Time to perform a task drops as a power of repetitions.
Snoddy (1926) observed and Newell and Rosenbloom (1981) formalized: as people repeat a task, the time it takes follows a power-law curve. Sharp early gains, gentler later ones, and a long tail where additional practice still pays — just less.
Plot time-per-task against repetitions on a log-log axis and you get a straight line. The biggest gains come early — minute one to minute ten changes everything. Beyond that, each doubling of practice produces a smaller relative gain, but never zero.
Time-per-task and reps form a sharp decay. The first ten tries shave off most of the cost. Reps a hundred through a thousand still bring the time down — but slowly, on a curve that flattens toward an asymptote rather than zero.
Most onboarding decisions hinge on this. The first hour of practice is worth more than the next ten. The next ten are worth more than the next hundred. But the curve keeps going — experts in any complex tool keep getting faster years in, just at a rate beginners would not notice.
Because most of the learning happens early, onboarding decisions have outsized leverage. A first-run experience that drops setup from 8 minutes to 2 moves the entire curve to the right — fewer drop-offs, faster time-to-value, more users who stay long enough to get to the flat part of the curve.
Practice curves are rarely smooth. Users hit plateaus — stretches where time stops dropping despite reps. Often the way out is a new technique (a shortcut, a different mental model), not more practice on the same one. Design for the plateau: surface the next move, do not just count reps.
The slope of the curve depends on the technique available. A user who only ever clicks the menu stays on a gentler curve than one who learns the keyboard shortcuts. Surface the faster path early and often — power-users save not from talent but from a steeper learning curve.
Ericsson's research: deliberate practice — focused, with feedback, at the edge of ability — produces a steeper curve than mindless reps. Speed of typing, speed of code, speed of design review all improve faster when feedback is immediate and the next-hardest target is in view.
Gladwell turned Ericsson's research into "10,000 hours" — but the original work was about specific kinds of deliberate practice in specific domains. Ten thousand mindless hours produce a flat experienced user, not a top performer. The curve still applies; the multiplier depends on practice quality.
Your product is used by users at every point on the curve. The first-day user needs visible defaults and a successful first task. The week-100 user needs shortcuts, bulk actions, and customization. Each cohort sits on a different part of the curve — and they all see the same UI.
Models can drop the early part of the curve to almost flat — and that changes who counts as an expert.
A first-day user with an assistant can produce expert-level output. The steep early part of the practice curve — the part where most users used to drop off — becomes almost vertical. The expert and the novice are now closer than they have ever been, at least on the output. Whether they are closer on judgment is a different question.
The model handles the mechanics; the human still has to know what good looks like. The curve for judgment — knowing which output is right, which is plausible nonsense, which is subtly off — has not been flattened by anything. Reps still build it, and beginners still need them.