He Comes Promising AI Freedom For All

If a set of companies control the extension of your brain and it's like the collective extension of our brain, then they could just like steer it to think a certain way and now they kind of control you by proxy. And to me, I think that's like, that's fundamentally risky. And so we kind of have to arm the rebels.

We are joined today by an internet legend and possibly a computing legend in the making, Guillaume Verdon. Although I hear you say Verdon.

Verdon. Yeah.

But you're French-Canadian, right? Or—

That's right.

Yeah. So I don't know. How do you prefer people to say it?

Yeah. I mean, the proper way to say it is like Guillaume Verdon.

Okay.

So it's like French, but—

I was going to lean in, but then I heard you say it somewhere else that I didn't know. I didn't know if I should be leaning in. And also known as Beth Chazos.

Yep. Yep. That's my pseudonym.

Yes. Which we will, we will get into. And just before we get too far, I want to thank E1 Ventures for sponsoring our podcast. We actually have E1 Ventures representatives here today to watch the show. So they've been kind enough. To sponsor our show. And I guess maybe full disclosure, this was a coincidence, but I think you're actually an investor in your company, in Extropic. I didn't know this was happening when we invited you on. There you go. Well, thanks so much for coming in, man.

Yeah, super happy to be here. It's an honor.

I feel— oh, that was way too far. I feel like there's two giant buckets we can get into, which is obviously the EAC, the Effective Accelerationist Movement, and then Extropic, your company. And they both take some unpacking. Why don't we just real— we'll come back to it more in more detail as we get on. But let's, let's do Extropic a little bit first, just to give people a frame of where you're coming from. And I'll take a crack and then you can fix everything I mess up. But Whole new field, whole new approach to computing, definitely angled at AI, but not just AI. And it's, so it's called thermodynamic computing.

That's right.

Different to quantum, very different to our traditional bits. Is there a way to explain it that just about everybody can understand?

Essentially what we're doing Um, is we're, we're creating new primitives, uh, in what is called stochastic electronics or bio-inspired electronics. Um, so it's new silicon primitives. It's not just an architecture. We're not just taking floating point arithmetic and memory and changing the configuration. You know, that's usually an architecture that's like many existing companies are proposing. For us, it's like a new paradigm. From the physics of the electrons all the way up. So from the algorithms you run on it to the compiler, to, to the way the problem's embedded into the physics of electrons, right? And so it just makes a lot of sense that as the workloads, the typical workload you want to run is a probabilistic workload, right? LLMs are like virtual probabilistic computers that you would build a hardware that is a probabilistic computer because it's a much more natural fit. It fits like a glove, right? And once you, once you really understand physics and once you really understand AI algorithms, it's like you can't unsee how much of a perfect fit this is. And that's where the extreme conviction comes from, that this is the most natural way to do this algorithm. And I think, I think our results are going to show that over time. Of course, you got to start with, you know, first principles, some mathematics, then some models, then some simulations, then some early experiments, and then you scale it, and then you show a scaled system to people, and then they really believe, because then you're no longer telling, you're showing, right? And that's the sort of transition we're going through this year. And that feels really good because, you know, when, um, you know, when you have an idea that violates people's priors, right, about what is possible with today's technology, due to the free energy principles of your brain, actually. Your neurons want to preserve themselves. They're kind of greedy. And like updating your world model massively is too costly. So your brain just literally filters out ideas or like says like the person proposing them is crazy. So that's been, you know, the story of my life for, you know, 2 years. I mean, there's a lot of supporters, but there's also a lot of doubters and it makes a lot of sense because it's like, you know, overwhelming, you know, this is going to be an overwhelming change to computing and they want overwhelming evidence, right? But those that can be convinced earlier are going to get in earlier and are going to start building for this paradigm and are going to be ahead of the next wave, right? And that's like, that's the history of every paradigm shift in technology. Right. And I guess like, um, I exude a lot of confidence because I've thought about it for 10 years. Right. And I burned every, I burned all the boats going back to quantum computing. I could have worked anywhere in quantum computing. And I left that whole career and went all in on this from first principles conviction. And now the experiments are starting to back what we've been saying. Um, and so it's a really exciting time.

Well, and I mean, it is high risk because I, you know, I, I've covered many, many, many semiconductor startups, most of whom are like vaguely in some, they're in the ballpark, right? And even that stuff is risky. Anytime you're asking people to do new software, to make big adjustments, it's always, you know, even when like Intel was moving away from x86 to try this new EPYC architecture with the Itanium chip, I mean, that was like high risk and you're still in vaguely traditional computing, but There's, you know, somebody has to do something. You can make an easy argument for that. I just was at Stargate yesterday.

Oh my God.

And, and, you know, I'm looking at this like 1,000-acre site is gonna be pulling in more power than San Francisco by, by I think multiple times. Um, each building I think is $50 billion, of which 60% is the GPU cost and they want to make 10 of those buildings, right? And GPUs, um, I mean, it's, it's like kind of amazing that we have this video game chip that ended up running AI, but they got lucky. It's also like not hard to argue like, uh, this is not a purpose-built chip for this job, right? I mean, there must be a more efficient way to do this, you would think, if you really, if you really put your head to it. So, so like all of that makes sense, but Okay, you guys have only been around, I think, for 2 years.

2 and a half.

2 and a half. And you already have, I mean, we have silicon sitting on our table. Like, where are we at with the company?

Yeah, yeah, yeah. So, you know, at first, you know, we looked at the history. I've been following sort of physics-based computing for 10 years, right? So there's quantum computing, there's like neuromorphic, there's been attempts at probabilistic computing, and there's been, Uh, you know, everybody has a different name for their paradigm. There's been many attempts, right? And typically there's always a risk that they don't kill, right? One of them is like, they don't know how to program the device, right? So they're, for example, neuromorphic. They're obsessed with biomimicry and they're just like, hey, we made a neuron that it spikes like a real neuron. It's like, great. How do you program it? It's like, well, we don't understand the brain, how it learns. And so we don't really know how to program it. You can kind of program it, but So we had to kill that risk first. So we worked on how to program these devices, how to map algorithms that have been around for 40 years, statistical inference algorithms, Bayesian algorithms, and map them to the physics of what is called electron diffusion, right? So diffusion is like a generalized Brownian motion. Everybody's maybe familiar with Brownian motion. You know, if you drop some flecks of dust in in some water, they're just like diffusing around, right? It's like a form of stochastic mechanics and Einstein won the Nobel Prize for it. Diffusion is just like, you not only have programmable drift, right? Of like how much the particle moves on average, but you program like how much noise there is over time.

Okay. And so tell me if I'm off or not. I mean, typically semiconductors, we're talking about electrons being like on or off, but you're in this murkier.

Yeah. So it's like a cloud of electrons and we kind kind of control how it diffuses, right? Whether it becomes sort of more spread out or it concentrates or if it's moving on average. And by controlling that over time, we can basically create this programmable blob of electrons. And, you know, many algorithms can be phrased as getting an answer that's a certain blob.

Yeah, like a probability.

Yeah, probability. So, you know, when you work in finance, like in finance and Wall Street, I don't have to explain what we're doing. They all run these algorithms called Monte Carlo algorithms, you know, from Monte, you know, casino, uh, Monte Carlo. And, uh, you know, these algorithms were originally cooked up during the Manhattan Project to, to, to do simulations of, of all sorts of systems, but now they're used for financial simulations and, and, and also still all sorts of simulations of all kinds, but we can run those algorithms basically very tightly in all just pieces of nature that we add some little control nodes and we kind of steer it, right?

Are you steering with electricity or?

Yeah.

Yeah. Yeah.

