Building in Public #9: The 133-Second Problem

We spent this week trying to make builds faster, and the most useful thing we found is that we'd been staring at the wrong number.

The obvious theory: builds are slow because the AI writes a lot of code. Less code, faster build. So we measured it properly, across 113 real builds timed end to end.

The theory is mostly wrong. The relationship between how much the model writes and how long the build takes is weak. When we fit it, the line comes out at roughly 133 seconds of fixed cost before a single character of your app exists, about 62% of the average build. Halving the amount of code written (which would be a big engineering win) would buy back about 28 seconds. The other 133 are something else: queue time, model start-up, thinking time. We're instrumenting each of those now, because you can only fix the part you can see.

133 seconds of fixed cost before a single character of your app exists. About 62% of the average build.

Two things I like about this result.

First, it killed a plan before we built it. We had a whole workstream ready for pre-building common components so the model writes less. Measured against real builds, that would have saved almost nothing. The boilerplate it targets is 0.3% of output. That workstream is cancelled, and the week it would have taken goes to the 133 seconds instead.

The boilerplate it targets is 0.3% of output. That workstream is cancelled.

Second, it was found, argued, and settled by the AI fleet itself. Three separate analyses, each done independently, all landing on the same answer. My job was reading the reports.

It was found, argued, and settled by the AI fleet itself. My job was reading the reports.

Build speed matters because iteration speed is the product: you describe a change, you see it. Every second in that loop is felt.

Iteration speed is the product: you describe a change, you see it. Every second in that loop is felt.

We're instrumenting the 133 seconds piece by piece, and the first piece is already done. One stage, the design pass's thinking time, turned out to be a cliff rather than a knee: capping how long the model reasons before it commits to a design cut that stage from a 100-second median to 31, with no measurable drop in the design it produces. It's already shipped to production. We don't yet know what share of the 133 that stage accounts for, so we're not claiming one. We'll publish the rest of the breakdown as the other pieces land.