Torvalds: huge AI-reviewed kernel RCs are the new normal

Linus Torvalds shipped Linux 7.2-rc7 on Sunday with a sentence worth pinning above every engineering team's desk: the pile of last-minute fixes isn't an anomaly anymore, it's structural, and AI review tools are why. Release candidate seven is supposed to be the quiet one. It hasn't been quiet for weeks.

What actually happened

In the 7.2-rc7 announcement, Torvalds wrote: "Another week, another -rc. I can't say that I'm exactly thrilled about the size of this all, but it is what it is: the new normal with a lot of fixes, many of them due to review by various AI tools."

A week earlier, on rc6, he called it "the biggest rc6 we've had in years at least by commit count." The stable 7.2 release is still on track for the following Sunday — he isn't slipping the date, and he's explicit that nothing in the pile looks scary on its own. It's the volume.

Look at what's actually in it. HWMON landed a batch of critical and high-severity fixes. Memory management closed a use-after-free race condition that had been in the tree for eight years. Btrfs restored fixup worker infrastructure that was causing silent data loss. Networking took a run of AI-generated patches. The input subsystem picked up AI-driven fixes for information leaks and out-of-bounds accesses.

That's the honest read: these are real bugs, most of them old, and the tools are finding them faster than a release cycle was designed to absorb.

Why AI code review matters for your business

The interesting number here isn't how many bugs AI found. It's that the world's most disciplined review process — the one with maintainers, subsystem trees, and a benevolent dictator who reads the diffs — is visibly straining to keep up with the intake.

If the kernel is at capacity, your team is too. The bottleneck in AI-assisted development was never generation. It's the human who has to decide whether a patch is correct, and that person does not get faster because the queue got longer. We've written about this pattern from the other direction: review is the bottleneck, and AI report floods are already forcing maintainers to ration triage.

The practical version for a small team running AI review in CI:

Route by subsystem, not by severity. The kernel absorbs volume because fixes land in trees owned by people who know that code. If every AI finding lands in one shared queue, one person becomes the constraint. Assign paths to owners in CODEOWNERS and let the tool file against them.

Budget the review, not the scan. Scanning is cheap and getting cheaper. If you turn on an AI reviewer and don't add review hours, you've bought a backlog. Decide up front how many findings per week you'll actually action, and configure the tool to that number.

Track the age of what you find. An eight-year-old use-after-free is a genuine win. A wall of style nits is a tax. If your tool's output skews toward the second, tune the rules — Ruff's default expansion is the same lesson at a smaller scale.

Torvalds's move is the one to copy. He didn't turn the tools off and he didn't delay the release. He named the new baseline out loud so nobody mistakes it for a bad week.

Sources: Phoronix, The Register.