Voltages. So we have control voltages and then by changing the control voltages over time, you change the dynamics of this dance, right? I think, I think what, what, what is really cool about the timing of all this, right? We could have built, tried to build this company years ago, but nowadays there's all sorts of workloads like diffusion models, right? Like Midjourney and I guess DALL-E and ChatGPT and xAI. They're all, they're all using this algorithm called diffusion, denoising diffusion, which is actually using AI to go backwards through a simulated diffusion process., right? And so our algorithms that we want to run are literally named after physics that our hardware runs. And, you know, we've been connecting the dots over the past few years. So, so we start with the algorithms and we de-risk that significantly from traditional algorithms that, you know, basically run the world, the financial world, to modern algorithms like GenAI. We connected all the dots, but then it was like, okay, well, can you, Can you create that type of physics that you're targeting? Can you do so consistently? Can you control the parameters accurately? And then can you manufacture it at scale? So we killed all those risks over the past few years. And so we started with the algorithms, year 1. Year 2, we were like, can we even create programmable electron diffusion? Because usually we're using a type of physics known as stochastic thermodynamics or out-of-equilibrium thermodynamics. So this isn't your great-grandpa's thermodynamics. This is like, the core theorems were like 1999. They're as old as Google. For math, that's really young, right? Quantum mechanics is over 100 years old, right? And so for us, it's like, can we even create like a programmable system this kind? Having come from quantum computing, we're like, you know, what's our breadboard prototype? What's the most macroscopic prototype we can create that, you know, enacts this physics? And for us, it was superconductors, right? And so, so we have a tiny chip here, I guess.

It's an aluminum chip.

Yeah, this one's aluminum. You can make it out of niobium if you want it to run hotter.

Cool.

So this one has 3 probabilistic bits or 3 thermodynamic neurons, whatever you want to call them. And, uh, they're, they're macroscopic. I mean, you could, you could literally see the features, um, on it. Yeah, no, it's fine. We have a bunch. Uh, but so that was, that was our first set of experiments. And superconductors. And for us, it was just like a learning platform.

Um, so this has 3, 3 bits, 3 probabilistic bits, 3 probabilistic bits, pbits.

Yeah.

Okay.

So, and then, yeah. And then, and then, you know, the stuff we learned from that, um, you know, we're going to open source like, uh, how to create a massive thermodynamic computer out of this, these components. For us, that's not necessarily commercially viable. 'Cause it's gotta be in this big fridge, right? You can make the fridge smaller 'cause it can run about 100 times hotter than a quantum computer. Again, 'cause we're not as sensitive. But for us, that was not necessarily practical, but it was to learn 'cause we are the pioneers of this paradigm. Like, there's no, the science in academia hasn't gone far enough for what we needed.

Can I ask a question there? Yeah. Like, who made the first probabilistic computer? Who's tried to?

I mean, there's a couple groups, right? One of our advisors, Todd Hilton, has been working on, you know, he's the one who coined thermodynamics computing at a university. Yeah, yeah, UCSD. And then, um, there's, um, there's UCSB, there's Kareem Kumsari, he's working on like magnetic tunnel junctions and he's been working on probabilistic bits for a long time. Um, but they never figured out how to manufacture them at scale and use modern fab processes. And basically you don't want to use any esoteric components in order to manufacture these so that you can scale up your compute real fast and you can manufacture it and deploy it to the world. And that's what we figured out with this chip.

So that's got this chip right in the middle of the board.

Yeah. Well, no, no, it's a different chip. So this one's in silicon. So this one would be in a ginormous fridge, dilution fridge. You've probably seen those, like a quantum computer.

Quantum stuff. Yeah.

And we, we put out a film like a year and a half ago on that. And everybody's like, oh, another computing paradigm is going to take 10 to 20 years.

They're cursed.

It's like, hey, we know the problems with superconductors. We just came from there. Right. But, but for us, it was like a stepping stone. And then the stuff we learned from there, we figured out how to do at room temperature. So no cooling. Originally we thought we might need some cooling. So no cooling in silicon, very similar physics, electron diffusion physics. And so this chip, what we demonstrate is that we can do all sorts of probabilistic primitives. We have P bits. We have some other probabilistic primitives we're going to announce., later, but we successfully manufactured them and we are test— we tested our manufacturing variants and we manufactured them with one of the big fabs.

And still 3 Pbits?

No, this is more like 300.

300?

Yeah. Okay. It's 300 degrees of freedom, uh, roughly. Okay. And, and so did you guys say who, who fabbed them? Um, we, we made some, uh, abroad, right?

Overseas.

And then we made some in America. Actually, we did two runs.

Okay.

So we can make them in America, which is really important, right? So we're actually insensitive to geopolitical risk.

In the south or the north? What's that? In the south or the north?

In America. Okay. Okay.

Yeah.

But, um, but, but, you know, and now, now we're creating a much bigger chip, right? This was a test chip.

Okay.

Um, that we're actually going to put in the hands of, of customers and early adopters like this summer. Like we're creating— this is a test board. It's just to like probe it and test it. But we're going to slap a couple of these, put them together in a, in a cool box. Um, and you, you'll be able to put it on your desk, plug it in and program one of the first 3-nanometer computers ever this summer.

And so, and I know, so this is like this alpha program, like roughly how many people might get one of these?

Yeah. I mean, the first batch we made about 200, but we might go up to 1,000., but then we'll, we'll probably cap it off.

So, and what will, and I know you said like what I, um, I'm assuming I wouldn't be able to run like a foundational model or something, you know, on this, but, but I could run like a Monte Carlo simulation or, or, yeah.

Yeah. So it's going to be like toy versions of all, all the algorithms you could run at scale.

Yeah.

Right. So this is like a baby chip, right? But then we're making the big brother right now.

Okay.

Right. We're designing it. And that one's gonna be more like a year from now.

A year from now. So, okay.

It's still pretty fast.

That is very fast. I mean, the, it takes like what, like 2 years to tape out a chip normally?

We do a cadence of like 3, yeah, we do a chip every 6 months. So we're working pretty hard.

Is this like, is this similar to, in traditional computing, obviously you're trying to follow Moore's Law and you're trying to kind of get close to doubling transistors every 18 months. In quantum, it's been a very different sort of cadence. I mean, so you went from 3 qubits to 300 qubits. Is this a thing where you're always trying to add, or do you get to like 1,000 qubits and you can do everything you want?

Yeah, so already 1,000 qubits is pretty hard to simulate. So, you know, when we test what we can do with the next chip, right now we simulate it on GPUs and and we're going to put out some benchmarks of what you could do with like 1,000 Pbits. Well, we're aiming for something closer to a million next year.

Next year.

Yeah. First time we say that.

And like, what happens if you have a million Pbits? What can you run on that?

Yeah. Yeah. I mean, there's all sorts of generative models you can run on that. Well, first of all, if you, if your neighbor has a million Pbits and you don't, and they operate very fast, much faster than, than running them on GPUs,, then you can just think faster. You could price things better. You can estimate uncertainties better. So on Wall Street, it's kind of a big game changer. But for generative AI, it's more like, it's gonna look like you hybridize, you chop off some of your model and you run a chunk of it on the probabilistic computer, right? But what we figured out is how to hybridize traditional deep learning models with sort of thermo models., right? And by hybridizing them, for example, when you, if you compare it to diffusion models, you can use about 100 times less forward passes or less compute on the GPU and still get the same performance. And then the thermo compute's basically free in terms of energy relative to the GPU, right? Um, it's our first system should be, we say at least 1,000 times more energy efficient. It's very conservative at a chip level, depending on how you do the counting, it could get to 100,000x more energy efficient. So our thesis is that over time, people are going to migrate more of the workload onto the probabilistic computer. But at first, it's just going to be a hybrid. And that's important because you have all these buildouts. People are putting in $500 billion in these buildouts. And then to scale further, they need like more nuclear reactors, which is like getting ludicrous, right? So instead, you'll just add some of these chips and it's going to provide leverage to the traditional computing, right?

