What's hillarious is people vibe coding massive AI apps thinking its not creating tech debt and security holes. If anything, AI is discovering existing holes, and creating them faster than ever before in history.

All those "paying down" technical debt with AI, I would advise reading the code _very_ carefully. Sometimes the "solution" which passes some test or adheres to some described functionality is incredibly narrow / brittle, and/or has failure cases that are non-obvious.

Yes I'm using $latest_model on high, its still not there. Unless your threshold for "fixed" is "the 1M line lean proof runs, we don't know how it works", it still requires a lot of attention, hand-holding, and scrutiny, especially when fixing issues / working around failures (where it tends to shortcut or invent workarounds).

A developer on my team introduced new features behind feature flags despite concerns about whether we should have those features and how they are going to be maintained. Changes touched 200+ files which is half of the files.

Turns out nobody wants to use them, they caused many subtle bugs in existing features and is a distraction for many things we do.

The developer left the team, and I am left with purging all of this.

The same can be said with human-written code. LLM code is by far better than the average developer’s code at this point, and there’s not enough top percentile developers to go around, so it’s a wonderful tool.

Also, those very average developers are also the ones doing the code review.