The compiler rule

On 13 October Linus Torvalds, who wrote Linux and still has the last word on it, answered a fellow developer on the kernel mailing list. The developer wanted the project to keep telling contributors what it did and did not want from AI tools. Torvalds did not.

From:        Linus Torvalds
Subject:     Re: [PATCH] docs: remove Assisted-by references
Date:        Tue, 13 Oct 2026
Archive:     lkml.iu.edu/2610.1/13255.html

I don't understand this whole idiotic concept of "don't use AI for X".

[... the rest is in the archive ...]

The rest of the message is a joke told with a straight face. He proposes new rules: no compilers for kernel submissions, so that every contributor proves they understand the machine by writing in archaic assembler; no editors, so that nobody can hide a typo with the backspace key. Then he drops the face. Rules about tools are stupid, he says; what he expects is that people do their best and understand what they are doing, and that understanding has little to do with what they used. The people he respects, he adds, are not the ones who shy away from new tools.11Linus Torvalds, "Re: [PATCH] docs: remove Assisted-by references", linux-kernel mailing list, 13 October 2026, message-id CAHk-=whTs09_ozN+e09GU68N1wPP56MUzDjrg9rnKxQUa0A_4Q. https://lkml.iu.edu/2610.1/13255.html (T1 -- the author's own message in the list archive)

The rule under discussion was modest. Since January the kernel's documentation had asked anyone who used an AI coding assistant to say so, with an "Assisted-by" line on the patch.22Sasha Levin, "docs: add AI Coding Assistants documentation", Linux kernel commit 78d979db6cef, committed 2026-01-06; simplified to "Assisted-by: LLM" by Christian Brauner in commit 816d9992d9ed, 2026-08-03. https://github.com/torvalds/linux/commit/78d979db6cef557c171d6059cbce06c3db89c7ee (T1 -- the kernel's own history) Ten months later, roughly one commit in ten carried the line, while everyone involved agreed the real use was far wider.33Sasha Levin, "[PATCH v2] docs: remove Assisted-by references", 13 October 2026: roughly 10% of the 18,533 non-merge commits in 7.3 through -rc7 carry the tag, "while actual LLM use is far more widespread". https://patchew.org/linux/20261013121501.3738437-1-sashal@kernel.org/ (T1 -- the patch author's own message, via a list mirror)

Biology has had this argument before

Swap the kernel for a cell and the email reads as a history of the life sciences.

By the reckoning of America's National Human Genome Research Institute, sequencing one human genome cost about $95m in 2001. Its figure for May 2022 is $524.62.44National Human Genome Research Institute, "DNA Sequencing Costs: Data": cost per genome $95,263,071.92 at 2001-09-30 and $524.62 at 2022-05-31, the latest value published; an estimate for a human-sized genome, not the cost of a specific project. https://www.genome.gov/about-genomics/fact-sheets/DNA-Sequencing-Costs-Data (T1 -- the funder's own data) The Human Genome Project, which produced the first reference, cost about $3bn.55National Human Genome Research Institute, "Human Genome Project Completion: Frequently Asked Questions": the projected cost was $3 billion, and "most agree that this rough amount is close to the accurate number". https://www.genome.gov/human-genome-project/Completion-FAQ (T1 -- the funder's own account) Nobody now asks a geneticist to prove her competence by reading a sequencing gel by hand, and no journal requires an "Assisted-by: sequencer" line. The question a referee asks of a variant call is the one that was always worth asking: is it real, and does it mean what you say it means?

