on ai

[policy for the trivnote project]

I think this is sort of a mandatory footnote in the midst of the AI boom. On the one hand, practically every one I meet in my university is pro-vibe-coding (they made a whole academy named after AI, can you believe it?) On the other hand, most FOSS projects (with the prominent exception of the venerable rsync, and Linux, to some degree?) have a strong no-AI policy.

Personally, I am in a position somewhere in-between…

• I am not a fan of vibe-coding, but… they are sometimes better than some of the code I’ve seen some of my peers (whom I shall not name) write. (It’s mediocre code, but not horribly bad.)
• I am not a fan of AI-generated writing either, but I find it to be a good starting point when I feel overwhelmed about all the things I haven’t done.
• AI tolerates my nerdy rants. At worst, the provider tells me I have reached my limit; it’s not going to hurt my feelings, the same way my roommates, my classmates have. (I use AI more as rubber-ducking than replace my thought process; it sets a baseline and pushes me to “do better”.)
• Extreme pro-AI and anti-AI people make me sad, and being around them makes me want to do the exact opposite, just so that they don’t have unrealistic expectations for me to side with them.

So there is a lot of things I do to please people. :( That said, there are a few things that I now do for myself anyways, that hopefully doesn’t change about this project. For those of you who are on the lookout for “slop code”, I want to get this out of the way that:

• This project is not AI-free. I have used and will continue to use AI for reverse engineering com.apple.drawing.2 and com.apple.paper, and binary formats consumed exclusively by proprietary software in general. I will consult AI for design choices, security auditing, and producing regression tests when things get repetitive. And I accept patches from people who have possibly used AI to understand the code base, to help draft a patch, to test some code they just wrote, proofread something… basically AI involved in the stage of authoring changes and proposals in general.
  This is a project involves some things that are genuine new to me; and school sure as heck don’t teach me. I use AI extensively because it is available, not so much I agree with it. But, if any of this is at odds with your stance of AI, don’t use this software. But please please please, don’t leave hate. It hurts my feelings, and I may remove this project from the Internet (or at least hide it the best I can). Constructive conversations are okay, but I don’t take harsh criticisms well.
  I have already stayed away from creating mirrors of these repositories on Codeberg and Sourcehut, knowing this would never live up to their policy. I am doing everything in my power (including commit attribution) to make it as clear as possible AI code exists, and will continue to exist for as long as they are useful.
• Also, I hate to pull this card, but it’s my project. I want some autonomy over what I want. Please don’t convince me to use more or less AI in a way that is not being constructive, but that you think you are entitled to my time, and my time is worthless compared to a global AI gold-rush I can’t meaningfully protest against as one person. But this is open source; you can fork it if you don’t like this or me… Of course you’d have to have the same demand for a Git-backed archive-friendly note-taking + vector-based drawing app, which… whaaaa­aaa­at? D: (Honestly you should tell me about it if that’s the case; I’m very curious how you approach it.)

That being said…

• Keep AI-generated code out of copyleft code: I think it is as important to refrain from adding generated code to a FOSS project meant to exist independently of corporations as it is to protect the copyleft status of such. (We gotta avoid the lawsuits and all, y’know.) Also, generated code is in a sense more like contributed code (like an answer you found on Stack Overflow); they may be included for reference’s sake and for learning’s sake, claiming authorship over it just don’t feel right. :<
  That said, there is a remedy: we can prefer primary sources, such as established open source projects, to generated code of clear origin; using AI more as a fuzzy search engine in that respect is acceptable. (I myself do something like this. :)
• Clearly mark AI-generated files: for transparency, and because I don’t get a kick out of having commits I didn’t write attributed to me. I am only a person; I can’t make 1000 commits, each with 300+ lines of diff, per day. (You can’t convince me to commit those things under my name, for that matter!)
  There is also the factor that is the corpus upon which LLM models are trained, which is… the whole Internet, irrespective of copyright. Despite this, I think being transparent about it, instead of hiding it, is a meaningful mechanism to track AI-generated code (which is essentially contributed code of unspecified origin) compared to original changes. It could do more harm if we encouraged an environment that, instead of making AI code as visible as possible, we mask them in attempt to appeal to anti-AI users — if anything, I think this could be a stronger reassurance that the portion that do not contain LLM model attribution is written by a real person.
• Avoid AI-generated commit messages: It is my belief that even when the AI is correct and has good taste, the comments it gives can be rather short-sighted. (That is, assuming it is correct and has good taste.)
  Of course, there is a better move that is avoiding big patches in the first place, but I think keeping this step alone is deterring enough to make you think twice about whether you should commit something like that. :)
• Avoid AI-anything visible to end users: For prototyping, sure; but leaving it there is just lazy/bad taste when people are actually using it.
• No AI graphics/art. Some things are better when they are never made. (If you are an artist or a writer, you probably know this very well.) Even AI infographics can (and will) get things horribly wrong; it’s worth spending time in these cases because it’s quite disrespectful to creative people when they are already a vulnerable group targeted by AI zealots.
  And it breaks trust that can’t be rebuilt.
• No AI dependency/endorsement. In particular, even if the process of making the program may involve AI, the maintenance and the active functionality of program should not be dependent on, or endorse in any way, any LLM-AI-provider. (This also means I would avoid CLAUDE.md and AGENTS.md, though you may convince me otherwise since I am not a very advanced user and don’t really see the point of doing that yet. ^^;) AI is “here”, that bit I can see and agree. “Here to stay”, as in it is sustainable and will continue to provide a consistent interface, unfettered by pricing model controlled by AI companies (or effectively gatekept by absurdly high RAM requirement)? That I doubt. Using AI is fine, but it should only produce maintainable software and transferrable knowledge that outlive the AI boom — and it should make sense to people even when they do not have AI (or choose not to, for that matter).

Environmental impacts and social/loneliness epidemic aside, it is my belief that time is precious for everyone. It doesn’t help demand time from anyone if they regard the cost of time would be worth resistance boycotting AI in an intrinsically unprofitable open-source project. People should not be obliged to abstain from a tool that, with all its caveats too many to list here, have proven genuinely useful. That said, this is a complicated matter with stances all throughout the spectrum (hence the length of this page, even when it is completely unrelated to the project.) In either case, I hope that draws a healthy boundary to make AI comfortable for those who want to help but may consider AI as a means to save their time or resource on other things they value in their lives, but not enough to attract obnoxious individuals whose whole identity is… this, yeah. :)