▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=RoHzXoolPTY
• Original title :: No Slop Allowed

ThePrimeagen reviews Codeberg's newly updated terms of use and companion blog post banning projects that "mostly consist of code written by generative AI tools," alongside a same-meeting ban on cryptocurrency projects. He reads both source documents directly and separates the arguments he finds legitimate from the ones he thinks are weak or badly worded.

The terms-of-use change (section 2.7) forbids sharing projects mostly made of generative-AI code. ThePrimeagen calls the wording itself poor: "mostly" is undefined (51%? a plurality?), and it creates ambiguity for projects that started fully hand-crafted but increasingly use LLM assistance with light edits — he cites Mitchell Hashimoto's Ghostty as an example of a project that would sit uncomfortably close to this line. The second stated concern, unclear copyright/license provenance, he considers the strongest argument: LLMs may have trained on copyleft-licensed code, and unknowingly incorporating copyleft output could force a project's license to change or expose it to jurisdictional risk (EU, Australia, etc.). He says if this were Codeberg's whole rationale, he'd fully support it. The third stated concern, insufficient security safeguards on AI-generated code, he pushes back on: high-profile CVEs and famous hacks (he mentions "Jia Tan" / the xz backdoor) predate mass LLM coding, and he argues LLMs (he singles out Chinese models as being more willing to flag issues than Western ones) can actually help find vulnerabilities rather than only introduce them.

Turning to Codeberg's blog post, he agrees with two of its arguments. First, "losing trust in each other": low-effort AI-generated contributions increase maintainer review burden and erode the one-on-one trust that open-source communities are built on, making maintainers warier of unfamiliar contributors and the community less welcoming overall. Second, the hardware/operational cost argument: Codeberg is self-hosted, so hardware costs have risen sharply (he cites an SSD example going from roughly $700 to 3,700 euros), compounded by AI-driven commit volumes and constant crawler traffic straining infrastructure — a real cost, both human-review and literal-hosting, that he considers a fair reason to restrict AI-heavy projects. He also notes approvingly that Codeberg states it won't use hosted code to train AI models, calling the resulting corpus "high-quality training data."

His strongest disagreement is with the blog's framing of "single-use" or throwaway vibe-coded software as a negative outcome of declining collaboration incentives. He argues single-use software is valuable precisely because not every idea needs to become a maintained, evolving project — most ideas are bad, and being able to quickly vibe-code a one-off to test an idea (rather than spending months hand-building it, as he did with his own "99" Neovim/LLM-agent project that ultimately didn't pan out) is a feature, not a problem. Software can simply be "done," and only the ideas that prove themselves should graduate into serious, actively maintained projects.

He closes by arguing Codeberg, as a small donation- and membership-funded nonprofit, shouldn't be expected to operate under the same norms as GitHub, a platform owned by a trillion-dollar corporation with its own incentives (including training AI on hosted code). He suggests there's room for different communities with different norms: places like Codeberg oriented around craftsmanship and community interaction, and other places suited for quick vibe-coded experiments.