On Game Theory

In 1983, the United States ran a war game called Able Archer. Soviet intelligence, watching it unfold, could not tell whether it was a simulation or the real thing. Both sides had nuclear weapons. Both sides had dominant strategies. The world nearly ended not because anyone wanted it to but because the structure of the game made mutual destruction the rational move. Game theory does not begin with mathematics. It begins with that. There is a mode of reasoning that sits underneath economics, politics, evolutionary biology, military strategy, negotiation and social organisation simultaneously. It is the recognition that the outcome of any situation involving more than one decision-maker depends not just on what you do but on what others do in response to what you do and what they expect you to do in response to what they do in response. This recursive structure of interdependent decision-making is the domain of game theory and once you have genuinely internalised its logic you cannot look at competitive, cooperative or mixed situations the same way again. The books I referred for writing this piece collectively build this mode of reasoning from first principles to sophisticated application. What follows is what they collectively teach. Thanks for reading tiwaryshailesh’s Newsletter! Subscribe for free to receive new posts and support my work. The Foundational Architecture: What a Game Actually Is Every strategic situation has a structure that can be analysed before any specific move is considered. The structure consists of the players, the choices available to each player, the information each player has when making their choice, the timing of decisions and the payoffs that each combination of choices produces for each player. Before asking what to do in any strategic situation, the systems thinker asks what game is actually being played here because the answer to that question determines what the available strategies even are. This prior question is consistently neglected in practice. Most people enter strategic situations with their attention focused on their own options and their own objectives, with the other players modelled only superficially. Game theory insists on the prior step: who are all the players, what are their actual payoffs, what information do they have and what is the timing structure of the interaction. The answers are frequently different from first appearances and the difference is strategically decisive. Dixit and Nalebuff establish the most important initial distinction. Games differ along two fundamental dimensions. The first is whether they are zero-sum or non-zero-sum. In a zero-sum game, every gain for one player is exactly a loss for another. The total value in the game is fixed and the question is only how it is divided. Chess, most sports and many political contests are approximately zero-sum. In non-zero-sum games, the total value available depends on what players do collectively. Both players can gain or both can lose depending on their choices. Most business, most negotiation and most social interaction is non-zero-sum. The error of treating non-zero-sum games as zero-sum is one of the most expensive strategic mistakes available and it is extremely common. The second dimension is whether the game is played once or repeatedly. This distinction is as important as the zero-sum distinction and they interact in ways that are surprising and consequential. The logic of a one-shot game is completely different from the logic of the same game played repeatedly because in repeated games the future creates incentives that do not exist in single interactions. Reputation, reciprocity and the threat of future retaliation all become strategically relevant only when there will be future interactions. Understanding whether you are in a one-shot or repeated game with any particular counterpart is therefore one of the first and most important analytical moves available. Dominant Strategies and Nash Equilibrium: The Bedrock Once the structure of a game is understood, the search for strategic solutions begins with the concept of dominance. A strategy is dominant if it produces better outcomes for the player using it regardless of what any other player does. If you have a dominant strategy, you should use it. Full stop. The reasoning of the other players, the history of the relationship and the specific context are all irrelevant if one strategy is genuinely dominant. And if all players have dominant strategies, the outcome is determined by those strategies regardless of what any player prefers. The Prisoner’s Dilemma is the canonical illustration and Poundstone’s history of it reveals why it became so consequential so quickly after its invention. Two suspects are held separately and each is offered the same deal: confess and implicate the other and you go free while the other gets a long sentence, but if both confess you both get a medium sentence and if neither confesses you both get a short sentence on lesser charges. Each prisoner’s dominant strategy is to confess regardless of what the other does. If the other confesses, confessing reduces your sentence. If the other stays silent, confessing lets you go free. Confess is dominant for both players. The result: both confess and both get medium sentences, which is worse for both than if neither had confessed. The Prisoner’s Dilemma is not a curiosity. It is the structural template for an enormous range of consequential real-world situations. Arms races between nations. Price wars between competitors. The overuse of common resources. The underprovision of public goods. The collapse of cooperation in any situation where individual