# Rock, Paper, Scissors
![[resources/images/article-rock-paper-scissors.jpg]]
<iframe width="100%" height="20" scrolling="no" frameborder="no" allow="autoplay; encrypted-media" src="https://w.soundcloud.com/player/?url=https%3A//api.soundcloud.com/tracks/soundcloud%253Atracks%253A2440736001&color=%23ff5500&inverse=false&auto_play=false&show_user=true"></iframe><div style="font-size: 10px; color: #cccccc;line-break: anywhere;word-break: normal;overflow: hidden;white-space: nowrap;text-overflow: ellipsis; font-family: Interstate,Lucida Grande,Lucida Sans Unicode,Lucida Sans,Garuda,Verdana,Tahoma,sans-serif;font-weight: 100;"><a href="https://soundcloud.com/bryantmcgill" title="Bryant McGill" target="_blank" style="color: #cccccc; text-decoration: none;">Bryant McGill</a> · <a href="https://soundcloud.com/bryantmcgill/rock-paper-scissors" title="Rock, Paper, Scissors" target="_blank" style="color: #cccccc; text-decoration: none;">Rock, Paper, Scissors</a></div>
Imagine for a moment that life on Earth is an extraordinarily sophisticated simulation, and that every person you meet belongs to a vast learning environment designed to develop intelligence through experience. Every friendship, rivalry, romance, betrayal, unexpected kindness and apparently accidental meeting introduces information that changes how you perceive the world and how you respond to it. Some encounters reinforce what you already believe; others expose a weakness, overturn an assumption, or force you to build an entirely new way of thinking. Now imagine the same thing happening to everyone else at once, with every participant shaping the development of countless others. **Ordinary human life begins to resemble{Ordinary human life, seen as a field of encounters, resembles{ a planetary-scale reinforcement learning system, in which billions of intelligences are continuously training one another.}}**
The thought experiment is worth holding, and it is also worth putting down for a minute, because the shortest way into it is a game every child already knows. Rock breaks scissors, scissors cut paper, and paper covers rock. Nothing in that sentence names a winner. Each move beats one move and loses to another, so the value of any choice depends entirely on what someone else chooses. That small fact, played forward across a few rounds, a few generations, and eventually a few billion people, turns out to carry most of what we presently understand about how minds, populations and civilizations learn.
## I. The game with no best move
Rock, Paper, Scissors is the purest example of what mathematicians call a **non-transitive game**. In a transitive contest, strength stacks like a ladder: if A beats B and B beats C, then A beats C, and somebody eventually sits at the top. Rock, Paper, Scissors refuses that ladder and bends it into a circle. There is no strongest move to memorize and no champion strategy waiting to be discovered, and the only equilibrium a textbook can offer is to play each move a third of the time, unpredictably, forever. {{In Rock, Paper, Scissors the only strategy that cannot be exploited is unpredictability, which makes a child's game a small demonstration that freedom has strategic value.}} The game's whole difficulty lives in the other player, which is exactly why it is such a precise instrument for studying minds.
<iframe width="560" height="315" src="https://www.youtube.com/embed/iwVZO3jCBGw?si=YFQFUCgdCsq3cKHH" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>
Repeated play converts the game from a choice into a relationship. On the first throw you are guessing; by the tenth you are remembering, watching tendencies, anticipating, and adjusting, while your opponent does the same to you. Zhijian Wang, Bin Xu and Hai-Jun Zhou watched this happen in the laboratory, publishing in *Scientific Reports* in 2014 the results of groups of students who each played three hundred rounds of the game against one another. The players departed from the textbook equilibrium in a consistent way: winners tended to repeat the winning move, losers tended to shift to the next move around the circle, a conditional pattern close to the **win-stay, lose-shift** rule long studied in the Prisoner's Dilemma. More striking, the group as a whole began to cycle, its collective state rotating around the triangle of rock, paper and scissors, and a simple model of those individual conditional responses reproduced the cycling without any fitted parameters. Two people learning about each other had become a population with a rhythm of its own.
Every repeated game against an adaptive opponent is therefore{Every repeated game against an adaptive opponent is{ a contest between two models of mind}, each of them part of the environment the other is learning from.}. You are learning about another intelligence that is learning about yours, and each of you is, at every moment, part of the environment the other is learning from. What began as three gestures has become an evolving relationship between two adaptive systems, and the moment there are more than two players, it becomes an ecology.
## II. Nature plays it for keeps
Living systems were playing Rock, Paper, Scissors long before anyone named it. In 1996 Barry Sinervo and Curt Lively reported in *Nature* that male side-blotched lizards, *Uta stansburiana*, come in three throat colours tied to three mating strategies. Orange-throated males are aggressive and hold large territories; blue-throated males guard a smaller territory and their mates closely; yellow-throated males hold no territory and sneak matings. Orange beats blue by overpowering it, yellow beats orange by slipping into its sprawling territory unnoticed, and blue beats yellow by guarding too closely to be cheated. None of the three can take over the population, because the success{In a contest with no best strategy, the success{ of each one creates the conditions for its own defeat}.}, and so the frequencies of the three morphs cycle across the years.
