<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/"><channel><title>Linux &amp; open source - Soursen</title><description>Linux and open source covers kernel releases and the projects the rest of the stack depends on, from the mailing list thread or the commit.</description><link>https://soursen.org/</link><language>en-gb</language><atom:link href="https://soursen.org/beat/linux-open-source/rss.xml" rel="self" type="application/rss+xml"/><item><title>Flathub Closed Its Forum. Chat Apps Are a Poor Replacement.</title><link>https://soursen.org/articles/flathub-closed-its-forum-chat-apps-are-a-poor-replacement/</link><guid isPermaLink="true">https://soursen.org/articles/flathub-closed-its-forum-chat-apps-are-a-poor-replacement/</guid><description>On October 23, 2026, Flathub made its Discourse forum read only and pointed people to Matrix and issue trackers. Days later, Discord announced it will put public server messages into web search. Together they show what a forum does that a chat room has never done well.</description><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On October 23, 2026, Flathub shut down the forum it had run since 2020. I want to walk through what Flathub said, what the archive it left behind shows, and why I think the move says more about how we treat questions than about one app store. I also want to cover Discord’s plan to make public server messages searchable, because it lands in the same month and changes one of the main arguments about chat.&lt;/p&gt;
&lt;p&gt;I should be clear about my own position. Flathub had good reasons, and I will lay them out fairly. I still think a chat room is a weak place to put support questions, and I will explain why with specifics.&lt;/p&gt;
&lt;h2 id=&quot;what-flathub-announced-on-october-23&quot;&gt;What Flathub announced on October 23&lt;/h2&gt;
&lt;p&gt;Flathub published a short post titled &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://docs.flathub.org/blog/discourse-deprecation&quot;&gt;Flathub Discourse is now read only&lt;/a&gt; on October 23, 2026, signed by Bart Piotrowski, who goes by Flathub Janitor. Existing threads stay available, and nobody can start a new topic or reply. The forum home page now describes itself as &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://discourse.flathub.org/&quot;&gt;a read only archive&lt;/a&gt; with its content preserved for reference.&lt;/p&gt;
&lt;p&gt;The post gives the reason in 3 sentences. The forum opened in 2020 as a place for asynchronous discussion outside GitHub issues and IRC. It never took off and became another issue tracker to watch. A forum where people post and wait for an answer that never comes, in the post’s words, is worse than no forum at all.&lt;/p&gt;
&lt;p&gt;The post also lists where each kind of question should go now. Packaging and submission help goes to the Flathub Matrix channel, and questions about Flatpak itself go to the Flatpak channel. Problems with an app go to the issue tracker linked as Manifest on the app’s Flathub page. Problems with Flathub itself go to the flathub/flathub repository. Flathub does not take app requests, so you ask the app’s developers to submit it or submit it yourself. Runtime projects such as freedesktop SDK can post announcements on the Flathub blog.&lt;/p&gt;
&lt;h2 id=&quot;why-a-forum-existed-in-the-first-place&quot;&gt;Why a forum existed in the first place&lt;/h2&gt;
&lt;p&gt;Flathub opened the forum in late January 2020. A &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lists.freedesktop.org/archives/flatpak/2020-January/001875.html&quot;&gt;message on the Flatpak mailing list&lt;/a&gt; dated January 25, 2020 says the Flathub Discourse had just opened that day. The reason appears in a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lists.freedesktop.org/archives/flatpak/2020-January/001876.html&quot;&gt;reply by Alexander Larsson on January 28&lt;/a&gt;. He called the lack of a Flathub specific forum a problem, because when the team changed how the infrastructure worked, it could not get the message to packagers in a good way. He pointed out that the flathub-admins list exists but carries low traffic limited to sysadmin matters, and that GitHub issues host some discussion without being a good communication tool or reaching the right audience.&lt;/p&gt;
&lt;p&gt;The forum became a working part of the project quickly. On April 1, 2020, Larsson &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lists.freedesktop.org/archives/flatpak/2020-April/001956.html&quot;&gt;posted a link to a Discourse thread&lt;/a&gt; for the Flathub vision statement, so people had a place for comments. By July 22, 2020, GNOME’s Felipe Borges was &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lists.freedesktop.org/archives/flatpak/2020-July/002015.html&quot;&gt;telling a newcomer&lt;/a&gt; that the forum had a specific section for Flathub application requests, and warning that the work is voluntary, so nobody might take the request. That category has no equivalent today, since Flathub’s 2026 post says it does not take app requests at all.&lt;/p&gt;
&lt;p&gt;GitHub’s own forum feature did not exist yet. GitHub announced Discussions in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.infoq.com/news/2020/05/GitHub-codespaces-discussions&quot;&gt;early May 2020 for selected public repositories&lt;/a&gt; and opened it to all projects as a public beta in December 2020, according to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://arxiv.org/pdf/2102.05230v3.pdf&quot;&gt;an academic study of the beta&lt;/a&gt;. It left beta on &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://github.blog/2021-08-17-github-discussions-out-of-beta&quot;&gt;August 17, 2021&lt;/a&gt;. By then Flathub’s forum had been running for more than a year and a half.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CUwqNBTt_ZJdakG.webp&quot; alt=&quot;A forum opened before GitHub Discussions existed. Flathub’s Discourse started 4 months before GitHub announced Discussions.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;727&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;A forum opened before GitHub Discussions existed. Flathub’s Discourse started 4 months before GitHub announced Discussions. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://docs.flathub.org/blog/discourse-deprecation&quot;&gt;Flathub blog, GitHub blog,&lt;/a&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://freedesktop.org/&quot;&gt;freedesktop.org&lt;/a&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://docs.flathub.org/blog/discourse-deprecation&quot;&gt;Flatpak mailing list&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-the-archive-shows&quot;&gt;What the archive shows&lt;/h2&gt;
&lt;p&gt;The archive is big enough to matter. The &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://discourse.flathub.org/categories&quot;&gt;forum’s category page&lt;/a&gt; lists 24 announcement topics, about 1,900 in General, 63 in Tutorials and Tips, and 101 in Site Help and Feedback. That adds up to about 2,100 topics of questions and answers from nearly 7 years of use.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.BY6fhFgG_1StpAl.webp&quot; alt=&quot;About 2,100 topics sit in the archive. Almost all of them live in the General category.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;669&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;About 2,100 topics sit in the archive. Almost all of them live in the General category. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://discourse.flathub.org/&quot;&gt;discourse.flathub.org&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The archive itself is a static copy. Its front page reads as &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://discourse.flathub.org/&quot;&gt;a read only archive&lt;/a&gt; with 71 pages of topics and links for Latest, Categories and Search, so the old answers stay reachable.&lt;/p&gt;
&lt;p&gt;The activity numbers explain Flathub’s frustration, and they also show something else. The &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://discourse.flathub.org/latest&quot;&gt;latest topics page&lt;/a&gt; in February 2026 lists threads with 7, 18, 12, 7 and 9 replies, which is modest traffic for a project of Flathub’s size. The same threads drew 5,825, 1,825, 1,058, 421 and 121 views. The thread about installing older versions had 7 replies and 5,825 views, a ratio of about 830 readers for each reply.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.BNyml2Vv_Z1A6pr3.webp&quot; alt=&quot;Readers outnumber repliers by 10x or more. Every one of these 5 threads had at least 13 views for each reply.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;698&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Readers outnumber repliers by 10x or more. Every one of these 5 threads had at least 13 views for each reply. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://discourse.flathub.org/latest&quot;&gt;discourse.flathub.org/latest&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;I read that as a sign the forum worked as a library. Few people replied, and many people read. That is the pattern a searchable archive is built for, and it is the pattern a chat room cannot reproduce once the messages scroll away.&lt;/p&gt;
&lt;p&gt;One older thread shows the same tension from the inside. In June 2022 a developer &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://discourse.flathub.org/t/where-is-the-right-place-to-ask-flatpak-developer-questions/2659&quot;&gt;asked where to put Flatpak developer questions&lt;/a&gt; because the mailing list seemed dead. A reply said the forum was fine because developers monitor it, and that live chat might be more responsive, with a link to the Flatpak Matrix channel. The thread has 4 replies, and a 2023 comment noted that Discourse chat was disabled on Flathub’s server, so Matrix was the recommendation. In that thread, chat was offered for speed and the forum was offered as the place developers watch.&lt;/p&gt;
&lt;p&gt;I only checked 5 threads for the view counts, so treat the ratio as an illustration. Flathub’s own judgment, that too few questions got answers, is a different measurement and I have no data to dispute it.&lt;/p&gt;
&lt;h2 id=&quot;why-chat-is-a-weak-home-for-support-questions&quot;&gt;Why chat is a weak home for support questions&lt;/h2&gt;
&lt;p&gt;I have 4 specific problems with answering support questions in a chat room, and I think each one is a mechanical limit of the format.&lt;/p&gt;
&lt;p&gt;The first is finding the earlier answer. A person with a question wants to know whether someone already asked it. On a forum, a search engine can answer that before the person ever signs up. In a chat room, they have to join first and then rely on the room’s own search. Element’s help page says that in encrypted Matrix rooms, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://element.io/help&quot;&gt;search only works on the desktop app&lt;/a&gt; and only after you turn on message search in settings. One university’s Matrix help page adds that &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://docs.matrix.kit.edu/en/messaging/search/&quot;&gt;unencrypted rooms are searchable&lt;/a&gt; without that step. So the answer depends on the room settings and the app a person happens to use.&lt;/p&gt;
&lt;p&gt;The second is duplicates. If people cannot find the earlier answer, they ask again, and the people who know the answer cannot find it either. A forum moderator can link the old thread. A chat helper has to remember where the answer was posted, or retype it.&lt;/p&gt;
&lt;p&gt;The third is structure. A chat room is one stream. A new question breaks into whatever conversation is already happening, and the reply feature that links an answer to its question only works when the person answering uses it. Skip it once and the asker returns to a wall of newer messages with no sign their question was answered.&lt;/p&gt;
&lt;p&gt;The fourth is shelf life. A question asked in a forum stays at the top of a category list and gets found by people who browse back a few pages. A question asked in chat is pushed down by the next 20 messages within minutes. If nobody answers it in the first hour, I think it is effectively gone.&lt;/p&gt;
&lt;p&gt;Chat does plenty of things well. It builds community, it works for quick back and forth, and it handles throwaway questions without ceremony. The trouble starts when a project expects it to double as a public knowledge base.&lt;/p&gt;
&lt;h2 id=&quot;discord-is-changing-the-premise&quot;&gt;Discord is changing the premise&lt;/h2&gt;
&lt;p&gt;The argument that chat history is invisible to search is shifting, and Discord is the reason. On October 27, 2026, Discord &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.implicator.ai/discord-public-chats-web-search.md&quot;&gt;announced plans&lt;/a&gt; to make some messages from its Discoverable Servers searchable on the web. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://redact.dev/blog/discord-public-server-messages-google-search&quot;&gt;Redact.dev’s explainer&lt;/a&gt; puts the number of Discoverable Servers at more than 63,000, with a minimum of 1,000 members and a minimum server age of 8 weeks.&lt;/p&gt;
&lt;p&gt;The rollout has 3 steps. Controls for admins and members went live on October 27. Every Discoverable Server gets a public web page on November 16, along with automated channel summaries. Individual messages from eligible channels start to appear in mid December, in a gradual rollout, according to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://pcworld.com/article/3273746/discord-is-letting-search-engines-read-public-messages-heres-what-to-do.html&quot;&gt;PCWorld&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.D-KeUSMt_Z1UXygj.webp&quot; alt=&quot;How Discord is putting public chat on the web. Messages begin to appear in mid December, and the member setting defaults on in the US.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;873&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;How Discord is putting public chat on the web. Messages begin to appear in mid December, and the member setting defaults on in the US. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.implicator.ai/discord-public-chats-web-search.md&quot;&gt;Implicator,&lt;/a&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://redact.dev/&quot;&gt;Redact.dev&lt;/a&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.implicator.ai/discord-public-chats-web-search.md&quot;&gt;, BleepingComputer, PCWorld&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The limits are worth knowing, because the first reactions I saw ignored them. Direct messages, group DMs and private servers are excluded. Messages sent before the mid December rollout are excluded. Published messages leave out usernames and avatars, but Implicator notes that identifying details in the text can remain. Members can switch off the sharing of their message content under User Settings and Data and Privacy, and that applies only to later messages. Messages that were already published stay unless the member deletes them. Search engines may keep cached copies longer, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.ibtimes.co.uk/discord-public-server-messages-searchable-web-1828365&quot;&gt;IBTimes reports&lt;/a&gt;, and removal can take a separate request. The member setting defaults off in the EU, the UK, Canada, Brazil, Australia and South Korea, and on in most other regions including the US. Teen accounts cannot turn it on.&lt;/p&gt;
&lt;p&gt;Discord’s pilots ran in the NBA 2K, Rust and Dead by Daylight communities, according to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bleepingcomputer.com/news/security/discord-to-let-search-engines-index-public-server-messages/&quot;&gt;BleepingComputer&lt;/a&gt;, and Implicator adds that the pilots used only administrator and moderator posts. The mid December rollout is the first time regular members’ messages will be published.&lt;/p&gt;
&lt;p&gt;On AI, Implicator reports that Discord denies selling messages to AI companies. A Discord staff member also acknowledged that anything on the public web can be accessed by people and bots. Redact.dev adds context about earlier scraping. A 2025 project called Searchcord claimed to have archived more than 63 billion messages from Discoverable Servers, and Spy.pet sold searchable archives of public server messages until Discord banned the accounts behind it.&lt;/p&gt;
&lt;p&gt;I find the stated motive the most telling part. Implicator reports that Discord group product manager Kelsey Duncombe-Smith wrote that web search would help people find answers currently locked inside Discord conversations. That is Discord agreeing that chat hides answers. The company is now building the index that a forum has by default.&lt;/p&gt;
&lt;p&gt;The idea is older than Discord’s announcement. A service called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://docs.answeroverflow.com/channel-settings/indexing&quot;&gt;Answer Overflow&lt;/a&gt; has let server owners index chosen Discord channels into Google through a sitemap since at least 2023. Communities that wanted a searchable archive have been bolting one onto chat for years.&lt;/p&gt;
&lt;h2 id=&quot;other-projects-made-the-same-move&quot;&gt;Other projects made the same move&lt;/h2&gt;
&lt;p&gt;Other projects closed Discourse forums before Flathub did, and the destinations differ. Keycloak &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.keycloak.org/2023/06/sunsetting-discourse&quot;&gt;sunset its Discourse on June 26, 2023&lt;/a&gt;, saying that managing multiple channels was a challenge and that fewer channels should allow more meaningful conversations. It pointed people to Google Groups, GitHub Discussions and Slack, and left the old threads accessible as an archive. Bluefin, a Linux distribution, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://universal-blue.discourse.group/t/bluefin-discussions-moving-to-github/9790&quot;&gt;announced on July 17, 2025&lt;/a&gt; that it was moving discussions to GitHub Discussions and announcements to a blog. One of the reasons it gave was that a single source of truth is easier for AI tools to index than content spread across several places, and another was that its Discord was nearing 20,000 members.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.Bhr9FMv2_NgAfu.webp&quot; alt=&quot;3 projects closed a Discourse forum, 3 different exits. Keycloak and Bluefin sent discussion to GitHub Discussions, and Flathub sent it to Matrix and issue trackers.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;698&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;3 projects closed a Discourse forum, 3 different exits. Keycloak and Bluefin sent discussion to GitHub Discussions, and Flathub sent it to Matrix and issue trackers. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.keycloak.org/2023/06/sunsetting-discourse&quot;&gt;Keycloak blog, Universal Blue forum, Flathub blog&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;That comparison shapes my view of Flathub’s choice. Two of the 3 projects moved discussion to GitHub Discussions, a public web venue. Flathub sent packaging questions to a chat room. I think that is the weakest part of the plan, because the questions Flathub says it wants to keep answering are the ones with answers worth keeping.&lt;/p&gt;
&lt;h2 id=&quot;the-issue-tracker-worry&quot;&gt;The issue tracker worry&lt;/h2&gt;
&lt;p&gt;Flathub’s post says the forum became another issue tracker to keep an eye on. I think that observation is accurate, and I doubt that closing the forum fixes it. People report a problem in whichever place feels least intimidating. An issue template with fields for reproduction steps and logs scares some people off, and a plain discussion thread does not. Take the discussion venue away and some of those reports move to places the project does not control.&lt;/p&gt;
&lt;p&gt;Flathub’s replacement for packaging help is the Matrix channel. I would not send a question that needs a durable answer there for the reasons above. The Flathub issue tracker works for problems with an app because the answer ends up attached to the report, and the tracker is public and searchable. That leaves general questions with the weakest home, and those are the questions a forum was best at.&lt;/p&gt;
&lt;p&gt;Flathub also made 2 decisions I would copy. It kept the old threads online as a read only archive, so the answers did not disappear with the venue. It also named a specific destination for each category of question. Closing a forum is easier to defend when the old content stays findable.&lt;/p&gt;
&lt;h2 id=&quot;2-things-from-this-generalize&quot;&gt;2 things from this generalize&lt;/h2&gt;
&lt;p&gt;First, match the channel to how long the answer stays useful. A question whose answer will help the next person belongs somewhere a stranger can find without joining anything. A question about what someone is doing right now belongs in chat. Most communities mix the two in one room and then wonder why the same questions repeat.&lt;/p&gt;
&lt;p&gt;Second, never close a venue without keeping its content readable. Flathub’s archive of about 2,100 topics keeps working for anyone who searches for a Flatpak problem. A deleted forum or a private chat server loses that value on the day it closes.&lt;/p&gt;
&lt;p&gt;This week, list the 5 questions your community asked most often last month and check whether a stranger could find each answer through a search engine without joining anything. Move the ones that fail onto a public page, and link that page from the chat room.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>flathub</category><category>linux</category><category>chatbots</category><category>seo</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Asahi Linux Dropped MIT for GPL After Alleged AI Code Laundering</title><link>https://soursen.org/articles/asahi-linux-dropped-mit-for-gpl-after-alleged-ai-code-laundering/</link><guid isPermaLink="true">https://soursen.org/articles/asahi-linux-dropped-mit-for-gpl-after-alleged-ai-code-laundering/</guid><description>On November 7, 2026, Asahi Linux said it is moving every repository to GPL v2 or later. Asahi alleges that unreleased code was lifted from personal forks and that some of it was rewritten with LLM help to hide its origin. The report names no one, and the legal and community questions are still open.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On November 7, 2026, the project that made Linux run on Apple Silicon Macs announced it was giving up the MIT license. I want to walk through what Asahi Linux actually said, what the license change does, and why the argument landed in the middle of a fight about AI generated code. I also want to be careful about what is established and what is only alleged, because most of the heat online comes from filling in names the report never mentions.&lt;/p&gt;
&lt;p&gt;Everything about the alleged copying comes from Asahi’s own words. I could not verify it independently, and nobody has been named. I am not a lawyer, so treat the license discussion as an explanation of public documents and statements, and get a lawyer’s view before acting on any of it.&lt;/p&gt;
&lt;h2 id=&quot;what-asahi-announced-on-november-7&quot;&gt;What Asahi announced on November 7&lt;/h2&gt;
&lt;p&gt;Asahi published its &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://noise.getoto.net/2026/11/07/asahi-linux-7-3-progress-report/&quot;&gt;Linux 7.3 progress report&lt;/a&gt; on November 7, 2026. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://braindetox.kr/en/posts/asahi_linux_progress_report_linux_7_3.html&quot;&gt;Braindetox’s summary&lt;/a&gt; lists the byline as James Calligeros and counts 9 sections, covering upstreaming work, M3 general availability, reverse engineering of the M4 and M5, the M3 GPU, Thunderbolt, games, microphones, Fedora 45 and the license change. The license change sits in the last section, which is titled “Learning from history”.&lt;/p&gt;
&lt;p&gt;The core announcement is short. Asahi said it has begun moving all of its repositories to the GNU General Public License version 2 or later, and that going forward all Asahi Linux code will be under strong copyleft licenses only. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/asahi-linux-drops-mit-for-gpl/&quot;&gt;It’s FOSS reported the move&lt;/a&gt; on November 8, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://runtimewire.com/article/asahi-linux-gpl-llm-rewrites-dhh-steinberger&quot;&gt;RuntimeWire&lt;/a&gt; published the same day. RuntimeWire stresses that the repositories have only “begun” to move, and it gives no list of affected repositories or a completion date. Braindetox’s read of the report agrees on both gaps.&lt;/p&gt;
&lt;p&gt;The stated reason is the part that got attention. Per Braindetox, Asahi said it develops in the open, including its finances and development branches, and that it originally chose MIT to encourage participation from the BSDs and other free operating systems. It now considers that approach no longer viable. Asahi said it learned of multiple attempts to use its permissive license to take over the project and monetize Linux on Apple Silicon, and that some in development code was taken from personal forks without attribution or rewritten with LLM help to obscure where it came from. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/asahi-linux-drops-mit-for-gpl/&quot;&gt;It’s FOSS quotes the report&lt;/a&gt; saying MIT “explicitly allows this, and has no carveouts to protect against bad faith actors.”&lt;/p&gt;
&lt;p&gt;Asahi did not name anyone, and RuntimeWire notes that the article presents the claims as Asahi’s allegations with no evidence, repositories or copied code attached. RuntimeWire also points out that Asahi has not explained how existing contributor owned code and the earlier dual licensing will be handled.&lt;/p&gt;
&lt;h2 id=&quot;why-the-mit-license-left-the-door-open&quot;&gt;Why the MIT license left the door open&lt;/h2&gt;
&lt;p&gt;MIT is one of the most permissive licenses in common use. Anyone can use, copy, modify, sell and relicense MIT code as long as the copyright and permission notice travels with it. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://sesamedisk.com/asahi-omarchy-license-fight/&quot;&gt;SesameDisk’s comparison&lt;/a&gt; lays out the difference with GPL v2, which also allows selling but requires anyone who distributes a derivative to provide the source and keep it under GPL v2.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DMn1wrMy_2atu6s.webp&quot; alt=&quot;What MIT and GPL v2 each allow. Both licenses let a company sell the code, and only GPL v2 blocks a closed fork.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;713&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;What MIT and GPL v2 each allow. Both licenses let a company sell the code, and only GPL v2 blocks a closed fork. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://sesamedisk.com/asahi-omarchy-license-fight/&quot;&gt;The MIT License, GNU GPL v2, SesameDisk&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;That one row about closed forks explains the whole decision. Under MIT, a company can take Asahi’s code, close it up and sell a product, and that is legal as long as the notice stays in. Under GPL v2, anyone who distributes a modified version has to release the source under the same license. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/asahi-linux-drops-mit-for-gpl/&quot;&gt;It’s FOSS makes the practical point&lt;/a&gt; that a closed, paid Apple Silicon distribution built on GPL code would be illegal, and that copying becomes easier to detect. Stripping attribution already violates MIT, since keeping the notice is its main condition.&lt;/p&gt;
&lt;p&gt;Asahi picked MIT on purpose. Its &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://asahilinux.org/copyright&quot;&gt;copyright policy&lt;/a&gt; says code written for Asahi itself typically uses a permissive license such as MIT, decided case by case, and that self contained modules such as an entire driver should be dual licensed under a permissive option so they can be reused elsewhere. The policy’s example header marks files as GPL 2.0 or later OR MIT. The page I read is the policy as it stood before the November report, and I could not tell whether it has been updated. The goal was reuse by other operating systems. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/asahi-linux-drops-mit-for-gpl/&quot;&gt;It’s FOSS notes&lt;/a&gt; that OpenBSD’s Apple Silicon support relies on Asahi’s m1n1 bootloader, which shows how real that goal was.&lt;/p&gt;
&lt;p&gt;Parts of Asahi were already GPL. The Linux kernel is GPL v2, so Asahi’s kernel work had to be, and SesameDisk reports that m1n1 has been GPL v2 since version 2.0.0. The change matters most for the userspace and tooling repositories that were MIT.&lt;/p&gt;
&lt;h2 id=&quot;wine-ran-the-same-experiment-in-2002&quot;&gt;Wine ran the same experiment in 2002&lt;/h2&gt;
&lt;p&gt;Asahi drew the comparison itself, and It’s FOSS reports the same parallel. Wine, the layer that runs Windows software on Linux, started under the MIT license. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Wine_%28software%29&quot;&gt;Wikipedia’s Wine entry&lt;/a&gt; says the project moved to the LGPL as of March 2002 because of concern that proprietary versions would not contribute their changes back. The people involved said so at the time. On February 6, 2002, CodeWeavers’ Jeremy White &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://winehq.org/pipermail/wine-devel/2002-February/003933.html&quot;&gt;posted a call for the change&lt;/a&gt; to the Wine developer list. In a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://winehq.org/pipermail/wine-devel/2002-February/004286.html&quot;&gt;follow up on February 15&lt;/a&gt; he said the BSD style license encourages forks such as WineX, that competitors can take contributed code without giving anything back, and that CodeWeavers had become unwilling to contribute its code under the current license. He added that he would not release code under a BSD license again.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Cedega_%28software%29&quot;&gt;Wikipedia’s Cedega entry&lt;/a&gt; describes the proprietary product by TransGaming, formerly WineX, that forked the last MIT licensed version of Wine in 2002. It is the same shape of problem Asahi describes, a permissive license being used the way it was written. Wine kept going, and today it ships under LGPL 2.1 or later. CodeWeavers sells CrossOver, a supported version of Wine, and contributes most of its Wine code back under the LGPL. Valve works with CodeWeavers on Proton, which is built on Wine.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.JUD0wI-x_1OfXP9.webp&quot; alt=&quot;Two license switches, 24 years apart. Both projects started permissive and moved to copyleft after a dispute about reuse.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;786&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Two license switches, 24 years apart. Both projects started permissive and moved to copyleft after a dispute about reuse. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://winehq.org/pipermail/wine-devel/2002-February/004286.html&quot;&gt;WineHQ mailing list archive, Wikipedia, Asahi Linux progress report 7.3 via braindetox&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The comparison has limits. The Wine forks were openly commercial and visible. Asahi’s allegation involves unreleased code and rewrites that hide their origin, which is mainly a provenance problem. A license answers who may distribute what, and the origin of a rewritten file needs a separate provenance record.&lt;/p&gt;
&lt;h2 id=&quot;what-else-was-in-the-73-report&quot;&gt;What else was in the 7.3 report&lt;/h2&gt;
&lt;p&gt;The license change closes a report that is mostly engineering, and the numbers are worth knowing because they show what is at stake. Per Braindetox, the macaudio audio driver and its Devicetree work total about 3,100 lines and reached the kernel mailing list this cycle. The AVD driver is about 13,000 lines, spent about 3 months downstream, and Asahi called it the shortest turnaround yet for a major SoC component. Initial Devicetrees for the M4 Pro, M4 Max and M5 also reached the list, and they only boot the SoC to a text console.&lt;/p&gt;
&lt;p&gt;M3 support moved out of the installer’s expert flag and is now generally available, based on almost no issues from early adopters. Enabling the display coprocessor was the most notable M3 improvement, bringing performance and battery gains and allowing sleep. The M3 GPU is still missing. Asahi says a new group is reverse engineering it, and that m1n1 can initialize it and draw a triangle, though Apple’s choices will need significant effort in Mesa.&lt;/p&gt;
&lt;p&gt;Thunderbolt works for USB3, XDomain and Thunderbolt Networking, and developer Sven has spent nearly 5 years on it. PCIe and DisplayPort tunneling, which most docks need, is unsupported, and the report adds the word “Yet”. On M5, macOS needs a component called Exclaves, so Asahi now runs the Secure Kernel in a separate virtual machine, and m1n1 spawns 3 VMs so macOS can be studied on the M4 and M5. On games, Asahi puts the penalty of its muvm virtual machine at about 20%, depending on workload, and it shows Wolfenstein: The New Order reaching 144+ FPS on an M2 Max Mac Studio outside it. The Triforce microphone beamformer became almost twice as fast.&lt;/p&gt;
&lt;p&gt;That is slow, careful, volunteer paced work, and it is the work the license change is meant to protect.&lt;/p&gt;
&lt;h2 id=&quot;the-apple-silicon-field-got-crowded-in-2-months&quot;&gt;The Apple Silicon field got crowded in 2 months&lt;/h2&gt;
&lt;p&gt;The timing makes sense once you line up what happened between September and November 2026. Omarchy announced a dedicated Apple Silicon effort called Omarchy M on September 11. Gravity Linux published an early alpha for the M4 Mac mini on September 18, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://modelfit.io/blog/gravity-linux-m4-mac-mini-local-llm/&quot;&gt;according to ModelFit&lt;/a&gt;. On October 23, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Borealis-Vulkan-Driver&quot;&gt;Phoronix reported on Borealis&lt;/a&gt;, a Vulkan driver from the Aurora Silicon project that is derived from Asahi’s Honeykrisp driver and lives in its own Mesa fork. Asahi shipped Fedora Asahi Remix 45 on November 3, the same day as Fedora 45, which the report calls the first time a Remix release matched upstream’s date. Phoronix reported on November 5 that &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Omarchy-4.5-Released&quot;&gt;Omarchy 4.5 had shipped with Apple Silicon support&lt;/a&gt;. Asahi’s report followed on November 7.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.t9rKzve1_2dADlP.webp&quot; alt=&quot;6 dates behind the license change. The orange dot marks the Asahi announcement.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;727&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;6 dates behind the license change. The orange dot marks the Asahi announcement. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Omarchy-4.5-Released&quot;&gt;Phoronix, ModelFit, Asahi Linux via braindetox, SesameDisk&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Support differs by project, and this is where the pace argument comes from. Asahi’s report says M3 support is generally available, the M3 GPU is still being reverse engineered, and the M4 and M5 are in reverse engineering. Phoronix says Omarchy 4.5 is reported to work on M1, M2, M3 and the MacBook Neo, and that it builds on work by Asahi Linux and others. Braindetox, summarizing Omarchy’s announcement, says M4, M5 and M6 are listed as in development with no schedule. Gravity’s alpha targets the base M4 Mac mini, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Gravity-Linux-M4-Mac-Mini-Alpha&quot;&gt;Phoronix lists&lt;/a&gt; suspend, display output on the USB-C ports and Thunderbolt as not working.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.Bl44bZj__1dCsIK.webp&quot; alt=&quot;Apple Silicon Linux support by project. Gravity’s alpha reaches the M4, which Asahi has not shipped.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;786&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Apple Silicon Linux support by project. Gravity’s alpha reaches the M4, which Asahi has not shipped. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://modelfit.io/blog/gravity-linux-m4-mac-mini-local-llm/&quot;&gt;Asahi Linux progress report 7.3 via braindetox, Phoronix, ModelFit&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;who-the-report-is-about&quot;&gt;Who the report is about&lt;/h2&gt;
