358
Linus Torvalds let Claude Code fix the Linux page cache. Then reviewers found a race. (thenewstack.io)
submitted 3 weeks ago by to c/technology@lemmy.world
165 comments fedilink hide all child comments
 

They got to him

you are viewing a single comment's thread
view the rest of the comments
[–] 121 points 3 weeks ago (29 children)

On Tuesday, he submitted a 10-patch series that delivered a 23-fold throughput improvement in a synthetic benchmark

Impressive, probably kind of like I can answer math questions very very fast.

Reviewers have since identified a race condition and raised concerns about memory reclamation.

That what I thought, I can answer math questions fast, but I don't claim to be able to answer them correctly.

  • permalink
  • fedilink
  • source
  • hideshow 29 child comments
    •
  • [–] 100 points 3 weeks ago (5 children)

    It seems that after the initial AI-generated patches there's been good discussion amongst the kernel devs about sorting out the issues, and a few revisions already that have tidied things up a bit.

    So if after all that there ends up being a good performance improvement directly from a revised AI-created patch, what do I do with my pitchfork? Still raised angrily? Half-mast? Do I just let the torch smoulder?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] 17 points 3 weeks ago (1 child)

    Wonder how much time it had taken without AI

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 18 points 3 weeks ago

    Torvalds had spent 18 months running a broken version of his own fix for the page-cache reference-counting bottleneck, while other developers had encountered problems with their approaches.

    The article mentions he had been working on solving the issue for 18 months already... so at least that long.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 7 points 3 weeks ago*

    I'm still mad because the source is still tainted. You can guarantee that the contributions of unwilling people, { who 1) did not intend their work to be used by a for profit LLM, and 2) actively do not want their work used by an LLM at all }, is part of these changes.

    Scraping isn't reading, etc etc. But also when did we just stop respecting people's wishes with the stuff they post? Like I made my own word scramble js script because a website owner didn't get back to me on whether I could use theirs, even though the script is right there to copy paste.

    These companies just did the yolo founder-disrupter impatient thing when they could have been chill and done this all legit.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 81 points 3 weeks ago (22 children)

    So you're saying the review process worked as designed?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 22 child comments
    •
  • [–] 40 points 3 weeks ago (18 children)

    Maybe it did, but apparently the code was not properly reviewed initially, so there could be other hidden problems.
    Of course that's the same for human made code, but there's a problem in only judging code from how it performs, and not actually understanding the code. And that seems to be the situation with a lot of AI code.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 18 child comments
    •
  • [–] 24 points 3 weeks ago (11 children)

    And particularly disturbing when the human understanding failure is from the original and lead Linux developer.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 11 child comments
    •
  • [–] 10 points 3 weeks ago* (last edited 3 weeks ago) (8 children)

    100% that is the worst place to have the problem originate.
    I have rarely disagreed with Linus Torvalds on things regarding the kernel, and maybe I'm wrong, but I strongly disagree with him on this issue. Contributing code nobody understands cannot be a good thing.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 8 child comments
    •
  • [–] -1 points 3 weeks ago* (6 children)

    The code still undergoes the exact same code review as before, because humans were the original slop coders.

    The standards needed for code to be accepted haven't changed at all. Even the code Linus personally is reviewed..

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 6 child comments
    •
  • [–] 4 points 3 weeks ago (5 children)

    The code still undergoes the exact same code review as before.

    That's not good enough, because it is well documented that AI often make mistakes and absolutely needs to undergo way stricter review.
    It is also well documented that although you can achieve some results quickly, AI code also quickly becomes impossible to maintain.

    If nobody understands the code, as it seems was the case here, and it is introduced anyway, the kernel is doomed to deteriorate.

    My wife has already noticed deteriorating quality in Debian, with system hangs after they loosened the requirements for AI code. The quality of Debian now is lower than it has been in a decade!

    Maybe you need to watch something like this:

    https://www.youtube.com/watch?v=l9ARF2KcDgc

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [+] 0 points 3 weeks ago* (last edited 3 weeks ago) (5 children)
    [removed by mod]
  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] 1 point 3 weeks ago* (last edited 3 weeks ago) (4 children)

    Lol. That guy banned me from his community

    Maybe because you are making stupid ignorant comments.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] -1 points 3 weeks ago* (last edited 3 weeks ago) (3 children)

    Or maybe his blog post falls apart completely by simply pointing out that AI isn't approving code, code doesn't go straight to main, and all code is manually code reviewed by human.