The Nobel committee has twice rewarded tooling. In 1993 Kary Mullis shared the chemistry prize for the polymerase chain reaction, a way to copy a stretch of DNA until there is enough of it to study.66The Nobel Prize in Chemistry 1993, one half to Kary B. Mullis "for his invention of the polymerase chain reaction (PCR) method". https://www.nobelprize.org/prizes/chemistry/1993/summary/ (T1 -- the awarding body) In 2024 Demis Hassabis and John Jumper shared it for predicting protein structures, with David Baker for designing new ones.77The Nobel Prize in Chemistry 2024, one half to David Baker "for computational protein design", the other half to Demis Hassabis and John Jumper "for protein structure prediction". https://www.nobelprize.org/prizes/chemistry/2024/summary/ (T1 -- the awarding body) The scale of that second change is massive. The AlphaFold database holds over 260m predicted structures; the Protein Data Bank holds 260,896 determined by experiment.88AlphaFold Protein Structure Database (EMBL-EBI and Google DeepMind): "over 260 million protein structure predictions", read 17 October 2026. https://alphafold.ebi.ac.uk/ (T1 -- the database's own count)99RCSB Protein Data Bank: 260,896 experimental structures, read 17 October 2026. The ratio of about a thousand to one is our arithmetic on the two counts. https://www.rcsb.org/ (T1 -- the database's own count) Predictions outnumber measurements about a thousand to one.

Each time, the same pattern. A tool takes over a step. Those who mastered the step suspect those who did not. The suspicion lasts until the work that the tool made possible becomes too good to argue with. Then the step is forgotten, and competence is measured one level higher.

The usage is already here

Anyone who thinks these tools are for drafting emails should look at who is using them. The most cautious institution in biology is the one that decides which medicines Americans may take. In May 2025 the Food and Drug Administration finished its first pilot of AI-assisted scientific review; a deputy director in its drug-evaluation office said the tool let him do review tasks in minutes that had taken three days.1010US Food and Drug Administration, "FDA Announces Completion of First AI-Assisted Scientific Review Pilot and Aggressive Agency-Wide AI Rollout Timeline", 8 May 2025; Jinzhong Liu, Office of New Drugs, CDER, on performing review tasks "in minutes that used to take three days". https://www.fda.gov/news-events/press-announcements/fda-announces-completion-first-ai-assisted-scientific-review-pilot-and-aggressive-agency-wide-ai (T1 -- the agency's own announcement) On 2 June it gave a generative AI tool, Elsa, to every employee, and said it was already using it to speed clinical protocol reviews, shorten scientific evaluations, summarise adverse events and pick high-priority inspection targets.1111US Food and Drug Administration, "FDA Launches Agency-Wide AI Tool to Optimize Performance for the American People", 2 June 2025: Elsa is "already" used "to accelerate clinical protocol reviews, shorten the time needed for scientific evaluations, and identify high-priority inspection targets", and "can summarize adverse events to support safety profile assessments". https://www.fda.gov/news-events/press-announcements/fda-launches-agency-wide-ai-tool-optimize-performance-american-people (T1 -- the agency's own announcement) When the drug regulator reads with a language model, the argument about whether such tools belong in serious work is over. The argument worth having is about how the work is checked.

Business has reached the same view. Anthropic, which makes Claude, publishes how its tools are used. Through its programming interface, which companies build into their own systems, 74% of use in late 2025 was work, and three-quarters of those exchanges handed the model a whole task rather than a part of one.1212Anthropic, "Anthropic Economic Index report: economic primitives", 15 January 2026, on November 2025 data: programming-interface use "overwhelmingly work-related (74% vs. 46%)", with "three-quarters of interactions classified as automation"; on the consumer app augmentation 52%, automation 45%. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report (T1 -- the company's own usage data; it makes the tool in question) Yet on the consumer app, more conversations used the model as a partner than as a substitute, 52% to 45%.1212Anthropic, "Anthropic Economic Index report: economic primitives", 15 January 2026, on November 2025 data: programming-interface use "overwhelmingly work-related (74% vs. 46%)", with "three-quarters of interactions classified as automation"; on the consumer app augmentation 52%, automation 45%. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report (T1 -- the company's own usage data; it makes the tool in question) People hand it the routine and keep the judgement. Torvalds's standard -- understand what you submit -- is already the habit of most of the people a rule would be written to police.

