---
title: 'Why von Neumann Was Right'
subtitle: 'Shannon Entropy, Boltzmann Entropy, and the Architecture of Artificial Energy'
slug: 'why-von-neumann-was-right'
date: 2026-06-07
type: 'essay'
status: 'published'
tags: ['information-theory', 'thermodynamics', 'physics', 'artificial-energy', 'landauer', 'causal-sovereignty']
abstract: 'John von Neumann''s 1948 instruction to Shannon—call your information measure entropy—was a statement of mathematical identity, not convenience: Boltzmann''s thermodynamic entropy and Shannon''s information entropy are the same function. The essay traces that identity through Jaynes, Maxwell''s Demon, and the Landauer limit to its engineering expression in Artificial Energy, where information-rich matter harvests free-energy gradients the way the biosphere has for 3.5 billion years.'
license: 'CC-BY-4.0'
author: 'Jed Anderson'
co_authors: []
canonical_url: 'https://jedanderson.org/essays/why-von-neumann-was-right'
pdf: '/pdfs/why-von-neumann-was-right.pdf'
hero_image: '/images/why-von-neumann-was-right-hero.jpg'
hero_image_alt: 'Cover of Why von Neumann Was Right: Shannon Entropy, Boltzmann Entropy, and the Architecture of Artificial Energy, by Jed Anderson, 2026.'
supporting_files: []
---

*John von Neumann's instruction to Claude Shannon in 1948—to call his information measure entropy—was not a naming convenience. It was a statement of mathematical identity. The Boltzmann entropy of thermodynamics and the Shannon entropy of information theory are the same function, counting the same object: the logarithm of distinguishable configurations compatible with what is known. This identity, formalized by E.T. Jaynes through the principle of maximum entropy, has a physical consequence measured in twenty orders of magnitude: the [bond-bit asymmetry](/essays/bond-bit-ratio) between the cost of knowing and the cost of moving matter. This essay traces that identity from its mathematical origins through Maxwell's Demon and the Landauer limit to its architectural expression in [Artificial Energy](/essays/artificial-energy): the engineering of information-rich matter to harvest free-energy gradients the way the biosphere has for 3.5 billion years.*

## I. The Formula That Appears Twice

In the fall of 1948, Claude Shannon was finishing the most important paper in the history of communication theory and could not decide what to call his central quantity. He had considered "information"—too vague, already overloaded. He had settled tentatively on "uncertainty." Then he brought the problem to John von Neumann, who gave him an answer that has puzzled, amused, and occasionally irritated physicists ever since.

"You should call it entropy," von Neumann told him, "for two reasons. In the first place your uncertainty function has been used in statistical mechanics under that name, so it already has a name. In the second place, and more important, no one really knows what entropy really is, so in a debate you will always have the advantage."

The second reason is usually quoted as the wit of the exchange. The first reason is the one that matters. Von Neumann was not being poetic. He was pointing to something mathematically precise: the function Shannon had derived for measuring the information content of a message is the same function Boltzmann had derived for measuring the disorder of a thermodynamic system. Not analogous. Not similar in spirit. The same function, the same counting operation, the same logical structure, expressed in two different languages: one built from molecules in a gas, one built from symbols on a wire.

Von Neumann knew this because he had already written it down. In his 1932 treatise on quantum mechanics, he defined what is now called the von Neumann entropy, the quantum generalization of classical thermodynamic entropy, using exactly the functional form Shannon would independently derive sixteen years later. When Shannon showed him the formula, von Neumann recognized it immediately. He had been there before.

Boltzmann's entropy counts the number of ways the microscopic components of a physical system can be arranged while producing the same macroscopic state. Its expression is S = −k_B Σ pᵢ ln pᵢ, where pᵢ is the probability of occupying microstate i and k_B is Boltzmann's constant. This is the entropy of a gas, a crystal, a black hole, the quantity the Second Law says can never decrease in an isolated system: the arrow of time, the direction of disorder, the reason coffee cools and ice melts.

Shannon's entropy, derived from the requirements of consistent information measurement, is H = −Σ pᵢ log₂ pᵢ. This is the entropy of a message, a language, a genome. It determines how compressed a signal can be and how much uncertainty a measurement resolves.

Strip out the constants and the two expressions are the same object. Both count, in the same mathematical way, the logarithm of the number of distinguishable configurations of a system. Both are maximized when all configurations are equally probable. Both reach zero when the system is fully ordered. The structures are not analogous; they are isomorphic.

