"Read 2000 lines (50 pages or so) from a book in an afternoon? Sounds like a wonderful and relaxing day."

Sure, because you're not checking it for errors. 50 pages of copyediting an academic book about economic history? Not a relaxing day.

> People (myself included) are exceptionally bad at writing software. We're so bad we've constructed an elaborate Tower of Babel

It's why software is an aggregate effort. Most software is bad but the good software becomes part of the stack and enough people contribute to this stack over time to create the systems that run the world today.

I don't think AI will be able to have this kind of long foresight to create solid building blocks of software, even at a large app level. This is one aspect agentic AI that is being underestimated. A lot of people are calling it "taste", but it's really engineering and knowing what will constitute a solid building block. Without this AI is providing a kind of rocky mine where you go in and mine for good enough software. You still have to sort through the rocks.