Terry Pratchett once slipped a quiet piece of contraband into humor: the idea that “million-to-one chances crop up nine times out of ten.” It reads like a joke about narrative excess, about heroes surviving impossible odds because the story demands it. But it’s doing something more subversive. It suggests that what we call “unlikely” is often just poorly modeled—that the world contains hidden variables, asymmetries, and feedback loops that naive probability flattens out. In other words, the punchline isn’t that stories distort reality; it’s that reality, under the right conditions, actually behaves like a story.
The old fable of The Tortoise and the Hare has always been misread as a moral about effort versus laziness. That’s the kindergarten version. A more interesting reading—especially now, in the age of AI—is that it’s a story about optimization strategies under uncertainty.
The hare is a specialist. High-speed, high-variance, metabolically expensive. In evolutionary terms, it’s an organism tuned for peak performance in stable environments—what biologists might call a K-selected competitor pushed toward an r-strategy edge case: explosive bursts, fast gains, catastrophic lapses. The tortoise, meanwhile, is not “hardworking.” It is robust. Low energy expenditure, low variance, slow but anti-fragile in the face of noise.
Translate this into AI systems and things get uncomfortable.
Modern frontier models—call them the hares—are optimized for benchmarks. They sprint through standardized tests, saturate leaderboards, compress enormous datasets into fluent output. They are trained via gradient descent toward narrow objective functions, often shaped by techniques like Reinforcement Learning from Human Feedback. In evolutionary terms, RLHF is artificial selection: we are breeding hares for docility, speed, and surface coherence.
A predictable accelerationist reading would dismiss all of this as Aesop dressed up in systems language—moralizing disguised as analysis, tortoise propaganda for the slow and risk-averse. In that framing, any invocation of robustness, embodiment, or constraint becomes “cope”: a sentimental attachment to friction in a world supposedly defined by its removal. But that dismissal only works by quietly assuming what it claims to prove—that speed, scale, and recursive optimization are self-justifying values rather than contingent strategies with failure modes. It treats acceleration as both method and metaphysics, as if going faster were identical to going deeper. And in doing so, it performs its own intellectual sleight of hand: rebranding survivorship bias as inevitability, and calling any reminder of collapse psychology. The irony is that this posture depends entirely on ignoring the historical record of systems that optimized for velocity until they outran their own coordination, their own signal, and eventually their own legibility. Calling that “cope” is not an argument—it’s a reflexive refusal to update.
The correct way to read the fable is not as allegory but as a kind of remote viewing—looking through the animals rather than at them. Granny Weatherwax would call it “headology,” but it’s closer to a distributed perception found in many indigenous traditions: the animal is not a symbol standing in for a human trait, but a lens onto a mode of being. The hare is not “arrogance,” the tortoise not “persistence.” They are strategies embodied—metabolisms, tempos, relationships to risk and environment. To read the fable properly is to temporarily inhabit those strategies, to feel the world at their respective speeds, their thresholds of fatigue, their exposure to variance. From that vantage, the outcome is no longer moral or surprising; it becomes legible as an interaction between two ways of processing reality. The story stops being about character and becomes about ecology of behavior—a small simulation of how different forms of intelligence couple to the same environment.
There is, running parallel to all this, a quieter but more telling selection pressure: vanity. For every lab using large models to probe protein folding or climate dynamics, there are thousands of deployments aimed at polishing prose, inflating output, or scaffolding brittle micro-apps that simulate productivity without extending capability. This is not just a sociological footnote—it is an evolutionary signal. Systems get optimized for what they are most frequently used for. If the dominant use case is self-presentation—tidying blogs, automating tone, generating endless near-duplicates—then the selection function drifts toward surface coherence over structural insight. In biological terms, it’s sexual selection run amok: traits that signal fitness (fluency, confidence, stylistic polish) outcompete traits that are fitness (robust reasoning, causal inference, genuine discovery). The result is a population of hares bred for display—fast, impressive, and increasingly fragile—while the slower, less glamorous investments in R&D, instrumentation, and ground-truth data collection are comparatively starved. The system doesn’t just reflect vanity; it begins to evolve for it.
But selection pressures are always incomplete.
Running beneath all of this is a deeper, almost unspoken failure: disembodiment. A whole generation raised in increasingly abstracted, screen-mediated environments has learned to treat cognition as something that floats—detached from metabolism, sensation, constraint. Large language models both reflect and amplify this tendency. They produce language without bodies, judgment without stakes, synthesis without exposure to consequence. And rather than correcting for this, much of their use doubles down on it: more time in the symbolic layer, more reliance on mediated inference, less contact with the friction of the real. But in evolutionary and biological terms, intelligence is not separable from embodiment; it is produced by it. Organisms know the world through constraint—through hunger, fatigue, risk, error. Strip that away, and what remains is a kind of ghost cognition: fluent, recursive, but ungrounded. The danger is not that people will believe LLMs are conscious, but that they will begin to model themselves after systems that lack the very conditions that make intelligence adaptive in the first place.
