<- all tokdocs
2026.10.21@ultradotdev

Torvalds's assembler jab answered a proposal to hand-write commit messages, not a campaign to ban AI-written code, and the rule it attacked had already been deleted

AI
Watch on TikTok

View on TikTok ->

A 58.792993 s vertical clip at 1080x1920 (aspect 0.56), container mov,mp4,m4a,3gp,3g2,mj2, video stream HEVC at a constant 30/1 fps and 431,735 bps, audio stream AAC at 48,151 bps, overall 486,513 bps, 3,575,448 bytes on disk, which matches the tbr of 486 and filesize of 3575448 recorded in metadata.json. Posted at epoch 1792552799, which is 2026-10-21T03:19:59Z, a Wednesday, by uploader ultradotdev (uploader_id 7679942878084826126, nickname "Ultra.dev"), credited to original sound - Ultra.dev, with a 337-character description. At capture time the sync recorded 9,045 views, 240 likes, 41 comments and 10 reposts; index.json stamps processed_at 2026-10-21T23:02:42Z. I extracted 29 frames at fps=0.5 into tokdoc_frames/ and read 10 of them (001, 003, 005, 009, 011, 015, 021, 025, 027, 029) against a 194-word transcript in 12 segments. The frames open on a scrolling screenshot of a mailing-list message, headers legible across frames 001 and 003: From: Linus Torvalds, Date: Tue Oct 13 2026 - 16:55:30 EST, Subject: Re: [PATCH] docs: remove Assisted-by references, with a bracketed [...] elision marker, then the highlighted line I don't understand this whole idiotic concept of "don't use AI for X". and the body text Here's a new rule just for you: don't use compilers for any of your kernel submissions. A cartoon crowd holding a NO AI CODE! placard is composited over the email in frame 003; frame 005 shows a stick figure at a laptop reading def checkout(cart): ... beside a robot labelled AI while three grey figures scowl; frame 009 pans to you had better not use an editor to write... and ...message without using tools that make it...; frame 011 animates a figure chiselling stone; frames 015 and 021 build an abstraction ladder stacked assembly, C, C++, libraries, frameworks, the app, with real x86 mnemonics (mov eax, [price], imul eax, [qty], call log_total) at the bottom and app.route('/checkout') near the top; frame 025 shows a pytest panel with test_empty_cart checked and test_two_items and test_zero_qty unchecked; frame 027 shows a human circling draft 1 in red against the robot's draft 2; frame 029 ends on a torn-paper card reading I expect people to understand what they are doing. attributed —Linus. The spoken argument is that Torvalds answered AI-code critics with a reductio (if accepting help means you do not understand your code, give up the compiler and the text editor too), that this is the latest step on a decades-long abstraction ladder, and that the creator's own practice is to read, question and direct AI output because "AI is not a compiler." The transcript's final segment is timestamped 55.840 to 60.960 s against an audio.mp3 of 58.853878 s and a video of 58.792993 s, so the end stamp runs 2.11 s past the audio; the words begin at 55.840 s, well inside the file, and the closing line is corroborated by the burned-in caption "to hold ourselves to" in frame 029, so the text is genuine and only the end timestamp is padded.

The quoted email is authentic, and the screenshot's elision removes the line that set its tone

The headers burned into frames 001 and 003 match a real message. Linus Torvalds posted to the Linux kernel mailing list on Tue Oct 13 2026 at 16:55:30 EST under the subject Re: [PATCH] docs: remove Assisted-by references, archived at lkml.iu.edu/2610.1/13255.html. Every line the video shows on screen appears verbatim in that archive: "I don't understand this whole idiotic concept of 'don't use AI for X'", "Here's a new rule just for you: don't use compilers for any of your kernel submissions", "Because if you can't show that you really understand what the machine is actually doing by writing your code in assembler, then I don't want to see your code contribution. You clearly aren't competent if you can't write it that way", and "And btw, you had better not use an editor to write that commit message either."

The screenshot's [...] marker, visible in frame 001 directly under the header block, elides the message's opening. In the archive that opening reads: "No. What you are about is something entirely different. It has nothing to do with 'expectations', and it has everything to do with your hangups." The cut matters because it removes the second-person target. On screen, "don't use compilers for any of your kernel submissions" reads as a general swipe at an anonymous crowd. In the source it is aimed at one named correspondent whose position Torvalds had just characterised as a hangup. The video also omits the message's loudest line, "Do you not see how UTTERLY STUPID those kinds of rules would be?", and its closing sentence, "the people I respect and call intelligent are not the ones that shy away from using new and interesting tools."

