Similarly, some people are championing agentic RTL design. But if the agent spits out a 1000 line RTL what is the confidence level? 50%? 70%? 90%? Manufacturing is expensive and a single defect/bug is even more so.
There is a rise in startups that are developing hardware CAD tools with AI, e.g. PCB autorouting, ASIC physical design, etc, etc. Almost all of them will fail because they are trying to solve a combinatorial problem using AI which is non-deterministic.