Despite all the controversy surrounding AI use, one thing is becoming increasingly obvious: it is finding its way into more and more parts of everyday software development. And apparently, Linux kernel development is no exception.
Linux creator Linus Torvalds has revealed that he used AI while tracking down a particularly nasty memory corruption bug in the kernel’s Intel Xe graphics driver, describing the whole process as a “debug session from hell.”
The issue affected Flat CCS memory handling in the Xe DRM driver and, on Torvalds’ own Intel Battlemage G21 system with 16 GiB of VRAM, caused a black screen as GDM repeatedly restarted the compositor.
The problem stemmed from a surprisingly small mistake. The get_flat_ccs_offset()
function calculated where the reserved CCS storage begins and then rounded that address up to a 128 KiB boundary. However, everything below that calculated address was subsequently handed to the VRAM allocator as usable memory. As a result, the compositor’s first GPU submission failed, GDM restarted it, and the machine remained stuck at a black screen. According to Torvalds, the actual fix was almost comically small. Getting there, however, was anything but simple.
Torvalds said the investigation involved 24 patches adding progressively more debugging information and 18 kernel boots before the problem was finally narrowed down. And this is where AI entered the picture.
According to him, AI performed much of the repetitive work during the investigation, adding debugging code and analyzing the resulting information as he guided the process. He noted that the AI repeatedly concluded that the problem was impossible to solve and suggested simply writing a report about it. But Torvalds, being Torvalds, kept pushing.
Once told to continue, the AI kept producing additional debugging instrumentation and analyzing the results until the underlying memory corruption was isolated. Torvalds even let the AI write the lengthy commit message documenting the problem once the investigation was complete.
The episode also fits remarkably well with Torvalds’ increasingly pragmatic public position on AI.
Back in October 2024, his view was considerably more skeptical. Speaking at the Open Source Summit in Vienna, Torvalds said AI was genuinely interesting and would likely change the world, but complained that the industry surrounding it was dominated by hype.
However, his position has become noticeably more concrete since then. During discussions over Linux kernel guidelines for tool-generated contributions in January 2026, Torvalds argued against turning kernel documentation into either a pro- or anti-AI manifesto. Instead, he wanted AI treated simply as another development tool, with the usual expectations around responsibility, review, and quality still applying.
By July 2026, Torvalds had made his position considerably more explicit. Responding to criticism of AI-assisted development on the kernel mailing lists, he stated that Linux is not an anti-AI project and described AI as another tool available to developers – one whose usefulness, in his view, is no longer seriously in doubt.
In short, Torvalds’ position is essentially that the technology should be judged by what it produces rather than rejected simply because AI was involved. And the latest Intel Xe debugging episode is probably one of the clearest demonstrations yet of what he means.
And after 24 debugging patches and 18 kernel boots, with the help of AI, the entire investigation eventually came down to changing an incorrect round_up()
into a round_down()
. Sometimes the smallest kernel fixes really do have the longest stories behind them.