Guillaume Verdon

Born in Montreal, Verdon developed an early fascination with cosmic engineering, space exploration, and the Kardashev scale. He pursued an honors double major in mathematics and physics at McGill University, graduating in 2014, before moving into graduate work at the University of Waterloo’s Institute for Quantum Computing and the Perimeter Institute under Achim Kempf.

In his 2017 master’s thesis, Probing Quantum Fields: Measurements and Quantum Energy Teleportation, Verdon focused on Quantum Energy Teleportation (QET)—a protocol in quantum field theory where an observer (Alice) makes a local measurement on a quantum vacuum state, transmits the classical result to a distant location, and allows another observer (Bob) to extract energy from the local vacuum at his site.

Energy is not transmitted through space. Instead, Alice’s measurement unlocks energy that was already present in the vacuum’s entangled zero-point fluctuations. Verdon extended QET using arbitrary-dimensional qudit probes and polynomially localized Hamiltonians, showing how spatial correlations in vacuum noise can be harvested.

The durable epistemic lesson of his early physics was simple: an apparent void contains structured, exploitable correlations. What conventional engineering suppresses as noise can be reclassified as a computational resource.

Transitioning into a PhD program in applied mathematics and quantum machine learning, Verdon began asking how quantum mechanics could perform optimization natively. In 2017 he co-founded Everettian Technologies, serving as Chief Scientific Officer.

At Everettian, Verdon co-authored research on Backwards Quantum Propagation of Phase errors (Baqprop) and Quantum Dynamical Descent. Rather than simulating gradients on classical hardware, Baqprop encoded error information into relative quantum phases, using phase kickback and quantum coherence to guide the system through complex energy landscapes.

The architectural doctrine of this era was uncompromising: do not spend energy simulating quantum physics with classical mathematics when quantum physics can execute the mathematics directly.

Theoretical mechanisms, however, remain distinct from demonstrated commercial advantage. While phase kickback offered elegant algorithmic structures, real-world NISQ-era quantum hardware remained severely limited by gate errors and decoherence.

Moving from academia into industry, Verdon joined Google Quantum AI and Alphabet’s X environment. He confronted the stark realities of Noisy Intermediate-Scale Quantum (NISQ) devices: gate errors, thermal decoherence, and the absence of fault-tolerant error correction.

Co-leading TensorFlow Quantum alongside Michael Broughton, Trevor McCourt, and Masoud Mohseni, Verdon led the development of a open-source framework released in 2020. TFQ combined Cirq quantum circuits with TensorFlow computational graphs.

Parameterized quantum circuits (PQCs) were embedded inside differentiable classical pipelines. Expectation values calculated from quantum measurements were passed directly into classical neural networks, allowing automatic differentiation and end-to-end optimization across hybrid classical-quantum models.

The achievement was structural: TFQ converted complex quantum mechanical operations into composable tensor abstractions, exposing physical operations to machine learning developers without pretending the underlying substrate was classical.

At Google and X, Verdon connected two distinct research programs to overcome the limits of generic quantum circuits:

A. Quantum Graph Neural Networks (QGNN).Untrained generic quantum circuits suffer from “barren plateaus”—vanishing gradient landscapes where optimization stalls. Verdon introduced Quantum Graph Neural Networks and continuous-variable spectral variants (QGCNN), assigning Hilbert spaces to graph vertices and coupling Hamiltonians to edges. By embedding problem geometry into the quantum ansatz, QGNNs translated classical graph symmetries into legal unitary operations.

B. Variational Quantum Thermalizer (VQT). In Quantum Hamiltonian-Based Models and the Variational Quantum Thermalizer Algorithm (2019), Verdon moved beyond zero-temperature ground-state optimization. Recognizing that real physical systems operate at finite temperature, VQT minimized Helmholtz free energy (F = E - TS) rather than energy alone.

This was the direct bridge between quantum mechanics, machine learning, and statistical thermodynamics. Optimization was no longer just finding a single lowest-energy state; it was thermalization into a target probability distribution.

Despite progress in quantum software, Verdon grew dissatisfied with the commercial time horizon of fault-tolerant quantum computing. He drew an analogy: quantum computing resembled nuclear fusion—physically real, but decades away from practical economics; he wanted the computational equivalent of fission.

Quantum hardware requires extreme isolation near absolute zero (0 Kelvin) to prevent environmental decoherence. Yet biology, fluids, chemical reactions, and living systems operate at room temperature in the mesoscale: hot, noisy, and driven.

Influenced by Jeremy England’s framework of dissipative adaptation, Verdon recognized that driven open systems naturally self-organize to absorb and dissipate energy. If an AI algorithm requires stochastic samples (for generative modeling, MCMC, or diffusion), generating pseudorandomness through deterministic GPU arithmetic wastes immense energy. A physical thermal system at room temperature relaxes directly into the target distribution.

In 2022, Verdon initiated the effective accelerationism (e/acc) movement under the Twitter persona “Beff Jezos.” The movement opposed centralized regulatory deceleration and advocated for market competition, variance, and technological expansion as thermodynamic imperatives.

