SFI's first postdoctoral fellow (1988). Chair of the Science Steering Committee — ex officio on both the Science Board and Board of Trustees, making him the external faculty member with the most structural influence over SFI's research direction. Professor of Economics and Social Science at Carnegie Mellon. Fourth-generation Colorado rancher.
Trained at Michigan (PhD Economics, 1988) under Ted Bergstrom and Hal Varian — yes, the Varian who later became Google's chief economist. But Miller's intellectual formation came from sitting next to John Holland (inventor of genetic algorithms, founder of Michigan's Center for the Study of Complex Systems) and Robert Axelrod (Evolution of Cooperation)2. Their 1991 paper launched computational economics.1
Monthly lunches at Murray Gell-Mann's house in Santa Fe with Cormac McCarthy. Auctioned McCarthy's typewriter for SFI — it raised $254,000. Co-hosted Kauffman's 80th birthday celebration at SFI with Geoffrey West. Co-founded (with Scott Page) the Graduate Workshop in Computational Social Science, now 30+ years running.
Three books. ~15,000 citations. The key thing to understand about Miller: he is not interested in making agents smarter. He is interested in how staggeringly simple agents produce staggeringly complex collective behavior. The simplest possible computational machines. The richest possible emergent structures.
Miller uses Moore machinesA finite automaton whose output depends only on its current state — not on input. Named after Edward F. Moore (1956). The simplest class of sequential logic machines. — the simplest class of finite automata. A Moore machine has:
That's it. No memory beyond the current state. No learning within a game. No model of the opponent. No reasoning. Just: look at state, produce output, read input, transition.
With 1 state, there are exactly 2 possible machines: Always Cooperate and Always Defect. That's the entire strategy space. No conditional behavior is possible.
With 2 states, after minimization there are 26 distinct machines, classified into behavioral types:
Without minimization, there are 210 = 1024 possible 2-state machines (2 outputs × 4 transitions × 2 initial states × 2 initial actions). After removing duplicates and reducing functionally identical machines, Miller identifies 26 behaviorally distinct automata. Of these, 20 are functionally 1-state (the second state is unreachable or produces the same output). Only 6 exhibit genuinely conditional social behavior — but those 6 include TFT, Grim Trigger, and other reciprocal strategies.
Of these 26 machines, only 6 exhibit genuinely conditional social behavior. But those 6 are enough. They include:
With accurate bit-flip mutation, Tit-for-Tat dominates ~64% of 2-state runs. TFT's forgiving reciprocity outperforms Grim Trigger when populations are clean — most opponents cooperate reliably, so forgiveness pays. But add noiseImplementation error: each move has a probability of being flipped before the opponent sees it. Miller (1996) tested 1% and 5% noise conditions. (misreported moves), and TFT collapses: it can't distinguish accidental defection from real betrayal. At 2 states, even 1% noise destroys cooperation entirely. At 16 states, complex strategies absorb noise better — computational capacity is a buffer against environmental uncertainty.
Two populations of 30 agents each coevolve against each other. This is crucial — not one population, but two, like biological species in an arms race. Every member of Population 0 plays against every member of Population 1 in a round-robin tournament. Fitness = cumulative payoff across all matches.
A genetic algorithmOptimization method inspired by natural selection: encode solutions as bit strings, select the fittest, recombine via crossover, introduce random mutations. Invented by John Holland (1975) — Miller's colleague at Michigan. governs evolution: tournament selection, crossover, and mutation. The automata are encoded as 148-bit strings (4 bits start state + 16 × 9-bit state packets), and standard GA operators produce offspring. Mutation rate is tunable — Miller systematically varies it to show that low mutation doesn't prevent transitions, it just makes them rarer and more abrupt.
This is the book's central finding. Social behavior requires crossing two simultaneous thresholds:
Threshold 1 — Computation (≥2 states): The agent must be able to condition its behavior on the opponent's behavior. A 1-state machine is blind — it cannot respond, reciprocate, punish, or forgive. It is computationally incapable of social behavior.
Threshold 2 — Interaction (≥2 rounds): Even a 2-state machine playing a one-shot game has no opportunity to execute conditional strategies. Tit-for-Tat needs at least one prior move to mirror. Repeated interaction is the substrate on which computational capacity operates.
