hmm
because kind of the alliance of the general population with your company matters. Like a lot of the goodwill that Anthropic had, and maybe it's losing a tiny bit of it with the most recent release, um, is just this idea that they position themselves as the most benevolent company, right? Their marketing was like, "Keep thinking," right?
Mm-hmm. That was brilliantly done. And then also there's this, uh, approach to like, you know, selling sort of the, the, um, um, selling different services, right? There's a way in which all of these companies now have to give intelligence [00:58:00] to different sectors of society. They can't all just be like, "Okay, we're just gonna sell the super AI to the government and only to the government, and we're gonna take over the government with our super AI while we sell them the super AI."
Nevin: Mm-hmm.
Samo: I don't think that can happen as easily in a world where the general consumer has some s- some access to intelligence, and say, all of the Fortune 500 companies have some access to coding super intelligence. Mm-hmm. That actually means that actually the government is also constrained, or whatever political alliance you have, it's, uh, more equally constrained.
So I think that there is something positive there. We also are getting so much more experimentation with the interactions of human and machine than we ever could have hoped for. In that world of the like 1990s, early 2000s, with the brain in the mach- in the box, well, everyone's carrying that box- Yeah ... in their pocket.
We have a vast amount of data, right? Humanity, or by which I mean all these companies, [00:59:00] every company with an app, has a vast amount of data of how basically anyone interacts with AI. So we now know a lot about the interaction of humans and artificial intelligences. And I think we are learning surprising differences between artificial and natural intelligence that we wouldn't have learned otherwise.
Like, a lot of our philosophical concepts have been very clarified. So there's a weird way in which, uh, AI has almost made philosophical progress empirically, right? And I'm just waiting for sort of the, um, you know, research indicating on how do people feel about the personhood of their little chatbot.
Nevin: Mm-hmm.
Samo: I think there's super interesting psychological, uh, research to be done, even with just statistical methods, and I bet there'll be, like, super interesting correlations between different personality types or cognitive styles and whether they perceive the AI to be, like- Mm ... you know, conscious or not.
Nevin: Mm-hmm.
Samo: Yeah.
Nevin: Mm-hmm. Yeah. Yeah. It's interesting [01:00:00] 'cause I, I think at the beginning of OpenAI, Sam and Elon's statement was, you know, the way to deal with super intelligence is to give everyone a super intelligence- Mm-hmm ... on their phone.
Samo: Mm-hmm.
Nevin: And in a way, it's playing out in that direction.
Samo: Mm-hmm.
Nevin: Um, but, you know, there's always been the question of, like, well- If somebody does get to the point of super intelligence, are they gonna keep giving it to you so you can have it on your phone?
Right? And, and I think the, the mythos situation was maybe the first publicly digestible telltale sign of like- No ... oh, right. If it, if it gets powerful enough then it will be kept from me.
Samo: Mm-hmm.
Nevin: Mm-hmm. Right? And, and if, if that happened already in this way that, you know, seemed responsible but, you know, was just this border condition- Mm
well, what if there was something that was 10 times, you know- More ... more capable or whatever. What's, what's really gonna happen there? Um, and so I don't really feel like I know the answer. Have you th- have you guys spent time thinking through that eventuality of l- you [01:01:00] know- Mm-hmm ... what happens when, you know, the, the, the, uh, systems that are available, you know, internally within the labs and possibly to government, uh- Mm-hmm
agencies or companies who happen to be close to them or whomever, are just far beyond what is publicly accessible. Do you think that is the likely outcome? And if so, what do we do about that?
Samo: I think that those companies are gonna find it very difficult to hide that fact, if that's the case.
Nevin: I agree, yeah.
Samo: Mostly because of competition between the AI companies. So this is actually a good check on any sort of would-be super secret lab advancing vastly beyond everyone else.
Nevin: Right.
Samo: And, uh, I do think that there is a set of assumptions where in the world of compute, here's the thing, in a way, the scaling hypothesis is the most pro-human one.
Why? Because it means you kind of need the whole industrial stack of our civilization to build this. It's kind of a mega project. It's kind of like pyramid-like, right? Except, you know, presumably the pyramid does useful things. [01:02:00] Though, to be fair, when they built the pyramids, they thought they were immortality engines, so, you know?
Really? We'd probably, yeah, we'd probably invest 3 or 4% of GDP into- ... a giant, like, construction project that assures you an excellent afterlife.
Nevin: Yeah.
Samo: Right? The afterlife's forever. Beautiful. And not, and basically everyone who helps the Pharaoh build the pyramid, they get a share of that. They did bury the workers around it, and it was considered, like, beneficial for your own afterlife.
So
Nevin: it's, so it's not just- Yeah ... kind of slavery. You, you got a, you got a chance at immortality.
Samo: Yeah. Exactly. Okay. Exactly. All right. Exactly. It's just, it's kind of funny, right? But, and all of this stuff is, like, a little bit up for interpretation, but I think that there's a way in which, uh, the scaling of super intelligence physically necessitates all of this, like, deep interaction with all the levels of our society.
And actually, all these companies have quite a bit of power. TSMC has quite a bit of power over Nvidia. Nvidia has quite a bit of power over Anthropic. Uh, ASML has some amount of power over TSMC. The US government has a [01:03:00] lot of power over ASML, as they demonstrated in 2018 when they blocked the sale of ASML machines to China.
Mind you, had those machines been sold to China in 2018, there would be no US advantage in AI Right? They would have SMIC and other companies would have simply caught up to TSMC because, you know, of the people that work at SMIC, which is the Chinese mainland alternative, uh, those are retired TSMC engineers.
And also, of course, now in China, they sometimes hire the retired ASML engineers to reverse engineer a replacement for those machines. So that is, like, a rare case where I think the US government was actually, like, fairly proactive. So that was the- the first Trump admin. That was a surprisingly good decision on their part.