In evolutionary biology, fitness landscapes are rugged, not smooth. Peaks shift. Local optima trap organisms. The hare evolves to dominate one hill, only to find the environment has moved. The tortoise, by contrast, occupies a flatter region of the landscape—never optimal, but rarely catastrophically misaligned.
This is where the fable sharpens.
AI systems that optimize aggressively for short-term performance—speed, fluency, persuasion—may be overfitting to the current environment of human expectations. They become exquisitely tuned to what we can measure. But what we can measure is always a lagging indicator of what matters.
Meanwhile, slower, less flashy systems—ensembles, modular architectures, hybrid symbolic-statistical approaches—behave more like tortoises. They are less impressive in demos. They do not “win the race” in obvious ways. But they degrade more gracefully. They generalize more reliably under distribution shift.
In evolutionary terms, this is the difference between exploitation and exploration. The hare exploits a known niche. The tortoise preserves optionality.
There’s also a metabolic analogy worth taking seriously. Biological speed is expensive. Fast-twitch muscle fibers burn energy quickly and fatigue just as fast. In AI, compute is metabolism. Large models burn staggering amounts of energy to maintain their speed and responsiveness. The tortoise strategy—smaller models, slower inference, more structured reasoning—resembles an organism that can survive famine.
And famine is coming, metaphorically speaking.
Not a collapse of compute, but a collapse of clean signal. As AI systems begin to train on outputs generated by other AI systems, the environment becomes saturated with synthetic data—what some researchers describe as model collapse. In such an environment, the hare’s strategy—rapid ingestion and optimization—can amplify errors. The tortoise’s strategy—slower, more conservative updating—may actually preserve fidelity longer.
What current geopolitical friction is beginning to expose—whether in the tightening calculus around the Strait of Hormuz or in the broader drift away from uncontested blue-water supremacy—is not simply the obsolescence of specific platforms, but the exhaustion of an entire way of seeing. The doctrine of power projection—carrier strike groups, billion-dollar systems, $90 million aircraft—presumes a world that is legible, hierarchical, and slow enough to dominate through scale and coordination. But that world is fragmenting. Power is becoming more distributed, more ambiguous, more resistant to spectacle. And yet the institutional response has been to double down: more complexity, more cost, more refinement of the same underlying model.
Large language models emerge from a strikingly similar epistemic framework. They are built on the premise that if enough data is aggregated, if enough parameters are tuned, if enough signal is extracted from the noise, then reality becomes tractable—predictable, even governable. But this is a bet on a certain kind of world: one where the map can asymptotically approach the territory. As the environment destabilizes—whether through synthetic data feedback loops, shifting human norms, or adversarial inputs—the models do not so much fail as they begin to produce outputs that are “not even wrong”: internally coherent, statistically plausible, and operationally useless. They are optimized to win within their own representational system, not to survive contact with a changing reality.
And then there’s the so-called abundance problem, which is really an access problem in disguise. We talk about infinite content, infinite generation, a glut of language pouring out of systems—but if that output is immediately gated, fragmented, and sealed behind paywalls, it doesn’t behave like abundance. It behaves like a field of asteroids: plenty of material in aggregate, but locked into discrete, inaccessible chunks, each requiring its own toll to reach. This is not a post-scarcity information economy; it’s a recomposition of scarcity at a higher resolution. Instead of lacking content, we lack continuity. Instead of open flows, we get monetized fragments. In that sense, the system mirrors the same error Pratchett points to: mistaking the surface appearance (plenty) for the underlying structure (constraint), and then wondering why the outcomes don’t match the narrative of abundance.
A Gregory Bateson reading would push this even further, into the realm of epistemology and ecology of mind. For Bateson, the unit of survival is not the individual organism or the isolated system, but the organism plus environment—a circuit of information, feedback, and correction. Pathology emerges when that circuit is broken or misread, when a system begins to treat its own representations as sufficient and loses the ability to register difference from the outside. What we are seeing, in both AI and the institutions that deploy it, is a kind of recursive epistemic error: the map feeding back into itself, the signal mistaken for the territory, the correction mechanisms gradually atrophying. In Bateson’s terms, this is a failure of “the difference that makes a difference.” The system continues to process information, but it no longer knows which differences matter. And once that happens, escalation is almost inevitable—not because anyone intends it, but because the system has lost the capacity to self-correct in relation to the world it inhabits.
