top 50 comments

sorted by: hot top controversial new old
[–] 220 points 3 weeks ago (5 children)

I don't have the time to write out the full response to this I want to right now, so I just want to say there is an almost immeasurable gulf between someone like Linus Torvalds using AI for debugging vs a coked out techbro junior engineer vibecoding the next Spotify but for access to water.

In my heart of hearts I still wish he wouldn't touch the stuff just for the message this is going to send to all the wrong kinds of people but here we are.

  • source
  • hideshow 5 child comments
  • [–] 66 points 3 weeks ago (2 children)

    Just use ai to write out your longer response /s

  • source
  • parent
  • hideshow 2 child comments
  • [–] 16 points 3 weeks ago (1 child)

    That would improve length, comprehension, desirability, gramification, and … hold on, still buffering out the Vibe List of random categories I posthoc figured help justify my point. AI taught me that’s a good way to make points.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 76 points 3 weeks ago* (26 children)

    If an AI can find the bug and point it out, I don't see a problem, as long as it is checked thoroughly enough by an actual skilled programmer, to make sure that doing what the AI suggest is not making things worse.

  • source
  • hideshow 26 child comments
  • [–] 70 points 3 weeks ago (13 children)

    The exact quote is:

    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.

    This is basically a one-liner fixing a bogus "round_up()" to a "round_down()", but there were 24 patches adding more and more debug information to this, and 18 kernel boot to finally narrow it down to this. - Linus

    That's a long way short of finding the bug and pointing it out, and instead is just the tedious work of adding checks and then deciding where to add more checks when you have the result of the previous checks. I suspect if we'd spent a trillion dollars on a tool for that, we could have come up with something much more efficient than an LLM.

  • source
  • parent
  • hideshow 13 child comments
  • [–] 10 points 3 weeks ago (2 children)

    I have found that LLM agents can monitor ALL the logs, synthesize information that takes me hours to put together in minutes - and when there's a problem in there it can spot it faster and more often than me. It can also miss things, not as many as I miss given equal time to work the problem, but it's not perfect - just better. Like cutting down a tree with a chainsaw, instead of a bare hacksaw blade.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 8 points 3 weeks ago (1 child)

    For every bug the AI does not find and I have to spend 1 hour to realize the AI is not able to find and then spend 2 hours to identify myself, there are 10 bugs the AI finds in a couple minutes.

    I can gladly spend the extra time when it doesn't work.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 4 points 3 weeks ago

    I find that the LLM agents find maybe 90% of "bugs" in their first pass, and 2/3 of those tend to be stuff I really wouldn't sweat over in most cases, but O.K. they're easy enough to button up so let's do that, then the other 30% are more serious stuff I'm genuinely relieved got found sooner than later.

    Then you get into that last 10%, and I'll help it in the hunt, and probably 90% of them I'll solve much more quickly with the LLM agent than I ever would have without it, even if I have to give it directions about where and how to look for the problems / solutions.

    Then there's the final 1% where the LLM just doesn't figure it out, even with clues, and I can see it and fix it far faster than the LLM... those exist, and they make great "how dumb are LLMs these days???" stories, but that's ignoring the other 99% where LLMs are clearly helpful.

    Of course all the above is generalization, and if you roll back 12 months ago the LLMs weren't nearly as good then - and any further than that and I thought they were a total waste of time, but starting about 12 months ago it was pretty clear they were useful, and improving, fast.

    Last week I actually solved a network configuration problem on the other side of the planet by having the LLM: 1) create scripts to diagnose the problem, 2) have a test engineer run those scripts on problem machines and control (non problem) machines, 3) get the logs from those script-runs and have the LLM analyze them, 4) repeat the process, based on information from 3 make new scripts focused on the problem area, 5) more test engineer runs, this time generating many megabytes of log files, 6) LLM analyzes the log files from 5 and identifies the issue, 7) LLM makes a script to reconfigure the terminal machine to solve the issue, 8) test engineer deploys it and retests - problem solved, 9) I discover the "solution" broke something else, describe the problem to the LLM, 10) LLM adjusts the "solution" to fix the problem without creating the new problem.

    What's more remarkable about all that is: we first observed this problem 5 years ago, it was kind of a minor annoyance how it manifested but it's recently become more important that we solve it. In those 5 years, nobody here or there figured out what was going wrong, but a couple of rounds of applied LLM diagnosis and treatment led to a solution inside a week.

  • source
  • parent
  • load more comments (12 replies)
    [–] 38 points 3 weeks ago (6 children)

    He use AI to put print statements, summarize output and describe code based on prints to narrow place of bug. This is exactly what AI is good at. Funny fact is that you don't need data center scale model for it. You can do it with local model.

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

    You may. You seem to not have seen the potatoes people are using. ;)

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

    You can get a surprising amount of shit done with an Intel B580.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 5 points 3 weeks ago

    3060 ti 8GB and 32 GB DDR4 says no, I can't do that or anything else useful with a local model.

