John von Neumann: The Man Who Designed the Modern Computer

Princeton, New Jersey. 1946. A small group of mathematicians and engineers are gathered in a seminar room at the Institute for Advanced Study. At the front of the room stands a man of medium height, slightly round, impeccably dressed in a three-piece suit despite the informality of the occasion. He is speaking quickly, filling a blackboard with equations, turning occasionally to address the room with a smile that suggests he finds the material as delightful as a particularly good joke.

The subject is the design of a new kind of computing machine — one in which the program, the instructions telling the machine what to do, will be stored in the same memory as the data it operates on. This is a simple-sounding idea. Its implications are enormous. Every computer ever built since that day — every laptop, every smartphone, every server farm, every AI training cluster — is built on the architecture being described at this blackboard.

The man at the blackboard is John von Neumann. He is forty-two years old. He has already done enough mathematics to secure a place in history several times over. He has at least twelve more years of extraordinary work ahead of him.

He will use every one of them.

A Mind That Arrived Fully Formed

John von Neumann was born on December 28, 1903, in Budapest, Hungary, the eldest son of Max Neumann, a prosperous banker, and Margaret Kann. He was born into a world of comfort, culture, and intellectual ambition — the Jewish professional class of fin-de-siècle Budapest, a community of unusual density and brilliance that would, over the following decades, produce a disproportionate share of the 20th century’s most important scientists and mathematicians.

Budapest in the early 1900s was a city in the full flower of its imperial confidence. Part of the Austro-Hungarian dual monarchy, it had experienced decades of rapid modernization and cultural flourishing. Its coffeehouses were alive with political debate, literary experiment, and scientific discussion. Its schools were rigorous and demanding. And its Jewish professional families — ambitious, educated, excluded from certain social circles but intensely invested in intellectual achievement — produced, in the generation born around the turn of the century, a cohort of genius that history has rarely seen concentrated in one place and time.

Von Neumann was identified as exceptional almost from infancy. By the age of six, he could divide two eight-digit numbers in his head and exchange jokes with his father in classical Greek. By eight, he had memorized a volume of the Budapest telephone directory for amusement. By twelve, he had mastered calculus. These stories have the quality of legend — they have been told so many times and with such embellishment that the facts are hard to separate from the mythology. But the people who knew von Neumann as a child and as a young man were consistent: his mathematical ability was not just extraordinary but of a qualitatively different order from what they had encountered before.

His memory was the feature most often remarked upon. Von Neumann appeared to remember everything he read, more or less permanently. He could quote at length from books he had read decades earlier. He could recall the details of mathematical proofs he had encountered once as a student. He could reproduce conversations, lectures, and papers with unnerving accuracy. Whether this was truly photographic memory in the neurological sense is debated by those who have studied his life. What is not debated is that his ability to hold vast amounts of information in accessible, organized form gave him a decisive advantage in every intellectual domain he entered.

He was, by all accounts, also a person of great warmth and humor. He loved parties, gossip, and dirty limericks in multiple languages. He drove badly and cheerfully, accumulating a remarkable record of minor accidents. He was an enthusiastic consumer of what he called “mental junk food” — detective novels, especially, which he read at great speed and in enormous quantities. He was a lavish host, throwing parties at his Princeton home that were famous for their food, their drink, and the quality of the conversation.

He was, in other words, formidably human alongside his formidable intellect — a quality that distinguished him from the popular image of the isolated, unworldly genius, and that made him effective in the practical, political, and organizational dimensions of the scientific world in ways that more purely theoretical figures were not.

The Education of a Prodigy

Von Neumann’s formal education was shaped by the tension between his extraordinary abilities and the conventional structures of schooling. The schools of Budapest — excellent by any standard — were simply not equipped for a child who mastered their curriculum years ahead of schedule. His parents navigated this by arranging for him to be tutored privately by university mathematicians while he continued to attend school in the normal way for the social and general educational benefits.

