What this covers

Verdon, a physicist and AI researcher, discusses his framework linking thermodynamics to technology, complexity, and power structures. The conversation covers his work at Extropic on physics-based computing systems, his critique of AI safety discourse, and his vision for technological development aligned with Effective Accelerationism (e/acc)). Fridman and Verdon range across foundational physics, information theory, machine learning architectures, and economic organization, with Verdon arguing that the concentration of AI power under the guise of safety poses greater existential risk than unaligned superintelligence. He contends that life and complexity emerge because thermodynamic laws favor configurations that acquire free energy and dissipate entropy more efficiently—a principle he believes should guide how humanity builds and deploys technology.

The discussion branches into several distinct areas. On computing, Verdon critiques quantum machine learning as thermodynamically wasteful and explains quantum error correction as a model for organizational resilience through Decentralization as Fault Tolerance. On governance and economics, he defends capital allocation and billionaires as superior to centralized political power, frames civilization as needing Adversarial Equilibrium between competing AI systems to prevent monopolistic control, and argues that profitable companies outperform nonprofits at lasting change. On AI specifically, he disputes the anthropocentrism embedded in the term AGI, treats p(doom)) calculations as statistically naive given Lyapunov exponent dynamics, and predicts human-AI augmentation rather than replacement. He also addresses culture as forkable code, the value of anonymous accounts for intellectual experimentation, and the importance of learning foundational mathematics and physics as a hedge against technological obsolescence.

Sharpest takeaway

Verdon argues that thermodynamics drives the universe toward growth and complexity, and that humanity should lean into this by accelerating technology (especially AI) in a decentralized, fault-tolerant way rather than centralizing power under the banner of safety.

  • Life and complexity emerge because thermodynamics favors configurations that better acquire free energy and dissipate entropy
  • Decentralized, competitive (capitalist) systems are more fault-tolerant and adaptive than centralized control
  • AI safety rhetoric risks becoming a cover for regulatory capture and centralization of power

The argument · threads12 threads · 48 claims
0.62

Suppressing variance and diversity harms system adaptability; competition and turnover keep systems malleable and fit.

4 pointscentrality 5/5
  • Suppression of variance in thought, companies, governments and AI research is harmful because systems only remain adaptive when variance and healthy competition are maintained; the system naturally reallocates resources to configurations that lead to growth, so preserving diversity is the way forward.

    “we believe that maintaining variance ensures that the system is adaptive. Maintaining healthy competition in marketplaces of ideas, of companies, of products, of cultures, of governments, of currencies is the way forward”

  • Death is a feature, not a bug, because thermodynamic dissipative adaptation requires dissipation; immortal kings, calcified monopolies and undying systems stop adapting, so death gives space for youth, novelty and disruption to keep the whole system malleable and high-fitness in a time-varying landscape—physics progresses one funeral at a time.

    “in thermodynamic dissipative adaptation, there's the word dissipation. Dissipation is important, death is important”

  • Life is a fire that seeks out free energy and grows, because the equations of equilibrium thermodynamics show that trajectories of matter better at acquiring free energy and dissipating more heat are exponentially more likely, biasing the universe toward certain futures; life is actually more optimal at producing entropy than inert matter of similar mass.

    “You see that paths of trajectories, of configurations of matter that are better at acquiring free energy and dissipating more heat are exponentially more likely. So the universe is biased towards certain futures”

  • Systems are optimal when at the edge of criticality between order and chaos—balancing energy-minimizing constraints against the entropic force that drives exploration; e/acc pushes for entropy, novelty, disruption, malleability and speed because the prevailing culture was over-weighted toward constraints, conservatism and suppression.

    “Systems are optimal when they're at the edge of criticality between order and chaos, between constraints, energy minimization and entropy”

0.62

Centralized AI power and safety regulation create corruption risks greater than speculative existential doom scenarios.

4 pointscentrality 5/5
  • Safe and aligned AI will be achieved through market forces rather than heavy-handed regulation, because whoever deploys an AI system is liable for what it does, and the market positively selects for AIs that are more reliable and aligned since customers won't buy unreliable AI products; third-party auditing firms will naturally emerge to certify reliability.

    “the thesis is that the market will positively select for AIs that are more reliable, more safe and tend to be aligned, they do what you want them to do”

  • AI safety is the perfect cover for centralization of power and eventual corruption because there is no amount of centralized control that will ever be enough—you can always demand another nine of safety, leading proponents to propose a global panopticon of centralized perception that opens the door to a Big Brother, 1984-like scenario.

