yeah he's really underselling the recent models' ability to do the "Designing the solution conceptually" part. I still have to be in dialogue with the AI - I ask a lot of questions, we iterate towards something - but I cover conceptual ground much more quickly this way. and then I still have to be the glue between "design" and "writing" steps, and have to manage that carefully or I get slop.

if you look at the change in capability over time, it looks like the AI are climbing this hierarchy. "Centaur" seems to already be giving way towards "research assistant". I hesitate to make predictions but I would not place money on things stabilizing here.