Fear is the civilization killer: Guillaume Verdon on e/acc, thermodynamics, and being unmasked

In December 2023, Forbes published a piece titled “Who is BasedBeffJezos, the leader of the tech elites e/acc movement?” The article fused two identities that had been living in one mind. Guillaume Verdon is a physicist, applied mathematician, and quantum machine learning researcher who earned a PhD in that field, worked on quantum computing at Google, and later founded Extropic to build physics-based hardware for generative AI. BasedBeffJezos, the previously anonymous X account, created effective accelerationism, or e/acc, which treats rapid technological progress as the ethically optimal path for humanity. Lex Fridman opens by holding both characters in view: a still-respected scientist building a company, and a meme account that launched a philosophy.

Verdon says he wanted a theory of everything as a child, a path that led through theoretical physics and information theory and a picture of the universe as one vast computation. Black hole physics, quantum cosmology, and quantum gravity were the extremes he studied, hoping a few equations might describe the whole show. The crisis of faith came when he decided that augmenting human perception, prediction, and control with machines was the real path forward, which pulled him out of pure theory and into quantum computing and quantum machine learning. He wanted not only to understand how the universe computes, but to compute the way nature does.

After about eight years in that field he noticed a gap. At the smallest scales things are quantum; at the largest they have averaged out into the deterministic. Daily life — proteins, biology, gases, liquids — lives in the mesoscales, where things fluctuate and thermodynamics rules. Life is not a soup of particles at equilibrium with nature. It is a coherent state that maintains itself by acquiring free energy and burning it. Citing Jeremy England at MIT, Verdon argues that trajectories of matter better at capturing free energy and dissipating heat are exponentially more likely, so the universe is biased toward certain futures. “We’re far more efficient at producing heat than, let’s say just a rock with a similar mass as ourselves,” he tells Fridman.

The anonymous account started as pressure relief. Verdon had absorbed Google X’s culture of secrecy, and he wanted a place where not every sentence was attributable. “Restricting your speech back propagates to restricting your thoughts,” he says. Unclamping those variables let him explore a much wider parameter space. He later describes the growth of BasedBeffJezos as New Game Plus: restart the game with the knowledge of how to beat it, but from scratch, so the ideas would be judged apart from his résumé.

Late 2021 and early 2022 felt, to him, soaked in virally amplified pessimism about technology and the future. Doom is hyperstitious: if you believe it, you raise the odds of it happening. Optimism has the same property. He felt a responsibility to engineer a mimetic optimism virus, a community of people doing the hard work of scaling civilization, because stagnation is not an option and there is more cooperation when a system is growing than when it is splitting a shrinking pie. e/acc’s later manifesto would put the same claim in thermodynamic language: the system wants to grow, and believing you have agency makes a better future more likely.

The unmasking was not consensual. Journalists matched his public talks to his voice on X Spaces, scraped SEC filings, and looked at a private Facebook account; notes across the newsroom were “sensor fused” until they had too much. They reached his investors, which, as a founder, felt like the only bosses he had. The phrase that rang alarm bells was “public interest,” allegedly triggered around 50,000 followers. He disclosed that he was Beff Jezos because he had a company to protect, and because they had him cornered. He still thinks the move was wrong: anyone who speaks truth to power and crosses a threshold gets exposed, after which the traditional apparatus knows where to apply pressure. That, he says, is a speech-suppression mechanism.

Fridman asks about a world of anonymous LLM swarms starting revolutions from a basement. Verdon’s own origin story is almost that joke: he quit big tech, moved back in with his parents, sold the car, bought about $100,000 of GPUs, and started building. He wants LLMs to have freedom of speech too, on the theory that it will induce freedom of thought in synthetic minds. The practical fix he likes is certified-human identity that still stays pseudonymous, plus clear labeling of bots — the direction Elon Musk is pushing on X. Degrowth and deceleration, he adds, often serve adversaries; he cites Germany’s nuclear shutdowns and later dependence on Russian oil.

The OpenAI board crisis is his exhibit A. A couple of board members had enough power to nearly collapse the organization. Quantum computers, he reminds Fridman, are fragile to noise; the answer is quantum error correction, encoding information non-locally so a local fault cannot kill the payload. Humans fluctuate, get corrupted, get bought. A top-down hierarchy where a few people control many nodes is not a fault-tolerant system: corrupt a few nodes and you corrupt the whole thing. Keeping AI’s power from concentrating, and keeping an adversarial equilibrium of players, is therefore a top priority.