The arrangement worked. By the time von Neumann was seventeen, he was already corresponding with professional mathematicians about research-level problems. By eighteen, he had published his first mathematical paper — on the theory of transfinite ordinals — in collaboration with Michael Fekete, one of his tutors. This paper was not a student exercise. It was a contribution to an active area of mathematical research, and it was recognized as such.

He enrolled simultaneously at two universities: the University of Budapest, where he was nominally a mathematics student, and the ETH Zürich, where he studied chemical engineering — a concession to his father’s pragmatic concern that pure mathematics might not be a reliable source of income. He appeared at Budapest only to take examinations, which he passed with distinction without attending lectures. He actually studied in Zürich, where the engineering curriculum gave him a grounding in applied science that would serve him well throughout his career.

In 1926, at twenty-two, he completed both degrees simultaneously — a PhD in mathematics from Budapest and a diploma in chemical engineering from ETH. The PhD thesis was on the axiomatization of set theory — one of the foundational problems of early 20th century mathematics, concerning the logical foundations on which the whole edifice of mathematics rested. The thesis was technically impressive enough that David Hilbert, then the most eminent mathematician in the world, invited von Neumann to Göttingen as his assistant.

Von Neumann spent time in Göttingen, the center of world mathematics in the 1920s, absorbing the most advanced work in quantum mechanics, mathematical logic, and the foundations of mathematics. He contributed to all three fields with remarkable speed. His work on the mathematical foundations of quantum mechanics — his 1932 book Mathematische Grundlagen der Quantenmechanik — provided the rigorous mathematical framework for quantum theory that it had previously lacked, and it remains the standard reference for the mathematical structure of quantum mechanics.

He was twenty-eight when he published it.

Princeton: The Institute for Advanced Study

In 1933, as the Nazis consolidated power in Germany and the brilliant world of Central European Jewish intellectual life began its catastrophic dissolution, von Neumann accepted a position at the newly founded Institute for Advanced Study in Princeton, New Jersey. He was one of the institute’s six founding faculty members — the others included Albert Einstein and Hermann Weyl — and he was, at twenty-nine, the youngest.

The Institute for Advanced Study was an extraordinary institution — a place explicitly designed to free the world’s best mathematical minds from teaching, administrative duties, and all practical obligations, so they could think without distraction. It had no students, no departments, no requirements of any kind beyond the implicit requirement that its members produce important work. It was, in a sense, the purest possible expression of the view that great mathematics happened when great mathematicians were left alone with their thoughts.

Von Neumann thrived there — and also, characteristically, exceeded its implicit mandate. Where the institute’s ethos was contemplative and purely theoretical, von Neumann was restless, practical, and drawn to problems where mathematics could do things in the real world. He consulted widely. He traveled constantly. He was involved in more different projects, in more different fields, than any other person at the institute. His colleagues sometimes found this breadth bewildering. His contributions were usually decisive.

In the late 1930s and early 1940s, the approaching and then actual world war drew von Neumann into the most consequential applied work of his life. He became a consultant to the US Army, then to the Manhattan Project. His specific expertise was in the mathematics of shock waves — the explosive dynamics of bombs and shells — which was both technically challenging and strategically crucial. He was, by several accounts, one of the most effective scientists working on the design of the atomic bomb, not because he had built it — that was Oppenheimer, Fermi, Teller, and the engineers — but because he had solved key mathematical problems in the design of the implosion mechanism that triggered the plutonium bomb.

The moral weight of this contribution is heavy and contested. Von Neumann was not naive about what he was doing. He worked on the bomb deliberately and energetically, with full understanding of its destructive potential. He was a hawk among hawks on the question of the Soviet Union — he believed, with a consistency that was sometimes alarming to his colleagues, that the United States should use its nuclear monopoly while it lasted, that waiting for the Soviets to develop their own bomb was a catastrophic mistake. The phrase “preventive war” — the idea of attacking the Soviet Union before it could match American nuclear capability — was associated with von Neumann in the late 1940s, though the story has been somewhat simplified in the retelling.

