# Algorithms to Live By **Brian Christian & Tom Griffiths** | [[Numbers]]  --- > "Err on the side of messiness. Sorting something that you will never search is a complete waste; searching something you never sorted is merely inefficient." Life is full of problems that are simply **hard**. Mistakes say more about problem difficulty than human fallibility. The 37% Rule guarantees a 63% failure rate—that's not you being bad at hiring or dating; that's mathematics. Computer science shows us that what looks like personal failure is often just computational intractability. Three insights that reframe decision-making: **"Before you can have a plan, you must first choose a metric."** We think we're optimising, but we're often optimising the wrong thing. Live by the metric, die by the metric. **"Err on the side of messiness."** The big pile on your desk is actually well-designed. Your messy desk isn't chaos; it's optimal caching. Display files by "Last Opened" not "Name." Keep running gear by the front door. **"Outcomes make headlines—but processes are what we have control over."** Even the best strategy sometimes yields bad results. If you followed the best possible process, you've done all you can. Computer scientists distinguish between process and outcome; we should too. --- ## Core Ideas ### [[Optimal Stopping]] **37% Rule**: Look without committing for 37% of your window, then leap at the next option better than all previous. Even optimal play fails 63% of the time—this reframes "failure" as inherent difficulty, not personal inadequacy. Time costs are real. The classical secretary problem assumes no cost for looking, but there's always a time cost. People tend to stop early in experiments, and that's not irrational—it's accounting for endogenous search costs. After searching for a while, we humans just get bored. It's not irrational to get bored, but it's hard to model that rigorously. The problem's deepest constraint is the nature of time itself. Hesitation is as irrevocable as action. We decide based on possibilities we've not yet seen. No choice recurs. We may get similar choices again, but never that exact one. ### [[Explore vs Exploit]] > "People tend to treat decisions in isolation, to focus on finding each time the outcome with the highest expected value. But decisions are almost never isolated, and expected value isn't the end of the story. If you're thinking not just about the next decision, but about all the decisions you are going to make about the same options in the future, the explore/exploit tradeoff is crucial to the process." The multi-armed bandit problem formalises the tension between gathering information and using what you know: **Win-Stay, Lose-Shift**: simple but reliably better than chance—choose an arm at random, keep pulling it as long as it pays off, switch if it doesn't. **Gittins Index**: formal justification for preferring the unknown—"the grass is always greener" is mathematically sound when you have time to exploit discoveries. The untested rookie is worth more than the veteran of seemingly equal ability, precisely because we know less about them. **Upper Confidence Bound**: optimism in the face of uncertainty is perfectly rational—assume the best until proven otherwise. The success of these algorithms offers formal justification for the benefit of the doubt. **Regret minimisation**: logarithmic—you make as many mistakes in your first ten tries as in the next ninety. Every year you can expect fewer new regrets than the year before. Young people should explore (random button-pressing, fickle interests). Old people should exploit (curated social networks, known favourites). Both are rational responses to their time horizons. The deliberate honing of a social network down to the most meaningful relationships is the rational response to having less time to enjoy them. ### [[Sorting vs Searching]] **Scale hurts.** Sorting 100 books takes longer than sorting two shelves of 50. Sorting involves steep diseconomies of scale. Sorting is a preemptive strike against future search effort—but only valuable if you'll actually search later. The effort expended on sorting materials is just preparation against the effort it'll take to search through them. Sometimes mess is more than just the easy choice. It's the optimal choice. **LRU (Least Recently Used)**: the best general caching strategy—assume history repeats backward. The nearest thing to clairvoyance is to assume that history repeats itself—backward. If you follow the LRU principle, the total amount of time you spend searching will never be more than twice as long as if you'd known the future. The mathematics of self-organising lists suggests something radical: the big pile of papers on your desk, far from being chaos, is actually one of the most well-designed and efficient structures available. What might appear to others to be an unorganised mess is, in fact, a self-organising mess. --- ## Key Insights **Rationality doesn't mean exhaustive analysis—it means knowing when to stop.** Only 9% of scheduling problems we understand can be solved efficiently; 84% are proven intractable. Intuitively, we think rational decision-making means exhaustively enumerating options, weighing each carefully, and selecting the best. But when the clock is ticking, few aspects of decision-making are as important as: when to stop. **Weighted Shortest Processing Time sets an implicit hourly rate.** Divide importance by duration. Only prioritise a task that takes twice as long if it's twice as important. In business contexts, divide each project's fee by its size, and work your way from the highest hourly rate to the lowest. **Responsiveness vs throughput.