@fvsch
I currently don't have the luxury to have a strict "AI, not even once" policy too (at work).
And as a former cogsci/AI student, I'm too interested by the topic – but disgusted by the industry.
I'm in line with the "extra background" too (lot of moral self-reflexions), and try to keep a critical and attentive posture.
I use GH copilot at work since the private beta, it's not a bad solution if you use VSCode:
- you can try a lot of models
- it's not too expensive
- you can cap your consumption
- you can hook local-running models (but to have an efficient local setup, you'll need a powerful machine, think "beefed-up gaming PC/Mac Studio/Pro " with loads of RAM and GPU cores).
It's not that "easy" to use, and being efficient with them is a bit far from the hype and inflated discourses you see online.
Andrej Karpathy (who I find always reliable and honest, despite being in the "pro" camp) summed it up well here:
https://xcancel.com/karpathy/status/2015883857489522876
Carefully craft your project's context file (claude.md, copilot-instructions.md, whatever).
Stay away from hand-off tools/workflow (often promoted by CLI tools), keep yourself in the loop, always challenge the answers and beware of models idiosyncrasies.
My most efficient uses are mostly:
- brainstorming partner / answering rubber duck, when I feel stuck or want another take on conception and debugging
- exploring unfamiliar or legacy codebases / tools (e.g. projects I didn't work on since several months, infra/ops tools)
- smart autocomplete for "glue code" (I always start by writing myself tests or types / typechecked skeleton first - it helps tremendously).