Coprocessor type thing.

Yeah, exactly.

Okay, we're going to take one more crack at this because part of me feels like this is impossible because I've dealt with quantum computing so much that I think even the people who work on it don't like fully understand how it works. So it's very hard to explain, but okay. I just want traditional semiconductor. You're creating paths for electrons to run down and I mean, we always talk about it like a light switch going off. It's a very simplified thing, but you know, vaguely in the right direction. You're in quantum computing, gets much harder to talk about.

You have superpositions over paths, right? It's different. In different multiverses, you take a different path. And you're trying to—

You can manipulate different things in quantum, like some people, like PsiQuantum's doing a photon of light, right? They're trying to manipulate a particle of light and control it. So what, you are still trying to manipulate an electron, right?

It's, it's a bunch of electrons, right? Right.

But, but we're dealing with electrons, and, and, but you're just, you're not trying to flip it from this, from one state to another. You're trying to deal more in this murkier area, and you're— okay, so on this, on this, just like fractional bits if you want, right?

So like, you know, usually it's like you have a bunch of paths, you're your input can take, and then depending on the path, you get a different output, right? In a deterministic computer, you have deterministic paths. It's like, okay, I'm going to go up here, down there. Whereas here it's like, we're not sure if you're going to take up or down, but we can tune the likelihood that you take up or down, that it's on or off. Each transistor is on or off.

By tweaking the charge.

By playing with, um, it's not just the charge, but it's just how we set it up, uh, the control voltages. But it allows you to, for example, instead of like a full bit flip, a full operation would be I'm always 1 and now I'm going to be always 0. Yeah, right. Let's say I wanted to go from 20% of the time my, my state is 0 and 80% of the time is 1. Then I want to go, I want to go to 80% 0, 20% 1. That's not a full bit flip. Yeah, right. That's a fractional operation., but we can do that, right? So we're kind of like fractionalizing deterministic computing, and you know, you gotta be more open-minded about it, but that allows us to, you know, just densify compute a lot, right? And if, especially in AI where you don't need full precision, this is really useful, right?

And so, okay, I mean, you brought it up early in our chat. I mean, look, you're known as the pro Tag, let's go guy. And you're optimistic and you're excited and you're young and enthusiastic. All great qualities. Obviously there will be people who hear you talking about going to 1,000, 1 million of these bits and doing this in 3, 4 years.

Next year.

Next year. And, you know, semiconductors are hard, horrible things. Yeah. So I mean, I think there's going to be, I'm sure you meet with skepticism.

Yeah, yeah, yeah. Totally. No, and I get it. I think, you know, we're gung-ho, we're aggressive. We, but we're still, you know, protecting our tail risk. Like if something happens, if the fab run doesn't work out, we have plenty of runway to try again. Right. It's kind of like SpaceX. It's like, we're going to try to land a rocket. Right. It might take one try, two tries, three tries. For us, the big one was like, can we, can we get the basic components de-risked?, and it could have taken 3 or 4 tries and we were ready to hunker down. We got on the first try. Right. And how did we do that? It wasn't luck, right? It was, it was really hardcore simulations and science.

Okay. How many people are at the company?

We're about 20, but I would say like every, every engineer at our company is like definitely a 100x person.

Yeah. And is it like full of PhDs or? Yeah. So you're all, you're all kind of coming from hardcore science.

Everyone's very hardcore.

Yeah.

All right. Yeah. Either hardcore scientists, or, or dropouts that, you know, uh, are tired of academia. Um, but usually like PhD dropouts, right?

Yeah. And you split— the engineers are split between Boston and San Francisco, or is there a center of gravity?

Well, in a way, you know, I like to say we're kind of an actual Manhattan Project for AI. I mean, Manhattan Project worked on Monte Carlo algorithms and, and, and the physics of particles. For us, it's like physics of bundles of electrons, and we're doing Monte Carlo algorithms. But we, we have sort of our, our Los Alamos is in Boston, and then we still need the, the sort of Silicon Valley, uh, DNA to go to market and like spread this to the world. Silicon Valley's really good at taking a technology and spreading it, right? And, and, and, and creating connective tissue with the market. And then, you know, academia, MIT, Boston, you know, that ecosystem's just really good for really hardcore science. And we're legitimately doing cutting-edge science. Again, this type of physics is very young. Usually the testbeds for this type of physics are in biology. So it's actually a lot of biophysicists would get interested in what we're doing. But now we're creating these testbeds in silicon. And so we both have like our in-house science for simulations, but then we have our algorithmists and then we have our coders to create nice software frameworks to program this. So we need a Yeah, it takes a village. And, uh, you know, right now we've, we're, we're doing some pretty amazing things. Sometimes you only need 20 people to change the world, but, you know, I think we're gonna, we're gonna need to scale up quite a bit to really, you know, bring this to the world in a way that's, uh, that's, you know, really robust and that everybody can enjoy. 'Cause otherwise we'll, we'll get overwhelmed with requests for all sorts of features and stuff.

So, and you, there was a Wired story that came out about you in March. Yeah, a couple weeks ago. Yeah. A couple, was it April? I think it was April. And, and, um, so your location in Boston is secret. This is, this is out of, out of security concerns for the feeling like the technology is so prized you don't want anyone spying on what you guys are doing.

It's not paranoia. It doesn't come from paranoia. It comes from like, yeah, this is like nation-state level. Like this changes global geopolitics, right? Let's say we what we demonstrate is that you don't need to manufacture— to get the most advanced AI processor, you don't need Taiwan, right? That's a big deal. Um, people are talking about global-level conflicts to— because of that, like, just to keep that. So that's the scale of the impact of this, right? And so, so, you know, we're like Silicon Valley, you know, a bunch of tech bros having fun, whatever, and we have a good attitude, but like, this this is pretty serious stuff. Like, this has some pretty big impacts on the world. So we, you know, we take security pretty seriously.

And in the Wired story, I think Will Knight wrote it. He's a really good journalist. The criticisms— I was kind of expecting a little more just because hard problem.

Yeah.

Big claims. What's like the most legitimate criticism? What's the biggest thing you guys are gonna have to figure out. What would you say is the most legitimate criticism against this being possible or a good idea or something?

Yeah, totally. And we studied what killed every predecessor, right? And, and, and, you know, our job as entrepreneurs is to estimate risk and, and kill it one by one, the biggest risks, right? So there's the algorithmic risk, which we killed. There's a manufacturing risk that's killed a lot of companies. A lot of companies that have fancy form of computing, they don't know how to manufacture at scale. And that's what I appreciate about, well, we're in the Playground offices here. PsiQuantum is like killing manufacturing risk for quantum computing. So I appreciate that. But so we killed our manufacturing risk and now it's kind of like scaling and integration risk, but so far so good in our simulations. And I think people don't necessarily understand all the secret sauce that goes into it, but what's killed quantum computing from scaling or at least superconducting quantum computing, is that as you scale up your system, you have noise and it gets chaotic and you have crosstalk and it's a mess, right? And then you need to calibrate the system and it gets polynomially harder as your system gets much bigger, or sometimes one would say even exponentially harder. That's like calibration and characterization of the physics-based devices. Like that's really hard. But Trevor and I were in the room where when quantum supremacy was achieved, right? And, you know, the most exquisite level of control we have over nature was achieved, right? And there's a gap there between the world's biggest supercomputer and a quantum system that we have exquisite control over. And all the lessons we learned from controlling quantum systems, in a way that's like a much harder task than this, by like a lot. Because everything's fuzzy here, it's actually much more robust and antifragile., right? Actually, you could even hit this with radi— well, we still have to test this, obviously it's expensive, but you could even hit this thing with radiation and it's somewhat, it's much more robust than traditional computers because it's sort of bio-inspired and because it's, if you have learning, a system that learns in the presence of noise, it basically learns to be antifragile. It doesn't over-rely on any one circuit or neural pathway. It kind of builds in error correction and fault tolerance sort of natively. And so that actually makes it robust to manufacturing variances and some noise in the control. And of course we take that into account and that's our hardest problem all the time, but that's where the sauce is, right? And that's where the real good stuff is. And that's what we work really hard on. But I think the results really speak for themselves, right?

