The Computer and the Conscience: On What Algorithms to Live By Actually Proves
Brian Christian and Tom Griffiths's Algorithms to Live By
I have been holding, in my mind, a specific image from Brian Christian and Tom Griffiths’s Algorithms to Live By: Johannes Kepler, Holy Roman Empire astronomer, discoverer of elliptical orbits, courting eleven women in sequence and returning, after much back-and-forth, to number five. The book presents this as something close to vindication—Kepler, without knowing it, was implementing a variant of the optimal stopping algorithm, allowing for recall. He explored, established a threshold, and leaped. He got a wife. He got six children. The marriage was happy.
The image is irresistible, and the book is full of them. A Nobel laureate splitting his retirement savings fifty-fifty between bonds and equities, not from ignorance of his own portfolio theory but from its correct application under uncertainty. A Mars rover procrastinating, ground engineers eventually diagnosing not laziness but priority inversion. A musician’s oblique strategy cards: randomness as structured escape from local maxima. Christian and Griffiths have written a book of beautiful correspondences between the problems computers face and the problems humans face, and they argue, mostly persuasively, that the solutions computer scientists have developed for those problems are transferable.
The book that results is one of the more genuinely illuminating popular science books of the last decade. It is also, examined carefully, something more interesting and more limited than it knows—a work that repeatedly proves rigorous results about specific formal problems and then applies them to human life with a speed and confidence the proofs don’t quite authorize.
The Structure of the Argument
Every chapter in Algorithms to Live By follows the same three-part architecture, and understanding the architecture is the key to evaluating the book’s claims.
First: here is a formally proven result in computer science. The 37% rule is the provably optimal strategy for the secretary problem when options arrive serially and cannot be recalled, your sole objective is selecting the single best option, and you can rank options relative to each other but have no cardinal information about their absolute quality. Second: empirical evidence that humans already approximate this strategy—Rapoport and Seale’s experiments, Tenenbaum and Griffiths’s prediction studies, Carstensen’s research on elderly social pruning—suggesting that human intuition is better calibrated to computationally optimal strategies than behavioral economics tends to assume. Third: a prescription. When you feel the urge to commit too early or explore too long, consult the algorithm.
This architecture has real intellectual force. The first part is mathematics—not metaphor, not analogy, not rhetorical illustration, but formal proof. The second part is published empirical research of varying rigor. The third part is inference: if the proof holds, and if human situations structurally resemble the problems the proofs address, then the strategies are good ones.
The question the book never quite settles, and perhaps cannot settle without dismantling its own premise, is whether that third move is always legitimate.
What the Proofs Actually Prove
The proofs are real. The 37% rule is mathematically optimal under its stated assumptions. The Gittens Index genuinely solves the multi-armed bandit problem with geometrically discounted payoffs. Least Recently Used eviction demonstrably outperforms alternatives under temporal locality. The Vickrey auction makes truthful bidding a dominant strategy. Exponential backoff stabilizes collision-prone networks. Finding Nash equilibria is computationally intractable in the technical sense of PPAD-completeness. These are not approximations or suggestions—they are theorems.
But theorems are conditional. They are optimal for their problem specification, and problem specifications are built on assumptions that real life is under no obligation to honor.
The secretary problem’s 37% rule is optimal when: options arrive serially and cannot be recalled, you have no cardinal information, you can rank options relative to each other, your sole objective is maximizing the probability of selecting the single best option, and the size of the pool is known. Relax any one of these—allow recall (as with Kepler), add cardinal information (income percentiles in dating), change the objective (satisficing rather than best-or-nothing), add a cost to searching (time, emotional exhaustion, competing priorities)—and the optimal strategy changes, sometimes dramatically.