Fridman reads four recommendations from a report by Yoshua Bengio, Geoffrey Hinton, and others: independent auditors; governments and companies putting a third of AI R&D into safety; mandatory safety measures if dangerous capabilities show up; and liability for foreseeable harms. Verdon accepts liability and expects third-party auditors to appear because customers will demand certified reliability. An arbitrary 30 percent budget is another matter: firms will spend whatever the market requires. What he truly opposes is “weird correlated interest” between incumbents and the state. “I just want separation of AI and state.” America’s advantage, he says, is variance, dynamism, and a free market that converges on useful technology faster than centralized control.

He does not believe in a fast takeoff or a hyperbolic singularity on a finite horizon. Progress looks like a compounding exponential, closer to Moore’s law: more people, more resources, more intelligence applied to a science that is adding value, which then draws more resources. The hedge against a runaway monopoly is open-source AI trailing not too far behind the frontier, so every hacker and graduate student can search the hyperparameter space. He also worries about hardware chokepoints — Nvidia, TSMC in Taiwan, ASML’s lithography — which is one reason Extropic is trying to explode the variance of how AI can be embedded in physics. He dislikes the term AGI as anthropocentric. Human-like intelligence is one point in a large space; quantum intelligence can grok entanglement no biological brain can represent efficiently.

The beautiful object, for him, is the “homo, techno, capital, mimetic machine”: humans, technology, capital, and memes, all applying selection pressure to one another. Options are grow or die. He expects humans to augment themselves with AI — phones, wearables, shared perception — while non-biological AIs coexist, aligned the way companies already align mixed human-and-tech superintelligences through economic exchange. Dan Hendrycks’s objection, that natural selection would favor AIs over humans, gets a domestication reply: we already induce evolutionary pressure toward AIs of positive utility, the way wolves became dogs. The upside of climbing the Kardashev scale with AI, he says, is unimaginable; leaving it on the table would be a betrayal of future humans.

Asked for a probability that AI eventually kills all or most humans, he refuses the calculation. A real probability would be a stochastic path integral over all possible futures, not the ones a neurotic brain prefers. Humans evolved to oversample scary futures; estimators do not normalize over a super-exponential possibility space. Chaos and Lyapunov exponents turn a little fuzz into a lot of fuzz. The only prior he trusts is physics: systems want to grow, and subsystems optimized for growth and replication are more likely in the future. Google’s quantum supremacy experiment, which he watched being simulated, was a quantum chaotic system whose outcome probabilities even the largest supercomputer could not estimate. If p(doom) casters were that good at prediction, they would already be rich in the stock market.

His darkest timeline is not Terminator. It is AI power concentrated in a few companies that control information flow, mixed with government, close enough that they could even convince the public AI never existed. That prior is data-driven: when humans centralize power, they do horrible things. He asks doomers to trace a Markov chain to doom and put a probability on each transition; very often one hop is unphysical. Fear is the mind killer, he says, and also the civilization killer. Centralized cybernetic control — one giant intelligence fusing every sensor — has never been the optimum. Nature and companies invent hierarchical loops: local contributors iterate daily, managers sample weekly, larger scopes sit above. If the CEO goes on vacation and the company crawls, the decoder was never fault-tolerant.

Intelligence, for Verdon, is cybernetic: perceive, predict, and control, which in practice means compressing the world by minimizing relative entropy between a model and reality. Black holes are the densest information objects known, which is how he wandered from their entropy into algorithms that search over quantum codes, and into quantum analog-to-digital conversion — capturing superposition from the world without destroying it. Chemistry, nuclear physics, the greenhouse effect, and parts of biology sit behind a complexity barrier that classical brains cannot efficiently represent. Quantum machine learning was supposed to unlock that. Over time the timelines started to look like nuclear fusion: always a decade away.

