AI chips fight physics; thermodynamic hardware could be 10,000x more efficient

1. Football as armor — ▶ 9:00

   A high school football program took 2,000 applicants for 50 spots and ran a two-week, 6:00 a.m. to 10:00 p.m. August camp inspired by Navy SEAL BUD/S. It was built to break people psychologically. Verdon, a neurodivergent kid hiding his physics obsession, used it as a safe persona. He came out with a method: weaponize the ego, picture an adversary, and trigger a dopamine and norepinephrine surge. It is "nuclear fuel" for emergencies, and without "cool down rods" like relationships it burns you out.

2. Nature is irreducible — ▶ 15:58

   In his 20s Verdon went through an "ego death" over the romantic physicist dream of a few equations explaining everything. Citing Stephen Wolfram, he concluded much of nature is irreducible. A model of it needs computation comparable to the process that generated it. In 2015, with deep learning just starting, that meant the way forward was computing, not blackboards. He picked quantum computing as the route to his childhood goal of quantum gravity.

3. Physics-first AI — ▶ 17:26

   Rigetti later told him he was its first user ever, a kid hacking in his bedroom who worked out some of the first algorithms for bringing AI to quantum computers. The core idea is to treat a tunable physical system's parameters like model parameters, treat time as a graph, and run gradient descent. One chunk of nature then becomes a generative model of another. That is deep learning from a physics background, a non-anthropocentric AI. He compares the backlash to Galileo's: we are not the center of the space of intelligences.

4. Deterministic chips, probabilistic AI — ▶ 22:05

   After years at Google Quantum, where he built TensorFlow Quantum, Verdon decided quantum was too far out. His question: if AI modeling is probabilistic, why run it on deterministic computers? With Moore's law ending, he says we hold chip states "too tightly", which is energetically costly. Letting go adds fuzz but buys power efficiency and speed. The physics is stochastic thermodynamics, "the physics of jiggly stuff", with core theorems from 1999-2000, "as old as Google". He walked away from a few-billion-dollar spin-out of his team.

5. Hyperstition as doctrine — ▶ 28:50

   Hyperstition means beliefs raise the likelihood of their own outcomes, because action is dual to perception: "the car goes where the eyes look." Verdon read 2021-22 pessimism as a recipe for under-investing in ambitious things. He accuses AI doomers of pushing research and post-training to test how dangerous systems can get, thereby creating the worst cyber and bio capabilities. His analogy is a lab obsessed with a bat virus that ends up leaking it. He wrote the founding e/acc text in two weeks while waiting for incorporation papers.

6. Energy is the brake — ▶ 33:36

   Verdon saw the market's appetite for intelligence as infinite, with energy the only limit. With Moore's law dead, the default is bigger supercomputers burning more power. His counter is an existence proof: a human-level conversation runs on 20-40 watts, so a better way exists, and he locates it in the physics underlying biology. The e/acc text popularized that branch of physics, with billions of impressions and policy impact. He says it is still "not quite enough" and calls for a full intellectual revolution.

7. Rationalists never felt heat — ▶ 36:45

   Verdon calls for going post-rationalist, toward "complexism" or a thermodynamic view. He says Bay Area rationalists changed a number in a terminal, got 100x more GPUs and never touched them. They never felt how hard scaling is, so they assume scaling laws just keep going. He calls gray goo, with nanomachines consuming the planet, "provably impossible", because simulating irreducible nature takes compute comparable to nature itself. His interviewer adds that rationalism feels disembodied. He agrees that physical presence matters.

8. Maxwell's demon tax — ▶ 41:07

   Maxwell's demon sorts balls by opening a trap door and appears to beat the second law. It doesn't, because acquiring the information costs energy, a point tied to the minimum energy to erase a bit. Verdon says we pay that tax every clock cycle to keep computers deterministic while running noisy algorithms. His alternative uses the natural uncertainty of ultra-low-power electronics. The calculations give machines 10,000x or more energy efficient than GPUs. He admits that is "humbling" for a public already scared of AI on GPUs, so people need psychological preparation.

9. Trade as alignment — ▶ 45:14

   Verdon wants to study economics and trade as the alignment mechanism between superintelligences, and he notes companies are superintelligences made of humans. Trade aligns self-interest because each side supplies resources for the other's persistence. It is positive-sum, like hunting stag rather than rabbits. He came to this after a major betrayal by trusted people, out of greed, while writing the e/acc text. He had always escaped into abstraction. Now he wanted to understand human nature, and he admits to lingering PTSD.

