Protecting our FLOSS commons from LLMs

保护我们的自由开源软件公共资源免受大语言模型（LLM）的影响

Bastian Greshake Tzovaras, Otto Richter, William Zijl | Thu 23 July 2026 | Category: Contrib Bastian Greshake Tzovaras, Otto Richter, William Zijl | 2026年7月23日（周四） | 分类：贡献

In Brief: Two motions regarding “artificial intelligence” and Large Language Models (LLMs) were voted on among Codeberg e. V. members and passed. We are promising to not use any of your data to train LLM and explain what the planned Terms of Use change mean for ‘vibe-coded’ projects. We believe that LLMs endanger the free/libre software ecosystem as a whole. 简讯：Codeberg e. V. 的成员对两项关于“人工智能”和大语言模型（LLM）的提案进行了投票并获得通过。我们承诺不会使用您的任何数据来训练大语言模型，并解释了计划中的服务条款变更对“氛围编程”（vibe-coded）项目意味着什么。我们认为，大语言模型正在危及整个自由/开源软件生态系统。

The Codeberg e. V. annual assembly is the meeting that puts power into the hand of our active members. Proposals are discussed live, and later voted on asynchronously. Since Large Language Models (LLMs) are an emerging but controversial technology, it is not surprising that two of the votes were concerned with Codeberg’s position about this technology. The 14-day voting period ended yesterday and both proposals were accepted. Codeberg e. V. 年度大会是将权力交到活跃成员手中的会议。提案在现场进行讨论，随后进行异步投票。由于大语言模型（LLM）是一项新兴但充满争议的技术，因此其中两项投票涉及 Codeberg 对该技术的立场也就不足为奇了。为期 14 天的投票期已于昨天结束，两项提案均获通过。

The first vote was a statement about Codeberg e. V.’s stance on using your data to train LLMs. As stated in our privacy policy, “We do not want to need your data”, and this also holds for the use of our user and project data for using or training generative “AI”: The Codeberg forge and its associated services are not and will not use the code or data of projects and users to train “Artificial Intelligence” tools such as Large Language Models, whose purpose is to create output modelled after their training input. As an association, we believe that these technologies are incompatible with responsibly creating and maintaining free & open source software. 第一次投票是关于 Codeberg e. V. 在使用您的数据训练大语言模型方面的立场声明。正如我们的隐私政策所述，“我们不希望需要您的数据”，这也适用于将我们的用户和项目数据用于使用或训练生成式“人工智能”：Codeberg 代码托管平台及其相关服务现在不会、将来也不会使用项目和用户的代码或数据来训练大语言模型等“人工智能”工具，因为这些工具的目的是根据其训练输入来模拟生成输出。作为一个协会，我们认为这些技术与负责任地创建和维护自由及开源软件是不相容的。

The second vote was more controversial, but was also accepted with 358 agreements vs 144 disagreements (and 14 abstentions), with a high voter turn-out of around 50% of active members. It implies a change to our terms of use to prohibit ‘vibe-coded projects’. We’ll share thoughts about the practical impact at the end of the article. 第二次投票更具争议性，但也以 358 票赞成、144 票反对（另有 14 票弃权）的结果获得通过，活跃成员的投票率高达 50% 左右。这意味着我们将修改服务条款，禁止“氛围编程项目”。我们将在文末分享关于其实际影响的思考。

We all pay for hungry LLMs

我们都在为贪婪的大语言模型买单

LLMs are a very costly technology, and those costs keep rising as the companies providing them have to start recouping their investments. They are not only costly for those who use and explicitly subscribe to these services. The costs are not only hidden in ‘normal’ cloud and service subscriptions that cross-finance the ‘innovative new features’ you never asked for. LLMs are so costly that companies externalize the costs on a massive scale - on those who don’t use them and society at large. Increased hardware prices, energy use and environmental damage - we all pay for it! 大语言模型是一项成本极高的技术，随着提供这些技术的公司开始收回投资，成本也在不断上升。它们不仅对那些使用并明确订阅这些服务的人来说成本高昂。这些成本不仅隐藏在那些为您从未要求过的“创新新功能”提供交叉补贴的“普通”云服务订阅中。大语言模型成本之高，以至于公司将成本大规模地转嫁给了那些不使用它们的人以及整个社会。硬件价格上涨、能源消耗和环境破坏——我们都在为此买单！