&lt;p&gt;Asahi’s report names nobody, and It’s FOSS says so directly, adding that it supplied the context about Aurora Silicon and Gravity Linux itself. Some readers jumped to Omarchy. I would hold off. The facts I can verify are narrower.&lt;/p&gt;
&lt;p&gt;Aurora Silicon’s Borealis driver describes itself in its own pull request as derived from Honeykrisp and as running on the M1 through M5 and the A18 Pro. Phoronix reports that the fork strips Mesa’s AI agent documentation and replaces it with a policy that allows AI contributions, and that the change was co authored by Claude Opus 5.5. That is a public fork that says plainly where it came from, and nothing public ties it to the unreleased code Asahi describes.&lt;/p&gt;
&lt;p&gt;Gravity Linux forked from Asahi over the LLM policy. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/gravity-linux-launch/&quot;&gt;It’s FOSS names&lt;/a&gt; Cody Ho and Niklas Sheth as the main drivers and describes the project as open source, selling nothing, crediting its AI use openly and planning to wind down once its driver work lands upstream. Phoronix says it uses OpenAI Codex and Claude Code, and that its release is described as clean room reverse engineering. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://modelfit.io/blog/gravity-linux-m4-mac-mini-local-llm/&quot;&gt;ModelFit adds&lt;/a&gt; that Gravity applies a tainting rule, under which a contributor whose model session has seen Apple disassembly or protected details cannot write the clean room implementation of that component. The 2 developers say in their own write up that they built an OpenGL ES 3.0 compliant driver for the M4 Mac mini in about &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://zeli.app/story/49717638&quot;&gt;1 month&lt;/a&gt;, which is a claim about speed that I could not test.&lt;/p&gt;
&lt;p&gt;Omarchy’s creator, David Heinemeier Hansson, has been publicly critical of Asahi’s AI policy. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://digg.com/tech/sq3kpn14&quot;&gt;Digg’s summary of his post&lt;/a&gt; says he respects what Asahi achieved, objected bluntly to the policy, and suggested someone start a new effort using the best agents to finish M3, M4 and M5 support. SesameDisk paraphrases his November reply as saying Asahi rejected upstreaming, that Omarchy had expanded Asahi’s work with Touch ID, M3 GPU, DisplayPort over USB-C, deep sleep and MacBook Neo support, and that the AI policy deserved mockery. Some commentary attributes further claims to him, including a pledge that Omarchy stays free. I could not match those to a primary post, so I leave them out.&lt;/p&gt;
&lt;p&gt;SesameDisk also looked at the Omarchy Mac repository and reports that it is MIT licensed, installs Asahi’s kernel and firmware packages through a package list, and does not vendor Asahi’s kernel source. Its grep found no Asahi copyright headers outside the README and a mirrorlist. That is one outlet’s check, and it does not speak to unreleased code either way. Omarchy’s own 4.5 announcement, as relayed by Braindetox, counts more than 740 merged pull requests from more than 180 contributors in 12 weeks.&lt;/p&gt;
&lt;h2 id=&quot;the-ai-policy-sits-underneath-all-of-it&quot;&gt;The AI policy sits underneath all of it&lt;/h2&gt;
&lt;p&gt;Asahi’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://asahilinux.org/slop&quot;&gt;generative AI policy&lt;/a&gt; says the use of generative AI tools in any contribution is expressly forbidden, and it promises an immediate and permanent ban for subsequent disregard. Its first stated concern is illegal output, since the popular generators are trained on an enormous corpus of text. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://modelfit.io/blog/gravity-linux-m4-mac-mini-local-llm/&quot;&gt;ModelFit quotes the policy&lt;/a&gt; as broadly forbidding generative AI for material contributions and summarizes the reasons as clean room and legal risk.&lt;/p&gt;
&lt;p&gt;That is a coherent position for a reverse engineering project, and a different project can reach the opposite one. The Linux kernel &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.tomshardware.com/software/linux/linux-lays-down-the-law-on-ai-generated-code-yes-to-copilot-no-to-ai-slop-and-humans-take-the-fall-for-mistakes-after-months-of-fierce-debate-torvalds-and-maintainers-come-to-an-agreement&quot;&gt;settled on a policy&lt;/a&gt; in April 2026 that allows AI assisted contributions with strict disclosure rules. Per Tom’s Hardware, contributors disclose AI assistance with a new Assisted-by tag, AI agents cannot use the Signed-off-by tag, and the human submitter answers for every bug or security flaw. My read is that this gap matters for upstreaming. Asahi wants its work in the mainline kernel, and the mainline kernel now accepts code Asahi’s own rules would reject. That tension will outlast any license.&lt;/p&gt;
&lt;p&gt;ModelFit adds a useful caution on the speed argument. It notes that The Register credits coding agents with compressing months of driver work into weeks, and calls it a demonstration with no controlled test of whether Asahi’s policy slowed progress.&lt;/p&gt;
&lt;h2 id=&quot;what-a-license-switch-fixes&quot;&gt;What a license switch fixes&lt;/h2&gt;
&lt;p&gt;I think the criticism of the move deserves as much weight as the defense. SesameDisk makes the cleanest case that the license and the AI policy do different jobs. GPL v2 governs distribution. Its text is silent on AI training, and a clean room contribution rule has to live in a separate policy, backed by contribution audits. Old releases stay available under their old license, and the change applies to new code and new releases.&lt;/p&gt;
&lt;p&gt;There is a second objection that is simply practical. If the concern is people who already ignore the MIT license, why would they respect the GPL. A determined bad actor ignores both. It’s FOSS agrees that the GPL cannot stop a determined bad actor, and says its effect is on the commercial payoff, since a closed fork built on GPL code is a license violation with a visible paper trail.&lt;/p&gt;
&lt;p&gt;The third objection comes from inside the free software world. OpenBSD developer Bryan Steele &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/asahi-linux-drops-mit-for-gpl/&quot;&gt;criticized the move on November 7&lt;/a&gt; on X, saying that relicensing m1n1 and the drivers seemingly without the consent of all prior contributors was misguided. SesameDisk argues that relicensing old contributions needs that consent, and that the public repositories it checked show no copyright assignment agreement. In its view, if consent was never obtained the relicense is shaky for those contributions, and if Asahi holds consent elsewhere the concern goes away. Another view, which I have also seen argued, is that MIT already permits sublicensing, so a project can distribute the combined work under GPL as long as the original notices stay. I do not know which reading a court would take, and Asahi has not published its process. That paperwork is what I would watch.&lt;/p&gt;
&lt;p&gt;It’s FOSS also offers a broader reading. As far as its author can tell, this is the first major open source project to move to copyleft specifically because of AI assisted code laundering, since earlier relicensings mostly targeted cloud resellers. The author expects other hardware enablement and reverse engineering projects might follow. That is a forecast, and I treat it as one.&lt;/p&gt;
&lt;h2 id=&quot;where-this-leaves-the-2-camps&quot;&gt;Where this leaves the 2 camps&lt;/h2&gt;
&lt;p&gt;Asahi’s policy page and its report point to a project that treats legal exposure as the main threat to a reverse engineering effort, which explains the slow pace and the AI ban. The projects building on its work move faster and say so openly, with AI agents doing a lot of the writing. Both camps depend on Asahi’s reverse engineering, and only 1 of them can send code back, because Asahi will not accept AI assisted patches.&lt;/p&gt;
&lt;p&gt;The license settles a narrower question, whether a closed commercial product can be built on Asahi’s new code. After November 7, the answer for new code is no. For code already released under MIT, the answer stays yes.&lt;/p&gt;
&lt;h2 id=&quot;2-things-from-this-generalize&quot;&gt;2 things from this generalize&lt;/h2&gt;
&lt;p&gt;First, pick a license for the worst reader your project will ever have. A permissive license is a decision to let anyone do anything within a short list of conditions, including things you will hate. If you would be angry to see a closed fork succeed, the license should have said so on day 1, and a relicense later only protects what you write afterward.&lt;/p&gt;
&lt;p&gt;Second, keep policy and license apart. A license decides who may distribute your code. A contribution policy decides what you will accept. Asahi has both, and the report shows how much each one can carry. Only 1 of them can say anything about where a file came from.&lt;/p&gt;
&lt;p&gt;This week, open the LICENSE file of 1 project you maintain or depend on and write down 1 behavior it permits that you would not want to see. If you find one, decide whether the license or the contribution policy should change.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>linux</category><category>ai</category><category>artificial-intelligence</category><category>ai-agent</category><category>code</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>For 15 Years, KDE Told You a USB Copy Was Done When It Had Barely Started</title><link>https://soursen.org/articles/for-15-years-kde-told-you-a-usb-copy-was-done-when-it-had-barely-started/</link><guid isPermaLink="true">https://soursen.org/articles/for-15-years-kde-told-you-a-usb-copy-was-done-when-it-had-barely-started/</guid><description>A progress bar that timed the copy into memory, while the stick lagged far behind, corrupted files for anyone who trusted it and pulled the stick. The fix took 4 merge requests and a Linux call so sharp its own manual calls it extremely dangerous.</description><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On October 27, 2026, KDE developer Méven Car published &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blogs.kde.org/2026/10/27/fixing-kio-copy-to-usb-stick-and-network-filesystems/&quot;&gt;a writeup&lt;/a&gt; of a bug he called arguably the worst he has chased in KIO, the framework behind Dolphin and most file operations in KDE. He called it his white whale. The bug is simple to state. When you copied a file to a USB stick, the progress bar and the “copy finished” dialog tracked how fast data reached memory. The stick itself was far slower. On a slow stick, the dialog could say done in a tenth of a second while the drive kept writing for more than a minute. Pull the stick out when the dialog closes, and your file is corrupt. Do it as a move, and the file can disappear from both places. The report sat open for 15 years. I want to walk through why the progress bar lied, why a correct fix was so hard that it needed kernel features that did not exist in 2011, and what actually changed in KDE Frameworks 6.32.&lt;/p&gt;
&lt;h2 id=&quot;why-the-progress-bar-was-wrong&quot;&gt;Why the progress bar was wrong&lt;/h2&gt;
&lt;p&gt;A file manager copies a file by reading it and handing the data to the write system call. On Linux, write returns as soon as the data is in the page cache, the part of memory the kernel uses to hold data on its way to a disk. The kernel writes it out to the real device later, when it decides to.&lt;/p&gt;
&lt;p&gt;On a fast internal SSD, later means a fraction of a second, so you never notice. Copy from your main drive to a second internal drive and the write lands almost at once. On a USB 2.0 stick that writes a few megabytes a second, later can mean a minute or more for a large file. KIO handed the data to the kernel, saw write return, and reported the copy as finished. The data was sitting in RAM.&lt;/p&gt;
&lt;p&gt;The original 2011 report caught it in the plainest way. The notification claimed about 50 MB/s to a USB 1.1 device, which physically tops out near 1.5 MB/s. Alan Stern, the kernel USB developer who looked at it, summed the behavior up in 1 line. It shows the copying operation is finished when it has not even been started.&lt;/p&gt;
&lt;p&gt;Car measured the same thing in 2026 on a USB 2.0 stick running at about 2.3 MB/s. Copying a 128 MiB file, the completion dialog appeared after 0.06 to 0.13 seconds. The data actually reached the stick 68 to 91 seconds later, across 4 runs.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.B-nf8TuF_2aIfnE.webp&quot; alt=&quot;The dialog reported done almost instantly while the stick kept writing for over a minute.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The dialog reported done almost instantly while the stick kept writing for over a minute. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blogs.kde.org/2026/10/27/fixing-kio-copy-to-usb-stick-and-network-filesystems/&quot;&gt;Meven Car, KDE Blogs, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The progress bar had the same flaw. It showed how fast KIO filled memory, which meant it could sit at 100% with the entire 128 MiB file still unwritten. For a removable drive, that gap is where data dies. You see the dialog close, you unplug the stick, and the write never finished.&lt;/p&gt;
&lt;h2 id=&quot;why-a-move-was-worse-than-a-copy&quot;&gt;Why a move was worse than a copy&lt;/h2&gt;
&lt;p&gt;A copy that gets interrupted leaves a corrupt file on the stick, and the original is still safe on your drive. A move is different. KIO deleted each source file right after it believed the copy was done, which meant it deleted the original while the data was still in memory and not yet on the device. Unplug during that window and the file is gone from both places.&lt;/p&gt;
&lt;p&gt;This is the reason Nate Graham raised the bug to the highest priority in 2021. A wrong progress bar is annoying. A move that destroys the only copy is data loss. It took another 5 years to fix, because doing it correctly without wrecking performance turned out to be genuinely hard.&lt;/p&gt;
&lt;h2 id=&quot;the-tools-to-fix-it-did-not-exist-in-2011&quot;&gt;The tools to fix it did not exist in 2011&lt;/h2&gt;
&lt;p&gt;The naive fix is easy. Flush every file to the device as you copy it, using a call like fsync, which waits until the data is on the disk. The problem is speed. Car measured an fsync on every file on 3,000 files of 16 KiB. It took 10 seconds. Without it, the same copy took 0.04 seconds on a local disk.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.DmNg_ulI_ZYSOfj.webp&quot; alt=&quot;Flushing every file to disk is correct and roughly 250 times slower on small files.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Flushing every file to disk is correct and roughly 250 times slower on small files. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blogs.kde.org/2026/10/27/fixing-kio-copy-to-usb-stick-and-network-filesystems/&quot;&gt;Meven Car, KDE Blogs, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;So a correct fix has to flush removable and network copies without flushing internal ones, and without slowing a copy of many small files to a crawl. That needs a way to wait on one part of a file on its own. POSIX, the portable Unix standard, only offers fsync, which waits for the whole thing. Linux has sync_file_range, which can wait on a specific range of a file. Kernel USB developer Alan Stern suggested it in the bug report 2 weeks after it opened in September 2011, and KDE maintainer Christoph Feck replied that it looked like exactly what they needed.&lt;/p&gt;
&lt;p&gt;The call is powerful and sharp. Its own Linux manual page warns that it is extremely dangerous and that careless use can corrupt data, because it does not flush file metadata or the drive’s own hardware cache on its own. Car tried twice in 2024 and, by his own account, got it wrong both times. In 1 attempt the sync_file_range path stayed commented out. In another, the code waited on the chunk it had just written, which had not started writing yet, so it never waited at all. Waiting on the window before the one just written is what makes it work, and no reviewer caught the mistake either.&lt;/p&gt;
&lt;h2 id=&quot;how-the-fix-actually-works&quot;&gt;How the fix actually works&lt;/h2&gt;
&lt;p&gt;KDE’s answer sits between the 2 extremes. Flushing after every write would be correct and very slow. Flushing only at the very end would leave the progress bar wrong for the whole copy. So KIO now writes in windows.&lt;/p&gt;
&lt;p&gt;The data is copied in chunks called windows. When a window fills, a helper thread hands it to the kernel with sync_file_range and waits for it to reach the device, while the main thread fills the next window. The helper thread earns its place on Windows style CIFS shares, where starting a writeback blocks the calling thread for about 157 milliseconds per 128 MiB window while the server grants write credits, which would otherwise stall the whole copy. The copy only waits when the window it wants to hand over is not written yet. At most 2 windows are ever in memory and unwritten, so the progress bar is never more than 1 window ahead of what is really on the device. The window is adaptive, sized to hold about 200 milliseconds of writeback based on the measured speed of the link. It starts at 1 MiB, can at most double each step and caps at 16 MiB, or 128 MiB on CIFS. Car tuned that target by hand. A 500 millisecond window left SMB folder listings at 4.2 seconds, and dropping it to 200 milliseconds brought them to 1.7. A window too small throttles a fast link, and a window too large never fills on a slow one.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.B_Xl200N_4akn0.webp&quot; alt=&quot;After the fix, the progress bar leads the device by at most 1 window, where it once led by the whole file.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;After the fix, the progress bar leads the device by at most 1 window, where it once led by the whole file. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blogs.kde.org/2026/10/27/fixing-kio-copy-to-usb-stick-and-network-filesystems/&quot;&gt;Meven Car, KDE Blogs, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;A window bounds how much is left in memory during the copy, and there is still something to flush at the end. For a removable device, KIO now makes the end of the copy wait for it. A file larger than 16 MiB is flushed with fsync, then its folder, before the copy reports success. By then the windows have already written almost everything, so the final flush is short. The completion dialog now appears when the data is really on the stick. On the same USB 2.0 stick, that 128 MiB copy that used to report done in about a tenth of a second now reports done at 28.1 seconds, which is when the data has actually landed. The progress bar, which once ran 128 MiB ahead of the device, now leads it by 3 MiB.&lt;/p&gt;
&lt;p&gt;Small files are the hard part, because a flush per file is the 10 second disaster above. So the worker flushes no small file on its own. After copying to a removable disk, it runs a single syncfs, which writes back the whole destination file system in one operation, and the copy shows “writing the copied data to the device” while that runs. If the flush fails, the copy fails, because the data may not be on the device. The overhead is small. Those 3,000 small files took 185.2 seconds to actually land on the stick before the fix and 183.9 after, so the copy now reports 183.9 seconds and tells the truth. Before, it reported 1.2 seconds and left the rest to chance.&lt;/p&gt;
&lt;h2 id=&quot;a-move-now-deletes-nothing-until-the-data-is-safe&quot;&gt;A move now deletes nothing until the data is safe&lt;/h2&gt;
&lt;p&gt;The data loss path is closed. A move to a removable device keeps every source file until the flush has succeeded, then deletes them. If the flush fails, or you pull the stick during it, every source stays, so nothing is removed from your drive until the move is genuinely done. A copy or move that you cancel or that hits an error still flushes the device first, and a move then deletes only the sources of files that actually made it across. A partial move is safe.&lt;/p&gt;
&lt;h2 id=&quot;telling-a-usb-stick-from-an-internal-disk&quot;&gt;Telling a USB stick from an internal disk&lt;/h2&gt;
&lt;p&gt;None of this should touch internal drives, because you do not unplug them and waiting on every copy would just make them slower for no reason. Car checked that it does not. Forcing the removable path onto a local fixed disk changed a copy of those same 3,000 small files from 1.54 to 1.55 seconds, so the machinery costs nothing when it is not needed. Deciding what counts as removable is harder than it sounds. The file system type does not tell you. exFAT can be a memory card, and FAT is also the EFI partition on your laptop. Most USB disks do not even report themselves as removable media.&lt;/p&gt;
&lt;p&gt;KIO reads the answer from the system instead. A new private function, added in merge request 2506, asks sysfs, the kernel’s device information interface, for the disk’s own removable flag, then for the USB port or PCI slot it hangs off, which the kernel marks as removable when it can be reached from outside the machine. It follows encrypted and LVM volumes down to the physical disk underneath, and it also treats SD cards as removable. Car notes this will not catch every exotic case, but it handles the normal one, which is someone pulling a stick out of a USB port on the front of their machine. Network mounts that are already known to be slow get the same windowed treatment, including NFS, SMB, Ceph, GlusterFS, 9p, sshfs and FUSE mounts such as rclone.&lt;/p&gt;
&lt;p&gt;That detection has to be cheap, because it runs on every copy. 2 earlier changes to how KIO looks up mount points made it affordable. The first, in KIO 6.30, caches mounts by the id that the statx call returns on Linux 6.8 and later, which made lookups by id about 700 times faster. The second, in 6.32, uses a newer call named statmount that avoids walking the path at all, and dropped a lookup that could take more than 15 seconds on a busy system to about 1 millisecond.&lt;/p&gt;
&lt;p&gt;A second speed regression got fixed along the way. Back in 2021, KIO switched its copy path from the old sendfile call to a newer one, copy_file_range, which is faster. Then Linux 5.19 made copy_file_range refuse to copy between 2 different file systems, failing with an error code, so those copies fell back to a slow byte by byte path. A new change, merge request 2520, restores a sendfile fallback for that case. A 1 GiB copy from a RAM disk to an ext4 disk dropped from 443 to 272 milliseconds, and the reverse direction from 531 to 278.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.DSFNWIit_2fmchb.webp&quot; alt=&quot;The fix leans on cheaper kernel lookups and a restored sendfile fallback so it can check every copy without slowing things down.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The fix leans on cheaper kernel lookups and a restored sendfile fallback so it can check every copy without slowing things down. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blogs.kde.org/2026/10/27/fixing-kio-copy-to-usb-stick-and-network-filesystems/&quot;&gt;Meven Car, KDE Blogs, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;network-shares-had-the-same-lie-with-a-worse-twist&quot;&gt;Network shares had the same lie with a worse twist&lt;/h2&gt;
&lt;p&gt;A copy to an NFS or SMB share leaves a lot of data in memory waiting to go to the server, so the progress bar runs ahead of what the server has received, the same flaw as the USB case. SMB adds a nastier twist. With its default setting, cache equals strict, even looking at a file waits until that file’s unwritten data has been sent. So listing the folder you are copying into makes the file manager wait behind the upload.&lt;/p&gt;
&lt;p&gt;Car copied a 256 MiB file to an SMB share over a 5 MB/s link and timed folder listings during the copy. Before the fix, the slowest listing took 27.4 seconds. After, the slowest took 1.70 seconds, because a listing now waits behind at most 2 windows of writeback, where it used to wait behind the whole queued upload. On a 4 GiB NFS copy, the median listing fell from 24 milliseconds to 11. The gap shows up even on a fast link, where the stakes are lower. On a 4 GiB SMB copy over a quick connection, the progress bar used to run as much as 1,334 MiB ahead of what the server had received. After the fix it leads by at most 128 MiB, 1 CIFS window, and the copy itself finished slightly faster, 6.06 against 7.15 seconds.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.CdS1G1QD_QBCbM.webp&quot; alt=&quot;On a slow SMB link, browsing the share during a copy went from almost half a minute to under 2 seconds.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;On a slow SMB link, browsing the share during a copy went from almost half a minute to under 2 seconds. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blogs.kde.org/2026/10/27/fixing-kio-copy-to-usb-stick-and-network-filesystems/&quot;&gt;Meven Car, KDE Blogs, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;this-was-never-only-a-kde-problem&quot;&gt;This was never only a KDE problem&lt;/h2&gt;
&lt;p&gt;The habit of reporting a cached write as a finished one runs through the whole Linux stack, which is why this is worth understanding even if you never touch KDE. The command line tool dd behaves the same way. Write an ISO to a USB stick with dd, watch it return to the prompt, pull the stick, and the image can be unbootable, because dd reported done when the data was in cache. The fix there is to add conv=fsync or to run sync afterward, which is exactly why the Arch Linux install guide tells you to run sync before removing the drive.&lt;/p&gt;
&lt;p&gt;Most other graphical file managers share the original flaw too. GNOME’s Files and the newer Cosmic file manager report a copy as complete once the data is in memory. The usual safety net is the eject or “safely remove” action, which does force the flush. The trouble is that plenty of people, myself included, just pull the stick when the dialog disappears. The value of the KIO fix is that it makes the dialog mean what people already assume it means.&lt;/p&gt;
&lt;h2 id=&quot;what-the-fix-cannot-fix&quot;&gt;What the fix cannot fix&lt;/h2&gt;
&lt;p&gt;One measurement during testing sent Car down a rabbit hole worth repeating, because it marks the edge of what a file manager can promise. While copying the small files to a cheap USB stick, listing a folder on that stick sometimes took about 15 seconds, with the new window turned on or off. The window was not the cause. The stick was. Roughly every 8 seconds, a single 120 KiB write to it took 4 to 7 seconds to complete, and the folder listing was stuck waiting on a page queued behind that write. KIO cannot fix a device that stalls for seconds at a time. What it can do is stop blaming the copy, and tell you honestly when your data is on the drive. It cannot make a slow stick write faster.&lt;/p&gt;
&lt;h2 id=&quot;why-it-took-15-years&quot;&gt;Why it took 15 years&lt;/h2&gt;
&lt;p&gt;A wrong progress bar hid a data loss bug for 15 years, and the delay came down to missing tools. The right call was named 2 weeks in, back in 2011, and the kernel features to use it safely and cheaply arrived much later. sync_file_range existed but was dangerous to use correctly. The cachestat call that reports how much of a file is still dirty in memory landed in Linux 6.5 in 2024, and it works 1 file at a time. The cheap statmount based mount lookups are recent too. The fix needed a developer willing to thread all of that through KIO’s copy path and get the ordering exactly right, after 2 failed attempts that passed code review.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-06.3uGxlRSX_Z1OlaUP.webp&quot; alt=&quot;From the first report to the shipped fix, the bug outlived KDE 4, the whole of KDE 5 and several kernel releases.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;786&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;From the first report to the shipped fix, the bug outlived KDE 4, the whole of KDE 5 and several kernel releases. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blogs.kde.org/2026/10/27/fixing-kio-copy-to-usb-stick-and-network-filesystems/&quot;&gt;Meven Car, KDE Blogs, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-this-means-for-you&quot;&gt;What this means for you&lt;/h2&gt;
&lt;p&gt;2 things from this are worth keeping. The first is that on anything that writes slowly, a progress bar tells you about memory, and only the eject action proves the data reached the device. Until you are on KDE Frameworks 6.32 or later, treat every “copy finished” dialog on a USB stick or a network share as a claim about memory, and use the eject or safely remove action before you unplug. It forces the flush that the dialog did not wait for. On the command line, add conv=fsync to dd or run sync, and wait for it to return before pulling the drive.&lt;/p&gt;
&lt;p&gt;The second is that the fix ships in KDE Frameworks 6.32, which is tagged and releasing on November 13, 2026, with the groundwork already in 6.30 from September. If you use Dolphin with USB sticks or network shares, this is the update that makes the copy dialog trustworthy. After it lands, a copy to a stick finishes when the data is actually on the stick, a move deletes nothing from your drive until the write has succeeded, and a network share stays usable while a copy runs. At that point you can close the dialog and unplug, which is all anyone wanted in 2011.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>kde</category><category>linux</category><category>llm</category><category>ai-agent</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Linux 7.4 Had the Biggest Merge Window in History, and Linus Torvalds Wrote Some of It With Claude</title><link>https://soursen.org/articles/linux-7-4-had-the-biggest-merge-window-in-history-and-linus-torvalds/</link><guid isPermaLink="true">https://soursen.org/articles/linux-7-4-had-the-biggest-merge-window-in-history-and-linus-torvalds/</guid><description>The kernel just took in nearly 20,000 commits in one cycle, a record. The reason is the same force straining its maintainers, AI. The release is the clearest look yet at what that help gives and what it costs.</description><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Linux 7.4 landed the largest merge window in kernel history. The numbers tell the scale plainly, 19,671 commits in a single cycle, past the old record of 16,267 set by 7.3, against a source tree now approaching 42 million lines. A lot of that growth traces to one cause, AI tools writing and suggesting code faster than people used to. I want to walk through this release carefully, because it is the clearest picture we have of what AI is actually doing inside the most important open source project on Earth, the good and the mess, and because the person demonstrating the good is Linus Torvalds himself.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DvPw4NQA_Z1pUJ13.webp&quot; alt=&quot;Linux 7.4 took 19,671 commits, past the old record of 16,267. Maintainers tie much of the rise to AI tools.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Linux 7.4 took 19,671 commits, past the old record of 16,267. Maintainers tie much of the rise to AI tools. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-7.4-Massive-Window&quot;&gt;Phoronix, kernel merge notes, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;torvalds-started-writing-code-again-with-claude&quot;&gt;Torvalds started writing code again, with Claude&lt;/h2&gt;
&lt;p&gt;For years Torvalds has mostly reviewed code rather than written it, saying yay or nay on what makes it into the kernel. In 7.4 he showed up as a contributor again, and he used &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.anthropic.com/claude&quot;&gt;Claude&lt;/a&gt; to do it.&lt;/p&gt;
&lt;p&gt;The cleanest example is a piece of housekeeping he had wanted done for a long time. He removed two old kernel functions, `__get_user()` and `__put_user()`, that move a single value between user space and the kernel. The versions with the underscores skip the pointer validation that the safe versions do, so they are faster and more dangerous, and modern helpers already do the job better. Torvalds described using Claude for the cleanup as treating it like &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linus-Kills-Get-Put-User-Linux&quot;&gt;a fancy sort of Coccinelle script&lt;/a&gt;, referring to a tool kernel developers use for automated source transformations. His own commit message was blunt about why the functions had to go. “The __get_user() and __put_user() functions are entirely historical,” he wrote, and “when mis-used (read: every case) they are not only slower than the right thing.”&lt;/p&gt;
&lt;p&gt;That was not his only AI assisted work in the cycle. The week before, he used Claude Code on a performance idea for the case where several CPU threads keep reading small amounts of data from the same file. He came up with the approach and wrote the boilerplate, Claude reworked it to cut unnecessary memory reference counting so CPUs can cache data more efficiently, and a Meta engineer finished and tested it. He also pointed Claude at the kernel’s associative array library, cleaned it up, and turned up three real bugs that had been sitting there since the library was first added in 2013.&lt;/p&gt;
&lt;h2 id=&quot;torvalds-thinks-the-model-for-ai-help-got-healthier&quot;&gt;Torvalds thinks the model for AI help got healthier&lt;/h2&gt;
&lt;p&gt;What makes this more than a stunt is the stance behind it. Torvalds has moved from wary to cautiously positive, saying the kernel is now using a much healthier model for AI contributions than the panic of a year ago suggested. His position is consistent. An experienced developer who understands the code can point an AI at legacy cruft, clean it up, and surface hidden bugs, and that is a real gain. The trouble starts when the person steering does not have that understanding, which is exactly what the rest of this release shows.&lt;/p&gt;
&lt;h2 id=&quot;the-performance-wins-are-large-and-specific&quot;&gt;The performance wins are large and specific&lt;/h2&gt;
&lt;p&gt;Before the mess, the wins, because 7.4 is stacked with them. The one that made me do a double take is the FAT32 file system. A small change to how it preallocates space, allocating multiple clusters in a single operation instead of one cluster at a time, made preallocation up to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-7.4-Faster-FAT32&quot;&gt;265 times faster&lt;/a&gt; in testing. A handful of modified lines, four files changed, and a decades old file system got hundreds of times quicker at one job.&lt;/p&gt;
