Sorry for swooping in this conversation as a regular contributor. >> From my understanding (what others have told me), AI generally does not >> produce good quality code though. So how is that a benefit to society? > > Well, in that case, those “others” are using them wrong or are just spreading > second-hand misinformation.
While all my previous code, translation and documentation contributions to KDE didn't involve "AI" (here I mean LLM-based stuff actually), I did experiment with some "AI"-based utilities in my own projects and some other projects that allows developers to use this sort of utilities and see if they are actually helpful, and turns out it actually is (in most cases). To put some examples, my experiments are: - I use sourcery.ai for code review with my 100% hand-written code, and it indeed catches some issues that are easy to miss out, especially when writing custom Qt proxy models which is pretty tricky and hard to get it right. - I use LLM-based code copilot utilities to simplify my existing hand-written code, and the result is also quite good, sometimes it will even let me learn some useful APIs that I previously didn't know existed at all. - I use ChatGPT and Gemini to ask them questions about the field that I was previously unfamiliar with (e.g. polkit, GTK stuff), while sometimes it will provide inaccuracy result but at least it will provide a direction that what is the next stuff I need to check out, it saves tons of times than manually reading and understanding. - I even used Gemini's Cavas feature to help me write a standalone User Script for once recently with my very own requirement, and the result is also actually okay for me, which usually will take me 2 days to hand-write it from scratch since I am not that familiar with modern JavaScript programming. While I rarely use LLM to generate large amounts of code for my real projects due to licensing concerns. - and yeah, I am not a native English speaker, I do use "AI"-based utility to help me understand documentations and papers in English, and damn the translation quality is way better than a traditional machine translation tool [1]. [1]: and actually, traditional machine translation tool can still count as "AI" as well, even thought it's not based on "LargeLM". In conclusion I do think "AI" can provide good quality code, code review, and other workflow improvements, so I think what we need to do is focus on the stuff that we really care about, which is licensing. I have to agree with: > Likewise, we have concerns about licensing, quality of contributions, wasting > contributors' time, excessive use of computing power, and more. "AI" is at > best a proxy for these, and at worst a wedge issue that can chase away valued > contributors like Jin while not doing KDE much good otherwise. Also, I'd like to comment on "excessive use of computing power". For LLM, training and/or using a model is indeed resource-consuming, but things do improve over time. For example, The Qwen3-4B model's quality is way better than an older 7B or even 14B models which consume significantly less resources, and such technology is still continuing to evolve. That's my 2 cents, I'll keep avoiding using any of "AI" utilities when contributing to KDE projects (and other open-source projects as well) before we have a consent about "AI" usages, and I hope this reply could provide some perspective to people who hardly use "AI", or say, LLM-based utilities.