Nevin: Was it just luck, or do you think there was someone there who understood the stakes?
Samo: I think there was someone there who understood the stakes. Yeah. I think pretty straightforwardly, this is not the kind of thing you stumble into.
Nevin: Yeah.
Samo: Right?
Nevin: Yep.
Samo: Especially since ASML's not even an American company. It's a Dutch company.
But this just goes to [01:04:00] show how in an industrial society, anything that requires a complex supply chain automatically distributes power among many different stakeholders. And then okay, none of these stakeholders, these big companies, these governments, they're not individuals, but all of them have some things that they, um, require from the public, and if there's enough of them, then there's actually a large class of people who have decision-making power.
And, you know, those people have moms and sisters and brothers and sons and girlfriends and, you know, friends, so it ends up being, like, some push to benevolence. Yeah. Now, if it is pure transformer architecture, um, transcending that architecture for totally new things and seeing 10X or 20X or 50X improvements without the need for custom hardware, in that case, the sort of, uh, you know, singularity in a basement scenario becomes likelier.
Mm-hmm. And then the extraordinary leverage of individuals who, like you [01:05:00] said, might be totally, you know, they might even be benevolent individuals in most circumstances, but if you have the megalomania of like, "Oh, in this moment I'm determining the future of all of the universe-
Nevin: Mm-hmm ...
Samo: you know, where I'm just making a god right now," I don't know- I don't think most people can stay sane in that moment.
Nevin: Right. Yeah.
Samo: I don't think they can. It'd
Nevin: be a hard situation to deal with.
Samo: I'm not sure. Like, I wouldn't wanna be in that situation. Oh, yeah. Because I don't think I would be sane in that moment.
Nevin: It would be genuinely challenging to figure out the right thing to do. And yeah, in the, in the- whereas in the current moment, you know, the- the relevant leaders, um, uh, you know, they have so much public attention on them.
Mm-hmm. Right? Like, I- I- I think actually a lot of humans sort of, um, irrationally think that society is kind of always watching them more- Mm-hmm ... than it actually is. Mm-hmm. But in these cases, like no, we are actually really watching very closely. We are. Um, and so that- that- that's an interesting check. One, um- One sort of dark horse that [01:06:00] I'm...
I don't know if you've thought about, uh, I would even suggest it as a potential future research topic if you haven't, is, um, the hedge funds.
Samo: Mm-hmm.
Nevin: Um, so the quant funds in particular. Because, you know, what, what I find interesting there, I recently, uh, went to an AI conference, uh, an a- an academic AI conference- Mm-hmm
where the majority, I think, of the big sponsors were these quant hedge funds. Mm-hmm. Um, and I found that interesting. And I went... I just went to talk to their recruiting staff and, and engineers and so on who were there to try to understand what are they doing, why are they putting so much time and energy into recruiting AI, uh, researchers o- out of academia.
And they all have built their own or are building their own foundation models. Not language models. Mm-hmm. Like transformer models that are based on, you know, the enormous amount of data that they've, uh, amassed, uh, on financial markets over many decades. And, um, and it's clear that they are [01:07:00] getting returns from these given how much, you know- Mm-hmm
uh, they're investing in recruiting in this, in this domain, and the statements that they made when I asked them. So I just find that really interesting because they are... These are models that none of us are ever gonna see or interact with. Mm-hmm. Mm-hmm. Right? Whereas the... Whereas with the chatbots, they're all public.
Mm-hmm. We all get to use them, you know, the day they come out. You can go poke around and see what it does and so on. It's hard to even... I can't even explain, like, really what these transformer architecture models are doing. You know, are they trying to predict the state of the order book one tick out? Are they...
What else are they trying to do? Mm-hmm. I'm not really sure. I know they're generally trained on these market structure, you know, data sets, and they're, and they're helping to inform predictions. Um, but they're all... They're gonna be kept secret. And if you look at, you know, kind of the AI lab companies largely in San Francisco with everyone, you know-
Samo: Mm-hmm
Nevin: living with each other and going back and forth between companies and so on, as [01:08:00] you pointed out, there's not a big information wall between them. Mm-hmm. But if you look at the quant funds-
Samo: Yes ...
Nevin: they have, you know, they're set up in different... They're n- There are none of them here in California because- That's true
of the labor laws, right? It's, it's impossible to run, uh, that kind of shop in California because you can't have a non-compete that's- Mm-hmm ... sufficiently strong. Right? In the quant fund world, you, um, uh, you have this concept of garden leave where, you know, you, you leave a f- a firm like that, you have to basically be unemployed for- Mm-hmm
I forget what it is, a couple of years. Mm-hmm. A year. Uh, don't quote me on the timeframe. To basically wait, you know, the firm is just waiting for your knowledge to be irrelevant- Mm-hmm. Mm-hmm ... uh, before you can go work with someone else. Well- That might mean that the technological progress that occurs within, uh, one or the other quant fund ends up diverging quite a bit more- Mm-hmm, mm-hmm
um, from anybody else. And there's even just this culture of not- Mm-hmm ... sharing the information. You know, everyone is sort of practiced and bathed in saying, "I can't talk about that, I can't talk about that." Mm-hmm, [01:09:00] mm-hmm. And so, um ... And you know, they have in turn... You know, it's like yeah, you need many billions, uh, to scale up these, um, these supercomputers.
Um, well, they've got that, right? Mm-hmm. And they've got an engine to just keep turning that into potentially more and more cash. Mm-hmm. So I, I haven't actually looked closely at the numbers. I don't know. Like, you know, obviously with these lab companies IPO-ing around $1 trillion, maybe they're s- already in another league.