    And his blog is pages of slop.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 4 points 3 weeks ago* (last edited 3 weeks ago) (2 children)

    AI isn’t approving code

    You really seem to believe that's a good point, but it isn't.
    It is an extremely well documented fact that critical/independent thinking declines with use of AI, and you apparently have zero understanding of that fact.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] -1 points 3 weeks ago* (last edited 3 weeks ago)

    Did you even read that 1 study you're repeating, or just click the link

    If you read the study, they ran the tests over 13 mins only. And, they didn't keep retesting to see if results changed over time either. If that ran the test over an hour, both results may have ended up at similar levels (we don't know because it only tested over a short time)

    You need to learn to stop exaggerating or actually read crap instead of just the headline lol

    I know it's the same study we both (or presumably just myself) read, because it was the only one. I wouldn't call that "extremely well" anything

  • permalink
  • fedilink
  • source
  • parent
  • [–] -2 points 3 weeks ago* (last edited 3 weeks ago)

    You seem to believe that's a great response. At this point, you're just arguing (which seems to be most of your Lemmy history).

    If you want to argue that a journalist with a slop blog and no evidence of being a developer anywhere is credible, that's up to you.

    All I see is obviously bad journalism from a guy who wants attention...

  • permalink
  • fedilink
  • source
  • parent
    • load more comments (-1 replies)
  • [–] 2 points 3 weeks ago (1 child)

    Is it an understanding failure, an unexpected interaction, or just an existing issue now exposed? A 23x performance increase could very well expose an existing race condition that systems weren't fast enough to run into.

    People make mistakes. Software bugs existing at all is proof of this. This is certainly not the first time a change Torvalds approved has an issue, and it won't be the last.

    There's a reason code is reviewed, and there's a reason people still go through after it is approved and verify. The process is working exactly as designed, the way it has for decades before AI coding.

    Expecting absolute perfection is ridiculous, the anti-AI circlejerk acting like only AI fucks up just makes them look ridiculous. Any developer could have developed this fix, and there would be no article. Nothing in this issue is specific to AI usage and treating it like it is somehow special just undermines arguments against AI slop and vibe coding in general.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 0 points 3 weeks ago (5 children)

    So, let me get this right. You're disapproving on how things were handled based on your speculation of how it went?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] 2 points 3 weeks ago (4 children)

    Not based on speculation, based on numerous analysis of the problems with using AI code you don't understand.
    Clearly Linus failed to understand the code, otherwise he wouldn't have suggested it for submission.

    You seem to have a problem with criticism of Linus, is it a religious thing?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] 0 points 3 weeks ago* (last edited 3 weeks ago) (1 child)

    You understand it less it sounds like.

    1. Race Conditions are common. This isn't new to AI. Nobody writes stories about them though unless they lead to security issues (like with XFS, which had one for 10 years and only fixed recently).
    2. There is a lot happening in the kernel, its not always a matter of looking in one tiny location. A race condition is pretty broad and can lead to many things.
    3. Patches get rejected all the time, for reasons including race conditions which are identified.
    4. If you check the post, other developers had tried and failed to fix the same issue. He's not the only one. It was just yet another failed attempt
    5. Optimisation is HARD, and always has been.
    6. Finally, you say "suggested it for submission". That code would go to a staging repo first for testing. He wasn't suggesting pushing it immediately to the main tree..
    7. Standard development tools (at least for my development work) wouldn't catch race conditions (maybe commercial ones / high end ones would).. AI is actually far better at detecting these things

    This isn't a unique or interesting story in any capacity.... It's boring and normal

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] -5 points 3 weeks ago (1 child)

    Of course of course. How dare you blaspheme against our Lord and Saviour, filthy infidel!

    But anyway, I'd like to be as enlightened as you. Could you name any of those numerous analyses that detail exactly what Torvalds did and what he failed to understand?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 0 points 3 weeks ago* (last edited 3 weeks ago)

    No I can't, because those are so general and commonly known, that if you somehow are unaware, you must have been living under a rock for the past year.
    The second option is that you are just trolling or for some reason is in denial.
    Either way I'm not wasting my time on any of those options.

  • permalink
  • fedilink
  • source
  • parent