It is pretty simple imo. AI (just like humans!) does best on well written, self contained code bases. Which is a very small niche, but also over represented in open source and subsequently by tech celebrities who tend not to work on “ugly code”.

I work on a giant legacy code base at big tech, which is one piece of many distributed systems. LLM is helpful for localised, well defined work, but nowhere close to what the TFA describes.