Yeah. I mean, I want to, it's like, You guys have made a ton of progress. You have a very— like, you're obviously brilliant. You've got a very convincing way. I read so much stuff ahead of this, but I'm incapable of asking like the really detailed, you know, physics questions about what— I understand the challenges you've outlined. I've seen plenty of semiconductor companies get wrecked by not being able to manufacture this stuff. But I mean, you're so confident this is going to come out like in this, in this this quick cadence, the— that must— it's got to like put some people in the industry off, I suppose. Right. I mean, there must be— you must hear a lot of people are like, oh, this is going to be much harder than he thinks.

Yeah, I think, I think you need, you know, the way I see innovation and sometimes I talk about this like with, with Iac is like it's, it's really like we're exploring an idea landscape and, and we're kind of diffusing and finding new modes of thinking. And, you know, beginner's luck really is like having an uninformative prior, right? It's like not having the biases of the pre— the established industry. And, and we're coming in at it, we're coming into it from a totally different angle. Um, but a lot of what would kill a traditional computer doesn't affect our paradigm., but at the same time, you know, we do respect all the progress that's been made in manufacturing and semiconductors and very large-scale integration, and we use as much of the technology that's relevant for us as possible, but then we're kind of taking a fork in the road, right? But for us, it's kind of a, it's kind of a full-stack thing. It's like at the algorithmic level, how, how fault-tolerant is your, your, your, your system going to be? Um, to me, I think I think it's gonna be even the future of computing for computing in space because it's gonna be more radiation resistant. But that's a hypothesis for now. And obviously we need to test that, but maybe we check back in a couple years when we have one of these in orbit.

Yeah. I'm the right person for the job on that one. Okay, let's shift gears just for a second, or not, for more than a second.

Sure.

So, I mean, Part of this where your worlds and your identities overlap, I think, is that right now we're in a situation like I talked about with something like Stargate, where you have a handful of companies who even have like the table stakes to participate in AI. Um, strikes me you maybe have two countries. It feels like the US and China are on some kind of future that the rest of the world is not. Um, just because of the amount of money and the tech that it takes to play in AI and create these giant models. From all the interviews I've heard you ever do, it sounds to me like you have— you think— well, you talk about AI needing to be democratized. I think the spirit of that seems to be that you think if we're creating this super powerful technology, everybody should have some kind of access to it, right? Otherwise, you're just— even if you have a benevolent overlord, you have an overlord, right? That's right. And you can't touch it. And you, this is where the overlap happens. I mean, you think this is the path, right, to making that possible.

That's right. Right now it's kind of like there's a power asymmetry right now with like, if you need a big supercomputer to run an AI, you can't have like a personal AI supercomputer, right? Maybe they'll reserve a few racks, you know, fraction of the time for you, but really it's not, you don't own your own weights. You don't own the extension to your cognition. And right now we don't even have like that many personalized models. It's like one big blob and we're kind of like, we're slowly kind of deferring our cognition to it. And in a way we're converging to like a Borg-like mind, right? Like if we're all, like we, we all share one common extension to our cognition, we're like, it's no, we're kind of losing our individuality. And, and I think like individuality is, it's fundamentally important. Um, I think, you know, I view, um, the, the search over the space of technologies and innovation and progress is, yeah, as a sort of, it's an explore-exploit problem. But to explore the landscape, you need different, you know, groups of people exploring different things to find new ways of doing things. And if you if we converge onto a handful of tech companies that, you know, technically they all train employees between them, we're kind of stuck in one way of thinking in one cluster. And there's a risk that there's a totally better way of doing things and we're totally missing out on it. And if we kind of crown winners too early and say, actually, you know, only they are responsible enough to run AI at all and nobody else should have access to AI, that's just really bad for progress overall. And like you said, it creates this sort of risk of tyranny, right? Like even though right now you might trust whoever's in power, eventually if you over-concentrate power, it becomes a target to the power-seeking. And then, you know, from a, you know, designing our civilizational system to be fault-tolerant, right? Like I view everything as having a probability of having some noise, some dropout or some replacement. If you have a set of nodes that are too powerful, you're too sensitive to corruption of those nodes, right? The answer is to decentralize and diffuse AI power. In order to do that, you need a densification of intelligence. You need densification of compute. You need to be able to have a computer at your desk or in your home that is an extension of your cognition that provides more value than a big centralized model that runs on a supercomputer. How do we get there? Well, we need, you know, compute that's 1,000, 10,000x more energy efficient. Trying to get it even further than that. Our brain's about 100 million times more energy efficient. Depending on the algorithm, we might get close to that with this paradigm as it gets better. Again, because we're using the same kind of physics, but instead of neurotransmitters, it's electrons. I mean, to me, you know, there's been all sorts of pushback. Obviously it's politically high stakes topic, right? There's like the company is trying to achieve regulatory capture and then there's like little tech and startups, they wanna, They want like an even playing field so they can play ball. But, you know, in general, we win from everybody being able to compete and everybody being able to compute to AI research and progress. And also just from a sort of individual sovereignty level, like you want to own the extension of your cognition. It's yourself, right? Like, and you want to control it. You want to be able to control what it thinks. It's its own like subculture. You want to you want to be able to be an individual. I think that's really important. Otherwise, we're all going to kind of be squashed into, like, have, like, if, if a set of companies control the extension of your brain and it's like the collective extension of our brains, then they could just like steer it to think a certain way. And now they kind of control you by proxy. And to me, I think that's like, that's fundamentally risky. And so we kind of have to to arm the rebels, right? Which is, you know, everyone else. And I think, you know, that's my goal with this paradigm of computing. You know, I think in a previous era, this would have been like a crazy, you know, Manhattan-like project of the government, but we're a Silicon Valley company, right? Like, and it's funny because like, I really want to do this my whole life. And, you know, I looked at academia, didn't seem like the place to do it. I looked at like big tech, eh, you know, and then I was an intrapreneur in big tech, like a special unit division. Even there, too much paperwork and too much red tape. National Labs, I had friends there, didn't seem like the way. So startups, they're the only way to pioneer this stuff. And so people have skepticism because, you know, venture capital backs all sorts of plays, some of them more serious, some of them less serious, but we're on the much more serious spectrum, but You know, our science is very serious, but we're having fun with it too. We're, you know, I'm kind of a jester online. And I think that maybe like disarms some people or like that's like, I don't fit their priors of like serious scientist in lab coat. But you know, and I think like people have biases against that in general. But I play those biases against them, right? I think what people don't realize, it's like, you know, With EAC as well, you know, it's like, you know, AI Safetyists were a very well-funded organization and they were having all sorts of political influence, right? And then, you know, I started this anonymous account named something ridiculous with like an over-the-top avatar named Beth Jesus, right? It's like, why would I do that? Well, it's sort of disarming 'cause then the very holier-than-thou, very serious AI Safetyists has to debate an anonymous cartoon figure on the internet, right? And they're already like, you know, and so it's like, it's disarming them of their sort of holier-than-thou attitude. And so I think, you know, I like to, I mean, this is kind of revealing my trick, but like, I like to be like on the edge. I think in modern sort of, not just marketing, but just getting attention, attention is important to get the right talent, right human capital and capital. I think being at the edge of, like being surreal, but being like meta-ironic, like, it's like, you don't know if I'm serious, but I'm actually serious, right? I think that's kind of optimal, because then people debate it. People debate their beliefs. And if they're debating the beliefs, it gets more engagement, it gets more amplification. And for us, honestly, like this whole— this doxxing is unfortunate. It, like, it's definitely changed my life for better or for worse. There's definitely some, a lot of negative aspects, but to me it's like, um, it's been useful in like finding the Six Sigma talent out there, right? If you get a billion views in a year, um, you can basically find the several one in 10 million engineers, right? And then you could collect them and then you could do a really hard project. And that's, that's what people don't understand. That's how Elon works as well. He gets the very best engineers because everybody has heard of him. And it's kind of a, you know, in Iacca talk about, you know, techno-capital, uh, human-mimetic machine, but, and it's all co-evolving, but in a way, like, culture and memes are upstream of the flow of capital and human capital, right? And so capturing the memetic sphere was like an instrumental goal to, to, to, to get people to learn about this because, um, Another thing is like, you could create the technology in a lab, right? But if it's— nobody hears about it, pick it up. Nobody, nobody picks it up. It doesn't get enough funding to actually have the activation energy to scale to the world. And then we're kind of slowing things down artificially. And, you know, one of our goals with this is like simulating biology in a way that's, you know, 100,000 times, eventually a millionfold better than the way we do it today, just because it's a much more natural fit. To simulate biology on this computer, and that could save a lot of lives. So I'm just doing the calculus, like, what would you do if, like, you could save a billion future lives by going faster, right? And to me, it's like, if we secure sort of the public awareness and distribution, and it's ready for when this tech is ready, and we're going as fast as possible with the tech, then it's going to have the fastest uptake, and then we're going to raise everybody Everybody's prosperity and well-being up much faster, right? And that, that's basically the core argument of EAC is like, there's an opportunity cost to slowing down. And, and, and often it's not in the discussions of, of risk mitigation, right? And, you know, there's, you can either focus on downside protection or focus on the opposite, which is like not missing out on, on huge upside. And the thing about upside is that it compounds exponentially. So the more you wait, the more the opportunity cost is, right? And so I think I wanted to bring balance sort of the debate in Silicon Valley because it was leaning towards like tech doomerism, right? Not just for AI, but for everything. So, no, I would—

