632 urgent fixes. 648 feature patches. Jakub Kicinski reckons up to half of that second number was written by a machine.
The networking pull for Linux 7.3 went out this week, and the part worth reading was not BIG TCP or the new wireless drivers. It was the cover letter, where the people who run one of the kernel’s busiest subsystems said out loud what they have been circling for months: Linux networking AI patches have them “completely overwhelmed”.
Kicinski’s own tally, from the pull request:
“Quick and dirty count suggests we (Paolo and I) merged a very similar number of net (632) and net-next (648) patches. This is not telling the full story either because 1/3 to 1/2 of the net-next patches also seem like AI-driven low priority fixes, cleanups and clarifications.”
The symmetry is the tell. net is the fixes tree, where things that are actually broken get repaired between releases. net-next is where new work lands over a full cycle. Those two are not supposed to move at the same speed, and historically they don’t. A fix stream that suddenly matches the feature stream patch-for-patch is a fix stream inflated by something that can produce a plausible-looking null check faster than a human can read one.
Fighting fire with fire #
The obvious response is a ban, and several maintainers have gone that way. The wireless side has already taken a hard line on unreviewed LLM output. netdev decided to do something stranger.
They got an AI token budget from Meta and now push submissions past multiple frontier models before a human ever opens them. It catches a share of the hallucinated nonsense, and Kicinski is careful not to oversell it: the models only go so far, and they cannot tell you whether a change was worth making in the first place.
The plan for next cycle goes further, and it is the bit that should interest anyone watching how open source absorbs this. Rather than pointing the models only at review, netdev wants them on the chores: babysitting patchwork, auto-replying to the same twenty process complaints maintainers type out every release, cleaning up commit messages, eventually applying patches that already carry a human review. The subsystem is being restructured so that machines handle the machine-generated volume, and humans are saved for the judgement calls.
Pruning helps too. Old networking drivers that almost certainly have no users left have been dropped in recent cycles, partly to shrink the surface area agents keep filing cosmetic fixes against. You cannot get fifty cleanup patches for code that no longer exists.
What actually shipped #
Quite a lot, for what it’s worth. BIG TCP now works inside UDP tunnels for VXLAN and Geneve, and MPTCP behaves better under extreme memory pressure. Wired drivers gain initial CXL support for AMD/Solarflare NICs, firmware flashing on AMD Pensando cards, and the skeleton of Intel’s iXD driver, which will eventually handle the Control Plane PIC function on E2100 and later IPUs and FNICs. Wireless picks up more Multi-Link Operation work, a new mm81x driver for Morse Micro’s long-range S1G radios, NXPWIFI for NXP hardware, and MT7928 support plus NAN on the MT7925 in MediaTek’s mt76.
One honest caveat: “seem like” is carrying real weight in that quote. Nobody has published a rigorous audit separating AI-authored patches from human ones, because there is no reliable way to do it. What the maintainers have is a pattern they recognise after years of reading this list, not a measurement. That is worth remembering before anyone quotes “half the kernel is AI now” at you.
The uncomfortable question is what happens if the curve keeps bending. Review capacity is the kernel’s scarcest resource and always has been. Generating patches got cheap almost overnight; reading them did not.
Details on the pull are on Phoronix, and the full submission is on the kernel mailing list.