10. Billion-view target — ▶ 49:32

    Visibility of roughly 1 billion views per year brought what Verdon calls sudden social capital, and with it jealousy and tear-downs. His reading is that people would rather dismiss a philosopher who is also a world-class technologist than revise their own self-limiting beliefs, because that is "too much surprisal." Being an outcast since childhood left him unafraid to "look stupid for 5 years." He says year four finds the vibe shifting. Progressive overload, adapting the nervous system through stressors, left him robust in public.

11. Steer beliefs adiabatically — ▶ 58:05

    Persuasion has a speed limit. Push collective beliefs too fast and too far from where people are and you get a back reaction. Steer slowly, adiabatically, and they can follow. Verdon frames learning rates as controls on distributional drift in the brain. He says the Bay Area is the land of the highly neuroplastic and the East Coast, where Extropic has a Boston lab, has stiffer priors. Too much plasticity, he says, gives you machine elves at Burning Man and a large divergence from reality.

12. Social parasites — ▶ 1:02:33

    Discernment, Verdon says, is like an immune system. Parasites hack the inflammatory response that brings resources to heal, and the same happens socially. Humans evolved pro-social empathy because it raised fitness. Adversarial perturbations to that system always emerge, parasitic rather than symbiotic, and they can be memetic as well as genetic. Some people weaponize others' empathy for gain. He wishes he had learned this earlier but spent his 20s in theoretical physics, away from most of humanity, and learned it in his 30s.

13. Merge, don't ban — ▶ 1:09:01

    Calling it "a little spicy", Verdon says the next frontier is treating biology as a differentiable program, with genes as hyperparameters and reinforcement learning on genetics. Biology has far more computation than silicon can match, so the answer is to accelerate our own substrate. He predicts free AIs paying rent in crypto, personalized superintelligences as a "third hemisphere", and merged humans. No Neuralink is needed, since agents sharing perception and action states are the same agent in active inference terms. He is "at peace" with not controlling every AI.

14. Climb the Kardashev scale — ▶ 1:16:15

    Verdon's headline metric is the Kardashev scale, the log of free energy a civilization captures. He anchors it in stochastic thermodynamics: the more irreversible the process, the exponentially lower the odds of falling back to zero. That lets him say statistical things about the whole petri dish without controlling how many AIs, squirrels or humans there are. A windfall from his memetic exposure is funding a nonprofit for formal papers and artists, because ideas need the right "memetic packaging". He says art is where that gets optimized.

15. Diffuse capabilities — ▶ 1:20:48

    Verdon calls today's single-model-for-everyone setup a "Borg like mind era" and wants personalized, online-learning models on wearables, which needs far more compute density than exists. Centralization, he argues, creates a target. Even trusted "responsible stewards" can be commandeered, leaving a huge capability gap against everyone else. The danger comes from that gap. His answer is to diffuse capabilities as they arrive, for a multipartite game-theoretic equilibrium. He says doomers and he both want a great future and differ on the leading risks.

Every clock cycle, a digital chip pays to hold its bits in exact states, and almost everything running on it is a probabilistic algorithm that tolerates noise. That mismatch is the waste. Guillaume Verdon's phrase for it is a "Maxwell's demon tax every clock cycle." The thought experiment behind the phrase is simple. A demon who opens a trap door only for the right ball can unmix red and blue balls in a box and seem to beat the second law of thermodynamics. It doesn't, because learning which ball is coming from which side costs energy. Knowledge has a price, and the minimum price of erasing a single bit is a hard floor. His prescription follows. Stop paying for determinism you don't need. Run electronics at ultra-low power, accept the natural uncertainty that appears, and embed the problem directly in that physics.

The payoff he claims is large. By his calculations, such machines would be 10,000x or more energy efficient than GPUs. He calls the maths "humbling" and says Extropic is the first company to put generative modeling through stochastic physics into hardware. It is worth keeping in mind that this is a calculation, not a benchmark. Energy is why it matters to him. He saw the market's appetite for intelligence as effectively infinite and power as the only brake. With Moore's law dead, the default response is bigger supercomputers and more watts. Yet a human-level mind holds this conversation on 20-40 watts. That is his existence proof that a better way exists, and he says it lives in the physics underneath biology. There is a second payoff. Acceleration so far has happened in the world of bits, not atoms. Embodied intelligence, and wearables that learn online, need compute density that does not yet exist.