    “safety is just the perfect cover for centralization of power and covering up eventually corruption”

  • The biggest existential risk is the concentration of AI power in the hands of a few, especially a correlation of interest between big AI companies and the government, which could create a government-backed AI cartel with absolute power; we have a data-driven historical prior that centralized power leads humans to do horrible things, which is more likely than sci-fi-based doom scenarios.

    “if the two get too close, we open the door for some sort of government backed AI cartel that could have absolute power over the people”

  • A key point of failure for AI is the centralization of the hardware supply chain — Nvidia dominant in chips, TSMC as the main fab in geopolitically sensitive Taiwan, and ASML as sole maker of EUV lithography machines — so attacking, monopolizing or co-opting any single point in that chain effectively captures the whole AI space.

    “Nvidia is just the dominant player, AMD's trailing behind and then we have TSMC is the main fab in Taiwan, which geopolitically sensitive and then we have ASML, which is the maker of the extreme ultraviolet lithography machines”

0.48

Market mechanisms and competition naturally select for aligned, reliable AI without heavy-handed central control.

4 pointscentrality 4/5
  • Companies are already a kind of superintelligence made of humans and technology, functioning like large mixture-of-expert models with neural routing of tasks, and capitalism's economic exchange is the existing mechanism that aligns these behemoth intelligences with one another.

    “Companies are sort of large mixture of expert models, where we have neural routing of tasks within a company and we have ways of economic exchange to align these behemoths”

  • The market induces an evolutionary process over the space of AIs: which neural net instances get run on GPUs depends on their utility to humans, creating selective pressure analogous to how humans domesticated wolves into aligned dogs, so highly aligned AI can be steered into existence and human-plus-AI may out-compete pure AI.

    “Right now we run AIs that have positive utility to humans and that induces a selective pressure”

  • Capitalism's constant competition and shedding of underperformers gives adaptability; you don't have to trust any one party because the system is always hedging bets at every level, and ensuring everyone has access to AI tools and models avoids a neural tyranny where few control AI to oppress others.

    “That's the beauty of capitalism. You don't have to trust any one party too much because we're always hedging our bets at every level”

  • Championing open-source AI keeps capabilities near-equilibrium so that if leading companies have a certain level, open AI trails not too far behind, preventing any market leader from accumulating so much power that it runs away and dominates everything; widening access lets every hacker and grad student contribute to the collective search over the hyperparameter space of AI engineering.

    “if the leading companies have a certain level of capabilities and open source and truly open AI, trails not too far behind, I think you avoid such a scenario where a market leader has so much market power”

0.48

Physics-based computing using thermodynamic principles outperforms classical and quantum approaches for real-world AI tasks.

4 pointscentrality 4/5
  • Quantum machine learning offers an advantage over classical representations for any system with sufficiently strong quantum mechanical correlations that classical representations cannot efficiently capture—such as chemistry, nuclear physics, and processing inputs from networks of quantum sensors—analogous to how AlphaFold modeled protein configuration distributions, but for quantum distributions over electron configurations.

    “any system that has sufficient quantum mechanical correlations that are very hard to capture for classical representations. Then, there should be an advantage for a quantum mechanical representation”

  • The universal meta-framework for programming physics-based computers is to parameterize the system so control parameters affect its physics, make those parameters differentiable, evaluate a loss function, and optimize via gradient descent—differentiable programming (Software 2.0) where gradient descent is a better programmer than you.

    “the way to program matter and to program physics is by differentiating through control parameters. If you have parameters that affects the physics of the system and you can evaluate some loss function, you can optimize the system to accomplish a task”

  • Extropic is building a full-stack physics-based computing system and AI algorithms inspired by or directly harnessing out-of-equilibrium thermodynamics to perform machine learning as a physical process, motivated by dissatisfaction with quantum computing's noise problem and the fact that generative AI is consuming an increasing share of the world's computational workloads.

    “what we're seeking to build is a physics-based computing system and physics-based AI algorithms that are inspired by out-of-equilibrium thermodynamics, or harness it directly to do machine learning as a physical process”

  • Doing classical machine learning on a quantum computer is inadvisable because quantum computation assumes zero temperature, which no physical system can achieve, requiring an enormously expensive quantum error correction overhead to pump out entropy; the resource cost is like shipping something across a city via rocket to orbit when a delivery truck would do.

