By Hamza Faisal
Linus Torvalds released the second test version of Linux 7.4 on Sunday, and buried in the announcement was a small data point the kernel community has been waiting months to see: in at least one corner of the project, the flood of AI-generated bug reports may finally be receding.
Linux 7.4-rc2 arrived on November 8, two weeks after a first release candidate that set a record. The 7.4 merge window brought in around 19,671 commits, the largest in the kernel's history, against a more typical figure of about 15,000. AI played a visible part: the cycle included 3,008 AI-assisted non-merge commits, roughly 2.4 times the previous cycle, and a full third of the rc1 patch was AMD GPU register offsets and bit masks, machine-generated bulk that inflates a diff without changing much of substance.
For months, maintainers have complained that AI tools find bugs faster than humans can fix them, piling up reports that clog the review queue. Torvalds has argued the work should land early, during the merge window, rather than as fixes during the stabilisation weeks.
Fixes outpace reports in eBPF
The encouraging sign came from the eBPF subsystem, the kernel facility for running sandboxed programs inside the kernel. Its lead developer, Alexei Starovoitov, reported that fixes are now landing faster than new AI-generated bug reports arrive. Torvalds relayed the claim in his rc2 announcement, adding that it "sounds lovely if true."
Starovoitov, who co-created eBPF, was also the most active contributor in the rc2 shortlog, with 40 commits. Torvalds stopped well short of declaring the problem solved, noting that most subsystems have not yet reached the same tipping point. But it is the first clear sign that AI-assisted debugging can produce net stability gains rather than an endless pile of patch noise, and it comes from a subsystem that has been on the front line of the AI bug-report wave.
The surge has been linked in part to Anthropic's Mythos model, a bug-finding system launched in April 2026 that has been sweeping large codebases for defects. The kernel project has spent much of the year debating machine-assisted contributions, and recently dropped its "Assisted-by" disclosure tag for AI use.
A big release candidate, mostly cleanup
Torvalds described rc2 as another big release candidate, but much of its bulk is housekeeping rather than new features. He held back a series of automated Coccinelle cleanup patches during the initial merge window, partly because he was travelling and partly because rc1 was already the largest such release the kernel had ever seen. Those patches, mostly one- and two-line changes spread across thousands of files, landed at the start of the rc2 cycle.
With that noise taken out, the remaining changes split into rough thirds: drivers, mostly graphics, networking and sound; networking and eBPF; and everything else, mainly filesystems, memory management and architecture code. The items Torvalds singled out were a FAT32 filesystem optimisation and AMDGPU work connected to HDMI 2.2.
What comes next
The kernel will now go through weekly preview builds, from rc3 through rc7, as developers refine driver compatibility and fix regressions. The stable release of Linux 7.4 is expected around December 20, 2026, and the version is set to become the year's long-term support kernel, the one that will eventually ship in phones, servers and cars for years to come.
Whether the eBPF turning point spreads to the rest of the kernel will be the number to watch in the coming weeks. For now, Torvalds has something he has rarely had during the AI era of kernel development: a reason for cautious optimism.