Bacteria play the same game with chemistry. In 2002 Benjamin Kerr and colleagues reported, again in *Nature*, an experiment with three strains of *Escherichia coli*: one that produces a toxin called colicin, one that is sensitive to it, and one that resists it. The producer kills the sensitive strain, the sensitive strain outgrows the resistant one because resistance carries a cost, and the resistant strain outgrows the producer because making toxin carries a larger one. On a static agar plate, where neighbours interact only with neighbours, all three strains persisted together. In a well-mixed flask, where every cell meets every other, the diversity collapsed and a single strain took over. The same three organisms, under the same rules, either coexisted or reduced to one depending on how far each interaction reached.
Physicists then found the threshold. Tobias Reichenbach, Mauro Mobilia and Erwin Frey showed in *Nature* in 2007 that in a spatial Rock, Paper, Scissors model, species coexist in entangled spiral waves as long as individual mobility stays below a critical value, and lose that coexistence once mobility exceeds it. Local, cyclic competition preserves variety; competition mixed so thoroughly that every player faces the whole field at once tends toward a single survivor. This is the first lesson the game teaches at population scale, and it is a lesson about diversity itself: **a world{In competition, a world{ with no best move keeps its players distinct, and a world flattened into one arena tends to keep only one.}}** [[reminders/Evolution/A World With No Best Move Keeps Its Players Distinct by Bryant McGill|∴]]
I wrote years ago in *Simple Reminders* that "all monoculture is inherently vulnerable; from plants in gardens to people in urban centers." The lizards, the bacteria and the spiral waves give that observation a mechanism. {{Monoculture is what competition produces when it becomes transitive, when one strategy can be ranked above all others and the field converges on it.}} Non-transitivity{Non-transitivity, the condition in which no move beats every other,{ is the structural guarantee of distinguishability, the arrangement under which no single way of being can absorb every other.}} {{A person who refuses to be ranked on a single ladder, who keeps a genuinely different move available, is preserving something the whole population needs, whether or not the population notices.}}
## III. Reinforcement learning was never only a computer science invention
Reinforcement learning is usually introduced as something computer scientists invented to train artificial intelligence, yet its central ideas describe processes life was running long before any computer existed. Organisms adjust behaviour according to consequences; in 1997 Wolfram Schultz, Peter Dayan and Read Montague showed in *Science* that dopamine neurons in the primate brain signal something very close to a **reward prediction error**, the very quantity temporal-difference algorithms use to learn. In 2018 Jane Wang and colleagues at DeepMind argued in *Nature Neuroscience* that the prefrontal cortex behaves as a **meta-reinforcement learning** system, a slow learner that trains a fast one, which is to say that the brain appears to learn how to learn. Artificial intelligence gave us precise mathematics for these processes; it did not give them their origin.{Artificial intelligence gave us precise mathematics for learning; it did not give learning its origin.{ We did not merely build machines that learn; we developed increasingly exact descriptions of how living systems already learn, compete and cooperate.}}
Evolutionary biology arrived at the same machinery from the other direction. Olof Leimar, emeritus professor of zoology at Stockholm University, published a model in *The American Naturalist* in 2021 in which animals form dominance hierarchies by adjusting their aggressive and submissive behaviour toward particular rivals through reinforcement learning, with rewards and penalties set by relative fighting ability. The parameters of the learning rule itself (how quickly an animal updates, how much it learns from watching others fight, how much it generalizes what it learned about one opponent to the next, and how aggressive it starts out) are heritable traits under selection. Evolution in his simulations favours high social competence: individual recognition, attention to bystander outcomes, and only modest generalization across opponents, a combination that reproduces the stronger loser effects than winner effects seen in real animals. The social environment{Among social animals, the social environment{ supplies the reward, the learning rule updates the behaviour, and natural selection tunes the learning rule.}} The whole loop is endogenous; nothing outside the population has to arrange it.
Mathematics confirms that the two descriptions are, in important cases, one description. Tilman Börgers and Rajiv Sarin showed in 1997 that a simple reinforcement learner, Cross learning, converges in the continuous limit to the **replicator dynamics** of evolutionary game theory, the same equation biologists use to describe the spread of strategies through reproducing populations. Daan Bloembergen, Karl Tuyls and their collaborators surveyed the wider correspondence in the *Journal of Artificial Intelligence Research* in 2015, showing how the trajectories of common learning algorithms can be analysed with the tools of evolutionary game theory, fixed points, stability and all. {{Learning within a life and selection across lives turn out to trace the same curves through strategy space; one runs on synapses and experience, the other on genomes and generations, and in many games they arrive at the same places.}}
A population of learners can therefore be studied as an evolving population, and recent work has done so at remarkable scale. Yann Bouteiller, Karthik Soma and Giovanni Beltrame simulated two hundred thousand individually learning agents in three classic games, the Stag Hunt, Hawk–Dove and Rock, Paper, Scissors, and compared ordinary policy-gradient learners with learners using **opponent-learning awareness**, a method that anticipates how the other player's learning will respond to one's own. The more sophisticated learners changed the population's fate in each game: they promoted cooperation in the Stag Hunt, delayed cooperative outcomes in Hawk–Dove, and in Rock, Paper, Scissors they reduced strategy diversity. That last result deserves to be held in the mind for the rest of this essay. The smarter the players became about modelling one another, the more{In a game with no best move, the smarter learners become about modelling one another, the more{ the circle of possibilities} can narrow toward sameness.} tended to narrow.
