the second hundred · metaphor 188
Is there a best way to choose — one system for deciding, searching, betting that beats every other, whatever the situation? The unwelcome answer is no. Averaged over every possible world, every method is exactly as good as every other.
We keep hunting for the one true method: the productivity system, the investing rule, the diet, the parenting philosophy, the hiring heuristic that finally works. And someone always seems to have found it — brilliant results, a clean story, converts. The suspicion this page makes precise is that the brilliance is real but local: it is a fit to one particular kind of world. Move the method to a differently shaped world and the same cleverness costs exactly what it earned.
There is a theorem for this, and it is unusually blunt. No search or prediction strategy is better than any other when its score is averaged over all possible problems. Not "usually tied" — provably, exactly tied. Every edge you have is borrowed from an assumption about which worlds are likely. Below is a small universe of problems where you can watch a clever method win, watch it lose, and watch the average refuse to break.
The flat average
Here is why the average cannot break. A "world" is an assignment of hidden payoffs to cells — and this little universe contains all of them, every one of the 720 ways to lay six distinct values on six cells. A method is just a rule for deciding which cell to open next, given what it has seen. The Sweep reads left to right; the Hill-climb chases whatever looked best so far; Bisect and Scout follow fixed scatters. Some are clever, some are dumb.
Run any one of them across all 720 worlds and something exact happens: the sequence of payoffs it uncovers runs through every possible ordering exactly once. A clever method just visits a lucky order on some worlds and an unlucky one on others — but summed over all worlds, luck is conserved. So the distribution of "best found so far" is identical for the genius and the fool. That is the No Free Lunch theorem, and the instrument computes it by brute force: not a formula printed on the screen, but 720 worlds actually searched, four methods, every probe counted.
What to try
Press New world a few times. On this world the Hill-climb finds the 6; on the next, its cleverness walks it straight past the peak and the dumb Scout wins. Flip enough worlds and the crown-count chips fill in — every method wears the crown on a healthy share of worlds, and none dominates. Then look down at the average bars: four numbers, and no light between them. Drag the budget and all four rise together, locked, to six decimal places.
Use Head to head to pit your favourite against a rival across all 720 worlds at once. You'll find your method genuinely beats the rival on hundreds of worlds — and loses on just as many by just as much. The wins are real. They are also, in total, worth exactly nothing. That gap between "wins a lot" and "wins on average" is the whole lesson: an edge is not an average, it is a bet.
The mapping
The theorem is about search algorithms, but the shape of it is about lives. A method that wins is a method whose assumptions happen to match the world it is run in. The Hill-climb wins on smooth worlds where good things sit near good things; it loses on deceptive ones where the peak hides behind a trough. Neither is smarter in the abstract. One is a bet that the terrain is gentle; the other is a bet that it is tricky.
So the guru with real results is not lying, and is not, on average, right. They have found a method well-fitted to the slice of worlds they operate in — a bull market, a certain kind of student, a company at a certain size. The failure is not their method; it is the story that the method is the method. Advantage is always on loan from the world, and the world can call the loan. The useful question is never "what is the best strategy" but "what is this strategy assuming, and is that true here?"
Read as life lessons
A method has no context-free quality; it has a fit. Every technique that helps you is quietly assuming your world is shaped a certain way. Name the assumption and you can check whether it still holds.
The theorem only bites when all worlds are equally likely. Real advantage exists — because real worlds have regularities. It is bought by matching your method to that structure, never by the method alone.
A strategy can beat every rival on hundreds of worlds and still be exactly average, because the losses cost as much as the wins. Track counts is not the same as track record.
In the wild
No learning algorithm is best across all datasets; each carries an inductive bias — an assumption about which patterns are likely. Performance is the bias meeting a matching world.
Every trading edge is a bet on a regime. The strategy that printed money in one decade's structure quietly bleeds when the structure turns — same rules, different world.
There is no universal predictor. A model tuned to one population, disease, or era can be worse than coin-flips elsewhere; the skill lives in the fit, not the machinery.
The honest model
Nothing here is scripted. On load, the page enumerates all 6! = 720 permutations of the payoffs 1–6 over six cells — the complete space of worlds. For each world and each of the four methods it replays the probe sequence, records the best value found within the budget, and averages. The four averages come out bit-for-bit equal (5.2500 at budget 3), which the readout confirms by printing the spread as literally 0.000000. The head-to-head counts are exact tallies over the same 720 worlds.
The reason is the clean core of Wolpert & Macready's proof: for any deterministic method that never re-probes a cell, the payoff sequences it uncovers, as the world ranges over all permutations, are themselves all permutations — each once. Averages built from "all orderings, each once" cannot depend on the method. The one honest caveat baked into the toy: it assumes every world is equally likely, which is exactly the assumption real life violates — see below.
The mapping, exactly
| Mathematics | Life |
|---|---|
| the space of all problems | Every possible world you might face — including the vast majority no one ever does. |
| a search strategy | A method, habit, or philosophy for deciding what to try next. |
| score on one problem | How well your method happens to fit this particular situation. |
| average over all problems | Your method with the world's regularities stripped away — where it is merely one of the crowd. |
| a non-repeating search | Not wasting moves on the already-known — and still unable to win everywhere. |
| a free lunch | A genuine, general edge — which the theorem says can only ever be bought with an assumption about the world. |
Where the metaphor tears
No Free Lunch averages over every possible problem, and almost all of those are structureless noise that no human ever meets. Our world is deeply regular — good things cluster, causes precede effects, yesterday resembles today. Over the thin, structured slice of problems we actually face, some methods really are better, durably. The theorem forbids a universal winner, not a local one; it is a warning against the word "always," not against skill.
Read carelessly, No Free Lunch counsels despair: nothing works, so why choose. That inverts it. The theorem says advantage must come from an assumption — so the move is to make the assumption explicit and test it, not abandon method. "This works because I expect the terrain to be smooth" is a checkable claim. Paralysis is just a worse bet, made silently.