Algorithms to Live By by Brian Christian & Tom Griffiths: Summary & Notes
by Brian Christian & Tom Griffiths
In one sentence
A tour through computer science's best solutions to hard decisions — when to stop searching, when to explore versus stick with a favorite, and how to organize and predict with limited information — applied directly to everyday human choices.
Key takeaways
- The 37% rule solves the "optimal stopping" problem: when searching through a fixed set of options one at a time with no do-overs (apartments, job candidates, even a parking spot), spend the first 37% of your search just looking — rejecting everyone — then pick the very next option that beats everyone you've seen so far. The 37% figure comes from 1/e (e ≈ 2.71828), and it maximizes your odds of ending up with the single best option.
- The explore/exploit tradeoff, borrowed from the multi-armed bandit problem, explains when to try something new (a new restaurant) versus stick with a known good thing (your favorite spot). The math says explore more early, when you have a long time horizon to cash in on what you learn, and exploit more as your remaining time shrinks — which is also a mathematical explanation for why novelty-seeking fades with age.
- Sorting algorithms reveal a real cost curve for organizing physical or digital stuff: the effort of fully sorting something (like alphabetizing a bookshelf) scales faster than the effort of just searching it unsorted, so past a certain collection size, it is mathematically rational to stay messy rather than sort.
- Caching and the LRU (least recently used) principle — the same rule computers use to decide what to evict from a small, fast cache — maps directly onto memory and decluttering: keep what you have used recently within easy reach, and let the rest migrate to slower storage (the back of the closet, a box in the attic, the bottom of the stack).
- Bayes's rule shows how to make predictions from very little data by combining what you already know (the prior) with a new observation. Applied to everyday forecasting — how long a Broadway run or a poem's popularity will last — it produces the "Copernican principle": in the absence of other information, assume you are seeing something at a random point in its lifespan, which means a good guess for remaining duration is proportional to how long it has already lasted.
- Game theory and randomized algorithms show why unpredictability itself can be a rational strategy: in adversarial situations (poker bluffs, penalty kicks, tax audits), being predictable is a weakness other people or systems can exploit, so deliberately randomizing your behavior can be the mathematically correct move.
Summary
Algorithms to Live By pairs everyday decisions with the branch of computer science built to solve their structural twin: house hunting and hiring map to the optimal stopping problem and its 37% rule, choosing a restaurant maps to the explore/exploit tradeoff, and tidying up maps to sorting and caching algorithms like LRU.
The book's case is that these aren't loose metaphors — the underlying math is the same, and computer science has already proven what the best strategy is under realistic constraints like limited time and incomplete information.
Reflections
The strongest parts of the book are the ones with a clean, provable answer — optimal stopping and the 37% rule especially, since the math produces a specific, checkable number rather than a vague heuristic. The frameworks hold up well as decision aids for real constrained-choice situations: apartment hunting, hiring, deciding how long to keep searching before committing to any option with a deadline.
The weaker fit is anything involving other people's feelings or genuinely unknown-unknowns, where the clean math of the underlying computer science problem stops mapping onto the messiness of the actual decision. The book is upfront about this — it treats the algorithms as useful starting heuristics under uncertainty, not as replacements for judgment.
Who should read this
- People who want concrete decision-making frameworks instead of vague advice — how long to search before committing, when to try something new versus stick with a favorite
- Engineers, programmers, and anyone who already thinks in terms of systems, tradeoffs, and optimization
- Readers who enjoyed Thinking, Fast and Slow or Superforecasting and want a computer-science-flavored companion to behavioral economics
- Anyone facing a live version of these problems right now — apartment hunting, hiring, or deciding how much to declutter
Favorite quotes
- "To try and fail is at least to learn; to fail to try is to suffer the inestimable loss of what might have been."
- "Never exploring is no way to live. But never exploiting can be every bit as bad."
- "Even the best strategy sometimes yields bad results—which is why computer scientists take care to distinguish between 'process' and 'outcome.'"
- "If you want to be a good intuitive Bayesian... you need to protect your priors. Counterintuitively, that might mean turning off the news."