Verdon mapped dissipative adaptation onto economics: human civilization and the “techno-capital mimetic machine” were framed as physical dissipative structures that grow to maximize entropy production and adaptive information processing.

In late 2023, Forbes publicly linked Verdon to the Beff Jezos pseudonym. Far from damaging his trajectory, the public revelation concentrated massive attention, capital, and technical recruiting power around his hardware startup, Extropic.

Critical distinction: A physical description of driven matter does not automatically produce a valid moral or political prescription. Treating the expansion of techno-capital as an inevitable thermodynamic law is Verdon’s analogy, not an established scientific result.

Verdon left Alphabet in 2022 and co-founded Extropic alongside Trevor McCourt and Christopher Chamberland, raising a $14.1 million seed round led by Kindred Ventures.

Extropic’s core primitive is the p-bit (probabilistic bit): an asynchronously fluctuating classical state driven by room-temperature thermal noise. In Extropic’s Z1 Thermodynamic Sampling Unit (TSU), fabricated in standard mixed-signal CMOS, thermal fluctuations are not shielded against; they are the clock and the random number generator.

Model parameters are configured as physical energy potentials coupling stochastic units. Instead of stepping through millions of clock cycles of matrix multiplications, the hardware network physically relaxes into its equilibrium Boltzmann distribution.

Extropic’s published claims cite potential energy efficiency gains of up to 10,000× for specific generative sampling tasks compared to GPUs. Crucially, this is a company-modeled benchmark claim for sampling, not an established result across general computing.

A physical hardware advantage is lost if software must decompose stochastic physical processes back into GPU-style deterministic instructions.

To solve this, Extropic developed Torx, a JAX-oriented software framework for stochastic differentiable programming. Torx defines directed factor graphs, Markov kernels, and continuous-variable SDE primitives (Fokker–Planck, Ornstein–Uhlenbeck, and jump-diffusion processes).

Accompanying compilers (“thermalizers”) map logical probability programs into undirected physical energy functions required by the TSU hardware.

Analog hardware suffers from device mismatch and physical fabrication disorder. Extropic addresses this through hardware-in-the-loop calibration, REINFORCE trajectory optimization, and context matching—learning chip imperfections rather than suppressing them. The chaos of the substrate is analytically domesticated, not eliminated.

To assess Verdon’s thesis rigorously, one must identify where thermodynamic hardware fails.

1. Transformer Linear Algebra. Modern LLM transformer forward passes require exact, high-precision matrix multiplications over continuous activation vectors. Although token selection at the output uses stochastic sampling, the bulk of transformer computation is strictly deterministic linear algebra. There is no direct, efficient mapping from high-precision transformer math onto a physical lattice of noisy p-bits without massive precision loss and error-correction overhead.

2. Complexity Theory Limits. Thermodynamic sampling does not evade complexity theory. Highly frustrated energy landscapes exhibit mode trapping, autocorrelation, and critical slowing down. Generating millions of raw physical samples per second is meaningless if those samples remain trapped in a single local energy minimum.

The true metric of hardware performance is energy per effective decorrelated sample, taking into account analog I/O overhead, DAC/ADC conversion, and compilation costs. Verdon’s work is coherent not as a universal replacement for digital GPUs, but as a specialized accelerator for workloads whose mathematical structure matches physical thermal relaxation.

Education. McGill University (B.S., honors double major in mathematics and physics, 2014). University of Waterloo / Institute for Quantum Computing & Perimeter Institute (M.S., 2017; thesis: Probing Quantum Fields: Measurements and Quantum Energy Teleportation; advisor: Achim Kempf). Applied mathematics PhD candidate in quantum machine learning.

Affiliations. Everettian Technologies (co-founder & Chief Scientific Officer, 2017–2019). Google Quantum AI / Alphabet X (Quantum AI Tech Lead, 2019–2022). Extropic AI (co-founder & CEO, 2022–present).

Technical Through-Line.Quantum energy teleportation → quantum error phase-kickback (Baqprop) → TensorFlow Quantum differentiable graphs → Quantum Graph Neural Networks & Variational Quantum Thermalizers → room-temperature p-bit thermodynamic hardware (Z1 TSU) → stochastic differentiable programming (Torx).

Collaborators Worth Naming. Achim Kempf (master’s advisor), Trevor McCourt (Extropic co-founder & CTO), Christopher Chamberland (Extropic co-founder), Michael Broughton (TFQ co-lead), Masoud Mohseni (Google Quantum AI).

Intellectual Influences. Jeremy England (dissipative adaptation), Hugh Everett III (Everettian relative states), Ludwig Boltzmann & Josiah Willard Gibbs (statistical mechanics), Richard Feynman (substrate-native simulation).

Tension. Physical isomorphism versus universal digital computing ambitions; raw sampling speed versus decorrelated effective sample size; thermodynamic scientific description versus civilizational political prescription.