AI Can Design Faster Chips — But Does It Really Understand Them?

Tech companies increasingly claim their AI agents (AI that can take actions on its own) can design computer chips — and sometimes they really do. But a good result doesn't prove the AI understood the machine. It might just be turning random dials until the score goes up, like a rat learning which lever drops food. That distinction matters, because only real understanding carries over to the next, different chip.

So researchers built an experiment called AutoTuring. They handed the same AI the same 15-part chip design space twice. The first time, the settings had real names and explanations — fan speed, memory, that sort of thing — plus readouts from a simulator. The second time, those same settings were just anonymous numbers between 0 and 1 with no meaning attached. Everything else stayed identical, so the only difference was whether the problem meant anything. The gap between the two runs is the measurement.

The results favored meaning. Working with real labels, the AI beat a simulated Nvidia H200 chip by 5.4%, and beat its blind twin by 12.3%, while making 70.1% fewer simulator calls. But here's the twist: adding a 'critic loop' — where the AI double-checks its own choices — closed most of the gap for the blind version, and gave the informed version nothing extra. In other words, knowing architecture and checking your work seem to be substitutes for each other, not partners.

The authors are careful: only five to six runs per setup, on one simulated chip. Still, the point is bigger than chips. As AI gets handed harder jobs — medicine, engineering, finance — we need to know whether it's thinking or just lucky. Good scores alone won't tell us.