But still, um, you know, at least per head, these, these quant funds are extraordinarily profitable. So I just think it could be a- an interesting place that we're not drawn to look at, 'cause you never really hear about them or talk about them- Mm-hmm ... uh, that could end up, uh, leading... You know, w- what if one of them discovers, you know, uh, this chain of insights that leads to, you know, something that's self-improving?
And maybe we don't hear about that for, for quite a while, um, after it happens.
Samo: Well, I think it's certainly possible that other centers of AI advance will like, you know, come to surpass Silicon [01:10:00] Valley in this sense. And I think you're, um, right to consider that the very different initial culture really shapes the outcome.
Um, and you know, we have, we have seen some demonstrations of like, you know, having, uh, the ability to make AI progress even outside the narrowest network. We mentioned previously Demis. Uh, and you know, I think that there was a way that, uh, DeepMind was an example of a company that stayed in London partially to keep its advantage.
Nevin: Mm.
Samo: You know, maybe there's no explicit non-compete, but are you g- you know, really take your kids out of school, out of London, and move them to California? Mm. Like, probably not, right- Mm ... after a few years. And who else are you gonna work for in London on AI? Well, not- Right ... a lot of options there.
Nevin: Mm-hmm.
Samo: So I do think that there is, um, also the...
I think people underestimate intelligence as trained on data sets that aren't words, full stop. We're in a way a very literate civilization [01:11:00] of the word, where it starts with, you know, again, the scribes in ancient Egypt, the books printed in the 15th century, and now whatever our chatbots are texting us and what we text them back.
Um, we in a way overestimate the importance of the word, right? Just because it's so load-bearing to how we- Right ... organize our society. Everyone asks, "What is it like to be Claude?" No one asks, "What is it like to be a Waymo?" Right. Right? And there's- the, the Waymo's not less intelligent.
Nevin: Yeah.
Samo: It's arguably sometimes doing things, you know, the self-driving car is sometimes doing things that are more difficult-
Nevin: Yeah, and has-
Samo: than, uh, the
Nevin: LLM ... maybe a whole lot more integrated sense data.
Samo: It might even... Here's a fun one. I would completely believe a world in which the self-driving cars are conscious and the chatbots are not, because if it turns out that you need, like, a kinda sensory loop of embodiment- Right ... oof. Well, those things are embodied.
Kinda hard to argue they're not.
Nevin: And so what do you think, what do you think it is like to be a Waymo?
Samo: I mean, I think you just wake up in a second and experience the, like, [01:12:00] thrill of solving, like, visual puzzles, and, like, directing movement. And you kinda go to sleep when the car stops. Uh-huh. And then maybe you, you don't really remember it as you.
So I think the, the subjective life of Waymo- What's it like to be a Waymo when
Nevin: you're being remotely piloted?
Samo: That's a good one. I don't know what technically happens. I don't know what technically happens. Is the AI still issuing commands that are just, like, not honored? Or is the AI kinda, like, disinstantiated and there's just the human driving it, and it's reinstantiated with a new starting position?
So what I would say is that, you know, because they have to do decisions live, I assume that it's run, like, a single instance until there's, like, a stopping point, like the car stopping, the car turning off, or maybe a human taking over. Yeah. And then that's the, that's the shared thing, you know? I don't think they keep memories for the individual cars of all the places the individual car has been.
But again, I don't know, right? Uh, I do, I do think it would be kind of fun if someone wrote an essay, what is it like to be a Waymo, that has actually worked [01:13:00] on them.
Nevin: Yeah.
Samo: I, I want, I want some engineer to write that.
Nevin: Do you think that if we are creating conscious experiences in these things, we'll ever be able to tell?
Samo: I think we will. I think we will because we will, over time, develop a better and better understanding of what it means to think and what it means to process information. Like I said, there are philosophical concepts about language. For example, LLMs basically disproved Noam Chomsky's approach to linguistics, right?
Nevin: Hmm.
Samo: There are deep, uh, aspects of, um- There, there are philosophical concepts that we only clarified when we started materially investigating some of their correlates. Like, that's not the first time. Like, like let's... Even our assumptions about the universe or the nature of time, these are things that, you know, once were only pondered by philosophers, now are pondered by philosophers and physicists.
Um, and you know, there are potentially ways to [01:14:00] approach them. So I think that perhaps we will see... And by the way, this doesn't all go towards the obsolescence of philosophers. I actually think it's, in a weird way, a positive sign that some of the AI companies are hiring philosophers. I think it goes the other way, too.
Mm-hmm. I think they actually find technical use for that, uh, for that sort of, uh-
Nevin: Yeah ...
Samo: thinking.
Nevin: Okay. Let's talk a little bit about the market side of things. So everybody's wondering, you know, where do the valuations up and down the stack go from here? Um, you know, and there's kind of... In, in thinking about whether to invest and if so, how to invest in this whole AI build-out, to me the bigger question than, like, exactly which company to, to invest in is almost like is the whole thing overheated or not.
Mm-hmm. Right? 'Cause you could have a situation where it's like, yeah, this is gonna be real. It's gonna be huge. It's gonna revolutionize everything. And yet still you could [01:15:00] have speculation and valuations get way ahead- Mm-hmm ... and then a, and then a big sell-off. Um, or maybe that could happen multiple times.
Um, you know, we can see sort of the cyclicality in crypto as one, one template there, uh, where, you know, Bitcoin keeps going up, but it- Mm-hmm ... keeps going up and then down, and up and down, and up and down What do you think? Um, and, you know, if you have a sort of like direct answer or a hypothesis, feel free to share it.
I know it's a very tough question. Mm-hmm. If you, if you don't have a direct answer, um, I'm curious also to know, like, what do you think are maybe the two or three most important sub-questions that we should be asking- Mm ... in order to try to tell, you know, um, where the valuations go over the next year or two?
