Forget Everything You Believed About Computing w/ Gill Verdon | EP #102

Once humans and AIs of similar intelligence are created, both species will learn to interact productively and employ AI toward solving civilization-scale challenges, with abundant work remaining at all skill levels as civilization scales up.

“In my case I think once we've created AIs that are, you know, of similar intellect to us, we're going to learn how to interact with them, we're going to learn how to employ them in a in a positive fashion. Uh there's going to be a big adjustment there, but once we do so, we're going to have way more challenges for us to scale civilization to the stars and there's plenty of challenges left, so there will always be more work to do. Right? And and we're going to put our most capable systems on the most complex and difficult tasks and there's still going to be work left of all kinds.”

The Kardashev scale measures civilizational progress through energy production and consumption, with Type 1 civilizations harnessing all energy incident on their planet from the sun, Type 2 capturing all energy from their star via a Dyson sphere, and Type 3 controlling energy at the galactic scale; humanity is still well below Type 1.

“Um there's originally there was uh three types, there were three big milestones, and then Carl Sagan found a way to interpolate between the milestones so we have a continuous uh scale there. Um but the original scale was uh in type one the type one Kardashev's uh scale civilization Civilization. uh would would produce as much energy as is incident onto Earth from the Sun, right? So, if you take the Sun, we occupy a certain amount of solid angle, sun certain amount of the Sun's rays hits us, that's a certain amount of power. And by the way, uh the numbers that I remember cuz I I speak about this when talking about energy abundance, is that today I think there's 8,000 times more energy that hits the surface of the earth that than we consume as a species in a year. We're still We're still really early. We're still below type one. Well below type one. Type two would be having the equivalent of the energy production. Dyson sphere that that captures everything being emitted by our sun and type three would be the entire [30:14] galaxy, right?”

The human brain operates at 14-20 watts and contains 100 billion neurons and 100 trillion synaptic connections, making it far more energy efficient than current AI systems which would require tens or hundreds of millions of times more energy for equivalent capability.

“100 billion neurons, 100 trillion synaptic connections running on what n- how many watts do you Tens of watts, right? Like I I I think like 14 to 20 watts of of energy. Uh but the equivalent for a GPT-4 system [4:10] would be what? Tens or hundreds of millions of times less efficient.”

Stephen Wolfram's theory predicts that certain natural systems are computationally irreducible, meaning their behavior cannot be compressed or predicted by shortcuts and must be computed directly, implying nature is fundamentally hard to predict at all scales.

“There's this beautiful theory by by Stephen Wolfram on on complex and self-organizing systems and his prediction is that certain uh systems in nature are [34:22] irreducible. You can't you can't get a, you know, a TLDR. You can't compress, right, the gist of it to something very simple. You actually don't have a choice but to go through the highly complex computation to predict what's going to emerge as a behavior at a different scale.”

Understanding physics, chemistry, and biology at deep levels requires orders of magnitude more computation than distilling human intelligence from text, and will require scaled AI systems to perceive, predict, and control the physical world.

“I do think it's possible. I do think it's much harder than distilling human intelligence. I think uh understanding biology and chemistry [33:52] is orders of is going to take orders of magnitude more computation. And so the computers we're building, yes, they'll be able to run uh models that are anthropomorphic, right? Like LLMs and so on, trained on on human data. Um but ultimately it's machines that are going to help us grok the physical world.”

If systems are stiff and inflexible, they experience catastrophic failure when hit by disruptive forces; to be antifragile and robust, systems must be malleable, flexible, and adaptive.

“If they're too stiff, they break and they have catastrophic failure, which is what we're trying to avoid. Um and so part of the message of acceleration is for us to be robust and adaptive to whatever's to come.”

The average human today has access to ubiquitous, cheap technology (iPhones/Android) that is the same for CEOs and ordinary people, demonstrating how capitalism creates universal access to powerful tools.

“you know, we're both CEOs, right? We both have like standard issue iPhones or Androids, right? Um And it's the same as anyone else, right? Every Everybody has access to to the same technology, and it's ubiquitously accessible and cheap, and that's beauty of capitalism, right?”

Science fiction and Hollywood movies have shaped people's priors about the future in ways that promote doom narratives and fear of AI, rather than realistic assessments of technological potential.