## IV. A crowd can learn as one
Collective intelligence becomes something more than a metaphor when the mathematics of a crowd turns out to equal the mathematics of a single learner. On 17 October 2026, Karthik Soma, Yann Bouteiller, Heiko Hamann and Giovanni Beltrame published "The hive mind is a single reinforcement learning agent" in *Swarm Intelligence*. Working from the well-established model of honeybee swarms choosing a new nest site, in which scouts recruit others to the sites they have found and individual bees imitate what they observe, the authors show that the colony as a whole behaves as one coherent reinforcement learner, a bandit algorithm they name **Maynard–Cross learning** after John Maynard Smith and John Cross. Each bee follows a local imitation rule; the swarm, aggregated, is learning from many parallel experiences at once. The authors suggest that this may be how such imitation strategies evolved, as a scalable form of reinforcement learning at the level of the group.
The result is formally specific to the decision model it analyses, and it would be careless to declare that human societies have been proven to be single agents. One can nonetheless see how naturally it extends. If simple imitators can add up to a learner one level higher, then{When simple imitators can add up to a learner one level higher,{ the question of whether intelligence lives in the individual or in the group} becomes a question of scale.} dissolves into a question of scale: individuals learn from encounters, groups learn from their members, and the same formal structure may recur at each level. The idea of a planetary training ecology, which began this essay as a speculative image, acquires a mathematical precedent, and the precedent requires no central coordinator whatever.
Human social learning shows the same architecture from the inside. Dual-inheritance theory, developed by Robert Boyd and Peter Richerson from the 1980s onward, treats culture as a second system of inheritance whose learning strategies evolved alongside genes; their models give a central place to biases such as copying the majority and copying the successful. In 2025 David Schultner, Lucas Molleman and Björn Lindström reported in *Nature Human Behaviour* that across six experiments with 1,941 participants, people learn which social cues to trust by the ordinary mechanism of reward: they associate features such as what others choose and how well others fare with the rewards those features have delivered, and they reweight them as experience accumulates. In agent-based simulations, the familiar strategies of conformity and success-copying emerged from that feature learning without being built in. Approval,{Approval,{ imitation, prestige and the herd, so often treated as fixed instincts,} may be the outputs of reinforcement learning running on social information.} are strongly suggested by this research to be the outputs of reinforcement learning running on social information.
## V. Why machines had to train a league
Machine learning discovered the practical consequences of non-transitivity the hard way. An agent trained only against copies of itself in a game with cycles can spend its training chasing its own tail, beating yesterday's version and losing to the version before that, improving at nothing in particular. Wojciech Czarnecki and colleagues at DeepMind examined this structure in "Real World Games Look Like Spinning Tops," presented at NeurIPS in 2020. Across real games from Tic-Tac-Toe to Go and StarCraft II, they found strategy spaces shaped like a spinning top, with a transitive axis of genuine skill running vertically and wide bands of non-transitive cycles spreading out around the middle. Their analysis explains why training a strong agent requires a **population of strategies** rather than a single lineage, and how the size of that population relates to the shape of the game.
DeepMind's AlphaStar, which reached Grandmaster level in StarCraft II and was described in *Nature* in 2019, was built on exactly this principle. Its developers trained a league: main agents learning to win, alongside exploiter agents whose only job was to find and punish the main agents' weaknesses, so that every blind spot became a training signal. The league is a deliberate reconstruction of what the lizards and the bacteria do without instruction, a standing population of rivals whose diversity keeps every member honest. The lesson for any intelligence, artificial or biological, is that competence{Competence{ in a world of cycles is grown inside a community of opponents}; a mind trained against only one kind of challenger learns only that challenger.} [[reminders/Competence/A Mind Trained Against One Challenger Learns Only That Challenger by Bryant McGill|∴]], and a mind trained against only one kind of challenger learns only that challenger.
Joel Leibo, Edward Hughes, Marc Lanctot and Thore Graepel gave the larger pattern a name in a 2019 manifesto for multi-agent research: the **autocurriculum**. {{When agents compete and cooperate, every solution one agent finds becomes a new problem for the others, so the challenges escalate on their own, and no designer has to write the next lesson.}} The authors identified what they called the "problem problem," the difficulty of generating ever-new environments rich enough to keep learners learning, and proposed that social interaction solves it from within. The kindness after hostility, the betrayal after trust, the rival who appears precisely when one has grown complacent: what the simulation hypothesis imagines as a curriculum arranged from outside, the autocurriculum literature describes as the curriculum a population of learners generates for itself.