He was a man of his time, shaped by the experience of watching his world — the world of European Jewish intellectual life — destroyed by a totalitarian state. His hawkishness about the Soviet Union was not bloodlust. It was the terror of a man who had seen what happened when a totalitarian regime was allowed to grow unchecked, and who was determined, by whatever means necessary, that it would not happen again. The calculations he made about nuclear deterrence were the calculations of someone who had already lived through the consequences of being unprepared.

None of this makes his contributions to nuclear weapons design morally straightforward. It does make them humanly comprehensible.

The EDVAC Report: Designing the Modern Computer

The contribution for which von Neumann is most directly relevant to the history of AI is not his mathematics, not his quantum mechanics, and not his work on the bomb. It is a document he wrote in 1945 that is known, slightly inaccurately, as the EDVAC Report — formally, the “First Draft of a Report on the EDVAC.”

The EDVAC was a computer being developed at the University of Pennsylvania under the direction of John Mauchly and J. Presper Eckert, the engineers who had also built ENIAC — the first general-purpose electronic computer. EDVAC was intended to be ENIAC’s successor, more flexible and more powerful. Von Neumann had become involved with the project as a consultant, and in June 1945 he wrote a draft report describing the logical design of the machine.

The report was circulated within the project and then, without von Neumann’s full authorization, distributed more widely. It became one of the most influential documents in the history of computing.

What the EDVAC report described was the stored-program architecture — the design principle that would define every general-purpose computer built since. The key idea was simple but transformative: the program, the instructions telling the machine what to do, should be stored in the same memory as the data the machine was operating on.

This sounds obvious in retrospect. But before the stored-program architecture, computers like ENIAC were programmed by physically rewiring — by plugging and unplugging cables that determined which operations would be performed in which order. This was slow, inflexible, and required significant human labor every time the machine was to perform a different task. A new computation required, in effect, rebuilding part of the machine.

The stored-program concept changed this fundamentally. If programs were stored in memory, they could be loaded and changed as quickly as data could be loaded and changed. The machine did not need to be physically reconfigured. You loaded a new program and the machine did something new. The same hardware could run any computation — limited only by memory, speed, and the ingenuity of the programmer.

There was a deeper implication that von Neumann clearly understood: if programs were stored in memory as data, then a program could, in principle, read and modify other programs — including, eventually, itself. The machine could be made to work on its own instructions the same way it worked on numerical data. This was the door through which self-modifying code, interpreters, compilers, and eventually the whole ecosystem of software would eventually walk. It was also, though this would take decades to fully develop, the door through which machine learning would eventually walk: programs that modified themselves based on experience.

The stored-program concept was not entirely von Neumann’s alone. Turing had described something very similar in his theoretical work on the universal Turing Machine, which could simulate any other Turing Machine by reading the description of that machine as data. The Manchester and Cambridge teams working on early British computers were developing similar ideas. The priority question is genuinely complicated, and some of the engineers on the EDVAC project felt, with some justification, that von Neumann had received credit for ideas that were at least partly theirs.

What is not complicated is the influence of the EDVAC report. It circulated widely. It was clear and well-written. It provided the conceptual framework in which the whole subsequent generation of computer builders worked. The architecture it described — now called the von Neumann architecture — is the architecture of essentially every general-purpose computer built in the decades since. It is the architecture of the computers on which AI programs run today.

Building MANIAC: The Computer at the Institute

Von Neumann was not content to design computers on paper. He wanted to build one. And in the late 1940s, he directed the construction of one of the first stored-program computers in the United States — the MANIAC, or Mathematical Analyzer, Numerical Integrator and Automatic Computer — at the Institute for Advanced Study in Princeton.

The project was characteristic of von Neumann in several ways. It was ambitious far beyond what most of his Institute colleagues thought appropriate for a pure research institution. It was technically sophisticated, incorporating design choices that were years ahead of other contemporary computers. It attracted the best engineers and mathematicians available. And it produced results that justified the ambition.

