I'm tired of "AI," but there's more to say
I'm tired of reading about LLMs, but I'm stuck with them anyway. Without further ado...
Code review is usable
I don't know that I could call this a killer app, but review in Codex is good enough that I can see it staying around for the long term. It's a quick way to get a different perspective. The task is so limited, compared to design, that there's much less room to get off track or do something completely broken.
Codex seems to do targeted reviews well; Sashiko's problems with "bringing up other bugs that could be fixed" (as reported on LWN) are her own.
Providers can pad the bill
Charging for output tokens creates a clear conflict of interest. Tune the model to be a bit wordier, and more money comes in. The expensive part was all the inference along the way.
There's another opportunity. Make inference intentionally mediocre, and let users send more input over more turns to generate more output. Ka-ching!
The deal may be altered at any time
New models are constantly coming out. Even randomization (temperature) changes the result from run to run. I simply cannot fathom how anyone could build a stable, trustworthy process on top of a model.
Search remains a terrible application
If there's any chance the model can be wrong, then it is simply unfit for purpose of searching. How am I going to validate the response, if not reading the actual search results? (And evaluating those for slop, of course.)