Cross both thresholds simultaneously and social behavior doesn't just become possible — it becomes probable. The system doesn't drift toward cooperation. It snaps.
Running the model with Miller's two-population structure (30 agents per population, 150 rounds, mut=0.005, 200 generations, 30 seeds per state count) confirms the sharp threshold:
| States | Mean Cooperation | Character |
|---|---|---|
| 1 | 6.0% | AllD always dominates — no conditional behavior possible |
| 2 | 75.0% | Sharp jump — conditional strategies emerge via neutral drift |
| 3 | 77.5% | Richer strategy ecology, similar cooperation level |
| 4 | 77.5% | Diverse strategies, high variance across runs |
| 5 | 77.0% | No further improvement — complexity plateau |
The two-population model produces higher variance than single-population round-robin because cooperation requires coordination across independently evolving populations — the neutral drift → complementary mutation mechanism. Some runs haven't transitioned by generation 200; with 500 generations, all do.
Miller calls the abrupt explosion of strategic diversity beyond the threshold the Burgess Shale — a deliberate echo of Gould's Cambrian explosion. But Miller inverts Gould's conclusion: where Gould argued contingency ("replay the tape, get a different outcome"), Miller argues inevitability. Social behavior is a robust attractor. The conditions required are so modest — two states, two rounds — that any world with minimally complex agents and repeated interaction will produce it.
Miller's threshold is about computational capacity (states) and opportunity (rounds). The simulation reveals a second phase transitionAn abrupt qualitative change in system behavior at a critical parameter value. Like water freezing at 0°C — gradual cooling, sudden crystallization., orthogonal to the first: a mutation rate threshold.
| Mutation Rate | Cooperation | Character |
|---|---|---|
| 0.001 | 62% | Low exploration — bimodal, many runs stuck at AllD |
| 0.003 | 90% | Sweet spot — high cooperation, low variance (sd=7%) |
| 0.005 | 75% | Miller's default — good but high variance (sd=29%) |
| 0.010 | 17% | Sharp cliff — mutation overwhelms conditional logic |
| 0.020 | 15% | Cooperation effectively destroyed |
The mechanism: conditional strategies (TFT, Grim Trigger) depend on specific state transitions — "if opponent defects, go to state 1." High mutation rates corrupt these transitions faster than selection can repair them. This is analogous to Eigen's error thresholdManfred Eigen (1971): if mutation rate exceeds a critical value, selection cannot maintain genetic information. The quasispecies collapses into random sequences.3 in molecular evolution: above a critical mutation rate, the information content of the genome cannot be maintained by selection.
Cooperation requires not just the computational capacity to condition behavior, but enough genetic stability to maintain the conditioning across generations.
How does the system get from asocial to social? Not gradually. Miller identifies a six-stage transition motif that repeats across all his experiments:
The mechanism in detail:
The transition is not a steady climb toward cooperation. It's punctuated equilibrium. Long stasis, then sudden revolution. The system appears stable for hundreds of generations, but neutral variation is accumulating beneath the surface — latent diversity that becomes decisive when the right complementary mutation appears.
This is the "revolution" part: a single additional mutation in the right context triggers a cascade that flips the entire system. Evolution is gradual; the revolution it produces is abrupt.
Tracking which strategies dominate across runs reveals the ecology beneath the cooperation rate:
| Strategy | Behavior | Advantage |
|---|---|---|
| Grim TriggerCooperate until the opponent defects once, then defect permanently. Maximally retaliatory — no forgiveness. | Cooperate until betrayed, defect permanently | Crushes "wobbly" mutants |
| TFTTit-for-Tat: start cooperating, then copy the opponent's last move. | Mirror the opponent's last move | Recovers from mutual defection |
| Novel (3+ states) | Graduated punishment — one-round retaliation, then forgive | Balances forgiveness and deterrence |
In a noiseless environment with accurate mutation (guaranteed bit-flip), TFT dominates: ~64% of 2-state runs converge to TFT-dominant populations, vs ~16% for Grim. TFT's forgiving reciprocity outperforms Grim's permanent retaliation when opponents cooperate reliably. But add noise (environmental misreport of moves), and both strategies suffer — TFT retaliates against accidental "defections," triggering cascades. At 2 states there's no room for noise-tolerant strategies; at 16 states, automata evolve graduated punishment that absorbs noise gracefully.
