AI reaching GDS is a capability demo. AI predicting the next EDA result, measuring its error, and beating a trusted baseline across repeated runs is engineering evidence. Moh Kolbehdari's proposed framework gives semiconductor teams a practical boundary between autonomous exploration and conventional signoff.
The sequence starts with the target workload, derives compute and memory behavior, defines a comparable engineering benchmark, structures exploration with design of experiments, fits a response model, predicts the next result, and verifies it in established EDA tools. The useful output is not the best frequency or smallest area from thousands of attempts. It is margin, prediction error, outcome distribution, and the shape of the power, timing, area, congestion, and thermal trade space. This removes the selection bias that lets one successful tapeout stand in for a reliable design process.
Teams evaluating autonomous EDA should reserve one representative block and a conventional signoff baseline, then require the agent to expose its full design history and held-out prediction error. The cost is a controlled evaluation before production access. AI EDA vendors that only show the winning run should not reach the production flow.