Torvalds Blames AI for a Full-Fat Linux 7.3-rc2

Linux 7.3-rc2 landed on September 6 as what Linus Torvalds called a full-fat release candidate, and he joked that everyone would blame AI. The second snapshot after a merge window is meant to be a quiet week. This one was not.

He listed late subsystem pulls, said nothing looked odd, then reached for the line that now writes itself. The joke works because the project already told critics to live with those tools, while USB maintainers were sorting a mailbox that still would not drain.

Torvalds Called This One a Full-Fat Rc2

On September 6 Torvalds posted his Sunday 7.3-rc2 announcement on LKML and admitted the week had fooled him. “This didn’t feel like a particularly busy rc2, but it clearly was,” he wrote. Rc2 is usually the quietest stretch, when people take a breather after the merge window and bugs take a while to show up. Not this time, he said. This is a “full fat” rc release, and it is not some single cause.

He tried the late Error Detection and Correction pull that had been forgotten in the merge window, then waved it off as too small. Several filesystems had sent fixes together. There was a sizable DRM pull with scattered fixes, plus networking and BPF, and a number of driver trees. “Nothing looks particularly odd,” he added.

It might be just random, but we’ll obviously all blame it on AI, because whether that’s really the cause or not, it’s an easy thing to blame;)

Linus Torvalds, Linux kernel maintainer, Linux Kernel Mailing List, September 6, 2026

The 7.3-rc2 Makefile version bump is the usual one-line EXTRAVERSION change, from -rc1 to -rc2, under the release name Baby Opossum Posse. One count of the shortlog attached to that mail put the drop at 615 commits from 292 contributors. Torvalds did not publish a figure of his own.

What Landed in Linux 7.3-rc2?

The unusually large snapshot is still a pile of ordinary kernel work. Drivers took most of the diff. Tooling fixes made up about 20 percent of the non-driver volume, with filesystems, core kernel, and networking filling the rest. No one has pointed to a broken architecture change sitting under that pile.

WHAT STACKED IN RC2

• Late EDAC pull: The error-detection tree that missed the merge window, which Torvalds called a fairly small thing on its own.
• Filesystems together: NTFS and XFS both sent adjustments, part of the cluster he said arrived at once.
• DRM and Nouveau: AMD’s DRM tree plus display fixes in Nouveau for NVIDIA Blackwell cards, including HDMI work tied to a 2.147GHz pixel-clock cap.
• Networking and BPF: Targeted TCP and IPv6 fixes, memory-safety corrections in the ksmbd SMB server, and a BPF set in the same window.
• Scheduler and alloc: A Cache Aware Scheduling fix for hybrid CPUs, and Kees Cook’s tree-wide move of more kmalloc() calls to kmalloc_obj().
• RandStruct and Rust: RandStruct is off by default when a usable Rust compiler is present, to break a circular dependency with Rust kernel support.

Nouveau developer Mohamed Ahmed described four separate Blackwell problems around the GPU System Processor, including HDMI vendor infoframes on GB20x hardware, HDMI GCP AVMute register offsets, and vblank interrupts. He said that work grew out of HDMI 2.1 bring-up, with features such as FRL, DSC, and VRR still waiting on the clock cap. That is hardware enablement, not a chatbot dumping drive-by cleanups.

Greg Kroah-Hartman’s USB Inbox Hit 1,732 Messages

Four days before rc2, Greg Kroah-Hartman, the kernel’s stable maintainer and USB lead, had already told developers the cycle would hurt. “Things don’t seem to be slowing down in kernel development world, even for old subsystems that should be ‘stable’,” he wrote on social.kernel.org, posting a Mutt status line from his USB todo mailbox: 1,732 of 4,807 messages, 78M. “This is going to be a rough -rc cycle.”

After a first cut at the easy ones, obvious bug fixes and older versions replaced by newer submissions, the same mailbox still read 1,094 of 4,170, 69M. He cleared 638 of the filtered USB notes and called what remained crazy. Pushing back on AI-generated submissions takes time even when a patch is clearly wrong, he said, because he does not want to throw out a genuine fix in the process.

For drivers/staging he has already drawn a harder line: no LLM patches except security bugs the submitter can prove on the hardware the driver actually talks to. He also said a huge majority of the current flood is static analysis finding really old minor things that can be triggered if you hold the code wrong. The tools are not only inventing junk. In March he described a sloppy prompt that returned 60 findings, about one-third wrong and two-thirds pointing at a real problem that still needed a human changelog.

The CVE ledger is the other half of that load. Kroah-Hartman has shown counts moving above 1,000 with Linux 7.0 and past 1,500 with Linux 7.2, after years in the Linux 6.x era when a cycle sat nearer 500. Maintainers are now using frontier models to filter hallucinated reports, which is a neat loop: the same class of tool that stuffed the mailbox is being hired to sort it.

The Tree Already Asks for Assisted-by Tags

The kernel did not wait for this rc2 to invent a paper trail. Its tool-generated content review rules say tools raise the volume of contributions while reviewer and maintainer bandwidth stays a scarce resource, and that knowing which parts came from a person and which from a tool is how the project protects that resource. Maintainers may treat a generated series like any other patch, reject it, demand extra tests, or drop it in priority behind human-written work.