Christian and Griffiths trace many of these variants, and the chapter on optimal stopping is admirably honest about the assumptions it’s making. But the book then applies the 37% heuristic to job hunting, house selling, and dating by narrative proximity rather than by verifying that those domains satisfy the required structural conditions. The reader who finishes the chapter knowing to spend roughly 37% of any search in non-committal exploration has learned something approximately useful. The reader who walks away believing they have derived optimal behavior from a mathematical proof has learned something slightly false.
The Empirical Claim, Examined
The book’s second move—humans already approximate optimal algorithms—is the most surprising and, in some domains, the most thoroughly validated. The Tenenbaum-Griffiths prediction experiments are the clearest case: subjects’ predictions for movie grosses, human lifespans, and congressional terms closely matched Bayesian posterior expectations calculated from actual population distributions. This is rigorous published research, carefully done, and the finding is genuinely striking. Human intuition carries accurate implicit priors for domains we regularly encounter, and those priors are approximately well-calibrated in the Bayesian sense.
But the book’s other empirical confirmations are less uniform. In optimal stopping experiments, humans stop early—around the 31st percentile rather than the 37th—and do so more than 80% of the time. In multi-armed bandit experiments, they over-explore. In scheduling, they fail to correctly weight tasks by importance per unit time. The pattern isn’t “humans approximate optimal algorithms universally.” It’s “humans approximate optimal algorithms in familiar predictive contexts and deviate systematically in contexts requiring serial commitment under uncertainty.”
The book tends to explain these deviations away rather than take them seriously as evidence against the central thesis. Early stopping in the secretary problem is explained as rational response to implicit time costs not modeled by the algorithm. Over-exploration in multi-armed bandits is explained by the world’s non-stationarity. These are plausible—even likely—partial explanations. They are also unfalsifiable in a way that should give us pause. Any deviation from optimal algorithm behavior can be attributed to a richer problem structure, which means the claim “humans are approximately Bayesian optimizers” is not really at risk from any data. A sufficiently expansive conception of the implicit problem being solved will rescue it.
The Digression That Changes Everything
The book’s eleventh chapter, on game theory, is the one that quietly dismantles the optimistic framing of the ten chapters before it.
The earlier chapters position computer science as a source of individual strategies: here is how you should stop searching, how you should allocate exploration and exploitation, how you should schedule your tasks. The implicit picture is one of a reasoning agent doing better or worse at solving well-defined problems, and algorithmic insight improving performance.
Game theory reveals a different class of problems entirely. The prisoner’s dilemma doesn’t yield better outcomes when each player reasons more carefully—it yields worse ones, because defection is the dominant strategy regardless of how well either player reasons. The tragedy of the commons isn’t caused by individual irrationality—it’s caused by a payoff structure where rational individual behavior produces collectively disastrous equilibria. Information cascades arise not from poor thinking but from the rational updating of beliefs on publicly available information, where the public information has decoupled from underlying reality.
And then the hardest result: finding Nash equilibria is computationally intractable. This is not a metaphor. It is a theorem—PPAD-completeness—that means there is no known polynomial-time algorithm for computing the equilibria that classical game theory assumes rational agents will reach. As Papadimitriou puts it, if an equilibrium concept is not efficiently computable, much of its credibility as a prediction of the behavior of rational agents is lost.
The implications are profound. The first ten chapters of the book ask: how should I reason, given the computational constraints I face? The eleventh chapter reveals that the most important determinants of outcome in many real situations are the rules of the games being played, not the strategies of individual players within them. The shopkeeper who calculates optimal vacation time under an unlimited vacation policy is playing a game with a bad equilibrium, and no individual algorithm makes that better. The answer isn’t smarter play—it’s mechanism design, which is to say, changing the game.
Christian and Griffiths know this. The chapter on mechanism design is one of the book’s best. But the book ends where it began—with individual computational kindness, with framing questions to minimize others’ cognitive load, with the personal practice of approximating good algorithms. The systemic insight is allowed to illuminate the individual chapters rather than revise them.
The Question the Book Keeps Almost Asking
There is a thought experiment that Algorithms to Live By sets up but doesn’t quite spring.