The IAS machine — as it was also known — was completed in 1952 and ran its first programs shortly after. It was used for an extraordinary range of calculations: weather prediction, nuclear weapons design, early work on fluid dynamics, and — significantly for AI history — some of the first experiments in what von Neumann called “automata theory,” the mathematical study of self-reproducing and self-organizing computational systems.

The IAS machine’s design was widely copied. Machines based on its architecture were built at research institutions across the United States and Europe, spreading the stored-program concept and making it the de facto standard for computer design. Von Neumann had not just described the modern computer. He had built a version of it that became the template.

During the construction and early operation of the IAS machine, von Neumann’s thinking about what computers could do was expanding rapidly. He was not interested only in numerical calculation — the solving of differential equations, the simulation of physical systems. He was interested in the deeper question of what computation could do in principle — what kinds of systems could be built, what kinds of behavior could emerge from computational processes.

This interest led him to some of the most original and farsighted work of his career: his theory of self-reproducing automata.

Self-Reproducing Automata: The Theory of Life and Machine

In a series of lectures and papers in the late 1940s and early 1950s — some published in his lifetime, others reconstructed from notes and published posthumously — von Neumann developed a mathematical theory of self-reproducing automata. It was, and remains, one of the most extraordinary bodies of work in the history of ideas.

The central question was: what is the minimum complexity required for a system to reproduce itself? This was not a biological question, though it was inspired by biology. It was a mathematical question. Von Neumann wanted to understand the abstract, logical structure of self-reproduction — to find the simplest system that could make a copy of itself, and to characterize what such a system required.

His approach was to work with cellular automata — a mathematical model that he developed in collaboration with Stanislaw Ulam, a fellow mathematician at Los Alamos. A cellular automaton is a grid of cells, each in one of a finite number of states, that evolves over time according to simple local rules: each cell’s next state is determined by its current state and the states of its neighbors. The rules are applied simultaneously to all cells, producing a new configuration, then applied again, and so on.

Von Neumann showed that in a sufficiently rich cellular automaton — one with enough states and complex enough rules — it was possible to construct a configuration that could read a description of itself and produce a copy. The self-reproducing automaton had two essential components: a constructor, which could build any machine whose description was provided, and a description of itself, which the constructor could use to build a copy.

This separation of the constructor from its description is profoundly important. The description was, in effect, a program — a set of instructions that, when read by the constructor, produced a specific output. The self-reproducing automaton worked by having the constructor read the description and build the described machine, then copy the description and insert it into the new machine. The new machine then had both the constructing capability and the description — and could reproduce again.

Von Neumann’s self-reproducing automaton was a theoretical system of enormous complexity — his design required cellular automata with twenty-nine possible states per cell, and the self-reproducing configuration contained over 200,000 cells. It was never computationally implemented in his lifetime. But the theory it embodied was immediately recognized as important.

For AI, the self-reproducing automaton demonstrated several things. First, it showed that self-replication was not a mystical property of life but a computational one — it could be precisely defined and theoretically constructed in a formal system. This was a powerful argument against vitalism, the view that life required some special vital principle beyond the physical. Life, at least in its most fundamental reproductive aspect, could in principle be an entirely mechanical phenomenon.

Second, the self-reproducing automaton introduced the concept of a universal constructor — a machine that could build anything whose description was provided. This is, in the most literal sense, the concept underlying 3D printing, molecular nanotechnology, and — more relevantly for AI — the idea of a general-purpose learning system that could be trained to perform any task by providing it with an appropriate description of that task.

Third, von Neumann’s work on automata theory was the direct precursor to the work of John Conway on the Game of Life in 1970, which demonstrated that extraordinarily complex and lifelike behavior could emerge from extremely simple cellular automaton rules. Conway’s Game of Life became one of the most important demonstrations in the history of artificial life research — a proof of concept that complexity and apparent purposiveness could emerge from simplicity and mechanism without any special design.