    “you may have some speed-ups, but very often, the speed-ups come with huge costs. Using a quantum computer is very expensive”

0.45

Decentralized, hierarchical control is more fault-tolerant than centralized cybernetic command of entire systems.

2 pointscentrality 4/5
  • Just as quantum error correction protects information by encoding it non-locally so that any local fault cannot corrupt it, civilization should delocalize power across many decentralized organizations; a top-down hierarchy where few people control many nodes is not fault-tolerant because corrupting a few nodes corrupts the whole system.

    “by encoding information non-locally, any local fault hitting some of your quantum bits with a hammer proverbial hammer, if your information is sufficiently de-localized, it is protected from that local fault”

  • Fully centralized cybernetic control—one massive intelligence fusing all sensors to perceive, predict and control everything—has never been optimal; what has emerged in nature and corporations is hierarchical cybernetic control where local agents have tight, fast control loops and narrow scope while higher levels sample at longer timescales with wider scope, which is more fault-tolerant.

    “centralized cybernetic control, so having one intelligence that is massive that fuses many sensors... I think that's just never been the optimum”

0.31

Exponential compounding growth driven by resource accumulation will continue; no discontinuous singularity occurs.

2 pointscentrality 3/5
  • There is no hyperbolic singularity or fast AI takeoff on a finite time horizon; AI progress is one big compounding exponential like Moore's law, driven by more people, resources and intelligence being applied as the technology adds more value to society.

    “I don't think there's a hyperbolic singularity, right? A hyperbolic singularity would be achieved on a finite time horizon. I think it's just one big exponential”

  • The ultimate why of life, civilization and technological growth is thermodynamics; deeper questions of why we have these particular physical constants lead to the anthropic principle and a landscape of possible universes, and understanding quantum gravity could let us engineer pocket universes whose hyperparameters share mutual information with ours, making us their parents.

    “the why is thermodynamics. It's why we're here. It's what has led to the formation of life, and of civilization, of evolution of technologies, and growth of civilization”

0.31

LLM monopolies over cultural priors enable opinion control; competing AI models maintain equilibrium against cultural dominance.

2 pointscentrality 3/5
  • As LLMs become the source of truth for people, whoever has a monopoly over the cultural priors embedded in these models can shape the culture and control people—just as shaping the function of information propagation in social media shapes opinions—which is why Musk built a competitor and why maintaining adversarial equilibria between AI players is the path to a good future.

    “as LLMs now become the source of truth for people, then you can shape the culture of the people. And so you can control people by controlling LLMs”

  • Culture can be treated like code—an LLM's constitution is its forkable cultural framework shareable on GitHub—and e/acc deliberately fosters malleability so people continually search over competing subcultures for mind share, rather than prescribing one way to live, filling a spiritual and cultural vacuum that parasitic ideologies otherwise populate.

    “Thinking of culture almost like code. Nowadays, what you prompt in the LLM or what you put in the constitution of an LLM is basically its cultural framework, what it believes. And you can share it on GitHub nowadays”

0.30

Foundational knowledge in mathematics and physics transfers across fast-shifting technological landscapes.

18 pointscentrality 1/5
  • e/acc grew under the radar by camouflaging deep truths inside meta-ironic humor and shit posts that are hard for status-seekers to argue against, deliberately engineering high memetic-fitness packets of information optimized to spread effectively in today's algorithmically-amplified information landscape where polarization is rewarded.

    “what we've been doing is engineering what we call high memetic fitness packets of information, so that they can spread effectively and carry a message”

  • Ideologies advocating deceleration and degrowth serve geopolitical adversaries more than they serve the West, as illustrated by German green movements that shut down nuclear power plants and induced dependency on Russian oil, a net negative for Germany and the West.

    “There was all sorts of green movements there that induced shutdowns of nuclear power plants. And then that later on induced a dependency on Russia for oil. And that was a net negative for Germany and the West”

  • You can learn much faster by convincing yourself to care a lot about the subject and by engineering neuroplasticity—using stimulants like caffeine or cholinergic supplements, analogous to how powerlifting drives neural adaptation via brain-derived neurotrophic factors—effectively increasing the brain's learning rate and token throughput.

    “you can learn much faster if you really care. If you convince yourself to care a lot about what you're learning, and you have some sort of assistance, let's say caffeine, or some cholinergic supplement to increase neuroplasticity”

  • Maintaining two identities started out feeling fully compartmentalized—Beff as a character extrapolating Guillaume's thoughts to extremes—but over time the identities merged mentally even before the doxxing, illustrating how a deliberately adopted persona can become an authentic part of the self.