&lt;p&gt;Networking got one of the biggest swings. Reworking how the stack handles mixed big TCP and standard TCP traffic pushed throughput up to 70% higher, alongside hardware pacing offload that lets certain network cards manage outgoing packet timing and take load off the CPU. On the scheduler side, a change to the EEVDF scheduler for short slice tasks, the quick bursts of CPU work that pile up on machines with huge core counts, cut latency by up to 55% on hardware like Nvidia’s Vera CPUs and high core count servers.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.BMuEhndZ_Zeu8rl.webp&quot; alt=&quot;Four of the headline speedups in Linux 7.4. Each one comes from a focused change rather than a rewrite.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Four of the headline speedups in Linux 7.4. Each one comes from a focused change rather than a rewrite. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-7.4-Massive-Window&quot;&gt;Phoronix, kernel merge notes, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;There is more underneath. Intel’s Battle Mage GPUs got ultra low latency scheduling that handles memory page faults better, trimming 30 to 40 microseconds off GPU copy operations. RISC-V gained runtime detection of efficient unaligned memory access, so distributions can get an 11% memory copy bandwidth gain without recompiling the kernel for specific chips. AMD Zen 6 groundwork lets the kernel drop the expensive SRSO security mitigation on chips that provide branch target buffer isolation in hardware, keeping the protection while removing its cost.&lt;/p&gt;
&lt;h2 id=&quot;the-speedup-ai-found-by-writing-ugly-code&quot;&gt;The speedup AI found by writing ugly code&lt;/h2&gt;
&lt;p&gt;The performance story that fits this release best is the kernel build itself. Lorenzo Stoakes, an ARM engineer and longtime kernel developer, posted a series of 23 patches to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/AI-To-Faster-Linux-Kernel-Comp&quot;&gt;speed up kernel compilation&lt;/a&gt;. He used an AI model to find the bottlenecks and then to work out fixes, and he was honest about the output. “It generated a lot of code, much of it hideous,” he wrote. The AI found single threaded steps across Kbuild, kallsyms, modpost, objtool, and the Rust build system that could be parallelized.&lt;/p&gt;
&lt;p&gt;The results are large. Full kernel builds run about 36% faster, incremental builds up to 70% faster, and builds with nothing to do up to roughly 90% faster. For people who rebuild the kernel dozens of times a day, that is a real change to the working day. Stoakes was careful to add that he checked correctness himself in both the build and the running kernels, which is the pattern Torvalds keeps describing, AI does the grunt work, a human who knows the system signs off.&lt;/p&gt;
&lt;h2 id=&quot;the-ai-flood-the-maintainers-are-drowning-in&quot;&gt;The AI flood the maintainers are drowning in&lt;/h2&gt;
&lt;p&gt;The other side of the same force is ugly. The sound subsystem maintainer opened his 7.4 pull by calling it, plainly, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-7.4-Sound&quot;&gt;the AI flood&lt;/a&gt;. Networking maintainers have described themselves as completely overwhelmed. AI coding agents are generating a torrent of kernel patches and bug reports, and a large share is low quality, unrealistic, or plain wrong, which someone experienced still has to read and reject.&lt;/p&gt;
&lt;p&gt;One feature in 7.4 captures the problem exactly. Greg Kroah-Hartman, effectively the second in command of the kernel, wrote a feature called Taint Force bind meant to flag the kind of nonsense configurations that AI assisted fuzzers were generating, where tools bind hardware devices to incompatible drivers at random and then file bug reports about the resulting crashes. The feature shipped, then broke legitimate GPU and PCI Express passthrough setups, so developers reverted it in the same release. An attempt to filter out AI generated noise created its own regression and had to be pulled. That is the tax in one story.&lt;/p&gt;
&lt;h2 id=&quot;everything-else-that-landed&quot;&gt;Everything else that landed&lt;/h2&gt;
&lt;p&gt;The rest of the release is the usual broad sweep that makes Linux run on nearly anything. Rust support grew by about 17,000 lines, including support for self referential data structures that make certain complex drivers safer to write. Valve funded a new driver for Nintendo’s Switch 2 controllers, including the Joy-Con 2 and the Switch 2 Pro Controller. Apple’s M4 and A18 chips got initial mainline support, much of it from the Asahi Linux project, a first step toward booting Linux on recent Macs. HDMI 2.1 finally got enabled by default in the AMD GPU driver after years of fighting the HDMI Forum’s licensing restrictions. The kernel even keeps fixing memory safety bugs in its floppy disk driver, three decades of hardware supported in one tree.&lt;/p&gt;
&lt;h2 id=&quot;what-this-release-actually-tells-you&quot;&gt;What this release actually tells you&lt;/h2&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.U6CjQ9VD_ThD9S.webp&quot; alt=&quot;The same cycle produced both outcomes. The deciding factor was whether someone who understood the code was steering.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The same cycle produced both outcomes. The deciding factor was whether someone who understood the code was steering. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-7.4-Sound&quot;&gt;Phoronix, kernel merge notes, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Strip away the individual features and 7.4 is a single argument with two sides. AI in experienced hands produced a record cleanup from Torvalds, a 265x FAT32 speedup, and a 36% faster kernel build. AI in careless hands produced a flood of junk patches that maintainers now spend real time filtering, and at least one well meaning filter that broke working hardware. The kernel did not choose between those outcomes. It got both at once, in the same cycle, and the difference every time came down to whether a person who understood the code was the one steering. That is the model Torvalds says is working, and 7.4 is the proof and the warning in one release.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>linux</category><category>ai</category><category>artificial-intelligence</category><category>linus-torvalds</category><category>claude</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Linux Deleted a $10 Billion Intel Chip in One Commit. Itanium Took 23 Years to Die</title><link>https://soursen.org/articles/linux-deleted-a-10-billion-intel-chip-in-one-commit-itanium-took-23/</link><guid isPermaLink="true">https://soursen.org/articles/linux-deleted-a-10-billion-intel-chip-in-one-commit-itanium-took-23/</guid><description>Intel and HP spent roughly $10 billion building Itanium, the chip that was supposed to replace x86 forever. Analysts projected $38 billion in sales by 2001. In Linux 6.7, released in January 2024, the kernel deleted about 65,000 lines of Itanium support in one go, because nobody was left to maintain it. I went through why the most expensive bet in CPU history ended as a pile of deletions, and what it taught Linus Torvalds about anyone who promises the future needs a clean break.</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;An entire CPU architecture got deleted in a single commit. That is the short version of how Itanium ended. In &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Intel-IA-64-Removed-Linux-6.7&quot;&gt;Linux 6.7&lt;/a&gt;, which shipped in January 2024, the kernel removed every line of support for Intel’s IA-64, about 65,000 lines, because not one person was left who could test it or answer for it. The chip had cost Intel and HP something like $10 billion to build, it was once projected to own the server market, and it exited the most used operating system in the world as a routine cleanup.&lt;/p&gt;
&lt;p&gt;I want to walk through how that happens, because the gap between the ambition and the ending is enormous, and the reasons are specific.&lt;/p&gt;
&lt;h2 id=&quot;the-idea-was-genuinely-clever&quot;&gt;The idea was genuinely clever&lt;/h2&gt;
&lt;p&gt;In the early 1990s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Hewlett-Packard&quot;&gt;HP&lt;/a&gt; looked at its own PA-RISC chips and at x86 and decided neither would keep scaling. So in 1994 it teamed up with &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Intel&quot;&gt;Intel&lt;/a&gt; to build something new. The idea was a real bet on a different way to run code.&lt;/p&gt;
&lt;p&gt;A normal processor works out, while it runs, which instructions can happen at the same time. That is out of order execution, and the scheduling hardware that does it costs transistors and power. Itanium’s pitch was to skip all that. Let the compiler decide ahead of time which instructions can run in parallel, pack them into bundles, and let the chip just execute what it was handed. A simpler chip paired with a smarter compiler. They called the approach &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Explicitly_parallel_instruction_computing&quot;&gt;EPIC&lt;/a&gt;, explicitly parallel instruction computing.&lt;/p&gt;
&lt;p&gt;On paper it was elegant. Move the hard thinking from the silicon, where it is expensive, to the compiler, where you only pay for it once. The whole industry believed it. HP and Intel announced &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Itanium&quot;&gt;Itanium&lt;/a&gt; in 1997 and promised chips in 1999. The analyst firm IDC projected around $38 billion in Itanium sales by 2001. Vendors wound down their own chip lines to get ready for it. Compaq dropped its Alpha line. Only one major player refused to climb aboard, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/AMD&quot;&gt;AMD&lt;/a&gt;, which turned out to matter later.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DN6ubNiF_27C1N7.webp&quot; alt=&quot;The forecast against the reality. IDC projected $38 billion in Itanium sales by 2001. The real number was a rounding error next to it.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The forecast against the reality. IDC projected $38 billion in Itanium sales by 2001. The real number was a rounding error next to it. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Itanium&quot;&gt;IDC 1997 forecast, contemporary reporting&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-compiler-could-not-do-the-one-thing-it-had-to-do&quot;&gt;The compiler could not do the one thing it had to do&lt;/h2&gt;
&lt;p&gt;The first chip, codenamed Merced, shipped in 2001, 2 years late and slow. The press already had a nickname ready. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.theregister.com/&quot;&gt;The Register&lt;/a&gt; had been calling it the Itanic. A slow launch alone does not kill a chip with this much money behind it. The design itself is where the trouble was.&lt;/p&gt;
&lt;p&gt;The whole scheme depended on the compiler knowing, at compile time, how long every memory access would take, so it could pack the bundles efficiently. That timing depends on what is sitting in the cache at that exact moment, which the chip knows while it runs and the compiler cannot know in advance. Asked to schedule something it could not predict, the compiler did the safe thing. It filled the bundles with empty instructions that do nothing, called no ops, so it would not break anything. A large part of the chip’s width ended up executing air. The elegant idea needed a compiler that could see the future, and nobody knew how to write one.&lt;/p&gt;
&lt;p&gt;The second problem was compatibility. Itanium could run existing x86 software, but through emulation, at a fraction of native speed. So the pitch to a customer was to buy an expensive new chip that ran the programs they already owned worse than the cheap chip they already had. That is a hard sell, and it left a door wide open.&lt;/p&gt;
&lt;h2 id=&quot;amd-walked-through-the-door-intel-left-open&quot;&gt;AMD walked through the door Intel left open&lt;/h2&gt;
&lt;p&gt;In 2003 AMD did the obvious thing Intel had refused to do. It stretched x86 to 64 bits and kept everything that already ran at full speed. It called the result &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/X86-64&quot;&gt;AMD64&lt;/a&gt;. No clean break, no throwing away your software, just more address space on top of the instruction set the whole world already used. Linux supported it within months.&lt;/p&gt;
&lt;p&gt;Then the decisive move. Microsoft agreed to build Windows for AMD64 and told Intel it would not ship a second, separate 64-bit version of Windows for a different instruction set. Intel was cornered. It could keep fighting its own customers with Itanium, or adopt its rival’s design. In 2004 it did the second thing, copying AMD’s extension and selling it in its own chips as EM64T, later renamed Intel 64.&lt;/p&gt;
&lt;p&gt;That decision is why, if you open the system details on almost any Intel machine today, the architecture is listed as x86–64 or AMD64. Intel’s own processors carry a competitor’s name on the instruction set, because the chip Intel built to kill x86 instead made AMD’s version of x86 the permanent standard.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.CxITDBjx_Z1pb2t1.webp&quot; alt=&quot;The timeline, from the 1994 partnership to the 2024 kernel deletion. The chip outlived its own relevance by about 20 years.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;1493&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The timeline, from the 1994 partnership to the 2024 kernel deletion. The chip outlived its own relevance by about 20 years. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Itanium&quot;&gt;Intel, HP, kernel history&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-deeper-reason-the-bet-was-doomed&quot;&gt;The deeper reason the bet was doomed&lt;/h2&gt;
&lt;p&gt;The compiler idea could not win for a structural reason, not because nobody tried hard enough. The whole premise was that a smart compiler could schedule instructions better than the chip could at runtime. That works only when the compiler can predict how the program will behave, and the single biggest variable in modern performance is memory latency, how long it takes to fetch data. A value already in the fast cache comes back in a few cycles. A value that has to come from main memory can take hundreds. Whether it is in the cache depends on everything the program did just before, which changes run to run and input to input.&lt;/p&gt;
&lt;p&gt;A compiler sees the program once, at build time, with no idea what will be in the cache when a given instruction runs. The chip sees the real state every cycle. So the very thing Itanium moved into the compiler, scheduling, is the thing that most needs live information the compiler does not have. Out of order execution on a normal chip is not wasteful overhead to be optimized away. It is the chip reacting to reality as it unfolds, which a static schedule cannot do. That is the part Itanium’s elegant diagram left out, and no amount of compiler research closed the gap.&lt;/p&gt;
&lt;h2 id=&quot;linux-carried-it-for-23-years&quot;&gt;Linux carried it for 23 years&lt;/h2&gt;
&lt;p&gt;Here is the part I did not expect. Linux ran on Itanium before the chip even shipped, because the industry paid to make it happen. In 1999 Intel, HP, and a group of Linux companies including Cygnus, VA Linux, SGI, Red Hat, and SUSE formed the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Trillian_project&quot;&gt;Trillian Project&lt;/a&gt; to port Linux to IA-64 ahead of the hardware, and they released the source in early 2000, a year before Merced reached customers. Debian, Red Hat, and SUSE all shipped it. For a while it was a real, supported platform with real money behind it.&lt;/p&gt;
&lt;p&gt;Then the slow retreat. Red Hat announced in December 2009 that it would drop Itanium in RHEL 6, which shipped in 2010, saying the number of Itanium servers sold was too low to justify the support. Debian carried it one release longer, keeping it in Debian 7 and dropping it from Debian 8 Jessie in 2015. Intel took its last Itanium orders by January 2020 and shipped the final chips, the Itanium 9700 Kittson, by July 2021. HP kept supporting its own Unix on the hardware through the end of 2025. By 2023, though, nobody on the kernel mailing list was answering for IA-64 anymore. The architecture was marked Orphaned in the kernel’s own MAINTAINERS file, which is the formal way of saying no company and no volunteer will speak for this.&lt;/p&gt;
&lt;p&gt;In September 2023 the engineer Ard Biesheuvel &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-6.7-To-Drop-Itanium-IA-64&quot;&gt;proposed removing it&lt;/a&gt;, with a patch note that said simply that the Itanium architecture is obsolete. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Linus_Torvalds&quot;&gt;Linus Torvalds&lt;/a&gt; accepted it, and the removal shipped in Linux 6.7 in January 2024. About 65,000 lines, gone.&lt;/p&gt;
&lt;h2 id=&quot;why-delete-working-code-at-all&quot;&gt;Why delete working code at all&lt;/h2&gt;
&lt;p&gt;Plenty of people assume old code gets deleted because it is old. That is not the rule, and Linux proves it constantly. The kernel only recently retired support for the Intel 486, a chip from 1989. It carried a graphics driver dating back to 1982 for years. Age alone does not get code removed.&lt;/p&gt;
&lt;p&gt;The actual rule is about testing. When nobody can run the code, every change to the kernel risks breaking it silently, and nobody will notice until it is too late. Every supported architecture is a tax on everyone else. Anyone who changes a core kernel interface has to update every architecture it touches, including the ones nobody owns or tests. The asymmetry tells the whole story. Itanium support was added because HP paid for it. It was removed because nothing showed up, nobody answered the mail, and the cost of pretending it still worked outgrew any benefit.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.BU2YlvRV_1fVE4r.webp&quot; alt=&quot;Age does not get code deleted. A lack of anyone to test and answer for it does.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;853&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Age does not get code deleted. A lack of anyone to test and answer for it does. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-6.7-To-Drop-Itanium-IA-64&quot;&gt;kernel maintainer discussion, 2023&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;torvalds-saw-it-early-and-the-history-proved-him-right&quot;&gt;Torvalds saw it early, and the history proved him right&lt;/h2&gt;
&lt;p&gt;Torvalds had been skeptical of Itanium’s whole premise for a long time. Back in February 2003, in a kernel mailing list debate, he put it bluntly. He said Intel had thrown out all the good parts of the x86 because people thought those parts were ugly, and that those were the parts that actually make x86 perform. He added that the Itanium 2 did not hold a candle to a Pentium 4 on any real world benchmark, and that IA-64 simply could not do the things a Pentium 4 did every single day. Information Age ran his view under the headline that Itanium is horrible, says Torvalds. The complaint underneath was about the arrogance of the opening move, asking everyone to throw away the software they already ran in exchange for a promise that the hardware never delivered.&lt;/p&gt;
&lt;p&gt;That distrust has aged well, and it generalizes. Any time a company tells you the future requires a brand new architecture and a clean break from everything you own, the burden is on them to make the switch painless, and they almost never do. The lesson of Itanium is that compatibility is not a limitation to escape. It is most of the value.&lt;/p&gt;
&lt;h2 id=&quot;the-idea-did-not-actually-die&quot;&gt;The idea did not actually die&lt;/h2&gt;
&lt;p&gt;The strange epilogue is that Itanium’s core idea outlived the chip. The notion of letting software decide the schedule ahead of time, instead of the hardware working it out live, came back in places where it fits. Parts of the GPU and AI accelerator world lean on exactly that approach, because their data patterns are predictable enough that you really can plan ahead. The idea works where the workload is predictable. It fell apart on general purpose CPUs running unpredictable code, which is the one job Itanium was built for.&lt;/p&gt;
&lt;p&gt;The other half of the Itanium story came back too, and got it right. Moving to a brand new architecture with emulation for old software is exactly what Apple did with its move to Arm, and what Qualcomm is doing in Windows laptops. The difference is that their emulation is fast and native software arrived quickly, which is the thing Itanium never managed. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/RISC-V&quot;&gt;RISC-V&lt;/a&gt; takes the opposite lesson in a different way, with an open instruction set that no single company can decide to abandon, which is its own answer to the problem of one owner walking away.&lt;/p&gt;
&lt;p&gt;Intel spent roughly a decade and around $10 billion trying to retire x86. What it left behind is a competitor’s name on its own chips and about 65,000 lines of deleted kernel code. Linux carried the thing for 23 years, longer than Intel itself kept wanting it, and let it go the day nobody answered the mail.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>linux</category><category>intel</category><category>hp</category><category>ai</category><category>artificial-intelligence</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>AI Is Quietly Dismantling the Thing That Kept People on Windows and macOS</title><link>https://soursen.org/articles/ai-is-quietly-dismantling-the-thing-that-kept-people-on-windows-and/</link><guid isPermaLink="true">https://soursen.org/articles/ai-is-quietly-dismantling-the-thing-that-kept-people-on-windows-and/</guid><description>For 30 years the reason you could not simply switch to Linux was the software. Photoshop, Office, the apps you needed and the file formats they locked you into. In October 2026 a single developer used Claude Code to rebuild Adobe’s core creative apps in Rust and gave them away. I think the moat that protected proprietary software just sprang a leak, and the legal fight over whether that is allowed is going to be the real story.</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;The pitch for Linux has always hit the same wall. You tell someone their machine would run faster, respect their privacy, and cost nothing, and they ask one question. Does it run Photoshop. The honest answer, for 30 years, was no, or only through a compatibility layer that might break the day Adobe shipped an update. That one gap, the software people need and cannot get on Linux, is the entire reason Windows and macOS still own the desktop.&lt;/p&gt;
&lt;p&gt;That wall is starting to come down, and the thing knocking it over is the same AI that the big software companies spent the last three years pushing on everyone. I find that funny, and I also think it is going to get messy fast.&lt;/p&gt;
&lt;h2 id=&quot;what-a-single-developer-did-in-october-2026&quot;&gt;What a single developer did in October 2026&lt;/h2&gt;
&lt;p&gt;In early October 2026 a developer named Brandon Thomas, the former CEO of a startup called storyteller.ai, released a project called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.omgubuntu.co.uk/2026/10/artcraft-free-photoshop-ai-clone&quot;&gt;ArtCraft&lt;/a&gt;, a set of open source apps meant to replace the Adobe Creative Cloud suite. Photocraft stands in for Photoshop, Filmcraft for Premiere, and Lightcraft for Lightroom. By his own account on Reddit, Thomas built it because he got fed up with Adobe’s fees and the hoops it made him jump through just to cancel his subscription. That is a petty reason to start, and petty reasons ship more software than noble ones do.&lt;/p&gt;
&lt;p&gt;The apps are written in Rust, they are free, and they ship builds for Linux, Windows, and macOS, including a native Linux client, which is the thing Adobe never bothered to make in 30 years. The cross platform work is real, not a wrapper. Photocraft uses Vulkan for GPU acceleration on Linux, Metal on macOS, and DX12 on Windows, with a WebAssembly build that runs in a browser through WebGPU. The code is on GitHub, and Ubuntu users can install it from a DEB package on Intel, AMD, or Arm.&lt;/p&gt;
&lt;p&gt;The detail that matters is how it was built. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Photocraft-Filmcraft-AI&quot;&gt;Phoronix reported&lt;/a&gt; that the apps were made with Claude Code and AI agents, and that the project leans on AI for most of the work. The project describes itself as a clean room reimplementation, though that term is doing a lot of load bearing work and I will come back to why.&lt;/p&gt;
&lt;p&gt;The same wave produced office suite clones that open and save Microsoft Office formats, and a separate effort called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Starling-Swift-Desktop&quot;&gt;Starling&lt;/a&gt;, a full Linux desktop environment with its own Wayland compositor, written in Swift and built by one person in about 6 months with AI help. None of these are finished. ArtCraft’s Photoshop compatibility is partial, the office clones are further along on spreadsheets than on presentations, and the whole category is alpha quality right now. Treat all of it as early.&lt;/p&gt;
&lt;h2 id=&quot;why-this-was-impossible-until-about-a-year-ago&quot;&gt;Why this was impossible until about a year ago&lt;/h2&gt;
&lt;p&gt;Cloning software is not new. You could always decompile a program, study how it works, and rewrite the same features. The problem was time. Reimplementing Photoshop by hand would take a competent team 10 to 15 years, and by the time they finished, Photoshop had 15 more years of features to catch up on. The target moved faster than any volunteer effort could run. That gap is exactly what made the moat hold.&lt;/p&gt;
&lt;p&gt;Here is the mechanism that changed. A model like the one behind Claude Code can read a feature, understand the behavior, and produce a working implementation in an afternoon instead of a month. Stack enough of those afternoons together and a job that was a decade of human work compresses into a few months. The speed is the whole point. It does not make the clone perfect, what it does is make the clone possible at all, and possible is the word that breaks the moat.&lt;/p&gt;
&lt;p&gt;The moat was never really the code. These companies built their defense on one fact, that copying all the work they paid developers to do would cost about as much as they paid to build it the first time. Once that cost falls by 2 orders of magnitude, the defense is gone. You can step over the wall instead of climbing it.&lt;/p&gt;
&lt;p&gt;We have 30 years of evidence for how hard that wall was. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/GIMP&quot;&gt;GIMP&lt;/a&gt; has been trying to be a free Photoshop since 1996 and, after all that time, still is not one for professionals, because a volunteer project cannot keep pace with a company that pays hundreds of engineers full time. LibreOffice and its ancestor OpenOffice have chased Microsoft Office compatibility for over 20 years and still trip on complex documents. Inkscape, Krita, Darktable, each is good, each took a decade or more, and none of them erased the lock in, because the proprietary tool kept moving and the free one could never fully catch up. That track record is the backdrop for why ArtCraft reaching rough parity on parts of Photoshop in weeks reads as a different kind of event. The thing that always made these efforts lose, time, is the exact thing that changed.&lt;/p&gt;
&lt;h2 id=&quot;the-moat-was-3-separate-walls-and-ai-hits-all-3&quot;&gt;The moat was 3 separate walls, and AI hits all 3&lt;/h2&gt;
&lt;p&gt;People talk about Linux adoption as one problem. It is actually 3, and the clone wave hits each one.&lt;/p&gt;
&lt;p&gt;The first wall is applications. If your job needs Photoshop or Office, you cannot leave the platform that runs them. Vibe coded clones attack this directly, same features, same file formats, same keyboard shortcuts, running on any OS you like.&lt;/p&gt;
&lt;p&gt;The second wall is the desktop itself. Even with perfect app clones, Linux still looks and behaves differently from Windows or macOS, and for a lot of people that alone is a dealbreaker. They cannot find their settings, the muscle memory is wrong, it feels foreign. A shell is far simpler to replicate than Photoshop though. It shows icons, opens apps, manages files, and you have the exact reference sitting right there to copy click for click. Starling proving that one person can build a whole desktop in 6 months with AI tells you a Windows style or macOS style shell on Linux is a matter of when, not if.&lt;/p&gt;
&lt;p&gt;The third wall is hardware. Drivers for Linux have always lagged because writing them by hand is slow and thankless. A hardware maker that once needed a dedicated team for weeks to write a Linux driver can now get a first draft from an AI, test it, and submit it. The initial work got cheap.&lt;/p&gt;
&lt;p&gt;The kernel itself made room for this, carefully. Linus Torvalds spent years dismissing AI, then changed his tune. In mid 2026 he told critics the kernel is not an anti AI project and called demands to ban AI generated patches pointless posturing. The project then adopted a formal policy. AI assisted patches are allowed, but they cannot carry the usual Signed off by tag that certifies authorship. They carry an Assisted by tag instead, so the use of AI is on the record, and the human who submits the patch stays fully responsible for its bugs, security holes, and licensing. Not everyone went this way. Gentoo and NetBSD banned AI generated contributions outright, and maintainers have already rejected a wave of low quality AI submissions to the driver staging area as slop. The door is open, with a bouncer on it, which is roughly the right setting for letting AI speed up driver work without drowning the kernel in garbage.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.D9_gJ_5w_Z1wty02.webp&quot; alt=&quot;The 3 walls that kept users on Windows and macOS, and how the clone wave hits each one.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;892&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The 3 walls that kept users on Windows and macOS, and how the clone wave hits each one. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Photocraft-Filmcraft-AI&quot;&gt;Soursen analysis&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-gaming-side-moved-too&quot;&gt;The gaming side moved too&lt;/h2&gt;
&lt;p&gt;Games were always a separate lock, and it is loosening from 2 directions. In 2026 Epic released a rebuilt Epic Games Store launcher that it openly called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://hothardware.com/news/epic-claims-vibe-coded-pc-launcher-dusts-steam-speed-tests&quot;&gt;vibe coded&lt;/a&gt;, AI assisted, and for the first time it runs natively on Linux. Epic claimed the new launcher reached the store page about 46% faster than Steam on a cold start. It is buggy, because it was built fast, but a native Linux client from a company that size is a real signal.&lt;/p&gt;
&lt;p&gt;From the other direction, game decompilation got dramatically easier, and people are porting console games to PC that were never meant to leave the console. Some of those ports are unofficial and legally shaky. The same technique can make porting games to Linux straightforward, so a game that blocks Proton or Wine for anti cheat reasons could get a native build that leaves the anti cheat baggage behind. Put the two together and the number of games that simply will not run on Linux keeps shrinking.&lt;/p&gt;
&lt;h2 id=&quot;what-this-does-to-linux-if-it-holds&quot;&gt;What this does to Linux, if it holds&lt;/h2&gt;
&lt;p&gt;Line all of it up and the conclusion is hard to avoid. In a few years there may be no hard technical reason left to stay on Windows or macOS if you do not want to. The apps run, through clones or native ports. The desktop can look like whatever you are used to. The hardware works, because the drivers got cheap to write. You could run a Linux system that looks the way you want, runs the software you want, and supports the hardware you own.&lt;/p&gt;
&lt;p&gt;The economics point one way. Who keeps paying $50 a month for 2 Adobe apps, locked onto a specific OS, when a free open source clone opens the same files, runs on any platform, and actually answers bug reports. Open source projects tend to respond to feature requests and fix bugs faster than a giant vendor answers a paying customer, because the people using the software are often the people writing it. Hardware makers could even start shipping their own Linux builds with a Windows style shell, since conforming to Microsoft’s specs and certifications stops being worth it once the alternative does everything Windows users want.&lt;/p&gt;
&lt;p&gt;I want to be careful here. This is the idealized version. It might take 1 year or 10, some of it might happen and some might not, and alpha quality clones are not replacements for daily work you get paid for yet. What changed is that the thing is now possible. A human doing it by hand made it impossible. An AI doing it makes it a question of effort and testing, not of whether it can be done at all.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.CHuOeBiE_Z28vDgt.webp&quot; alt=&quot;Where the clones stand in late 2026. Early and uneven, which is the honest state of it.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;720&quot; height=&quot;400&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Where the clones stand in late 2026. Early and uneven, which is the honest state of it. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.omgubuntu.co.uk/2026/10/artcraft-free-photoshop-ai-clone&quot;&gt;project statements and press coverage, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-legal-fight-is-the-actual-story&quot;&gt;The legal fight is the actual story&lt;/h2&gt;
&lt;p&gt;None of this matters if the courts kill it, so this is where the real battle sits. There are 4 separate legal questions tangled together, and the answers point in different directions.&lt;/p&gt;
&lt;p&gt;First, the license terms. Plenty of software agreements ban decompiling the program or using AI to analyze it. Nobody knows yet whether those clauses hold up, or whether anyone can even prove an AI studied a program as opposed to a human copying the visible features by hand, which has always been legal.&lt;/p&gt;
&lt;p&gt;Second, the clean room question, which is the term I said was doing heavy lifting. A clean room implementation is the classic legal way to copy a product without copying its code. One team studies the original and writes a specification of what it does, a second team that has never seen the original builds from that spec alone, and a wall between them keeps the copying honest. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Clean-room_design&quot;&gt;Phoenix Technologies used exactly this method&lt;/a&gt; in the 1980s to reverse engineer the IBM PC BIOS, then licensed its clean copy to any manufacturer who wanted one, which launched the entire PC clone industry and is a large part of why Intel and Microsoft got the market this article is about in the first place. The method works because the people writing the new code provably never touched the old code. ArtCraft calls itself clean room, but an AI model may have been trained on the original source or on code very close to it, so whether anything an AI writes can honestly be called clean room is an open question no court has answered. That single word may decide whether these projects are legal copying or infringement wearing a nicer label.&lt;/p&gt;
&lt;p&gt;Third, whether the output can even be owned. The US Copyright Office said in 2025 that purely AI generated material has no human author and cannot be copyrighted, while AI used as a tool with real human creative input can be. For a clone that leans heavily on AI, that creates a strange situation where the open source license laid on top of the code might not fully bind, because you cannot license what you may not own.&lt;/p&gt;
