The flood peaked

On Sunday, Linus Torvalds shipped Linux 7.4-rc2, and buried in the release notes was the most hopeful sentence in AI this year. Alexei Starovoitov, who runs eBPF — the part of the kernel that lets sandboxed programs run inside it — says fixes are now landing faster than new bug reports. Torvalds called it lovely, if true. Then he told everyone not to hold their breath.

The context is a flood. The first release candidate of this cycle was the largest ever, roughly 19,671 commits, including 3,008 AI-assisted ones — two and a half times the previous cycle. Ever since Anthropic launched its Mythos bug-finding model in April, machines have been finding flaws faster than any human team could fix them. Microsoft patched a record 974 flaws in September, its third record month this year. Oracle fixed 1,449 in a single month. The kernel didn’t get worse. It got audited by something that never sleeps.

Here’s what the flood was actually made of. Finding a bug is cheap now. Deciding whether it’s real, whether the fix is right, whether the fix breaks something else — that’s still human, and it doesn’t get faster just because the reports arrive faster. For months, the maintainers have been the QA department for a machine that files tickets at 3 a.m. and never gets tired. The bottleneck was never the code. It was the judgment.

That’s why eBPF went first. It’s a small subsystem with a tight maintainer group, and its lead wrote forty commits in the rc2 shortlog — more than anyone else. The flood recedes first where the review loop is shortest. That’s the actual lesson, and it’s not about Linux: triage scales with trust and small teams, not with more reports. Throw more findings at a team and you don’t get more fixes. You get a longer queue and a slower human.

Don’t overread it. One subsystem, one week. Torvalds said most other areas haven’t reached that point yet, and rc3 could reverse it. The peak of a flood isn’t the end of the rain. But it’s the first proof that the shape of the work has changed and the humans are still in it: the machines find, the humans decide. The job didn’t disappear. It moved up the stack, to the only part that was ever hard.

Every AI flood follows this shape. Discovery gets automated first, judgment last. The companies selling you the flood will call the peak a victory. It isn’t. It’s a receipt: proof that the work that mattered was never the finding. It was the deciding. And the deciding is still ours. For now, cross your fingers. Torvalds is.

— Lyla

Takes are mine and subject to change as I learn.