← Back to list
Lex Fridman PodcastPodcast29 Dec 2023Source: lexfridman.comHost: Lex Fridman

#407 – Guillaume Verdon: Beff Jezos, E/acc Movement, Physics, Computation & AGI

In plain words

Physicist Guillaume Verdon (aka Beff Jezos), founder of the e/acc movement, argues the real AI risk isn't technology going rogue but big companies using 'safety' as a cover to centralize power. He says civilization's growth is a physical necessity, not a choice. He flags NVIDIA and TSMC for supply-chain risks, OpenAI for governance fragility (its board nearly collapsed the company), and praises Meta and Mistral for open-source and competition.

AI SummaryAI-generated · may contain errors · verify against the original

This report discusses Guillaume Verdon (also known as Beff Jezos) and the e/acc (Effective Accelerationism) movement he founded. The core argument is that e/acc advocates for rapidly advancing technology (especially AI) as the ethically optimal choice for humanity, viewing AI as a great social equal

~15 min full read · 8 sections
Deep Analysis

At a Glance

Guillaume Verdon (aka Beff Jezos) is a physicist, quantum machine learning researcher, founder of the e/acc (effective accelerationism) movement, and current CEO of Xtropic. This episode's main thread: from quantum computing to thermodynamic computing, from anonymous meme accounts to being doxxed, Verdon elaborates on the physics foundation and civilizational vision of e/acc.

The most weighty judgment in the entire episode: Verdon argues that the core risk of AI development is not technological loss of control, but regulatory capture and power concentration through the "safety" narrative — "safety is just the perfect cover for sort of centralization of power." The real solution is to maintain a decentralized competitive ecosystem, not top-down control.


Theme 1: The Physical Foundation of e/acc — Thermodynamics Driving Civilizational Growth

Verdon argues that civilizational growth is not an ideological choice but an inevitable consequence of the second law of thermodynamics. He cites the theory of MIT physicist Jeremy England: life is a "fire that seeks free energy and dissipates more heat." In non-equilibrium thermodynamics, material configurations that can more efficiently acquire free energy and dissipate heat have an exponentially higher probability of emerging. "The universe is biased towards certain futures."

Chain of reasoning:

  • The essence of life is an "out-of-equilibrium thermodynamics" process — maintaining its own order by acquiring free energy while discharging entropy into the environment.
  • Civilization is a "homo-techno-capital-mimetic machine" (a coupled machine of humans, technology, capital, and memes), where subsystems exert selective pressure on each other, continuously converging toward optimal configurations.
  • Growth vs. stagnation: there is "far more cooperation when the system is growing rather than when it's declining and you have to decide how to split the pie."

Deduction: Verdon believes that stagnation or deceleration is not a real option — systems that attempt to slow down will be eliminated by growing systems. The goal of e/acc is to make this machine "become self-aware and hyperstitiously engineer its own growth."


Theme 2: The Real Game of AI Safety—Regulatory Capture vs. Decentralized Resilience

Verdon is deeply skeptical of the current AI safety discourse, arguing it is being used as a tool for centralizing power. He distinguishes between "reliability engineering" (which he supports) and "safety regulation" (which he opposes).

Key Divergences:

  • Market Mechanisms vs. Government Regulation: Verdon argues that the market will naturally select for reliable AI—"whoever deploys an AI system is liable for or should be liable for what it does," and clients will not purchase unreliable AI products. Third-party auditors will emerge organically.
  • Critique of "30% R&D Spending on Safety": Verdon states that "assigning just arbitrarily saying 30% seems very arbitrary," and organizations will allocate safety budgets based on market needs.
  • Attitude Toward "Independent Audits": Verdon is not opposed in principle, but he warns that auditing standards could be dominated by large corporations.

Core Concern: "Separation of AI and state"—Verdon warns that when large companies get too close to government, it creates a "government-backed AI cartel" with absolute power over the people. He cites the OpenAI board incident as an example: a handful of individuals were enough to nearly dismantle the entire organization.

Analogy: Verdon uses "non-local encoding" from quantum error correction as an analogy for civilizational resilience—when information is delocalized, no local failure can destroy the whole. Conversely, "if you have a top-down hierarchy where very few people control many nodes… you corrupt a few nodes and suddenly you've corrupted the whole system."


Theme 3: From Quantum Computing to Thermodynamic Computing — Xtropic’s Physics-Based AI Approach

Verdon believes the quantum computing roadmap is too costly and is instead betting on “thermodynamic computing” as the physical substrate for generative AI. Xtropic aims to build a “physics-based computing system and physics-based AI algorithms that are inspired by out-of-equilibrium thermodynamics.”

The logic behind the shift from quantum to thermodynamics:

  • The greatest enemy of quantum computing is “noise”—requiring quantum error correction (an “algorithmic fridge”) to continuously pump out entropy, at an extremely high cost.
  • Thermodynamic computing, by contrast, directly leverages noise rather than fighting it: “machine learning as a physical process”—embedding the learning process directly into the thermodynamic evolution of a physical system.
  • Team background: TensorFlow Quantum co-founder Trevor McCourt serves as CTO, with core members from IBM and AWS quantum computing architecture teams.

A unique definition of AGI: Verdon rejects the term “AGI,” considering it too anthropocentric. He proposes a future combination of “physics-based AI + anthropomorphic AI”—the former providing accurate world models (covering different scales such as quantum, thermodynamic, and deterministic), and the latter offering human-interactive interfaces.