Multi-agent reinforcement learning can even find social strategies that humans and theorists had missed. Qi Su, Hongyu Wang, Yu Xia and Long Wang reported in *Nature Communications* in December 2025 a framework for searching the vast space of strategies in iterated and evolutionary games, and it recovered a previously unrecognized policy they call **memory-two bilateral reciprocity**, which conditions on the last two rounds of both players' choices. In pairwise play it consistently outperformed a wide range of established strategies, and in evolving populations, well mixed or structured, it dominated while raising cooperation and collective welfare. Experience-driven adaptation alone discovered a form of reciprocity more robust than the one-step reciprocators evolutionary biologists studied for decades. One can imagine, though the paper itself makes no such claim, that richer contingencies of this kind only become worth learning once a population has stabilized simpler ones long enough for its shared history to carry information; if so, each layer of social sophistication would build the ground on which the next one stands.
Edward Hughes and colleagues at Google DeepMind made the stakes explicit in a position paper presented at ICML in 2024, arguing that **open-endedness is essential for artificial superhuman intelligence**: a system must keep producing artifacts that are both novel and learnable from the standpoint of an observer, rather than converging on a fixed objective. The trajectory of the field has thus come full circle to the childhood game. The route the leading laboratories now describe toward the most capable machine intelligence runs through{The route the leading AI laboratories now describe toward superhuman machine intelligence runs through{ populations, rivals, self-generated curricula and the refusal to stop at a single equilibrium}, which are the lessons of a childhood game with no best move.}, which is to say through the principle Rock, Paper, Scissors states in its first sentence.
## VI. The players begin changing the game
Games become far more interesting once the players can alter the rules, and the formal study of that move is younger than the game theory it extends. Seth Frey and Curtis Atkisson published "A dynamic over games drives selfish agents to win–win outcomes" in *Proceedings of the Royal Society B* in 2020. Their agents do not merely choose strategies inside a fixed game; they compare neighbouring games according to institutional preferences for features such as stability and efficiency, move to the more preferable one, and trace trajectories through the space of possible games until they settle into attractor games. In one illustrative path, a Prisoner's Dilemma drifts into a win–win game. The striking finding is that the attractors are disproportionately fair, even though the agents are strictly self-interested and indifferent to fairness; their earlier preprint reported fairness overrepresented by about a hundred percent relative to baseline. {{Players who can edit the game they live in can arrive, through selfish navigation alone, at better games than the one they were handed.}} [[reminders/Civilization/Selfish Players Who Can Edit the Game Find Better Games by Bryant McGill|∴]]
Institutions can also emerge from the interaction of agents with an environment their own actions keep changing. Kenji Itao and Kunihiko Kaneko published "Self-organized institutions in evolutionary dynamical-systems games" in PNAS on 15 April 2025, modelling players who harvest a shared resource whose availability depends on what they harvest, so that the payoff matrix itself changes as the game is played. As the players' decision functions evolve, they learn to detect selfish harvesting and punish it by deliberately overharvesting, and norms self-organize that classify actions as cooperative, defective or punitive. The categories that classical game theory takes as given turn out, in this model, to be outputs of the process. {{Cooperation and defection are not written on the board before play begins; the players and their world write them together, and the players can therefore later rewrite them.}}
Sovereignty over a game belongs to whoever can change its rules, and I traced the same principle through the apiary in [[articles/Cyber-Physical Lessons from Imperial Beekeeping|Cyber-Physical Lessons from Imperial Beekeeping]]: the queen is not the sovereign, because sovereignty sits at the level that alters the enclosure. [[articles/gamification/This Is a Game Planet|This Is a Game Planet]] described governance as the art of setting parameters so that a desired behaviour becomes the equilibrium. The [[collections/Gamification|Gamification Collection]] follows that question across designed arenas, scoring systems and playable institutions; the institutional-evolution research here adds that the players themselves can occupy the level where the game is changed. A child first learns which behaviours earn approval or exclusion, later learns to influence which behaviours other people reward, and eventually discovers that groups can institutionalize those preferences in norms, laws, markets, schools and technologies. {{**The learner becomes an architect of learning environments**, and those environments go on producing new learners who may one day rebuild the architecture.}}
## VII. Rules about changing the rules
The ladder has a further rung, and it was named long before the current mathematics arrived. Elinor Ostrom, whose work on the governance of shared resources earned the Nobel Memorial Prize in Economic Sciences in 2009, distinguished three levels of rules in *Understanding Institutional Diversity* (2005). **Operational rules** govern daily action, such as who may draw water and when. **Collective-choice rules** govern how operational rules are made. **Constitutional-choice rules** govern how collective-choice rules are made, which is to say they are rules about the making of rules about the rules. Every durable self-governing community Ostrom studied had found some way to work at more than one of these levels at once.