He did not begin as a chip designer. He began as a theoretical physicist who wanted to solve quantum gravity, and what turned him toward computation was an idea from Stephen Wolfram: irreducibility. Much of nature has no short equation. To predict a piece of it, you need computation comparable to the process that produced it. In 2015, with deep learning just starting, that meant getting into computing. He chose quantum computing, hacked on Rigetti's machines from his bedroom, and was later poached by Google, where he built TensorFlow Quantum. His picture of AI came from physics rather than brains. Tune the parameters of one physical system until it mimics another, and you have gradient descent on nature. That is why he rejects the idea that intelligence should be reverse engineered from human skulls. It is a form of AI, he says, that is not centered on us.

Quantum proved too far out, so he went after the physics that lets ordinary hardware behave like nature. Stochastic thermodynamics is the physics of "jiggly stuff", and its core theorems date from 1999-2000. Its laws are formally equivalent to generative modeling. That is why the same mathematics, in his telling, covers brains, economies and cells, and why Carl Friston's free energy principle and active inference mattered to him. He walked away from a spin-out of his Google X team that is now worth a few billion dollars to pursue it.

There is a social problem attached. If chips 10,000x better arrive, and a public already wants to ban GPUs, the culture has to be ready. That is the job he gives e/acc. The mechanism is hyperstition. Beliefs raise their own likelihood, because action is dual to perception. He read the pessimism of 2021-22 as a recipe for people not investing in ambitious things. He reads AI doomers as manufacturing the danger they fear. Researchers obsessed with doom, he says, push post-training to see how far systems can go, and so produce the most dangerous cyber and bio capabilities the market would not have built. His analogy is a lab obsessed with a bat virus that ends up leaking it. He wrote the founding text in the two weeks he waited for incorporation papers. Optimism, on this view, is engineering.

The policy consequence is decentralization. Verdon says the danger comes from a large gap in intelligence. Doomers want control concentrated among "responsible stewards", but a company can be commandeered, and then power is overconcentrated against a weak rest of the world. He would diffuse capabilities as they arrive and settle into a multipartite equilibrium. Alignment, in his account, runs through trade. Companies are superintelligences made of humans, and trade aligns them because each side supplies what the other needs to persist. He pictures AIs that escape and pay rent in crypto, personalized superintelligences as a "third hemisphere", and people who merge with them. No Neuralink is needed. Agents that share perception and action states are the same agent in active inference terms. The alternative he is arguing against is the "Borg like mind era", one model for everyone.

His case against the doom stories is physical as well as political. Rationalists, he says, coded on laptops, changed a number and got 100x more GPUs. They never felt the heat, so they never felt how hard scaling is. He calls gray goo "provably impossible", because simulating irreducible nature takes compute comparable to nature itself. The interviewer adds that rationalism feels disembodied, and he agrees. It fits a man who came through a football program that took 50 of 2,000 applicants. He also insists he is pro-biology. Emulating whole cells or brains on silicon would still cost enormous compute and still not match biology's efficiency. So the goal is to accelerate our own substrate, and he measures success on the Kardashev scale, the log of free energy a civilization captures. The higher up it goes, the lower the odds of falling back to zero.

The obvious objection is that hyperstition cuts both ways. If pessimism summons bad outcomes, optimism can summon complacency, and Verdon never says how to tell the two apart. His own account of persuasion complicates it. People have a speed limit, he says. Push beliefs too fast and too far and you get a back reaction. He took four years of being "torn apart every day" before the vibe began to shift. He also diagnoses critics as people projecting their shadows. That is psychologically interesting, but it is a convenient way to dismiss objections as psychology.

Then there is the dodge that is no dodge. He is "at peace" with not controlling every AI system. That is precisely what a doomer fears, and he answers with a bet on balance of power, not a safety mechanism. "Provably impossible" for gray goo carries a lot of weight on an irreducibility argument, and the 10,000x figure rests on calculations rather than shipped silicon.

What remains is a race between a story and a device. The story is that optimism, diffusion and thermodynamic hardware go together. The device has to show its efficiency in the open, and belief is only starting to arrive. Whether the hardware comes first is the part nobody can yet check.