- "Sometimes mess is more than just the easy choice. It's the optimal choice."
- Paraphrase, not a verbatim quote: the authors describe a "brain fart" — a mental blank or retrieval failure — as functionally the same thing computer scientists call a cache miss.
FAQ
What is the 37% rule?
The 37% rule is the solution to the "optimal stopping" problem: when you must evaluate a fixed set of options one at a time and decide immediately whether to accept or permanently reject each one, the best strategy is to reject the first 37% of options outright (using them only to learn the range of what's available), then choose the next option that beats everything you saw in that initial phase. The 37% comes from 1/e (e ≈ 2.71828) and it maximizes the probability of picking the single best option in the whole set.
What is Algorithms to Live By about?
Algorithms to Live By, by Brian Christian and Tom Griffiths, argues that many everyday decisions — house hunting, hiring, scheduling, tidying up, predicting the future — are structurally identical to problems computer science has already studied and solved, such as optimal stopping, the multi-armed bandit problem, sorting, caching, Bayesian inference, and game theory. Each chapter takes one computer science concept and shows how its proven strategies apply to ordinary human choices.
What is the explore/exploit tradeoff?
The explore/exploit tradeoff, drawn from the "multi-armed bandit" problem, is the balance between trying new options to gather information (exploring) and choosing the option you already know is good (exploiting). The book explains that the mathematically rational balance depends on how much time you have left to benefit from what you learn: explore more when your horizon is long, and exploit more as it shortens — which the authors use to explain why people tend to try new things when young and gravitate to familiar favorites later in life.
How does Bayes's rule apply to everyday predictions in this book?
The book uses Bayes's rule to show how to make reasonable predictions from very little data by combining a prior expectation with new evidence. Its central example is predicting how much longer something will last (a play's run, a career, a relationship) using the "Copernican principle" — assuming you're observing it at a random point in its timeline, which implies a longer track record so far is itself evidence of more time still to come.
What does the book say about caching and decluttering?
The book applies the LRU (least recently used) caching algorithm — the rule computers use to decide what to evict from limited fast-access memory — to physical organization. The principle: keep what you've used most recently within easy reach, and let everything else migrate to slower, less accessible storage, since recency of use is a strong predictor of whether you'll need something again soon.
Is Algorithms to Live By worth reading if I don't have a computer science background?
Yes. The book is written for a general audience — Brian Christian is a writer, not a computer scientist by trade — and each concept is introduced through its real-world decision problem first, with the computer science explained in plain language alongside it.
Detailed Notes
Click to expand the full detailed notes →
Detailed Notes
Click to expand the full detailed notes →
The Core Idea
Brian Christian and Tom Griffiths, a writer and a cognitive scientist, spend the book pairing everyday human dilemmas with the computer science subfields built to solve their structural twins. Deciding where to eat is a version of the explore/exploit tradeoff that computer scientists study as the multi-armed bandit problem. Deciding whether to marry the next person you date is a version of the "secretary problem," an optimal stopping puzzle with a known best strategy. The book's throughline is that many decisions people treat as matters of personality or intuition are actually well-defined computational problems, and computer science has already worked out the math for how to approach them under uncertainty and time pressure.
Optimal Stopping and the 37% Rule
The clearest and most quoted idea in the book is optimal stopping, illustrated with the "secretary problem": imagine interviewing a fixed number of candidates for a job, one at a time, and you must accept or reject each on the spot — no going back. The proven optimal strategy is to interview and reject the first 37% of candidates purely to calibrate your sense of the field, then hire the very next candidate who is better than everyone you saw in that initial phase. That number, roughly 37%, is 1/e, where e is Euler's number (about 2.71828). Followed correctly, this strategy gives you a 37% chance of ending up with the single best candidate out of the whole pool — which is also, not coincidentally, the best success rate any strategy can guarantee in the worst case.