Game Theory: The Mathematics of Strategic Interaction

Von Neumann’s contributions to AI cannot be fully understood without discussing his work on game theory — the mathematical study of strategic interaction, which he developed with the economist Oskar Morgenstern and published in their 1944 book Theory of Games and Economic Behavior.

Game theory was not, in the first instance, an AI project. It was an attempt to put economics on a rigorous mathematical foundation by treating economic behavior as strategic decision-making — as the choices of rational agents who were each trying to maximize their own outcomes in situations where the outcomes depended on the choices of others.

The key conceptual tool von Neumann developed was the minimax theorem, which he first proved in 1928. In a two-player zero-sum game — a game where one player’s gain is exactly the other’s loss — the minimax theorem showed that there was always an optimal strategy: a strategy that minimized the maximum possible loss, or equivalently, maximized the minimum possible gain. The optimal strategy might involve randomization — deliberately playing randomly among several options according to specific probabilities — but it always existed and could in principle be calculated.

Game theory became foundational for AI in several ways.

Most directly, it provided the mathematical framework for game-playing AI programs. Chess, Go, poker, and other games are strategic interactions, and game theory gives the conceptual vocabulary — optimal strategies, minimax search, Nash equilibria — that AI game-playing programs use. The minimax algorithm that underpins chess engines is a direct computational implementation of von Neumann’s minimax theorem. When AlphaGo defeated the world Go champion in 2016, it was using reinforcement learning techniques that had grown from the same mathematical soil that von Neumann had cultivated.

More broadly, game theory provided a framework for thinking about rational decision-making in general — about what it meant to be a rational agent, what rational agents would do in various situations, and how the interactions between rational agents produced collective outcomes. This framework has been enormously influential in AI research on multi-agent systems, on the design of AI systems that need to interact with humans or other AI systems, and on the theoretical foundations of machine learning.

Game theory also, less directly but perhaps more importantly, contributed to the conceptual vocabulary of AI ethics. Many of the most important questions in AI alignment — how to ensure that AI systems act in ways aligned with human values and human interests — are game-theoretic questions. How do you design an AI system whose incentives are aligned with the people who use it? What happens when multiple AI systems interact? How do you ensure cooperation rather than defection in systems that are capable of strategic reasoning? These are questions that von Neumann’s framework helps formulate, even if it does not by itself answer them.

The Brain and the Computer: Von Neumann’s Last Work

In the final years of his life, increasingly ill with the bone cancer that would kill him at fifty-three, von Neumann turned his attention to a question that had fascinated him for years and that he had never found time to address fully: the relationship between the brain and the computer.

The result was a series of lectures he gave at Yale in 1956, which were unfinished at his death and published posthumously in 1958 as The Computer and the Brain — a short, dense, remarkable book that is at once a technical comparison of biological and artificial computation and a meditation on the nature of mind.

Von Neumann approached the comparison with characteristic rigor. He examined the known facts about neural computation — the speed of neural signals, the number of neurons, the connectivity of neural networks — and compared them systematically with the characteristics of contemporary computers. The comparison was illuminating in both directions. Computers were faster than neurons at individual operations. Neurons were vastly more numerous and more interconnected than the components of any existing computer. The brain was a massively parallel system; computers were sequential. The brain was reliable despite the unreliability of its individual components; computers were unreliable whenever their components failed.

The most interesting and prescient part of the book was von Neumann’s discussion of the logical language of the brain. He observed that the brain’s computational processes could not, as far as anyone could tell, be described in the formal logical languages that had been developed for computers — in the Boolean algebra of classical logic, or the arithmetic of digital computation. The brain seemed to use a different kind of mathematics — perhaps a statistical one, operating on probabilities rather than certainties, on distributions rather than discrete values.

This was a prescient observation. The dominant paradigm in AI in the late 1950s and for the next several decades was symbolic AI — formal logic, explicit rules, discrete computation. The neural network tradition — statistical, probabilistic, operating on continuous rather than discrete representations — was the minority view. Von Neumann, in his last work, was suggesting that the neural approach was more likely to be right, that the brain’s mathematics was statistical rather than logical.