** Decide your minimum acceptable responsiveness, then be no more responsive than that. Try to stay on a single task as long as possible without decreasing your responsiveness below the minimum acceptable limit. Giving yourself more time to decide doesn't necessarily mean you'll make a better decision—it does guarantee you'll consider more factors and risk overfitting. **When the future is foggy, you don't need a calendar—just a to-do list.** The weighted version of Shortest Processing Time is a good general-purpose scheduling strategy in the face of uncertainty: each time new work comes in, divide its importance by the time it'll take to complete. If that figure is higher than for your current task, switch; otherwise stick with it. **Laplace's Law for estimating with small data.** (wins + 1) / (attempts + 2). Small data is big data in disguise—our priors are surprisingly accurate. We appear to carry around in our heads surprisingly accurate priors about movie grosses, poem lengths, political terms of office, and human life spans. We don't need to gather them explicitly; we absorb them from the world. > "Something normally distributed that's gone on seemingly too long is bound to end shortly; but the longer something in a power-law distribution has gone on, the longer you can expect it to keep going." Good predictions require good priors. Our judgements betray our expectations, and our expectations betray our experience. What we project about the future reveals a lot—about the world we live in, and about our own past. > "Failing the marshmallow test—and being less successful in later life—may not be about lacking willpower. It could be a result of believing that adults are not dependable: that they can't be trusted to keep their word, that they disappear for intervals of arbitrary length." **Overfitting is idolatry of data.** It's a consequence of focusing on what we've been able to measure rather than what matters. The greater the uncertainty, the bigger the gap between what you can measure and what matters, the more you should watch out for overfitting—the more you should prefer simplicity, the earlier you should stop. **Randomness helps escape local maxima.** There is a deep message in the fact that on certain problems, randomised approaches can outperform even the best deterministic ones. Sometimes the best solution to a problem is to turn to chance rather than trying to fully reason out an answer. In recruitment, an unconventional rule nobody else uses beats a "better" rule everyone uses (finds undervalued talent). Whether it's jitter, random restarts, or being open to occasional worsening, randomness is incredibly useful for avoiding local maxima. There is a virtue in making slightly random decisions that don't conform to established rules. **Protocol is the foundation of connection.** The Greek *protokollon*: "first glue." Exponential Backoff works for networking and flaky friends—retry with increasing patience. AIMD (Additive Increase, Multiplicative Decrease): push to failure, cut in half, repeat. > "We've all had the experience of talking to someone whose eyes drifted away—to their phone, perhaps—making us wonder whether our lackluster storytelling was to blame. In fact, it's now clear that the cause and effect are often the reverse: a poor listener destroys the tale." **Tactical dropping of balls is critical to getting things done under overload.** We use the idiom of "dropped balls" almost exclusively in a derogatory sense, implying laziness or forgetfulness. But the tactical dropping of balls is a critical part of getting things done under overload. The problem isn't that we're "always connected"—it's that we're always *buffered*. Tail Drop > long queues: better to reject messages than create expectations you can't meet. **The equilibrium strategy isn't necessarily the strategy that leads to the best outcomes.** Even when we can reach an equilibrium, just because it's stable doesn't make it good. This has emerged as one of the major insights of traditional game theory: the equilibrium for a set of players, all acting rationally in their own interest, may not be the outcome that is actually best for those players. Unlimited vacation sounds generous, but from a game-theoretic perspective, it's a nightmare. Everyone wants to take just slightly less vacation than each other, to be perceived as more loyal. The Nash equilibrium is zero. **If the rules of the game force a bad strategy, change the game.** By and large we cannot shift the dominant strategies from within. The solution has to come from somewhere else—mechanism design. Seek out games where honesty is the dominant strategy (the revelation principle: any game requiring strategically masking the truth can be transformed into a game requiring nothing but honesty). Then just be yourself. --- ## Connects To - [[The Fifth Discipline]] – systems thinking, delays, feedback loops, balancing exploration (learning) with exploitation - [[Better, Simpler Strategy]] – metrics and measurement; "before you can have a plan, you must choose a metric" echoes WTP/WTS clarity - [[Alchemy]] – challenges pure logic; algorithmic thinking is rational but not always sufficient for human systems - [[High Performance Habits]] – process over outcome thinking - [[Dead Companies Walking]] – explore/exploit applies to industry life cycles and when to abandon failing strategies --- ## Final Thought *Algorithms to Live By* is permission to stop feeling guilty. Your restlessness in your twenties wasn't flightiness; it was rational exploration. Your 63% failure rate at hiring wasn't incompetence; it was mathematics. The book's real gift is vocabulary: explore vs exploit, process vs outcome, regret minimisation, optimal stopping. These frames help you see that many apparent personal failures are actually optimal responses to hard problems. And once you see the problem clearly, you can stop trying to be perfect and start being strategic about where to invest effort.