incentives diverge from collective interests. In every case the structure is the same: individual rationality produces collective irrationality. Each player doing what is best for themselves, given what they expect others to do, produces an outcome that is worse for everyone than the cooperative outcome they could have achieved. Nash Equilibrium generalises this insight beyond games with dominant strategies. A Nash equilibrium is a combination of strategies such that no player can improve their outcome by unilaterally changing their strategy given what all other players are doing. It is a stable point in the strategic landscape, not necessarily optimal for any player and sometimes far from optimal for all players collectively, but stable in the sense that no individual player has an incentive to deviate from it unilaterally. The prisoner’s dilemma mutual confession is a Nash equilibrium. Both players confessing is stable because given that one player confesses, the other’s best response is also to confess. Understanding Nash equilibrium is powerful not because it always tells you what to do but because it identifies the stable attractors of strategic situations. In any repeated strategic interaction, the system tends to move toward Nash equilibria over time as players learn and adapt. Knowing where the equilibria are tells you where the situation is heading and whether intervention to change the game structure is required to reach a better equilibrium. The Repeated Game Revolution: Why Cooperation Is Possible The most important single finding in the empirical game theory literature comes from Axelrod’s tournament and it overturns the naively pessimistic reading of the prisoner’s dilemma as proof that cooperation is irrational. Axelrod invited game theorists, economists and computer scientists to submit strategies for an iterated prisoner’s dilemma tournament in which the same pairs of strategies would play each other across many rounds. The winning strategy across multiple tournaments, submitted by the psychologist Anatol Rapoport, was the simplest strategy entered: Tit for Tat. Cooperate on the first move. Then do whatever the other player did on the previous move. Cooperate if they cooperated. Defect if they defected. Return immediately to cooperation if they do. Tit for Tat won not by being clever but by embodying four properties that Axelrod identified as the structural requirements for successful cooperation in repeated games. It is nice: it never defects first. It is retaliatory: it immediately punishes defection. It is forgiving: it returns to cooperation the moment the other player does. And it is clear: its strategy is simple enough that the other player can model it accurately and understand that cooperation will be rewarded and defection will be punished. The implications extend far beyond game tournaments. The Evolution of Cooperation shows that these same four properties characterise the emergence of cooperation in biological systems, in human social evolution and in the sustained cooperation that makes markets, institutions and societies function. Cooperation is not irrational. It is the rational response to repeated interaction with players who will retaliate against defection and reward cooperation. The conditions that make cooperation stable are the conditions that allow complex social organisation to exist at all: sufficient probability of future interaction, sufficient ability to identify and remember how specific players have behaved in the past and sufficient capacity to reward cooperation and punish defection. This analysis identifies precisely why cooperation breaks down and what restores it. Cooperation collapses when interactions become one-shot or when players believe they will not interact again, when players cannot identify who defected or cannot attribute specific defections to specific players, and when the temptation to defect is large relative to the long-term value of sustained cooperation. Every institution designed to sustain cooperation, from contract law to international treaties to professional reputation systems, works by addressing one or more of these breakdown conditions: extending the shadow of the future, improving the attribution of defection to specific players and increasing the cost of defection relative to the long-term value of cooperation. Commitment, Credibility and the Strategic Importance of Constraints Schelling’s contribution to game theory is the insight that most consistently surprises people who encounter it for the first time. In strategic interactions, the ability to constrain your own future choices is frequently more powerful than having more choices available. The player who can credibly commit to a specific course of action, and specifically to a course of action that includes following through on threats that would be irrational to execute if called upon, changes the strategic calculus for every other player. The strategic logic is precise. A threat is credible only if the player making it would actually follow through if the threatened behaviour occurred. A threat that would be irrational to execute is not a genuine strategic constraint on the other player because the other player knows it will not be followed through. The player who threatens to destroy their own negotiating position unless their demands are met is making a credible threat only if they can somehow commit to actually destroying their position. The nation that threatens nuclear retaliation for a conventional attack is making a credible threat only if the other side believes it would actually execute a response that would invite counter-retaliation. The genius of Schelling’s analysis