I (and many people) have been thinking a lot about "what are the best tasks to do using AI." My personal framework I've been using is, you should use AI to generate code when: 1. The code is easy to validate 2. It is not important that a human understands it

I find that the best thing to AI generate is something on the level of a pure function (easy to validate + not necessarily important to understand implementation as long as you understand the interface). I've tried doing things like generating whole services or applications, but those often violate rule 1 - it's hard to validate an entire application behaves "correctly" when "correctly" isn't really well defined - i.e., are there memory leaks, is it secure, is it able to be monitored, etc.

I'm curious of others thoughts on this topic. The larger the task you complete using AI, the more time savings, but also the more likely that a catastrophic error exists in the generated code. A side effect of following these rules is there is not that much of a productivity gain - maybe 20-30% at most.

I'm really interested if anyone has found a way to strike a balance here - significant speedup without losing correctness.

For me, I feed the local LLM all the regexp/bash one-off scripts. Review everything because rm -rf is only funny at most once, if ever.

[dead]

A fun first project would be something to cultivate bitter almonds, and then isolate the cyanide compounds from them.

I don't know what you'd do with the cyanide compounds. But maybe you could get some inspiration from the current attitude toward "AI" products.