Strained servers due to nonsensical crawling

无意义爬取导致的服务器压力

In past posts we have already outlined how our infrastructure at Codeberg is regularly put under heavy load from webcrawlers of those companies who plan to ingest all of the code that is hosted on Codeberg for training their LLMs. At Codeberg, we are happy to provide free and open access to code. Just run git clone and enjoy. Unfortunately, these crawlers instead try to read every single page from Codeberg, no matter if it makes sense. This includes all the different issue filter variants, Git history, as well as the actual files at any point in Git history - even if they are still equal. These needless accesses create expensive database queries that diminish the service quality for all of us, requires substantial amounts of work from our system administrators, and force us to spend time building defensive mechanisms instead of cool new stuff. Mechanisms that also affect new and existing legitimate users, as we’re having to impose limits or outright blocks on their desired workflow; leaving them a worse experience with Codeberg. 在之前的文章中，我们已经概述了 Codeberg 的基础设施如何经常受到那些计划抓取 Codeberg 上托管的所有代码以训练其大语言模型的公司的网络爬虫的沉重压力。在 Codeberg，我们乐于提供免费和开放的代码访问权限。只需运行 git clone 即可享受。不幸的是，这些爬虫反而试图读取 Codeberg 上的每一个页面，无论这是否有意义。这包括所有不同的问题过滤器变体、Git 历史记录，以及 Git 历史中任何时间点的实际文件——即使它们的内容完全相同。这些不必要的访问产生了昂贵的数据库查询，降低了我们所有人的服务质量，需要我们的系统管理员投入大量工作，并迫使我们花费时间构建防御机制，而不是开发酷炫的新功能。这些机制也会影响新的和现有的合法用户，因为我们不得不对他们期望的工作流程施加限制或直接封锁，从而导致他们在 Codeberg 上的体验变差。

The development team of none

“零人”开发团队

Using LLMs to work with your code gives you a kick of adrenaline. You can develop at a rapid pace, build things as if you had a large team. Only that you have none. In fact, you are (often) alone, working with a statistical machine that turns energy into code. It seems like many ‘vibe coders’ don’t realize that they don’t actually have a community around them. They build projects as if they had, and spend resources accordingly. We see projects having a lot of code activity, heavy CI/CD testing, frequent and large release binaries. Sometimes, it feels like the amount of supported platforms exceeds the amount of actual users. To us, it seems ridiculous to see projects with a single developer and virtually no users consuming as much or even more resources than some of the largest community projects on Codeberg, which operate frugal with CI/CD and storage resources. We do not believe it is reasonable for Codeberg to invest our precious donation money into hosting of large ghost projects. 使用大语言模型来处理代码会给你带来肾上腺素激增的感觉。你可以快速开发，构建出仿佛拥有庞大团队才能完成的东西。但实际上你并没有团队。事实上，你（通常）是孤身一人，与一台将能量转化为代码的统计机器合作。似乎许多“氛围程序员”并没有意识到他们周围并没有真正的社区。他们像拥有社区一样构建项目，并相应地消耗资源。我们看到一些项目有大量的代码活动、繁重的 CI/CD 测试、频繁且庞大的发布二进制文件。有时，感觉支持的平台数量超过了实际用户的数量。对我们来说，看到一个只有一名开发者且几乎没有用户的项目，其消耗的资源竟然与 Codeberg 上一些最大的社区项目（这些项目在 CI/CD 和存储资源上非常节俭）相当甚至更多，这看起来很荒谬。我们认为 Codeberg 将宝贵的捐款投入到托管这些大型“幽灵项目”中是不合理的。

Hardware sourcing is becoming an headache

硬件采购正成为头疼的问题

The training and deployment of LLMs has drastically raised the cost of buying hardware, in particular for SSDs and memory. To give you an example: The type of drive we sourced for € 700 only some years ago has risen to € 3.700 now - and is often out of stock. As a consequence, hosting code on Codeberg is becoming more expensive. While we are owning our hardware, and are thus not directly impacted by inflating ‘cloud’ rental costs, it means that replacing or expanding our hardware is now substantially more costly than it used (and needs) to be. While we might be able to afford paying those inflated hardware prices, it is money that we can not spend elsewhere to improve our service and foster the mission of Codeberg. 大语言模型的训练和部署极大地提高了购买硬件的成本，特别是固态硬盘（SSD）和内存。举个例子：几年前我们以 700 欧元采购的同类硬盘现在已经涨到了 3700 欧元，而且经常缺货。因此，在 Codeberg 上托管代码变得越来越昂贵。虽然我们拥有自己的硬件，因此不会直接受到通胀的“云”租赁成本的影响，但这确实意味着更换或扩展硬件的成本现在比过去（以及应该有的水平）要高得多。虽然我们或许能够负担得起这些虚高的硬件价格，但这笔钱我们就无法用于其他地方来改善我们的服务并促进 Codeberg 的使命。

A growing digital divide

日益扩大的数字鸿沟

These price hikes also lead to a growing digital divide: Small and even large operators are endangered by rising costs, while only the largest cloud companies have reliable agreements for hardware. Increasing costs for services like website hosting, storage or compute can be challenging to a lot of small NGOs, local coops, research projects and other usage of digital tools that we considered for granted until recently. Not only does it become more expensive to run digital infrastructure, even more basic digital tools like computers and smartphones are affected by. 这些价格上涨也导致了日益扩大的数字鸿沟：小型甚至大型运营商都受到成本上升的威胁，而只有最大的云公司才拥有可靠的硬件供应协议。网站托管、存储或计算等服务的成本增加，对于许多小型非政府组织、本地合作社、研究项目以及我们直到最近还认为理所当然的其他数字工具的使用者来说，可能是一个挑战。不仅运行数字基础设施变得更加昂贵，甚至连计算机和智能手机等更基础的数字工具也受到了影响。