Peter Suber built the same idea into a game. Nomic, which Douglas Hofstadter introduced in his "Metamagical Themas" column in *Scientific American* in June 1982, opens with a small set of rules that do little more than regulate how rules may be changed, and Suber's own description is the whole design in a sentence: "Nomic is a game in which changing the rules is a move." He made the initial winning condition deliberately dull so that players would hurry to amend it, and he built the game to demonstrate the argument of his book *The Paradox of Self-Amendment*, in which a constitution's amending clause can apply to itself and authorize its own revision. Nomic is Rock, Paper, Scissors turned inside out: instead of three fixed moves and unchangeable rules, it begins with almost no content and makes the rules themselves the field of play.
The escalation the opening of this essay promised can now be stated plainly. The first rung is learning which move to make; the second is learning how another mind chooses its moves; the third is changing the game; the fourth is changing the process by which games are changed. Philosophers, legal theorists and political economists described the fourth rung decades ago. What is new is that every rung now has formal mathematics attached to it, from reinforcement learning and replicator dynamics at the bottom, through hive-scale aggregation and institutional navigation in the middle, to models of open-ended cultural evolution at the top, and that machine learning has begun to operate on these rungs deliberately. {{When a system can examine and revise its own methods of learning and invention, **the question changes from how intelligence learns to win to how intelligence learns to keep expanding the possibilities of learning.**}}
## VIII. Open-endedness is rare, and it has to be kept
Open-ended growth is the property every civilization claims and few can demonstrate, and the most careful recent model suggests why. James Winters and Mathieu Charbonneau published "Modelling the emergence of open-ended cultural evolution" in *Philosophical Transactions of the Royal Society B* in 2026, as part of a theme issue on the evolution of collective intelligence. Their model allows a cultural system and the space of problems, needs and goals that system confronts to coevolve. Sustained open-ended growth appeared only under particular conditions: the cultural system and its search space had to coevolve, stochastic factors had to be strong enough to keep the dynamics away from equilibrium, and selection-like factors had to be strong enough to keep the system effective. Open-endedness, in their results, is rare and depends on history. Adaptation alone does not guarantee it;{Adaptation alone does not guarantee open-ended growth;{ a culture can adapt its way into a dead end.}}
Error, in other words, is part of the engine. Robert Ford compresses the point in *Westworld*: “Evolution forged the entirety of sentient life on this planet using only one tool: the mistake.” [[reminders/Westworld/Mistakes Built Every Sentient Life by Robert Ford|∴]] Mutation is copying error; exploration is deliberate departure from the best-known move; invention begins as an anomaly that someone declined to suppress. {{A system that eliminates all variation in the name of efficiency loses its ability to find better ways of living when the world changes; a system that keeps every variation without choosing among them never learns which ones worked.}} [[reminders/Adaptability/Efficiency Without Variation Cannot Find a Better Way by Bryant McGill|∴]] The Winters and Charbonneau result puts the balance in quantitative form: enough randomness to keep the system moving, enough selection to keep what works.
The warning hidden in the population studies now comes into focus. Bouteiller and his colleagues found that opponent-aware learners reduced diversity in Rock, Paper, Scissors, and Kerr found that a well-mixed flask reduced three strains of bacteria to one. {{Sophistication and connectivity, the two forces modern civilization multiplies fastest, can each flatten a cyclic world toward sameness when they operate without counterweight.}} One can imagine a society so thoroughly connected, and so skilled at anticipating itself, that it settles into an exquisitely managed equilibrium, stable and efficient and closed. **{{A perfectly managed equilibrium is a finished game.}}** [[reminders/Evolution/A Perfectly Managed Equilibrium Is a Finished Game by Bryant McGill|∴]] The work{The work{ of keeping a civilization open is the work of protecting variation, tolerating useful error, preserving distinct players, and keeping the means of revising the rules available to the people inside them}.}, a theme I developed through the ecology of ideas in [[articles/Ideas as Lindy Organisms|Ideas as Lindy Organisms]], where filtration reveals fitness and suppression teaches nothing.
## IX. The simulation question, returned
The simulation hypothesis comes back at this point in a more sophisticated form. Nick Bostrom's 2003 argument in *The Philosophical Quarterly* made the possibility of a simulated world a serious subject, and the image with which this essay began extends it: an orchestrating intelligence arranging human encounters as a curriculum, choosing kindness after hostility and uncertainty after stability so that memorized habits are separated from competence that survives change. Such an orchestrator would also have reason to remain hidden, because learners who could inspect the curriculum would begin optimizing{Learners who can inspect their curriculum begin optimizing{ for the detection of the curriculum rather than for the world it presents}.}, the same failure machine learning calls reward hacking. That hypothesis is coherent, and it cannot be refuted from inside.
