The issue is that the downsides are massive and the upsides pretty mid.

Oh look, I can whip up a pretty generic website or utility from natural language. That's.... something I guess... The more... ambitious applications of CodeGen I've found to be "different" rather than "better", a different huge pain in the ass than the usual huge pain in the ass, and it's a matter of taste which one someone prefers, and I don't think the "All-in Agentic" flow actually is any faster when you get at this "pain in the ass" point.

On the downside, well pretty much his whole post.

Note that this specifically is the GenAI stuff, broader AI like machine vision is arguable more balanced (intrusive surveillance, but also useful for medical imaging, spotting wildfires early, stuff like that).

So I guess the concern is that by saying "if not for all the bad, it could be pretty good" may be fine, but in practice is used to provide cover for a lot of the bad, because they would argue to not risk throwing out the baby with the bathwater.