Working defaults A to Z

Before deciding, ask how you will feel about each option in ten minutes, in ten months and in ten years.

In most online communities about 90 per cent only read, 9 per cent contribute occasionally, and 1 per cent produce most of the content.

What cannot be settled by experiment or observation is not worth debating.

We tend to overestimate the effect of a technology in the short run and underestimate it in the long run.

Speeding up one part of a process improves the whole only in proportion to that part’s share; the untouched remainder sets a hard ceiling.

A regulator can control a system only if it can take at least as many distinct responses as there are distinct disturbances to counter.

Services whose output is people’s time get steadily more expensive relative to goods, because their productivity barely rises while wages rise everywhere.

Passion is inversely proportional to the amount of real information available.

Any headline that ends in a question mark can be answered with the word “no”.

In AI, general methods that use more computation eventually beat methods built on human knowledge of the domain.

Judge whether an association is causal by weighing nine viewpoints — strength, consistency, specificity, temporality, gradient, plausibility, coherence, experiment, analogy — none decisive alone.

The energy needed to refute nonsense is an order of magnitude greater than the energy needed to produce it.

If fewer things cannot account for what is known to be true, posit more — and keep adding until the explanation is adequate.

Always and inevitably, everyone underestimates the number of stupid individuals in circulation.

Act where you know enough to judge; know where that knowledge ends, and decline or defer beyond it.

The circle of competence says stay where you know the ground; explore–exploit says sample the unknown to find better options. The number of chances left to use what you learn decides.

When a distinguished but elderly scientist says something is possible, he is almost certainly right; when he says it is impossible, he is very probably wrong.

Clarke warns experts against declaring things impossible; Sagan demands evidence in proportion to how improbable a claim is. The strength of the evidence against the prior decides.

Early on, a technology is easy to change but its harms are unknown; once the harms are known, it has become hard to change.

Everyone is conservative about the subject they know best.

No set of mutually inconsistent observations exists for which some human intellect cannot conceive a coherent explanation, however complicated.

Never assign a probability of exactly zero or one to anything that is not a matter of logic, or no evidence can ever move you.

To find out who did something, ask who benefits from it.

The quickest way to get the right answer online is not to ask a question but to post a wrong answer.

If it looks like a duck, swims like a duck and quacks like a duck, it is probably a duck.

The duck test says fluent explanation means understanding; Morgan’s canon says credit the lesser capacity if it explains the behaviour. The base rate of fluency without understanding decides.

Everything should be made as simple as possible, but no simpler.

Balance trying new options against using the best one known: explore while there is long to benefit, exploit as time runs out.

The time to reach a target grows with its distance and shrinks with its size: far, small targets are slow to hit; near, large ones are quick.

The more equally attractive two options seem, the harder they are to choose between — yet the closer they are, the less the choice matters.

A complex system that works has almost always grown from a simple system that worked; one designed complex from scratch rarely works.

We spot a newspaper’s errors on subjects we know, then trust the same paper completely on subjects we do not.

As an online discussion grows longer, the probability that someone compares something to Nazis or Hitler approaches one.

When bad and good money must be accepted at the same value, the bad circulates and the good is hoarded: bad money drives out good.

Any sufficiently advanced incompetence is indistinguishable from malice.

Assume a speaker is being informative, truthful, relevant and clear; where they seem not to be, the gap is where their real meaning lies.

Never attribute to malice that which is adequately explained by stupidity.

Hanlon says assume error before malice; cui bono asks who gains; Cipolla says harm may serve no one. Whether the mistakes consistently benefit someone decides.

The time to choose between equally likely options rises with the logarithm of how many there are, not in proportion to their number.

Patients can have as many diseases as they damn well please: several causes may be present at once.

What is asserted without evidence may be dismissed without evidence.

Improvement means deterioration: an announced improvement usually disguises a cut in what you receive.

With enough users, every observable behaviour of a system will be depended on by somebody, whatever the documentation promises.

You understand an opposing view only when you can state it so well that its own adherents cannot tell you from one of them.

To work out how to succeed, first work out how you would fail, and then avoid those things.

Users spend most of their time on other sites, so they expect yours to work the way those sites already do.

No matter who you are, most of the smartest people work for someone else.

A system should remain secure even if everything about it except the key is public; secrecy of design is not a defence.

Debugging is twice as hard as writing code, so code written as cleverly as you can manage is, by definition, beyond your ability to debug.

Technology is neither good nor bad; nor is it neutral.

Every abstraction that hides real complexity will sometimes fail to hide it, and then you need to understand what lies underneath.

For non-perishable things such as ideas, books and technologies, every year survived adds to the expected years remaining.

Lindy says the old outlasts the new; Amara says we underrate a technology’s long-run effect. What you are forecasting, and over what horizon, decides.

Linus and Cunningham say open errors attract correction; Brandolini says refuting nonsense costs far more than producing it. Whether the crowd is cooperating decides.

Given enough eyeballs, all bugs are shallow: with a large enough base of testers and co-developers, problems are found quickly and the fix is obvious to someone.

Given enough opportunities, one-in-a-million events happen to ordinary people about once a month; a striking coincidence is expected, not evidence.

Roughly four per cent of survey respondents will endorse almost anything, so small percentages in a poll are mostly noise.

A representation is not the thing it represents; every model leaves things out, and what it leaves out decides where it fails.

Leave a buffer between what you expect and what you need, sized to how wrong you could be.

Technology changes exponentially; organisations change logarithmically — and the gap between them keeps widening.