A companion document tells contributors to put an Assisted-by tag for machine help in the commit message, in the form Assisted-by: LLM plus any specialized checkers such as coccinelle or sparse. AI agents must not add Signed-off-by. Only a human can certify the Developer Certificate of Origin, and that human owns the patch. The assistant is told to build a reproducer, drop unverified reports, and say so when it could not test a fix, because maintainers currently waste too much time on those.

Those tags sit in git, which means they are countable. A public tracker that greps Assisted-by trailers has been doing that on mainline since January 1, 2026. The page is honest about the hole: it measures policy compliance, not actual AI usage, and it misses patches that landed without disclosure. The tags are voluntary, trivial tool use is out of scope, and a maintainer who rewrites a generated diff by hand need not say so. Torvalds still did not put a count in the rc2 mail. The instrument exists. The joke remains unmeasured.

Why the Wink Landed This Time?

On July 15 Torvalds put his foot down on the linux-media list after a fight over an agentic review helper named Sashiko. “Linux is not one of those anti-AI projects,” he wrote, “and if somebody has issues with that they can do the open-source thing and fork it. Or just walk away.” He called AI a tool like the other tools the project already uses, and said that was no longer in question. He would loudly ignore people who try to stop others from using it. He also granted the pain: it is a somewhat painful tool for maintainer workloads, and it keeps finding embarrassing bugs. “It’s not like natural intelligence is always all that great either.”

In May, during Linux 7.1-rc4, he had already declared the private security list almost entirely unmanageable. Researchers running the same models against the same tree were filing duplicate vulnerability reports in private, where nobody could see what had already been sent. “AI detected bugs are pretty much by definition not secret, and treating them on some private list is a waste of time for everybody involved,” he wrote, and the docs now push those finds onto a public path. He asked reporters who wanted to add value to read the documentation, create a patch, and put real work on top of what the model did.

So the Sunday wink is not a conversion. It is the same man who opened the door, then watched the mailbags swell, then declined to pretend he could name one file that made rc2 fat. The easy blame is also the policy. People reaching for “the dog ate my homework” are late: the kernel already treats AI as both a useful finder of old USB bugs and a machine for filling a Mutt folder to 78M.

AMD Headers Took a Third of Rc1

The 7.3 cycle was already large before this snapshot. Linux 7.3-rc1, released August 30, sat at 40.98 million lines on a cloc pass: 30,937,090 lines of code, 4,912,366 comments, and 5,134,418 blank lines, up from 40.42 million in Linux 7.2. That is about 560,000 lines in one merge window. Torvalds said it was not the biggest rc1 the project has had, but it was up there in commit count, and he did not think that surprised anybody anymore.

The fat in that diff was not a mystery either. “This is yet another one of those releases with a big AMD GPU register dump,” he wrote. This time it was the AMD DCN6 register headers, dumped as C #defines, plus code for the new generation. All in all, that AMD GPU code and header dump accounted for about a third of the whole rc1 patch. Mask that block out, he said, and the rest looks pretty normal.

KERNEL SIZE AT 7.3-RC1

The AMD directory, which holds AMDGPU, AMDKFD, display code, and the headers, is now roughly 16 percent of the kernel, up about 170,000 lines from 7.2. That growth is register maps and new IP blocks, the same pattern that has been stuffing diffs for years. It is also why a “full fat” rc2 on top of a 40.98 million line rc1 reads heavier than the calendar thinks it should, even when the shortlog is fixes.

Torvalds spent part of the rc1 week calling himself a grade A nincompoop for doing a system upgrade in the middle of the merge window and then chasing the breakage. He said that mess was cleaned up without lasting damage. Rc2’s size is a different kind of self-own: a process that now invites machine help, then shrugs when the weekly tar file comes in heavy.

October 18 Is the Target, With One Spare Week

Kernel.org’s own cadence is a two-week merge window and then about seven weeks of weekly release candidates, with rc7 usually last and an extra rc8 if the tree is still noisy. From 7.3-rc1 on August 30, seven weeks lands on October 18, which is the date in circulation for stable 7.3. If the incoming patch rate does not ease, that can slip to October 25.

THE 7.3 CALENDAR

1. August 30, 2026: Torvalds releases 7.3-rc1 at 40.98 million lines after the merge window.
2. September 2, 2026: Kroah-Hartman posts the USB todo mbox at 1,732 of 4,807 and warns of a rough -rc cycle.
3. September 6, 2026: 7.3-rc2 ships; Torvalds calls it full fat and points the easy blame at AI.
4. September 21, 2026: Kernel Recipes opens in Paris for three days, with Kroah-Hartman already teasing a talk on the CVE counts.
5. October 18, 2026: Planned stable 7.3, with October 25 held as the extra week if rc patches stay loud.

By September 8 he was already merging urgent x86, tracing, BPF, USB, and kmalloc_obj pulls toward the next snapshot. Testers and distribution maintainers are asked to run 7.3-rc2 from kernel.org on drivers and tooling. The tree is available now. The wink will not shrink the tarball.