This is a post about how I personally use AI. This is not a post about the ethical/societal/environmental issues. This is not because I do not think there are any, but because I do not wish to talk about them in this post. This is also not necessarily a judgment of other people's usage; these are my personal preferences. Conversely, do not read this post to imply that I condemn or commend things that I do not mention. That being said.
personal software projects
No usage for code. I don't have any personal projects that I want badly enough to overcome my dissatisfaction with the feeling of non-understanding that using an agent would give me. I avoidusing big frameworks for the same reason, a lot of the time; e.g. in my React projects I never used create-react-app. This is also one of my biggest complains about Nix as infrastructure, and why I tend to like the "assemble your desktop environment" model of Linux systems as opposed to more "batteries included" ones like KDE or Dank.
I would consider using a local model (we have some pretty decent hardware at my apartment), or possibly an open-weight one run remotely, if I had any large scale automated refactors to do. But I don't.
I do, however, use the free Claude model for things like coming up with names or "give me some ideas for how to structure this". For my current project, I asked it how to have a Nix package with a Rust frontend and a TypeScript backend with the compiled JS embedded into the binary, but served separately during dev time, and the technique it gave me basically just worked. And now that technique is in my personal repertoire, and in my notes.
personal writing
None. It violates the purpose. I would possibly consider passing it through a spellchecker/grammar checker solely for the purpose of catching unambiguous errors (as opposed to stylistic choices).
An observant observer may notice that the diction of this post is quite different than my usual. This is an intentional choice, although the reasons for that choice are somewhat opaque to myself. It is, in any case, not a signifier of a difference in process.
professional software projects
This is where the "large scale refactors" from above comes in. I've thrown agents at tasks that are not amenable to AST-based manipulation, such as "in this folder, find all functions that pass in a Foo, Bar, Baz, and remove the Foo and Bar parameters, accessing those values from the Baz instead; fix all callsites.". This is not a productive use of my time to manually do, and doing this makes the interesting work easier. I've also used it to quickly check whether a couple ideas were viable, generating maybe 100-200 lines of code each time.
In all cases I manually reviewed the code and would consider myself fully responsible for any breakage as if I had typed it myself, of course.
I would not use these models for building out nontrivial larger projects because I believe that skill would atrophy, and because reviewing code is always less pleasant than writing it.
professional writing
Very little. I've passed documentation through an LLM for clarity (since my personal desire for style is somewhat out weighted by professionalism here, but only somewhat). I wouldn't use it when talking with people. I also once had a longer document run through an LLM against my wishes, which annoyed me, but I didn't find fighting it to be a good use of my time.
personal image generation
I have used it for generating a couple avatars as self-representation, typically with a 1-2 hour amount of "effort" each time (including time spent in an image editor doing manual postprocessing) with a 50% duty cycle: roughly half the time was spent waiting for generation. This was with a local model several years ago, so if I did this again I would likely spend much less time waiting. I've also considered using a local model for generating computer wallpapers. I've used "corporate" models for quick "I want a quick visualization of this character doing this thing" images, none of which left the context of a Discord DM or similar.
I am intentionally not addressing whether or not any of this counts as 'art'.
socialization
I have talked with Claude on one occasion about a particular mental hangup I had been bothered by for quite some time. The advice I received was useful, and at least partly driven by the process of speaking the anxiety, in the same style as rubber-duck debugging or tarot. I would not tell a non-local model anything I would not feel comfortable in a training set, and anything "severe" I would not consider a local model to be sufficient for.
I find the idea of constructing a "personality" that a persistent model would inhabit to be interesting, but I would require local control, and I suspect that there is a nonzero chance I would develop some kind of emotional attachment in excess of what I would consider healthy. So I haven't experimented with this.
conclusion
AI is a land of contrasts.