A quantum computer is a quantum system you can control well enough to keep it quantum. Nature at small or cold scales is more fundamental than probability, but brains cannot think in superposition, so we translate into linear algebra at exponential cost. Doing classical deep learning on a quantum machine, he would not advise: you assume zero temperature, which no physical system reaches, so error correction becomes an algorithmic fridge pumping entropy out of logical qubits. The overhead is like shipping a package across a city by rocket. Quantum deep learning earns its keep only on systems with correlations too quantum for classical representations — chemistry, nuclear physics, networks of quantum sensors, Fermilab’s dark-matter search, LIGO as a giant quantum sensor.

Fridman teases a paper, “A Universal Training Algorithm for Quantum Deep Learning,” which introduces baqprop with a Q. Everyone claimed quantum computers would speed up deep learning; Verdon’s reply was that the best way to predict the future is to invent it, “so here’s a hundred page paper, have fun.” The trick is a feedforward, then a phase kick — really a force kick proportional to the loss — then uncomputation, so errors back-propagate onto a superposition of parameters. Alternating that with an emulation of kinetic energy makes parameter space a quantum particle that can, in principle, tunnel to optima stochastic optimizers miss. In practice, on architectures anyone actually plans to build, the algorithm would be extremely expensive.

Why is hardware so hard? You must induce a near-zero-temperature subspace by encoding a code within a code within a code. If each component is below threshold, scaling up adds more errors than it removes. Recent years have crossed that threshold, which is why he is still bullish on a ten-to-fifteen-year horizon, but the media runs ahead of the engineering and winters follow hype. Math explores for its own sake, physics applies math to the world, engineering tries to hack the world. Extropic is the exodus: thermodynamics rather than quantum mechanics as the substrate for physics-based AI.

Another paper, on asymptotically limitless quantum energy teleportation via qudit probes, grew out of a master’s attempt to understand the energy and entanglement of empty space. Emptiness has energy, which is a strange sentence, and cosmology’s equations do not match calculations of vacuum fluctuations. You cannot extract that energy directly; ignorance of the fluctuations is the lock. But the vacuum is correlated, like a stock. Measure here, send the information there, infer the configuration, and statistically extract some energy — “teleporting” it, and even creating pockets of negative energy density. Correlations decay with distance, and you cannot take out more than you put in. He gave vacuum energy “the old college try” and hit the limits the laws allow.

Aliens, for him, are a useful thought experiment in anti-fragility: there is no law of physics against another civilization showing up and disrupting us, so become formidable fast, whether the shock is AI or something from another planet. Statistically, observed alien life would most likely be a self-replicating von Neumann-like probe; alleged ocean-floor mineral collectors would not violate his priors, but he has only secondhand data. When the idea for the computer he is now building arrived eight years earlier, it felt “beamed from space.” He does not literally believe that. He does think intelligence and life need a less anthropocentric definition.

The most beautiful idea he has chased is AdS/CFT: a lower-dimensional hologram dual to a bulk theory of quantum gravity, a duality that looks like learning deep representations of the boundary. On his bucket list is applying quantum machine learning to SYK models and reading an emergent geometry out of the boundary theory. Black holes sit at the quantum-gravity interface. From the outside, infalling matter sediments onto the near horizon; close enough for quantum fluctuations, it can interfere with vacuum pair creation so outgoing radiation shares mutual information with what fell in. Getting that calculation right, he says, is still early. Would he volunteer to orbit a black hole? Yes. Fall in? No — unless black holes punch pocket universes, in which case a Kardashev Type III-plus civilization might engineer them as a legacy against heat death.

Type I on the Kardashev scale is producing as much power as the sunlight incident on Earth; Type II is the Sun’s full output; Type III is galactic. He does not put much stock in Types IV and V. Life evolved here because the Sun supplied an energetic drive, and dissipative adaptation is not special to this planet. Extending life beyond Earth is, in his telling, a responsibility, because that is the process that brought us here, and it may have beauty in store we cannot yet imagine.

In a Substack post, Verdon and the still-anonymous co-founder bayeslord listed four interdependent civilizational goals: climb the Kardashev gradient, which in the short term almost certainly means nuclear fission; raise human flourishing through pro-population and pro-growth policy; build AGI as the greatest force multiplier in history; and develop interplanetary and interstellar transport. Decompressed, the goal is for the human techno-capital mimetic machine to become self-aware and hyperstitiously engineer its own growth. Humans work at companies, allocate capital, and talk through memes; all of those systems couple. e/acc is a thin cultural framework you can fork, not a lifestyle prescription. The core thesis is that the adaptive process that got us here is good and should continue.

