LLMs still seem bad at making some key "give up and start from scratch" call that seasoned software engineers are better at making. Latest example: I was implementing (w/ Opus 4.8max) a feature to add offline mode to an internal app. I did a PR, it got mired in an endless review-patch cycle with codex and claude.
I started a new PR: asked it to look at the previous one, learn from the revisions and comments, and start clean slate. It still got kind of stuck. I nudged "try a state machine for X", that made it much better. Still, it needed a third clean slate PR to get it done without bugs.
LLMs still seem bad at making some key "give up and start from scratch" call that seasoned software engineers are better at making. Latest example: I was implementing (w/ Opus 4.8max) a feature to add offline mode to an internal app. I did a PR, it got mired in an endless review-patch cycle with codex and claude.
I remember when I was writing software by hand in ye olde days that there's a feeling of being "too locked in" that feels bad and then you take the broader view and go like ah yes I was trying to make something work that's just a bad structure, let's try something new. This move is something LLMs are not good at. (something something metarationality)
I started a new PR: asked it to look at the previous one, learn from the revisions and comments, and start clean slate. It still got kind of stuck. I nudged "try a state machine for X", that made it much better. Still, it needed a third clean slate PR to get it done without bugs.
LLMs still seem bad at making some key "give up and start from scratch" call that seasoned software engineers are better at making. Latest example: I was implementing (w/ Opus 4.8max) a feature to add offline mode to an internal app. I did a PR, it got mired in an endless review-patch cycle with codex and claude.
Users recall feeling too locked in when coding by hand to argue that LLMs are not good at making start-from-scratch decisions in software engineering.