mastodon 4.7.3A Japanese developer and incoming UCLA Data Theory student. Interested in Open Source, Nix, and Zig. I share some thoughts and my writing.
Hi there, Fediverse! I'm Koichi.
I'm a Japanese Data Science student in the U.S, and an Open Source enthusiast. I'd like to share my project progress and my writing. Hope to connect with others who share the same interest. I'm interested in and learning Zig, ML, Linux, Nix/NixOS, and RiverWM. NixOS is pretty tough, but I think I need to keep learning patiently. It's pretty fun and at least working. Look forward to sharing more!
This was so refreshing to read, and it made me love Zig even more as a learner. From my reading, I think the essential distinction is whether the contributors actually think about and acknowledge the changes (and the design of those changes) they're making.
In light of that, the AI ban itself isn't the "perfect" borderline, I think. It's because there are a lot of developers who can utilize AI as a tool, driven by their own thinking, decisions, and understanding (using AI for searching parts of code, online documentation, and less hallucination-prone code generation with good human review would be good examples). So I understand Zig's policy as having a clear boundary on that, even if it's not perfect as intended. But at the same time, I can think of two things for software development in general (not directly to Zig's) with great respect and love for the current Zig's AI policy.
One is that it might be okay to allow the core development team to use AI since they're trusted not to misuse it. Second, it's more intrinsic if there were a good metric to deterministically check contributors' understanding. Not perfectly aligned with the second point, but letting them write the PR comment in their own words could be a more mitigated solution than a strong AI ban, if I allow AI use while upholding the philosophy I felt was behind that.
I'd be glad if anyone could share their thoughts on this.
If you use terminal coding agents, what complaints do you have? I want to build a new, highly lean and extensible one in Zig and Lua. I'd like to hear as many opinions as possible.
I've been using OpenCode and recently tried Pi. I like the idea of Pi, but it's in the npm ecosystem, so there's a security concern. It's minimal but made for the author himself (as he says it's not entirely bare-bones), so it's still opinionated (tree, HTML share, etc.). I highly respect the two agents and their maintainers, and I love them.
I just found out that OpenCode’s web and desktop interfaces can display LaTeX. It’s just a web view, so it works for sure. Now I no longer need ChatGPT Plus for my math/physics learning and only use Kimi K2.6, etc., and store all the history locally! It’s still pwd-based, so I can let the LLM summarize progress in Markdown, and can git-track that repo!
Since the last post, I've been using LLM for researching and learning, while actually reading the websites they cite as needed; not vibecoding. It’s genuinely great, and this setup only costs $10 a month with a bit of Exa web search API as needed! Love to open weights models!
What’s the best use of AI in software development?
I know it’s a huge and “hot” topic. Some prefer artisanal, and some are fully addicted. As a learner myself, I prefer to set a clear rule: If you can write the code or think of the structure needed, use AI to shorten the time typing. Prototyping, non-essential projects, and tasks that require specific knowledge and searching are the only exceptions. An example of a non-essential project would be a personal website. I’m not interested in front-end development, so I let them do the CSS and stuff (I gotta admit this is vibecoding)
If I don't know what to write and only ask AI high-level questions, it simply means I can't review and refactor the code. It’s easy to be AI-pilled, and it’s easy to be anti-AI, but we should find the best use of it as a tool.
I think we're almost at the point where, if we don't vibecode and use an LLM as our extended keyboard, we don't need the frontier models. The more intelligent the frontier models get, the lower the cost of the same level of intelligence we need for such a use case.
GPT5.5 Low, MiMo v2.5 Pro, and MiMo v2.5 are the ones I love using, but no matter what models they are, we've reached the threshold, and the price is gonna get lower. Vibecoding, coding without learning, and throwing away your thinking are the HUGE exceptions for that.
Neovim is getting its built-in package manager, vim-pack, in v0.12. The boundary of what to make built-in might be difficult for projects that take a plugin-first approach, like Neovim. Stability is certainly a benefit, and some might say it’s core functionality, so it's better to have it built in. But it might take away the opportunity for competition, and it might be better for the project to provide only the basic APIs. I think vim-pack is a move, given they’re trying to make things more out-of-the-box. Probably I'll migrate when 0.12 is out