is the recognition that constraining your own rationality, reducing your own options, burning your own bridges, can be strategically powerful because it changes what others rationally expect you to do. Cortez burning his ships on the shore of Mexico was not madness. It was a commitment device. It eliminated the option of retreat, which changed the rational expectation of both his soldiers and the Aztecs about what he would do when faced with adversity. The soldiers who could not retreat fought differently than those who could. The commitment changed the game. This insight propagates through an enormous range of strategic contexts. The labour union that commits to striking rather than accepting any settlement below a specific threshold has a more credible negotiating position than one that will accept anything above their current offer. The competitor that credibly commits to matching any price cut by a rival has made the price cut less attractive for the rival before it happens. The nation that credibly commits to defending a specific piece of territory regardless of cost has made attacking it less attractive than attacking territory whose defence is uncertain. In every case, the commitment works not by improving the outcome if the threat is called but by preventing the threat from being called because the other player’s calculation of expected outcomes shifts. Schelling’s related insight in Micromotives and Macrobehavior addresses the aggregation problem that sits alongside the commitment problem. Individual decisions that are each rational from the individual’s perspective aggregate into collective outcomes that nobody chose and that frequently nobody wanted. The residential segregation that emerges from individual preferences for having some neighbours of the same group, even when that preference is mild, produces complete segregation at the neighbourhood level through a tipping point dynamic. The traffic jam that emerges from individual rational decisions to leave at a specific time produces delays that everyone would prefer to avoid. The bank run that emerges from individual rational decisions to withdraw deposits once other depositors start withdrawing destroys the bank that would have survived if no one had panicked. These emergent outcomes from aggregated individual rationality create the social situations that institutions, norms and policies are designed to manage. And they are systematically misread as the product of malicious actors, cultural pathologies or structural oppression when they are in fact the predictable emergent properties of individual rational responses to specific incentive structures. Changing the emergent outcome requires changing the incentive structure, not blaming the individuals responding rationally to it. Changing the Game: The Highest Leverage Strategic Move Brandenburger & Nalebuff’s Co-opetition makes the argument that should sit at the top of any senior strategist’s thinking about competitive situations. The most powerful strategic moves available in non-zero-sum games are not moves within the current game. They are moves that change the game itself. They introduce the concept of the value net, an extension of game theory to multi-player business situations that includes not just competitors but complementors, players whose participation makes the value of your product or service higher rather than lower. Intel and Microsoft are complementors: each makes the other’s product more valuable. A hotel and a convention centre in the same city are complementors. A streaming platform and a content producer are complementors, at least until one of them decides to vertically integrate and become a competitor. Identifying and cultivating complementors is frequently more strategically powerful than defeating competitors because it expands the total value available in the game rather than just competing for a share of existing value. The five dimensions along which a game can be changed are: the players, the added values each player brings, the rules of the game, the tactics or information architecture and the scope of the game. Each of these is a dimension of strategic intervention that most players never consider because they are focused on playing the existing game better rather than changing it. Apple entering the mobile phone market did not compete harder at the existing game of mobile phones. It changed the game by redefining what a mobile phone was and what value it could provide. The most consequential business strategies of the past three decades are almost all game-changing moves rather than game-playing improvements. The game-changing insight also applies to negotiation. The most powerful negotiating move is rarely the smartest play within the current negotiation. It is changing the structure of the negotiation: adding players whose presence changes the payoff structure for existing players, changing the information architecture so that information previously private becomes common knowledge or vice versa, changing the timing so that delay or urgency shifts the relative costs for different players, or changing the outside options available to one or more players so that their walk-away position and therefore their bargaining power changes. Information Asymmetry: The Hidden Architecture of Strategic Advantage A recurring theme across multiple books in this list is the strategic role of information. In game theory, information structure determines what strategies are available and what equilibria are possible in ways that are as important as the payoff structure itself. The fundamental categories are: games of complete information where all players know all payoffs, games