@flancian @khinsen @dch I like generating code with llms for learning by for example generating dynamic notebooks around a core that is doing the actual work (and thus deserves greater scrutiny). I don’t particularly care about the notebook code since I can fairly easily review that the “boundary” is legit (by reading the code, but also logging and instrumentation).

There’s so much “peripheral code” that doesn’t require “runs in production with 6 9s” or “properly does the stats required” that however makes the job of getting a high quality core software so much easier.

Reproducible data snapshots, making sharing and onboarding easy, logs and traces, dashboards, proper archival, robust workflow runners, up to date validated docs, all of these require “boring” code yet provide incredible amounts of value.

There’s this premise that if you use an llm you have to go full in and first apply it to the core code, both on the megahype and the anti-ai sides.