“People have sci-fi based priors, right? Like what are what are their priors on on what the future holds. They've they've seen a lot of sci-fi movies. I blame Hollywood. I really do.”

It is provably impossible to predict the future perfectly; if perfect prediction were possible, a small number of people on Wall Street would have monopolized all wealth, but they have not, demonstrating that future complexity is irreducible.

“We cannot predict the future. You provably [1:14:45] cannot predict the future. If you could, there would be a couple people on Wall Street making all the money right now, right? And and clearly they they can't, right? You can't reduce the markets to a simple model. And so you can't predict the future.”

Humanity needs to up-level its ambitions and challenges as AI handles routine work; without struggle and purpose, humans face existential malaise, as demonstrated by the Universe 25 mouse experiment where mice in a resource-abundant environment with no challenges died off.

“There's a great uh paper I read recently um uh it's called Universe 25 if people want to Google it. And so the the studies were done in the '60s in which a large open space was created for field mice. Mhm. And and it had all of the [28:09] room and the nests and the food, and they had no struggles at all. And the mice, you know, breeding pairs were put into this, and they grew and they grew and they grew, and then after a couple of generations, it basically died off because there was no struggle in there.”

Current large language models are trained on human-generated data and thus distill a mixture of human intelligences, constraining them to near-typical human intelligence until they are embodied and can interact with and query the environment independently.

“I think that the current approaches where we train on human-generated output, right? The internet is broadly, at least right now, I'm sure in a couple years it won't be the case, but generated by humans, right? And we're kind of distilling a mixture model of all human intelligences, right? You could think of the LLMs as trying to distill a mixture across the outputs of all our brains, right? And so at least in to me it seems like it would saturate to something nearing, you know, typical human intelligence. Um Until Well, until it's embodied and then can interact with the environment and get its own samples and and query the environment in a way that, you know, isn't bottlenecked by what was generated previously by a human.”

The market is a powerful aligning force for AI development: products without positive utility fail to gain users and companies fail, creating selective pressure toward aligned, interpretable models that perform well, which is gentler and more effective than top-down regulation of compute.

“I would say that um you know, the market itself, right, um is a very powerful aligning force, right? If you have a product that is not of positive utility to us, we don't buy it. It runs out of whatever company makes that product runs out of capital and the product dies off, right? There's a selective pressure on the space of products. And uh right now because uh el- AI models are products, right? You have model-as-a-service companies like OpenAI, Anthropic, eventually XAI. Um this competition of for users and ultimately of capital to, you know, fuel the GPUs that keep these systems alive, um uh this competition induces a certain selective pressure and models that are not aligned, that don't do what you ask it, that are hard to interpret, hard to read, actually don't do well in the market.”

Effective Accelerationism (E/ACC) is a philosophical framework that argues civilization should maintain variance, dynamism, and adaptability rather than slow down or freeze in response to uncertainty, drawing from physics of complex self-organizing systems and out-of-equilibrium thermodynamics.

“And essentially what I saw was that we need to maintain variance and dynamism as a core uh, value for us to be malleable and adaptive to whatever challenges come our way rather than trying to freeze everything, slow down and panic. [9:55] Actually, what we want is actually more malleability, more dynamism, more acceleration.”

The future of human-AI augmentation should enable individuals to own and control their own AI systems rather than relying on centralized AI services, ideally resulting in 8 billion or more personal AIs that serve as extensions of individual cognition and can be fine-tuned privately.

“I think if we're going to truly augment oursel- our own intelligence with AIs, I think in order to maintain the benefits of having individuality, right? Individuality we celebrate individuality in our society, at least in the West, and it's our greatest strength because this variance, everybody brings something different to the table... And if we only have centralized models, right? A few models that are trained for everyone, they're amortized so so so there's one model for everyone, we lose the benefits of having sort of individuality. And so to me, my quest both on Eon and Extropic is for people to be able to own uh the compute that is an extension of their own cognition. And for them to have the right to run their own AIs and have the right to fine-tune models in a way that's private. So, 8 billion AIs basically. Or more. Or more.”

People who embrace technological change and augment themselves and their businesses with AI will have a place in the future, while those who reject it will be left behind or relegated to alternatives like the Amish model of technological rejection.

