It's because they've been in FAANG silos with near zero experience in the pre-SDLC (basic research, practical application phases) through to the post-SDLC (tech transfer, commercialization etc) with most having near zero direct experience inside actual QA teams (with QA being beneath a lot of engs). A true AI pipeline hums when this is real end to end cycle. One of the reasons for that is based on how well it manages many-to-many connections inside vector space. One more thing, most soft devs have near zero comprehension of how language models work so leveraging them to manage complexity is very foreign with most not understanding that language models apply far beyond just human language and into any sequence of data e.g. DNA/RNA or just about anything you can chop up. This kind of scale is also a new concept to most.