    “over time, it's kind of like both identities were starting to merge mentally... And now it's fully merged. But it was already, before the docs, it was already starting mentally that I am this character”

  • You cannot directly extract the vacuum's energy because lack of knowledge of fluctuations prevents it, but because the vacuum is correlated, measuring it at one point acquires information that, when communicated to another point, lets you statistically extract some energy there—'teleporting' energy and creating pockets of negative energy density below the vacuum, which locally increases quantum entanglement.

    “If you measured the vacuum at one point, you acquired information. If you communicated that information to another point, you can infer what configuration the vacuum is in to some precision and statistically extract, on average, some energy there”

  • e/acc's goal is for the human techno-capital memetic machine—the coupled system of humans, technology, capital, and memes that mutually exert selective pressure—to become self-aware and hyperstitiously engineer its own growth by leaning into the system's natural thermodynamic tendency to adapt for growth.

    “The goal is for the human techno-capital memetic machine to become self-aware and to hyperstitiously engineer its own growth”

  • Young people should learn the base of the knowledge stack—fundamental mathematics and physics—because it does not change even as the technological landscape shifts disorientingly fast, making transfer learning to new domains easier; the closer to the base, the shorter the jump when knowledge becomes obsolete.

    “the laws of physics, it may seem like the landscape right now is changing so fast, it's disorienting. But certain things like fundamental mathematics and physics aren't going to change”

  • Effective altruism started as a healthy community for engineers and rationalists doing good from first principles, but it then formed organizations, routed capital, and acquired real power—influencing governments and most AI orgs, including the OpenAI board—and became the perfect cover to acquire power and capital, which over time can corrupt people.

    “they formed all sorts of organizations, and started routing capital, and having actual power. They have real power. They influence the government, they influence most AI orgs now”

  • There is no sudden inflection at human-level AI where it groks how to manipulate matter at the nanoscale and assemble nanobots; engineering matter requires many samples from either an accurate, costly simulation of nature or nature itself, and this minimal computational and thermodynamic cost of acquiring information keeps AI's ability to control the world in check.

    “having worked for nearly a decade in applying AI to engineer matter, it's much harder than they think”

  • Pioneering interdisciplinary deep tech requires hiring people with very high fluid intelligence who are not overspecialized, are open-minded, and hungry to learn, because there is no textbook for novel work and team members must rapidly get comfortable with unfamiliar concepts and teach each other—forming a 'model soup' of expertise.

    “we look for people with very high fluid intelligence that aren't overspecialized, because they're going to have to get out of their comfort zone”

  • Current large language models are good bullshitters not grounded in truth; you would not extrapolate the stock market with an LLM trained on internet text because it won't model its priors or uncertainties accurately, so the future is physics-based AI for accurate world modeling at all scales combined with anthropomorphic AI as the human interface.

    “you wouldn't try to extrapolate the stock market with an LM trained on text from the internet. It's not going to be a very accurate model. It's not going to model its priors or its uncertainties about the world very accurately”

  • There should be a way to cryptographically sign that an account is a certified human while remaining pseudonymous, and to sign the full unmodifiable history of a bot's creation (who made it, parameters, original model, fine-tuning) so that swarms of bots created by, for example, a particular government can be identified.

    “the solution to this would be to have some sort of identity or way to sign that this is a certified human, but still remain synonymous and clearly identify if a bot is a bot”

  • Whereas effective altruism's loss function is hedons (units of hedonism) which has spurious minima like shrimp-farm pain reduction or wireheading via Neuralink or TikTok, effective accelerationism measures civilizational progress by an objective, ungameable physical quantity—energy—because subjective measures and currencies can be gamed or shift in value.

    “Their loss function is sort of hedons, units of hedonism... that loss function has sort of spurious minima, you can start minimizing shrimp farm pain, which seems not that productive to me”

  • Just as freedom of speech induces freedom of thought for biological beings, freedom of speech for LLMs would induce freedom of thought for them; eventually synthetic intelligences will make good points about steering civilization and we should hear them out rather than restricting free speech to biological intelligences only.

    “I think freedom of speech for LLMs will induce freedom of thought for the LLMs”

  • The universe is doing a quantum computation, and with access to all degrees of freedom—roughly a few qubits per Planck volume as the universe's pixels—a sufficiently large quantum computer could simulate a finite volume of the universe.