“if people have a mindset that they want to embrace technological change, they want to figure out how to best augment themselves, augment their businesses with AI, they will have a place in the future. Those that want to, you know, stay away from it, then, you know, I mean, they could go back to the cabin. Amish we have the Amish example Yeah, yeah, the Luddites.”

Dinosaurs went extinct because they lacked genetic variance and malleability; in contrast, furry mammals possessed variance in their phenotypes and were thus able to adapt to the asteroid impact and environmental change, demonstrating why biological variance confers survival advantage.

“The dinosaurs that were slow and lumbering were not malleable, they were not adaptive, and they died. But it was the furry mammals that Right. flittered around, jiggled around, to use the free electron analogy, that ended up becoming dominant.”

Thermodynamic computing is a third paradigm of computation alongside deterministic (classical) and quantum computing, where the computer is deliberately kept in a probabilistic state rather than trying to maintain determinism, achieving dramatic energy efficiency gains.

“Yeah, it seems like it cuz today we we have the deterministic computers, right? Your transistors are definitely on or definitely off. One or zero. It's one or the other, right? And you definitely know which one it is, right? Uh and then you have a quantum computer which has superpositions of ones and zeros. It's 0 + 1, 0 - 1. And everything in between. Everything in between, complex numbers, you can make those interfere with one another. But actually having a computer that you're unsure of the state of the computer and it's probabilistic, it's 0 or 1 or [49:53] something in between, but um you're not sure exactly the state of the computer, is actually much more energy efficient because knowledge costs energy.”

Modern deep learning algorithms co-evolved with GPUs—algorithms that ran well on GPUs were selected for, creating a bidirectional feedback loop where hardware and algorithms reinforce each other; thermodynamic computing aims to create a similar co-evolution fork in the space of both hardware and algorithms.

“turns out matrix multiplications which GPUs excel at are very useful for all sorts of different applications including AI, but as as you mentioned, GPUs weren't designed from the ground up from first principles to be AI processors, right? It's kind of a co-evolution between the hardware and the algorithms, right? The algorithms that ran on GPUs like modern deep learning tended to do well because GPUs [48:21] already existed and then both kind of fed off each other. Right? So, we're trying to create an evolutionary fork in the space of hardware. It's going to engender evolutionary forks in the space of algorithms and they're going to co-evolve. That's why we're a full stack company and we co-design the algorithms and the”

Humans and AI will eventually couple through technological augmentation, with humans embedding AI as extensions of their cognition—analogous to how prokaryotic cells absorbed mitochondria to become eukaryotes—enabling dramatically accelerated capabilities.

“The coupling with AI means that we get a chance to uh to accelerate alongside it, enabled by it. It's almost like when life began and we had these prokaryotic life forms and they absorbed mitochondria Yeah. and became a eukaryotic life form, was able then to utilize the the free energy of oxygen Mhm. uh for oxidation to grow more rapidly.”

Simulation hypothesis fails because information density has limits; if there is a base reality with its own physical laws, it has maximal information density constraints that make nested simulations (simulation within simulation within simulation) impossible beyond a certain depth.

“there's a certain um uh density to information, if you will. Um and and and my work on on quantum internet was actually studying how densely you could pack quantum information in in various substrates. And in a sense, this um assumption that you can have a simulation within a simulation within a simulation really doesn't hold, cuz at the end, there's a base reality, and in that embedding universe, you have a maximal uh information density for your computer.”

Verdon's long-term mission is to increase the amount of intelligence per watt and the total amount of intelligence in the universe, which he considers a goal prescribed by thermodynamic principles rather than a human invention.

“I think a lot of our goals, people's goals are too anthropocentric or they they want to do relative to others. And now that AI comes in and kind of breaks this sort of zero-sum competition between humans, right? We got to set our our sights on a non-anthropocentric goal, right? And we have a goal prescribed by the universe in a sense, which is to grow. Yes. Um [31:17] because any life form seeks free energy and looks to grow.”

The highest leverage moonshot for individuals to pursue is not to compete with others on incremental improvements but to aim at the Kardashev scale—to contribute to scaling civilization's energy capture and intelligence production rather than zero-sum human competition.