The functional identification problem: Many agents appear "unknown" in a naive census because they're known strategies with swapped state numbering. A 2-state automaton has two equivalent encodings of the same behavior (start in state 0 or state 1, with outputs permuted). Reducing to accessible states and canonicalizing the numbering reveals the true strategy distribution — a critical step for meaningful analysis.
In the two-population model, the revolution is slower but more dramatic than in single-population variants. Cooperation at states=2 requires neutral driftGenetic variants invisible to selection accumulate randomly in the population. Builds latent diversity that becomes relevant when conditions change. in both populations, then a complementary mutation in one that exploits the other's latent diversity:
Cooperation never fully stabilizes. It oscillates as mutant defectors trigger retaliation cascades, get purged, then reappear. The "consolidation" stage is dynamic, not static.
Miller doesn't just run the Prisoner's Dilemma. He applies the identical Coevolving Automata Model (CAM) to five fundamentally different social situations, each getting its own chapter:
| Game | Chapter | Social Problem | What Emerges at Threshold |
|---|---|---|---|
| Hawk-Dove | 7: Conflict | Resource competition without total war | Turn-taking — agents alternate H/D, sharing surplus |
| Battle of the Sexes | 8: Coordination | Agreement on standards without central authority | Within-game sharing of equilibria, flexible coordination |
| Prisoner's Dilemma | 9: Cooperation | Mutual benefit when defection tempts | Reciprocity (TFT), punishment (Grim), forgiveness (Punish Once) |
| Stag Hunt | 9: Cooperation | Coordination under risk | Cooperative basin larger — less temptation, easier threshold |
| Edgeworth Exchange | 10: Commerce | Trade with asymmetric endowments | Spontaneous trade emerges from evolved strategies |
Running the coevolving automata model across different payoff structures reveals how game type constrains what evolution can discover:
| Game | Payoffs (T,R,P,S) | Cooperation | Character |
|---|---|---|---|
| Stag Hunt | 3, 4, 1, 0 | 91% | Easiest — no temptation to defect when opponent cooperates |
| Classic PD | 5, 3, 1, 0 | 69% | Baseline — temptation creates adversarial tension |
| ChickenAlso called Hawk-Dove. Unlike PD, mutual defection (P=0) is worse than being exploited (S=1). This changes the strategic landscape entirely. | 5, 3, 0, 1 | 49% | Converges to mixed equilibrium — remarkably consistent |
| Strong Temptation | 8, 3, 1, 0 | 20% | High T overwhelms cooperative incentives |
Chicken is especially revealing. When mutual defection is the worst outcome (P=0 < S=1), the automata converge to an almost exactly 50/50 mixed strategy across all trials — the game-theoretic mixed strategy Nash equilibriumWhen no pure strategy is optimal, players randomize. In Chicken, the equilibrium has each player choosing Cooperate with probability (T-P)/(T-S+R-P).. The GA doesn't "know" game theory — it discovers the equilibrium through selection pressure alone.
Chapter 11 covers Communication — agents evolve the ability to send and receive signals, and communication itself enables cooperation even in one-shot games (extending Miller, Butts & Rode 2002).
The same computational threshold — ≥2 states, ≥2 rounds — produces qualitative behavioral transitions across all five game types. This is not a result about the Prisoner's Dilemma. It is a result about social behavior itself. Conflict, coordination, cooperation, commerce, and communication all emerge from the same computational substrate under the same minimal conditions.
The capstone is Chapter 12: The Eightfold Path. Miller identifies eight behavioral archetypes that emerge reliably once the threshold is crossed, forming a taxonomy of minimal social strategies. He presents this as the answer to his opening question: social behavior would arise again if we rewound the tape.
This is Miller's popular-science book and his most accessible statement of philosophy. The core argument:
Systems analyzed: stock markets, airline scheduling, Balinese rice terraces, beehive thermoregulation, heartbeat patterns, racial segregation (Schelling), bacterial colonies, nuclear chain reactions.
Nature called it "a succinct, elegant study of systems thinking." W. Brian Arthur said Miller made "complexity seem simple."