&lt;p&gt;Fourth, and most settled, the interface itself. US courts have repeatedly held that you can copy function without copying expression. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Google_LLC_v._Oracle_America,_Inc.&quot;&gt;Google won its case against Oracle&lt;/a&gt; at the Supreme Court in 2021, where reimplementing the Java APIs was ruled fair use. Going back further, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Lotus_Development_Corp._v._Borland_International,_Inc.&quot;&gt;Lotus v. Borland&lt;/a&gt; established that a menu command hierarchy is not copyrightable. In the EU the ground is even friendlier, since a 2012 ruling in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/SAS_Institute_Inc_v_World_Programming_Ltd&quot;&gt;SAS Institute v. World Programming&lt;/a&gt; held that the functionality of a program and its interface are not protected by copyright, and that license terms cannot stop you from studying how a program works to build something else. A plain menu and palette UI like Photoshop’s is not an original creative work, so copying the layout, without copying the icons or artwork, sits on solid ground in both places.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.Ch_2_xyO_ZomBJl.webp&quot; alt=&quot;4 legal questions hanging over the clones, from the most settled to the most open.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;859&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;4 legal questions hanging over the clones, from the most settled to the most open. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Google_LLC_v._Oracle_America,_Inc.&quot;&gt;US and EU case law, US Copyright Office 2025&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-companies-will-not-need-a-courtroom&quot;&gt;The companies will not need a courtroom&lt;/h2&gt;
&lt;p&gt;Even where the law is unsettled, the big vendors have a faster weapon. The DMCA takedown. Adobe or Microsoft can send GitHub a notice claiming a repo infringes, GitHub generally complies and pulls it, and the developer then has to file a counter notice and wait, possibly fighting a court case to get the code back. A project can vanish from the main place people find it without any ruling on whether it actually broke the law. Expect a lot of this.&lt;/p&gt;
&lt;p&gt;Jurisdiction will matter too. The EU rulings make these clones look close to safe for a team hosting the code in Europe, while US case law on decompilation clauses is more mixed. So the same project might be legal in one place and exposed in another, and where a lawsuit lands could decide its fate more than what the code actually does.&lt;/p&gt;
&lt;h2 id=&quot;the-part-that-could-hurt-open-source&quot;&gt;The part that could hurt open source&lt;/h2&gt;
&lt;p&gt;Not all of this is good news for the open source world, which is the part I think gets missed in the excitement.&lt;/p&gt;
&lt;p&gt;A clone that copies Photoshop exactly might bury the open source tools that tried to do something different. Why learn a new way to edit images when a free clone gives you the exact Photoshop you already know. The projects that built their own approach, their own interface, their own ideas, could get flattened by clones of the proprietary thing, because people reach for what is familiar. The skills angle cuts the same way. Learning a Photoshop clone teaches you Photoshop, which transfers to a paid job. Learning a genuinely different tool teaches you that tool.&lt;/p&gt;
&lt;p&gt;Proprietary file formats win too. If a free clone opens and saves PSD and the Office formats perfectly, people stop caring about open document standards, because they can already exchange files with everyone. The formats end up readable by anything, which is good, but they also stay the default, which quietly keeps the proprietary format in charge.&lt;/p&gt;
&lt;p&gt;Then there is the AI question hanging over the whole thing. Plenty of developers do not want to use AI to build their software, for reasons that are easy to respect, so the clone wave may split the open source world between projects that embrace it and projects that refuse. Both positions are defensible. The split is going to be real.&lt;/p&gt;
&lt;h2 id=&quot;where-i-land-on-it&quot;&gt;Where I land on it&lt;/h2&gt;
&lt;p&gt;I do not love AI. It has real costs, it was often built by scraping work without consent, and a lot of what it enables is genuinely harmful. Watching it get pointed back at the companies that pushed it hardest, the ones that locked users in and treated proprietary formats as a cage, is the rare case where I find the whole thing satisfying rather than grim.&lt;/p&gt;
&lt;p&gt;The lock that kept people on Windows and macOS was never really technical. It was the cost and time of rebuilding 30 years of proprietary software, and that cost just collapsed. Whether the clones survive the lawsuits, the takedowns, and their own alpha quality bugs is the open question. What is already settled is that rebuilding the software is no longer the impossible part. For Linux, that is the opening it never had before.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>ai</category><category>ai-agent</category><category>macos</category><category>windows</category><category>linux</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>How Intel, Microsoft, and Linux Took Down the UNIX Server Kingdoms</title><link>https://soursen.org/articles/how-intel-microsoft-and-linux-took-down-the-unix-server-kingdoms/</link><guid isPermaLink="true">https://soursen.org/articles/how-intel-microsoft-and-linux-took-down-the-unix-server-kingdoms/</guid><description>Going into the 1990s, a serious enterprise server meant a proprietary UNIX system running on custom RISC silicon, and it cost a fortune. By 2005 most of that market had been eaten by cheap Intel boxes running either Windows or Linux. I went back through how that happened, because it is one of the cleanest examples in computing history of a business model getting beaten on economics while its technology stayed ahead.</description><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;If you walked into a large company’s data center in 1992, you were looking at machines from &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Sun_Microsystems&quot;&gt;Sun Microsystems&lt;/a&gt;, IBM, Hewlett-Packard, or Silicon Graphics. Each one sold you a closed stack, their own RISC processor, their own flavor of UNIX, their own support contract. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/SPARC&quot;&gt;Sun&lt;/a&gt; ran SPARC chips and Solaris. IBM ran POWER chips and AIX. HP ran PA-RISC and HP-UX. Silicon Graphics ran MIPS chips and IRIX. These systems ran into the hundreds of thousands of dollars, and once you bought one you were locked in, because moving your software to a rival’s proprietary UNIX was close to impossible. Buy in, get stuck.&lt;/p&gt;
&lt;p&gt;These machines were genuinely good, which matters for the whole story. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Reduced_instruction_set_computer&quot;&gt;RISC&lt;/a&gt; chips, reduced instruction set designs that did fewer things but did them fast, held real performance leads for years. The UNIX variants running on them were mature, scaled to dozens of processors, and ran the databases and transaction systems that banks, airlines, and telecoms could not afford to have fall over. This was not a case of a bad incumbent getting what it deserved. The best engineering in the market was sitting inside these boxes.&lt;/p&gt;
&lt;p&gt;That whole arrangement looked permanent. Within about 13 years it was mostly gone. I want to walk through why, because the story has three main characters, Intel, Microsoft, and Linux, and each of them wanted something different out of the same transition. Intel wanted to sell more chips. Microsoft wanted to own the server operating system. Linux wanted nothing, which is part of why it won.&lt;/p&gt;
&lt;h2 id=&quot;intel-had-a-playbook-and-wanted-to-run-it-again&quot;&gt;Intel had a playbook and wanted to run it again&lt;/h2&gt;
&lt;p&gt;Intel had already won the desktop. The way it won matters, because it tried the exact same move on servers. High sales volume on PC chips improved its semiconductor yield. Better yield cut costs, which raised profit, which threw off cash it plowed back into better factories. That flywheel gave Intel a manufacturing lead that AMD and Cyrix could not match. By the mid 1990s Intel looked at the enterprise server market, saw machines selling for 100 times the price of a PC, and decided its volume economics could grind that market down too.&lt;/p&gt;
&lt;p&gt;The math behind that confidence is the engine of the whole story. A chip costs a fortune to design and a fortune to build a factory for. After that, each additional chip off the line costs almost nothing. Sun shipped SPARC chips in the tens or hundreds of thousands per year. Intel shipped x86 chips in the tens of millions. Spread those enormous fixed costs over 100 times the units and your per chip cost collapses, which funds the next factory, which widens the lead again. A proprietary RISC vendor could out design Intel on any given year and still lose, because it could never match Intel on cost per transistor. That gap only grew as chips got more expensive to make.&lt;/p&gt;
&lt;p&gt;The problem was that Intel made chips, not systems. To get into servers it needed an operating system that ran on its hardware and that corporate buyers trusted. In 1993 that pointed straight at Microsoft and a brand new product called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Windows_NT&quot;&gt;Windows NT&lt;/a&gt;.&lt;/p&gt;
&lt;h2 id=&quot;windows-nt-was-built-from-scratch-for-this&quot;&gt;Windows NT was built from scratch for this&lt;/h2&gt;
&lt;p&gt;Regular consumer Windows in that era was a graphical shell sitting on top of MS-DOS, held together with tape. NT was different underneath. A team led by &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Dave_Cutler&quot;&gt;Dave Cutler&lt;/a&gt;, who had built &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/OpenVMS&quot;&gt;VMS&lt;/a&gt; at Digital, wrote NT from the ground up for workstations and servers. Microsoft sold it on reliability, security, and real multitasking, the things enterprise buyers cared about.&lt;/p&gt;
&lt;p&gt;The first version, NT 3.1, shipped in 1993 and ran on the Intel 386. It worked, but it did not light the world on fire. It was immature next to IBM’s OS/2, Novell’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/NetWare&quot;&gt;NetWare&lt;/a&gt;, and the established UNIX variants, and it demanded 16 megabytes of RAM, which was a lot in 1993. NT Server sold modestly that first year. Microsoft kept grinding, the way it tends to, shipping new features and new versions and slowly wearing the competition down. Sales roughly tripled the next year, still trailing the UNIX incumbents but climbing fast.&lt;/p&gt;
&lt;h2 id=&quot;the-pentium-pro-gave-intel-its-opening&quot;&gt;The Pentium Pro gave Intel its opening&lt;/h2&gt;
&lt;p&gt;In late 1995 Intel shipped the chip it actually believed could crack servers, the 32-bit &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Pentium_Pro&quot;&gt;Pentium Pro&lt;/a&gt;. It carried 5.5 million transistors and ran around 150 to 200 MHz, a big leap over the original Pentium’s 3.1 million transistors and 60 MHz. What shocked people was that the Pentium Pro beat the fastest RISC chips on some integer benchmarks. The faster RISC parts like the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/R10000&quot;&gt;MIPS R10000&lt;/a&gt; caught back up quickly, but the point had been made. A so called desktop architecture was now in the same conversation as the expensive stuff.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.MTtmLSSl_2w7N04.webp&quot; alt=&quot;Transistor count and clock speed, original Pentium versus Pentium Pro.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Transistor count and clock speed, original Pentium versus Pentium Pro. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Pentium_Pro&quot;&gt;Pentium Pro, Wikipedia&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Pair the Pentium Pro with Windows NT 4.0, which also landed in 1996, and vendors suddenly had a ready made server standard. It matched the performance of some RISC integrated systems at a fraction of the price. Intel started circulating reference boards it called Standard High Volume Servers, basic designs partners could copy to get products out faster. The first generation boards took 1 to 4 Pentium Pros, 1 to 2 gigabytes of memory, and hundreds of gigabytes of storage. At the launch event Intel showed Pentium Pro machines ranging from a $2,600 workstation up to a $20,000 server, which was already an order of magnitude cheaper than the proprietary UNIX boxes.&lt;/p&gt;
&lt;p&gt;By 1996 NT 4.0 was growing fast and taking real server market share, according to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/International_Data_Corporation&quot;&gt;IDC&lt;/a&gt; tracking at the time. Wintel, the Windows plus Intel pairing, was building the same kind of base in servers that it already owned in consumer PCs.&lt;/p&gt;
&lt;h2 id=&quot;intel-and-microsoft-did-not-actually-trust-each-other&quot;&gt;Intel and Microsoft did not actually trust each other&lt;/h2&gt;
&lt;p&gt;Here is the part that gets left out of the tidy version. Intel knew that chaining its enterprise future to Microsoft was risky, because the two companies fought constantly under the surface. In 1995 Intel pushed a multimedia software standard it called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Native_Signal_Processing&quot;&gt;Native Signal Processing&lt;/a&gt;, NSP, which offloaded audio and video work onto the CPU through its own APIs. Intel’s logic was simple, make people want faster chips. Microsoft saw NSP as middleware that stepped around Windows and onto its turf, and killed it. Bill Gates emailed Andy Grove to make the displeasure clear, and Microsoft publicly refused to support it. The whole thing resurfaced later during Microsoft’s antitrust trial as an example of how it leaned on partners.&lt;/p&gt;
&lt;p&gt;The deeper tension was structural. Intel’s business depends on selling you a new, faster chip every couple of years, so it wants software that exploits every new capability immediately. Microsoft’s business depends on servicing its entire install base, so it moves slowly and guards backward compatibility. Intel shipped a 32-bit chip, the 386, in 1985. Microsoft did not ship a fully 32-bit consumer OS until Windows 95, 10 years later. NT also capped out at 4 to 8 processors per instance in that era, while high end UNIX could scale to 64, which kept NT out of the very top of the market. Intel wanted a partner that would not hold its silicon back. Around 1998, one appeared on the radar.&lt;/p&gt;
&lt;h2 id=&quot;where-linux-actually-came-from&quot;&gt;Where Linux actually came from&lt;/h2&gt;
&lt;p&gt;To understand why Linux mattered to Intel, you have to back up to the free software movement. In 1983 &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Richard_Stallman&quot;&gt;Richard Stallman&lt;/a&gt;, a programmer at the MIT AI Lab, started a project to build a free operating system that followed UNIX’s design without copying its code. He called it &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/GNU&quot;&gt;GNU&lt;/a&gt;, a recursive joke standing for GNU’s Not UNIX. Over several years he and a growing community built the toolchain, the compilers and utilities and shells. One piece kept slipping, the kernel, the part that sits between the OS and the hardware and manages memory, processor time, and storage.&lt;/p&gt;
&lt;p&gt;That kernel arrived from an unexpected direction. In August 1991 a Finnish student of 21 named &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Linus_Torvalds&quot;&gt;Linus Torvalds&lt;/a&gt; posted to a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Minix&quot;&gt;MINIX&lt;/a&gt; Usenet group that he was writing a free hobby OS and wanted feature ideas. It drew contributors, grew, and became &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Linux&quot;&gt;Linux&lt;/a&gt;. Torvalds shifted from coder to maintainer, reviewing other people’s contributions and merging them, a model famously described as a bazaar in Eric Raymond’s essay &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/The_Cathedral_and_the_Bazaar&quot;&gt;The Cathedral and the Bazaar&lt;/a&gt;. It let Linux ship new code faster than any single company could.&lt;/p&gt;
&lt;p&gt;In 1992 Linux adopted Stallman’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/GNU_General_Public_License&quot;&gt;GPL&lt;/a&gt; license. This turned out to be the thing that saved it. The GPL let anyone read and modify the code as long as they open sourced their changes when distributing. That stopped any company from forking Linux, adding proprietary extras, and selling a closed competitor, which is exactly the fragmentation that had splintered UNIX into 30 incompatible variants. It also reassured companies that work they did on the kernel would benefit everyone and not get locked away. IBM would eventually have hundreds of employees contributing code. Torvalds later called going GPL the best decision he ever made.&lt;/p&gt;
&lt;p&gt;The fragmentation point deserves a closer look, because it is why UNIX was vulnerable in a way that looked invisible from inside. UNIX started at Bell Labs in the early 1970s, then spread through universities and companies that each added their own extensions. By the 1990s there were around 30 incompatible flavors, each with its own quirks, its own APIs, its own subtle breakage. Software written for Solaris failed to run on AIX without changes. A vendor selling a database had to port and test against every one of them, which was expensive, so it supported a few and ignored the rest. That splintering was the direct result of everyone being allowed to take the code private and diverge. The GPL made divergence pointless. If you improved Linux and shipped it, your improvement had to come back to everyone, so there was one Linux instead of 30 Linuxes. A database vendor ported once. That single trait, one target instead of many, is a large part of why application makers eventually found Linux cheaper to support than the UNIX zoo.&lt;/p&gt;
&lt;h2 id=&quot;distributions-turned-a-kernel-into-a-product&quot;&gt;Distributions turned a kernel into a product&lt;/h2&gt;
&lt;p&gt;A kernel on its own is a car engine with no car. Fine for enthusiasts, useless for a company that just wants a working server. So people built distributions, bundles that packaged the Linux kernel with the GNU utilities and other free software into something you could actually install. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Slackware&quot;&gt;Slackware&lt;/a&gt; and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Debian&quot;&gt;Debian&lt;/a&gt; came early. So did &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Red_Hat&quot;&gt;Red Hat&lt;/a&gt;, started by Marc Ewing in 1993 and named after a red and white striped lacrosse cap he wore in college.&lt;/p&gt;
&lt;p&gt;Red Hat wrapped Linux in the things a professional product needs, like the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/RPM_Package_Manager&quot;&gt;RPM&lt;/a&gt; package manager for adding, removing, and updating software cleanly. In 1995 Ewing and Bob Young founded Red Hat the company, earning money from selling boxed CDs plus training, support, and services. Young had a line he loved about the business model. Anyone with tomatoes, sugar, and vinegar can make ketchup, yet Heinz owns most of the market. You were not paying for the code, which was free. You were paying for the trust, the packaging, and the phone number to call when something broke.&lt;/p&gt;
&lt;h2 id=&quot;linux-grew-because-it-was-reliable-and-cheap-helped&quot;&gt;Linux grew because it was reliable, and cheap helped&lt;/h2&gt;
&lt;p&gt;By mid 1997 Linux had an estimated 3 to 5 million users worldwide. Tiny next to Windows at around 100 million and the Mac at 50 million plus, but it had come from maybe a few hundred users in early 1992. A Red Hat box retailed around $50. An NT license ran from roughly $150 for a single upgrade to over $1,000 for a 10-seat package. A full UNIX system with hardware ran anywhere from $25,000 to a quarter of a million dollars and up.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.ofyQkfTh_1biWF6.webp&quot; alt=&quot;Roughly what each option cost to put a server in service in the late 1990s. Log scale, because the spread is enormous.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Roughly what each option cost to put a server in service in the late 1990s. Log scale, because the spread is enormous. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Linux&quot;&gt;price points cited above, contemporary trade press&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Price was not the whole story though. The thing people actually loved about Linux was reliability. One early user called it almost crash proof, which in server terms means uptime, which means money. Sophisticated users loved that they could read and change the source, something Microsoft would never allow with Windows. People ported Linux to a Palm Pilot. Researchers at Los Alamos wired 68 PCs together with it to simulate atomic shock waves, an early taste of the cluster computing that would later define supercomputing.&lt;/p&gt;
&lt;p&gt;That clustering thread turned into something big. In 1994 two researchers at NASA Goddard, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Thomas_Sterling_%28computing%29&quot;&gt;Thomas Sterling&lt;/a&gt; and Donald Becker, wired 16 ordinary PCs together with cheap Ethernet and ran Linux across them, and called it a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Beowulf_cluster&quot;&gt;Beowulf cluster&lt;/a&gt;. The idea was simple and a little heretical. Instead of buying one enormous supercomputer, buy a pile of commodity machines and make them act like one. Linux was the natural choice because it was free, you could run it on as many nodes as you wanted without licensing math, and you could read and modify the networking code to make the cluster go faster. This is the same scale out logic that would later eat the server room, showing up first in science labs where budgets were tight and the people running the machines could write their own code.&lt;/p&gt;
&lt;p&gt;It had real gaps. There was no good Linux version of Microsoft Office or the business software companies ran on. Support was a question mark unless you paid Red Hat or &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Caldera_%28company%29&quot;&gt;Caldera&lt;/a&gt;. And it carried UNIX’s reputation for being hard to use. A manager might ask how you could possibly run a critical server on free software, and the honest answer was that Linux was probably already running somewhere in their building without anyone in management knowing.&lt;/p&gt;
&lt;h2 id=&quot;intel-noticed-the-thing-it-did-not-plan-for&quot;&gt;Intel noticed the thing it did not plan for&lt;/h2&gt;
&lt;p&gt;In early 1998 Intel started getting odd requests from national labs like Fermilab, asking it to port its math libraries to Linux. Before that Intel had barely touched Linux, maybe fixing a bug here and there. When it checked with the OEMs building white box servers for small internet service providers, the picture snapped into focus. Linux adoption was exploding. Close to 20% of public web servers were running it, a surprising share for an OS with basically no desktop presence.&lt;/p&gt;
&lt;p&gt;That number went straight to Andy Grove, who made it clear Intel needed to do something. Intel executives sat down with Red Hat and came away baffled, because Red Hat did not own the core product it sold and anyone could get it free. Then the pieces clicked. Intel increasingly felt that betting everything on Windows NT was dangerous, especially with its new 64-bit server chip, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Itanium&quot;&gt;Itanium&lt;/a&gt;, then codenamed Merced, in development. Intel worried NT would shackle Itanium the way Windows had lagged its 32-bit chips for a decade. Microsoft was not helping its own case, having slipped NT 5.0 from 1998 to 1999 before renaming it &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Windows_2000&quot;&gt;Windows 2000&lt;/a&gt;. Backing Linux gave Intel a fast moving OS it could tune to its silicon and hold over Microsoft’s head.&lt;/p&gt;
&lt;h2 id=&quot;the-summer-of-1998-flipped-the-enterprise&quot;&gt;The summer of 1998 flipped the enterprise&lt;/h2&gt;
&lt;p&gt;Several big companies went public on Linux and open source in a few months. In June 1998 IBM, one of the most conservative firms in tech, folded the open source &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Apache_HTTP_Server&quot;&gt;Apache&lt;/a&gt; web server into its e-commerce suite. Apache is not Linux, but most Apache instances ran on it, and IBM’s blessing gave open source its first real foothold in the enterprise.&lt;/p&gt;
&lt;p&gt;Oracle made the sharpest turn. On 6 July 1998 InfoWorld ran a piece where Oracle, Sybase, and Informix all said they had no plans to port their databases to Linux, citing weak customer demand. Three days later Oracle announced it was porting &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Oracle_Database&quot;&gt;Oracle8&lt;/a&gt; to Linux. A big reason for the reversal was a panel in Santa Clara that July called the Future of Linux, where over 900 people showed up to see Torvalds and others argue about whether open source was a fad and push vendors to port their software. People later marked that panel as a turning point. By September, Computer Associates, Corel, and Informix all had Linux ports in the works.&lt;/p&gt;
&lt;h2 id=&quot;intel-put-money-on-the-table-and-microsoft-went-ballistic&quot;&gt;Intel put money on the table and Microsoft went ballistic&lt;/h2&gt;
&lt;p&gt;Intel was at that July panel. Its engineer Sunil Saxena gave a talk calling Linux the OS of choice for internet service providers while naming its weak spots, scalability, high end drivers, and large memory support. He said Intel wanted to help. On 29 September 1998 Intel backed that with cash, joining Netscape and the VC firms Benchmark and Greylock in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Red_Hat&quot;&gt;investing in Red Hat&lt;/a&gt;. Intel pledged engineering help and a dedicated liaison to the Linux community, and said it would join Linux International.&lt;/p&gt;
&lt;p&gt;One underrated domino fell right after. A Benchmark partner sat on Dell’s board, and through him Red Hat got in front of Michael Dell. Dell invested too, and in early 1999 announced it would preinstall Linux on workstations and servers, the first major vendor to do it. Suddenly Linux stood on the same shelf as Windows and NetWare.&lt;/p&gt;
&lt;p&gt;When news of Intel’s Red Hat investment broke, the tech press lost its mind over what it meant for the Intel Microsoft relationship. Red Hat executives later recounted that Microsoft went ballistic, and that Intel took angry weekend calls from its counterparts in Redmond. Before 1998 Microsoft might have been able to pressure Intel out of the deal. It no longer could, because in May 1998 the US government had charged Microsoft with &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/United_States_v._Microsoft_Corp.&quot;&gt;monopolistic conduct&lt;/a&gt; over bundling Internet Explorer with Windows. Under that scrutiny Microsoft had to behave, and it even pointed at Red Hat as proof it did not have an OS monopoly.&lt;/p&gt;
&lt;p&gt;Internally the mood was different. In late October 1998 a set of leaked Microsoft memos, nicknamed the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Halloween_documents&quot;&gt;Halloween documents&lt;/a&gt; and published by Eric Raymond, called free software like Linux a credible long term threat to Windows. Microsoft reportedly set up a team to convince customers Linux was not ready, put up a Linux Myths web page picking apart claims about Linux performance and reliability, and in April 1999 sponsored &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Mindcraft&quot;&gt;Mindcraft&lt;/a&gt; benchmarks that showed NT beating Linux badly as a file and web server. Those benchmarks were real and stung, because the early Linux kernel genuinely did not scale well past 2 processors.&lt;/p&gt;
&lt;h2 id=&quot;intel-played-switzerland&quot;&gt;Intel played Switzerland&lt;/h2&gt;
&lt;p&gt;Intel dodged the fight by declaring itself OS neutral. Linux, NetWare, Solaris, Windows 2000, Intel wanted all of them running well on its chips. It even tried to herd the 30 or so UNIX variants into writing shared drivers for Intel hardware through a Uniform Driver Initiative, which went nowhere. Intel later formalized this as a Port of Choice strategy, though &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Pat_Gelsinger&quot;&gt;Pat Gelsinger&lt;/a&gt; called it a long standing approach as early as 2001. The OS makers had to accept it, because Intel silicon had no real alternative.&lt;/p&gt;
&lt;p&gt;The hardware kept advancing. In mid 1998 Intel shipped the Pentium II &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Xeon&quot;&gt;Xeon&lt;/a&gt; with a large 512 KB L2 cache suited to servers. In March 1999 the Pentium III Xeon arrived, and later that year, around August, Intel shipped the Profusion chipset that let vendors glue 8 processors into one box, breaking the old 4-way ceiling that had been choking data centers. That multiprocessor support needed the OS to cooperate, which NT and Windows 2000 could do and early Linux could not yet. High end UNIX still held the very top, where big databases and CAD systems needed more memory than 32-bit x86 could address, but the low and middle ground was now wide open.&lt;/p&gt;
&lt;h2 id=&quot;the-web-scaled-out-not-up&quot;&gt;The web scaled out, not up&lt;/h2&gt;
&lt;p&gt;Sun welcomed Linux at first, in line with its open branding, and ported its Java kit to Linux. Its attitude was that it did not care whether a box ran Linux or Solaris as long as it was a SPARC box. That assumption was the mistake. It did not have to be a SPARC box. Serving websites, email, and firewalls is not the same workload as running a giant database, and companies worked out that scaling out with lots of cheap Intel Linux nodes made far more financial sense than scaling up one enormous UNIX machine.&lt;/p&gt;
&lt;p&gt;The dotcom crash accelerated everything. Venture money dried up, frugality became the rule, and companies bought second hand gear from dead startups or switched to Linux. Amazon did a major Linux migration in 2001 and told the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.cnet.com/tech/tech-industry/amazon-saves-millions-with-linux-switch/&quot;&gt;SEC it cut technology costs by around $17 million in one quarter&lt;/a&gt;. An Amazon spokesman, asked how, said the company stayed closed mouthed about back end technology because it was steep tuition and it did not want others going to school on its dime.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.KADdHy51_10n53n.webp&quot; alt=&quot;Gartner figures for 2002, the year the direction became obvious. Linux server hardware revenue climbing while UNIX fell, even though UNIX was still far larger in absolute dollars.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Gartner figures for 2002, the year the direction became obvious. Linux server hardware revenue climbing while UNIX fell, even though UNIX was still far larger in absolute dollars. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.theregister.com/2004/02/27/server_vendors_cheer_strong_q4/&quot;&gt;Gartner 2002 server figures, contemporary trade press&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;An E-Trade case study from that era captured the tradeoff cleanly. A Sun 4500 Enterprise server cost $250,000 and served 300 to 400 users. A Linux Intel box cost $4,000 and served 180. The fix for the capacity gap was obvious, buy more boxes, and you still came out tens of thousands of dollars ahead. Starting in 2001, Linux Intel machines got good enough to steadily replace Sun in everything from airline fares to commodities trading. Sun tried its own low cost scale out play with the Netra, priced as low as $1,000, but commodity x86 still gave more performance per dollar, and Sun never formed a real answer to the shift.&lt;/p&gt;
&lt;h2 id=&quot;linux-climbed-up-the-ladder&quot;&gt;Linux climbed up the ladder&lt;/h2&gt;
&lt;p&gt;After taking the low end, Linux went after the big machines. In late 2000 IBM announced a roughly $1 billion Linux effort with 1,500 employees dedicated to it, a stunning commitment from the company that had written AIX. In January 2001 &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Linux_kernel_version_history&quot;&gt;Linux 2.4&lt;/a&gt; arrived and closed real performance gaps with high end UNIX and NT. The multiprocessor weakness the Mindcraft tests had exposed got worked on hard, and 2.4 let Linux use 16 or even 32 processors with some effectiveness. It did not beat high end UNIX overnight, but solid 4 and 8 CPU support cleared the path into the middle of the market.&lt;/p&gt;
&lt;p&gt;The economics got brutal for the incumbents. A Dell server with 4 Xeons running Red Hat at $24,000 cost half what a comparable Sun Fire UNIX box did and performed just as well. Red Hat and IBM went straight at Sun’s Solaris customers. HP and SGI started offering their own proprietary hardware customers a path to Linux, reasoning it was better to keep the relationship than lose it entirely. In 2002, a hard year for servers overall, Gartner reported Linux server hardware revenue grew sharply while UNIX revenue fell. UNIX was still the bigger number in absolute dollars because the machines cost more, $17 billion against Linux’s $2 billion that year, but the trend line had turned. By 2004 the average selling price of a Sun server had dropped roughly a third in two years.&lt;/p&gt;
&lt;h2 id=&quot;x86-went-64-bit-and-the-last-wall-came-down&quot;&gt;x86 went 64-bit and the last wall came down&lt;/h2&gt;
&lt;p&gt;One of the last real advantages UNIX had was memory. A 32-bit x86 machine can address only about 4 gigabytes of RAM, far below what big databases and scientific workloads need to hold in memory. Intel’s bet here was Itanium, a clean break 64-bit architecture. It spun up &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.redhat.com/en/about/press-releases/35&quot;&gt;Project Trillian&lt;/a&gt; in 1999 to get Linux ported to Itanium early, and released the IA-64 Linux source in February 2000, a genuine milestone for open source development.&lt;/p&gt;
&lt;p&gt;Itanium did not win. The company that kicked off the real 64-bit transition was AMD, not Intel, with the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Opteron&quot;&gt;Opteron&lt;/a&gt; in April 2003. AMD’s approach, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/X86-64&quot;&gt;x86–64&lt;/a&gt;, extended the existing x86 instruction set to 64-bit while staying backward compatible with all the 32-bit software people already ran. In 2004 Intel swallowed its pride and adopted AMD’s extensions in its Nocona Xeon. Once x86 could address dozens or hundreds of gigabytes of RAM, the last major architectural reason to buy UNIX was gone.&lt;/p&gt;
&lt;h2 id=&quot;why-microsoft-kept-growing-anyway&quot;&gt;Why Microsoft kept growing anyway&lt;/h2&gt;
&lt;p&gt;Through all this, plenty of people predicted Linux would take down Microsoft on the server. It did not. Windows Server sales kept climbing. In 2005 Windows server revenue matched UNIX for the first time, and passed it the next year. Linux kept growing too, but its growth came mostly from eating UNIX and from brand new markets where Windows had never had a foothold, supercomputing, high performance compute, the web, and later cloud.&lt;/p&gt;