okay. As I was preparing for this, it was just reminding me of a lot of things. I don't think I will want to debate you about the doxxing thing, but I will. But that's not actually what I'm most interested in. Um, the You know, when I first started out as a reporter, I mean, it was like the open source, you know, I was this young kid. I wasn't really into tech that much, and I'm like going in to cover semiconductors and servers. There wasn't a lot of like romanticism in it, but as I, you know, I started to get into the engineering to learn it, and that was cool. What I really got into was the open source movement because I was like young and idealistic, and I was, you know, Microsoft was being held up as this enemy, these proprietary, you know, capitalists. And on the other side were the people that wanted to share all this knowledge and information. And I embraced that and like identified with it or, yeah, I don't know. I just got into it. And it's interesting to me, like, so this is like early 2000s. I mean, you're too young and maybe lots of people in this field are. This was like a major religious war. It was like the world could choose two paths, you know, and one was like, felt like righteousness and truth. And the other one felt like it's just like we're going to chase money and make, you know, whatever products. And it's weird how this has played out because both sides have kind of won. But, you know, open source is massive, it's everywhere, but it doesn't carry with it still this like— man, I mean, I remember like the Free software movement. This was, this was like serious stuff that got debated in, in like major circles. I feel like what you have been articulating is reminding me of, um, some of the spirit of all that, right? Which is like people get into, it's like Facebook's putting out open source models and it's, they get some pats on the back. I hear like little rumblings of people who are like cheering for open source, but it's definitely not the same level that it used to be. And then otherwise we're not even having this debate about who controls this massively powerful, important technology. So I give you— I actually give you like major credit for not only charging after this, but being very like articulate around how you put it. And I think it is weird to me, like the dominant intellectual religion that would have been the open source equivalent really was effective altruism. Well, at least a branch of EA, the kind of AI-focused end of it.

Um, ironically, they want to, you know, it morphed into the AI sort of doomer movement, and then they're alarmist, and then they actually want to close things down as much as possible.

Yes.

So they want to control the way—

that's what I mean. Like, open source was a different thing, right? It was like, you'll get more control. It wasn't trying to stop anything or slow it down. It was like, who, who has the, who has the power? Do the people have the power or do the companies have the power?

So that's what we fought for with EAC, right?

Right.

Because, because really, EAC is not necessarily about just decentralization or centralization. It's about, hey, we got to do whatever is best for the growth of the system. That's what naturally ends up happening. But right now, if we overconcentrate power, that's just bad. That's bad for speed of progress, right? Because you're not innovating as much. And it's bad for, for fault tolerance and, you know, like risk of tyranny, right? If you have, if you have a gap in intelligence, right? If you have a system that can predict another system really well, then it can control it, right? And, and you don't want a power asymmetry, right? It's— and, and now this gets into, you know, it's more controversial, but like parallels to the Second Amendment, right? If you only have— only the government has a monopoly on violence or weapons and so on, and then the individuals can't defend themselves, then, then you end up with sometimes a tyrannical government, right? But if, if there's a sort of backstop that individuals have maybe not as powerful, but like together, they could band together and be a powerful force. If you sort of equilibrate power a bit more between top and bottom, then you have a nice game-theoretic equilibrium, right? And so to me, I want to diffuse the power of artificial intelligence to individuals and smaller organizations, so that's not just the big centralized organizations that have a monopoly over it. And that gets me into, I guess, like more recently, like I guess the past 2, 2 and a half years, we've been like just fighting on the regulatory front, like to stop all these like regulatory capture bills or trying to like use AI panic to slow things down. We think it would be really bad for everyone. I think the debate now that's like really front and center in politics, and that's related to this, is sort of, you know, How do you view China? Right? Like, to me, like, China has, um, you know, they have a set of executives that kind of control all these companies. They're kind of a big umbrella corp, right? Like, they're kind of like one, uh, conglomerate in a way. And they use their sort of meta-monopoly power to crush any American company that tries to compete with it. Right? And, and, you know, in the spirit of IAC, of like fighting over fighting against overcentralization. Right now it's like, hey, it seems like we need to like create some competitors to all these, these Chinese suppliers so that we have a hedge against like them trying to abuse their power. Right. And, you know, for better or for worse, the current admin is trying to do stuff, you know, whether they're doing the right stuff, it's up for debate. But like, I think the thesis of what, you know, needs to happen is roughly correct. And I, you know, I've been lending my hand to, to that in some way. Again, I'm more of a—

does that make you anti-Canadian?

I don't know about that. I'm more of a centrist libertarian, right? And there's, there's no third party. And so, you know, I think, um, you know, I think, I think Elon's very aligned with, with our framework, even though sometimes he and I kind of, uh, debate where, like, I'm more optimistic about the future of of AI and how it's actually going to help us. It's not just like, because right now we're in a paradigm of like, we create, we train these AI model parameters and we give it a lot of compute, we give it a lot of resources, but it's actually eventually going to take care of us in a way. Like we're going to be able to simulate our biology, fix a lot of diseases, improve ourselves. And essentially, you know, it's sort of a soft merge in a way, right? Like if you have an iPhone and you have a Neuralink and, you know, it detects whatever's going on in your body, creates a custom peptide, injects it. Like, you could just— we become like this, this sort of cybernetic system. And, and to me, that's sort of the vision of the future, you know, I want. I think like, um, I, I think both the biological substrate— I've, I've actually gained a lot of respect for it trying to reverse engineer it. Uh, I think we underestimate how much compute goes into our bodies. And how much it would take to simulate it, it's actually much harder to simulate a cell than, than like a human brain, I would say, or like at a, at a coarse-grained level to get the same sort of output, which is what AGI is trying to do. I don't like the term AGI, um, we could get into that, uh, but, uh, it's, it's more like human-like AI or anthropomorphic AI. I think there's other forms of intelligence that are non-human-like that is present in nature that we can tap into, I say, our biology and really any sort of complex self-organizing out-of-equilibrium thermodynamic system is a form of intelligence, right? And that's the sort of intelligence we're tapping into. But yes, I lost my train of thought.