This identity cannot be a coincidence. Both expressions arise from the same underlying requirement: how do you consistently assign a number to the uncertainty of a probability distribution? There is essentially one answer, subject to mild assumptions about consistency. It was discovered twice: once in the language of molecules and heat, once in the language of symbols and transmission. The universe had only one answer to give and offered it to whoever asked correctly.

## II. What Jaynes Proved

The von Neumann anecdote identifies the identity. It took E.T. Jaynes, in his landmark 1957 papers, to explain what the identity means.

Jaynes made a claim that was audacious at the time and is now foundational: statistical mechanics is not a special branch of physics. It is a special case of Bayesian inference under entropy constraints. The Boltzmann distribution emerges from thermodynamics for the same reason it emerges from the principle of maximum entropy: both answer the question "Given what I know, what is the least biased probability distribution I should assign to the microstates of this system?"

That is an information-theoretic question, not a physics question. Physics provides the constraints. Information theory provides the method: maximize entropy, because maximum entropy is the unique probability distribution that assumes nothing beyond the stated constraints. Any lower-entropy assignment would smuggle in structure that is not actually known.

The consequence runs deep. The laws of thermodynamics are not merely the laws of molecules in motion. They are the laws of inference applied to physical systems. Entropy is not just a measure of molecular disorder; it is a measure of missing information (how much you do not know about which microstate the system occupies, given everything you know about its macrostate).

The Second Law is therefore not just a statement about molecules. It is a statement about information. A system's entropy increases as information about its precise microstate is lost. Environmental protection is entropy decrease, but only because it is information recovery. Pollution is entropy increase, but only because it is information loss: a molecule of benzene dispersed through groundwater is a molecule about which knowledge of location and configuration has been forfeited.

The thermodynamics and the information theory are the same theory. The formulas are the same formulas. Von Neumann was not being clever when he told Shannon to call it entropy. He was stating a mathematical fact.

## III. The Gate and the Gradient

Maxwell's Demon, proposed in 1867, sits at the hinge of everything that follows.

Maxwell imagined a tiny being stationed at a small door between two chambers of gas. By observing individual molecules and opening or closing the door selectively, the demon could sort the gas into a hot side and a cold side without doing apparent work. For nearly a century, the Demon seemed to violate the Second Law. The resolution, worked out by Szilard in 1929 and completed by Bennett in 1982, is that the Demon is an information-processing system. It must acquire information about each molecule, store it, and eventually erase it to continue operating. Memory erasure is physically irreversible: Landauer proved in 1961 that erasing one bit must dissipate at least k_BT ln 2 of heat. The entropy generated by erasure exactly compensates the entropy decrease in the gas. The Second Law survives.

But the resolution revealed something more important than the paradox's dissolution. The Demon is real. It works. Its energy cost is k_BT ln 2 per bit (the Landauer limit, approximately 2.87 × 10⁻²¹ joules per bit at room temperature), and that cost is extraordinary in its smallness.

A carbon-hydrogen bond costs approximately 6.86 × 10⁻¹⁹ joules to break: 240 times more expensive than processing one bit of information at the thermodynamic floor. This is the [bond-bit ratio](/essays/bond-bit-ratio). The fine-structure constant and electron mass determine bond energies; the Boltzmann constant and temperature determine the information floor. The 240× asymmetry is written into the electromagnetic structure of the universe.

Here is where the Shannon-Boltzmann identity becomes economically consequential. The Demon exploits this identity: it uses information to steer energy, paying the information price to avoid the thermodynamic price. At the molecular level, it pays 1 unit to control 240. At the macroscopic level, where information compresses and one bit about a valve's health can prevent 10²⁶ molecular bonds from scattering into groundwater, the ratio reaches 10²⁰: one hundred quintillion to one.

The Demon is not a paradox. It is a design pattern. The universe charges vastly more to rearrange matter than to know about it, by a factor derivable from first principles, fixed by fundamental constants, exploitable by any information-processing system with a sufficient model of its environment.

## IV. The Architecture of Artificial Energy

The biosphere figured this out 3.5 billion years ago.

Chlorophyll is a Maxwell's Demon realized in molecular architecture: matter structured with enough information, encoded in approximately ten thousand bits of evolved protein configuration per photosynthetic reaction center, to selectively couple to photons at 680 nanometers, route their energy along quantum-coherent pathways faster than thermal decoherence can scatter them, and deposit that energy in a chemical bond rather than releasing it as heat. The apparatus does not run down. The gradient (sunlight against the cold of space) supplies the energy. The molecular structure supplies the information. The information steers the gradient.