&lt;p&gt;Microsoft’s defense was not a better benchmark. It was lock in through &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Active_Directory&quot;&gt;Active Directory&lt;/a&gt;, released with Windows 2000, which Ben Thompson of &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://stratechery.com/&quot;&gt;Stratechery&lt;/a&gt; has called one of the most valuable free products of all time. Active Directory authenticates and authorizes every user and asset in a Windows network. An admin can onboard a new manager, verify credentials, and grant exactly the right access, all from one place, and users log in to everything with one identity. That turned a pile of random Windows machines into a single managed organization, and it gave Microsoft ownership of identity inside the company. Identity reaches so deep into how an organization runs that ripping it out costs far more than swapping an OS. Microsoft’s real moat was never the kernel.&lt;/p&gt;
&lt;h2 id=&quot;the-pattern-underneath-and-why-it-still-shows-up&quot;&gt;The pattern underneath, and why it still shows up&lt;/h2&gt;
&lt;p&gt;Strip the names off this story and you get a shape that repeats across the industry. A high margin incumbent sells an integrated product where the value is bundled together and the price reflects the whole bundle. A challenger commoditizes one layer of that bundle, makes it cheap or free, and the margin drains out of the layer above. Joel Spolsky wrote about this in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/&quot;&gt;Strategy Letter V&lt;/a&gt; under the phrase commoditize your complement. Make the thing next to yours cheap, and demand for your thing goes up.&lt;/p&gt;
&lt;p&gt;Intel wanted the operating system to be a cheap complement to its chips, which is why it funded both Windows alternatives and Linux. A free, good OS running on Intel silicon meant customers spent their money on Intel hardware instead of on a Sun or HP bundle where the margin went to someone else. Linux wanted nothing in return, so it was the perfect complement, there was no vendor on the other side of the table trying to capture value back. IBM spent roughly $1 billion making Linux better not out of charity but because a strong free OS undercut Sun and Microsoft while IBM sold the hardware, services, and consulting around it. Everyone pushing Linux forward was commoditizing someone else’s product to protect or grow their own.&lt;/p&gt;
&lt;p&gt;The same move keeps happening. Cloud providers gave away or cheapened the managed database and the operating system so you would rent their compute. Android was handed out free so the hardware and the search box around it would sell. When you see a large company investing heavily in making something free and good, the question to ask is which adjacent thing they actually sell, because that is almost always where the money is going.&lt;/p&gt;
&lt;h2 id=&quot;why-unix-actually-lost&quot;&gt;Why UNIX actually lost&lt;/h2&gt;
&lt;p&gt;UNIX never fully died. There are still banking and healthcare systems running Solaris, AIX, or HP-UX today, applications too risky or too expensive to move. But the kingdoms are gone.&lt;/p&gt;
&lt;p&gt;I do not think UNIX lost on technology. In raw capability the high end UNIX machines were ahead for years. It lost on business model. The vertically integrated vendor, one company controlling chip, OS, and hardware and charging a fortune for the bundle, worked fine until Intel showed up with volume economics that no single vendor could match on cost. Windows NT hit UNIX first and hard. Linux hit it harder, because it offered everything UNIX had, portability, reliability, the UNIX design, without the lock in or the price. Once Linux got good enough, the UNIX vendors themselves had to make the switch, selling customers the migration off their own platforms. The empire was not conquered so much as undercut, one cheap box at a time.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>intel</category><category>microsoft</category><category>linux</category><category>unix</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>The 23-Year Lawsuit That Claimed to Own Part of Linux Just Died</title><link>https://soursen.org/articles/the-23-year-lawsuit-that-claimed-to-own-part-of-linux-just-died/</link><guid isPermaLink="true">https://soursen.org/articles/the-23-year-lawsuit-that-claimed-to-own-part-of-linux-just-died/</guid><description>On 10 August 2026 a US appeals court threw out the last live appeal in a legal fight that started with a forgotten 1998 UNIX project, ran through a $5 billion lawsuit, a bankruptcy, an asset sale, and a fresh filing in 2021. For 23 years a private company kept claiming it owned code sitting inside the Linux kernel. I went through the whole saga, because it is the closest the open source world ever came to a genuine extinction event, and because the thing that finally killed it was not a ruling on who was right.</description><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;If you run almost any server today, you run Linux. Something like 9 in 10 of the machines serving the internet do. That dominance looks inevitable now. For most of the last two decades it was not, because a legal claim hung over the kernel’s core code, the kind of claim that, if a court had ever agreed with it, could have forced every Linux distribution on earth to either pay a private company or rip out code that had been shipping for years.&lt;/p&gt;
&lt;p&gt;That claim is now, for all practical purposes, dead. On 10 August 2026 the US Court of Appeals for the Second Circuit rejected the final appeal, and the company behind it is left with a rehearing petition its opponent puts at a 0.03% chance of success. I want to walk through how a handshake deal from 1998 turned into a 23-year shadow over the most widely used operating system in the world.&lt;/p&gt;
&lt;h2 id=&quot;it-started-with-a-unix-project-almost-nobody-remembers&quot;&gt;It started with a UNIX project almost nobody remembers&lt;/h2&gt;
&lt;p&gt;In October 1998 four companies signed on to something called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Project_Monterey&quot;&gt;Project Monterey&lt;/a&gt;. IBM, the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Santa_Cruz_Operation&quot;&gt;Santa Cruz Operation&lt;/a&gt;, Intel, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Sequent_Computer_Systems&quot;&gt;Sequent&lt;/a&gt; agreed to build one unified UNIX that would run across many processor architectures, including Intel’s upcoming 64-bit Itanium and IBM’s own POWER chips. IBM brought POWER support from its AIX system. Santa Cruz Operation brought the 32-bit Intel side. Sequent brought multiprocessing from its DYNIX system. Intel brought money and Itanium expertise.&lt;/p&gt;
&lt;p&gt;The logic was sound for 1998. UNIX had splintered into roughly 30 incompatible variants, and a single unified standard looked like the fix. The whole thing was aimed at Itanium, which everyone expected to take over the server market.&lt;/p&gt;
&lt;p&gt;Two things then went wrong at once. Itanium slipped badly, missing its first delivery date by about two years, which left the Monterey software without the hardware it was built for. Meanwhile Linux, which nobody in that room had treated as a serious threat, started growing faster than anyone predicted. Between 1998 and 2001 the market that Project Monterey was designed to serve got absorbed by free open source alternatives before the project ever shipped at scale. IBM sold just 32 licenses of the Monterey product in 2001 and fewer the next year.&lt;/p&gt;
&lt;h2 id=&quot;ibm-read-the-room-and-switched-sides&quot;&gt;IBM read the room and switched sides&lt;/h2&gt;
&lt;p&gt;IBM did the rational thing. It saw Linux winning and pivoted hard, shifting its strategic weight from proprietary UNIX to Linux and pouring roughly $1 billion into Linux development. In 2001 it wound down its involvement in Project Monterey.&lt;/p&gt;
&lt;p&gt;Here is the detail that mattered later. Before IBM left, it had contributed code developed during the Monterey work into Linux itself. IBM’s partners in the joint venture would later argue that IBM only ever held a non-exclusive license to that code, meaning IBM could use it but did not own it outright, and so had no right to hand it to an open source project where anyone could copy it. IBM saw it differently. That disagreement sat quiet for about two years.&lt;/p&gt;
&lt;h2 id=&quot;2003-the-lawsuit-that-scared-an-entire-industry&quot;&gt;2003, the lawsuit that scared an entire industry&lt;/h2&gt;
&lt;p&gt;By 2003 the Santa Cruz Operation had sold its UNIX business and the buyer had renamed itself the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/SCO_Group&quot;&gt;SCO Group&lt;/a&gt;. In March 2003 SCO &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/SCO_Group,_Inc._v._International_Business_Machines_Corp.&quot;&gt;sued IBM&lt;/a&gt; in Utah, claiming IBM had improperly folded confidential UNIX source code into Linux in breach of their agreement. SCO started the damages demand at $1 billion and later raised it, eventually seeking &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.pinsentmasons.com/out-law/news/sco-claims-copyright-infringement-and-up-to-5-billion-from-ibm&quot;&gt;up to $5 billion&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The suit landed like a bomb. SCO was asserting ownership over code that had been shipping inside Linux for years, running on servers across the entire internet. In August 2003 SCO said it would start sending invoices to companies running Linux, demanding per processor license fees for software those companies thought was free. SCO backed off the invoicing that October after being warned it could face mail fraud exposure, but the message had already gone out. If SCO’s theory held up in court, Linux distributions worldwide were shipping software that legally belonged to one private company. That is as close to an extinction event as the open source world has faced.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.C_8T6uCa_mquU3.webp&quot; alt=&quot;The 23-year timeline, from the 1998 UNIX project to the 2026 ruling that effectively ended it.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;1493&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The 23-year timeline, from the 1998 UNIX project to the 2026 ruling that effectively ended it. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Timeline_of_SCO%E2%80%93Linux_disputes&quot;&gt;SCO-Linux disputes timeline, Wikipedia&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-claims-fell-apart-slowly&quot;&gt;The claims fell apart, slowly&lt;/h2&gt;
&lt;p&gt;The courts were not persuaded. Over years of litigation SCO’s arguments came apart piece by piece. A key 2007 ruling found that &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Novell&quot;&gt;Novell&lt;/a&gt;, not SCO, actually owned the UNIX copyrights SCO had been waving around, which cut the legs out from under the whole case. IBM filed heavy counterclaims. In September 2007 SCO filed for Chapter 11 bankruptcy, and the Linux related claims were largely lost or rendered moot.&lt;/p&gt;
&lt;p&gt;Most of the industry assumed the story was over. It was not, because intellectual property does not die when a company does. It becomes an asset that gets sold.&lt;/p&gt;
&lt;h2 id=&quot;the-claim-gets-a-second-owner&quot;&gt;The claim gets a second owner&lt;/h2&gt;
&lt;p&gt;In 2011 a company called UnXis bought SCO’s UNIX business and copyrights out of the bankruptcy estate. UnXis later renamed itself &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Xinuos&quot;&gt;Xinuos&lt;/a&gt;. For a while Xinuos ran the actual UNIX products it had bought, OpenServer and UnixWare, as a going business.&lt;/p&gt;
&lt;p&gt;Then on 31 March 2021 Xinuos filed a fresh lawsuit in the US District Court of the Virgin Islands, this time naming both IBM and Red Hat. The argument was the old Monterey theory in new clothes. Xinuos claimed IBM had only ever held a non-exclusive license to the Project Monterey code and had breached that license by contributing it to Linux, and that it owned specific copyrighted code blocks that originated in Monterey and ended up in the kernel. Red Hat got pulled in as the distributor of a commercial Linux product built on that allegedly tainted code.&lt;/p&gt;
&lt;p&gt;I want to be precise about why this was dangerous, because the stakes were not abstract. A copyright claim against code in the Linux kernel does not stop at IBM. The kernel is the one component every Linux distribution ships, from Red Hat Enterprise Linux and Ubuntu down to the hundreds of small volunteer projects tracked on &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://distrowatch.com/&quot;&gt;DistroWatch&lt;/a&gt;. If a court ruled that portions of that kernel belonged to a private company, the liability would ripple out to everyone who distributes, uses, or modifies the code, which is effectively everyone.&lt;/p&gt;
&lt;h2 id=&quot;what-ibm-actually-argued&quot;&gt;What IBM actually argued&lt;/h2&gt;
&lt;p&gt;IBM’s defense had a striking feature. IBM did not dispute that it had contributed code from the Monterey period into Linux. What it disputed was that Xinuos owned anything that let it block that contribution or charge for it. IBM pointed to a prior settlement that SCO had reached, which IBM said broadly covered all the Linux related claims before Xinuos ever bought the assets. Xinuos said that settlement did not apply to its claims.&lt;/p&gt;
&lt;p&gt;The District Court for the Southern District of New York, where the case ended up being heard, dismissed it. Xinuos appealed to the Second Circuit.&lt;/p&gt;
&lt;h2 id=&quot;the-ruling-and-the-technicality-that-decided-it&quot;&gt;The ruling, and the technicality that decided it&lt;/h2&gt;
&lt;p&gt;On 10 August 2026 a three judge panel of the Second Circuit rejected Xinuos’s appeal and upheld the dismissal. The reasoning is the part worth sitting with. The court did not rule that IBM was in the right about ownership. It ruled that Xinuos had waited too long. Although Xinuos framed its claim as a licensing dispute, the court found the real question underneath was code ownership, and that ownership question was time barred under the statute of limitations. The claim was too old to litigate.&lt;/p&gt;
&lt;p&gt;As &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.tomshardware.com/software/linux/decades-long-linux-ownership-dispute-effectively-dead-after-xinuos-appeal-rejected-us-court-of-appeals-halts-the-legal-wrangling-over-ibms-and-red-hats-use-of-project-monterey-unix-code&quot;&gt;Tom’s Hardware&lt;/a&gt; put it in its coverage, IBM does not even dispute taking the code from Project Monterey. It disputes that Xinuos owns anything that makes that taking legally actionable. The court agreed, and the appeal was gone.&lt;/p&gt;
&lt;p&gt;For the open source world the result is enormous. The exact scenario SCO started threatening in 2003, a private company using copyright to assert control over code in the Linux kernel, has now been rejected at every level of federal court where it has been tested. The Second Circuit is one of the most respected appeals courts in the country on copyright, so its word carries real weight.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.SUEzGQ_G_ZMc1KP.webp&quot; alt=&quot;Where Linux stands in 2026, the dominance the lawsuit never slowed. Figures are approximate.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Where Linux stands in 2026, the dominance the lawsuit never slowed. Figures are approximate. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.top500.org/&quot;&gt;TOP500 and industry data&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;why-it-is-not-100-over-and-why-that-barely-matters&quot;&gt;Why it is not 100% over, and why that barely matters&lt;/h2&gt;
&lt;p&gt;There is a wrinkle. Xinuos has said it will petition for an en banc rehearing, meaning the full bench of Second Circuit judges would reconsider the case rather than the three judge panel that heard it. IBM’s lawyers put the odds of that petition being granted at 0.03%, which is three hundredths of 1%, not 3%. An en banc rehearing is the legal equivalent of the last option you reach for when nobody is taking you seriously, and a 0.03% success rate is about as close to over as lawyers get in public while a case is technically still pending.&lt;/p&gt;
&lt;p&gt;There is a subtler open question too. Because the court decided on timing rather than on the merits, it never actually ruled on whether a company holding a non-exclusive license can contribute that code to an open source project without the owner’s permission. That legal principle was sidestepped rather than settled. For Xinuos it does not matter, since its path to litigate the question is closed. For the broader architecture of how code gets contributed to Linux, it means one specific question is still, strictly speaking, unanswered.&lt;/p&gt;
&lt;h2 id=&quot;the-defense-that-quietly-made-linux-hard-to-kill&quot;&gt;The defense that quietly made Linux hard to kill&lt;/h2&gt;
&lt;p&gt;One part of this story rarely gets told, and it is the part I find most interesting. While SCO was filing suits, the open source world built legal armor specifically to survive this kind of attack.&lt;/p&gt;
&lt;p&gt;The &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Open_Invention_Network&quot;&gt;Open Invention Network&lt;/a&gt;, founded in 2005 partly in response to the original SCO case, holds a large portfolio of patents and runs a binding framework where members cross license their patents to each other around Linux. IBM, Red Hat, Google, Sony, Philips, and many others take part. The point is to make attacking Linux expensive and risky, because a company that sues over Linux exposes itself to the network’s pooled defensive patents. A troll targeting Linux faces more than the one company it sued. It faces 20 years of accumulated legal infrastructure built for exactly that moment.&lt;/p&gt;
&lt;p&gt;That armor is also where the real vulnerability shows. Xinuos ran into IBM’s legal team and lost. Had the same theory been aimed at a small volunteer maintained distribution with no lawyers, no network membership, and a budget measured in donations, the outcome could have looked very different. The protection that shields Red Hat and Ubuntu is not automatically there for the hundreds of tiny distributions that also ship the same kernel.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.DRksBgq6_17sNXw.webp&quot; alt=&quot;Both ship the same Linux kernel, so both can be targeted by the same claims. Only one side can afford to answer.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;871&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Both ship the same Linux kernel, so both can be targeted by the same claims. Only one side can afford to answer. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://distrowatch.com/&quot;&gt;Soursen analysis&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-threat-that-outlived-this-one&quot;&gt;The threat that outlived this one&lt;/h2&gt;
&lt;p&gt;The Xinuos case was about copyright. The quieter, growing risk is about patents, and it hits small projects harder.&lt;/p&gt;
&lt;p&gt;Patent assertion entities, companies that buy patents purely to extract licensing money through litigation threats and never build a product, have been circling open source for years. In 2019 a troll called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/GNOME_Foundation&quot;&gt;Rothschild Patent Imaging&lt;/a&gt; sued the GNOME Foundation over its Shotwell photo app, claiming it infringed a patent. GNOME chose to fight, and the case settled in 2020 with the troll handing over a license and releasing its claims, but the episode rattled the community as a warning. Since then smaller projects have reported demand letters claiming that ordinary software practices, cryptographic protocols, interface patterns, networking behavior, infringe some patent.&lt;/p&gt;
&lt;p&gt;The mechanism is the same one the Xinuos case exposed from the copyright side. A large distributor like Red Hat, or Ubuntu’s parent Canonical, can absorb the cost of answering a demand letter and fighting if it has to. A volunteer project cannot. The letter does not have to win to work. The cost of responding, legal review, compliance checks, possible indemnity for users, can exceed what a donation funded project can carry. The threat itself is the weapon, and the lawsuit behind it may never need to come.&lt;/p&gt;
&lt;p&gt;Groups like Unified Patents and its open source work, backed by the Linux Foundation and others, exist to challenge bad patents before they reach that expensive stage, and they have knocked out some trolls. They cannot reach all of them, and they have little visibility into the quiet demand letters landing in the inboxes of projects too small to make noise.&lt;/p&gt;
&lt;h2 id=&quot;what-the-23-years-actually-bought&quot;&gt;What the 23 years actually bought&lt;/h2&gt;
&lt;p&gt;From March 2003, when SCO first sued and threatened per processor fees, to 10 August 2026, when the Second Circuit closed the appeal, a legal cloud sat over Linux’s core code for 23 years. The cloud was always more vapor than substance. Courts rejected the core claims again and again. The ownership theory never survived real scrutiny. But rejected and definitively resolved are different states, and the litigation kept finding new vehicles, a bankruptcy, an asset sale, a fresh plaintiff, to stay technically alive.&lt;/p&gt;
&lt;p&gt;The open source world moved forward anyway, treating the claims as meritless, and Linux went from a curiosity to running the overwhelming majority of the world’s servers during the exact window the lawsuit was pending. The case ending well is worth something. The structural weakness it exposed, that the smallest distributions have none of the legal protection the big ones rely on, is still sitting there waiting for the next plaintiff who picks an easier target.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>linux</category><category>ibm</category><category>ai</category><category>artificial-intelligence</category><category>tech</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Siemens Deleted a Popular Open Source Project on Its Way Out the Door, and the AGPL Made Sure That…</title><link>https://soursen.org/articles/siemens-deleted-a-popular-open-source-project-on-its-way-out-the-door/</link><guid isPermaLink="true">https://soursen.org/articles/siemens-deleted-a-popular-open-source-project-on-its-way-out-the-door/</guid><description>In October 2026, months after buying Altair for $10.6 billion, Siemens retired the OpenRadioss website and deleted its GitHub repository, pointing users to a paid product. The code was under the AGPL, so within days a Rocky Linux founder had forked it into a new project called OpenCourant. I went through what happened, why deleting the repo was the self defeating part, and what the license actually guaranteed.</description><pubDate>Sun, 25 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Open source has a recurring fear with a clumsy name, EEE, for embrace, extend, extinguish. A company adopts an open project under a permissive license, builds a user base on its version, quietly adds features that lock people in, and eventually takes the whole thing proprietary. The more mundane version is simpler, a company buys a project, promises to keep it open, and later decides it would rather you pay. It does not happen often, because the backlash is bad and the license usually does not allow a clean exit. In October 2026 it happened to OpenRadioss, and the way Siemens did it turned a quiet retirement into a public mess.&lt;/p&gt;
&lt;p&gt;If you visit the old OpenRadioss GitHub page now, you get a 404. The website no longer loads the project either, it redirects to a Siemens page telling you the future of Radioss is a paid product called Simcenter. I want to walk through how a widely used open source tool got deleted, and why the delete button was the mistake.&lt;/p&gt;
&lt;h2 id=&quot;what-radioss-is-and-how-it-got-here&quot;&gt;What Radioss is and how it got here&lt;/h2&gt;
&lt;p&gt;Radioss is a finite element solver, software that simulates how physical structures behave under extreme, fast, highly non-linear events, car crashes, impacts, and blasts. If you have seen a slow motion simulation of a vehicle crumpling in a safety test, that is the kind of problem Radioss handles. It has been an industry tool for decades, developed by the French firm Mecalog with roots going back to the 1980s.&lt;/p&gt;
&lt;p&gt;The ownership changed hands the way these things do. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Radioss&quot;&gt;Altair Engineering&lt;/a&gt;, a company based in Michigan, acquired Mecalog and its Radioss technology in 2006 and renamed it Altair Radioss. Then in September 2022, Altair did something generous. It kept its commercial product, and it also released a community edition under the GNU AGPL license, a genuinely open source version called OpenRadioss. For 4 years that version was the one students, researchers, and engineers reached for when they needed serious crash simulation without a commercial license. It built a real community around teaching and research.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CHwZZGnT_9rpco.webp&quot; alt=&quot;Figure 1. The ownership chain. Radioss came from the French firm Mecalog, was acquired by Altair in 2006, opened as OpenRadioss under the AGPL in 2022, and passed to Siemens when it bought Altair, after which the open version was shut down.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1006&quot; height=&quot;575&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. The ownership chain. Radioss came from the French firm Mecalog, was acquired by Altair in 2006, opened as OpenRadioss under the AGPL in 2022, and passed to Siemens when it bought Altair, after which the open version was shut down. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Radioss&quot;&gt;Author diagram based on Radioss project history&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;siemens-buys-the-company-then-the-project&quot;&gt;Siemens buys the company, then the project&lt;/h2&gt;
&lt;p&gt;The trouble started with an acquisition. In October 2024 Siemens announced it would buy Altair for about $10.6 billion, a deal that closed in 2025. Radioss was a small piece of that, Siemens was mainly buying Altair’s broad portfolio of simulation, high performance computing, and data science software to fill gaps in its own lineup. When you buy a company that size, you absorb everything it owns, including an open source project with an active community.&lt;/p&gt;
&lt;p&gt;At first nothing changed. Siemens renamed the commercial product to Simcenter Radioss and folded it into its Simcenter suite, which is a normal thing to do after an acquisition. OpenRadioss kept running, development continued, and the community had no particular reason to worry. That held for about a year.&lt;/p&gt;
&lt;h2 id=&quot;the-part-that-made-people-angry&quot;&gt;The part that made people angry&lt;/h2&gt;
&lt;p&gt;On 1 October 2026, the situation changed. Siemens retired the OpenRadioss website and, shortly after, removed the GitHub repository, as reported by &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Siemens-Ends-OpenRadioss&quot;&gt;Michael Larabel at Phoronix&lt;/a&gt;. The site now redirects to a blog post titled the future of Radioss is Simcenter. The post frames it in the language these announcements always use, that after 4 years of successful community driven research, Siemens is consolidating its efforts and entering a new chapter, focusing future investment on the Simcenter portfolio to provide support for production grade simulation. The FAQ tells visitors who relied on the free version to explore Simcenter Radioss and its R&amp;amp;D program through Siemens.&lt;/p&gt;
&lt;p&gt;Read plainly, it is a redirect from a free tool to a paid one, built in part on 4 years of unpaid community research and development that the previous owner had committed to keeping open. That alone would have annoyed people. Deciding to stop supporting an open source project is a company’s right. Announce it, leave the code where it is, and let whoever wants to maintain it carry on.&lt;/p&gt;
&lt;p&gt;That is not what happened. Siemens deleted the repository. A project that thousands of people depended on returned a 404, and Phoronix noted the repo was removed, not archived, when archiving would have preserved it in read only form at no cost. There was no practical reason to delete it. It reads as a deliberate attempt to make a clean copy harder to find and fork, an extra kick on the way out that was never going to win anyone over. Nobody who just got locked out of a tool they relied on responds by happily buying the paid version.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.6-Rh1wZY_Z1kIRRW.webp&quot; alt=&quot;Figure 2. 2 ways to leave an open source project. Walking away leaves the code in place for the community to maintain. Deleting the repository strands users, and under a copyleft license it does not even stop the fork.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1096&quot; height=&quot;575&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. 2 ways to leave an open source project. Walking away leaves the code in place for the community to maintain. Deleting the repository strands users, and under a copyleft license it does not even stop the fork. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Siemens-Ends-OpenRadioss&quot;&gt;Author diagram of the 2 exit paths&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;why-the-agpl-made-the-delete-pointless&quot;&gt;Why the AGPL made the delete pointless&lt;/h2&gt;
&lt;p&gt;The delete button did not work the way Siemens might have hoped, because of the license. OpenRadioss was released under the GNU AGPL, a strong copyleft license. Once code is published under it, anyone who received a copy has the right to use, modify, and redistribute it, and that right does not expire when the original owner changes its mind. Deleting the central repository does not revoke the license on every copy already out in the world. It just makes the most convenient copy harder to reach.&lt;/p&gt;
&lt;p&gt;So the code was never really Siemens’ to pull back. Any engineer, lab, or company that had cloned the repository, and in a 4 year old project with an active user base that is a lot of them, still held a fully licensed copy they were free to republish. A permissive giveaway can sometimes be quietly walked back. A copyleft project, once it is out, stays out. That is the entire point of the license, and it is why the deletion looked less like a power move and more like a company forgetting which license it was dealing with.&lt;/p&gt;
&lt;h2 id=&quot;rocky-linux-picked-it-up-within-days&quot;&gt;Rocky Linux picked it up within days&lt;/h2&gt;
&lt;p&gt;The fork that matters is called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://opencourant.org/&quot;&gt;OpenCourant&lt;/a&gt;, and it is serious because of who is behind it. It was announced on 1 October 2026 by Brian Clemens, a founder and vice president of the Rocky Enterprise Software Foundation and Rocky Linux. This is not a random one person fork with no staying power. Rocky Linux was started in 2020 by Gregory Kurtzer, who had co-founded the original CentOS, after Red Hat turned CentOS into the less stable CentOS Stream and left a lot of enterprises without the free, bug-for-bug compatible version of Red Hat Enterprise Linux they had been running production systems on. Rocky Linux filled that hole and has kept a major enterprise distribution going ever since, with a commercial sponsor in Kurtzer’s company CIQ. The same organization taking over OpenRadioss means the fork has real institutional weight behind it, people who have already proven they can run a serious project long term after a corporate owner pulled the rug.&lt;/p&gt;
&lt;p&gt;The handover was not clean, precisely because of the deletion. The team built OpenCourant from the latest open source code they could obtain and kept the AGPL license. The problem was that a proprietary input reader library had never been kept in git, and the newest copies they could recover from public sources were months older than the rest of the code base expected. Because the repository had been deleted outright, there was no up to date backup to pull from, so the project put out a call asking anyone with a recent OpenRadioss package to help recover the missing pieces. A community member who had archived the original package came forward, the team verified it against an independently recovered binary of the same build, and the gap started closing.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.CpNaz7UC_zaHPU.webp&quot; alt=&quot;Figure 3. The timeline, from open to forked. OpenRadioss ran openly for 4 years, Siemens deleted it on 1 October 2026, and the Rocky Linux team announced the OpenCourant fork the same day, shipping builds within days.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;983&quot; height=&quot;544&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. The timeline, from open to forked. OpenRadioss ran openly for 4 years, Siemens deleted it on 1 October 2026, and the Rocky Linux team announced the OpenCourant fork the same day, shipping builds within days. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://fossforce.com/2026/10/rocky-linux-rescues-openradioss-from-siemens-dumpster/&quot;&gt;Author timeline based on Phoronix, FOSS Force, and OpenCourant records&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;this-keeps-happening-and-the-fork-keeps-winning&quot;&gt;This keeps happening, and the fork keeps winning&lt;/h2&gt;
&lt;p&gt;OpenRadioss is the latest entry in a pattern that has been building for years, a company that owns a widely used open source project decides it wants more money from it, changes the terms, and the community forks the code and walks away. The specifics differ, but the shape is the same, and the forks have been doing well.&lt;/p&gt;
&lt;p&gt;In 2021 &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thenewstack.io/what-happens-to-relicensed-open-source-projects-and-their-forks/&quot;&gt;Elastic&lt;/a&gt; moved Elasticsearch off the open Apache 2.0 license, and Amazon forked it into OpenSearch, now a Linux Foundation project. In August 2023 HashiCorp moved Terraform to the restrictive Business Source License, and within 2 weeks users forked it into OpenTofu, also under the Linux Foundation. In March 2024 Redis abandoned its BSD license, breaking a 2018 pledge to keep it, and 8 days later contributors forked Valkey from the last open version, again under the Linux Foundation, and many of Redis’s own external contributors moved over to it. Each time, a neutral foundation ended up hosting a fork with more organizational diversity than the relicensed original.&lt;/p&gt;
&lt;p&gt;OpenRadioss is a sharper version of the same story, and the license made the difference. Those earlier cases were relicensing, a company changing the terms on new versions while the old open code stayed available to fork. Siemens went further and deleted the code outright. The twist is that it did not matter. Because OpenRadioss was AGPL, a strong copyleft license, the deleted code was already free in every copy that had been downloaded, so the fork was as clean as the others despite the repository being gone. A company can change a license going forward. It cannot un-free code that is already out under copyleft.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.Du36WjE7_uiGqc.webp&quot; alt=&quot;Figure 4. OpenRadioss joins a run of open source projects taken commercial, each answered by a community fork under a neutral foundation. Elasticsearch, Terraform, and Redis were relicensed, OpenRadioss was deleted, and all 4 were forked.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1124&quot; height=&quot;576&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. OpenRadioss joins a run of open source projects taken commercial, each answered by a community fork under a neutral foundation. Elasticsearch, Terraform, and Redis were relicensed, OpenRadioss was deleted, and all 4 were forked. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thenewstack.io/what-happens-to-relicensed-open-source-projects-and-their-forks/&quot;&gt;Author chart based on relicensing and fork reporting&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;where-it-stands-now&quot;&gt;Where it stands now&lt;/h2&gt;