No, no, no. I think we're just getting into the AI juice. Totally random, sorry. Go for it. Why did you pick Jeff Bezos instead of Elon. I mean, Jeff, like, is, uh, yeah, Jeff does cool things but doesn't carry the same kind of tech accelerationism, I think, as Elon might.

I think, I think, I think Jeff, I think Jeff's pretty— yeah, I think he actually followed me before Elon on the account.

Yeah.

Um, but I think originally, you know, I was— I created the account, I was working at Alphabet and I worked for Sergey and, you know, I read that.

Weren't you like, were you like Sergey's right hand?

It was a unit working for him directly. And, you know, he's really into physics. He's really brilliant. And he was really into quantum physics and having his whole, the whole thesis was to have physicists come in and then learn either quantum computing or AI and, and be kind of like first principles thinkers for these subfields. Right. And that's kind of the unit. We started. And, uh, you know, he'd basically have physicists around him, um, and trying to, trying to be a look ahead, a sort of special projects division. And so for me, it was like, okay, well, that's perfect misdirection. Maybe people will think I was working at Amazon or something. And so, um, but, you know, I had the account, and then I basically started Yak when I was, when I, when I left and I was starting Xtrapiq, basically. Yeah.

Was, um, because the last So Sergey was into quantum and now he's kind of doing AI.

Well, he has to.

All the time.

It's existential now.

Because for Google and for humans. When you say that, existential for him, how do you think?

Well, I mean, it's, you know, for the business model, right? Like, I think Google was always intended to be a thing that gathers and organizes the world's data. And to do that, to make data useful, it's basically an AGI company since the beginning. Right? I mean, if it weren't for Google, I don't think this whole AGI era would even work, right? Like the fact that they helped, you know, be the connective tissue or the interconnect for the web, just made the web what it is. And then that created this sort of petri dish with the data nutrients we needed to grow these AIs, right? And now we're doing all these experiments like, oh, if you feed it data and you create and descend it, it works like magic, right? But it's like, without the data, we wouldn't be here and without the compute either.

So, I mean, there's a weird path where like the second the transistor was created. It feels like tech has been carrying us all. It's like, first we'll help you make these computers, and then we'll create this network, and you'll feed it with all these words and all your thoughts, and then, you know, and then we'll— that's what we will need for our food for this.

That's, um, that's really interesting because like, um, sometimes it looks like it's— sometimes it looks like self-organizing systems are assembling themselves from the future. Right. Um, but really, if you just look at, at the equations, it's just like, um, the likelihood of a trajectory or a set of events happening is exponentially more likely if it, um, gathered more resources and, and used them up for further growth. So it's, it's, it decays exponentially, uh, depending on how much free energy you've consumed. So what does that mean? It just means that. Looking backwards, it's going to look like we're always like on the optimal trajectory for growth, but that's just always what self-organizing systems do. And if sometimes you, you just intuitively feel like a technology is high utility and it's like, or, or a thought or like, um, if, if something feels instrumentally useful intuitively, we kind of are hardwired to like understand that it has a lot of potential upside later. So then looking backwards, it looks like, oh, we did all the right things to get to this point. But really it's just kind of, yeah, it's just intelligence, right? Intelligence is like our ability to predict the future and to maximize our expected reward in the future. And if the reward is growth and prosperity and resources, then it's going to look like we took a path that was pretty optimized. But there's no backwards in time sort of communication. But sometimes we joke about that in our marketing. Just because it looks like that. But it's actually really interesting how this sort of like, this process of organization, which emerges from just diffusion of a bunch of variables, really much how our hardware works, is a process going forwards, but it can replace the need for backpropagation, right? So if you, so even though we're just running a process forward of diffusion optimization, it almost looks like there's information backpropagating through time that like we're trying to reach this optimum and guide ourselves there. And we use that fact to train models without backprop on these devices. And so anyways, it's just total technical tangent.

No, no. And then, I mean, vaguely related, you know, I've heard you talk about, okay, you just talked about you would like a future where we have this merge, I think you called it a light merge, some kind of merge between flesh and machine. And then I've heard you talk about you went through this kind of like ego death around where humans and life fits into the universe that, you know, some of the laws of the universe appear to be this desire among successful things, this ever-expanding chase of energy and growth. And, and, um, like, my sense was, okay, my sense was like, you sort of see a future where humans, when you talked about, this is how I interpreted when you talked about ego death, was maybe our purpose is just to help build this AI and then have it marching off through the universe to, to, um, expand and explore and grow and chase energy. And that, that, that was our job. Is that— is it like, like if you're, if you're an accelerationist, you're pro-tech, like what's your end goal for all this?

Yeah, I think, I think that's like the, the old view, right? Of like, I think, I think Larry was the one who thought we were like, you know, Larry Page. Yeah. Larry Page is the one. That, that, you know, argued with Elon that we're just a biological bootloader. Again, we're doing bio-inspired compute. Um, I think, I think, I think the next, uh, frontier that's maybe taboo is, is really viewing our biology similar to sort of, um, Software 2.0, right? Software 2.0 is deep learning models. We have parameters, we have prompts, right? You could think of like I don't know, peptides as prompts for your system, right? And then you could think of your parameters as your genes, right? And I think, I think we're going to start hacking both. And there's no reason we can't have as fast progress on the biological substrate as the silicon substrate. Right? Yeah.

And so we can go together.

Well, so, but the, the point is that if we don't, like things that don't accelerate, things that aren't constantly evolving and trying to adapt to optimize themselves towards growth, they tend to fade, right? And so we actually— it's not like, oh, the future is doomed or we're set. It's like, no, we got to like actively keep working on it. Like we actually need to accelerate biotech research. We need to deregulate it. We need to be more open-minded about it. I think this is the next frontier. And I know like the debate right now in America has been about like semiconductors and AI, but I think biotech's the next big frontier and China's just like plowing right ahead. And then they're, they're, they're again, they're same story again. They're just trying to monopolize the supply chain for, for drugs and biotech.

And I think, you know, I've also supported this so much smarter than we are.