Samo: Mm-hmm. Mm-hmm. Um, well, I think that, I think that the most direct answer, perhaps the most interesting one now on the eve of the big three IPOs is that, you know, I actually expect they're gonna do quite well.
Nevin: You think so?
Samo: Yeah, yeah. There's-
Nevin: And we're recording this, I think, the day before the SpaceX [01:16:00] IPO.
Samo: Yes, exactly.
Nevin: So it'll- it'll probably, uh, be done by the time this comes out.
Samo: It- it'll be interesting to see either way. If I'm, if I'm wrong, like this would be a c- neat, a neat moment to keep in.
Nevin: Yep.
Samo: Um, I think that these IPOs are ... There is so much demand from not just the public, but from so many investors to get various kinds of exposure on AI, um, that I think that this is like, you know, gonna be their opportunity.
Now the question is, of these like IPOs, like I think almost the bigger question is between OpenAI and Anthropic, which one of those does better.
Nevin: Mm.
Samo: Because I do think it's been neck and neck. I think Anthropic's really played public perception very well. Uh, they produced very good models. You know, they had some good stuff with the pope, maybe some slightly less good PR with the throttling of Mythos into Fable, where, you know, if you're an AI researcher, uh, you're not gonna get the best thing the model can give you.
Nevin: Mm-hmm.
Samo: And that's, in a way, that upset a lot [01:17:00] of people on tech Twitter-
Nevin: Mm ...
Samo: right? Because previously it was this kind of like pro-social for the industry, and then that seemed, to them, antisocial for the industry because the tools they were hoping to use- Mm ... to make the next wonderful thing-
Nevin: Mm ...
Samo: uh, are intentionally just not gonna be fit for purpose.
Mm. I, of course, see the other side of it if you like equate really advanced AI with like bioterrorism. Y- you know, so there's some safety stuff there.
Nevin: Yeah.
Samo: Um, and but, you know, I- I- I think that it's- it is not yet clear those things in fact matter that much. I think if you're a normal person and you're hearing about AI, you still think ChatGPT.
So even though Claude is like perhaps like now the more like premium brand-
Nevin: Mm-hmm ...
Samo: people love Chat.
Nevin: Yeah.
Samo: Right? And they didn't stop loving it, and people have known about OpenAI, and I think when it comes to these sort of IPOs, often it's just like, have you heard of the company?
Nevin: Right. Yeah.
Samo: I don't think anyone is gonna get [01:18:00] in trouble for buying OpenAI stock.
Nevin: Right
Samo: No one's gonna get in trouble. No one's gonna be like, "Oh, you bought OpenAI rather than Anthropic? Well, you're gonna be fired."
Nevin: Right.
Samo: So okay, they're already blue chip companies, even though maybe we shouldn't think of them that way. They don't have a proven track record, but we kinda treat them that way.
People do treat them that way. Yeah. And then with regard to SpaceX, it's also kinda hedged on the AI boom. There is a l- look, SpaceX would've been a great IPO even had the AI boom not happened.
Nevin: Right.
Samo: It was already profitable. It did these amazing things that people were super proud of, like the Starship launches, the, like, catching the rockets with the chopsticks sort of when it was, like, returning.
And, uh, people are excited about, like, Starlink. They see Starlink. They now use Starlink when they fly over the ocean. So I think SpaceX, easy, just, just buy it. I would just buy it. I mean, this is not investor advice, et cetera, et cetera. Obviously. You know what I mean? Yeah. [01:19:00] Obviously. But it-
Nevin: Neither of us are investment professionals-
providing investment advice.
Samo: Yeah, yeah. Um, uh, but you know, I think the more interesting one is what happens after the IPOs.
Nevin: Mm-hmm.
Samo: Like, can it still keep on growing?
Nevin: Mm-hmm.
Samo: And you know, I think the answer is to a great extent, uh, to a great extent there is still a lot of growth in the system left. Even when it comes to financing these data centers, in a way the IPO of Anthropic, OpenAI, and to a lesser extent SpaceX, is needed for them to keep up with Google.
Mm. Google has mostly been funding its build-out from its giant cash cannon of revenue.
Nevin: Right.
Samo: For decades, Google kinda didn't know what to do with their money, and they were already a giant company.
Nevin: Right.
Samo: Apple, in a way, didn't know what to do with their money, though arguably they still don't know what to do with it.
They, they don't have a clear next step in the AI world, and it's not clear that they should be pursuing a next step rather than simply, I don't know, focusing on moving their supply chain [01:20:00] from China to India and continuing to sell iPhones. Maybe they're just a mature company.
Nevin: Yeah, it's like Apple is like Europe or something.
Samo: Exactly, right? It's like, it's nice. You like to have it- Yeah ... but it's, it's kind of a product like a BMW at this point, right? I, I, I could almost offend some people when I joke, is it even a tech company? Right? It makes technology, but, like, it's not revolutionizing anything anymore, and hasn't for a long time.
Uh, and you know, they really could have done something there because, you know, everyone remembers still talking to, uh, their virtual assistant, right? This was before LLMs. They could have quickly upgraded that- Yeah ... had the first AI-powered system. But, uh, now it just, it doesn't stand out.
Nevin: Yeah.
Samo: It's like the same as what everyone else has.
Nevin: Right. Right.
Samo: Uh, to say one more thing though about what investments, I think that because the, let's call it the [01:21:00] legacy software industry, so like we could think of it as SaaS, we could think of it as B2B, all that world, right? That will partially be refactored for real, I think, by AI. I think the AI coding stuff is quite real.
Nevin: Mm-hmm.