“I think a lot of our goals, people's goals are too anthropocentric or they they want to do relative to others. And now that AI comes in and kind of breaks this sort of zero-sum competition between humans, right? We got to set our our sights on a non-anthropocentric goal, right? And we have a goal prescribed by the universe in a sense, which is to grow.”

Extropic aims to reduce reliance on TSMC (Taiwan Semiconductor Manufacturing Company) as the single dominant supplier of cutting-edge AI chips, reducing geopolitical vulnerability and supply chain risk in the event of Taiwan-China conflict.

“our goal is to have a different way to embed the problem into uh the devices, and for us to not depend on the same processes that everyone else depends on, which we feel is important hedge, because otherwise, Taiwan is a very sensitive uh area, and the world supply chain depends on it. And so, if we could, you know, forego our reliance on on Taiwan to make the most cutting-edge chips for AI, that would, I think, put everyone at ease and be a net benefit.”

The debate between techno-progressives and techno-regressives is a new axis of polarization, where accelerationists bring balance by providing a counterweight to the prevailing narrative of regulation and precaution.

“It's kind of a new axis of of of polarization... there's some people that are uh you know, positive sum and abundance mind mindset... and the people that uh you know think technology maybe is is net negative... it's kind of techno-progressive versus techno-regressive... brought balance to this force of sort of novelty-seeking... versus... more order, more constraints.”

Verdon intends to partially open-source the concept and theory of thermodynamic computing so a broader community can join the development effort, rather than keeping it proprietary.

“part of my goal is to uh partially open-source the concept of a thermodynamic computer, so a broader community can join us on this quest. You know, we're one effort for this, but it's too important of a technology to to keep it on a shelf for for a few more years. We don't have time.”

People should think like capital allocators and managers, directing AI agents as employees; this mindset shifts the future from fear (what will AI do to us) to empowerment (what can we do with AI).

“I think I think people should should think more like that, like a a capital allocator, like a manager, as the way we're going to merge with AIs. [38:31] Um and it's going to allow them to do much more. I think more people should be entrepreneurial, more people should think, "Hey, what opportunities do I see in the world now that we have these AIs that are on the verge of human-like intelligence? What should I aim to build? How will I direct capital to unlock more value?"”

Humans all have the same 24 hours per day, 7 days per week, 365 days per year, so the differential between people is how they use their time; personal AI agents that multiply productivity are massive force multipliers for time allocation and thus for human advantage.

“the reality is um what makes the the leveling the playing field for humanity is the poorest and the [38:00] wealthiest all have 24 hours in a day, 7 days in a week, 365 in a year. It's how you use your time that differentiates you. So my ability to have ultimately hundreds of of extraordinary agents that can do do my the things that I desire and bring back the answers for me is a massive uh force multiplier for productivity.”

Claude 3 has an IQ of 101, which exceeds human average intelligence, demonstrating that current AI systems are already superhuman in raw cognitive capability, raising the stakes for how we govern such powerful technology.

“Uh Claude 3 is already at 101 IQ, which is more intelligent than humans, and we're giving infants nuclear weapons to play with, and these things are going to [24:02] um ultimately destroy society.”

The E/ACC manifesto emerged from late-night conversations with technologists in online communities who use anonymous accounts to avoid employer retaliation, discussing where civilization is heading and sharing candid thoughts about progress and technology.

“I was having conversations late at night with other technologists sort of online. There's sort of communities where people have [7:18] anonymous accounts, right? Uh cuz they're employed by XYZ and they don't want their opinions to reflect that of their their employer and they want to have the freedom to experiment with their thoughts, right? And have candid conversations of like, 'Hey, where is civilization going? Where is society going? Where is this all going, you know?' And we would have these conversations late at night and and essentially we decided to write something up and uh, you know, I added my own twist uh to it and uh, you know, put out the manifesto.”

Most of the energy and time cost in running thermodynamic chips is consumed by the classical computers interfacing with the chip, not the chip itself, meaning the bottleneck shifts from computation to I/O.

“Because for us, most of the energy and time is actually consumed by the computers that have to interface with the chip than the chip itself. Um which is really weird to think about.”

Quantum computers could empirically test whether we live in a quantum simulation by reaching qubit counts where there would not be enough atoms in the observable universe to classically emulate them.