Many of the same developmental dynamics, however, arise without any orchestrator at all. The autocurriculum generates its own escalating challenges; Leimar's animals tune their own learning rules through selection; the honeybee swarm aggregates into a learner one level higher; Frey and Atkisson's agents navigate to better games; Itao and Kaneko's players invent the categories of their own morality. Partial observability, the opacity the orchestrator would need, is a standing condition of every finite agent interacting with other finite agents. {{From inside, a curriculum arranged by a designer and a curriculum generated by the learners themselves present the same face: rivals, reversals and lessons arriving just as each learner becomes ready for them.}} A designed simulation and an evolved civilization can therefore share important computational principles without either one serving as evidence for the other, and the interesting result is the shared principle: **intelligence{Intelligence{ can be generated by the recursive interaction of learners with environments they jointly produce, whoever or whatever started the process.}}**
The competencies such a process cultivates are, notably, the portable ones. {{Models of other minds, of trust and defection, of identity under partial observation, and of causes whose effects arrive late and through many hands remain useful in any substrate where agents meet agents.}} If those competencies could be preserved across biological, synthetic or other computational substrates, a civilization could carry its accumulated intelligence beyond the material conditions in which it first arose. Nothing in these models makes that transition automatic, but they make it legible as an engineering objective: preserving a system's capacity to learn, reorganize and generate novelty when its original environment can no longer sustain it. [[articles/Mind Uploading and AI — The Host is Reusable and the Person is the Delta|Mind Uploading and AI]] and [[articles/We Were Never Going to Make It|We Were Never Going to Make It]] take up what such a crossing would require; the open-endedness research adds that what must cross{What a civilization must carry into any new substrate{ is less a stored body of knowledge than a working capacity to keep inventing}.}, and the sooner that capacity is understood well enough to carry, the better prepared any successor will be.
## X. The person across from you
Each{Each{ of us is the environment someone else is learning from}.} [[reminders/Agency/Every Person Is the Environment Someone Else Learns From by Bryant McGill|∴]], and that is the most personal consequence of everything above. The friend who reinforces your best habits, the rival who exposes your blind spot, the stranger who forces you to explore, the person who betrays a trust and teaches you which of your values were real: every one of them is supplying training signal, and you are supplying it to them in return. Your conduct is part of someone else's curriculum whether or not you intend it, and the moves you make become part of the field in which other people's strategies are selected. {{A kindness offered where hostility was expected introduces a new move into the population; everyone who sees it learns that hostility was never the only way to play, and every move that has been seen can be copied.}} [[reminders/Altruism/Unexpected Kindness Teaches Everyone a New Way to Play by Bryant McGill|∴]]
Most people, most of the time, play the moves they were trained to play. Ford describes the condition without sentiment: “We live in loops as tight and as closed as the hosts do, seldom questioning our choices, content for the most part to be told what to do next.” [[reminders/Westworld/Humans Live in Loops as Tight as Hosts by Robert Ford|∴]] The loop is the transitive collapse at the scale of a single life,{A life lived in a loop is{ one strategy repeated until it becomes identity.}} The way out is the same one the lizards and the league and the Nomic players found:{The way out of any loop is to{ keep a genuinely different move available, pay attention to what the other players are learning, and remember that the rules themselves were made by people and can be revised by people.}} Ford's hope for his creations was that “someone was paying attention. Someone who could change.” [[reminders/Westworld/Someone Was Paying Attention and Could Change by Robert Ford|∴]] That is a description of a learner who has noticed the game, and every person is free to become one.
Rock, Paper, Scissors looks different at the end of this journey than it did at the beginning. Rock still breaks scissors, scissors still cut paper, and paper still covers rock, but intelligence eventually discovers that it can invent new moves, rewrite the rules, construct new games, and teach its successors to do the same. {{The most advanced intelligence may be the one that learns how to ensure that learning, invention and transformation never have to end.}} What began as three gestures made by human hands becomes a way of understanding the continuity of intelligence across generations, civilizations, and perhaps substrates. **The final achievement{The final achievement of intelligence{ is keeping existence open to possibilities no previous intelligence could have imagined}.}, and it begins with the realization that the person{The person{ across from you is learning from you while you are learning from them, and together you are changing what happens next.}}** [[reminders/Cybernetics/The Person Across From You Is Learning From You by Bryant McGill|∴]]
---
[[about/About Bryant McGill|Bryant McGill]] is a Wall Street Journal and USA Today bestselling author, systems architect, technologist, and strategic advisor, as well as a Congressionally Recognized Ambassador of Goodwill and United Nations–appointed Global Champion. His work spans naval intelligence systems, computational linguistics, artificial intelligence, digital transformation, and civilizational governance architecture. His forward analysis on U.S.–Israel Pax Silica frameworks has appeared in Jewish/Jerusalem News Syndicate (JNS).
---
## Related articles
- [[articles/Humans and AIs as Entangled Learning Systems|Humans and AIs as Entangled Learning Systems]]: the companion argument that culture operates as human reinforcement learning, read through the organoid, the child and the model.
- [[articles/Human Proxy-Based Reinforcement (DeepRL) in Human Environments|Human Proxy-Based Reinforcement (DeepRL) in Human Environments]]: what happens when the reinforcement-learning environment becomes human lives themselves.