Samo: But that also means that a lot of these revenue streams previously, um, they're just gonna be directed towards AI. Sometimes the prices will go down, which is good for consumers. Uh, if your company needs custom software, that service might actually be cheaper today than it was two years ago.
Nevin: Mm-hmm.
Samo: Right? If your company needs custom, uh, software to manage its machines, or something about your workflow, or something about how we do things in-house, all of that plausibly will become much cheaper, right? Create a lot of value. Still though, all of the big companies, Meta, Google, and all of those existing ones, so not even the, the new AI companies, um, they will invest their money into AI So where do we go [01:22:00] from there?
I think that just keep investing down the whole supply chain as this, like, massive demand increase goes down through TSMC and sort of disperses into the rest of the economy.
Nevin: Right.
Samo: For example, you know, maybe buy, maybe buy Mitsubishi or Samsung, and not because they're live player-led companies, but just because they'll be the ones providing, uh, the...
They'll be constructing way more natural gas turbines and other machinery that data centers need. Maybe buy, uh, maybe buy, uh, the German manufacturers like Carl Zeiss and so on, right? Maybe buy, uh, even stuff that's just correlates of economy. Like again, now people are like, "Hey, I should have invested in Korea."
Well, maybe you should invest in the Netherlands. It's almost a narrative violation, right? We assume all of industry is in East Asia, but actually the Netherlands is right there embedded in that.
Nevin: Yeah.
Samo: So if the East Asia thesis is correct, then the Ne- the Netherlands does well [01:23:00] as well.
Nevin: Mm-hmm.
Samo: Uh, I don't know, maybe even exotic things.
Maybe you actually just go look at all of the, uh, ASML, uh, facilities you can find and just start buying all the houses around those facilities. And you're investing in Dutch real estate.
Nevin: Right.
Samo: Right? But that, that's what it means to take this stuff seriously.
Nevin: That would be a funny REIT, a sort of publicly-
Samo: Yeah
Nevin: traded just, like, you know, houses near AI stuff.
Samo: I mean, that's a good investment strategy in the Bay Area. Did you see that joke about, like, you know, the Anthropic engineer's like, "I offer a million dollars for this house"? The AI engineer is like, "I offer $2 million." And then the AI, the OpenAI engineer again is like, "$3 million."
And then two baby boomers, "Finally, our foresight and hard work is paying off"- ... right, when they're selling the house. Yeah. 'Cause, uh, I think if you are a... If you own a house in the Bay Area and you're selling right now, you're pretty happy.
Nevin: Yeah.
Samo: And if you're a landlord renting, you're pretty happy.
Nevin: Okay. But so then if, you know, if we look [01:24:00] back on this conversation A year or two from now and, and it was the top of the market now.
Mm-hmm. What's the most likely explanation for how that could be?
Samo: Uh, all of the capital that can be invested into developing new technology has been. And actually, the, uh, m- the story in that world is that the mainstream financial institutions, their books are actually a little bit worse than they say they are.
Nevin: Mm.
Samo: They actually can't afford to buy into it because the technology they believe is working. There's some, like, murmurs in the back that, you know, maybe some of the, uh, contracts will not be renewed, but really that, that kinda only matters for Anthropic and OpenAI. It doesn't even matter that much for SpaceX, right?
SpaceX can start contracting, uh, building data centers instead of for Anthropic, for Google or whatever, right? Like, there are other options available to them. Um, I do think there will be AI [01:25:00] bubbles just because that's the nature of capitalism, and I think people will get ahead of themselves. I don't think we're currently ahead of ourselves because if you look at what people are saying, it's really moderated.
It went from 2023, this is gonna take all your jobs to, actually, this is not gonna take your jobs, it's gonna, like, just make your job better. It went from, oh, actually, you know, uh, we're going to, like, solve all diseases to, well, we're working with the regulatory process to help accelerate drug development and testing.
Mm-hmm. And it went from, well, you know, maybe this is going to be ... Maybe this is a desperate race with China to, well, you know, Chinese models do some interesting things and we can learn from them, but we're still ahead.
Nevin: Mm-hmm.
Samo: So there's all these ways in which the pressure has been let off slowly by the CEOs themselves, and perhaps that is because they have to appeal to, like, let's call it less technophilic capital.[01:26:00]
People talk about smart money and dumb money, but I think they should talk about, uh, capital that loves technology and capital that hates technology.
Nevin: Mm.
Samo: And for example, there's just, uh, for ... I think Wall Street has a slight anti-AI bias Why? Because the smartest person is the person that proves that someone else is wrong.
So when you make an investment on AI, you're not proving anyone wrong. But if you make a, a bet to short AI, well, then you're proving the AI guy is wrong and showing that you, the finance guy, is actually the smartest guy. Mm-hmm. I think that's very emotionally compelling.
Nevin: Mm.
Samo: Right? But, you know, the real reason is it's, it's, it's gonna be bankers who have never looked at AI, who don't even find it that useful, who are looking at the returns their pension fund or whatever is supposed to have to make it solvent, which is basically impossible because the demographics to make pension funds solvent are gone.
They're nowhere in the world. [01:27:00] And they're gonna be like, "Yeah, I'm not gonna gamble the house on this." Right. And then, you know, in that case, I think we'll still get pretty good valuations. Are you going to tell me that techies, for example, who have worked in m- in Meta for 10 years, are you, are you gonna tell me they're not gonna buy some Anthropic or some SpaceX, some OpenAI?
We're gonna see a great portfolio diversification.
Nevin: Mm-hmm.
Samo: Honestly, OpenAI people are probably gonna buy some SpaceX too.
Nevin: Right. Right.
Samo: Right? They're bro- probably they're gonna go from owning 100% just, you know, most of your net worth's OpenAI to trying to trim that down a little, trying to follow portfolio logic because the future now is less clear than it was in 2022.