He did not live to see neural networks triumph. He did not live to see deep learning, or to see the systems trained on vast datasets that would vindicate his intuition about the statistical character of neural computation. He died in February 1957, nine months after the Dartmouth Conference, eight months after the IAS machine was shut down, with The Computer and the Brain unfinished.

He reportedly spent his final months in the hospital still thinking, still working, dictating equations and ideas to the people who visited him. His friend and colleague Herman Goldstine described visiting von Neumann toward the end and finding him in remarkable intellectual form — still sharp, still engaged, still full of ideas — even as his body failed. He was trying to finish the work, trying to get it all out before the time ran out.

He did not quite manage it. The book ends mid-thought. A chapter he intended to write — on the statistical character of neural computation — was never begun.

The Shadow Side: Power and Responsibility

John von Neumann’s life raises questions about the relationship between intellectual power and moral responsibility that have not become less urgent with time. He was one of the architects of the atomic bomb. He was a vocal advocate for preventive nuclear war. He sat on the boards and advisory committees that made some of the most consequential and dangerous decisions of the Cold War era. He was, in the political dimension of his life, a hawk of extraordinary influence.

How should we weigh this against his intellectual contributions? The question is not comfortable, and it does not have a comfortable answer.

One answer — the one his defenders most often give — is that von Neumann was a realist who understood the stakes of the Cold War more clearly than most, and that his hawkishness was not irresponsibility but a calculated response to a genuine and terrible threat. The Soviet Union was a totalitarian state that had already killed millions of its own citizens. Von Neumann had watched, from a distance, the destruction of European Jewish civilization by another totalitarian state. His determination that the Western democracies should not be similarly unprepared was not irrational.

Another answer — one that his critics have offered — is that the willingness to calculate about the acceptable costs of nuclear war, to think of cities and populations as variables in an optimization problem, represented a failure of moral imagination. That his extraordinary analytical power was, in this domain, deployed in service of conclusions that no adequate moral framework could support.

Both answers have something to them. Von Neumann was a man of his time, formed by experiences that produced a specific and understandable worldview. He was also a man of extraordinary influence who used that influence in ways whose consequences were not always benign. The two things are both true, and the tension between them does not resolve neatly.

For AI, the shadow side of von Neumann’s legacy raises questions that are directly relevant to the present. The people who build powerful technologies — and von Neumann built some of the most powerful technologies in history — bear a special responsibility for thinking about how those technologies will be used and what consequences they will produce. Von Neumann thought about this, in his way — he was not naive or incurious about the implications of what he was building. But his thinking was shaped by his politics, his fears, and his institutional position in ways that produced conclusions that not everyone would endorse.

The researchers who build AI systems today face similar questions. The systems they build are becoming powerful enough to reshape economies, influence politics, conduct surveillance, and potentially one day make decisions about war and peace. The moral frameworks they bring to this work — the values they hold, the risks they attend to, the consequences they consider — will shape what gets built and how it gets used. Von Neumann’s example is a reminder that extraordinary intellectual power does not automatically produce wise moral judgment, and that the two capacities need to be developed together.

What Von Neumann Left Behind

John von Neumann died on February 8, 1957. He was fifty-three years old.

He left behind the mathematical foundations of quantum mechanics. He left behind the theory of games and economic behavior. He left behind the stored-program computer architecture. He left behind the theory of self-reproducing automata and cellular automata. He left behind The Computer and the Brain — incomplete, but still one of the most important early attempts to think seriously about the relationship between biological and artificial computation.

He left behind, in the form of the IAS computer and the many machines built on its design, the practical infrastructure on which the digital revolution was built. Every computer program ever run — including every AI program — has run on hardware whose logical architecture traces back to the EDVAC report of 1945.

