[jvanderbot]

On your "Scares the shit out of me" comment.

Use AI like a CNC machinist uses a mill. You're still in the loop, but break it into manageable "passes" with testing touchpoints. These touchpoints allow you to understand what's going on. Nothing wrong with letting AI oneshot something, but it's more fun and less ennui to jump in and look around and exercise some control here and there. And, on larger systems, this is basically required. (for now, perhaps).

This is how I do it now: https://jodavaho.io/posts/ai-useage-2025.html

[vardalab]

Exactly this! I am a retired EE just messing around with AI in my homelab datacenter and that has been my approach as well. Amazing force multiplier, I can finally create more or less what I want with sw based on first principles and basic systems engineering approaches by just guiding AI. I have used golang and ansible and terraform and typescript, languages I never had time to learn and now I can create a working tools/solutions to whatever is my need at the moment. Other day my STT subscription app became too laggy so I asked Claude to spin up an endpoint on one of my GPU boxes, create a proxy server, intercept transcript cleanup calls and create traces in langfuse, setup a prompt eval framework, etc. We make a plan, iterate on it, I usually get Codex or Gemini in on the call as well and in a couple of hours I have a good enough solutions for my personal needs. This would have been probably a weekend or more project before. Skepticism here does remind me a bit of when I learned how to use handtools for woodworking. Ultimately it was nice to be able to make mortises by hand with chisel but damn if using festool dominoes is not that much more productive.