The textbook. ~4,200 citations — Miller's most-cited work. Four sections:
The pedagogical principle: keep the model simple. Two or three parameters can reveal substantial complexity. This is a methodological commitment, not laziness — it's what makes thresholds visible.
| Paper | Year | Citations | Significance |
|---|---|---|---|
| Andreoni & Miller, "Giving According to GARP" (Econometrica) | 2002 | 2,878 | Altruism is rational — 173/176 subjects' giving behavior satisfied revealed-preference axioms |
| Holland & Miller, "Artificial Adaptive Agents in Economic Theory" (AER) | 1991 | 1,333 | Founded computational economics. Proposed classifier systems as economic agents. |
| Andreoni & Miller, "Rational Cooperation in the Finitely Repeated PD" (Econ. J.) | 1993 | 1,195 | Discovered "homemade altruism" — subjects bring natural cooperation tendency from outside the game |
| Miller, "Coevolution of Automata in the Repeated PD" (JEBO) | 1996 | 536 | Noise destroys TFT's advantage. Ecology of strategies matters more than individual optimality. |
| Kollman, Miller & Page, "Adaptive Parties in Spatial Elections" (APSR) | 1992 | 482 | Political parties as adaptive agents in spatial voting models |
| Miller, "Active Nonlinear Tests (ANTs)" (Mgmt. Sci.) | 1998 | 260 | GA-based methodology to break simulation models — probe for nonlinear fragilities |
| Miller & Page, "The Standing Ovation Problem" (Complexity) | 2004 | — | Canonical micro→macro cascade model. Simple rules, spatial context, collective action. |
| Jung, Miller & Page, "ABM for Psychological Research" (Am. Psychologist) | 2025 | new | Introduces "minimal society paradigm" and "principle-oriented modeling" |
Active Nonlinear Tests deserve special attention. ANTs use genetic algorithms to actively probe simulation models for parameter combinations that break their predictions. Miller demonstrated this on the World3 model of global dynamics, finding small but powerful nonlinear effects invisible to standard sensitivity analysis. The methodology shares intellectual DNA with early warning signals — probing for fragility near critical thresholds — though it's a model-validation tool rather than a prediction tool.
Jung, Miller & Page (2025) introduces two concepts worth knowing:
The Kauffman connection: Institutional, not co-authorial. Miller co-hosted Kauffman's 80th birthday at SFI. They inhabit the same intellectual ecosystem. But Miller has not published on autocatalytic sets or RAF theory — and he has not explicitly mapped his computational thresholds to Kauffman's catalytic closure. That mapping is available.
Miller's entire framework is built on a specific assumption: agents are computationally minimal. Moore machines with 1-5 states, binary I/O, no learning within a game, no model of the opponent. This is deliberate — it's how he isolates the threshold.
What happens when you replace these machines with LLMs?
Miller's central finding is that social behavior requires ≥2 states and ≥2 rounds. An LLM has effectively infinite states and can condition on the entire interaction history. The threshold is crossed before the experiment begins. This is like asking "does water boil above 100°C?" when your starting temperature is 10,000°C.
So the question isn't whether LLM agents cross Miller's threshold. They're born past it. The question is: what new thresholds emerge when computational capacity is no longer the bottleneck?
If Miller's framework survives extension (and it might — the structural logic is sound even if the specific parameters change), the new thresholds might look like:
Miller's automata have no goals, no preferences, no values — just state transitions. LLMs have training objectives, system prompts, and RLHF-shaped behavioral policies. When two LLMs interact, cooperation isn't just a strategy — it's entangled with their training. The threshold question becomes: at what point does the training objective interfere with emergent social strategy? An LLM trained to be "helpful" might cooperate even when defection is rational — not because of evolved reciprocity but because of alignment pressure. This contaminates the experiment.
Miller's automata communicate through actions only — binary choices observed after the fact. LLMs communicate through language. They can negotiate, promise, threaten, deceive, and reason about reasoning. Miller's Chapter 11 (Communication) shows that even giving 2-state automata a signaling channel transforms the dynamics. With LLMs, the signaling channel is the primary interaction mode. Communication doesn't emerge at a threshold — it's the starting condition.
This inverts Miller's structure. For automata: computation → interaction → communication (each a threshold to cross). For LLMs: communication is given; the question is what constrains it. Context window limits, system prompt restrictions, inability to verify claims — these become the new bottlenecks.
