19 October 2026 · Built with Claude

Human & AI
on a tandem

Two riders, one bike. You decide who learns faster, and whether the AI learns continuously or only in occasional releases. The shared score is how much they train against the same tests and specification. Marginal gains make the bike quicker. Only when both riders are good, and in time with each other, does something new happen.

Wobbling off the start line

Try leaving the shared score low: the AI's own score can climb, but less of it helps the bike (Goodhart's law: a measure that becomes a target stops measuring well). Push it up and each rider learns from the other. With releases on, the AI is frozen between updates, so the human does most of the adapting in the meantime.

The AI rides as a pelican, in honour of Simon Willison's habit of testing each new language model by asking it to draw a pelican riding a bicycle.