The book extends this to apartment hunting (look at the first 37% of places you tour without committing, then take the next one that beats the rest), and more loosely to dating and hiring, while being explicit about the assumptions the math depends on: you know roughly how many total options there are, options arrive in random order, and a rejected option is gone for good. When those assumptions break — for example, you can sometimes go back to an old option, or you don't know how many candidates you have left to see — the ideal cutoff point shifts, but the underlying "look, then leap" structure still applies.
Explore/Exploit and the Multi-Armed Bandit
The explore/exploit tradeoff comes from the "multi-armed bandit" problem: a gambler facing several slot machines with unknown, different payout rates has to decide how to split pulls between exploring (trying machines to learn their payout rates) and exploiting (playing the machine that currently looks best). The book walks through the Gittins index, a formula developed in the 1970s that assigns each option a single number balancing its known value against the value of the information you'd gain by trying it, and through simpler heuristics like the upper confidence bound approach used in modern recommendation and ad-serving algorithms.
The practical upshot: explore more when you have a long time horizon left to benefit from what you learn (why young people are told to travel and switch jobs), and exploit more — stick with known favorites — as the remaining time to capitalize on new information shrinks (why older adults often gravitate to the same restaurants and the same friends, a pattern psychologists call socioemotional selectivity theory). The book frames this as a rational response to a shrinking time horizon, not simply a loss of curiosity.
Sorting, Caching, and Organizing Physical Things
Sorting algorithms — the computer science of putting things in order — turn out to have a cost structure worth knowing before you alphabetize your bookshelf or file your email. Comparison-based sorting has a mathematically proven lower bound: you cannot sort n items with fewer than roughly n log n comparisons. Searching an unsorted list, by contrast, only costs O(n) — you just check everything once. Since sorting is more expensive than a single search, it only pays off if you expect to search the same collection many, many times. For anything you rarely need to find, staying messy and doing a slow linear search when you actually need something can beat the upfront cost of maintaining full order.
Caching addresses a related problem: what do you keep close at hand when you cannot keep everything close at hand? The LRU (least recently used) eviction policy — used in real computer memory hierarchies, from CPU caches to web browsers — says: when space runs out, discard whatever has gone the longest without being used. The book applies this directly to closets, garages, and desks: the items you have used most recently are the ones worth keeping in the most accessible spot, and everything else can migrate to slower "storage" (a closet, a box, the garage) without much practical cost, because the odds you'll need it again soon are, statistically, low.
Bayes's Rule and Predicting the Future from Thin Evidence
Bayesian reasoning is the book's tool for making calibrated predictions from very little data. Bayes's rule combines a prior belief with new evidence to produce an updated belief. The book's recurring illustration is predicting how long something will last — a Broadway play's run, a friend's life expectancy, a politician's career — using what statisticians call the Copernican principle: absent other information, assume you are observing the thing at a random point in its total timeline. That assumption implies a simple rule of thumb Christian and Griffiths call the "multiplicative rule": a reasonable guess for total remaining duration is proportional to how long the thing has already existed. Something that has run for 100 days is more likely to run for a while longer than something one day old, without needing any information beyond the one data point of its current age.
The same chapter warns that Bayesian updating is highly sensitive to the quality of the prior belief you start with, which is why the book — echoed in one of its more quoted lines — argues protecting a well-calibrated prior sometimes means limiting exposure to noisy, unrepresentative information (their example is the news), since bad priors compound into bad predictions no matter how good your math is afterward.
Game Theory and the Case for Being Unpredictable
The closing chapters turn to game theory and randomized algorithms, covering situations where another agent is actively trying to anticipate and counter your choices — poker, penalty kicks, tax audits, price wars. In these adversarial settings, being predictable is a liability: any consistent pattern in your behavior is something an opponent can learn and exploit. Computer science has a formal answer for this, borrowed from mixed-strategy game theory: deliberately randomizing your choices, in the right proportions, can make you un-exploitable in a way no fixed strategy can. The book also touches on computational complexity itself — some problems (like scheduling or certain negotiations) are provably hard to solve perfectly no matter how much computing power you have, which the authors treat as a reason to make peace with good-enough heuristics rather than chase an unreachable optimum.
