AI 輔助寫程式,該拿來把常用開源軟體變好用
We don't need more apps, we need better open source
每天都有人用 AI 做出新的待辦清單或筆記 App,但許多開源軟體真正難的是安裝和維護:設定步驟太多、升級容易出問題,少數人使用也可能很吃記憶體,讓沒有專職工程師的非營利組織只好繼續付錢使用大公司的服務。作者建議把 AI 輔助寫程式的力氣用在改善大家已經在用的工具,例如讓安裝更簡單、程式跑得更省資源、錯誤訊息更清楚,並讓備份和升級更容易處理。例如,作者提到 uv 在開發者的測試中,安裝 Python 套件比 pip 快 8 到 10 倍。這類改進能幫到許多現有使用者,也讓小型組織更有機會自己管理資料。作者也提醒,使用 AI 幫忙改程式時,要先和維護者溝通、提出實測結果,並親自負責後續修正;大量品質不佳的回報或一次送上太大的改動,反而會增加維護者的負擔。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:

Every day, a dozen people post a new app they built with AI over a weekend: a habit tracker, a to-do list, another note-taking app. I’m glad! More people building things is one of the best parts of this moment.
But far less of that energy goes into the open source software we already rely on.
People call this new way of building “vibecoding.” I hate the name; it makes a real shift in who gets to write software sound like a joke. I’ll call it AI-assisted coding, and I think it could aim somewhere more useful.
Free and open source software already does most of what my clients and I need. What it often can’t do is run well on a cheap server, install in one step, or stay up without forum threads at 11pm.
Where I’m coming from
I self-host a lot of my own stuff, and I help progressive nonprofits with their tech for a living. I run a Matrix server (Synapse, Element and a moderation bot) for a community with thousands of users. I also run apps on Coolify and Hetzner, keep my paperwork in Paperless-ngx, and own a Mac Mini that’s always one download from a full disk.
The software is usually great. Running it is the hard part. Install guides assume you know five other tools. Upgrades break things. Apps eat 4GB of RAM to serve a handful of people. Error messages tell you nothing.
My nonprofit clients would love to leave Big Tech for privacy, cost and owning their data. But the open source option usually needs a part-time sysadmin they don’t have, so they pay for the corporate tool.
Four jobs worth your weekend
- Make it faster. uv took a job Python developers do every day, installing packages, and made it 8 to 10 times faster than pip in its makers’ benchmarks (up to 115 times with a warm cache). Find the slow part of something people already use, measure it, fix it.
- Make it easy to install. Too many apps need a database, a cache, a background worker, a reverse proxy and a long config file before you see a login screen. Aim for one command, sensible defaults and a setup that works the first time on a cheap server.
- Make it friendly to run. Installing happens once; running it happens every day after. Give it a simple admin screen, safe upgrades, built-in backups, errors that say what to do next, and an easy, secure way to manage logins and add admins.
- Make it lighter. If a tool runs happily on a $5 server, an old laptop or a Raspberry Pi, far more people can use it. Lighter software costs less to run and uses less energy.
Why this matters
A new to-do app usually helps the few people who find it. Making a tool millions already use 20% lighter helps all of them, every day.
This is also the work that slides when maintainers are tired, and many are unpaid volunteers. We saw the risk in 2024. Someone spent years earning the trust of the worn-out solo maintainer of xz, a compression tool inside most Linux systems, then slipped a backdoor into it.
I think this unglamorous work suits AI well: profiling slow code, writing tests that make big changes safe, testing installers, rewriting cryptic errors, writing the missing docs. It takes patience more than brilliance, plus a person who checks every line.
It matters politically too. If organizations are going to own their data instead of renting it from a few giant companies, open source has to be something a 20-person nonprofit can run. I want AI’s gains shared widely. Better public tools that anyone can use, change and share are one way to get there.
Don’t be a jerk about it
Maintainers have already seen the wrong way. The curl project ended its bug bounty in January 2026 after a flood of fake, AI-generated security reports. By April, its lead developer wrote that the fakes were mostly gone and almost every report used AI, most of them very good. The new problem is volume. Even good work can bury a maintainer. So:
- Use the thing first and find what actually hurts.
- Talk to maintainers before writing code. Nobody wants a surprise 4,000-line pull request.
- Bring numbers. “Startup went from 40 seconds to 4, here’s how I measured it” beats “faster.”
- Keep changes small and tested.
- Say you used AI, and own the result. Stick around to fix what breaks.
- If they say no, help downstream with a better installer or setup guide, and keep it current.
Pick one tool
Pick one open source tool you use. Find the most annoying thing about installing or running it. Spend a weekend fixing it, with the people who maintain it. Then tell me about it.
Sources
- uv: Python packaging in Rust, Astral
- The end of the curl bug-bounty, Daniel Stenberg
- High quality chaos, Daniel Stenberg
- XZ Utils backdoor, Tukaani Project
Subscribe to
Mission Control ➡️
I share thoughts on online tools, learnings from my nonprofit advocacy consulting business, projects I'm launching, Artificial Intelligence, and personal updates — directly to your inbox.