What it looks like from here

This site is a small demonstration of the point. The full English text of China's 2026-2030 plan for its pharmaceutical industry was rendered by a machine pipeline, and then every number, threshold and negation in it was matched against the Chinese characters, and a model from a different family was asked to find the errors.1313BigBio, "China's pharmaceutical industry plan for 2026-2030, side by side", 25 September 2026, with its method in "Chapter, verse and checksum". https://bigbio.ai/articles/china-pharma-15th-fyp-plan/ (T1 -- our own published work) A pitch deck's claim that two patents had been issued was put to the patent office's own records, which said both had expired; the check took minutes, where it once took a lawyer's afternoon.1414BigBio, "Pipe dreams", 8 August 2026. https://bigbio.ai/articles/pipe-dreams/ (T1 -- our own published work)

The surprise, for me, has been the direction of travel. The more I use these tools, the closer the work gets to fundamentals. I spend my days with plain text, primary registries and checksums, lower-level and simpler instruments than I have used at any point in my career. The machine fetches; I decide what to ask, whom to ask, and whether the answer can stand. A search engine returns what has been written about a thing. This is something else: a library that answers back, and a document that can be rewritten the moment the record changes. It is, admittedly, a little alarming. So, in its day, was a computer on every desk, which Steve Jobs liked to call a bicycle for the mind.1515Steve Jobs used the bicycle analogy in a talk filmed around 1980 and, in a 1990 interview, called the computer "a bicycle of the mind"; no earlier written source was found. https://www.computerhistory.org/highlights/stevejobs/ (T3 -- filmed talk via a museum archive; the 1990 wording via transcripts) The new tools are rather faster than a bicycle.

Drop the label, keep the method

Torvalds's email settled the tag question for Linux: the documentation dropped it the next day.1616Jonathan Corbet, reply to the v2 patch, 14 October 2026: "OK, I have applied this". https://lkml.iu.edu/2610.1/15868.html (T1 -- the documentation maintainer's own message) A tag is a meaningless confession: it names the tool and stops.

Biology invented a better habit, and it is older than any of these tools. A paper's methods section names the sequencer, the software and the version, not to apologise for them but so that a reader can judge the result and repeat it. That is the standard worth carrying into the age of the language model. The question to ask of any piece of work, a kernel patch, a variant call, a diligence memo, is not "was AI used?" but "what was checked, against which source, and can I check it too?" This site tries to pass that test: each piece says how it was made in the spirit of a methods section, not a confession, and every figure links to the record it came from.

Strip away everything that is not essential and that is what remains. Not whether a model, a sequencer or a compiler was involved, but whether the person can say what the tool did, show what was checked, and stand behind the result. The tools will keep changing; today's will look quaint within a decade. The question will not. Can you think? Torvalds asked it of his contributors. It is worth asking of anyone who hands over an answer, including us.

Footnotes

1. Linus Torvalds, "Re: [PATCH] docs: remove Assisted-by references", linux-kernel mailing list, 13 October 2026, message-id CAHk-=whTs09_ozN+e09GU68N1wPP56MUzDjrg9rnKxQUa0A_4Q. https://lkml.iu.edu/2610.1/13255.html (T1 -- the author's own message in the list archive) ↩

2. Sasha Levin, "docs: add AI Coding Assistants documentation", Linux kernel commit 78d979db6cef, committed 2026-01-06; simplified to "Assisted-by: LLM" by Christian Brauner in commit 816d9992d9ed, 2026-08-03. https://github.com/torvalds/linux/commit/78d979db6cef557c171d6059cbce06c3db89c7ee (T1 -- the kernel's own history) ↩

3. Sasha Levin, "[PATCH v2] docs: remove Assisted-by references", 13 October 2026: roughly 10% of the 18,533 non-merge commits in 7.3 through -rc7 carry the tag, "while actual LLM use is far more widespread". https://patchew.org/linux/20261013121501.3738437-1-sashal@kernel.org/ (T1 -- the patch author's own message, via a list mirror) ↩

