Codeberg's LLM ban — which parts are actually reasonable
- YT :: https://www.youtube.com/watch?v=RoHzXoolPTY
- Original title :: The Anti-Github
Prime reads Codeberg's Terms of Use change and its blog post "Protecting our FLOSS commons from LLMs" end to end, deliberately against the grain of the internet reaction. His verdict is split: the copyright, trust, and hosting-cost arguments are legitimate, the vague wording and the "no safeguards against harmful code" claim are weak, and the swipe at single-use software is the one part he rejects outright. Underlying frame — a donation-funded nonprofit should not be held to the same expectations as GitHub, and different communities are allowed different rules.
Why Codeberg is in the news
GitHub has spent roughly two years absorbing bad press: downtime, poor UX, action runners, general disdain from ordinary users. Alternatives gained momentum, Codeberg among them. Its members then voted to disallow projects consisting mostly of LLM-generated code, which produced the predictable wave of outrage and donation-cancellation threats. Prime notes that the very next ticket in the same meeting banned cryptocurrency projects as harmful to Codeberg's reputation, and nobody complained about that one.
The wording problem
The new clause is section 2, paragraph 7: you must not share projects that mostly consist of code written by generative AI tools. Prime calls "mostly" unworkable — 51%? 75%? a plurality? — and asks where a hand-crafted project that has since shifted to heavy LLM use with light human edits falls. His example is Ghostty, whose author has been open about extensive LLM use for the past year. Without a line count or a measurement, enforcement reduces to selective judgment.
Where he agrees
Copyright. The strongest argument against LLM-generated code, in his view. Models may have trained on copyleft repositories, and copyleft is viral: inadvertently reproducing that code obligates your project to the same license. Jurisdictions disagree on how this works, and a small nonprofit reasonably does not want to litigate it. He says if this were Codeberg's whole reason, he would fully support the ban.
Trust and community. Maintainers absorbing well-meaning, low-effort AI contributions spend disproportionate review time, which makes them warier of strangers, which makes communities less welcoming and raises the barrier for newcomers who genuinely want in. Prime calls this a real fracture and agrees it is worth defending against.
Hosting cost. Codeberg self-hosts. An SSD that used to be ~$700 is now ~€3,700. Add agent-driven commit volume — the Cursor and Anthropic papers describe thousands of commits per hour, work that once took tens of thousands of engineers — plus constant crawler traffic, and LLMs become a straightforward net negative on the cost of hosting code.
He also flags the no-training-on-your-data pledge as good, with a joke that it makes Codeberg a repository of hand-crafted, grass-fed, artisanal training data that someone will inevitably want.
Where he disagrees
"Little safeguards against harmful code." Software was already full of CVEs and spectacular breaches long before LLMs. He suspects generated code is roughly as insecure or slightly more so, but calls that his personal bias rather than a demonstrated fact. He also points out LLMs are useful for review — with the caveat that Fable and OpenAI models often refuse security analysis, so the Chinese open-weight models end up doing the work of letting you audit your own code, with decent hit rates.
Single-use software. The blog treats disposable, never-maintained one-off tools as a symptom of decay. Prime's visceral objection: single-use software is good and there should be more of it. Software can be done. Most ideas are bad, and cheap disposable attempts are how you find that out. His own example is a Neovim LLM-interaction tool he hand-coded over roughly two months before concluding the idea did not work — in hindsight he should have vibe-coded it hard until he knew whether he liked what he was seeing. Build for yourself, throw away what does not land, promote what does.
Takeaway
The honest reading is messier than either "based" or "worst thing ever". Codeberg has real constraints — a donations-and-membership budget, self-hosted hardware, a volunteer maintainer pool — that GitHub, backed by a multi-trillion-dollar company that uses the platform to train the AI aimed at your job, simply does not. Different expectations are warranted, and different kinds of community are allowed to exist: go to Codeberg for the interaction and to get better as an engineer, go elsewhere to vibe something out.