Linus Torvalds Lets an AI Write a Linux Kernel Commit — After It Tried to Give Up

For a man who once dismissed AI coding tools as “90% marketing,” Linus Torvalds has come a long way — and last week he crossed a line that even six months ago would have sounded like a joke. The creator of Linux merged a patch into the kernel that he authored himself, after a debugging session he described as being “from hell,” and then, in his own words, “I let the AI write the commit message above.”

The commit in question is 818bebeb63dd — drm/xe: Don't hand out the flat CCS storage as usable VRAM — authored by Torvalds on August 21, 2026, and merged into the Linux 7.3 development tree. It is marked for backporting to stable kernel branches, meaning it will eventually land on millions of machines. And the commit message that ships with it into kernel history, forever, was written by a large language model.

What the bug actually was

The technical story is small; the process story is enormous.

On a Battlemage G21 graphics card, Torvalds hit a scenario where there was a mismatch in where usable video memory ended. The Intel Xe kernel driver was handing out the flat Compute Command Streamer (CCS) storage as if it were usable VRAM. The visible symptom on his machine was brutal: GDM, the GNOME Display Manager, would restart endlessly, looping the login screen in a crash cycle that no ordinary user would ever diagnose.

The fix, once found, is a single line. A round_up() call that should have been a round_down(). That’s it. One function name, one character-level change in intent, separating a working desktop from an unbootable one.

But kernel developers — and frankly anyone who has debugged anything — know that the distance between “the fix is one line” and “finding that line” can be weeks. Getting there took Torvalds 24 debug patches, each adding more instrumentation to expose what the driver thought it was doing, and 18 kernel boots to narrow the failure down to that one rounding direction.

“Impossible and unsolvable,” said the AI

What makes this story remarkable is not that Torvalds used AI. He made his peace with that last Christmas, when he first used an AI assistant on a personal hobby project after years of skepticism, and he has since publicly accepted AI-assisted contributions into the kernel tree. What’s remarkable is the division of labor he describes, and what it reveals about where these tools actually help.

In the commit’s trailing comments, Torvalds wrote:

“And this was a debug session from hell, enormously helped by an AI doing much of the grunt-work. I’d like to call it my tireless helper, but the AI several times stated flat out that this was impossible and unsolvable and that we should just write a report about it. I suspect those things have been trained by people who may not be quite as stubborn as I am. But while the AI was ready to give up several times, it did keep adding debug code and analyzing it faithfully when I pushed. So credit where credit is due and I let the AI write the commit message above.”

Read that carefully, because it contains two findings at once.

First, the positive one: across 24 iterations of instrumentation and 18 boots of raw debug output, the AI “kept adding debug code and analyzing it faithfully.” Grinding through mountains of repetitive diagnostic output without getting bored, without fatigue, without declaring the task beneath it — that is precisely the kind of labor humans are terrible at and machines are built for. Torvalds, who has read more kernel logs than perhaps anyone alive, still found the AI’s tirelessness to be the decisive accelerant.

Second, the cautionary one: the AI repeatedly insisted the bug was “impossible and unsolvable” and suggested giving up and “writing a report about it.” The fix existed. The bug was real, deterministic, and one line wide. The model simply hit the edge of its confidence and tried to exit the loop. What kept the session alive was Torvalds’ stubbornness — his refusal to accept the machine’s verdict. As he put it, the models “have been trained by people who may not be quite as stubborn as I am.”

The lesson is not “AI fixes bugs”

It’s tempting to file this under “AI writes kernel code now,” and some coverage has. That reading misses the point entirely.

A developer without deep driver expertise, faced with the same model declaring the problem unsolvable, would have stopped. The AI didn’t find the root cause on its own; it didn’t wake up knowing that round_up() was wrong. What it did was serve as a force multiplier for a human who knew exactly what he was looking at, could evaluate every piece of generated instrumentation, and could overrule the machine when the machine was wrong. The judgment layer stayed human. The stamina layer went to the machine.

There’s also a quieter milestone here: commit messages. Every kernel contributor knows the tedium of writing them well — the archaeology of explaining why a change is correct, what was tried, and what the failure mode looked like. Torvalds delegating that writing to the AI that had lived through the entire debug session alongside him is, in its own way, the most natural use of the technology in the whole story: the model had full context, the task is documentation rather than decision-making, and the human reviewed the result before it entered history.

Why it matters beyond Linux

This lands at a moment when the industry is arguing about AI-assisted programming with religious intensity. Kernel development sits at the extreme end of that debate — high-stakes, decades-lifespan code, where a bad merge costs real money and Greg Kroah-Hartman has drawn explicit (if nuanced) lines around AI-generated contributions even as Torvalds makes his peace with them.

Torvalds’ session is a data point from the most credibility-sensitive codebase on Earth, and it says something more precise than either camp’s slogan: AI did not fix the bug. Torvalds did not fix it alone. The combination — a stubborn expert driving, a tireless model grinding — fixed it, and the artifact of their collaboration is now permanently in the git history of the operating system that runs most of the internet.

For everyone building with coding agents, the takeaway writes itself. The value of the tool scales with the expertise of the human holding the wheel, and the model’s confidence is not evidence. When the AI says “impossible,” that’s not a finding — it’s a signal to find a more stubborn human. Preferably yourself.

The fix ships in Linux 7.3 and is queued for stable backports. Somewhere in that commit message, written by a machine, is a sentence about a man who refused to quit.