Building with AI: The Compound Advantage

I built most of this site through Claude Code as a deliberate workflow: the components, the type errors that blocked deploys, and the notes themselves all went through the same loop of prompting and review. The obvious claim to make about that is speed, and the best available evidence does not support it. METR ran a randomized controlled trial on sixteen experienced open-source developers across 246 real tasks in their own repositories and measured them 19% slower with AI tools than without. Those developers had forecast a 24% speedup before they started, and once the work was done they estimated they had been 20% faster.

What changed for me was the filter. When a first version costs little enough, an idea stops being killed at the worth-building stage and starts being killed by contact with something running, and the second filter is both later and far more informative, because a running version answers questions a plan cannot. The compounding lives in the attempt count rather than in any single task, which is exactly what a stopwatch on one task cannot see: it measures the task in front of it and says nothing about how many got tried. METR's own February 2026 update reports that developers are likely more sped up now than in early 2025, while selection effects leave the size of that increase uncertain. I have no measured attempt count of my own to put against theirs. Count the things you tried, not the speed you felt.