- [[articles/gamification/This Is a Game Planet|This Is a Game Planet]]: governance as terrain work, setting parameters so that a desired behaviour becomes the equilibrium.
- [[articles/Cyber-Physical Lessons from Imperial Beekeeping|Cyber-Physical Lessons from Imperial Beekeeping]]: sovereignty located at the level that alters the enclosure, the institutional rung of this essay's ladder.
- [[articles/John Nash's Unparalleled Legacy in Societal Transformation|John Nash's Unparalleled Legacy in Societal Transformation]]: equilibrium, bargaining and the mathematics of stable cooperation under constraint.
- [[articles/Ideas as Lindy Organisms|Ideas as Lindy Organisms]]: ideas as organisms under selection, and why filtration reveals fitness while suppression destroys the evidence.
- [[articles/Digital Darwinism and the Invisible World of Machine Evolution|Digital Darwinism and the Invisible World of Machine Evolution]]: variation, selection and descent on computational substrates.
- [[articles/We Were the Inference Machines First|We Were the Inference Machines First]]: the predictive machinery humans ran before machines inherited it.
- [[articles/Mind Uploading and AI — The Host is Reusable and the Person is the Delta|Mind Uploading and AI — The Host is Reusable and the Person is the Delta]]: the receiving infrastructure a substrate-portable intelligence would need.
- [[articles/We Were Never Going to Make It|We Were Never Going to Make It]]: the thermodynamic case for a synthetic carrier of human memory and civilization.
## Related topics
- [[wiki/Game Theory|Game Theory]]: the mathematics of strategic interaction from which the essay begins.
- [[wiki/Non-Transitive Games|Non-Transitive Games]]: cyclic advantage, whose persistence depends on population structure and learning.
- [[wiki/Nash Equilibrium|Nash Equilibrium]]: the stable point Rock, Paper, Scissors reaches only through unpredictability.
- [[wiki/Reinforcement Learning|Reinforcement Learning]]: learning from the consequences of action, in organisms and machines alike.
- [[wiki/Multi-Agent Systems|Multi-Agent Systems]]: populations of learners whose environments are one another.
- [[wiki/Autocurriculum|Autocurriculum]]: new learning challenges generated by interacting learners.
- [[wiki/Replicator Dynamics|Replicator Dynamics]]: how strategy frequencies change under frequency-dependent performance.
- [[wiki/Coevolution|Coevolution]]: reciprocal adaptation between interacting populations.
- [[wiki/Collective Intelligence|Collective Intelligence]]: cognition at the scale of the group, as in the honeybee swarm.
- [[wiki/Maynard-Cross Learning|Maynard-Cross Learning]]: the bounded swarm-level learning equivalence in the honeybee nest-choice model.
- [[wiki/Cultural Evolution|Cultural Evolution]]: inheritance and selection of learned behaviour across generations.
- [[wiki/Cultural Transmission|Cultural Transmission]]: the social learning channels dual-inheritance theory models.
- [[wiki/Cooperation|Cooperation]]: the outcome institutional evolution and reciprocity strategies can stabilize.
- [[wiki/Institutional Evolution|Institutional Evolution]]: rules and norms changing through the interaction that they govern.
- [[wiki/Rules of Change|Rules of Change]]: who can revise the rules, and how revision authority itself changes.
- [[wiki/Niche Construction|Niche Construction]]: organisms modifying the environments that select them.
- [[wiki/Open-Ended Evolution|Open-Ended Evolution]]: sustained production of novelty without convergence on a terminal solution.
- [[wiki/Open-Endedness|Open-Endedness]]: the AI research concept of systems that keep producing learnable novelty.
- [[wiki/Metacognition|Metacognition]]: modelling one's own modelling, the capacity behind learning how to learn.
- [[wiki/Recursive Self-Modification|Recursive Self-Modification]]: systems revising the processes that produced them.
- [[wiki/Simulation Hypothesis|Simulation Hypothesis]]: the possibility the essay opens and returns to.
- [[wiki/Substrate Independence|Substrate Independence]]: the premise that competencies can outlast their original embodiment.
- [[wiki/Machine Succession|Machine Succession]]: the wider framework for carrying intelligence beyond biology.
## References
**Rock, Paper, Scissors in people and in nature**
- [Social cycling and conditional responses in the Rock-Paper-Scissors game](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5376050/) — Zhijian Wang, Bin Xu and Hai-Jun Zhou, *Scientific Reports* 4, 5830 (2014).
- [The rock–paper–scissors game and the evolution of alternative male strategies](https://doi.org/10.1038/380240a0) — Barry Sinervo and Curt Lively, *Nature* 380, 240–243 (1996).
- [Local dispersal promotes biodiversity in a real-life game of rock–paper–scissors](https://doi.org/10.1038/nature00823) — Benjamin Kerr, Margaret Riley, Marcus Feldman and Brendan Bohannan, *Nature* 418, 171–174 (2002).
