▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=h3DlwC1nex0
• Original title :: The Codeberg Situation | TheStandup

Prime, TJ DeVries, Trash Dev and Casey Muratori discuss Codeberg's new terms banning projects that mostly consist of generative-AI code. The panel finds Codeberg's stated reasoning sound — hardware and infrastructure costs, copyright uncertainty, and above all the erosion of trust between maintainers — but doubts the rule is enforceable. The larger half of the episode attacks the "you'll be left behind" reaction: a diversity of approaches is insurance for the whole industry, and people demanding everyone adopt LLMs are arguing against their own interest.

What Codeberg actually changed

Codeberg is a self-hosted, FLOSS-focused code host. Its terms now say you must not share projects mostly consisting of code written by generative AI tools (Claude, OpenAI Codex, etc.), because such projects have unclear copyright status and few safeguards against harmful code. A companion blog post, "protecting our FLOSS commons from LLMs", lays out the motivation. A membership vote backed it roughly 2:1 — with the caveat that only about half the eligible members voted, so it is two-thirds of the voting half.

Reasons given:

• Hardware costs. Self-hosting means they absorb the SSD price spike (~700 → ~3500 euros), and LLM-driven activity eats that capacity.
• Infrastructure load. Huge machine-generated repos with a single interacting human cost disproportionately versus a project like Ghostty where many people collaborate; the expensive repos generate no community.
• Copyright status of generated code is unclear.
• Single-use software. They would rather host things built to be reused than one-off generated tools.
• Loss of trust. Low-effort AI pull requests train maintainers to assume every new contributor is generating slop, which makes them hostile to newcomers and degrades the community.

The trust argument lands hardest

Prime calls the community-erosion point the most interesting and an "unknown known" — a bias he already had without noticing. Anything over three sentences now reads as generated, so he stops giving people a fair hearing. Trash Dev reports the same asymmetry: he spent the week clearing tech debt with LLMs, then got an obviously autogenerated PR on his own project and was almost angry about it, without being able to explain the psychology.

Casey generalizes: this is the standard "I use it responsibly, other people don't" bias, familiar from driving, drinking, gun ownership. Most likely your AI output is about as good as everyone else's, and you are grading yourself up and them down. TJ adds a norms twist — he does not mind trashing his own codebase, but going into someone else's repo is like being a guest in their house, so a higher standard applies.

Enforceability

Nobody on the panel knows how "mostly" gets adjudicated. Trash Dev's worry is that it collapses into reputation: established open-source people are trusted, newcomers cannot prove a negative and get presumed guilty. Ghostty is the concrete edge case — Mitchell Hashimoto has talked openly about generating patches with AI, so what is the slop threshold?

Casey separates two goals. If you only want quality, the rule you need is simply "nobody should be able to tell an LLM did this" — a human reviewed it, it is not verbose, it has no obvious mistakes. Codeberg is going further because it wants to invest in people rather than in the LLM economy — the same reason Andrew Kelley gave for Zig refusing generated contributions: they want future contributors who spent real time in the codebase. Casey's only counterproposal is a monetary one: charge heavy generators rather than ban them, so the policy can be tuned. But choosing not to participate at all is a legitimate call.

"Left behind" is the weak argument

The panel spends the back half dismantling the backlash. The claim is that if most programmers come to work through LLMs, a host that bans generated code will not have enough contributors to matter. Prime traces the rhetoric to a Silicon Valley pattern — in 2021 the same line was used about not building with React.

TJ's counter is that "left behind" presumes everyone's goal is to be at the frontier. Codeberg can be a sustainable place where humans share code without needing every new programmer or VC-grade growth. He notes people using Notepad++ beat him on every metric he could name while he was tuning Neovim; success by your own standards is the measure. Stack Overflow is offered as the counterexample of a business that did collapse — down to roughly ten questions a day.

Casey's version is broader: a population that reflexively declines the new thing is insurance. Humanity does not know in advance when it is doing something fantastically stupid, so having some fraction go another way is a feature, even if they turn out to be wrong. Homogenizing everyone onto one method eliminates our greatest strength. Prime's practical corollary: if you love AI you should want people still writing code by hand, both as untainted training data and so the craft is not lost. Casey's analogy — telling people not to write a game engine because Unity exists is arguing against your own interest, since those are the people who built Unity.

The episode also quotes George Hotz's post "I love LLMs, I hate the hype" against the "window is closing / perpetual underclass" framing: it is negative-valence hype designed to make you feel bad about yourself, and the people pushing it feel that way inside all the time.

Other threads

• Gell-Mann amnesia applied to AI: people confidently declare LLMs great at the areas they do not know and weak at the one area they do, then forget the pattern. Your problem domain is not everyone's — Prime would automate a CRUD SaaS aggressively, and would not generalize from it.
• The TJ analogies the panel keeps returning to: Amish chairs you can still buy, playing drums instead of using a drum machine, playing chess even though the engine is stronger, watching baseball even though a robot could hit better. Wanting to do the thing is sufficient justification, and other people doing it is not in your way.