Linus Torvalds Uses Claude Code to Rework the Linux Page Cache, Tripling Small-Read Throughput
It may or may not be merged. Nonetheless it's certainly a step forward into genAI adoption for our beloved OS.
I would really like to know what's going to happen: better code or just the appearance of it?
https://hwbusters.com/news/linus-torvalds-uses-claude-code-to-rework-the-linux-page-cache-tripling-small-read-throughput/Open linkView original on programming.dev
10 replies
he also introduced at least one bug/vulnerability in the form of a race condition.
HIS FIRST VIBED COMMIT. The commit that at least should have been the most carefully scrutinized before handing it in. Then he left the cleanup to other maintainers and went on a vacation.
Linus needs to step down, he has fucking lost it.
For real? Do you have further info or a link?
I wish it wasnt. Here:
https://woof.tech/@techokami/117446164348546588
Yea... Linus should stay on vacation.
Wouldn't he need some time and trials to perfect the art?
You mean Dirty COW? CVE-2016-5195 The race condition lived undetected in the source tree for nine years until security researcher Phil Oester discovered it in October 2016.
Unlike this race condition in the age of LLM agents which never made it into a stable release.
Linus used to shame large companies for not contributing to open source, now he shames devs for not using the large companies closed source software.
Claude models are really. I've been used them for so long. My colleagues too. I noticed a major improvement in code quality and enablement of developers. Now front end developers can take on more serious back end work in my team. As an analogy for this, I expect better outcome for Linux.
Guys, AI could perform outstanding hacks and can enable now mass hacks in a way never seen before. It also enables serious security scans and uncovered lots of serious security flaws. So why wouldn't it help get better kernel?
Note that finding flaws is not, in general, the same mode as creating correct or well-made things. AI is plausibly good at the former, but this does not necessarily translate to the latter.
Agreed, although I worry about this use case the most for "brain rot" - that security pen testers will quit thinking and just use LLM agents instead. The state of the art needs both, not just one or the other.
Coders with brain rot only have to make working code that fulfills all the requirements, security scans have a much higher burden, and they are also more directly linked to expensive problems when they fail in their task.