Linus Torvalds Used AI to Trace a Linux Graphics Bug, Finding More Along the Way
Twenty-four debugging patches and 18 kernel boots helped Linus Torvalds track down a fault that repeatedly restarted a Linux display manager.
Twenty-four debugging patches and 18 kernel boots helped Linus Torvalds track down a fault that repeatedly restarted a Linux display manager. The Linux creator used AI to make unfamiliar code easier to decipher—and the investigation exposed additional bugs.
The problem was in the Intel Xe graphics driver for Battlemage G21 cards. Torvalds, who now generally reviews and merges contributions rather than writing kernel code himself, took on the fix with AI assistance, according to a report.
AI as a debugging aid
The one-line patch was the end of a much longer hunt. Torvalds used a series of debugging changes and repeated kernel starts to narrow down the cause of the restart loop. The exercise also surfaced other problems in the process.
Torvalds has described AI as useful for tasks he finds difficult, and said he enjoys using it. But he has drawn a line between experimenting on a personal project and applying AI to software with real-world consequences.
“You need to be very careful using AI when you’re doing something real and important,” he said during a discussion at Open Source Summit Europe.
Human understanding still matters
His position differs from developers who reject AI-generated contributions outright. System76, for example, has barred AI-generated content from its COSMIC desktop project. Torvalds, by contrast, sees AI as a tool that can help developers work through code—but not as a substitute for understanding it.
At Open Source Summit North America, he pushed back against claims that AI will eliminate the need for programmers. For long-lived projects, he argued, developers need to grasp both the generated code and the architecture around it, rather than relying on prompts alone.
That distinction carries particular weight for the Linux kernel, where unchecked generated patches and bug reports could add to the burden on maintainers. In Torvalds’ graphics-driver investigation, AI helped make the code more legible; the debugging still took 24 patches and 18 boots.