The book describes optimal stopping, explores its variants, shows that humans approximate it under certain conditions and deviate from it under others, and prescribes the 37% rule as a correction for our worst instincts. It does this for caching, scheduling, explore-exploit, Bayesian prediction, overfitting, and the rest.
Now ask: what would it mean for a human life to be, in some deep sense, optimizable in the way a computer program is optimizable?
The computer scientist’s instinct is to formalize the objective function—what are you trying to maximize?—and then ask what strategy achieves the maximum. This is how scheduling theory works: choose a metric (minimum lateness? minimum sum of completion times? minimum weighted lateness?), and then the optimal algorithm follows. The book makes this explicit: pick your problems, it says. The most radical act is choosing which metric to optimize.
But the question of which metric to optimize is not itself a computational question. Darwin’s pro-con list is a beautiful example precisely because it shows the limit of the approach: the man who invented natural selection turned to a ledger sheet to decide whether to marry, listed children and books and the loss of freedom, and ultimately was moved not by the balance of the columns but by the thought that it was intolerable to be a neuterer bee working and nothing after all. The algorithm compressed. The longing spoke.
The book’s central metaphor—that computers and humans face the same problems—is partly true and immensely generative. But it quietly occludes a difference. Computers don’t have difficulty choosing their objective functions. We do. The problem of what to care about, which the algorithms assume solved, is not a problem computers face at all—and it is arguably the hardest problem human beings encounter.
What the Book Actually Achieves
None of this is a refutation of Algorithms to Live By. It is, instead, a characterization of what kind of achievement the book represents.
Its genuine contributions are three.
First, it gives readers a vocabulary for problems they had no language for. The explore-exploit tradeoff, once named, is visible everywhere. The price of anarchy quantifies the cost of decentralization. Overfitting names the danger of optimizing too precisely for available data. These concepts have genuine explanatory power, and Christian and Griffiths present them clearly.
Second, it provides a corrective to a simplistic narrative about human irrationality. Behavioral economics, in its popular form, has generated a story in which human cognitive errors are bugs—departures from rationality that need to be corrected. The algorithmic perspective suggests an alternative: many of these apparent errors are optimal responses to genuinely hard problems, solutions developed under the constraints of limited information, limited time, and uncertain environments. We stop early in optimal stopping problems partly because waiting has implicit costs not modeled by the classical formulation. We explore more than Gittens prescribes partly because the world is non-stationary. This is not a complete vindication of human cognition, but it is a useful corrective to the bug-hunting framing.
Third, and perhaps most durably, the book makes the case for computational kindness. If verification is cheaper than search—the practical implication of the gap between P and NP, even if the formal proof remains elusive—then offering someone a specific proposal to accept or reject is genuinely kinder than asking them to generate the optimum themselves. This is small but real, and it scales from scheduling meetings to designing cities.
The book is honest about its limitations in a way that comes through in the texture of the writing even when the argument doesn’t flag them explicitly. The language hedges appropriately—”suggests,” “roughly,” “in practice”—and the authors are clearly enjoying the correspondences they’ve found without insisting those correspondences are tighter than they are.
What they have written is not a manual for optimal human living derived from mathematical proof. It is something more interesting: a set of formal results from computer science applied to human experience with enough intellectual honesty to illuminate both what the results prove and where the analogy breaks down. The correspondences are genuinely illuminating. The caveats are genuinely necessary.
The algorithm gives you the 37% threshold. It does not tell you what you are searching for, or what it means to find it, or how to bear the 63% failure rate when you have followed the best possible process and still lost. That part remains stubbornly outside the proof.
Algorithms to Live By: The Computer Science of Human Decisions by Brian Christian and Tom Griffiths. Henry Holt, 368 pp.
Tags: Algorithms to Live By Brian Christian, optimal stopping secretary problem, overfitting human decision-making, Nash equilibrium computational intractability, computer science behavioral economics rationality