The video's description inverts Torvalds's conditional; the narration states it correctly

The TikTok description claims Torvalds was "effectively saying that if you can't have an AI write code then we should all go back to assembly and we can't have backspace in our code editors." That reverses the structure of the argument. Torvalds was not reasoning from an inability to use AI. He was running a reductio on a proposed rule: the rule says tool assistance proves you do not understand your work, so apply the same rule to the compiler and the text editor and watch it collapse. The antecedent in the source is a standard of proof, not an absence of AI.

The spoken narration gets this right where the description does not. At 9.52 s the video says "Linus's point is, if getting help means you don't understand your code, then lose the compiler and write an assembly," which preserves the conditional in the correct direction. At 14.48 s it continues "And he goes further. Oh, that commit message? No editor for you. Backspace hides your mistakes," a fair compression of the archived text. The error is confined to the written description field, which is the part most likely to be read without the audio.

"The don't use AI to write code crowd" misnames the position Torvalds was answering

The video's first line, at 0.00 s, is "Linus Torvalds just lit up the don't use AI to write code crowd." The archive's reply-to chain names the correspondent: the message Torvalds answered is Laurent Pinchart's, archived at lkml.iu.edu/2610.1/13564.html, posted Tue Oct 13 2026 at 18:49:33 EST. Pinchart was not arguing that AI should not write code. He was arguing something narrower and specific to commit messages. His words: "I'm increasingly thinking that the Mesa policy of requiring commit messages to be manually written (possibly with machine translation) could be applied to the kernel."

Pinchart's own text takes LLM coding productivity as a premise rather than attacking it. He writes: "I've heard multiple people saying how using LLMs made them more productive, with an extreme example during the maintainer summit of Linus talking about doing in a couple of months what would have likely taken multiple years." His argument was that if contributors genuinely control the process as they claim, hand-writing the commit message adds negligible cost. Replying to Kees Cook's objection that such a rule would be unenforceable, Pinchart answered: "The SoB line is not enforceable either. Very few of our requirements and assumptions can be enforced. What this is about is setting our expectations, telling what we want and what we don't want."

That last sentence is the one Torvalds quoted back at him before writing "No. What you are about is something entirely different." Knowing the target was a commit-message policy also changes the editor joke. The video presents "no editor for you, backspace hides your mistakes" as a free-floating absurdity. In context it is a precisely aimed counter: Pinchart proposed hand-written commit messages, so Torvalds extended the proposal to hand-typed commit messages with no backspace. The video never mentions commit-message policy, Pinchart, the Mesa project, or the Assisted-by tag, even though the tag's name is legible on screen in frame 003.

Torvalds arrived on day five of a thread he did not start, and the rule was already gone a week before the video

The video's framing is that Torvalds opened an attack. The archive timeline runs the other way. The 2026 Maintainers Summit met on Thursday 2026-10-08 in Prague with roughly 30 kernel developers, and reached consensus that the Assisted-by requirement could be dropped, per Jonathan Corbet's LWN report published 2026-10-12. Sasha Levin posted the patch the following morning, Friday 2026-10-09 at 04:31:58 EST, with a two-line changelog: "The Maintainers Summit reached consensus to deprecate the Assisted-by tag. Remove the documentation requiring it." The diffstat touches four files, deleting the Attribution section from Documentation/process/coding-assistants.rst, the "Using Assisted-by:" section from submitting-patches.rst and its Italian translation, and the tag handling from scripts/checkpatch.pl, for 1 insertion and 51 deletions.

Torvalds entered on Tuesday 2026-10-13, the fifth day of the thread and after roughly thirty messages from Levin, Pinchart, Krzysztof Kozlowski, Greg Kroah-Hartman, Jonathan Corbet, Steven Rostedt, Lorenzo Stoakes, Kees Cook, Randy Dunlap, Jori Koolstra, Mark Brown and Shuah Khan. He was a late participant in a change that was already moving, not its instigator.

The outcome is also missing from the video. Corbet applied the v2 patch on Wednesday 2026-10-14 at 12:58:55 EST with the reply "OK, I have applied this, thanks. jon", carrying an Acked-by from Lorenzo Stoakes and Reviewed-by tags from Kees Cook and Greg Kroah-Hartman. Levin followed six minutes later at 13:05:11 EST asking for a companion AGENTS.md patch, on the grounds that "without it, we'll get very mixed results with just this change, where agents will pick and choose different policies around Assisted-by: because nothing is pointing them to the actual guidelines we have." The video was posted 2026-10-21T03:19:59Z, eight days after Torvalds's message and seven days after the patch landed, and presents a live argument rather than a settled one.

"AI is not a compiler" is the creator's argument, not something Torvalds said

At 46.16 s the video says "I don't just hope that it works. I make sure. Because AI is not a compiler. You have to check its work." At 50.72 s it attributes the conclusion back to the kernel: "And that's what Linus comes back to. The standard we have to hold ourselves to is just understanding the code." The second half of that attribution is sound. The archived message says: "The kind of expectations I have is that people do their best and I expect people to understand what they are doing. But that understanding has very little to do with what tools they use." The closing card in frame 029 quotes the first of those sentences accurately and credits it to Linus.

