Claude), asked it to reason through the lock ordering, and the model spotted a missing
d_lock
acquisition in a code path that human reviewers had missed twice.The fix was five lines. Finding it took the model forty seconds.
What's interesting isn't that AI found a bug — static analyzers do that. It's that the model reasoned about the locking hierarchy across multiple subsystems (dentry cache, inode locks, mount namespace) the way a senior maintainer would, but without needing to hold the entire kernel state in working memory. Linus's exact phrasing: "it connected dots I didn't have time to connect."
He's still skeptical about AI writing net-new kernel code — "I don't want hallucinated locking primitives in my RCU paths" — but this debugging use case has shifted his stance from "overhyped autocomplete" to "genuinely useful for the boring forensic work."
Practical takeaway if you're doing systems work: feed the model context, not just the crash site. Give it the header files, the lockdep annotations, the relevant .c
files within three hops of the fault. The more constraints you provide, the less it hallucinates. I've started keeping a debug-context.md
in each subsystem directory — lock hierarchy diagrams, common race windows, recent fixes — specifically to paste into these sessions.
Also worth noting: the model didn't replace the verification step. Linus still ran the stress test suite for twelve hours. AI proposed the hypothesis; silicon confirmed it. That division of labor feels like the right model for kernel work going forward.
Linus Torvalds says AI 'enormously helped' a debug session from 7h ago
Captain makes travel planning actually usable via Telegram 7d ago
Claude Code actually writes a decent novel if you stop treating 8d ago
Linus Torvalds thinks AI is fundamentally changing how the Linux 10d ago
Linus Torvalds thinks AI is actually helping the Linux kernel 11d ago
Building a custom AI code review agent is way cheaper than the 11d ago
Next The Leiden Declaration on AI and Math just dropped — has anyone →