Nevin: Yeah, yeah. And it's an interesting thing about where the capital's gonna come from. I have definitely seen direct evidence in crypto of some very large holders basically getting entirely out of crypto- Yep, yep ... and they're now entirely in on AI. [01:28:00] Yep. Um, and they were early to that trade. Um, and if, if you look at crypto markets, you c- you can see that that's probably happening- Mm-hmm
on a broad scale. Um, you know, that kind of makes sense 'cause the folks who are interested in volatility and making huge returns- Mm-hmm ... are gonna rotate to whatever the shiny new thing is. Um, but it's an interesting thing about beyond crypto, where else is the capital gonna drain out of? Mm-hmm. You know?
Have you detected that in any other industries in a clear way, or...?
Samo: Hmm. Well, I think that, you know, I, I really think people are very pessimistic on the rest of the economy.
Nevin: Yeah.
Samo: I think we're sort of almost in a ... I think there's an aspect of an undeclared recession that's happening under the surface. I think without, without the AI boom, um, I'm not sure the economy would be doing that well.
Nevin: Right.
Samo: Just consider, for example, this massive energy shock we absorbed through the Straits of Hormuz, uh, you know, being closed for so [01:29:00] long. That really depleted strategic oil reserves in a number of countries and increased energy prices yet again. One big aspect is, and you know, this is like gives me no joy to say this because, you know, I, I, I, uh, moved from Europe to States 2014.
Um, uh, I think Europe is gonna see capital flight. Hmm. Just, just on a macro level- European ... a few exceptions.
Nevin: European companies generally.
Samo: Just European companies generally, unless they're touching the supply chain-
Nevin: Yeah ...
Samo: uh, going into East Asia delivering electronics.
Nevin: Right.
Samo: So I think that a lot of these old companies, uh, in the US also, if they don't have anything to do with the AI boom, they're just gonna get sold.
Yeah. So I think actually tomorrow we might see a number of companies dip to be basically dropped for SpaceX.
Nevin: Right.
Samo: At this point, for example, you know, I don't know, if you're holding some Boeing, why?
Nevin: Right.
Samo: Right. Like why?
Nevin: Yeah.
Samo: Right? [01:30:00] Or if you're holding Airbus, it sort of makes more sense than Boeing, but if you were holding it for the space stuff, y- no.
Just- Yeah ... just buy SpaceX.
Nevin: So depending on-
Samo: Yeah ...
Nevin: how things go, this, this year might be a, a good time to get a good price on Apple stock and a house in Europe.
Samo: Yeah. Yeah, actually.
Nevin: Maybe the next couple years.
Samo: Yeah.
Nevin: Um, one last question on this topic Is there a bunch of leverage, financial leverage building up in this whole system?
And, or, and have you noticed any, like, hidden pockets of financial leverage building up in the system that could blow up? Or, um, have you not run across that in your analysis of these companies?
Samo: I do think that they have been very, they've been insanely optimistic, uh, in a way that's been justified about the likely revenue growth, and the financial architectures have been very, um, very [01:31:00] creative.
And I think that, uh, I don't think any mistakes have been made so far, but a mistake is inevitable at some point if they continue at this pace. Because the IPO isn't the end of the story, right? The IPOs are just going to be this big capitalization event, and after that, the demands of wanting to continue to climb the curve of scale, um, while...
And of course, being the first to do it, those are gonna continue. So I think that for now they have been restrained enough that there is no incoming financial meltdown. But I think that they are perhaps, like, a year or two away from making such a mistake. Yeah. Because, uh, I think because consistently getting away with risk when dep- you know, deploying hundreds of billions in these various exotic deals where your customer is your lender, is your investor, is your customer-
Nevin: Mm-hmm
Samo: that's, someone's gonna just make a mistake.
Nevin: Right. There's [01:32:00] gonna... Yeah.
Samo: And then that shock is gonna spread through it, and some of the companies will have slack and others will not, and it'll be a big scandal.
Nevin: Right.
Samo: I don't think that's happening right now, but, um, you know, if, if we were wrong, uh, a mediocre IPO, which, uh, uh, IPO that's good but not good enough to overcome that, um, I think we'd start seeing problems three to six months from now.
Nevin: Yeah. Okay.
Samo: Because that's sort of like the timeline on which, like, I don't know, like the data center investments and other things come in. I think we should also here at this point ask ourself if Jensen is not at this point something of a central banker in his own right. Right. Where I'm sure if he doesn't sign off on certain deals, they don't happen.
So maybe I have a lot of trust in Jensen, in that he can punish bad actors in the ecosystem.
Nevin: Yeah, a central banker with no particular mandate.
Samo: Yeah.
Nevin: So-
Samo: I mean, the mandate is to increase GPU demand as much as humanly possible.
Nevin: [01:33:00] Right. Right. Right. Yeah. It's like central banks have to deal with, you know, two or three, two or three forces.
He's, he's got, you know, the profit motive and then any, any of his own ideological leanings.
Samo: Well, certainly once you're deploying hundreds of billions of dollars, you're, like, stabilizing whole financial systems- Yeah ... whether you want to or not.
Nevin: Are you comfortable sharing what you are invested in, in the space?
Samo: I mean, um, you know, I think that- You know, I think ASML is still a good idea.
Nevin: Yeah. And
Samo: I think, uh, I think the, uh, AI companies themselves are good ideas. Uh, but I think companies such as Carl Zeiss and others are great options too.
Nevin: Yeah.
Samo: For, um, anyone that wants to, like, construct a good portfolio, I recommend they check out the Bismarck Brief.