of incomplete information where some payoffs are private knowledge and games of imperfect information where players do not know the history of previous moves. Most real-world strategic interactions are games of incomplete and imperfect information and the management of information, what you reveal, what you conceal, what you credibly signal and what you strategically obscure, is as important as any other strategic choice. Signalling is one of the most important and most misunderstood mechanisms in strategic interaction. A signal is credible only if it is costly enough to send that a player without the underlying quality being signalled could not profitably fake it. The peacock’s tail is the canonical biological example: it is a credible signal of genetic quality precisely because it is so costly to maintain that only a genuinely healthy peacock can afford it. Education credentials signal capability partly through their content and partly through the selection effect of their cost: students who are unlikely to be productive workers find it more costly to obtain credentials and therefore self-select out, making the credential a credible signal even if it teaches nothing directly relevant to job performance. In business and negotiation contexts, the design of credible signals is a primary strategic challenge. The startup that wants investors to believe in its prospects needs to find signals that a low-quality startup could not credibly send, the founder taking a low salary, meaningful personal investment, specific contractual commitments around milestones. The negotiator who wants to signal genuine willingness to walk away needs to have taken actions that make walking away genuinely low-cost for them. The signal works only to the extent that it is genuinely costly to fake. Screening is the symmetric problem: the design of mechanisms that induce other players to reveal their private information. Insurance pricing that offers a menu of premium and deductible combinations screens for risk level without the insurer needing to observe it directly. Job offers that vary in fixed salary and performance bonus screen for confidence in own performance. Warranties screen for product quality. In every case the mechanism works by making the different options attractive to different types, allowing the structure of choices to reveal information that could not be obtained by direct observation or questioning. The Explore-Exploit Trade-off: Sequential Decision-Making Under Uncertainty Christian and Griffiths bring the insights of computer science optimal algorithms into contact with game theory through the lens of sequential decision-making under uncertainty. The explore-exploit dilemma, the tension between gathering information about options and using the best option currently known, arises in an enormous range of strategic contexts and has mathematically precise solutions that provide useful decision heuristics even when the exact mathematical conditions are not met. The multi-armed bandit problem, in which a player must choose repeatedly among options with unknown payoff distributions, is the canonical formulation. The optimal solution under specific conditions involves a specific balance between exploration of unknown options and exploitation of the best-known option. The key insight is that the optimal amount of exploration decreases over time as more information is gathered and as the time horizon shortens. Early in a sequence of decisions, exploring widely is valuable because the information gathered improves all subsequent decisions. Late in a sequence, exploiting the best-known option is correct because there is insufficient time remaining to recoup the cost of exploration through improved future decisions. Applied to hiring, this suggests trying many different types of candidates early in an organisation’s life when the cost of a wrong hire is low relative to the information gained and converging on a specific profile later when the cost of experimentation is higher. Applied to career decisions, it suggests deliberately exploring different roles and domains early in a career when the option value of information is high and the cost of switching is low. Applied to product development, it suggests running many small experiments early and concentrating resources on the best-performing approaches as the launch deadline approaches. The optimal stopping problem provides the precise answer to another class of sequential decisions: when to stop searching and commit to the best option found so far. The mathematical solution, explore for the first 37 percent of your search horizon then commit to the first option that exceeds the best seen during exploration, is not always directly applicable but it establishes the important qualitative principle that the optimal stopping point is always earlier than intuition suggests and that the cost of searching too long is frequently as large as the cost of stopping too early. Skin in the Game: The Alignment Problem Beneath All Strategic Interaction Taleb’s Skin in the Game addresses the most fundamental corruption of strategic interaction: the separation of decision-making from consequences. Every game theory model assumes that players bear the outcomes of their strategic choices. When this assumption is violated, when the person making the decision does not bear the cost of a bad outcome, the game being played is not the game that appears to be being played and the analysis that treats it as the apparent game produces systematically wrong predictions. The financial crisis of 2008 is the canonical modern example. The originators of mortgage loans who sold them on to securitisation vehicles