&lt;p&gt;By the Tuesday after the shutdown, OpenCourant was shipping builds for the 3 platforms Radioss had supported, Linux x86–64, Linux arm64, and Windows x86–64. One reader function is still missing across all platforms, and the project is asking anyone holding an OpenRadioss package from August or September 2026 to check whether it contains the piece. You can get involved through the OpenCourant GitHub, the project’s discussions, or the Rocky Linux community channels, and the team has specifically said it wants to hear from former OpenRadioss maintainers and community leaders.&lt;/p&gt;
&lt;p&gt;The speed of the recovery is the real story. A company with a $10.6 billion acquisition behind it deleted a project, and within days a volunteer backed organization had a working, rebranded, actively maintained replacement on 3 platforms. That is what a copyleft license buys a community, the ability to route around an owner who changes the rules.&lt;/p&gt;
&lt;h2 id=&quot;the-part-worth-remembering&quot;&gt;The part worth remembering&lt;/h2&gt;
&lt;p&gt;A company doing something does not mean its engineers agree with it. A Siemens employee commented publicly that watching this was sad, that some of them had been pushing hard for open source culture inside the company, and that setbacks like this one were disheartening. That rings true at most large companies, where a decision from the top does not reflect the people who have to watch it land. Plenty of engineers inside Siemens almost certainly wanted this handled differently.&lt;/p&gt;
&lt;p&gt;The lesson for anyone depending on a corporate backed open source project is that the license is what you are actually relying on, more than any company promise. Altair’s pledge to keep Radioss open meant nothing once Altair no longer owned it. The AGPL, on the other hand, meant everything, because it is the reason OpenCourant exists at all. If you build on an open project owned by a company that could be acquired, check the license, keep your own copy of the code, and know that a strong copyleft license is the thing that lets the community survive a new owner who would rather you paid.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>siemens</category><category>open-source</category><category>projects</category><category>ai</category><category>artificial-intelligence</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>MacOS Quietly Vanished From the Official Unix Registry</title><link>https://soursen.org/articles/macos-quietly-vanished-from-the-official-unix-registry/</link><guid isPermaLink="true">https://soursen.org/articles/macos-quietly-vanished-from-the-official-unix-registry/</guid><description>For almost 20 years, “macOS is certified Unix” was a real bragging right. I dug into what that certification actually required, and the honest answer is that no Mac anyone uses has ever met it. Here is what disappeared from the Open Group’s list in October 2026, and why it matters less than the headlines suggest.</description><pubDate>Wed, 21 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;For most of 2 decades, one of the quiet flexes in computing was that macOS is certified Unix and Linux is not. I heard it from university professors, from developers who came up in the 90s, from people who wanted a machine that was Unix without the rough edges. Then, sometime between October 7 and October 11, 2026, Apple’s entries slipped off the front page of the Open Group’s Unix registry, with no announcement, no blog post, nothing. I went looking at what that certification was, what it actually required, and whether losing it means anything. The answer turned out to be more interesting than the disappearance itself.&lt;/p&gt;
&lt;h2 id=&quot;what-actually-happened-on-the-registry&quot;&gt;What actually happened on the registry&lt;/h2&gt;
&lt;p&gt;The Open Group owns the Unix trademark. It is the body that gets to say a given operating system is Unix or is not Unix, and it keeps a public list of certified products. If you pull up a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://web.archive.org/web/2026*/https://www.opengroup.org/openbrand/register/&quot;&gt;Wayback Machine&lt;/a&gt; capture from October 7, 2026, Apple is there, macOS version 26 Tahoe, both the Intel and Apple Silicon builds. A capture a few days later shows only IBM, Hewlett Packard Enterprise, and the SCO Group. Apple is gone from the index.&lt;/p&gt;
&lt;p&gt;Here is the odd part. macOS is not fully gone. Filter the registry by the specific standard, Unix 03, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.opengroup.org/openbrand/register/brand3720.htm&quot;&gt;macOS Tahoe is still listed there&lt;/a&gt;, with its certificate dated 29 August 2025. So the certificate exists, the product record exists, but the entry vanished from the main list that aggregates everything. Nobody at Apple or the Open Group has said why. It could be a removal, it could be a site bug, though a bug that selectively drops only Apple for a week is a strange kind of bug. Several people who &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://dev.to/axrisi/macos-tahoe-and-unix-certification-checking-the-exact-product-record-4mfd&quot;&gt;tracked the exact product record&lt;/a&gt; landed on the same cautious read. The record stands, the front page listing is gone, and the reason is unexplained.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.ByYLJfjB_ZG2rhQ.webp&quot; alt=&quot;Figure 1. macOS disappeared from the main Unix registry index between October 7 and 11, 2026, while its Unix 03 product record stayed live. No announcement accompanied the change.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1120&quot; height=&quot;601&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. macOS disappeared from the main Unix registry index between October 7 and 11, 2026, while its Unix 03 product record stayed live. No announcement accompanied the change. Source: Open Group registry, Wayback Machine captures&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;why-macos-was-ever-on-that-list&quot;&gt;Why macOS was ever on that list&lt;/h2&gt;
&lt;p&gt;To see why this is funny more than tragic, you have to know why Apple bothered in the first place. macOS has been Unix certified since OS X 10.5 Leopard, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://appleinsider.com/articles/07/08/01/apples_leopard_gains_unix_03_certification&quot;&gt;registered in 2007&lt;/a&gt;. That was a moment when Unix still carried weight. The people who had lived through the Unix peak of the 80s and 90s were still running labs and making purchasing decisions, and calling your product Unix was a direct line to developer credibility.&lt;/p&gt;
&lt;p&gt;Steve Jobs understood this better than anyone, and he said it out loud, repeatedly. Back in 2001 he was already telling reporters that Apple would become the largest Unix supplier in the world by volume by the end of the year. At Macworld New York in 2002 he put it plainly from the stage. macOS is “the number one supplier of Unix in the world. Bigger than Sun, bigger than Linux, bigger than anyone.” Apple was selling a point and click machine to people who had been wrestling with the MS DOS command line, and the Unix label was how Jobs told developers it was serious underneath.&lt;/p&gt;
&lt;p&gt;The technical lineage was real, which is part of why the pitch worked. Apple did not fake a Unix core. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Jordan_Hubbard&quot;&gt;Jordan Hubbard&lt;/a&gt;, who co-founded FreeBSD in 1993, joined Apple in July 2001 to run its BSD technology group and later became Director of Unix Technology. The macOS kernel, XNU, is a hybrid of the Mach kernel and BSD code that has been synchronized with FreeBSD, and a large part of the user space traces back to FreeBSD too. The joke inside the name says the rest. XNU stands for “X is Not Unix,” which is the kind of thing only people who deeply respect Unix would name their Unix like kernel.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.CkrycIs5_ZgQoeK.webp&quot; alt=&quot;Figure 2. The Unix lineage under macOS is genuine. XNU blends the Mach kernel with BSD code, and the user space descends from FreeBSD, all shipped as the open source Darwin core.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1120&quot; height=&quot;544&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. The Unix lineage under macOS is genuine. XNU blends the Mach kernel with BSD code, and the user space descends from FreeBSD, all shipped as the open source Darwin core. Source: XNU and Darwin lineage&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;There was a fight over the label, too. The Open Group sued Apple over its use of the Unix name for Mac OS X, and Apple countersued, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.mactech.com/2003/06/12/apple-being-sued-over-unix-trademark/&quot;&gt;arguing the term had become generic&lt;/a&gt; and the trademark should be invalid. The case went through discovery and mediation without a public settlement. What Apple did in the years after was simpler than winning. It just paid for the certification and kept paying, year after year, long past the point where it mattered to the marketing.&lt;/p&gt;
&lt;p&gt;Apple kept the mark almost continuously for nearly 2 decades, with 1 gap. Leopard 10.5 was certified in 2007, Snow Leopard 10.6 followed, then 10.7 Lion was never registered as a Unix 03 product, a lapse that Andrew Josey, the Open Group’s standards lead, later confirmed. Apple picked it back up with Mountain Lion 10.8 in 2012 and held it straight through to macOS 26 Tahoe, whose certificate is dated 29 August 2025. Then came the quiet exit from the index in October 2026.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.Dd9x3atk_SEyKm.webp&quot; alt=&quot;Figure 3. Apple held Unix 03 certification from Leopard in 2007 almost continuously, with 1 lapse at Lion in 2011, through Tahoe in 2025, then slipped off the registry index in October 2026.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1111&quot; height=&quot;509&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. Apple held Unix 03 certification from Leopard in 2007 almost continuously, with 1 lapse at Lion in 2011, through Tahoe in 2025, then slipped off the registry index in October 2026. Source: Open Group register and Apple certification histor&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-the-certification-was-ever-for&quot;&gt;What the certification was ever for&lt;/h2&gt;
&lt;p&gt;To judge whether losing the mark matters, it helps to know why the mark exists. Through the late 1980s and early 1990s, Unix went through what people later called the Unix wars. AT&amp;amp;T had System V, Berkeley had BSD, and nearly every hardware vendor shipped its own incompatible variant, so a program written for Sun’s Unix would not simply run on IBM’s or HP’s. The fragmentation was brutal. The database maker Informix at one point supported more than 100 different Unix systems. Software had to be ported again and again, which is the opposite of the portability Unix was supposed to deliver.&lt;/p&gt;
&lt;p&gt;The fix was a shared standard. 2 rival camps, the Open Software Foundation and Unix International, buried the hatchet and formed the Common Open Software Environment alliance in March 1993, and the Unix trademark passed to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Single_UNIX_Specification&quot;&gt;X/Open&lt;/a&gt; later that year. That group, which merged into the Open Group in 1996, produced the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://unix.org/unix_history.html&quot;&gt;Single Unix Specification&lt;/a&gt;, the actual document the certification tests against. The versions line up with the marks. Unix 95 arrived in 1994 with 1,168 interfaces, Unix 98 in 1997 with 1,434, Unix 03 in 2002 with 1,742, and the current Unix V7 in 2008 with 1,833. The whole point was to let a vendor prove, on paper, that its system met a common bar so buyers could trust their software would run.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.2LuTXICi_ZCslce.webp&quot; alt=&quot;Figure 4. The Single Unix Specification grew through 4 marks, from Unix 95 to Unix V7, each adding interfaces. macOS is certified to Unix 03, the 2002 revision.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1101&quot; height=&quot;591&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. The Single Unix Specification grew through 4 marks, from Unix 95 to Unix V7, each adding interfaces. macOS is certified to Unix 03, the 2002 revision. Source: Single Unix Specification history, The Open Group&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The core of that standard is POSIX. The Single Unix Specification’s base is, by the Open Group’s own description, technically identical to the POSIX standard, which is the thing that actually decides whether your code ports. That distinction matters for the rest of this story, because it separates the part of Unix certification that changed from the part that did not.&lt;/p&gt;
&lt;h2 id=&quot;the-certification-almost-nobody-can-actually-meet&quot;&gt;The certification almost nobody can actually meet&lt;/h2&gt;
&lt;p&gt;Now the part that reframes the whole thing. Apple has been paying to certify a configuration of macOS that effectively does not exist on anyone’s machine. In January 2025, Thom Holwerda published a piece on OS News with a blunt title, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.osnews.com/story/141633/apples-macos-unix-certification-is-a-lie/&quot;&gt;Apple’s macOS Unix certification is a lie&lt;/a&gt;, and the evidence is sitting in Apple’s own public conformance statements.&lt;/p&gt;
&lt;p&gt;At the end of each conformance statement there is a field where the vendor lists the additional steps needed to make the system pass. For macOS 15, the version right before Tahoe, the list is long. To pass Unix 03 you have to disable System Integrity Protection, the security layer that stops even privileged code from modifying system binaries. You have to enable the root account, which is off by default. You have to turn on core file generation, disable timeout coalescing, mount the file system with strict access time updates, and reformat your APFS volume as case sensitive, which on its own requires a reinstall because APFS ships case insensitive. You have to disable Spotlight, the background file indexer. Then, to pass the UUCP tests, you copy a set of ancient Unix copy binaries into a new location and set their permission bits by hand.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.BA7pyfp7_ZG20qo.webp&quot; alt=&quot;Figure 5. The changes Apple documents to make macOS 15 pass Unix 03 certification. Almost no real Mac has any of these, let alone all of them.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1120&quot; height=&quot;667&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 5. The changes Apple documents to make macOS 15 pass Unix 03 certification. Almost no real Mac has any of these, let alone all of them. Source: Apple conformance statements, via OS News analysis 2025&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;On top of those changes, Apple holds 4 temporary waivers, granted at the Open Group’s discretion for implementation errors judged minor, valid for 12 months before the vendor is supposed to fix them. Those waivers are only visible to paying members, so outsiders cannot even read what was waived. Stack it all up and the conclusion is hard to dodge. There is effectively no macOS system in daily use, outside the certification test rig itself, that actually meets the Unix 03 standard Apple pays to claim.&lt;/p&gt;
&lt;h2 id=&quot;how-that-compares-to-a-real-certification&quot;&gt;How that compares to a real certification&lt;/h2&gt;
&lt;p&gt;The reason this stands out is that other certified systems do not need any of it. For fairness, Holwerda checked the other conformance statements, in case everyone cheats and the whole process is theater.&lt;/p&gt;
&lt;p&gt;The only operating system currently certified to Unix V7, the latest version of the spec, is IBM’s AIX 7.2 or later. Its conformance statement contains exactly 1 required change, setting a socket listen queue length to 5 with a single command. That is the whole list. HP-UX 11.31 for Itanium has remarks in its appendix, but they are informative and specify no changes. Solaris has a pile of remarks about differences between its x86 and SPARC builds, again with no required changes. 1 command for AIX, nothing for the others, against a wall of security teardowns and reinstalls for macOS.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-06.DV-8qY11_Z1KgNS4.webp&quot; alt=&quot;Figure 6. What each certified Unix requires to pass. AIX needs 1 command. macOS needs its entire security model dismantled plus 4 waivers.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1090&quot; height=&quot;591&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 6. What each certified Unix requires to pass. AIX needs 1 command. macOS needs its entire security model dismantled plus 4 waivers. Source: Open Group conformance statements, 2025&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;There is also a version gap worth naming. macOS sits on Unix 03, a spec released on 30 January 2002, despite the name. The current mark is Unix V7, tied to the 2008 revision of the spec, and the only thing on that list is AIX. So even the certification macOS holds is a generation behind the current standard, and it meets that older standard only in a torn down configuration nobody runs.&lt;/p&gt;
&lt;h2 id=&quot;the-certification-used-to-buy-real-money&quot;&gt;The certification used to buy real money&lt;/h2&gt;
&lt;p&gt;There was a time when this label was not a vanity badge. In the 1980s and 1990s, standards compliance was written into government purchasing rules. The US federal standard &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Microsoft_POSIX_subsystem&quot;&gt;FIPS 151–2&lt;/a&gt; required certain government computer purchases to be POSIX compliant, and that requirement was strong enough that Microsoft built an entire POSIX subsystem into Windows NT, shipped in 1993, purely so NT machines would stay eligible for federal contracts. Windows NT 3.5, 3.51, and 4.0 were all certified against FIPS 151–2. Microsoft did not add POSIX to NT because anyone loved it. They added it because contracts demanded it.&lt;/p&gt;
&lt;p&gt;Unix certification sat in the same world. For a stretch, being able to say your system was certified Unix or POSIX compliant was a gate to government labs, universities, and defense work that would not sign a purchase order otherwise. That is the market Apple was partly chasing when it certified Leopard in 2007 and when it sold its Xserve rack machines to institutions. The mark was a key to a room with budget in it.&lt;/p&gt;
&lt;p&gt;That room mostly emptied out. Linux, which never carried the Unix trademark, became acceptable and then dominant in exactly those institutional and server markets, and procurement rules relaxed to accept de facto standards, not just certified ones. Once the buyers stopped asking for the certificate, the certificate stopped being worth chasing, which is the quiet economic story under Apple’s registry entry going stale.&lt;/p&gt;
&lt;h2 id=&quot;why-apple-stopped-caring-and-why-it-barely-matters&quot;&gt;Why Apple stopped caring, and why it barely matters&lt;/h2&gt;
&lt;p&gt;Apple did not suddenly lose the technical ability to be Unix. The Darwin core, the XNU kernel, the BSD user space, the whole foundation, none of that changed when the registry entry vanished. What changed is the audience. The Apple of 2007 was courting developers, and Unix was a credential worth buying. The Apple of 2026 sells to visual creatives who have never typed csrutil in their life and do not care whether a trademark body in a registry somewhere blesses the kernel. There is no mention of Unix in Apple’s marketing anymore. It does not appear in release notes. For a company Apple’s size the certification fee is a rounding error, so cost is not really the story either. Someone may simply have looked at the renewal and asked why they were still paying for a label no customer reads.&lt;/p&gt;
&lt;p&gt;There is a deeper point underneath, which several engineers have made for years. Because the Unix copyright and the Unix trademark ended up owned by different organizations, the word Unix no longer cleanly means one thing. The BSDs are Unix derived in every sense except the narrow one that copyright law cares about. macOS is BSD derived through and through. If anyone seriously challenged the trademark in court on the grounds that the term has become generic, which is close to the argument Apple itself once made, the Open Group might well lose it. In normal usage, when a developer says a system is Unix, they mean it behaves like Unix and descends from Unix, not that someone paid a certification fee this year.&lt;/p&gt;
&lt;h2 id=&quot;what-this-actually-changes-for-you&quot;&gt;What this actually changes for you&lt;/h2&gt;
&lt;p&gt;If you use a Mac, nothing changes. Your shell works the same, your BSD tools work the same, the POSIX behavior most software actually depends on is untouched, because that was never the thing the Unix 03 certificate was really testing. POSIX compliance, the standard that genuinely matters for portability, is a separate question from the Open Group’s Unix branding, and it is the one worth watching if Apple ever starts trimming. The registry entry was always more flag than foundation.&lt;/p&gt;
&lt;p&gt;What the disappearance does is puncture a particular flex. For the handful of people who liked to say macOS is Unix and Linux is not, the first half of that sentence just got weaker, and it was already resting on a certification that described a machine none of them were running. Linux, which was never certified and never paid the fee, meanwhile became the practical standard that the entire server world runs on. The label and the reality drifted apart a long time ago. October 2026 just made the paperwork match.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>unix</category><category>apple</category><category>registry</category><category>macos</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>The Linux Kernel Just Dropped Its AI Disclosure Tag, and the Fight Over It Says More Than the…</title><link>https://soursen.org/articles/the-linux-kernel-just-dropped-its-ai-disclosure-tag-and-the-fight-over/</link><guid isPermaLink="true">https://soursen.org/articles/the-linux-kernel-just-dropped-its-ai-disclosure-tag-and-the-fight-over/</guid><description>At the Maintainer Summit in Prague, the people who run the Linux kernel reached a rough consensus to deprecate the Assisted-by tag, the little line in a commit that says AI helped write it. I went through who proposed it, the real numbers behind it, and the argument it set off, including Linus Torvalds calling the whole idea of AI disclosure rules utterly stupid, because the kernel is quietly becoming one of the most AI touched projects in open source.</description><pubDate>Wed, 21 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;For about a year, a Linux kernel commit could carry a small line called Assisted-by, a tag that said AI was used somewhere in making that change. It was meant to be a signal, a way to know a model had a hand in the code. On 8 October 2026, at the Maintainer Summit in Prague, a room of roughly 30 senior kernel developers reached a rough consensus to get rid of it. The next day, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-Dropping-Assisted-By-Req&quot;&gt;Sasha Levin&lt;/a&gt;, an Nvidia engineer who co-maintains the stable and long term kernels, sent the patch to deprecate the tag and remove the documentation requiring it. Jonathan Corbet queued the change into the docs branch. The tag is on its way out.&lt;/p&gt;
&lt;p&gt;The decision itself is small. The argument around it is the interesting part, because it exposes how the kernel’s leadership actually thinks about AI, and how much AI is already in there.&lt;/p&gt;
&lt;h2 id=&quot;what-the-tag-was-supposed-to-do&quot;&gt;What the tag was supposed to do&lt;/h2&gt;
&lt;p&gt;The idea behind Assisted-by was data. If you want to know how much AI is flowing into the kernel, which developers lean on it, how that changes over time, a tag in the commit gives you a record you can count. Sasha Levin was in favor of it originally for exactly that reason, to track how things are changing, not to police anyone. A commit with the tag told you a model helped. A commit without it told you, in theory, that none did.&lt;/p&gt;
&lt;p&gt;That theory broke down fast, and the reasons it broke are why the tag is being dropped.&lt;/p&gt;
&lt;h2 id=&quot;where-the-tag-came-from-in-the-first-place&quot;&gt;Where the tag came from in the first place&lt;/h2&gt;
&lt;p&gt;The Assisted-by tag was written into kernel policy less than a year ago. On 23 December 2025, a document called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lilting.ch/en/articles/linux-kernel-ai-coding-assistant-policy&quot;&gt;coding-assistants.rst&lt;/a&gt;, about 59 lines, was committed to the kernel tree, approved by documentation maintainer Jonathan Corbet and built on consensus reached at the 2025 Maintainer Summit. Sasha Levin championed it, and the naming went through a few rounds before it stuck. His first RFC in July 2025 floated a Co-developed-by tag, then after suggestions from David Hildenbrand and Konstantin Ryabitsev it became Assisted-by. The kernel’s own patch checker, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://checkpatch.pl/&quot;&gt;checkpatch.pl&lt;/a&gt;, at first rejected the new tag, and Levin had to send a second version to make the tooling accept it.&lt;/p&gt;
&lt;p&gt;The policy did more than name a tag. It set the rules that still hold even with the tag going away. Only a human can add the Signed-off-by line, because that line certifies the Developer Certificate of Origin, the legal statement that you have the right to contribute this code. An AI agent cannot sign off, full stop. The human who submits AI assisted code reviews it, checks it is compatible with the kernel’s GPL-2.0-only license, adds their own Signed-off-by, and takes full accountability for it. The tag format itself was specific, AGENT_NAME:MODEL_VERSION followed by any static analysis tools in brackets, with the policy’s own example reading Claude:claude-3-opus coccinelle sparse. Ordinary tools like git, gcc, and your editor were never meant to be listed. Even in that original policy, Levin said enforcement was deliberately avoided. The tag worked as a forensic record for spotting bug patterns by model, and skipping it carried no penalty.&lt;/p&gt;
&lt;h2 id=&quot;the-3-reasons-it-is-going-away&quot;&gt;The 3 reasons it is going away&lt;/h2&gt;
&lt;p&gt;Levin’s own analysis is the clearest look at the problem. He went through 18,533 regular commits across a recent development cycle and found that about 10% carried the Assisted-by tag. He is the first to say that figure almost certainly understates real AI use, because the tag only catches the people who bothered to add it.&lt;/p&gt;
&lt;p&gt;That gap is the whole issue, and it comes down to 3 things. First, plenty of developers do not know the rule exists, so they never add the tag even when a model helped. Second, some developers know and leave it off on purpose. Third, maintainers rarely chase a missing tag, so there is no real enforcement. Put those together and a commit without the tag tells you almost nothing. The absence could mean no AI, or it could mean someone who did not know, forgot, or chose not to say. A tag that cannot be trusted when it is missing is not giving you clean data, it is giving you a floor and a shrug.&lt;/p&gt;
&lt;p&gt;A second objection has nothing to do with accuracy. The tag often named a specific commercial product, the model or the agent used. A lot of people were uncomfortable turning kernel commit history into free advertising for proprietary AI companies. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-Dropping-Assisted-By-Req&quot;&gt;Greg Kroah-Hartman&lt;/a&gt;’s position captured the compromise, make up a generic name if you have to, all he wanted was the signal that a model was involved so he could review the change with that in mind, and if someone lied about it, that was on them.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CE-eTRFd_1zgQPB.webp&quot; alt=&quot;Figure 1. Why the Assisted-by tag stopped being useful. Across 18,533 commits in a recent cycle, about 10% carried the tag, and the missing 90% is a mix of no AI, developers who did not know the rule, and developers who left it off, so the absence of a tag means little.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1150&quot; height=&quot;631&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. Why the Assisted-by tag stopped being useful. Across 18,533 commits in a recent cycle, about 10% carried the tag, and the missing 90% is a mix of no AI, developers who did not know the rule, and developers who left it off, so the absence of a tag means little. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-Dropping-Assisted-By-Req&quot;&gt;Author chart based on Sasha Levin’s kernel commit analysis&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-people-who-wanted-to-keep-it&quot;&gt;The people who wanted to keep it&lt;/h2&gt;
&lt;p&gt;The consensus was rough, not unanimous, and the pushback is worth hearing because it is not coming from people who hate AI. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thenextweb.com/news/torvalds-ai-assisted-by-tag-utterly-stupid-open-source-summit&quot;&gt;Laurent Pinchart&lt;/a&gt;, one of the more AI skeptical voices among the maintainers, made a practical point. For him the useful thing was the absence of a tag, because it gave him a reason to ask a submitter whether they had forgotten to add it, and in many cases the submitter simply went quiet. Someone who sends a patch that looks machine generated, gets asked to tag it, and then ghosts the reviewer has told you something. Pinchart did not want the tag removed without some other required disclosure in its place, otherwise submitters just push back on the question and there is nothing to point to.&lt;/p&gt;
&lt;p&gt;Pinchart raised a second idea that cuts deeper. Requiring a human written commit message, he argued, is a cheap filter. Most LLM generated commit messages are easy to spot today because they run more verbose than a person would bother with, and asking for a human written message is a quick way to ignore patches from what he called meat proxies, people passing a model’s output through as their own. He also named the obvious counterargument himself and answered it. When someone says the tech will improve and the tells will vanish, his reply was that the process can change when that happens, there is no need to rewrite the rules now for a problem that does not exist yet.&lt;/p&gt;
&lt;p&gt;Kees Cook, another longtime maintainer, agreed the tag was unenforceable, and then said something more honest than most of the thread. His own recent commit logs, like his code, are a mix of his work and an LLM’s, and he trims and edits and reorganizes until the line between the 2 is gone. He admitted he had already failed a reverse Turing test, assuming a colleague’s genuinely human, genuinely verbose commit messages were AI generated. His read was that the point where you can clearly tell human from model is either here already or arriving fast. Once you cannot reliably tell, a disclosure tag built on the honor system is worth even less.&lt;/p&gt;
&lt;h2 id=&quot;then-torvalds-ended-the-discussion&quot;&gt;Then Torvalds ended the discussion&lt;/h2&gt;
&lt;p&gt;Linus Torvalds does not weigh in on everything, and when he does it tends to settle the direction. On 13 October 2026 he &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thenextweb.com/news/torvalds-ai-assisted-by-tag-utterly-stupid-open-source-summit&quot;&gt;replied to Pinchart&lt;/a&gt; on the mailing list, and he was blunt. He said the disclosure argument had nothing to do with expectations and everything to do with people’s hang ups, and he called the whole concept of don’t use AI for this specific thing idiotic.&lt;/p&gt;
&lt;p&gt;Then he reached for an analogy. Here is a new rule just for you, he wrote, do not use compilers for any of your kernel submissions, because if you cannot show you really understand what the machine is doing by writing your code in assembly, he did not want to see your contribution. He kept going, and you had better not use an editor to write your commit message either, because an editor lets you backspace and hide the mistakes you made while typing. He was being facetious, obviously. The point underneath was serious. You already do not understand everything the machine does when you use a compiler, and you trust it anyway, so singling out LLMs as the tool that proves you do not know what you are doing makes no sense. The expectation, he said, is that people do their best and understand what they are doing, and that understanding has very little to do with which tools they used.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.DUFuexCH_Z2j7ffR.webp&quot; alt=&quot;Figure 2. The kernel’s AI stance, by who holds it. From Torvalds, who committed code built with Claude Code, through pragmatists like Kroah-Hartman and Cook, to skeptics like Pinchart who still want disclosure, no major figure is calling for an outright ban.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1054&quot; height=&quot;525&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. The kernel’s AI stance, by who holds it. From Torvalds, who committed code built with Claude Code, through pragmatists like Kroah-Hartman and Cook, to skeptics like Pinchart who still want disclosure, no major figure is calling for an outright ban. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thenextweb.com/news/torvalds-ai-assisted-by-tag-utterly-stupid-open-source-summit&quot;&gt;Author diagram of kernel maintainer positions, October 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;This was not an abstract position for Torvalds. A few days earlier he had &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/linus-torvalds-claude-code-linux/&quot;&gt;published a branch&lt;/a&gt; where he used an AI coding assistant himself. The series rewrote part of the page cache, the code that handles many threads reading the same file at once, where all those reads were hammering a single shared counter. His 8 patches skip that per read step and let memory leaving the cache wait briefly before reuse so readers can copy safely. On his own laptop, 8 threads reading 1 file went from about 10 million reads a second to between 30 and 33 million, with some protections turned off for the test. The commit credits Claude Code, under his direction, with the code, the test programs, the review, and a first draft of the message. Torvalds is not claiming the model did the engineering. He directed it, ran the measurements himself, and takes responsibility for the result. He has been clear that he still thinks the engineering matters and that a model cannot just do everything on its own.&lt;/p&gt;
&lt;h2 id=&quot;how-much-the-ground-has-already-shifted&quot;&gt;How much the ground has already shifted&lt;/h2&gt;
