I've been a late starter with AI programming, partly from bad experience (shit/annoying code) and partly ideologically (privacy/sovereignty mainly). Now I'm trying to adjust to using it responsibly. I feel like it's finally pretty reliable for a lot of problems. EG I don't think I'll ever have to waste my life fucking around with Nvidia drivers ever again, which I'm glad to say. I also don't think I'll ever have to manually program matplotlib graphs ever again. I would feel like an idiot if I was making basically any simple/disposable code or experiment manually in 99% of cases. It can do them basically unreviewed, but I get antsy letting it do more and more complicated/permanent stuff.
On the other side, I definitely feel like I need to be waaay more intentional about learning. I don't automatically learn the codebase by outputting working code anymore. That avenue for learning has just completely disappeared which is kinda horrifying. I need to actually sit down and dedicate time+effort to reading PURELY for the purpose of understanding, not to solve a specific problem in mind. Prompting for explanations can help a little bit but I don't feel like it's enough, i need to write my own notes/guides/resources etc. Similarly on the other side I need to actively review "my own" code.
Even for hobby code, normally I would just push everything immediately bc I knew exactly what was there and I could write a commit message on the spot. but not really anymore sadly. It also makes some of the debugging a lot less satisfying, sometimes when I get into a rut with claude code i'm just fucked or I have to abruptly leave autopilot because I don't have the knowledge to get out of it. But overall I think it's still (probably) a net positive in terms of satisfaction since I like the result and I can try more ideas (ML has a lot of trial-and-error experiments).
I bought a remarkable paper pro to help me take notes and read ML papers. That's been quite pleasant and has brought back some of the pre-AI feelings/creativity, so I can recommend that (you could probably just buy nice paper/journal if you took more effort). Before full-auto AI like claude, I felt like my codebase was my "territory" that I could paste "approved" code into (I'm a youngin so I didn't have much experience in company codebases until recently lol, I think that bubble was going to burst anyway), but now it doesn't really feel that way. But having the paper/notetaker still has that feeling of sovereignty.
I also put claude into a rootless vscode devcontainer which has eased my paranoia about AI looking at the rest of my computer etc. but when I deal with nvidia drivers I still take my chances with AI lol. I looked at Qubes for more compartmentalisation but ehh idk maybe someone in the comments can tell me their experience with it. i'm on debian at the moment and I like minimising issues especially with hardware lol, i heard compatibility was a big thing for Qubes. I'm also lowkey scared to open my monero wallet on my desktop since running claude code and opencode outside the devcontainer ngl, not sure if there's any tips for that either. my phone has no AI on it anyway so the cakewallet is chill. I don't have lifechanging money in it anyway but still.
anyway here's a completely jacked dude I had on my computer
I'm not doing a whole lot of coding yet with AI, but I am starting to run local AI, such as running
Ollama with qwen3.5:2b. My main use case is just a quick chatbot, and this seems to be about the minimum possible to get that experience. According to gemma 4 12b (run via nano-gpt for local use you hit diminishing returns after ~30b for a conversational model like i want. I have submitted a couple of features to the meshcore open app and gotten them accepted but I was using Kimi K3 for that and I was using it from nano-gpt.