Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-23 · high
The model called the bug unsolvable, then produced a fix once Torvalds pushed back, in a commit landed August 22.
Why now: Torvalds' commit and Simon Willison's related note on verification both landed on August 22.
Linus Torvalds landed a Linux kernel fix on August 22 crediting an AI with doing much of the grunt work in what he called a debug session from hell. The commit stops the drm/xe graphics driver from handing out flat CCS storage as usable VRAM. In the message he writes that the model kept insisting the problem was impossible to solve, then kept adding diagnostic code anyway whenever he pushed back.
That's the detail worth sitting with. The usual complaint about coding agents is that they hallucinate fixes with confidence. This is the opposite failure. A model refused, confidently, and it was wrong to. Nothing in the refusal itself signaled that. Torvalds jokes that his own stubbornness outlasted the AI's training, but that's also the whole method: push once more before accepting "impossible."
Simon Willison made a related argument the same day. His claim is that the job isn't reading every line of a diff, it's confidently instructing an agent and confidently verifying the result, because line-by-line review was never the fastest way to catch a bad change. Torvalds' commit is what that looks like from the inside: instrumenting and re-checking instead of taking the model's first answer.
A Hacker News thread adds a rougher data point. Several people running Claude Code report worse outcomes on high reasoning effort than medium, including one case where a one-file config update turned into 43 minutes of pulling containers and building test suites. Anecdotal, unverified, but the shape matches. More compute spent isn't more judgment applied. Treat a model's "impossible" as a hypothesis, not a verdict, and go check it yourself.
Each link below shares sources, entities, or timing with this story.
Simon Willison uses Claude Code / Shared entities / Same source domain / Shared topic / Earlier coverage
Linked by a graph relationship (Simon Willison uses Claude Code); both cover August, Claude Code, Linux, Simon Willison; reported by the same outlet (simonwillison.net).
Linked by a graph relationship (Simon Willison uses Claude Code); both cover August, Claude Code, Simon Willison, There; reported by the same outlet (simonwillison.net).
Linked by a graph relationship (Simon Willison uses Claude Code); both cover August, Claude Code, Simon Willison; reported by the same outlet (github.com, simonwillison.net).
Linked by a graph relationship (Simon Willison uses Claude Code); both cover Claude Code, Simon Willison, There; reported by the same outlet (github.com, simonwillison.net).
Linked by a graph relationship (Simon Willison uses Claude Code); both cover August, Claude Code, Hacker News; reported by the same outlet (github.com).
Simon Willison uses Claude Code / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Simon Willison uses Claude Code); both cover Claude Code, Hacker News, Half, There; overlapping topics (code, model).
Simon Willison uses Claude Code / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Simon Willison uses Claude Code); both cover August, Claude Code, There; overlapping topics (agent, code, line).
Linked by a graph relationship (Simon Willison uses Claude Code); both cover August, Claude Code, Hacker News; overlapping topics (agent, code, commit).