AI generated microarchitecture could be fun, although I assume they do something similar already (What's the optimal cache parameters, number of execution pipelines etc. etc.)