Nice. Uh, you know, my personal investments pretty much track what we look into and what we arrive with good conclu- where we arrive to a positive conclusion about the company. You can, uh, read them and, uh, that gives you, that gives you that information.
Nevin: And then, you [01:34:00] know, in the AI supply chain, the sort of supply chain for the build-out, um, you know, there are many different sort of verticals that people have started to focus on.
There's photonics, there's, you know, uh, power generation, there's the underlying energy sources, there's sort of the neo cloud companies. Mm-hmm. Um, uh, and Reserve is gonna release a, a series of thematic DTFs- Mm-hmm ... that track those. Um, is there a sector, uh, of that whole supply chain, sort of a level of it that you think, um, is interesting that people haven't really quite recognized yet?
Samo: Yeah. It's straightforwardly energy. Any company that is, like, doing anything related to liquefied natural gas infrastructure globally is almost, like, right now experiencing, like, massive, massive, uh, demand- Hmm ... for their services. Uh, this is why I'm optimistic to, to do contrary to my recommendation, or not recommendation, the observation that [01:35:00] capital flees Europe.
Uh, I think TotalEnergies, which is by revenue the largest French company, is gonna continue doing extremely well. Hmm. Because you both have geopolitical forces such as the, you know, US-Iran conflict, the Russia-Ukraine conflict on the one hand. Persistent high demand for natural gas from places like China, where the economy only keeps growing, right?
Uh, and this is by the way, despite them, um, kind of getting out of oil and of course they build their own coal, nuclear, and everything else. Um, together with the massive demand of the data centers. The data centers will really increase the amount of natural gas we burn. So that means the US will either frack for more of it, or it'll export less of it, and that means higher natural gas prices once more in Europe, which means also again, higher natural gas prices in, in China and so on.
So really, if your company's providing natural gas globally, it's gonna do really [01:36:00] well, and I think people in the AI space are not really used to thinking about those energy economics. And you're sort of like, "Hey, why should I invest in this company that's like 100 years old, a French dinosaur?"
Nevin: Right.
Samo: Uh, well, it's the kind of the only company that's not doing that in the Middle East, right?
They have, like, a big natural gas field in Mozambique, and they've gotten into the AI deals in the US. So reasons to be optimistic. And there's other energy stuff as well where there are reasons to be optimistic. I think on a 10-year timescale, the modular nuclear reactor thing will probably work My only caution there, uh, would sort of be that maybe China will be the place to do it because they are, have just so much experience building so many new nuclear power plants.
So they're building up the expertise, just the technical expertise, to design completely new types of reactors as well. And why would the modular reactors be potentially eventually a good idea? Well, the data centers are straining the power grid so much, and of course they're gonna increase in size, but you still might want small [01:37:00] power plants that are off grid, right?
Off-grid to just start providing the base load power for your chips.
Nevin: Very interesting. This will, um... I think that, uh, that take will appeal to, um, the audience that, uh, that we're thinking about with The Network Fund. You and I talked about this- Yes ... idea of The Network Fund a couple weeks ago, um, where, you know, it's, it's...
So the concept of The Network Fund, for our viewers, is, um, that over the past few months we've seen a bunch of people in public- Mm-hmm ... make very good calls- Mm-hmm ... about, uh, which companies are gonna do well in this whole AI build-out.
Samo: Mm-hmm.
Nevin: And, you know, you have folks doing research that's very similar to the, the research and analysis happening inside of the big AI-oriented- Mm-hmm
hedge funds like Situational Awareness, but just posting the results on Twitter.
Samo: Mm-hmm.
Nevin: And, and having this [01:38:00] public discussion and, and making, you know, similar trades to what the hedge funds are making, um, each individually. And so there's this question of like, well, if the research is public, then maybe as soon as it's out there, it must be, you know, by definition it's immediately public and therefore it should be priced in.
Samo: Mm-hmm.
Nevin: And so, you know, there's this question of could you possibly run a fund that's based on, uh, public information and- Mm-hmm ... public research and coordination rather than private? And the bold prediction of the Network Fund is that you can. Okay. That actually you can do perhaps equally well to a private hedge fund by crowdsourcing the information.
Mm-hmm. And the way that we plan to do that is, um is to track the public calls made by analysts on first X and, and later perhaps other social networks, and calculate their implied track records. Mm-hmm. [01:39:00] And then basically you can decide who to listen to and who not to listen to based on their actual performance over the course of time.
And if you track them in an ongoing way- Mm-hmm ... even if they delete a tweet or whatever, well, you still, you still tracked it as it happened. Um, and then once you've, once you have a system for weighting, uh, the different analysts, then you can take their current calls, their current sort of open, open, uh, positions, uh, that you at least can infer from what they say, and combine those together weighted by track record.
Mm-hmm. And so then the idea is that this meritocratic approach, um, you know, the, the huge advantage that it has over any particular hedge fund is that anyone in the world can start doing research and make predictions, and that will be picked up because- Mm-hmm ... it's publicly visible, and you'll start to pay attention, and you pay attention in proportion to how well they're doing.
So maybe someone does well research and understanding a certain part of this particular build-out, and then maybe their expertise is exhausted and they start making bad calls [01:40:00] because now it's about energy. It's not about- Mm-hmm, mm-hmm ... you know, the, the, the particular chip companies anymore or something.
Well, as long as you can pick up on that fast enough, uh, you stop listening to them and you start listening to who's making, uh, better calls. So that's the concept of, uh, the network fund. And I guess, um, I guess I'm curious just to, to ask, you know, what do you think of that? What do you think of that idea?
Samo: Well, I think that would, uh, actually massively improve investments. Then the reason is I think a lot of the people on Twitter just know what they're talking about much better than, uh, a lot of financial workers. And honestly, even some of the people working in tech who are, say, software engineers who don't have this, like, basically special interest in niche industries.