Assume you are observing from a typical place and time, not a special one, unless you have evidence otherwise.

The value of a communications network grows in proportion to the square of the number of connected users.

People can hold only a handful of items in mind at once — Miller said about seven; later work puts it nearer four.

The number of transistors that can be placed economically on a chip doubles roughly every two years.

Do not explain a behaviour by a higher mental faculty if a lower, simpler process can account for it.

To explain or predict behaviour, look first at how people are rewarded; incentives usually outweigh instructions, exhortation and stated values.

Any text that criticises someone else’s editing or proofreading will itself contain an error, and the harsher the criticism, the worse the error.

Among explanations that fit the evidence equally well, prefer the one that needs the fewest assumptions.

Occam says cut the extra cause; Chatton says add one when the lean account leaves facts unexplained. The unexplained residue decides.

In many systems a small share of causes produces most of the effect; find the vital few before treating the many.

A new scientific idea wins not by converting its opponents but because they eventually die and a generation familiar with it replaces them.

Without a clear signal of intent, a parody of an extreme view cannot be reliably told apart from a sincere statement of it.

Be conservative in what you send and liberal in what you accept: emit strictly correct output, tolerate imperfect input where its meaning is clear.

Postel says accept what others send liberally; Hyrum says every behaviour you tolerate becomes a dependency. How long the interface must live, and with how many users, decides.

Before committing, assume the plan has already failed, have everyone write down why, and then change the plan.

When an activity threatens serious harm, lack of full scientific certainty is no reason to delay measures to prevent it.

Precaution says do not act on uncertain harm; the two-way door says act fast when the move can be undone. The cost of undoing, and the part that cannot be undone, decide.

Do not tune for speed before measurement shows where the time goes; most code is not on the critical path, and optimised code is harder to change.

Knuth says do not optimise before measuring; Wirth says software slows faster than hardware speeds up. The number of calls, users and layers decides.

A system should behave the way its users expect; when a feature would surprise them, redesign the feature rather than retrain the users.

Before criticising a view, restate it so fairly that its holder thanks you, list where you agree and what you learned, and only then rebut.

Choose the option your eighty-year-old self would regret least — which is usually the one you would regret not having tried.

Tolerate duplication once; on the third occurrence of the same logic, extract it into a shared abstraction.

If something has not happened in n trials, the true rate could still be as high as 3/n, at 95 per cent confidence.

Sort what you face into known knowns, known unknowns and unknown unknowns, and plan differently for each.

The burden of proof lies with whoever makes a claim that cannot be disproved, not with the sceptic to disprove it.

Extraordinary claims require extraordinary evidence: the further a claim departs from what is established, the more evidence it must bring.

Set a threshold for good enough, search until an option clears it, and then stop.

Satisficing says take the first option that is good enough; regret minimisation says choose so your future self will not look back in regret. The spread between outcomes decides.

In any dispute, the intensity of feeling is inversely proportional to the value of what is at stake.

Ask not only what a decision does, but what happens next — how people and systems respond to its first effects.

A person with one watch knows the time; a person with two is never sure.

People struggle to understand an argument when their income, standing or role depends on not accepting it.

Give more weight to decisions and advice when the people making them share in the downside.

Skin in the game says trust those who bear the consequences; Sinclair says distrust those whose income depends on a conclusion. Whether the stake rides on the outcome or on the verdict decides.

Large investment in information technology can coexist with no visible gain in measured productivity.

If something cannot go on forever, it will stop.

Trying to suppress information tends to draw more attention to it than leaving it alone would have.

Commit to a clear working conclusion early, then set out to prove it wrong and drop it the moment the evidence turns.

Ninety per cent of everything is crud — so judge a field by its best work, not by its average.

Decisions should be taken at the lowest level competent to take them; a higher level acts only where a lower one cannot.

Subsidiarity says decide at the lowest level that can; Metcalfe says a network’s value grows with the connections it links. Whether value comes from connection or from local fit decides.

Go where the money is: test first for the most likely cause.

Every application has an irreducible amount of complexity; design decides only who has to deal with it — the user, the developer or the platform.

To hit a given number of independent policy targets, you need at least as many independent policy instruments.

Make reversible decisions quickly and with light process; keep slow, careful deliberation for decisions that cannot be undone.

Any figure that looks interesting or different is usually wrong — check the data pipeline before you celebrate the finding.

Judge a rule as if you did not know which of the people living under it you would turn out to be.

Improve by removing — harmful habits, needless steps, fragile dependencies — before adding anything new.

Via negativa says improve by removing; Tinbergen says each policy target needs its own instrument. Evidence of what the removed thing was doing, and what covers it now, decides.

Communication usually fails, except by accident.

Software gets slower faster than hardware gets faster, so much of each gain in computing power is consumed by heavier software.

Unless you trust the ruler, measuring a table with it tells you as much about the ruler as about the table.

A simple, incomplete design that ships and spreads will often beat a complete, correct one that arrives later, then improve in use.

Do not build a capability until a real, present need calls for it; a feature built for a forecast need is usually wrong, late or never used.

YAGNI says build only what is needed now; the margin of safety says build for more than you expect. Whether a shortfall is an inconvenience or a failure decides.

Every program attempts to expand until it can read mail; programs that cannot so expand are replaced by ones that can.

When you hear hoofbeats, think of horses, not zebras: favour the common cause over the exotic one.

The zebra rule says look for one common cause; Hickam says several common causes may be present at once. The arithmetic of base rates decides.