Spotting barrels when the ammunition does most of the visible firing

Yesterday’s post was about barrels and ammunition. This one is a related thought.

Chip Huyen on Lenny’s podcast last week said something that fits right next to it. The engineers seeing the biggest gains from AI tools are the ones who were already high performers. AI did not flatten the curve. It steepened it.

Which is just the barrel thing again, is it not. The tool multiplies what was already there.

Gergely Orosz had a take along the same lines. Senior engineers take to multi-agent work very naturally, he said, because they spent years overseeing parallel work as the go-to code reviewer. The instinct does not come from Claude or Cursor. It comes from the reps before the tools existed.

The identity problem

So here is what I keep getting stuck on. If AI is a multiplier on existing skill, the people with skill pull further ahead. Fine. But it also means two PRs can land that look identical, compile identical, pass the same tests. And only one of the authors actually understands the code.

Output used to tell you who the barrels were. It does not anymore. It is the floor now, not the ceiling.

So how do you spot a barrel when the ammunition does most of the visible firing?

Where I have landed so far

The signal is in the path, not the artifact. Did they push back when the spec was wrong. Did they notice when the model confidently solved the wrong problem. Did they ship something rough and iterate, or did they wait for someone to tell them what good looks like.

The engineers who treat AI output as a draft to be scrutinized produce better outcomes than the ones who treat it as an answer to be accepted. That distinction is invisible in the final diff but obvious if you watch the process.

This is genuinely an open question for me, especially for folks doing hiring loops right now. The old signals are breaking down and the new ones have not fully formed yet.