These articles tend to overstate “using AI to design” technical systems when conceptually they’re really fuzzing the performance characteristics of different variations on an algorithm against a reference implementation, and validating its correctness.

I’m coming from software so I’m not sure how much simulation/verification can be done automatically in chip design, but I’d bet money this sort of process is what they’re talking about, similar to the alpha evolve paper from a bit ago. Rather than going “hey chatgpt how do I design this chip??”