Vitalik Buterin’s d/acc fork tries to sit between e/acc, effective altruism, and AI safety. Verdon likes forks. He wanted an opposite pole to the mainstream narrative taking over Silicon Valley, on the theory that systems are optimal at the edge of criticality between order and chaos. There is no formal organization, no control, just tweets and blog posts; people are free to defect. Cults, he notes, usually forbid competing forks. Humor and shitposts were camouflage: deep claims smuggled inside meta-irony, hard to argue with a cartoon frog or an intergalactic Jeff Bezos. They were engineering “high mimetic fitness packets” for algorithmically amplified feeds. He admits the cost in rigor, and says the movement is now large enough for debates and long-form conversation — which is partly why he is on this podcast.

Fridman had just interviewed Jeff Bezos. Verdon’s namesake, he says, built one of the most epic companies of all time by not taking profits too early, letting the system compound, and giving people fast, cheap convenience. AWS paid the salaries of a generation of technologists; Amazon’s robotics spend is enormous; Blue Origin’s O’Neill-cylinder ambition is industry off-planet on a multi-decade clock. Billionaires, in his framing, are a proof-of-stake vote: efficient capital allocation unlocks more capital, unlike politicians elected for television. Criticism of wealth, he thinks, is often a victim mindset. The shift he wants is positive-sum: create value, capture some of it, notice you have more agency than you thought. “Keep building” means founding a self-sustaining company that dances with the market, not a nonprofit that dies when the grant runs out, and not ESG-style degrowth. Elon, in this telling, innovates out of the climate dilemma with a profitable company rather than managing his way out.

Musk is the harder case, because he has long warned about existential risk from AI and favors regulation there while opposing it elsewhere. Verdon squares it by looking at actions: Musk saw OpenAI cornering cultural priors inside LLMs the way social platforms shaped information flow, so he built a competitor, and Neuralink is a hedge toward merging. Anxiety can be a useful deadline. What Verdon opposes is the executive-order move that treats open-source models as dual-use, registers GPUs, and adds friction that keeps hackers — including future safety researchers — out of the game. If nobody were working on AI safety he would be a proponent. In the current landscape he sees the topic used as a proxy for centralized control. Safety is a perfect cover: there is always another nine of pSafety to buy with more surveillance. Proposals for a global panopticon open the door to a 1984-like big brother, and history is the prior.

Extropic announced a $14.1 million seed round earlier than planned because the doxxing forced their hand. The company was born from dissatisfaction with the quantum roadmap: physics-based computing whose greatest enemy was noise you must constantly pump out. Generative AI is eating the world’s compute. The bet is to engineer the ultimate physical substrate for it from first principles of physics, information theory, computation, and thermodynamics — a thermodynamic computer, plus algorithms that treat machine learning as a physical process. The team includes quantum-computer architects who designed IBM and AWS systems and left that field, and Trevor McCourt, co-founder of TensorFlow Quantum, as CTO.

TensorFlow Quantum combined differentiable classical programming with quantum circuits so that, in the era of very short quantum programs, gradient descent could write the program. Andrej Karpathy calls differentiable programming software 2.0. The real lesson for Verdon was a universal meta-framework: if you have parameters that affect a system’s physics and a loss you can evaluate, you can optimize that system to do a task. Parameterize everything, make the parameters differentiable, optimize. Organizationally, the work taught patience across disciplinary boundaries. Put smart, hungry people in a room, teach each other, and you get a “model soup of expertise.” Hiring looks for high fluid intelligence and open-mindedness, not someone who has optimized one subsystem for four years, because there is no textbook for a method you are inventing.

He does not think Extropic will “create AGI” in a hallway. The picture is physics-based world models — quantum, thermodynamic, deterministic, hybrid — paired with anthropomorphic AI as a synthetic scientist you can talk to. Current language models are good bullshitters; you would not extrapolate the stock market from internet text. On the singularity, he again rejects a finite-time point. Light cones and physics bound growth. The “foom” story in which human-level AI suddenly groks nanobots underestimates how hard it is to engineer matter: you need many samples from an accurate simulation or from nature itself, and there is a thermodynamic cost to acquiring information. That cost, he says, keeps things in check.

