And I agree. Because ultimately we don’t need that much code in the first place. We need robust data sets.
AI models will enable the data driven machine state dream. Chips that self improve models will boot strap from them and rely on humans to iteratively improve updates.
Coding like it’s 1970 in the 2020s and beyond is not that high tech.
Software developers do X and Y. AI thing can now do X, so it's used for that, and it's cheaper, so the number of projects increase because you get more demand at a lower price. Those projects each need someone to do Y.
And even with reviews you can currently plausibly automate only the code correctness check part, the juicy part of reviews is always manual testing of the change and doing the logical reasoning if the change is doing a meaningful thing. And no, the ticket with the spec is not a reliable source of this info for an LLM as it's always just a partial understanding of the concept.