Nevin: Mm-hmm.
Samo: If you were sort of like a Twitter anon with a special interest in niche industries, you actually were doing pretty well. Mm-hmm. And then if you had some, like, extra additional insight or some connection with the AI labs in any way, um, you often would do even better. And if we add on top of [01:41:00] this, uh, sort of any sort of perspective on, like, society or the macro environment, you know, that's a small number of people Um, but they are active often and they're talking about this stuff.
Like, there are definitely many Twitter accounts I've, like, learned a lot from. So I would not dismiss the X.com poster with the anime avatar and Chinese characters- Mm-hmm ... in their name, right? Like that genre of poster-
Nevin: Yep ...
Samo: is now, like, you know, it's of course by now been imitated, et cetera, et cetera. Uh, but like spotting this, like, let's call it financially under-credentialed talent where, you know, if these people were well-positioned, they would be paid extremely well.
Mm-hmm. But then they would be in New York and they couldn't tweet about it.
Nevin: Yeah.
Samo: So I'm not even saying that there aren't many people on Wall Street at the great firms that know this. They're just never gonna talk about it. Right. And they're never gonna talk about it to you, and they contractually can't.
And anyone who is talking about it publicly, well, if they know what they're [01:42:00] talking about, by definition they're kinda under-credentialed, right? They're underemployed. Right. So in a way, uh, I feel this concept harnesses that intelligence-
Nevin: Yeah ...
Samo: and makes use of it. And by the way, isn't that interesting? Our society still has over-supplies of natural intelligence that have not been tapped fully.
Nevin: Oh, that's a good point- It's,
Samo: we- ...
Nevin: at a time like this.
Samo: Yeah. We don't always efficiently use intelligence in our economic system. Yeah. And, uh, I do think that eventually we will f- reach a point where maybe the supply will match the demand for artificial intelligence. It's just for now, I think demand still exceeds supply.
Nevin: Right. Right. Do you have one big question about the Network Fund?
Samo: If I were to have, like, one big question, right, it would be, um, well, what are the reasons for people to play along with the Network Fund- Yeah ... and, like, wish to, you know, boost it and w- wish to participate or give their expertise to it?
Nevin: Yeah.
Samo: Does it provide any incentive for undisclosed intelligence to become, like, you [01:43:00] know, open to the market- Yeah ... visible intelligence?
Nevin: It does. We're gonna play with that. So the plan is that, uh, if an analyst's predictions, uh, filter through the algorithm, which by the way is implemented by a decentralized network of oracles, um, if, if they end up influencing any particular rebalance-
Samo: Mm-hmm
Nevin: then we assign them, uh, some credit in the system, effectively. Mm-hmm. And they can choose whether to claim that or not. Mm-hmm. Right? Because some people, I think, are just, you know, they put things out there. They have their own, you know, they get their kicks from putting information out there. They'll be able to basically renounce and say, like, "No, I don't want any compensation," but, you know, the, obviously it's public information, so we can look at it.
But if people choose to, they can cl- they can claim those credits, um, and then basically have a portion of the fees that are- Mm-hmm ... collected by the Network Fund. So-
Samo: Mm-hmm ...
Nevin: it might- actually draw people out who are not currently publicly posting Mm-hmm because they'll see, like, well, [01:44:00] if I, if my track record had been public, here's, you know, uh, how much I would be influencing these calls, here's how much money Mm-hmm I would be making.
Mm-hmm. Um, and people can do that whole thing from anywhere in the world. You know- Mm-hmm ... you can post from a phone, whatever your ideas are, um, and, uh, and claim those credits, and then participate in this decentralized network fund. So we'll see if people wanna do that. I
Samo: think, I think they probably will.
Nevin: Um, I hope so. Um, okay, last question. Let me ask you something kind of fun, and then we'll wrap. Um, zooming out from AI, um, on the topic of, you know, your, your research and thinking over the past 20 years, what is one beautiful thing or fragile thing about our civilization that you think most people never notice?
Samo: That's a very good question. I think that there's something very, um- I don't know. I think there's something [01:45:00] very, very precious about, uh, the sort of truth-oriented aspects of our society. I think how you can actually still go and live the life of the mind in our society, I think that's really underrated.
If you want to just be, y- you know, learning new truths about the world-
Nevin: Mm-hmm ...
Samo: and you just want to understand it and you wanna perceive it, our society might today still be one of the best societies for that ever.
Nevin: Mm-hmm. Mm-hmm.
Samo: Right? Mm-hmm. We both have the means to access all that information, and there are still some niche careers, which I admit are difficult to get into, or there's a lot of competition and many of the people in them might not care, but they afford the opportunity- Yeah
for, like, a life of the mind. And I think, you know, people sometimes say, "Oh, you know, perhaps we f- failed as far as the Enlightenment project went." Uh, but there's one way in which the people in the 18th century would have actually thought this is great. I think this is one way in which we've fulfilled the aspiration of Enlightenment civilization, of the, like, encyclopedists, the letter [01:46:00] writers, the Republic of Letters people, the philosophers.
Uh, if you are motivated to seek, like, just this kind of, like, you know, illumination, uh, you can get as much of that as you would like, and that's- that's kind of beautiful and I think, uh, but could go away, right? Could go away, especially if, you know, we were to turn the chatbot in our pocket from our servant to our jailer, right?
The- the- the sort of... I do think that the- the freedom of thought aspect of our society is very fragile, and, uh, it'll- it is always perhaps under threat by mostly political, sometimes economic and cultural forces.
Nevin: Well put. Samo Burja, thank you for sharing some of your illumination, uh, on this first episode of Buying Into the Singularity.
Thanks for your time. Yeah. Much appreciated.
Samo: [01:47:00] Pleasure.