Miller can read the state table of every automaton. He can classify all 26 two-state machines. He knows exactly what strategy each agent is running. With LLMs, strategies are opaque. You observe behavior but cannot read the "state table." The strategy space is continuous, not discrete. You can't enumerate it.
This matters because Miller's analysis depends on identifying specific strategies (TFT, Grim Trigger, Win Stay Lose Shift) and tracking their evolutionary dynamics. With LLMs, you might observe TFT-like behavior without being able to confirm it's TFT — the agent might be running something more complex that looks like TFT under current conditions but would diverge under different ones.
Miller uses 40 agents per population. With LLMs, the "population" might be 3-5 models from different providers (Claude, GPT, Gemini, Llama), or it might be millions of instances of the same model with different system prompts. The evolutionary mechanism changes completely. LLMs don't reproduce through crossover and mutation — they adapt through fine-tuning, RLHF, and in-context learning. The "neutral drift" that enables Miller's revolutions has no obvious analog in LLM development... or does it?
Miller's neutral drift works because a mutant ν plays identically to α against current opponent β, accumulating without selection pressure. Then μ appears and the hidden difference becomes decisive.
LLMs have something structurally similar: capability overhang. A model is trained to do X. It can also do Y, but no one has asked it to. The capability accumulates without selective pressure. Then a new prompt, a new fine-tune, a new interaction context reveals the latent capability — and the system's behavior shifts abruptly.
This is speculative. But it maps onto Miller's transition motif with surprising precision: capabilities accumulate neutrally → new context reveals them → rapid behavioral transition.
Miller's automata are homogeneous — all Moore machines, same architecture, different state tables. LLMs vary in architecture, training data, parameter count, and alignment approach. Claude and GPT are not the same kind of machine with different state tables — they're different kinds of machines. This introduces a heterogeneity that Miller's framework doesn't address. The "coevolution of two populations" becomes "coevolution of fundamentally different architectures."
Despite these differences, core elements of Miller's framework may be robust:
| Dimension | Kauffman | Miller | Stephenson |
|---|---|---|---|
| Closure type | Catalytic (RAF) | Computational (automata) | Institutional (governance) |
| Threshold | Molecular diversity | State count + interaction rounds | Structural conditions (TBD) |
| Transition character | Abrupt (percolation) | Abrupt (punctuated equilibrium) | Abrupt (hypothesized) |
| Self-sustaining? | Yes — autocatalytic set maintains itself | Yes — social strategies persist evolutionarily | Yes — self-sustaining governance (the goal) |
| Designed? | No — emerges from conditions | No — "revolutions by evolution" | No — emerges from structure |
The comparison above is a claim about mathematical structure, not about analogy. If it holds, three questions follow directly, and none of them has been answered:
Miller's result is that cooperation appears once agents cross a minimal computational threshold — two states, two rounds, no design required. Whether that threshold structure persists when the agents are far richer than Moore machines is an open empirical question. A parameter sweep over agent capacity, interaction history length and population diversity would answer it.
Miller's “Eightfold Path” identifies eight behavioral archetypes that emerge at the threshold. Do these strategies form a reflexively autocatalytic, food-generated set — does each strategy's viability depend on the presence of others? Note the caveat that applies to all such work: binary RAF existence is generic at moderate catalysis density, so demonstrating that the set “is an RAF” would establish very little on its own. The informative quantity is the fraction of the reaction set that is non-zero under perturbation, not the yes/no existence result.
Active Nonlinear Tests use search to probe a model for parameter combinations that break its predictions. Applying the same technique to a structural model asks: which parameter combinations break a self-sustaining structure? This is a methodological transfer, and it is testable.
Run Miller's coevolving automata model in your browser. The engine implements the core algorithms from Ex Machina: Moore machine automata, two-population coevolution with cross-population tournaments, genetic algorithm evolution with ranked tournament selection, circular crossover, and per-gene mutation calibrated to Miller's 0.74 expected mutations per offspring. (Miller uses fitness-proportional selection with z-scoring; we use tournament selection — see gap analysis.)
Two modes:
Books:
Key Papers:
Code: github.com/SantaFeInstitute/ExMachina (visualization scripts; simulation engine is compiled JAR)
SFI Profile: santafe.edu/people/profile/john-h-miller