Yeah I'm in an adjacent field (FPGA) and I'm not sure what they're talking about. The "non-code" parts are even EASIER for AI. For instance I just finished a requirements sprint. Half a dozen engineers over months for something that could probably have been done by GPT-4 with a DOORS plugin in an afternoon. We'd have a review to validate its work but that's still a MAJOR disruption in manpower need.

I think there's a big disconnect between how it works right now versus how it will work (or even can now if you set it up right). The next version will do it more or less how we do it. It's not a 0 shot approach. It'll write the code, compile, fix warnings and errors, and then write unit tests to validate the code and hand you the work package.