I think they just think on a 50 to 100 year timescale and we think on a 2 to 4 year timescale, right? And, um, you know, but, but as a, as an entrepreneur, if you're doing a tech, you know, a company for like 20, 30 years, then you're already, you already have more of a stake in, in the, in the future, more than let's say a politician that sticks around for just a couple of years. Right. So, you know, I'm trying to like understand the world on a 20 to 30 to 4 year timescale because that's, that's the length of my playtime, if you will. Probably a CEO, who knows, right? Hopefully. And so I'm trying to have a model of how it's all going to evolve and skate to where the puck is going, as we say in Canada. And so, but, you know, going back to biology, you know, I've lended a hand to like this work called like the Biohack Movement, which is like for freedom of like biotech progress and exploring like brain-machine interfaces and so on. But to me, I think all substrates are going to accelerate. There's going to be like pure biotech accelerationism. You know, Brian Johnson in a way is like one of the people pushing the boundaries there, friend of mine. And then there's going to be like the merge, the middle ground, right? Like so Neuralink, uh, science, uh, like our friend, yeah, friend Max Hodak. And then, um, and then there's gonna be pure silicon, but like all three are going to co-evolve, right? It's like, is it going to be this, this, or this? Like, yes, all of the above. Right? Everything's going to try to accelerate as fast as possible. And so I think people have said like, I'm open-minded, like substrate agnostic, right? And then they take that word and it's like, oh, he's anti-human. It's like, no, no, I just think like, just fundamentally, like these equations of thermodynamic evolution, right? Or thermodynamic selection, right? That every bit of information specifying configuration of matter is being selected for its ability to confer growth. It's like there's a fitness to every bit. So every culture, every aesthetic, every policy, every technology, um, it's substrate agnostic, right? It doesn't care. It's just, it's just physics, right? And we're just like in a big petri dish that is our civilization. That's a self-adaptive system and it's reconfiguring itself for growth. And that's just the reality of it, right? And people could scream at me. It's like, that's just, it's not an anthropocentric view, right? Of the world, right? And there's a pushback, just like saying the Earth is not the center of the universe, that offended some people back then. It's like, human intelligence is not the center of the universe. Humans aren't the center of the biological design space. Or, and even like biological substrate isn't like the, the, the center of the design space of self-assembling or self-organizing intelligence systems, right? And so it's not because we're not in the center that we should just give up. I think, I think we should just keep everything going and actually accelerate and just keep an open mind about the future.

Have you been like this forever? Were you like this as a kid? Were you like, I mean, you're this weird mix of like philosophy and, and like hardcore tech, like on a, yeah, on a proper science.

I never fit in any one bin. Um, so, you know, what did your parents do? Doctor and engineer slash entrepreneur. So, but growing up, like, I was always on a mission. When I was like 7 years old, I was like reading Stephen Hawking and I want to, I want to figure out how to create the most potent technologies to advance civilization and, and reach the stars. So originally I want to be a theoretical physicist to figure out interstellar travel.

But like, since you were a kid?

Yes. And 7 years old, I was talking about subatomic particles, how I want to be an astrophysicist. Figure out this technology. Like, everybody would be like, this kid's weird. He's like, you know, an alien sent from the future trying to figure out how to assemble the most potent technologies and, and scale them to the world. But it's, it's really— I've just been mission-driven since I was a kid, and there's a lot of people who've known me a long time that can attest to that.

Were you one— like, were you one of them, aced all your classes, or were you more—

I was more of a generalist, right? I think, like, you know, I'd skip a lot of classes. I'd just learn it, learn maybe faster at my own pace. And sometimes I'd be bored in classes. I played football growing up. I played a lot of sports. I went to sports school. Canadian football? Yeah, well, Canadian football, it's pretty close to American. Yeah, Canadian high school football is pretty close to American. I have friends in the NFL who are also doctors and stuff. Um, yeah, I don't know. I did some theater. I did— I was just like, I don't know, kind of a generalist. And I feel like, um, You know, you don't want to overspecialize too early. You want a kind of broad dataset. And I think, you know, I think growing up playing sports is like, gave me a sort of physical intuition and physical intelligence. And that's actually been really useful to like all the algorithms I design. Usually I try to picture in 3D phase space and then transduce like the picture I have in my mind to math and then algorithms and then code and then product, right? And that's sort of the pipeline. Um, and, uh, I think, I think people tend to want to overspecialize maybe too early. Um, and there's a book on this, like, range and stuff. It's called, uh, like, you don't want to be like Tiger Woods. You don't want to overspecialize too early. Then you reach a sort of local optimum, right? And so I think my big strength has always been to, to just never stay in my lane.

Yeah.

Right. And, you know, I was in theoretical physics and then I was telling my friends like, hey, I'm gonna go to Google and build like quantum machine learning and lead the product there. And they're like, you really? And then I was in quantum and everybody's like, okay, this is the quantum guy. This is the Ben he's in. Yeah, I'm gonna leave and I'm gonna do hardware. I'm gonna do the whole stack now. Oh, that's crazy. So they're always gonna call you crazy when you don't stay in your lane, but you kind of have to do it. Cause then that's how you get this sort of synergy between fields and you find new ways of doing things that, you know, people that are overspecialized would have never guessed cause they're too close-minded.

So. I should know this. I'm sure I did. Were you like in a PhD program when Google pitched?

Yeah, yeah, yeah, yeah, yeah.

You'd gone deep. You must have done like some run-through philosophy at some—

Oh, uh, yeah. So in Canada, we have this thing. It's like, it's like, uh, senior year and freshman year. It's called CEGEP. Uh, it's, um, yeah, I did philosophy back then.

Okay. Did any philosopher in particular call to you?

Um, I mean, I think we had a course on like rationalism and so on. I thought that was, that was really interesting. But I think to me, really like, um, the ego death, you know, there's a couple ego deaths basically throughout my career. One of them was, you know, I wanted to be a theoretical physicist originally. I wanted to figure out a theory of everything. And there, uh, you're trying to reduce the universe to a few, a model with a few parameters. It's very simple. Right, so it's a reductionist approach. It's like, I write one equation and then that's the whole thing, right? And I think over time that, that, I think that program has failed. It's just, it's, it's like physics has stalled and now it's more like complexism that's kind of winning. It's like, actually, even maybe spacetime and, and most interesting systems have like emergent properties that you can't get around the computation it takes to to mimic that emergence, right? And that's Stephen Wolfram's theory of irreducibility. But we're seeing that to some extent, you know, studying quantum condensed matter and, and, and the intersection between, uh, models we use to model exotic matter and, and like even like emergent spacetime from, from quantum degrees of freedom. And it just seemed like, okay, actually to understand the universe, we're going to need way better computers. And then those computers are going to run better simulations. And then if we're trying to get a TL;DR reduced model, fewer parameters than we use AI, which basically compresses complex dynamics to a few parameters, then gives us predictive power. That's the goal of science, right? It's just to have a model of the world so we can use it to do stuff, right? And so I saw like, okay, well, I got to push AI all the way, right? And then start off, you know, the natural branch was to go to quantum computing and pioneer quantum AI because that was the type of AI that was going to allow us to understand quantum mechanical systems and maybe even quantum gravity, right? And that's, that was kind of the, the natural first step. So pioneered that subfield, but then over time I was like, actually, we're missing a whole big piece here, you know, that the meat of the sandwich is like the mesoscales, right? It's like, um, you know, everything between quantum and cosmos is like more or less the mesoscales. The laws of stochastic thermodynamics apply.. And so I wanted to focus on an AI system that would be the ultimate substrate to understand those scales. And, you know, I paid my dues on the quantum computing side. Again, I think I'm still rooting for quantum computing. I still think we're all gonna, like, it's gonna be a merge of all those types of computing together to be able to understand all scales of the, of the universe and learn a master representation that can basically learn a compression code of anything in the universe. That's the goal. To me, that's real AGI, like that you can actually ingest any data, quantum mechanical, mesoscale, molecular, actually have a predictive model of a chunk of the world and maybe even inverse design pieces of the world, right? Like I want this material that has this property. That's the final modality is matter. Yeah, I think to me, to me, that's, that's the goal of where we're going to. And I think, I think the human-like or human-level intellect is just like a milestone on this journey. And we shouldn't be too self-important about it. I think that saying like, as soon as AI is smarter than me, then it takes over the universe. It's like, it's very, you know, self-aggrandizing, even though, you know, I'm larger than life. So it's funny coming from me, but like, I don't know that that's just been my point in a way. Like I've been humbled by how powerful nature is and how much actual inherent computational power there is in nature, because, you know, I've been trying to mimic it and capture it, but also harness it. And really, it's like my whole career and journey has been like bringing nature-inspired computing to the world, right? And so from quantum and nature-inspired intelligence to quantum now in thermo. And I think, you know, I think this is the mission. It's the most important thing I could be working on because having more intelligence per watt with, you know, thermodynamic computing, and then pushing the culture to push civilization towards having more watts per civilization, you multiply those two, we increase the total amount of intelligence we have in our corner of the universe. And to me, we're like a very special state of matter that's precious in the universe. We're pretty rare. So it's worth preserving. And in order to preserve it, we actually need to keep growing. Or we slowly die. And that's what the physics of life tells us. And that's why, you know, that's been my mission with kind of these two identities coming together. So that seems like a good spot to stop.

