Algorithms to Live By

Their argument is that many everyday problems have exact formal analogues that computer science has already studied, and that the optimal strategies are known, counter-intuitive, and transferable.

Their framing is careful: these are not metaphors. Choosing when to stop looking at apartments is the same mathematical problem as optimal stopping, and the solution applies.

These are not analogies. The apartment search is optimal stopping, and the known solution actually applies.

Devmark’s reading — not a quotation from the book

Their opening result, and the best known. When you must choose from a sequence of options with no going back, the optimal strategy is to observe roughly the first thirty-seven per cent without committing, then take the next option better than everything seen.

Its appeal is that it gives a defensible answer to a class of decisions usually made by feel, and the success rate — also about thirty-seven per cent — is an honest statement of how well anyone can do.

Their treatment of the multi-armed bandit problem: how long to keep trying new options versus committing to the best one found.

Their key insight is that the answer depends on how much time remains. Exploration is rational when there is time to benefit from what you learn and irrational when there is not — which explains why the old sensibly return to known favourites and the young sensibly do not.

Their argument that sorting is only worth its cost if you will search repeatedly, and that many domestic and organisational sorting efforts are unjustified by the number of subsequent searches.

Their treatment of caching is the more useful half: least-recently-used eviction is provably near-optimal, which is a strong argument for keeping what you last used nearest to hand rather than organising by category.

Their survey of scheduling theory produces specific answers. To minimise total completion time, do the shortest task first. To minimise lateness, do the one due soonest. To minimise the number of late items, a different rule again.

Their most useful finding is that the objective determines the algorithm, so a person who cannot say what they are optimising cannot be scheduling well by any standard.

The right scheduling rule depends entirely on what you are minimising. Most people have never named it.

Devmark’s reading — not a quotation from the book

Their explanation of how to forecast a duration from a single observation, using the distribution the quantity is drawn from.

The practical content is the distinction between normally distributed quantities, where the expected remainder shrinks as you wait, and power-law quantities, where it grows — which is why waiting longer for a delayed train and a long-running project have opposite implications.

Their argument that more data and more complex models are not always better, and that the same applies to human deliberation.

Their conclusion — that under uncertainty, simpler rules and earlier stopping frequently outperform more elaborate reasoning — is the philosophically interesting part, and it connects to the heuristics literature.

Their most original social suggestion: recognising that asking someone an open question imposes computation on them, and that offering constrained options is a kindness rather than a presumption.

"When would suit you?" is expensive; "Tuesday at two or Thursday at ten?" is cheap. It is a small idea that changes behaviour immediately.

Not every mapping holds. Some human problems differ from their formal analogues in ways that matter — preferences change during a search, information is not independent, and the cost of a wrong answer is not symmetric.

The authors are generally careful about this, but several chapters press the analogy harder than the correspondence supports, particularly in the relationships material.

The explore/exploit framing, which is the most broadly applicable idea in the book: the right balance between trying something new and returning to a known good option depends almost entirely on how much time remains to benefit from what you learn.

And computational kindness, which is free, immediate, requires no mathematics, and improves every scheduling conversation you have for the rest of your life by moving the work from the other person to you.