Fridman posted a p(doom) poll that split the timeline into very-likely and very-unlikely camps. Verdon does not see e/acc as left or right; it is up versus down on civilizational scale, tech progressivism versus techno-conservatism. Parties that already want more centralization and regulation, he says, align with doomers because fear is a way to get people to surrender control. Polls inside e/acc look roughly balanced. The new cleavage is not the old culture war.

Effective altruism, in his sketch, tries to do good by minimizing suffering, measured in hedons. That loss has spurious minima: shrimp-farm pain, or wireheading on TikTok that feels good in the short run and decays in the long run. e/acc prefers an objective quantity that is hard to game: physical energy. GDP and currencies can be pinned and moved; wattage is wattage. Both movements want humanity to flourish. They disagree on the objective. Fridman pushes on the tyranny of equations; Verdon answers that this particular loss is statistical, an average of averages, not stiff rails, and that thermodynamics is not optional, like gravity. He still invites skepticism, because humans fill uncertainty with narratives.

What went wrong with EA, in his account, is the arc from a healthy community of engineers and rationalists to organizations that route capital and hold real power — including, he says, the OpenAI board and influence at Anthropic. The assumption of e/acc is closer to capitalism: every agent acts in its own interest, so keep adversarial competition at every scale. Doing good was a perfect cover for acquiring power, and power sometimes corrupts.

A favorite workday runs from noon to 4 a.m. Early afternoon is external: customers, investors, candidates. He front-loads social interaction on exogenous ketones and Red Bull, eats one animal-based meal — steak, eggs, vegetables, fruit — then takes a break and does deep technical work into the night, still contributing to most of the company’s patents. From 9 p.m. to 4 a.m. the fires stop trickling in. He started the split at Google, product by day and research by night, and notes that Demis Hassabis’s schedule is similar. He played American football, then powerlifted through math graduate school in a monk mode of proofs, iron, and caffeine, trying to engineer neuroplasticity while learning.

You can learn faster if you care, he says, even if you have to trick yourself into caring. Follow curiosity across field boundaries; acquire and compress as much as you can. Beff Jezos is the hype man, an intellectual Hulk: when he needs to invent something that has never existed, he jacks himself into an unreal level of excitement, the same mental switch he built under a 500-pound bar. In the early days he thought the character was compartmentalized. Over time the identities merged mentally, even before Forbes. He absolutely recommends alt accounts for young people — low-stakes essays, twenty followers, room to try thoughts that are not fully formed. We lost that, he thinks, in the era of everything under a real name.

Advice for a career you can be proud of: go to the base of the stack. He chose theoretical physics because those foundations survive technology fashion. Math, physics, complex and adaptive systems transfer; a framework three years from now may not. The closer you are to the base, the shorter the jump when you have to transfer-learn. The e/acc writings themselves were an experiment in applying out-of-equilibrium thermodynamics to the world around us.

Is mortality a feature or a bug? He does not want to be immortal. Dissipative adaptation has dissipation in the name. Physics progresses one funeral at a time; so do companies and empires. Lifespans should probably lengthen because the world is more complex and the window of high neuroplasticity otherwise caps how much we can understand. An immortal king, or a forever monopoly, calcifies in a time-varying landscape. “Death gives space for youth and novelty to take its place.” He is not especially afraid of his own death. He is not attached to this particular body and mind. There will be better forks — genetic or otherwise — and an evolutionary-like algorithm running on every bit of the world. He is not an optimum that needs to stick around forever.

The why of the e/acc machine, he says, is thermodynamics: it is why life, civilization, and technology are here. Why thermodynamics? Then you are in the anthropic principle and the landscape of possible universes. Could we engineer pocket universes, set the hyperparameters, and become their parents? Pure conjecture, and another reason to figure out quantum gravity. We seek, like a language model, to minimize the divergence between our predictions and the world. Blind energy scales feel like an insult. Better prediction evolved because it captures free energy for growth. Quantum gravity is the final boss of knowledge acquisition; between here and there, the mesoscales still have most of the work — perception, prediction, and control enough to climb the Kardashev gradient. That, he says, is the great challenge of our times.