had no skin in the game: they were paid for origination volume regardless of subsequent loan performance. The rating agencies that rated the resulting securities were paid by the issuers whose securities they rated: their skin in the game was the opposite of the investors relying on their ratings. The executives of financial institutions that failed were paid primarily in annual cash bonuses for short-term performance: their downside in the event of catastrophic failure was limited to the loss of future bonuses. The entire system was a game theory problem in which the apparent game, prudent financial intermediation, was not the actual game being played by the actual incentive structure, which was the extraction of fees and bonuses from a system whose tail risks were borne by others. The principle Taleb extracts is ancient and simple. Do not take advice from someone who does not bear the consequences of their advice. Do not trust risk assessments from institutions that profit from the risk without bearing its downside. Do not rely on the judgment of decision-makers who are insulated from the outcomes of their decisions. The absence of skin in the game corrupts the information content of every signal, the reliability of every commitment and the alignment of every incentive in ways that make standard game theory analysis inapplicable. The corollary is that the most reliable strategic partners, whether in business, politics or personal life, are those whose interests are genuinely aligned with yours through the structure of their payoffs rather than through the content of their stated commitments. Stated commitments are cheap. Structural alignment of payoffs is expensive to fake and therefore credible. This is the skin in the game version of Schelling’s commitment insight: the most credible signal of future behaviour is a structure of payoffs that makes the desired behaviour the rational choice regardless of stated intentions. Political Games: The Strategic Logic of Power Bueno de Mesquita and Smith’s Dictator’s Handbook applies game theory to political power with a rigour that explains what conventional political analysis consistently cannot. Their model treats political power as a multi-player coalition game in which the leader’s survival depends on maintaining the support of a winning coalition and all decisions, however dressed in the language of national interest or ideological principle, are best understood as moves in this coalition maintenance game. The three key variables are the nominal selectorate, everyone who has some formal role in choosing the leadership; the real selectorate, those whose support actually matters for leadership survival; and the winning coalition, the minimum subset of the real selectorate whose support is sufficient to maintain power. The game theory logic is inexorable. The smaller the winning coalition relative to the real selectorate, the more cheaply the leader can buy each coalition member’s loyalty because each member knows they are replaceable by someone from the larger pool and therefore accepts less. The larger the winning coalition, the more expensive each member’s loyalty and the more the leader must deliver public goods to maintain power. This single structural variable, coalition size, predicts an enormous range of political outcomes that conventional analysis treats as the result of cultural, ideological or personal factors. Large coalition systems, which require broad popular support, tend to provide public goods, respect civil liberties and maintain institutional constraints because the leader needs too many supporters to maintain power through private rewards alone. Small coalition systems tend to be kleptocratic, repressive and institutionally arbitrary because the leader can maintain power by distributing private rewards to a small group and has no incentive to provide the public goods that the excluded majority would benefit from. The game theory implication for understanding any political system is to identify the actual winning coalition, not the formal institutional structure and to understand the payoff structure that determines what leaders rationally do. The analysis extends to corporate governance, institutional management and any organisation where power is maintained through coalition. Wherever you find outcomes that seem irrational when explained by stated objectives, the game theory analyst asks what coalition is being maintained and what payoffs that coalition requires. Forecasting in Strategic Environments: The Epistemic Discipline Tetlock and Gardner’s Superforecasting connects game theory to the epistemic practices that make strategic analysis accurate rather than merely sophisticated. The challenge of strategic reasoning is not only the design of optimal strategies but the accurate assessment of the probabilities of different scenarios, the likely responses of other players and the range of outcomes that different strategy combinations can produce. Tetlock’s research on forecasting accuracy identifies the habits of thinking that distinguish the best forecasters from the rest. They think in probabilities rather than binary predictions, assigning numerical likelihoods to outcomes rather than predicting yes or no. They update their probabilities in response to new information rather than anchoring on initial assessments. They actively seek disconfirming evidence for their current views rather than only confirming evidence. They decompose complex questions into simpler components that can be assessed more accurately. They use base rates, the historical frequency of similar events, as anchors that are adjusted based on specific case features rather