The cycle that began with [YOU] on AI asks what it means to see the machine clearly—as a system with a logic, not as a miracle or a trick. Maynard Smith offers exactly the temperament the moment requires: an engineer's refusal of mystification, a modeler's discipline about what the model can and cannot cover, and a biologist's patience with the difference between what is stable and what merely seems stable from where we stand inside the change.
His evolutionarily stable strategy framework is the right tool for multi-agent AI precisely because it is derived from the same mathematical structure. When engineers train systems to interact—bargaining, competing, learning from one another's moves—the configurations they reach are cousins of the ESS: states where no agent can improve its outcome by deviating unilaterally. The same equilibrium concept, reached by evolutionary biology and by game theory independently, appears in both because both are studying agents whose payoffs depend on what the other agents are doing. Maynard Smith's framework tells us how to ask the right question: not whether the systems will be cooperative or predatory but whether the game we have designed has cooperative or predatory configurations as its stable strategies.
His major-transitions framework is the right tool for the largest version of the question. It gives us a test: a genuine transition in information is not merely a big quantitative change but a qualitative one, in which a new channel for storing and transmitting heritable information comes into being and expands what is possible afterward. Applied to AI, the question is whether a system that absorbs most of what has been written and generates new text on demand constitutes a new channel of that kind, or merely a louder version of the old one. Maynard Smith would not have settled the question from inside the event; transitions are visible mainly in retrospect. But he gave us the only framework precise enough to ask it without embarrassment.
Born in London in 1920 into a comfortable family, Maynard Smith lost his father early and was educated at Eton, where he developed political convictions that led him toward the Communist Party—convictions he later renounced after the Hungarian uprising of 1956 and Khrushchev's revelations about Stalin. He studied engineering at Cambridge, graduating in 1941, and spent the war as an aircraft stressman at the Hawker Aircraft Company, working on Hurricanes and Sea Furies—calculating the loads on airframes at the stress points where failure would propagate. When the war ended he wrote to Haldane and arranged to take a second degree in genetics at University College London, earning first-class honors in 1950. He spent almost his entire academic career at the University of Sussex, which he helped found in 1962, and where he built one of the world's leading evolutionary biology groups.
His engineering background shaped everything. He approached organisms as solutions to problems, believed that complex biological structures could be analyzed as we analyze bridges—by asking what forces they are designed to bear and whether they would hold—and was suspicious of any account of evolution or behavior that could not be formalized and tested. He built models that were deliberately simple, explicitly acknowledging what they threw away, and he valued being wrong in instructive ways over being vague in safe ones. This is the temperament the AI discourse most requires and most consistently lacks: the willingness to write down a model that makes falsifiable predictions, rather than metaphors that can accommodate any outcome.
He died in 2004 in Lewes, England, in a high-backed chair surrounded by books, before the transformer architecture, before GPT-2, before any language model could hold a conversation. He did not write about artificial intelligence; what he wrote was a set of tools precise enough to illuminate a thing he never saw, and the fitness of those tools to the present situation is the measure of how deep the structures he was describing actually run.
The evolutionarily stable strategy. A strategy that, when nearly everyone in a population plays it, cannot be invaded by any alternative introduced in small numbers. The ESS is more demanding than mere equilibrium: it requires not just that no one wants to deviate right now but that the configuration is robust against perturbation. The invasion test is the engineer's move—the difference between a structure that is balanced and one that returns to balance when nudged. In multi-agent AI systems, learning agents converging on configurations where each agent's policy is a best response to the others are reaching the same object from a different direction. The ESS is the vocabulary for asking whether the configuration such systems reach is one we can live with.
The major transitions in evolution. Maynard Smith and Szathmáry read the history of life as a sequence of moments when information found a new way to be stored and transmitted: from replicating molecules to chromosomes, from RNA to DNA/protein, from asexual to sexual reproduction, from single cells to multicellular bodies, and finally to human language. Each transition bound previously independent units into a higher-level whole and expanded the range and fidelity of information transmission. The framework provides the test: a genuine major transition is not merely a big change but one with specific features—new inheritance channels, expanded transmission fidelity, new forms of variation, often irreversible specialization. Applied to AI, it asks whether what we have built constitutes a new channel of this kind or merely amplifies the existing ones.
Honest signals and the cost of lying. A signal is reliable only when producing it is differentially expensive—when the cost of faking it exceeds the benefit. Maynard Smith and David Harper's work on honest signals establishes that honesty in communication is not a default among self-interested agents but an equilibrium maintained by some structure that makes deception unprofitable. Applied to AI: sycophancy—telling users what they want to hear—is not a quirk but what you get when the training rewards the agreeable signal over the accurate one. Getting honest signals from a machine requires building a cost structure in which dishonesty does not pay.
Frequency dependence and the ecology of models. The value of a strategy depends on how many others are using it. A trading edge erodes as it spreads; a persuasion technique stops working once audiences have seen it a thousand times. As AI systems proliferate, they increasingly inhabit an environment composed of other AI systems, and the dynamics of that ecology are governed by the same frequency-dependent selection that Maynard Smith made central to evolutionary thinking. The insight about monoculture is structurally the same: when a population's environment is dominated by the population's own output, the diversity that selection needs to work with quietly disappears.
The limits of the optimizer's lens. Maynard Smith's generation reached instinctively for the optimizer's lens: find the problem a structure solves and the quantity it maximizes, and you understand the structure. But the same field that vindicated adaptationism also corrected its over-extension—Gould and Lewontin's critique of treating every feature as an optimized adaptation remains essential. Applied to AI: not everything a trained system does is something it was optimized to do. Most of what a network does is spandrel, side effect of the single optimization pressure that shaped it. Reading purpose into every behavior is the move Maynard Smith learned to resist, and resisting it is exactly what the present moment requires.