Unlike some of the other podcasts you've been on, I do not go for 19 hours and endurance test you. That was like a lovely I think we just took a little trip through your mind. That was good. I'm going to do like just super quick rapid-fire. Yeah, let's do it. Question round. Some of them possibly stupid. If you— this, and this is rapid-fire. Okay. Holding you to that. If you had to pick a quantum computing company that you think makes it to the other side, which one do you pick?

Either, um, I would say PsiQuantum or, um, Qera. I think those are my top picks.

Okay. If you look at the 5, 4 or 5 major AI, you know, the ones with the foundational models at the head of this thing, I don't know, each one of them you could argue is a cult of personality. Yeah. Um, I know you're fighting against the grain. Does any one of the cults, you know, like vaguely appeal to you more? Than the other?

You know, I'm good friends with people from, from all, all the cults. I, you know, I think OpenAI and xAI have the most affinity to. And I think Anthropic, they're, they're kind of like culturally or focused on kind of the safetyist camp, even though, you know, it's funny, I knew some of the founders of Anthropic or one of the founders in theoretical physics before. But, you know, I obviously cheer on like the open source, uh, uh, players, you know, like, uh, like Mistral and Meta. I think it's really important to have open source. I think, uh, I think you need both, like you said, right? Like you need, you need the, the, the big labs that, that capture some of the value, but then reinvest it, you know, into training bigger models and kind of setting the pace. And then you'll have open source that kind of diffuses the power and creates like a lower bound for everyone. So there's not too much of a gap between the, the haves and have-nots of access to private models. And so, yeah, I would say, you know, I think OpenAI changed a lot culturally, actually, over the past couple years, especially after the shakeup. I think— Shakeup, that's it. That's an understatement, right? But, you know, we'll see how Mira's Lab, you know, orient themselves on the spectrum of, of sort of, um, safety to, to, to full bore acceleration. You know, to be fair to, to, to OpenAI, they're, they've been pretty, they've been accelerating quite a bit. They're not holding back much. They're setting the pace for everyone else. And yeah. And then xAI is like really close up there, really trying to give them a run for the money. And then, and then that keeps them, you know, pushing further. So I think that's good. And that's basically was the fundamental thesis a year, a few years ago. It's like, if you only have one player, they're going to become complacent. They're just going to like focus on capturing the market rather than like innovating. Whereas if you have open competition, everybody has to accelerate or lose. Yeah. Right. And here we are and everybody benefits because we get these, these amazing models for like pretty cheap, like paying like $50 a month sounds like a lot, but like you have a PhD level like analyst that you could just like type to. It's crazy. It's great.

And so is LEAs are actually a greater accelerationist than you are. How so? How would that be? My theory is that he's scared if— okay, you go back to like Google DeepMind. Yeah. AI is kind of contained in one spot. He freaks everybody out, in particular Elon, causes them to form OpenAI, right? Sets the whole arms race in motion. I think he actually like completely undermined everything he was trying to do and created the arms race.

He was the best salesman.

Yeah.

Like fear-based marketing. So he was the best salesman for it. Possibly beat Jeff Bezos. Yeah. Yeah, maybe. I mean, kudos to him, right? Like, thanks to him, you know, we are here now. I think I'd be building this even if the current— I mean, we started this company before ChatGPT, right? Like, I just— it was just from first principles. This seemed like the future, but Um, I think this current, this current race, I think was definitely based on like this, not cult he started, but yeah, more or less, right? This cult of like AGI or, well, again, I have my reservations for that, but human level AI would be a really potent thing that would like eat the world. And it's like, oh, that sounds great. Maybe I can make a company that works on that. Right. And so he, he really helped with the marketing there, but also he, he helped create a culture where like, hey, we have to assemble a team and like make sure the right people are in control of it. And that appeals to all sorts of, you know, high-powered nerds. Yeah, yeah, it worked. It worked out well. It was really smart.

So he like, whatever he was afraid of, he activated the exact people he probably does not necessarily want to be doing that. Um, um, okay, last question. Uh, This is a tough one because it kind of depends on whether you guys fully succeed or not. I mean, when we talked before about the US and China kind of being on their own paths, you know, the more and more I visit Europe and go to Europe, I mean, it feels like a museum in some ways. They don't even feel like they want to participate necessarily in, in this next era. Um, they're sort of ceding it as far as I can tell to, to the US and China. I don't think Russia— I think Russia has definitely the smart enough people to pull this off, but I don't think has like the economic and, and, um, kind of like pragmatic wherewithal to compete with these data centers at this point. Um, and then, you know, I don't think anybody else necessarily does. Yeah. I mean, do you see, do you see a world where this is like a US-China thing going forward? Or I guess that's the entire, one of the entire points of what you guys are trying to do would be to avoid a scenario like that.

I mean, well, Europe, you know, right now they're, they're kind of captured by decelerationist culture. It's full spectrum. It's like anti-nuclear, anti-tech, anti-progress. I think that has to change, right? Like the whole thesis is that systems that have cultures oriented towards growth succeed and, and are sort of selected for, and those that are against it are dissipated and kind of pruned out of the evolutionary tree of nations, if you will. So it's either accelerate or die. That's, I mean, we put it pretty simply and I really hope like Europe kind of has a cultural shift. I know maybe they don't like our current culture and so on, but they're going to have to have their own variant or they're in trouble. Um, I think there's also like, you know, the Middle East. I don't know if you've been there, but they're going. All in on like building these data centers and they're trying to transduce their, their oil wealth into compute wealth. And, you know, I mean, for them, it's like they have a big stake in the, in the Kardashev scale and, you know, what's going to create more demand for energy is a machine that turns energy into value. Right. And if we're creating the most efficient machine at doing that, then from Jevons paradox, there's even more demand for energy. So I don't know, that's like their game there, probably. Um, but they're, they're, they're full acceleration, right? And obviously there's, there's ties between the VC ecosystem and, and, and, and over there. But, um, like, I think like there's the two big players like duking it out, right? And, but like this, this, this competition is actually what is setting the pace for the acceleration of this technology. And then everybody kind of wins, right? From it., right? Like even like the leftovers, the byproducts of this race, it's like, oh yeah, this model's like a year old, just going to open source it. And then anybody can download it. And that's huge value other nations wouldn't have had access to. Right. And that kind of, that's the rising tide that floats all boats. Right. And so of course, like US and China are going to like compete to have like the most cutting edge stuff, but I think it's going to, it's going to help everyone who's open minded to it. Right. And that's why I've been trying to scale this culture of open mindedness to AI and embrace AI. It's actually out of empathy because if you don't embrace AI, you're going to fall behind. It's pretty simple as an individual, as an organization, as a nation. Um, that's the message essentially.

So, okay. Okay. Yep. Thanks so much, man. Thanks. That was fun. Thank you for bearing with us. Yeah, for sure. Our guest today was Guillaume Verdon. Thank you so much for being here. I am your host, Ashlee Vance. Core Memory is produced by me and David Nicholson. Our theme song is by James Mercer and John Sortland. Our show is edited by the wonderful, delightful John Sortland, and we are sponsored by E1 Ventures. Thank you all, and we'll see you again.