&lt;p&gt;Go back to 2023, about a year after ChatGPT landed, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://gamingonlinux.com/2026/07/linux-creator-linus-torvalds-puts-foot-down-on-anti-ai-comments&quot;&gt;Torvalds was far more reserved&lt;/a&gt;. In an on stage interview with Dirk Hohndel, he described large language models as autocorrect on steroids, software that predicts the next likely word and extrapolates, not very intelligent in his view. He expected LLM written code would show up in kernel submissions, maybe already was on a small scale, and he was not especially worried, because automation had always helped people write code. His hope was modest, that AI might catch the obvious stupid bugs, the ones that do not need any higher intelligence to spot, and flag patterns that look off so a human can say that was not what I meant. Asked about hallucinations, he said he saw the bugs humans produce every day, so he was not losing sleep over the machine’s mistakes.&lt;/p&gt;
&lt;p&gt;The distance between that 2023 answer and the October 2026 commit is the real story. In 2023 the tech produced boilerplate and images with 6 fingers, and anyone could tell at a glance what was machine made. It did not plateau. By late 2026 the top level maintainer of the Linux kernel is shipping AI assisted performance work and defending the right to use the tools in public.&lt;/p&gt;
&lt;h2 id=&quot;the-kernel-is-already-an-ai-heavy-project&quot;&gt;The kernel is already an AI heavy project&lt;/h2&gt;
&lt;p&gt;If you are anti AI and you run Linux, the numbers are hard to get around. By the kernel’s own rough count, more than 1 in 10 recent commits carried an AI tag, and the honest reading is that the true share is higher and climbing every month. That is not 10% of code accepted into the kernel, it is a floor on the new commits being sent in with AI somewhere in the process. Even if every maintainer stopped using these tools today, the AI assisted code already merged is not coming back out.&lt;/p&gt;
&lt;p&gt;AI is also already doing review work, not just writing code. A tool called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Sashiko-Linux-AI-Metrics&quot;&gt;Sashiko&lt;/a&gt;, built by Google engineers for agentic review of kernel changes and running on Gemini and other models, gave its status update at the Linux Plumbers Conference in late 2026. In under a year it had run more than 191,000 patch reviews across 99 mailing lists, made over 19 million autonomous Git lookups, and drawn more than 7,600 replies from over 1,100 kernel developers. This year alone, 1,277 commits merged into the mainline kernel cite Sashiko, and 463 Linux kernel CVEs, security vulnerabilities, cite it as well. A machine is reading kernel patches at a scale no human team could match and catching real security bugs in the process.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.BphSQywo_1D6TPJ.webp&quot; alt=&quot;Figure 3. Sashiko, an AI code review tool built by Google engineers, in under 1 year on the kernel mailing lists. More than 191,000 patch reviews, over 19 million automated lookups, and 463 kernel CVEs this year that cite it.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;936&quot; height=&quot;519&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. Sashiko, an AI code review tool built by Google engineers, in under 1 year on the kernel mailing lists. More than 191,000 patch reviews, over 19 million automated lookups, and 463 kernel CVEs this year that cite it. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Sashiko-Linux-AI-Metrics&quot;&gt;Author chart based on Linux Plumbers Conference metrics, 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Some subsystems will say no AI, we do not want it near our code, and the leadership has made clear that is fine, a maintainer can set that rule for their own corner. The project as a whole is going the other way. Torvalds controls the top of the tree, his position is not ambiguous, and as long as he runs the kernel you will see AI used in it. The leadership has effectively decided the benefits outweigh the costs, and that the slop a model produces is no worse than the slop people produce on their own.&lt;/p&gt;
&lt;h2 id=&quot;what-this-actually-changes&quot;&gt;What this actually changes&lt;/h2&gt;
&lt;p&gt;Dropping the Assisted-by tag does not make the kernel more or less AI driven than it already is. It removes a signal that was never reliable and replaces it, for now, with an informal convention of explaining AI use in the cover letter of a patch series, outside the permanent commit record. That is less visible and less countable, which is exactly what Pinchart was unhappy about. Responsibility for the code still sits with the human who submits it, and an AI agent still cannot sign off on a patch in its own name.&lt;/p&gt;
&lt;p&gt;The real thing to watch is whether the kernel keeps any durable way to measure how much of itself is being written with AI, because the one experiment it ran, voluntary tagging, just got called a failure by the person who proposed it. If you contribute to the kernel, read the subsystem’s own stance before you send anything a model helped you write, because the top of the tree and an individual maintainer can want very different things.&lt;/p&gt;
&lt;h2 id=&quot;sources-and-further-reading&quot;&gt;Sources and further reading&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-Dropping-Assisted-By-Req&quot;&gt;Linux kernel developers plan to remove the Assisted-by tag requirement, Phoronix&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thenextweb.com/news/torvalds-ai-assisted-by-tag-utterly-stupid-open-source-summit&quot;&gt;Torvalds calls AI disclosure rules utterly stupid as Linux drops its tag, The Next Web&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://gamingonlinux.com/2026/07/linux-creator-linus-torvalds-puts-foot-down-on-anti-ai-comments&quot;&gt;Linus Torvalds puts his foot down on anti AI comments, GamingOnLinux&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/linus-torvalds-claude-code-linux/&quot;&gt;Linus Torvalds says he likes using AI while his new kernel branch credits Claude Code, It’s FOSS News&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Sashiko-Linux-AI-Metrics&quot;&gt;Google’s Sashiko AI has completed 191k patch reviews and is cited on hundreds of kernel CVEs, Phoronix&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;Linux kernel source tree gets official AI coding assistant policy, Lilting&lt;/li&gt;
&lt;/ul&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>linux</category><category>open-source</category><category>ai</category><category>software-engineering</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Linux 7.3 Made BTRFS Up to 5.7x Faster, and It Did It by Deleting Code. Here Is How</title><link>https://soursen.org/articles/linux-7-3-made-btrfs-up-to-5-7x-faster-and-it-did-it-by-deleting-code/</link><guid isPermaLink="true">https://soursen.org/articles/linux-7-3-made-btrfs-up-to-5-7x-faster-and-it-did-it-by-deleting-code/</guid><description>The headline on Linux 7.3 is a filesystem that got dramatically quicker, and almost every gain came from stripping out a bottleneck that was quietly costing more than it saved. I went through the actual patches, the fio numbers, and the maintainers behind them to explain what changed and why it matters for anyone running Linux.</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Most kernel releases are a long list of small things. Linux 7.3 has a theme, and that theme is speed won by taking code out. BTRFS, the copy on write filesystem a lot of people still write off as slow, picked up performance gains between about 3x and 5.7x on specific workloads, and the pattern behind them is almost always the same. A safety mechanism was doing far more work than the situation needed, and a developer replaced it with something lighter. I want to walk through the real changes, with the numbers from the patches themselves, because this release is a clean lesson in where filesystem time actually goes.&lt;/p&gt;
&lt;h2 id=&quot;the-57x-win-came-from-deleting-a-1-millisecond-nap&quot;&gt;The 5.7x win came from deleting a 1 millisecond nap&lt;/h2&gt;
&lt;p&gt;Start with the single most dramatic number. BTRFS uses something called a log tree to make fsync fast. When a program calls fsync, it is asking the system to guarantee that a file is safely on disk before continuing, and the log tree lets BTRFS promise that without writing out a full transaction every time.&lt;/p&gt;
&lt;p&gt;On spinning disks, BTRFS used to pause for 1 jiffy before committing the log. A jiffy is the kernel’s basic timer tick, usually somewhere between 1 and 10 milliseconds depending on configuration. The idea was reasonable on paper. If several tasks were logging at once, waiting a moment let the system bundle their writes together, which is kinder to a hard drive that hates lots of tiny requests. The problem is that the wait itself became the bottleneck. 10 milliseconds is an eternity next to everything else in that path, and when multiple tasks were logging, that forced nap turned concurrent work into a single file queue.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lore.kernel.org/linux-btrfs/&quot;&gt;Filipe Manana&lt;/a&gt;, the BTRFS developer who did most of the heavy lifting this release, stripped the sleep out entirely. The result, measured with a fio random write test doing fsync on every write, went from 2,647 KiB per second to 14.9 MiB per second. That is about 5.7x faster, and all it took was removing the code that told the system to wait.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CHxVwBHF_2avnB7.webp&quot; alt=&quot;Figure 1. Removing a 1 jiffy sleep from the BTRFS log commit path took a fio fsync workload from 2,647 KiB/s to 14.9 MiB/s, about 5.7x faster.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1112&quot; height=&quot;576&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. Removing a 1 jiffy sleep from the BTRFS log commit path took a fio fsync workload from 2,647 KiB/s to 14.9 MiB/s, about 5.7x faster. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lore.kernel.org/linux-btrfs/&quot;&gt;Filipe Manana BTRFS patch set, Linux 7.3&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;skipping-work-you-can-prove-is-unnecessary&quot;&gt;Skipping work you can prove is unnecessary&lt;/h2&gt;
&lt;p&gt;The second BTRFS win follows the same logic. When BTRFS runs a full fsync, it checks for holes in the file. A hole is a gap you create when you skip ahead in a file before writing, and the log tree has to know a hole exists, because without that record a crash recovery would read whatever random bytes sit underneath and treat them as real data.&lt;/p&gt;
&lt;p&gt;Detecting holes means scanning every extent in the file to confirm there are no gaps. On a small file that is nothing. On a heavily fragmented 10 GB file that might hold 20,000 extents, scanning all of them on every fsync is brutal, and for most files it is wasted effort because the file has no holes at all. The fix is a short piece of math that proves a file cannot have a hole without walking its extents. If the no holes option is on, preallocation is off, and the logical size of the file is less than or equal to the bytes actually allocated on disk, there is no room for a hole to hide, so the scan gets skipped. On fragmented files, that change produced roughly a 5x speedup, again by doing less.&lt;/p&gt;
&lt;p&gt;Direct IO got the same treatment. BTRFS now uses an iomap bounce buffer when a direct IO request is misaligned, correcting it in a small buffer in place, where before it would give up and fall back to the slow buffered IO path through the page cache. That one change moved direct IO performance from about 50% of the theoretical limit to about 95%.&lt;/p&gt;
&lt;p&gt;There is a fourth BTRFS change in the same spirit, and it is the one that fills out the low end of that 3x range. Copy on write filesystems have a habit of rewriting metadata when a write lands out of order, a problem called cow amplification, and an earlier release used an xarray to track in progress metadata buffers and prevent it. Maintaining that xarray turned out to be heavier than the problem it solved. Replacing it with a small local LRU list, a lighter structure holding a handful of buffers, restored performance to its pre fix level while still preventing the amplification, worth about 3x on the affected path. 7.3 also flips the BTRFS free space cache to version 2 by default, retiring the older v1 format.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.BWi4rVkO_1IJrcU.webp&quot; alt=&quot;Figure 2. 4 BTRFS changes, 4 speedups, all from removing or avoiding work. Log commit sleep removal, fsync hole detection skipping, a lighter extent buffer tracker, and direct IO via a bounce buffer.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1120&quot; height=&quot;591&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. 4 BTRFS changes, 4 speedups, all from removing or avoiding work. Log commit sleep removal, fsync hole detection skipping, a lighter extent buffer tracker, and direct IO via a bounce buffer. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-7.3-Btrfs&quot;&gt;Linux 7.3 BTRFS changes, Phoronix and patch data&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-memory-allocator-stopped-making-cores-fight&quot;&gt;The memory allocator stopped making cores fight&lt;/h2&gt;
&lt;p&gt;Away from filesystems, 2 memory management changes landed that matter for anyone using compressed swap, which is to say most modern desktops and plenty of servers running ZRAM or zswap. Both are about lock contention, the situation where many CPU cores stall waiting for the same lock.&lt;/p&gt;
&lt;p&gt;The first touches zsmalloc, the allocator behind ZRAM and zswap. Freeing compressed memory used to require grabbing a pool wide lock just to look up which size class a chunk belonged to, and that pool is shared across the whole system for zswap. Under memory pressure, every core trying to free memory would pile onto that one lock. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-7.3-MM&quot;&gt;Wenchao Hao of Xiaomi&lt;/a&gt; added the size class directly to the object so the lookup no longer needs the lock. On a Raspberry Pi 4B the whole page out and free sequence ran up to 1.83x faster, with a smaller gain on a 20 core Intel box.&lt;/p&gt;
&lt;p&gt;The second is kernel samepage merging, or KSM, the feature that scans memory, finds identical pages, and merges them to save RAM. On a database server that carves its memory into tens of thousands of small regions, KSM’s reverse mapping walk was treating every region as a candidate for a single target page, which meant huge scans. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lwn.net/Articles/1074053/&quot;&gt;Xu Xin of ZTE&lt;/a&gt; passed a precomputed page index into that walk so it jumps straight to the right candidate using an existing interval tree. In a test with 20,000 regions sharing 1 structure, the lock hold time dropped from 705 milliseconds to about 1.44 milliseconds. Half a second of freezing, gone.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.BaknHXVx_Z2ikvPC.webp&quot; alt=&quot;Figure 3. The KSM reverse mapping fix cut the anon_vma lock hold time from 705 milliseconds to about 1.44 milliseconds in a 20,000 region test, ending multi hundred millisecond stalls.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1091&quot; height=&quot;576&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. The KSM reverse mapping fix cut the anon_vma lock hold time from 705 milliseconds to about 1.44 milliseconds in a 20,000 region test, ending multi hundred millisecond stalls. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lwn.net/Articles/1074053/&quot;&gt;Linux 7.3 memory management, via LWN&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;letting-a-core-walk-away-from-an-interrupt&quot;&gt;Letting a core walk away from an interrupt&lt;/h2&gt;
&lt;p&gt;There is a matching story in how cores coordinate. When the kernel needs every core to do something, like flushing a translation lookaside buffer on x86, it sends an interprocessor interrupt, an IPI, and waits for the target cores to finish. The core that sent the IPI could not be preempted while it waited, because the list tracking which cores were done was a single shared buffer that had to be protected.&lt;/p&gt;
&lt;p&gt;So if core 1 waited on core 2, and core 2 was itself stuck waiting on another IPI, you could end up with a chain where the whole machine stalled on 1 core. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-7.3-SMP&quot;&gt;ByteDance engineers&lt;/a&gt; gave each task its own IPI tracking mask, replacing the single shared one per core, which removes the need to block preemption during the wait. In a benchmark measuring how fast a high priority task can start running after an IPI, latency fell from 17 milliseconds to 1.5 milliseconds, roughly a 90% reduction. That is the kind of number that matters for audio, for real time control, and for anything that cannot tolerate a scheduler hiccup.&lt;/p&gt;
&lt;h2 id=&quot;a-filesystem-whose-only-job-is-to-fail&quot;&gt;A filesystem whose only job is to fail&lt;/h2&gt;
&lt;p&gt;Not every change in 7.3 is a benchmark. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://kernelnewbies.org/Linux_7.3&quot;&gt;Christian Brauner&lt;/a&gt;, who maintains the virtual filesystem layer, shipped a new pseudo filesystem called failfs, and it does exactly 1 thing. It fails. Ask for its parent directory, fail. Try to access the root, fail. Try to mount it, clone it, or run a binary inside it, every one of those returns operation not supported.&lt;/p&gt;
&lt;p&gt;The point is the strongest sandbox possible. Containers restrict what a process can reach, but until now there was no way to fully remove storage from a container. failfs, paired with a new fchroot system call that chroots into a file descriptor, lets you drop a process into a void where the only path to the outside world is a file descriptor you explicitly hand it. If a process cannot even find the filesystem, there is nothing there to exploit. The name makes it sound like a gag, but it gives containers a way to hand a process a storage environment with nothing in it at all.&lt;/p&gt;
&lt;h2 id=&quot;a-cleaner-way-to-write-without-filling-the-cache&quot;&gt;A cleaner way to write without filling the cache&lt;/h2&gt;
&lt;p&gt;Databases have a long running fight with the kernel’s page cache. The page cache keeps recently touched data in memory to speed up future reads, which is great until your program manages its own cache better than the kernel can, at which point the page cache is just wasting RAM on data the program will never reread. There is a per IO flag, RWF_DONTCACHE, that says write this through but do not keep it around, and 7.3 extends it to raw block devices, where before it only worked on files. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://phoronix.com/news/Linux-7.3-Block&quot;&gt;Tal Zussman&lt;/a&gt; enabled it, which helps databases like the ones that talk straight to a block device with no filesystem in between.&lt;/p&gt;
&lt;p&gt;Getting there surfaced a problem that kept showing up across the kernel. Dropping a page from the cache requires a lock, and acquiring that lock can put the task to sleep, but the completion of a disk write runs in interrupt context, where sleeping is forbidden. Different subsystems had each invented their own workaround, pushing the cleanup onto a separate work queue. 7.3 adds a shared solution, a flag on the block layer’s basic unit of work that defers the completion to normal task context when needed. Once it existed, it got adopted everywhere the same pattern appeared, in zswap writeback, in F2FS, in XFS, in the iomap layer. One helper replaced a half dozen copies of the same careful hack.&lt;/p&gt;
&lt;h2 id=&quot;everything-else-worth-knowing&quot;&gt;Everything else worth knowing&lt;/h2&gt;
&lt;p&gt;A few more changes round out the release. EXT4 got a faster iomap based direct IO path for simple requests, which pushed a sample workload from 1.92 million to 2.19 million IOPS. The base M3 Apple Silicon support from 7.2 expanded to cover the M3 Pro, Max, and Ultra, which is early boot scaffolding and not yet daily driver support, with the Asahi Linux team still ahead of mainline for anyone actually running Linux on Apple hardware. On the GPU side, work landed to keep games from crashing when they need more VRAM than the card has, tested on Indiana Jones and the Great Circle running in 9 GB on an 8 GB GPU at a usable frame time.&lt;/p&gt;
&lt;p&gt;The release also deleted a pile of dead code, including a 3G PCMCIA card driver last mentioned on a forum in 2011, a readonly SGI filesystem unmaintained for over 20 years, and the Veritas filesystem that once let Linux read old Unix partitions. Less unmaintained code means a smaller attack surface and fewer drags on the subsystems that have to keep working around it. Andrew Morton, who maintains memory management, also started using AI to draft the plain English summaries at the top of his pull requests, with a human still reviewing every one, which is about the lowest risk place in the kernel to let a model help.&lt;/p&gt;
&lt;h2 id=&quot;the-release-itself-is-getting-bigger-and-ai-is-part-of-why&quot;&gt;The release itself is getting bigger, and AI is part of why&lt;/h2&gt;
&lt;p&gt;7.3 is also a data point about how the kernel is being built now. It landed on 18 October 2026 with 15,267 commits in the merge window, the second busiest in the project’s history, behind only 7.0. The codebase crossed roughly 40.98 million lines, up from 40.42 million in 7.2, and a single release candidate added around 560,000 lines, a large chunk of it AMD display headers.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.D3oq8S8G_Z1LMYcW.webp&quot; alt=&quot;Figure 4. Linux 7.3 shipped 15,267 commits, the second busiest merge window ever, and pushed the codebase to about 40.98 million lines.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1120&quot; height=&quot;583&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. Linux 7.3 shipped 15,267 commits, the second busiest merge window ever, and pushed the codebase to about 40.98 million lines. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.linuxcompatible.org/story/linux-kernel-73-released-record-commit-count-and-aidriven-growth&quot;&gt;Linux 7.3 release reporting&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Linus Torvalds has tied some of that growth directly to AI. LLM coding agents are finding bugs and helping write fixes, and the same tooling is speeding up the removal of old, unmaintained drivers, which is part of why 7.3 could delete so much dead code at once. Torvalds has said publicly that he uses Claude Code himself to make code more readable and to surface extra bugs. The community is now working out formal guidelines for AI assisted contributions, including a 2026 proposal from Sasha Levin of NVIDIA, and there is active debate about dropping the tag that currently marks AI generated patches. Andrew Morton, who maintains memory management, already uses a model to draft the plain English summaries at the top of his pull requests, with a human reviewing every one.&lt;/p&gt;
&lt;h2 id=&quot;what-all-of-this-adds-up-to&quot;&gt;What all of this adds up to&lt;/h2&gt;
&lt;p&gt;The biggest wins in 7.3 came from developers like Filipe Manana looking at a cautious old mechanism, proving it was doing more than the job required, and taking it out. That is where a lot of filesystem and kernel performance actually hides, in the safety code written for a worst case that rarely happens. The release that found all of it is also the second largest the kernel has ever shipped, built by more people and more tooling than any before it, which is worth watching as its own story over the next few cycles.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>btrfs</category><category>linux</category><category>kernal</category><category>ai</category><category>devlopment</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>systemd Wants to Know Which App Is Talking to It. That Is a Bigger Deal Than It Sounds</title><link>https://soursen.org/articles/systemd-wants-to-know-which-app-is-talking-to-it-that-is-a-bigger-deal/</link><guid isPermaLink="true">https://soursen.org/articles/systemd-wants-to-know-which-app-is-talking-to-it-that-is-a-bigger-deal/</guid><description>A proposed component called systemd-appd would give the Linux desktop a shared way to identify running apps and check their permissions, the thing Android and macOS have had for years. I went through the actual proposal and the debate around it, because the technical fix is reasonable and the question of where it lives is the part people are actually fighting about.</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;The Linux desktop has never had a clean answer to a simple question. When an app asks a system service for your camera, your microphone, or your location, which app is actually asking, and is it allowed to. A new systemd component called systemd-appd is a proposal to fix exactly that, and the reaction to it says as much about how Linux is governed as about the code itself. I want to walk through what it does, the real mechanism underneath it, and why a sensible piece of plumbing turned into an argument.&lt;/p&gt;
&lt;h2 id=&quot;the-problem-is-that-linux-cannot-reliably-tell-who-is-asking&quot;&gt;The problem is that Linux cannot reliably tell who is asking&lt;/h2&gt;
&lt;p&gt;Picture a sandboxed app, packaged as a Flatpak, asking PipeWire to access your microphone. The service on the other end needs to know 2 things. Which process is this, exactly, and which application does it represent. Linux has been surprisingly bad at both.&lt;/p&gt;
&lt;p&gt;The classic way to identify the process on the other end of a socket was to ask the kernel for its process ID. The trouble is that a process ID can be recycled. The process you were talking to exits, the number gets handed to something else, and now your security check is pointing at the wrong program. That is a real race condition that has bitten real software. The modern fix is a mechanism called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lore.kernel.org/all/ZDoEG0VF6fb9y0EC@google.com/T/&quot;&gt;SO_PEERPIDFD&lt;/a&gt;, which hands a service a pidfd, a stable handle that refers to one specific process and cannot be confused with a later one that happens to reuse the number. Services like polkit and the system journal are the exact cases it was built for.&lt;/p&gt;
&lt;p&gt;Knowing the process is only half of it. You still have to map that process back to an application identity, which sandbox launched it, what it is called, and what it is permitted to do. Flatpak, the packaging format that bundles an app with its libraries so the same build runs on any distribution, makes this harder because the sandbox deliberately hides the app from the services it talks to. That gap is what systemd-appd is meant to close.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CitWv-8W_bbRq2.webp&quot; alt=&quot;Figure 1. A sandboxed app asks a service for the microphone. Without a shared identity service, the service cannot reliably tell which app is asking or what it is allowed to do.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;744&quot; height=&quot;325&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. A sandboxed app asks a service for the microphone. Without a shared identity service, the service cannot reliably tell which app is asking or what it is allowed to do. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://github.com/systemd/systemd/pull/43885&quot;&gt;Author diagram based on the systemd-appd proposal&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-systemd-appd-actually-proposes&quot;&gt;What systemd-appd actually proposes&lt;/h2&gt;
&lt;p&gt;The component was drafted as &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://github.com/systemd/systemd/pull/43885&quot;&gt;a systemd pull request&lt;/a&gt; by Adrian Vovk, one of the developers working on the next generation of Flatpak, and it had been discussed openly for about a year before that. The systemd project describes it plainly. systemd-appd is a centralized place to keep track of the user’s apps, and it lets session services like D-Bus, PipeWire, portals, and Wayland compositors reliably identify apps and learn about their current permissions.&lt;/p&gt;
&lt;p&gt;The flow is straightforward. An app registers when it starts, systemd-appd records its identity and its permissions, and any service that needs to make a security decision can query appd over a well defined interface and skip the guesswork. When a request lands, the service can ask appd which application this is, then check whether that application has been granted the camera or the microphone, and hand the user a clear choice. The permission information becomes something the whole desktop shares, so each service no longer reconstructs it on its own.&lt;/p&gt;
&lt;p&gt;Vovk has been clear that appd is tightly coupled to systemd by design, since it leans on systemd’s own tracking of the user’s service hierarchy and talks to it over internal IPC, which is the argument for shipping it inside systemd and not as a separate project. It is also described as a prerequisite for 3 things the Flatpak developers want next, nested sandboxing where a sandboxed app can safely spawn another sandbox, proper PipeWire support, and finally removing the old D-Bus proxy that Flatpak has leaned on for years.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.BKHpB-Io_Z1VtCkq.webp&quot; alt=&quot;Figure 2. How systemd-appd fits in. The app registers once, and every session service queries one identity and permission authority with no guesswork of its own.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;744&quot; height=&quot;395&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. How systemd-appd fits in. The app registers once, and every session service queries one identity and permission authority with no guesswork of its own. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://github.com/systemd/systemd/pull/43885&quot;&gt;Author diagram based on systemd PR 43885&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;how-the-identity-actually-attaches&quot;&gt;How the identity actually attaches&lt;/h2&gt;
&lt;p&gt;The mechanism underneath is worth seeing, because it is where the systemd coupling comes from. When an app launches, appd places it in a systemd unit that follows the XDG naming convention for applications, the standard scheme that gives an app a stable reverse domain name like org.mozilla.firefox. Putting the app in that unit does more than organize it. The app’s identity gets tagged onto the control group, the cgroup, that contains its processes, so the identity travels with the process tree, where a separate lookup table could drift out of sync.&lt;/p&gt;
&lt;p&gt;Services then talk to appd over a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/systemd-appd&quot;&gt;Varlink interface&lt;/a&gt;, a simple request and reply protocol systemd already uses for its internal APIs. A service holds a pidfd for the process that contacted it, hands that to appd, and gets back the application identity and its current permissions. The reason Adrian Vovk argues appd belongs inside systemd is exactly this chain. appd leans on systemd’s own tracking of your service hierarchy and its cgroup machinery, and a component that reaches that deep into systemd internals is awkward to ship anywhere else.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.CmjzEw-o_ZL0IBi.webp&quot; alt=&quot;Figure 3. How the identity attaches. The app launches into a systemd unit, its identity is tagged onto the cgroup, and any service hands appd a pidfd over Varlink to learn who the app is and what it may do.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;734&quot; height=&quot;400&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. How the identity attaches. The app launches into a systemd unit, its identity is tagged onto the cgroup, and any service hands appd a pidfd over Varlink to learn who the app is and what it may do. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/systemd-appd&quot;&gt;Author diagram based on systemd PR 43885 and Phoronix&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;One detail matters for the panic that followed the announcement. Vovk has said appd is not meant to become a hard requirement for D-Bus or PipeWire. Those services would query appd for permissions when it is present and enforce them, and keep working when it is absent. The dependency is softer than the headlines made it sound, though a non systemd distribution that wanted the same features would still need to provide a drop in replacement exposing the same interface.&lt;/p&gt;
&lt;h2 id=&quot;this-is-the-model-android-and-macos-already-have&quot;&gt;This is the model Android and macOS already have&lt;/h2&gt;
&lt;p&gt;None of this is new as an idea. It is the app permission model that mobile and macOS users see constantly. On macOS you get the prompt asking whether Chromium can find devices on your local network, and you can tap deny and watch the browser keep working fine, because the request was optional and the system gave you the choice. That machinery has a name, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://eclecticlight.co/2025/11/08/explainer-permissions-privacy-and-tcc/&quot;&gt;TCC&lt;/a&gt;, short for Transparency, Consent, and Control, and it traces each request up the call chain to the app actually responsible for it before deciding.&lt;/p&gt;
&lt;p&gt;Android is the cleaner parallel, and it is worth seeing exactly how it works, because it solves the same problem systemd-appd is chasing. When you install an Android app, the system hands it its own Linux user ID. The kernel then enforces isolation by that ID, so one app cannot touch another’s files, and every granted permission is recorded in a central store that services read. When an app asks for your location, the location service asks the permission service whether that app’s user ID holds the location permission, a stable check that a recyclable process ID could never give. Checking a stable identity is the whole point, because a process number can be recycled under you, and it is the same move systemd-appd is making on the desktop. Windows has its own version of this for packaged apps.&lt;/p&gt;
&lt;p&gt;Linux has had pieces of this for a while, portals that broker access to files and devices, polkit for privileged actions, the sandbox boundaries that Flatpak draws, but it has lacked a single shared notion of application identity that all of those pieces agree on. Each guard at each layer has its own way of working out who you are. systemd-appd is an attempt to give them a common answer, so a permission dialog on Linux can finally say this request comes from this app, and this app is or is not allowed to do that.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.BIbHDSLy_2aIV5o.webp&quot; alt=&quot;Figure 4. Android, macOS, and Windows each give apps a strong identity that system services can check. Linux has the pieces but no shared identity authority, which is the gap systemd-appd targets.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;744&quot; height=&quot;345&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. Android, macOS, and Windows each give apps a strong identity that system services can check. Linux has the pieces but no shared identity authority, which is the gap systemd-appd targets. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/systemd-appd&quot;&gt;Author comparison of platform app identity models&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-spying-fear-misses-the-real-objection&quot;&gt;The spying fear misses the real objection&lt;/h2&gt;