4. National Human Genome Research Institute, "DNA Sequencing Costs: Data": cost per genome $95,263,071.92 at 2001-09-30 and $524.62 at 2022-05-31, the latest value published; an estimate for a human-sized genome, not the cost of a specific project. https://www.genome.gov/about-genomics/fact-sheets/DNA-Sequencing-Costs-Data (T1 -- the funder's own data) ↩

5. National Human Genome Research Institute, "Human Genome Project Completion: Frequently Asked Questions": the projected cost was $3 billion, and "most agree that this rough amount is close to the accurate number". https://www.genome.gov/human-genome-project/Completion-FAQ (T1 -- the funder's own account) ↩

6. The Nobel Prize in Chemistry 1993, one half to Kary B. Mullis "for his invention of the polymerase chain reaction (PCR) method". https://www.nobelprize.org/prizes/chemistry/1993/summary/ (T1 -- the awarding body) ↩

7. The Nobel Prize in Chemistry 2024, one half to David Baker "for computational protein design", the other half to Demis Hassabis and John Jumper "for protein structure prediction". https://www.nobelprize.org/prizes/chemistry/2024/summary/ (T1 -- the awarding body) ↩

8. AlphaFold Protein Structure Database (EMBL-EBI and Google DeepMind): "over 260 million protein structure predictions", read 17 October 2026. https://alphafold.ebi.ac.uk/ (T1 -- the database's own count) ↩

9. RCSB Protein Data Bank: 260,896 experimental structures, read 17 October 2026. The ratio of about a thousand to one is our arithmetic on the two counts. https://www.rcsb.org/ (T1 -- the database's own count) ↩

10. US Food and Drug Administration, "FDA Announces Completion of First AI-Assisted Scientific Review Pilot and Aggressive Agency-Wide AI Rollout Timeline", 8 May 2025; Jinzhong Liu, Office of New Drugs, CDER, on performing review tasks "in minutes that used to take three days". https://www.fda.gov/news-events/press-announcements/fda-announces-completion-first-ai-assisted-scientific-review-pilot-and-aggressive-agency-wide-ai (T1 -- the agency's own announcement) ↩

11. US Food and Drug Administration, "FDA Launches Agency-Wide AI Tool to Optimize Performance for the American People", 2 June 2025: Elsa is "already" used "to accelerate clinical protocol reviews, shorten the time needed for scientific evaluations, and identify high-priority inspection targets", and "can summarize adverse events to support safety profile assessments". https://www.fda.gov/news-events/press-announcements/fda-launches-agency-wide-ai-tool-optimize-performance-american-people (T1 -- the agency's own announcement) ↩

12. Anthropic, "Anthropic Economic Index report: economic primitives", 15 January 2026, on November 2025 data: programming-interface use "overwhelmingly work-related (74% vs. 46%)", with "three-quarters of interactions classified as automation"; on the consumer app augmentation 52%, automation 45%. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report (T1 -- the company's own usage data; it makes the tool in question) ↩ ↩²

13. BigBio, "China's pharmaceutical industry plan for 2026-2030, side by side", 25 September 2026, with its method in "Chapter, verse and checksum". https://bigbio.ai/articles/china-pharma-15th-fyp-plan/ (T1 -- our own published work) ↩

14. BigBio, "Pipe dreams", 8 August 2026. https://bigbio.ai/articles/pipe-dreams/ (T1 -- our own published work) ↩

15. Steve Jobs used the bicycle analogy in a talk filmed around 1980 and, in a 1990 interview, called the computer "a bicycle of the mind"; no earlier written source was found. https://www.computerhistory.org/highlights/stevejobs/ (T3 -- filmed talk via a museum archive; the 1990 wording via transcripts) ↩

16. Jonathan Corbet, reply to the v2 patch, 14 October 2026: "OK, I have applied this". https://lkml.iu.edu/2610.1/15868.html (T1 -- the documentation maintainer's own message) ↩