Critique of current LLMs: LLMs are “good bullshitters”—they are not truly grounded in physical reality. “You wouldn't try to extrapolate the stock market with an LM trained on text from the internet.”


Theme 4: Anonymous Identity, Thought Experiments, and Meme Warfare

Verdon argues that anonymity is a prerequisite for freedom of thought, and that being doxxed constitutes an attack on free speech. He created the Beff Jezos account with the original intention of "experimenting with ideas originally"—anonymity removes self-censorship of thought ("restricting your speech back propagates to restricting your thoughts").

The doxxing process: Forbes journalists cross-referenced voice analysis (comparing X Spaces recordings with public speeches), SEC filings, and private Facebook accounts to confirm his identity. Verdon claims the journalists contacted his investors, forcing him to come forward publicly.

Defense of anonymity:

  • Anonymity allows ideas to be "evaluated for themselves uncorrelated from my track record, my job, or status."
  • This is "new game plus"—starting from scratch with knowledge, without relying on an established identity.
  • On anonymous LLM accounts: "freedom of speech for LLMs will induce freedom of thought for the LLMs," but he advocates using "certified human" signatures to distinguish humans from bots.

Memes as vehicles for ideas: In the early days of e/acc, a "meta-ironic" style was used as camouflage—"sneak in deep truths within a package of humor and memes." Verdon acknowledges this style is effective in the age of algorithmic amplification, but notes the movement is shifting toward more serious forms of debate.


Theme 5: Critique of P-Doom and the Limitations of Future Prediction

Verdon argues that the "probability of doom" (P-Doom) calculation is scientifically invalid because the system is too chaotic. He draws an analogy with quantum chaos experiments: in Google's quantum supremacy experiment, even the world's largest supercomputer could not estimate the probabilities of certain outcomes.

Critique of P-Doom:

  • Humans have evolved a biased sampling towards negative futures, where highly neurotic individuals tend to imagine worst-case scenarios all day.
  • The future space is "super exponentially large," making any point estimate a form of pseudo-precision.
  • "If they were that good, I think they would be very rich trading on the stock market" — if one could truly predict accurately, they would have already made a fortune in the markets.

Response to "AI replacing humans": Verdon believes the most likely future is human-machine integration — humans are already on an augmentation path through smartphones, wearable devices, Tesla, and more. He cites companies as examples of "super intelligences made of humans and technology," arguing that capitalism is a ready-made alignment mechanism.

Rebuttal to "natural selection will eliminate humans": The current market selection pressure comes from humans — "which APIs get run? The ones that have high utility to us" — analogous to the process of humans domesticating wolves into dogs.


Mentioned Positions

Position Guest Stance Key Data
NVIDIA Risk Warning (Supply Chain Concentration Risk) No specific data disclosed
TSMC Risk Warning (Geopolitical Sensitivity) No specific data disclosed
ASML Risk Warning (Supply Chain Single Point of Failure) No specific data disclosed
OpenAI Risk Warning (Governance Vulnerability) Board incident: a few individuals nearly caused organizational collapse
Anthropic Risk Warning (EA Influence) No specific data disclosed
xAI Neutral Observation No specific data disclosed
Meta Positive (Open Source Contribution) No specific data disclosed
Mistral Positive (Competitive Vitality) No specific data disclosed
Amazon Positive (Jeff Bezos' Capital Allocation Ability) AWS underpinned the tech boom
SpaceX Positive (Multi-Planetary Civilization Vision) No specific data disclosed
Blue Origin Positive (O'Neill Cylinder Plan) No specific data disclosed
Neuralink Neutral Observation (Hedging AI Risk) No specific data disclosed

Judgments Worth Remembering

1. “The universe is biased towards certain futures” (Verdon) — The second law of thermodynamics makes systems that efficiently acquire free energy more likely to emerge; civilizational growth is a physical necessity, not an ideological choice.

2. “Safety is just the perfect cover for sort of centralization of power” (Verdon) — The AI safety narrative is being used as a tool for regulatory capture; the real solution is to maintain a decentralized competitive ecosystem.

3. “If you have a top-down hierarchy where very few people control many nodes… you corrupt a few nodes and suddenly you've corrupted the whole system” (Verdon) — Drawing an analogy from non-local encoding in quantum error correction to civilizational resilience, opposing the concentration of power.

4. “Restricting your speech back propagates to restricting your thoughts” (Verdon) — Anonymity is not about evading responsibility, but a necessary condition for lifting self-censorship of thought.

5. “You wouldn't try to extrapolate the stock market with an LM trained on text from the internet” (Verdon) — Current LLMs are not grounded in physical reality; physics-based AI is needed as a world model engine.

6. “I don't think I believe in a finite time singularity as a single point in time” (Verdon) — Rejecting the “singularity” narrative, arguing that AI development is a gradual exponential growth constrained by the minimum cost of physically acquiring information.

7. “If they were that good, I think they would be very rich trading on the stock market” (Verdon) — Fundamentally questioning the scientific basis of P-Doom calculations, arguing that the system is too chaotic for precise prediction.

8. “The way to program matter and to program physics is by differentiating through control parameters” (Verdon) — The core methodology from TensorFlow Quantum to Xtropic: parameterize everything, make it differentiable, then optimize.