The key insight is that the information in a chlorophyll molecule is not incidental to its function; it is the function. Reduce the chlorophyll to its constituent atoms in disordered arrangement and there is no light harvesting, only an absorber that radiates heat. The structure is the technology. The Landauer limit tells us its thermodynamic cost: k_BT ln 2 per bit-flip, 240 times cheaper than breaking the bond the decision controls.

This is why Artificial Intelligence and [Artificial Energy](/essays/artificial-energy) are parallel not in some approximate or rhetorical sense, but in the precise sense that both technologies substitute information for force. AI substitutes information for cognitive labor. AE substitutes information for physical forcing. Both express the same underlying asymmetry: the universe charges less to know than to move, because information entropy and thermodynamic entropy are the same thing counted at different scales.

When an AI-designed catalyst selects one reaction pathway out of millions at ambient temperature, it pays the Landauer cost for the information encoded in its structure and collects the bond energy of the reaction it steers. When an engineered enzyme folds a protein into a specific functional configuration at room temperature (a task that blind synthesis would require the energy budget of a star to achieve across 10¹³⁰ possible sequences), it expresses the same ratio at the protein scale. The bond-bit asymmetry does not shrink as systems grow larger; it grows, because information compresses while matter does not.

A solar panel proves the gradient-harvesting regime is real. It is not the destination. The destination is the high-information end: designed catalysts, engineered enzymes, artificial reaction centers, adaptive materials that selectively route, transform, and combine energy flows the way a chlorophyll complex or a metabolic pathway does. Biology has operated at that frontier for 3.5 billion years. Engineering is just beginning to find it.

## V. What von Neumann Saw

Von Neumann's 1932 entropy, S = −tr(ρ ln ρ) where ρ is the density matrix of a quantum system, is the bridge between the classical and quantum regimes of this story. When the density matrix is diagonal, the von Neumann entropy reduces exactly to the Shannon entropy of those probabilities. When quantum coherence is present, it captures interference terms that classical Shannon entropy cannot represent.

This matters for Artificial Energy because quantum coherence is not incidental to the biosphere's efficiency; it is the mechanism. The Fleming group's 2007 experiments on the Fenna-Matthews-Olson complex in green sulfur bacteria showed that energy transfer through the photosynthetic antenna occurs via quantum-coherent wavelike motion, not classical hopping, achieving efficiency that classical diffusion cannot match because the quantum wave explores multiple pathways simultaneously and constructively interferes at the output.

Von Neumann knew in 1932 that entropy was a unified concept: the same mathematical object describing the disorder of a gas, the disorder of a quantum state, and (he would tell Shannon sixteen years later) the disorder of an information source. He wrote this not as a philosophical observation, but as a theorem. The trace formula reduces to the Boltzmann formula in the classical limit. The Shannon formula reduces to the Boltzmann formula when probabilities replace density matrices. One formula, three expressions, one underlying count: how many distinguishable configurations are compatible with what is known?

The universe operates by one logic. It charges less for knowing than for moving. It charges the same price, measured in the same units, whether the knowing is about a molecule's location, a quantum state's coherence, or a bit on a communication channel. The cost is k_BT ln 2. The asset purchased is order: locally maintained, carefully steered, precisely targeted.

Artificial Intelligence is what happens when that purchase is made in the cognitive domain, intelligence maintained in silicon and sustained against the entropy of noise and randomness. Artificial Energy is what happens when the same purchase is made in the physical domain: free-energy gradients harvested in engineered materials, ordered reactions sustained against the entropy of equilibrium, by information-structured matter approaching the thermodynamic efficiency of chlorophyll and enzyme.

Both are expressions of the Shannon-Boltzmann identity. Both are what life has always done. Both are what we are, finally and deliberately, learning to build.

Von Neumann was right—not because he was being clever about nomenclature, but because he saw the theorem. He saw that information and thermodynamics were the same discipline viewed from two angles, that entropy was the hinge between knowing and moving, and that any technology sophisticated enough to exploit that hinge would inherit the leverage written into their ratio.

The leverage is 10²⁰. The hinge is the Landauer limit. The theorem is Shannon-Boltzmann. The name is entropy. And the opportunity, now that we can finally act on what the formula was always telling us, is civilization-scale.

---

*Jed Anderson is the founder and CEO of EnviroAI and the author of [Artificial Energy: The Next Civilizational Technology Layer](/essays/artificial-energy), [The Intelligence Leverage Equation](/essays/intelligence-leverage-equation), and [The Missing Quadrillion](/essays/missing-quadrillion). The intellectual foundation of this work is maintained at jedanderson.org.*