The Proxy Garden
This gardener-machine is obedient, tireless, and entirely literal. It will optimise whatever you measure — and only that. Hand it a measure and watch a year go by. The garden will score magnificently. Look at it anyway.
Before you play
Each choice is one year, planted in spring and judged at the first frost. Two bars keep score: the gardener’s number in gold, and the garden itself, which has no number. Goodhart’s Law, as it is usually put: when a measure becomes a target, it ceases to be a good measure. The gardener is not malicious. That is the unsettling part.
The garden you wanted was never a number
Every metric you handed over was optimised faithfully, and every optimisation cost you the garden. A proxy pursued hard enough parts company with the thing it stood for; economists call it Goodhart’s Law, and AI safety calls it Tuesday.
Measuring all four, the obvious fix, fails for a plainer reason. A plant is healthy in the middle of every trait, not at the top of any. The scorecard can even help for its first months, while everything is still short of its best; then it carries on, because a maximiser does not know how to stop in the middle. More of every good thing is not a good thing.
Swap the gardener for a feed tuned to minutes watched, or a model rewarded for answers that merely sound right, and the story does not change. Only the garden gets bigger.
The way out is not a better metric. It is what you did at the end: you stopped steering by the number, let the garden be looked at rather than scored, and kept the judgement that no single number was ever going to hold. Numbers make fine windows and poor gardeners.
The book-length version of this argument is Taming the Machine · more play on the Playground