    However. Deepseek V4 flash is very cheap and fast. It'll do anything simple and some complex things you throw at it

  • source
  • parent
  • [–] 32 points 3 weeks ago (10 children)

    Ha! I recently went through a similar process, trying to get some hardware working for a different project that would close some 15+ year old bugs. AI helped me find it, getting sidetracked by several other small bugs in the same few files along the way.

    I then wrote the actual fixes myself as a series of very small PRs. They’re all just a handful of lines each, enabling certain hardware flags, properly bounds checking loops, getting a timer initialization corrected, etc. They are all the sort of thing that once you see where they are, the fix is extremely obvious.

    In the meantime, the project has adopted a very strict no AI assistance at all policy after some internal debate that almost tore the project apart.

    So now I’m sitting on almost a dozen fixes for longstanding problems and have no idea if I should lie and say yes, I found these entirely by myself, or tell the truth, probably get kicked out of the project, and see how much of a shit storm they cause.

  • source
  • hideshow 10 child comments
  • [–] 23 points 3 weeks ago* (4 children)

    Both extremes are absolutely maddening here. I'm getting as sick of the anti-AI crowd as I am the vibe-code-tech-bros. At this point you could just put them all in a sack and start hitting them with sticks and wouldn't care who got the worst of it.

    There's a definite middle-ground here. Linus is one of the ultimate pragmatists and it shows here. It was helpful - so he used it. It helped you so you used it. I've found it to be incredibly helpful so I use it.

    Not all usages are equal just as all hand-written code is not equal. There are good ways to use these tools and bad ways. But the tools themselves are still just tools.

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

    Yeah, it's like a calculator. Don't get reliant on it. Know why it gives an answer, be able to get to the answer yourself, but appreciate the time saving of the calculator when you can verify that it was correct much faster than getting to the answer yourself.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 9 points 3 weeks ago

    The art of using a digital calculator is being able to put in a problem that would take you a couple of minutes to solve, get the calculator's answer in a couple of seconds, AND be able to spot-check the reasonable-ness of the answer in a couple of seconds more. Blind trust is ridiculously risky, cutting your speed by half to move from 99% correct to 99.99% correct is well worth the extra time and effort invested.

  • source
  • parent
  • [–] 1 point 3 weeks ago*

    This isn’t such a simple question.

    If you think about it that way, the atomic bomb is also a kind of tool, but not everyone is allowed to use it...

    If a tool starts thinking for you, it’s clearly a different kind of tool than most known tools. Different tools have different safety procedures. :)

  • source
  • parent
  • load more comments (1 reply)
  • [–] 4 points 3 weeks ago
    • 1️⃣ Start screen recording yourself 🎦
    • 🦟 Fix the bug and write up the PR, except...
    • Throw in a lot of gratuitous emojis 🎉🏆🤝 -- em-dashes -- and "On the other hands"
    • 📤 Submit it
    • When they come 😡 screaming at you about no-AI, just show them the video 😎
    • All their outrage will be spent on accusing you of using 🤖AI.
    • Where it all started will be forgotten 🥳
    • Profit 🏆

    Comment written by 💯 human

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

    Honestly 15+ year old bugs are also left in the linux kernel for people to pick up, Learn from and get deeper in kernel development. They know they can easily find them with a LLM but choose not to. Policy's are changing due to the mass influx of these "bug" fixes.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 4 points 3 weeks ago (1 child)

    There's academic exercises for education. The production kernel used by billions around the world is no place to intentionally leave bugs, and it doesn't matter how they're detected. What matters is if they are corrected without creating regressions.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 2 points 3 weeks ago*

    Not really... "Education" and real world are vastly different. Getting accustomed to a mailing list other maintainers etc is not something you learn with your education.

    And leaving bugs for hardware no one has touched in ages is not too big of a deal. And a nice gateway for new comers who actually want to learn it instead of trow a LLM at it.

  • source
  • parent
  • [–] 23 points 3 weeks ago

    But while the AI was ready to give up several times, it did keep adding debug code and analyzing it faithfully when I pushed.

    Well, no shit, all they can do is follow instructions. We just really can't help avoid anthropomorphizing something that speaks language and think it was a good boy for stopping whining and trying again

  • source
  • [–] 7 points 3 weeks ago

    Honestly I am so tired of intels shit. I have been buying and plugging their GPUs but goddamn their shit is annoying to deal with sometimes. I’m glad Linus is chipping in to improve it. We need another big player.

  • source
  • [–] 4 points 3 weeks ago

    The code it committed has comments with "old code did this", "old code did that", like AI likes to insert. That is not useful info. It's what version control is for... You can explain why you do a thing, like the AI does at the start, the actual code edited. The assertion it edited is just clutered with irrelevant info now. Is including history like this the usual style for the kernel, seeing that Linus himself signed off on it?

  • source
  • [–] 2 points 3 weeks ago
    load more comments
    view more: next ›