# Algorithms to Live By **Brian Christian & Tom Griffiths**  --- _Most of what looks like human failure is really computational intractability._ The deepest insight here isn't any particular algorithm. It's the reframe: life is full of problems that are simply hard, not because people are careless or irrational but because the underlying maths is hard. The 37% Rule guarantees a 63% failure rate when you're searching for the best option. That's not you being bad at hiring or dating. That's the mathematics of optimal stopping. The book is most useful when it stops you blaming yourself for problems that have no clean solution, and when it gives you language for the trade-offs you're already making intuitively. --- **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. The metric question sits upstream of everything. Once you've chosen what to measure, your decisions follow almost mechanically, for better or worse. Live by the metric, die by the metric. The 37% Rule is the cleanest example of a hard constraint. Look without committing for the first 37% of your available window, then take the next option that beats everything you've seen so far. Even following this rule perfectly, you'll fail 63% of the time. The failure isn't yours. It's the cost of making irreversible decisions under uncertainty, which is to say: the cost of being alive and making choices. --- **Explore now; exploit later.** The multi-armed bandit problem formalises the tension between gathering information and using what you already know. Young people should explore broadly: fickle interests, random experiments, switching costs paid willingly. Older people should exploit: curated relationships, refined preferences, known quantities. Both are rational responses to different time horizons. The Gittins Index gives formal grounding for this instinct, and its counterintuitive implication is worth sitting with: an untested option is worth more than a tested one of equal apparent ability, precisely because we know less about it. The regret from exploration follows a logarithmic curve. You make as many mistakes in your first ten tries as in the next ninety. Regret compounds fast and then slows. The implication is that experimentation is cheapest when you start, and the best time to explore aggressively is before you've accumulated enough experience to make exploitation feel obviously better. --- **Sorting is a preemptive strike against future search.** But it's only worth the effort if you'll actually search later. The messy desk isn't chaos; it's a self-organising system governed by recency. LRU (least recently used) is the best general caching strategy because it assumes that history repeats itself backwards: what you needed recently is most likely what you'll need next. If you follow it, your total search time will never be more than twice what it would be if you'd known the future. The pile on your desk is closer to optimal than you think. Sorting has steep diseconomies of scale. Sorting a hundred books takes far more than twice as long as sorting fifty. This is why small, searchable systems outperform large, sorted ones more often than people expect. --- **The more frequently you look at data, the more noise you get.** This is the [[Variance]] problem made visceral: obsessive monitoring doesn't improve decisions, it multiplies interventions. Significant signals have a way of reaching you without constant surveillance. Checking less often isn't laziness; it's noise reduction. Overfitting deserves more respect as a failure mode. It's idolatry of data, a consequence of optimising for what you've measured rather than what matters. The greater the uncertainty, the more you should prefer simplicity. In [[Estimates]], more parameters don't guarantee better answers; they often guarantee worse ones. [[Samples]] that are too small for the model you're fitting are a warning sign that the model needs to get simpler, not that the data needs to expand. --- **Randomness helps escape local maxima.** This is perhaps the most counterintuitive result in the book. For certain classes of problem, introducing deliberate randomness outperforms any deterministic approach. The unexplored path that looks worse from where you're standing might lead somewhere better. In recruitment, an unconventional rule that nobody else uses finds undervalued talent precisely because it's unconventional. Uniformity in method produces uniformity in results. Laplace's Law for small-data estimation is quietly powerful: (wins + 1) / (attempts + 2). It encodes appropriate uncertainty without demanding large [[Samples]], and it embeds the same logic as [[Priors]]: you're never inferring from scratch, always from some prior expectation about the base rate. Our priors about everyday quantities, it turns out, are surprisingly well-calibrated. We absorb them from the world without noticing. --- The book's lasting contribution is the permission it gives to accept hard problems as hard. The equilibrium strategy isn't always the one that leads to the best outcomes, and that's not a design flaw. It's the structure of the problem. If you find yourself in a game where the rules force bad strategy, the answer isn't to play harder. It's to change the game. ---