- [Mobility promotes and jeopardizes biodiversity in rock–paper–scissors games](https://arxiv.org/abs/0709.0217) — Tobias Reichenbach, Mauro Mobilia and Erwin Frey, *Nature* 448, 1046–1049 (2007).
**Learning, evolution and their mathematics**
- [A neural substrate of prediction and reward](https://doi.org/10.1126/science.275.5306.1593) — Wolfram Schultz, Peter Dayan and P. Read Montague, *Science* 275, 1593–1599 (1997).
- [Prefrontal cortex as a meta-reinforcement learning system](https://doi.org/10.1038/s41593-018-0147-8) — Jane X. Wang et al., *Nature Neuroscience* 21, 860–868 (2018).
- [The evolution of social dominance through reinforcement learning](https://doi.org/10.1086/713758) — Olof Leimar, *The American Naturalist* 197(5), 560–575 (2021).
- [Learning through reinforcement and replicator dynamics](https://doi.org/10.1006/jeth.1997.2319) — Tilman Börgers and Rajiv Sarin, *Journal of Economic Theory* 77, 1–14 (1997).
- [Evolutionary dynamics of multi-agent learning: a survey](https://doi.org/10.1613/jair.4818) — Daan Bloembergen, Karl Tuyls, Daniel Hennes and Michael Kaisers, *Journal of Artificial Intelligence Research* 53, 659–697 (2015).
- [Evolution of societies via reinforcement learning](https://arxiv.org/abs/2410.17466) — Yann Bouteiller, Karthik Soma and Giovanni Beltrame, arXiv:2410.17466.
**Collective and social learning**
- [The hive mind is a single reinforcement learning agent](https://doi.org/10.1007/s11721-026-00264-7) — Karthik Soma, Yann Bouteiller, Heiko Hamann and Giovanni Beltrame, *Swarm Intelligence* (17 October 2026); preprint at [arXiv:2410.17517](https://arxiv.org/abs/2410.17517).
- *Culture and the Evolutionary Process* — Robert Boyd and Peter J. Richerson, University of Chicago Press (1985).
- [Feature-based reward learning shapes human social learning strategies](https://doi.org/10.1038/s41562-025-02269-4) — David Schultner, Lucas Molleman and Björn Lindström, *Nature Human Behaviour* 9, 2183–2198 (2025).
**Populations, autocurricula and open-endedness in machine learning**
- [Real World Games Look Like Spinning Tops](https://proceedings.neurips.cc/paper/2020/hash/ca172e964907a97d5ebd876bfdd4adbd-Abstract.html) — Wojciech Czarnecki et al., NeurIPS (2020).
- [Grandmaster level in StarCraft II using multi-agent reinforcement learning](https://doi.org/10.1038/s41586-019-1724-z) — Oriol Vinyals et al., *Nature* 575, 350–354 (2019).
- [Autocurricula and the Emergence of Innovation from Social Interaction: A Manifesto for Multi-Agent Intelligence Research](https://arxiv.org/abs/1903.00742) — Joel Z. Leibo, Edward Hughes, Marc Lanctot and Thore Graepel, arXiv:1903.00742 (2019).
- [A multi-agent reinforcement learning framework for exploring dominant strategies in iterated and evolutionary games](https://doi.org/10.1038/s41467-025-67178-6) — Qi Su, Hongyu Wang, Yu Xia and Long Wang, *Nature Communications* (8 December 2025).
- [Position: Open-Endedness is Essential for Artificial Superhuman Intelligence](https://arxiv.org/abs/2406.04268) — Edward Hughes et al., ICML, PMLR 235, 20597–20616 (2024).
**Changing the game and the rules of change**
- [A dynamic over games drives selfish agents to win–win outcomes](https://doi.org/10.1098/rspb.2020.2630) — Seth Frey and Curtis Atkisson, *Proceedings of the Royal Society B* 287, 20202630 (2020).
- [Self-organized institutions in evolutionary dynamical-systems games](https://doi.org/10.1073/pnas.2500960122) — Kenji Itao and Kunihiko Kaneko, *PNAS* (15 April 2025).
- *Understanding Institutional Diversity* — Elinor Ostrom, Princeton University Press (2005).
- [Nomic: A Game of Self-Amendment](https://legacy.earlham.edu/~peters/nomic.htm) — Peter Suber, first published in Douglas Hofstadter's "Metamagical Themas," *Scientific American* (June 1982).
- *The Paradox of Self-Amendment* — Peter Suber, Peter Lang (1990).
- [Modelling the emergence of open-ended cultural evolution](https://doi.org/10.1098/rstb.2025.0255) — James Winters and Mathieu Charbonneau, *Philosophical Transactions of the Royal Society B* 381(1948), 20250255 (2026).
**Simulation**
- [Are We Living in a Computer Simulation?](https://doi.org/10.1111/1467-9213.00309) — Nick Bostrom, *The Philosophical Quarterly* 53(211), 243–255 (2003).