▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=Y8cY1DEURtE
• Original title :: Codeberg, are you serious?!?

Weeks after Theo publicly moved to Codeberg, donated thousands of dollars and urged viewers to follow, Codeberg's members voted 358–144 (≈50% turnout) to amend the Terms of Use and prohibit projects that mostly consist of LLM-generated code. Theo walks through the pull request, the public forum threads, and the announcement blog post "Protecting our open-source commons from LLMs", agreeing with the narrow resource-cost argument buried at the bottom and rejecting nearly everything above it as performative politics. He frames the change as anti-progress for free software and pleads with Codeberg to reconsider, while explicitly telling viewers not to harass any Forgejo or Codeberg maintainers.

What was actually voted on

Two motions passed at the 2026 annual assembly. The first — Codeberg will never use user data to train LLMs — Theo calls entirely reasonable and supports without reservation. The second adds a Terms of Use clause: you must not share projects that mostly consist of code written by generative AI tools, including services such as Claude and OpenAI Codex, on the grounds that such projects have unclear copyright status and few safeguards against harmful code. A separate motion in the same session banned cryptocurrency projects, which Theo — a self-described crypto hater — still thinks should not have been banned.

The banned categories per the blog: projects created autonomously by LLM agents, projects written and maintained with heavy LLM use, projects whose resource consumption far exceeds what the involved humans could produce by hand, projects tied to the LLM ecosystem (LLM-written tools for LLM usage), and contributions that violate a project's own AI policy. Theo endorses only the resource-disproportion one. Explicitly safe: projects with an active community, projects with significant pre-LLM history, maintainers unknowingly accepting LLM contributions. Codeberg also says it will not spend resources auto-scanning content, so small experiments and personal scripts are discouraged but tolerated.

The transparency complaint

The deliberation happened in a members-only forum Theo cannot read; only the linked pull request and post-hoc summaries are public. Codeberg's answer is that it is volunteer-driven and there is no capacity to publish ongoing discussion, and that anyone can join for $12/year to influence policy. Theo argues there is a large group who want visibility into pending policy without becoming members, and finds the analogies offered by a Presidium member — you cannot read Stephen King's draft, you cannot read a lawmaker's unfinished bill — a poor fit for a project whose entire premise is transparency. He points at kernel mailing lists as the counterexample of a project that does all of this in public.

He is harshest on the tone of the proposal itself, which describes LLMs as detrimental to maintainability, backed by corporations that support war crimes, trained on scraped data, environmentally destructive, and burying the web in soulless content. Theo reads that register as punching rather than communicating, and says the private forum likely exists partly because that kind of writing would be embarrassing in the open.

The security and quality argument

Theo's central technical rebuttal: the claim that vibe-coded projects lack safeguards is backwards. He polls his chat on which is more likely to be secure — a random package generated with current frontier models, or one from a random below-average developer — and ~90% pick the AI-generated one. His reasoning is that billions of dollars of tuning have raised the floor for everyone at once, whereas a weak human developer improves only through years of individual experience. He dates his own change of mind to Opus 4.5, saying he would have voted the other way a year earlier.

He also notes that Claude Code and Codex both refuse to build overtly malicious software, while nothing but law stops a human from doing so — and that banning LLM-assisted patching does not stop anyone from using LLMs to find exploits in Codeberg-hosted code from the outside. On copyright, he says that if anyone can demonstrate licensed code being regurgitated verbatim without internet access, that is a strong lawsuit worth bringing.

The environmental and cost sections

Theo dismisses the water argument outright — data-center water use is a closed loop and roughly a thousandth of US corn irrigation — and calls the electricity framing performative given that nuclear appears nowhere in the document. He details Germany's nuclear phase-out (from over 20% of generation to zero), the resulting coal increase, domestic generation falling from 633 TWh to 509 TWh, and the country becoming a net energy importer more dependent on Russian supply. His complaint is not that energy concerns are illegitimate but that framing them as "AI bad because water" crowds out the real conversation about green generation, which he wants to have.

Where he agrees fully: Codeberg's observation that single-developer projects with no users are consuming CI, storage, and release-binary resources comparable to the largest community projects, and that hyper-aggressive AI crawlers strain servers. He says if the announcement had opened with that — here is the share of our CI burned by user-less generated projects — he would have supported restrictions and donated more. Prime's coverage of the same post adds the concrete number Codeberg cited: an SSD that used to cost around $700 now runs about €3,700.

The trust argument

Theo grants that maintainers face increased workload from well-meaning low-effort LLM contributions, and faults Anthropic and OpenAI for not building tooling to help maintainers absorb that wave early. But he turns Codeberg's own observations around: if it is now hard to distinguish LLM-maintained projects from expert-maintained ones, that is a statement about quality convergence; if contributors are hiding their LLM use, the hostility in documents like this one is why; and if genuine contributors are being falsely accused of using AI, the culture Codeberg is encouraging produced that.

On the claim that LLMs push people toward disposable single-use software instead of collaboration, he cites his own AI-assisted open-source releases — Lawn Video (a self-hostable Frame.io alternative) and t3.codes (an open harness supporting Open Code and open-weight models) — as counterevidence.

Takeaway

Theo closes with a horseshoe-theory framing: Codeberg now restricts what he may publish more than the Apple App Store does, which makes it a host of freely available software rather than a free-software organization. He quotes Linus Torvalds — the useful-or-not question is settled, and the answer to AI's costs is to make the tools help maintainers rather than singing "la la la I can't hear you" — and says the contrast explains why few serious projects have migrated to Codeberg. He is disappointed, has no replacement recommendation beyond self-hosting Forgejo, and says he will apply for membership to argue the case from inside.