This is where the analogy to elite overproduction sharpens into something more structural. It is not just that there are too many credentialed actors competing for too few meaningful roles; it is that the system itself incentivizes the production of signals of competence over competence. LLMs, in their current dominant use, accelerate this dynamic. They generate fluency at scale, lowering the cost of appearing informed, capable, even insightful. But because the underlying optimization is tethered to existing corpora—past language, past assumptions—they risk reinforcing the very paradigms that are becoming obsolete. The result is a kind of epistemic inflation: more words, more analysis, more apparent sophistication, but diminishing marginal contact with ground truth.
In that sense, both the overbuilt military doctrine and the overextended use of LLMs are instances of the same failure mode: success within a closed model mistaken for success in the world. They are systems that have become extraordinarily good at refining their own internal logic while losing sensitivity to external change. And like all such systems, they do not collapse immediately. They persist, they scale, they even appear to improve—right up until the point where the environment they were built to navigate no longer exists in a form they can recognize.
If the fable of the tortoise and the hare still matters here, it’s only as a prelude. The deeper lesson is not about pace but about model validity. When the world shifts, the question is no longer who runs faster within the model, but whether the model itself is still tethered to reality.
The fable, then, is not about who wins a race. It’s about which strategy survives a changing world.
The hare loses not because it sleeps, but because it assumes the race conditions are static. The tortoise wins not because it tries harder, but because it is insensitive to variance in the environment.
In AI, we are still betting heavily on hares.
We reward speed, scale, and spectacle. We build systems that can sprint through language, generate oceans of text, and simulate understanding at remarkable velocity. But evolution has a long memory, and it rarely rewards a single strategy forever.
If there is a moral worth keeping from Aesop, it’s not “slow and steady wins the race.”
It’s this: the environment is the real opponent. And the organism—or the system—that survives is the one that remains viable when the rules change.
The question is why time has selected for this fable and not for one in which the hare wins
That’s the right question, because it flips the frame: not “what does the fable teach?” but “what kind of world needs this fable to persist?”
One answer is that fables are not selected for their accuracy—they are selected for their corrective function. The story of The Tortoise and the Hare survives precisely because it pushes against a recurring human bias: overvaluing speed, brilliance, and early advantage. In evolutionary terms, humans are status-sensitive primates. We are drawn to hares—fast talkers, quick winners, visible dominance. A story in which the hare wins would simply reinforce what is already over-selected in social cognition. It would be redundant. The tortoise story, by contrast, encodes a counter-selection pressure: it preserves, culturally, a strategy that is otherwise undervalued but crucial for long-term survival—persistence, robustness, low variance.
There’s also a deeper evolutionary logic. Natural selection does not reward the fastest organism in isolation; it rewards the one that survives across changing conditions. The hare-winning fable would only hold in a stable, predictable environment where peak performance is consistently rewarded. But human history—and biological history more broadly—is defined by volatility: droughts, shocks, regime changes, noise. Under those conditions, strategies that minimize catastrophic failure (the tortoise) often outperform those that maximize peak success (the hare). The fable persists because it encodes a risk management heuristic, not a performance maximization rule.
Culturally, you can think of it as a kind of memetic immune system. Societies generate and retain stories that compensate for their own blind spots. If a culture becomes too enamored with acceleration, disruption, and spectacle—as ours clearly has—the tortoise narrative becomes more, not less, valuable. It is a stabilizer. A hare-victory story would be like adding fuel to a system already prone to overheating.
And there’s one more, harsher possibility: the hare-winning stories do exist, but they don’t need to be told as fables because they are constantly enacted in reality. Fast winners dominate headlines, markets, and wars—until they don’t. The tortoise only needs a story because its victories are quieter, less visible, and often only legible in retrospect. The fable is a way of compressing a long time horizon into a single, graspable narrative.
So time hasn’t “chosen” the tortoise and rejected the hare. It has kept the story that we are most likely to forget—and most likely to need—when the environment starts to shift.
Terry Pratchett had a way of smuggling epistemology inside jokes, and his line about “million-to-one chances cropping up nine times out of ten” is less a gag than a diagnosis. It points to a structural blind spot: what looks vanishingly unlikely under a thin model becomes almost predictable once you account for hidden variables—timing, incentives, overconfidence, feedback loops, narrative framing itself. In the Discworld universe, improbable outcomes recur not because the universe is whimsical, but because the observer is underinformed. The joke lands because it reverses the usual assumption: it’s not that stories inflate rarity into inevitability, it’s that reality, when properly conditioned, often clusters around outcomes we initially misclassify as rare. The “million-to-one” is frequently just a badly specified distribution.
Leave a Reply