He left behind, in the form of his students and collaborators and the institutions he helped build, a tradition of approaching computation with mathematical rigor and practical ambition simultaneously — of not being content with either pure theory or practical engineering but insisting on both at once.

And he left behind a question — the question posed but not answered in The Computer and the Brain — about the relationship between the computer’s kind of intelligence and the brain’s kind of intelligence. About whether the formal, logical, sequential architecture of the von Neumann computer was really the right model for intelligence, or whether intelligence required something more like what the brain was doing — something statistical, parallel, approximate, and emergent.

The deep learning revolution of the 2010s provided a provisional answer: something more like the brain. The neural network architectures that produce modern AI — transformers, convolutional networks, recurrent networks — are not von Neumann machines in their logical character, even though they run on von Neumann hardware. They are closer to what von Neumann himself was groping toward in The Computer and the Brain — statistical, distributed, emergent from the interaction of many simple elements rather than the sequential execution of explicit rules.

Von Neumann was, in his last work, beginning to see past the architecture he had created. He was beginning to understand that the stored-program computer, brilliant as it was, might not be the end of the story — that the brain was doing something different, and that understanding that difference might be the key to building machines that were truly intelligent.

He ran out of time before he could work it out. Others have spent the decades since trying to finish the thought.

The Martian in Princeton

Von Neumann’s Hungarian physicist friends — Leo Szilard, Eugene Wigner, Edward Teller, and others who had made similar journeys from Central Europe to the United States — were sometimes called the Martians by their American colleagues. The joke was that they were so obviously not from this planet that there was no other explanation.

The joke was affectionate, but it pointed at something real. These men had grown up in a culture of intellectual intensity that produced a kind of mind not commonly seen — broadly educated, deeply mathematical, effortlessly multilingual, comfortable with abstraction and equally comfortable with practical application, formed by the experience of having their world destroyed and rebuilt and destroyed again by forces of history that admitted no sentimentality.

Von Neumann was the most Martian of the Martians. His speed of calculation, his range of knowledge, his ability to move between pure mathematics and practical engineering and political strategy without apparent effort — these were the qualities that made people who worked with him feel, occasionally, that they were in the presence of something not quite human.

His colleague Herman Goldstine, who worked closely with him on the IAS computer, wrote that being around von Neumann was to be in the presence of a mind operating at a different level — not faster in the way a faster processor is faster, but different in the way a different kind of processor is different. He did not just calculate more quickly. He saw structures and relationships that others simply did not see, as if the mathematical landscape were more transparent to him than to everyone else.

This quality — the sense that von Neumann was operating with a different cognitive architecture, not just more hardware — is relevant to AI in an interesting way. One of the questions AI research is trying to answer is whether human intelligence is a matter of degree or a matter of kind — whether sufficiently powerful AI will be continuous with human intelligence, just faster and with more memory, or whether it will be qualitatively different, operating with cognitive structures and capacities that are genuinely alien to human experience.

Von Neumann was, in a human body, a data point for the latter possibility. His colleagues genuinely could not always follow his reasoning — not because they lacked mathematical training but because the connections he drew were not connections they would have drawn, the structures he saw were not structures they were looking for. He was, by most accounts, not especially good at explaining his reasoning in ways that made it fully accessible to others. He could demonstrate results. He could not always help others see what he saw.

Whether AI systems will someday be similarly opaque — producing outputs that are correct but inexplicable, drawing connections that human minds cannot follow — is one of the active concerns of AI safety research. Interpretability — the project of understanding what AI systems are actually doing internally — is partly motivated by the worry that systems operating with radically non-human cognitive structures might be right for wrong reasons, or wrong in ways we cannot detect, or pursuing goals we cannot fully understand.

Von Neumann was not a safety risk. He was a colleague, a human being, embedded in human institutions and human relationships. But he was a reminder that intelligence, even human intelligence, could operate in ways that were genuinely difficult for other intelligent beings to fully follow — and a preview, perhaps, of the interpretability challenges that more powerful AI systems will eventually pose.