&lt;p&gt;A chunk of the online reaction jumped straight to systemd is building a tracker to spy on you. I think that reading is wrong. appd stores which apps are running and what they are allowed to do so that your own desktop can enforce permissions, which is the opposite of surveillance, it is the machinery that lets you deny the microphone in the first place. Android and macOS keep the same kind of record, and nobody calls their permission manager spyware.&lt;/p&gt;
&lt;p&gt;The objection worth taking seriously is about centralization. systemd started life as an init system, the first process that boots your machine and starts your services. Over the years it has absorbed login session management, network configuration, DNS resolution, container management, and more. Each addition had a reasonable justification, and each one also made systemd harder to avoid. Adding an application identity and permission authority continues that pattern, moving one more piece of core desktop infrastructure inside the systemd tent.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.By2Se1Tp_Z1Ulth.webp&quot; alt=&quot;Figure 5. systemd began as an init system and has steadily absorbed login, networking, DNS, and container management. appd would add app identity and permissions to that list.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;744&quot; height=&quot;335&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 5. systemd began as an init system and has steadily absorbed login, networking, DNS, and container management. appd would add app identity and permissions to that list. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/systemd-appd&quot;&gt;Author timeline of systemd scope&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-part-that-actually-affects-distributions&quot;&gt;The part that actually affects distributions&lt;/h2&gt;
&lt;p&gt;This is the part that reaches real users. If Flatpak comes to depend on systemd-appd, and the Flatpak developers have said appd is a prerequisite for the features they are building, then distributions that do not use systemd have a problem. Alpine, Void, Guix, and Devuan all ship without it, and a cross distribution packaging format depending on 1 specific init system cuts against the entire promise of Flatpak, which is build once and run anywhere.&lt;/p&gt;
&lt;p&gt;There is precedent for how this goes. When systemd’s login manager, logind, became a requirement for modern desktops, non systemd distributions forked it into a standalone daemon called elogind just to keep their desktops working. A similar split could happen again, with those distributions maintaining a downstream shim, a replacement, or simply losing full Flatpak support. The counterargument is real too. systemd already manages the user’s service hierarchy and already has the cgroup machinery appd needs, most mainstream distributions already run it, and reinventing that foundation somewhere neutral would be a large amount of work for a problem many users will never notice.&lt;/p&gt;
&lt;h2 id=&quot;where-this-actually-stands&quot;&gt;Where this actually stands&lt;/h2&gt;
&lt;p&gt;systemd-appd is early. It is a draft proposal under review, current stable Flatpak does not require it, and no distribution ships it yet. What it represents is clearer than its final form. Linux is trying to grow a real application identity layer, the thing that lets a permission prompt mean something, and the most practical place to build it happens to be the component that already tracks your running services. That convenience is also the concern, because the more the desktop leans on systemd, the harder systemd becomes to replace. Both of those things are true, and the next few months of review will decide which one the Linux world weighs more heavily.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>systemd</category><category>linux</category><category>kernal</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>The Internet Runs on Code Maintained by People You Have Never Heard Of.</title><link>https://soursen.org/articles/the-internet-runs-on-code-maintained-by-people-you-have-never-heard-of/</link><guid isPermaLink="true">https://soursen.org/articles/the-internet-runs-on-code-maintained-by-people-you-have-never-heard-of/</guid><description>The phone in your hand depends on a time zone file one UCLA lecturer updates in his spare time, a compression library written by the man who ran a Mars rover, and a shell whose entire public history is 1 person’s commits. I went through the real names, dates, and funding numbers behind the software holding up modern computing, because the gap between how much we lean on these projects and how little we pay for them is worse than the famous cartoon makes it look.</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.Dl0U3WbJ_Z1DC9JB.webp&quot; alt=&quot;Illustration for the section “The Internet Runs on Code Maintained by People You Have Never Heard Of.”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;385&quot; height=&quot;489&quot;&gt;&lt;/p&gt;
&lt;p&gt;You have probably seen the XKCD comic numbered 2347. A towering, precarious stack of blocks labeled “all modern digital infrastructure,” with the whole thing balanced on 1 tiny block at the bottom, captioned “a project some random person in Nebraska has been thanklessly maintaining since 2003.” People love to edit it, usually getting it wrong by stacking company logos that do not actually depend on each other. The real point is a dependency chain. The giants everyone knows all rest on small pieces of code written and maintained by a handful of people, often 1 person, who almost nobody can name. I want to put real names and real numbers to that bottom block, because once you see them, the comic stops being funny.&lt;/p&gt;
&lt;p&gt;I went looking for how deep this goes, and it is deeper than the cartoon suggests. A 2024 study by the Linux Foundation and Harvard estimated that if the most widely used open source software had to be rebuilt from scratch, it would cost about $8.8 trillion. The people maintaining a lot of it are not on anyone’s payroll. The same body of research has found that a small number of projects carry an outsized share of the load, and that the median open source developer is doing it in their own time. So this is not a story about a few unlucky projects. It is the normal condition of the code under almost everything you touch.&lt;/p&gt;
&lt;h2 id=&quot;the-clock-on-your-phone-comes-from-1-text-file&quot;&gt;The clock on your phone comes from 1 text file&lt;/h2&gt;
&lt;p&gt;Anything on your phone that shows a time has to know where you are, your alarm, a boarding pass, a meeting in another country, the day your salary lands. Every Android phone, every iPhone, and most of the servers on the internet get their local time from the same place, a single file called the time zone database. It holds every clock rule a government has ever announced, including historical ones going back more than a century, because if you want to render a timestamp from 1895 correctly, the rule for 1895 has to be in there.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Tz_database&quot;&gt;Paul Eggert&lt;/a&gt;, a computer scientist at UCLA, keeps that file current, largely in his spare time, and has been its editor since 2005. When a government changes its rules, a maintainer has to type the new entry in, by hand, ahead of the deadline. Governments do this constantly and often on short notice, a few weeks of warning before daylight saving shifts or a region moves to a new offset, and every one of those changes has to land in the file before it takes effect or clocks in that region quietly go wrong. The file lives on GitHub with a short list of people who have ever contributed to it, and the day to day work comes down to Eggert and 1 collaborator, Tim Parenti. That is the team behind the clock on several billion devices.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.D7JrYt43_ZepQL0.webp&quot; alt=&quot;Figure 1. A few named individuals maintain software that ships on billions of devices. The time zone database, behind every phone clock, is kept current mostly by 2 people.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1120&quot; height=&quot;621&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. A few named individuals maintain software that ships on billions of devices. The time zone database, behind every phone clock, is kept current mostly by 2 people. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Tz_database&quot;&gt;Author chart based on project histories and GitHub data&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;It gets stranger. In 2011, an astrology software company called Astrolabe sued Eggert and Arthur David Olson, the researcher who had coordinated the database since 1986, claiming its historical data came from an atlas the company held a copyright on. For a short period the project’s mailing list and files went dark, and the people who maintain the world’s timekeeping data suddenly had to think about legal exposure. The &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.eff.org/cases/astrolabe-v-olson&quot;&gt;Electronic Frontier Foundation&lt;/a&gt; stepped in, IANA took over hosting the database, and Astrolabe dropped the suit and apologized in February 2012. A company that sells date and time software tried to sue the people who maintain the world’s date and time data, and for a few weeks the foundation of civil timekeeping on the internet was a legal target.&lt;/p&gt;
&lt;h2 id=&quot;the-backdoor-that-almost-shipped-to-every-linux-server&quot;&gt;The backdoor that almost shipped to every Linux server&lt;/h2&gt;
&lt;p&gt;The thin staffing is a security story as much as a funding one, and the clearest proof arrived in March 2024. XZ Utils is a compression tool installed on effectively every Linux system. For years its lead maintainer was &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/XZ_Utils_backdoor&quot;&gt;Lasse Collin&lt;/a&gt;, who ran it as an unpaid hobby project and had been open about his limited time and long running health struggles.&lt;/p&gt;
&lt;p&gt;Starting around 2021, a contributor using the name Jia Tan began sending useful fixes and slowly built trust over more than 2 years, while sock puppet accounts pressured Collin in public about how slowly the project was moving. By 2023 Jia Tan had maintainer level access. In February 2024 a carefully hidden backdoor went into the project, built to let an attacker run commands on any server that used the compromised version. It was caught almost by accident. Andres Freund, an engineer at Microsoft, noticed that SSH logins on a test system were using a little more processor time than they should, around half a second longer, chased the anomaly, and found the backdoor before the poisoned version reached most production systems. It was logged as CVE-2024–3094 and rated the maximum severity of 10 out of 10. The identity behind Jia Tan has never been established publicly, and the operation remains one of the most patient supply chain attacks on record. Had Freund not been curious about half a second of CPU time, the backdoor would likely have shipped into the stable releases of major Linux distributions and onto servers worldwide.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.CaWkLFHb_Z1R85Qn.webp&quot; alt=&quot;Figure 2. 3 incidents where thinly staffed core projects became global risks. Shellshock in Bash, Heartbleed in OpenSSL, and the XZ backdoor, with how long each flaw sat before discovery.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1120&quot; height=&quot;576&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. 3 incidents where thinly staffed core projects became global risks. Shellshock in Bash, Heartbleed in OpenSSL, and the XZ backdoor, with how long each flaw sat before discovery. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/XZ_Utils_backdoor&quot;&gt;Author timeline of open source security incidents&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-maintainers-barely-anyone-pays&quot;&gt;The maintainers barely anyone pays&lt;/h2&gt;
&lt;p&gt;Walk through the rest of the stack and the same shape keeps appearing. 1 or 2 people, enormous reach, very little money.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thestack.technology/core-js-maintainer-denis-pusharev-license-broke-angry/&quot;&gt;Denis Pushkarev&lt;/a&gt; maintains core-js, a compatibility library that runs on a large share of the most visited websites in the world, downloaded tens of millions of times a week. His own fundraising brought in about $57 a month, and payments got harder as financial firms stopped processing donations to Russia. When he added a plea for support to the library’s install message, millions of developers saw it, and a lot of them complained about the message while few actually helped. He also served about 10 months in prison in 2020 following a fatal road accident, a matter of public record that he has spoken about himself, and the project kept limping along on its tiny donations throughout.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.theregister.com/2026/02/03/sudo_maintainer_asks_for_help/&quot;&gt;Todd Miller&lt;/a&gt; has maintained sudo, the command that grants administrative access on nearly every Linux and macOS machine, since 1993. A corporate sponsor funded that work for over a decade until the arrangement ended in 2024, and by early 2026 Miller was publicly looking for a new sponsor to keep development going, with the work slowing to mostly bug fixes in the meantime. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Zlib&quot;&gt;Mark Adler&lt;/a&gt; cowrote zlib, the compression library inside PNG images, web pages, Git, Android, iPhones, and Chrome, back in 1995. His other job was managing NASA’s Spirit rover on its way to Mars. zlib has no sponsor page at all.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/SQLite&quot;&gt;Dr. Richard Hipp&lt;/a&gt; wrote SQLite in 2000 after a database on a US Navy destroyer kept failing whenever a separate server went down, so he built one that lives in a single file with no server at all. It is now in every Android phone, every iPhone, every Mac, every copy of Windows, every major browser, and countless apps and games, with the project estimating over 1 trillion databases in use. SQLite is open source but not open contribution, a deliberate choice where only the core team commits code, and the team funds itself by selling support through Hipp’s company. It is one of the few in this list that chose its own sustainable path on purpose.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.C7B3S1YQ_ZtaLuy.webp&quot; alt=&quot;Figure 3. Reach versus people. Each of these projects ships on billions of devices or across most of the web, and the regular work is done by a very small number of maintainers.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1121&quot; height=&quot;576&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. Reach versus people. Each of these projects ships on billions of devices or across most of the web, and the regular work is done by a very small number of maintainers. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/SQLite&quot;&gt;Author chart based on project histories&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-man-who-held-up-the-worlds-email-encryption&quot;&gt;The man who held up the world’s email encryption&lt;/h2&gt;
&lt;p&gt;In February 2015, ProPublica published a piece headlined that the world’s email encryption software relies on 1 guy who is going broke. The guy was &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.propublica.org/article/the-worlds-email-encryption-software-relies-on-one-guy-who-is-going-broke&quot;&gt;Werner Koch&lt;/a&gt;, a German developer who wrote GnuPG in 1997 and had kept it alive almost single handedly ever since. GnuPG is the encryption that secures email, software signatures, and a long list of Linux tools, and it is the implementation Edward Snowden used when leaking his documents. Koch had been making about $25,000 a year from it since 2001, a fraction of what he could earn in industry. German government funding dried up in 2010, and in 2012 he had to let his only programmer go. A fundraising drive in December 2014 had reached about $43,000 against a goal of $137,000.&lt;/p&gt;
&lt;p&gt;Then the article ran, and the money arrived. Koch hit his funding goal within days, the Linux Foundation’s Core Infrastructure Initiative gave him a $60,000 grant, and Facebook and Stripe each pledged $50,000 a year to the project. For almost 18 years the padlock on a huge amount of secure communication had rested on 1 man working himself close to broke, and it took a viral article to change that.&lt;/p&gt;
&lt;h2 id=&quot;25-years-from-the-bug-going-in-to-anyone-noticing&quot;&gt;25 years from the bug going in to anyone noticing&lt;/h2&gt;
&lt;p&gt;The bash shell runs on Linux servers and Macs everywhere, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Shellshock_%28software_bug%29&quot;&gt;Chet Ramey&lt;/a&gt; has maintained it since around 1990, alongside his job in the networking group at Case Western Reserve University. Every change in bash’s public history is his, including the official security fixes.&lt;/p&gt;
&lt;p&gt;In September 2014, Stéphane Chazelas reported a bug to Ramey that let anyone run commands on a server by sending it a specially shaped piece of text. It went public as Shellshock, under CVE-2014–6271. The vulnerable line had gone into bash on 5 August 1989. 25 years passed between the flaw entering the code and anyone reporting it, in software that sat on a huge share of the world’s servers the entire time, maintained by 1 person on the side of a university job. Within days of the disclosure, attackers were scanning the internet for vulnerable machines, and more flaws turned up in the same code as people finally looked closely at it.&lt;/p&gt;
&lt;p&gt;The pattern of famous flaws in underfunded core code is not new. In April 2014, a bug called Heartbleed in OpenSSL, the library that puts the padlock on a large share of the web, let attackers read passwords and private keys straight out of server memory across a significant share of secure sites. At the time, the OpenSSL Software Foundation was taking in roughly $2,000 a year in donations, and 1 person worked on it full time. It took a global security panic to change that. Within weeks the Linux Foundation’s Core Infrastructure Initiative gathered pledges from 13 major technology companies, $100,000 each per year for 3 years, a pool of almost $4 million, and OpenSSL finally got full time paid developers and a professional audit.&lt;/p&gt;
&lt;h2 id=&quot;the-weekend-16-volunteers-held-the-internet-together&quot;&gt;The weekend 16 volunteers held the internet together&lt;/h2&gt;
&lt;p&gt;In December 2021, a flaw turned up in Log4j, a logging library used in a staggering range of Java software, from enterprise servers to the game Minecraft. The bug, Log4Shell, let an attacker run code on a server by getting it to log a single crafted string, and it was rated the maximum severity. It was described at the time as one of the most serious vulnerabilities in years, present in hundreds of millions of devices.&lt;/p&gt;
&lt;p&gt;Log4j was maintained by the Apache Logging Services team, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.infoworld.com/article/3645131/how-developers-scrambled-to-secure-the-log4j-vulnerability.html&quot;&gt;16 volunteers&lt;/a&gt;, all unpaid, working in their spare time across nearly every time zone. When the exploit went public on 9 December, they spent the weekend scrambling to ship fixes while the rest of the world scrambled to patch. Gary Gregory, one of the maintainers, described working on about 4 hours of sleep. Christian Grobmeier, a former member of the team, said his old colleagues had families and other obligations and still spent the whole weekend on it. The first fix was not good enough, so they shipped another, then another, versions 2.15, 2.16, and 2.17 arriving over about 10 days. The security of a large slice of the internet’s servers ran, for that weekend, on a handful of exhausted volunteers and a private Slack channel.&lt;/p&gt;
&lt;h2 id=&quot;the-money-is-starting-to-arrive-unevenly&quot;&gt;The money is starting to arrive, unevenly&lt;/h2&gt;
&lt;p&gt;The picture is not uniformly bleak. Some of these projects have built real funding, and the difference almost always comes down to a maintainer treating the business side as part of the job.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://curl.se/&quot;&gt;Daniel Stenberg&lt;/a&gt; has run curl since 1996, the data transfer tool inside phones, cars, TVs, and shipped in Windows since 2018. He works on it full time because companies pay for commercial support, the project takes in a meaningful annual budget through its open collective, and Germany’s public fund for open source has paid for specific work on it. In his 2025 review, Stenberg noted that other contributors had now added more lines to curl than he has, something he had wanted for years, because a project that depends on 1 person is fragile no matter how healthy its funding looks today.&lt;/p&gt;
&lt;p&gt;Institutional money has started filling some of the gap. Germany’s Sovereign Tech Agency has funded dozens of open source projects since 2022, including a direct investment in Log4j after the scare. The Alpha Omega project, backed by the Open Source Security Foundation, handed out millions last year, much of it to security engineers at foundations like Python and Ruby. The city of Munich began paying a developer to work on Expat, an XML parser, in 2026. Nick Wellnhofer raised a low six figure sum over the decade he maintained libxml2, the XML library on Android phones, iPhones, and in Chrome, and when he stepped down in late 2025 a new maintainer picked it up within a day. These are real improvements, and they are nowhere near the scale of the dependence.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.CzMXl-Z7_1Iby3t.webp&quot; alt=&quot;Figure 4. The same move keeps saving these projects, money arriving only after a public crisis. OpenSSL, GnuPG, and Log4j all got sustained funding within weeks of a security event or a viral article, not before.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1210&quot; height=&quot;601&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. The same move keeps saving these projects, money arriving only after a public crisis. OpenSSL, GnuPG, and Log4j all got sustained funding within weeks of a security event or a viral article, not before. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Core_Infrastructure_Initiative&quot;&gt;Author chart based on CII, ProPublica, and Sovereign Tech reporting&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-you-can-actually-do-about-it&quot;&gt;What you can actually do about it&lt;/h2&gt;
&lt;p&gt;The honest summary is that a lot of the software your devices depend on is maintained by 1 or 2 people who are underpaid or not paid at all, and the fixes that exist are patchy and often arrive only after something breaks badly. Some people will tell you AI is going to maintain this code soon. Maybe it will, eventually. The projects are understaffed now, and the risks are present now, so betting the world’s infrastructure on a future that may or may not arrive is not a plan.&lt;/p&gt;
&lt;p&gt;If you build on open source, which is to say if you build anything, the useful move is small and specific. Pick 1 of the projects your own stack quietly depends on, the compression library, the XML parser, the shell, the time zone file, and find out who maintains it and whether they have a way to be funded. If your company uses it, sponsor it. A recurring sponsorship of a few hundred dollars a month, trivial against most engineering budgets, is life changing money for a solo maintainer. If you have time, become a contributor to a project that has 1 or 2 people doing the work, so there is a second set of hands if the first pair walks away. The Nebraska maintainer in the cartoon is real, there are dozens of them, and they have been holding this up for years while almost nobody noticed.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>open-source</category><category>software-development</category><category>security</category><category>infrastructure</category><category>supply-chain</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>The Linux Kernel Just Stopped Labeling Its AI Code. The Real Story Is Who Pays for the Compute.</title><link>https://soursen.org/articles/the-linux-kernel-just-stopped-labeling-its-ai-code-the-real-story-is/</link><guid isPermaLink="true">https://soursen.org/articles/the-linux-kernel-just-stopped-labeling-its-ai-code-the-real-story-is/</guid><description>At the October 2026 maintainers summit, the Linux project dropped its requirement to tag AI assisted patches. The maintainers had reasonable reasons. The harder question is what it means that the compute reviewing and writing kernel code is donated by the same companies betting everything on AI.</description><pubDate>Thu, 15 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;In October 2026, at the kernel maintainers summit in Prague, the Linux project decided to stop asking contributors to disclose when they used an AI tool to write code. The Assisted-by tag, introduced earlier in the year so that AI generated patches would be labeled, is being removed. I think that decision, and the fight around it, says more about where open source is heading than any benchmark does.&lt;/p&gt;
&lt;p&gt;I want to walk through what actually changed, the real reasons the maintainers gave, where the serious disagreement is, and the structural question underneath all of it, which is whether a project that leans on corporate funded AI compute can still call itself free.&lt;/p&gt;
&lt;h2 id=&quot;what-actually-got-decided&quot;&gt;What actually got decided&lt;/h2&gt;
&lt;p&gt;The facts first, because this story has been told in a lot of heated ways. In April 2026 the kernel began allowing AI assisted contributions on the condition that you tagged them, with something like Assisted-by, so the provenance was visible. At the Prague summit in October, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lwn.net/SubscriberLink/1099594/fba1c2268c3e2fb5/&quot;&gt;reported by LWN&lt;/a&gt; and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-Dropping-Assisted-By-Req&quot;&gt;Phoronix&lt;/a&gt;, the maintainers reached consensus to drop that requirement, and a patch removing it from the documentation went to the mailing list.&lt;/p&gt;
&lt;p&gt;The reasons the maintainers actually gave are mundane, not sinister. Jeff Layton, a Red Hat kernel developer, laid out 3. Plenty of contributors ignored the tag anyway. The tag doubled as free advertising for proprietary LLM vendors, and it was never clear why the project needed to collect the information in the first place. Linus Torvalds, for his part, put the emphasis on judgment. He said he expects people to do their best and to understand what they are submitting, and that understanding has very little to do with which tools they used. That is a narrower position than the headlines suggest.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CLFjRJPu_Z1N5YQb.webp&quot; alt=&quot;Figure 1. The kernel went from tagging AI code to dropping the tag in under a year.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;744&quot; height=&quot;311&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. The kernel went from tagging AI code to dropping the tag in under a year. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.theregister.com/ai-and-ml/2026/10/14/linux-kernel-devs-drop-obligatory-confession-of-ai-tool-usage/5303140&quot;&gt;LWN, Phoronix, The Register, 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-disagreement-is-real-and-it-is-not-settled&quot;&gt;The disagreement is real, and it is not settled&lt;/h2&gt;
&lt;p&gt;Not everyone agreed, and the dissent is the interesting part. Laurent Pinchart argued against removing the tag without some other disclosure mechanism, on the grounds that rules like this set expectations even when some people ignore them. Take away the only signal of how much AI is in the code, and you lose the ability to audit provenance later, which matters for a codebase that runs on billions of devices.&lt;/p&gt;
&lt;p&gt;There is a licensing angle underneath that worry too. AI models are trained on large bodies of code under various licenses, and the kernel is GPL. If a model reproduces something it learned from incompatible code, a labeled patch at least gives you a thread to pull. Remove the label and that thread is gone. I do not think anyone has a clean answer here, which is exactly why some maintainers wanted to keep the signal.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.N39FGLjZ_ZfIjGK.webp&quot; alt=&quot;Figure 2. The 2 sides of the tag debate, both with real points.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;744&quot; height=&quot;348&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. The 2 sides of the tag debate, both with real points. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lwn.net/SubscriberLink/1099594/fba1c2268c3e2fb5/&quot;&gt;LWN summit coverage, 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-networking-subsystem-is-the-canary&quot;&gt;The networking subsystem is the canary&lt;/h2&gt;
&lt;p&gt;The clearest picture of the strain is in networking, and it comes straight from a maintainer’s own words. In the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-7.3-Networking&quot;&gt;Linux 7.3 networking pull request&lt;/a&gt; in August 2026, subsystem maintainer Jakub Kicinski wrote that the team was “completely overwhelmed, of course,” by the flood of patches. His stated fix was more automation, not less. The message thanks Meta for securing enough LLM budget and access to run reviews with multiple frontier models on each patch, which he said helps eliminate some hallucinations, and floats letting the models handle more of the busy work, including applying patches that already carry trusted reviewers’ tags.&lt;/p&gt;
&lt;p&gt;Sit with that for a second. A core Linux subsystem is being flooded with machine generated patches, and the maintainers’ answer is to point more machines at the problem, paid for by one of the largest AI companies in the world. In a later cycle the networking tree &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-7.3-rc7-Networking&quot;&gt;diverted more than 50 fixes&lt;/a&gt; to the next kernel while still being swamped by AI fueled churn. Whatever you think of AI coding, the volume problem is real and documented.&lt;/p&gt;
&lt;h2 id=&quot;the-structural-question-underneath-it&quot;&gt;The structural question underneath it&lt;/h2&gt;
&lt;p&gt;Here is the part that actually worries me, and it has nothing to do with any one person’s opinion of chatbots. The Linux Foundation is funded overwhelmingly by large corporations. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.linuxfoundation.org/press/google-renews-platinum-membership-with-the-linux-foundation-continuing-its-ongoing-support-for-the-open-source-community&quot;&gt;Platinum members&lt;/a&gt;, which include Google, Meta, and Microsoft among others, pay around $500,000 a year each, and the Foundation’s annual revenue crossed $300 million in 2026. Those same companies have enormous commercial stakes in AI being everywhere.&lt;/p&gt;
&lt;p&gt;When the compute that reviews and increasingly writes kernel code is donated by Meta, and the organization stewarding the project is funded by the companies with the most to gain from AI adoption, the word free starts to strain. Big corporate backers fund the Foundation, those backers have a vested interest in AI, and a major project that stayed AI skeptical would be a problem for them. That is just how the incentives line up, no conspiracy required, and it is worth saying out loud.&lt;/p&gt;
&lt;p&gt;There is a counterweight worth stating fairly. Corporate money has funded Linux kernel development for 2 decades, and the kernel got better, not worse, over that time. Paid developers from Red Hat, Intel, Google, and others do a large share of the serious work. Corporate funding is not new, and it is not automatically capture. The new thing is that the funding now buys compute that produces code directly, which shortens the distance between a sponsor’s interest and what lands in the tree.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.Cxn9x4eo_UzMv4.webp&quot; alt=&quot;Figure 3. Who funds the Linux Foundation, and the disputed question of how much reaches the kernel.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1069&quot; height=&quot;598&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. Who funds the Linux Foundation, and the disputed question of how much reaches the kernel. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://techcrunch.com/2018/06/27/google-ups-its-linux-foundation-membership-to-the-500000-year-platinum-level/&quot;&gt;Linux Foundation, TechCrunch, 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-exits-and-why-none-of-them-are-clean&quot;&gt;The exits, and why none of them are clean&lt;/h2&gt;
&lt;p&gt;So where do you go if you do not like this? The honest answer is nowhere comfortable. You can freeze on a long term support kernel and buy a few years, which is the easiest move and what a lot of people are doing. You can switch to the BSDs. NetBSD has a reputation for rejecting AI slop, but it supports far less hardware, gaming through Wine is hit and miss, and the Nvidia drivers are rough, roughly where Linux was around the year 2000. FreeBSD is friendlier day to day but more open to AI assisted contribution than NetBSD.&lt;/p&gt;
&lt;p&gt;Forking the kernel sounds clean and is not. Copying the code is the easy 1%. The hard 99% is rebuilding the institutions around it, the maintainers, the review process, the testing infrastructure, the trust. Anyone who says just fork it is skipping the part that actually matters. Distributions already carry their own patch sets and lag mainline for stability, so in practice the pressure valve is more likely to be big distros quietly diverging than a clean community fork.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.4yP418-2_14MyNm.webp&quot; alt=&quot;Figure 4. The options people are weighing, none of them clean.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;744&quot; height=&quot;354&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. The options people are weighing, none of them clean. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.phoronix.com/news/Linux-Dropping-Assisted-By-Req&quot;&gt;Author summary of community discussion, 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-i-would-actually-watch&quot;&gt;What I would actually watch&lt;/h2&gt;
&lt;p&gt;Strip away the heat and a few concrete things will tell you how this goes. Watch whether the share of AI assisted commits keeps climbing and whether anyone outside the project can still measure it now that the tag is gone. Watch whether a serious kernel regression gets traced back to a machine written patch that no human fully understood, which is the failure mode the skeptics keep naming. Watch whether a major distribution forks or freezes in protest, because that is the signal that the disagreement moved from the mailing list to the shipping product.&lt;/p&gt;
&lt;p&gt;Linux is still the least bad way off Windows for most people, and the kernel is not going to fall apart next week. My concern is slower than that. A project defined for decades by humans understanding every corner of their code is handing more of that understanding to tools funded by companies with their own agenda, and it just removed the one label that let outsiders see how much. None of this is a reason to panic. It is a reason to pay attention, and to keep asking who pays for the compute.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Linux &amp; open source</category><category>linux</category><category>ai</category><category>artificial-intelligence</category><category>google</category><category>meta</category><author>hello@soursen.org (Shivashish Yadav)</author></item></channel></rss>