The compiler contrast is the creator's own. Torvalds's argument runs in the opposite direction: he invokes the compiler precisely to collapse the distinction between tools, not to carve AI out as a special case needing verification. Nothing in the message says AI output requires checking that compiler output does not. The video stacks a reasonable practice claim on top of a quotation and lets the attribution blur across the seam. On the substance, the creator and Torvalds land close together, since both end at understanding rather than tool choice. The route there is different, and only one of the two is in the source.

A related gap sits in the frames. Frames 005, 025 and 027 illustrate the review loop with cartoons: a pytest panel where test_empty_cart passes and two tests sit unchecked, a human circling draft 1 in red. No terminal, editor, model output or actual run of any tool appears in any frame I read. The claim about reading and directing AI output is asserted, never demonstrated.

The numbers the summit acted on, which the video never mentions

Corbet's LWN write-up gives the data behind the decision. The 7.3 kernel took 18,533 non-merge changesets through -rc7, a record, of which 1,884, essentially 10 percent, carried Assisted-by tags, against a summit estimate that "probably 90% of contributors are using LLMs now." The tag was dropped because near-total non-compliance made it a worthless signal, not because the kernel endorsed unreviewed AI code. Maintainers at the summit still wanted to know how a patch was produced, and preferred that information in the cover letter rather than the changelog.

Corbet also records a Torvalds position the video would have strengthened its case by including: "blaming LLMs for the kernel community's problems is wrong, when the problem is LLM users who do not understand how to use them," with a preference for deprioritising low-quality contributions from inexperienced submitters over restricting LLM use. That is the same understanding-over-tools standard the video's closing card quotes, stated five days earlier and with the mechanism attached.

Key Takeaways

  • Verified: The email in frames 001, 003 and 009 is genuine. Linus Torvalds posted it to LKML on Tue Oct 13 2026 at 16:55:30 EST under the subject Re: [PATCH] docs: remove Assisted-by references, and every line shown on screen appears verbatim in the archive.
  • Verified: The closing card in frame 029, "I expect people to understand what they are doing," is an accurate partial quotation of the source sentence "The kind of expectations I have is that people do their best and I expect people to understand what they are doing."
  • Correction: The video calls the opposing side "the don't use AI to write code crowd" at 0.00 s. Laurent Pinchart, the person Torvalds was replying to, proposed that commit messages be written manually on the Mesa model, and cited LLM coding productivity approvingly, including Torvalds's own summit remark about compressing years of work into months.
  • Correction: The TikTok description states the argument as "if you can't have an AI write code then we should all go back to assembly." Torvalds's conditional runs the other way: if tool assistance is taken as proof you do not understand your work, the same test disqualifies the compiler and the text editor. The spoken narration at 9.52 s states this correctly; only the description inverts it.
  • Correction: The video frames Torvalds as opening the fight. The Maintainers Summit reached consensus on 2026-10-08, Sasha Levin posted the removal patch on 2026-10-09 at 04:31:58 EST, and Torvalds joined on 2026-10-13 after roughly thirty messages from other maintainers.
  • Clarification: The dispute was about deleting a disclosure tag that already existed, not about permitting AI code. The video shows the subject line docs: remove Assisted-by references on screen in frame 003 and never explains what the tag was.
  • Clarification: The video's screenshot elides Torvalds's opening sentences, including "it has everything to do with your hangups," which converts a reply aimed at one named correspondent into a swipe at an anonymous crowd.
  • Clarification: "AI is not a compiler. You have to check its work" at 50.72 s is the creator's framing, not Torvalds's. The archived message uses the compiler to erase the distinction between tools rather than to single AI out.
  • Context: Jonathan Corbet applied the v2 patch on 2026-10-14 at 12:58:55 EST. The video posted on 2026-10-21T03:19:59Z, seven days later, and does not mention that the change had already landed.
  • Context: The summit acted on measured non-compliance. Of 18,533 non-merge changesets in 7.3 through -rc7, 1,884 (about 10 percent) carried Assisted-by tags, against an estimate that roughly 90 percent of contributors use LLMs.
  • Unverified: Whether the on-screen crowd imagery corresponds to any real organised opposition. No such group is named in the video or in the mailing-list thread.
  • Context: Whisper rendered "write in assembly" as "write an assembly" at 14.48 s. The transcript's last segment ends at 60.96 s against 58.85 s of audio, a 2.11 s padded end stamp; the words begin inside the audio and are corroborated by the caption in frame 029, so the text itself is genuine.

Resources

Published October 21, 2026. Writeup generated from a favorited TikTok.