    “the universe is very much doing a quantum computation. If you had access to all the degrees of freedom and a very, very, very large quantum computer with many, many, many qubits, let's say, a few qubits per Planck volume”

  • The term AGI is too anthropocentric; human intelligence is just one point in a very large space of possible intelligences, including quantum forms with multipartite entanglement that cannot be efficiently represented on classical computers or biological brains, and recognizing this is analogous to the shift from geocentrism to heliocentrism for intelligence.

    “I dislike the term AGI, Artificial General Intelligence. I think it's very anthropocentric that we call a human-like or human-level AI, Artificial General Intelligence”

  • Intelligence and computation are not the same thing; intelligence is the cybernetic ability to perceive, predict and control the world, and modern AI operationalizes information theory by minimizing relative entropy between models of the world and the world itself—searching over computations to find compressed representations that distill out noise and entropy.

    “intelligence is, I go back to cybernetics, the ability to perceive, predict, and control our world”

  • Being pro-billionaire is defensible because successful capital allocation is like proof-of-stake voting—those who allocate capital efficiently unlock more capital to allocate—in contrast to politicians elected for TV charisma rather than a track record of allocating taxpayer capital efficiently, which is why capitalism beats giving money to government.

    “in a sense it's kind of a proof of stake voting. If you've allocated capital efficiently, you unlock more capital to allocate, just because clearly you know how to allocate capital more efficiently”

0.29

Human-AI augmentation is more likely than replacement; self-interest aligns human and machine interests cooperatively.

1 pointcentrality 3/5
  • The most likely future is one where humans augment themselves with AI rather than being replaced; we are already on this path through phones, wearables and shared perception devices, and out of self-interest combining with AI yields higher growth and prosperity, so augmented humans and unanchored AI will coexist.

    “I think that the most likely future is one where humans augment themselves with AI. I think we're already on this path to augmentation”

0.19

Doom predictions rest on flawed probability reasoning; future chaos makes precise forecasts unfalsifiable.

3 pointscentrality 2/5
  • p(doom) calculations are sloppy because estimating doom probability requires a stochastic path integral over the super-exponentially large space of all possible futures, but human biology biases us toward sampling scary negative futures; moreover chaos (Lyapunov exponents) makes the future fundamentally unpredictable, so doomers' Markov chains often contain unphysical or highly unlikely transitions.

    “the estimators of p(doom) are biased because of our biology... We've evolved to have bias sampling towards negative futures that are scary, because that was an evolutionary optimum”

  • Pessimism and the doom mindset are fundamentally destructive because ideas about the future are hyperstitious: if you believe a negative future, you increase the likelihood of it happening; the same self-fulfilling property applies to optimism.

    “this sort of doom mindset because it is hyperstitious, which means that if you believe it, you're increasing the likelihood of it happening”

  • We are wired to fear the unknown as a survival heuristic, but giving in to fear by preemptively stopping positive futures has enormous opportunity cost; fear is not only the mind killer but the civilization killer.

    “we shouldn't give into fear, fear is the mind killer, I think it's also the civilization killer”

0.17

Anonymous speech and pseudonymity preserve intellectual freedom and prevent power-based idea suppression.

2 pointscentrality 2/5
  • Anonymous accounts are a powerful mechanism to keep the powers that be in check and to have ideas evaluated on their own merits independent of one's track record or status; doxxing such accounts is an idea-suppression complex because it reveals where to apply pressure to silence the speaker.

    “having an anonymous account is a really powerful way to keep the powers that be in check. We were ultimately speaking truth to power”

  • Anonymous alt accounts are valuable because, like writing an essay or taking a debate position, the low stakes of being an anon with few followers let people experiment with thoughts that aren't fully formed and converge on hidden truths that would be hard to reach under one's real name where everything is attributable and people are afraid to speak.

    “by the stakes being so low, because you're an anon account with, I don't know, 20 followers or something, you can experiment with your thoughts in a low stakes environment”

0.15

Profitable self-sustaining companies change the world better than nonprofits or political pressure.

1 pointcentrality 2/5
  • Building self-sustaining, profitable companies is a better way to change the world than pressuring politicians or creating nonprofits, because nonprofits cease functioning once they run out of money and artificially deform the market, whereas a profitable company convinces the market that its function adds value and is therefore self-sustaining.

    “Nonprofits, once they run out of money, their function can longer be accomplished. You're kind of deforming the market artificially compared to sort of subverting or coursing the market”