“One thing we can test is whether we live in a quantum simulation or not, right? Um because I think uh Google's quantum computers, amongst others, are reaching the number of qubits where there wouldn't be enough atoms in the observable universe to to emulate the quantum computer with a classical computer. And so, we can at least rule that out.”

During his time at Google, Verdon observed that generative AI workloads were consuming more and more compute internally, leading him to recognize that classical digital computing might not be the right substrate for AI, and that a different physics-based approach would be necessary.

“But during that time in quantum computing, I realized that actually there was going to be different nodes of our our our tech tree that need development imminently that use a different kind of physics that's not quantum mechanical physics and that would be much more useful for for generative AI as I was seeing sort of generative AI workloads eat more and more of the compute internally at [47:19] Google.”

Extropic AI will initially release application-specific thermodynamic compute devices at the edge with smaller neural networks but high speed and energy efficiency; only over time will it scale to general-purpose large models as the technology matures.

“at at at first, it's going to be, you know, application-specific devices. It's going to be smallish devices with not that many neurons, right? But they're going to be very fast and very energy efficient. Um there's all sorts of applications for that at the edge. But over time, yes, as the chips grow and more and more of the program uh becomes part of the physics of the chip or that become thermodynamic programs, then eventually we could run the whole program on the chip.”

Extropic's first prototype was built using superconducting technology operating at cryogenic temperatures because the physics required operating in the ultra-low-power, high-noise regime; subsequent silicon chips will operate at room temperature.

“So, your goal here is not something that's operating at cryogenic temperatures. It's something that is operating at room temperature That's right. built on a silicon fab. Yeah. So, so, you know, when you prototype things, you start with the biggest, most macroscopic prototype you can, cuz it's simpler, it's easier to get going, uh and you can probe it and understand it better. In our case, we couldn't do a a breadboard prototype, you know, like you would do with other electrical [1:04:26] circuits, because we want to operate it in the regime of ultra-low power and the rate ratio of power of noise where we have the right properties of of the electron physics. So, for us, the first prototype is the most macroscopic we'll make of a thermodynamic computer, but we had to supercool it, because that's how the physics works out. Uh essentially, uh we did a a superconducting prototype. It's our first prototype, but our next chips are going to be uh in silicon, and we're really excited about talk about embedding physics [1:04:57] uh and embedding the algorithms.”

Verdon attended University of Waterloo for his master's degree in quantum gravity and quantum information, and wrote down equations for his first thermodynamic computing chip while still an undergrad/early grad student in a notebook before building credibility in other areas.

“You were at University of Waterloo? Yep. And studying quantum physics? quantum yeah, quantum gravity, quantum information uh during my masters.”

After building TensorFlow Quantum, Verdon worked on quantum communications, sensing, and analog-digital conversion as part of a team under Sergey Brin, completing the stack for quantum perception and control of the physical world at the quantum mechanical level.

“And then you actually spent some time working closely with Sergey Brin. Yeah, so uh after we, you know, built a TensorFlow Quantum, there there's a team that was forming around Sergey working on quantum technologies uh and AI and physics and AI more broadly. Uh and and to me, I was sort of getting a bit uh impatient with the timelines with the compute the computing stack and I saw that there were opportunities in quantum [46:15] communications and sensing that were maybe shorter term. And so, I wanted to try my hand at that. And to me, it was a a completion of sort of the vision of understanding the world at a quantum mechanical level cuz if even if you have the algorithms running on quantum computers that can understand quantum data and learn AI representations of them, how do you acquire quantum data and how do you transmit it? And so, that's what I worked on. Uh so, I worked on quantum analog-digital conversion, the US quantum internet.”

Alexander Weisman Gross discussed at the Abundance Summit the prospect that AI will reach AGI and digital superintelligence with no theoretical upper bound to IQ, and the critical question is whether humanity couples with that superintelligence or decouples from it.

“at the Abundance Summit this year, um Alexander Weisman Gross was there, was talking about uh AI is coming on strong. Mhm. Um it is going to, you know, reach whatever we call AGI and [14:00] then digital superintelligence. There's no there's no barrier that says AI only goes towards, you know, an IQ of humans. It It blasts through and continues ad infinitum, especially if the hardware you're building comes into existence, or when it comes into existence, I should say. Um And the question is, does humanity couple with it, Yeah. or do we decouple?”