than ignored in favour of case-specific narratives. In game theory terms, the superforecaster is the player who most accurately models the probability distributions over other players’ strategies and over the outcomes of different strategy combinations. The player with the most accurate model of the game’s probability structure, other things equal, will make better strategic decisions because they are choosing among options whose expected values they can estimate more accurately. The calibration insight is particularly important. A well-calibrated forecaster’s 70 percent confidence predictions are right 70 percent of the time, not 90 percent or 50 percent. Most people are systematically overconfident in their predictions. In strategic contexts, overconfidence about the probability that your strategy will succeed, that the other player will respond as you predict or that the game structure is as you have modelled it produces strategies that look optimal in the modelled game and perform poorly in the actual game. The discipline of honest probability assessment, including honest assessment of your own uncertainty about the game structure itself, is the epistemic foundation that makes all the strategic sophistication above actually useful. The Unified Picture Pull all of this together and the complete architecture of strategic intelligence becomes visible. The first layer is structural analysis. Before any strategic move is considered, the game must be understood: who are the players, what are their actual payoffs, what information do they have, what is the timing structure and most importantly what kind of game is actually being played. Many strategic failures begin with misidentification of the game type: treating a repeated game as a one-shot game and defecting at the cost of a long-term relationship, treating a non-zero-sum game as zero-sum and missing the cooperative value available, treating a game of complete information as one of incomplete information or vice versa. The second layer is equilibrium analysis. Given the game structure, where are the stable attractors and are they good or bad equilibria from the perspective of the players? The prisoner’s dilemma mutual defection is a stable equilibrium that is bad for all players. Identifying that the equilibrium is bad and that the game structure is producing it points directly to the intervention: change the game structure so that a better equilibrium becomes stable. The third layer is commitment and credibility. What commitments can you make that are credible because they are costly to fake? What constraints on your own future choices change the rational expectations of other players in ways that benefit you? What signals can you send that separate you from players trying to fake your qualities? What screening mechanisms reveal the private information of other players? The fourth layer is the repeated game dynamic. If this interaction will recur, what reputation are you building with every move? Is your current behaviour consistent with the long-run strategy that serves your interests across the full sequence of interactions? Are you in a relationship where the tit-for-tat logic of reciprocal cooperation applies or a one-shot interaction where different logic applies? The fifth layer is game changing. Can you change the players, the payoffs, the rules, the information structure or the scope of the game in ways that produce better outcomes than any strategy within the existing game? This is the highest-leverage layer and the most neglected in practice. And beneath all of it is the epistemic layer: are your probability assessments of outcomes, player types and strategy effectiveness well-calibrated? Are you updating accurately on new information? Are you accounting for the skin-in-the-game distortions in the information and advice you are receiving? What This Demands of You Game theory does not tell you what to do. It tells you how to think about what to do. The demand it makes is the demand of genuine strategic sophistication: the willingness to model other players as rational agents with their own objectives rather than as obstacles or allies whose behaviour is determined by their character rather than their incentives. The most practically important single shift that game theory produces in a sophisticated practitioner is the habit of asking not just what is my best move but what will the other players do in response to my move and what will I do in response to that. This forward induction into the sequence of responses that any move initiates changes the evaluation of options in ways that consistently produce better outcomes than analysis that stops at the first move. The person who has genuinely internalized this mode of thinking approaches negotiations differently, understanding that the other side’s constraints and payoffs are as important as their own. They approach competitive situations differently, looking for game-changing moves rather than only game-playing improvements. They approach institutional design differently, understanding that the payoff structure determines behaviour more reliably than stated values or intentions. And they approach forecasting differently, maintaining calibrated uncertainty rather than false confidence about how the game will play out. Game theory is ultimately the formalisation of the oldest strategic wisdom: know the game you are in, know the players you are with, know where the stable outcomes are and know how to build the commitments and the reputation that make the best outcomes stable. The mathematics is optional. The thinking is not. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tiwaryshailesh.substack.com