So it feels like now we've sort of gone full circle! Curious if anyone else has noticed this as well.
This is fascinating because maybe a year ago I was prompting for this approach explicitly because the models at the time struggled with doing large scale refactors but still generated small scripts reasonably well.
The most interesting case is when the question hints at something the product should support but doesn't yet. Dangerous because even when the ask is real, the asker can't articulate what they actually need. Waiting until multiple teams ask is usually worth it.
Over the last 10 weeks I've had more security bug reports for Perfetto's trace processor than ever: 17 real bugs, all found by AI. Mostly mechanical, a few nudged overdue cleanups. A year ago none would've been caught.
Wrote up how the experience at
lalitm.com/post/perfett...
For years I've followed a rule at work I never bothered to write down: when an engineer asks me a "weird" question, I don't answer the first version of it.
Wrote about this at lalitm.com/post/dont-an...
One weird thing I've observed with Fable vs Opus: whenever I get it to do a refactor it seems to reach for writing as-hoc Python scripts instead of directly doing a refactor.
It sounds like the XY problem at first, but I think XY stops one step short. It treats the wrong question as a puzzle to decode. The confusion that produced the wrong question is usually worth more than the answer would have been.
