Login
You're viewing the techdon.dev public feed.
  • Nov 7, 2026, 10:59 AM

    The trillion dollar monsters are all focusing on *finding* issues at high speed in large volumes. Hardly any of them care about *fixing* these problems with any sense of quality.

    💬 28🔄 9⭐ 0

Replies

  • Nov 7, 2026, 11:02 AM

    @bagder they feel the urge to show spectacular results. Side effects do not really count.

    💬 0🔄 0⭐ 0
  • Nov 7, 2026, 11:04 AM

    @bagder They don't care about any of that, they just want credit (at the bank). These two are mostly inversely related btw. in this and other fields. The best they can do is clap if you fix things.

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Nov 7, 2026, 1:07 PM

    @bagder if youve always just hired someone else to buy everything, you dont realize that quality is worth anything

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Nov 7, 2026, 1:23 PM

    @bagder Yea, for marketing & stonks go less down.
    As if any of them really care about security.
    They only security thry care about is to not be defaced, and for no legal trouble to come their way.

    💬 0🔄 0⭐ 0
  • Nov 7, 2026, 1:39 PM

    @bagder isntb that in part because folks use sub optimal prompts. "Find me a vulnerability" and the outstanding pattern matching that AI is today does that. A better prompt might ask for a simple fix. That said, i would almost never trust the solution, without careful review and testing

    💬 0🔄 0⭐ 0
  • Nov 7, 2026, 1:41 PM

    @bagder Would you agree, that with a finite amount of important software, there are a finite number of outstanding bugs?

    If so, if we look at this in cumulative terms of the percentage of AI-findable bugs found, while it looks exponential now, it will asymptotically approach 100% once we pass the inflection point?

    In other words, the bugpocalypse is a myth, spread by you know who to sell more tokens and infect your workflows? Or can someone point out the flaw in my reasoning?

    💬 1🔄 0⭐ 0
  • Nov 7, 2026, 3:43 PM

    @bbuk @bagder
    Not quite: machine learning algorithms are just oversized statistical models, so at best they'll only ever approach the false positive rate, never zero.

    Additionally, again because they're statistical models, they'll tend to make more mistakes with non-boilerplate code. So the more difficult the code, the larger that false positive rate gets; script kiddies will tend to have fewer problems, but that was never a pain point to begin with.

    💬 0🔄 0⭐ 0
  • Nov 7, 2026, 3:58 PM

    @australopithecus @bagder To your first point, that's part of why I said asymptotically, the other part is the long tail. My broader point is there is no bugpocalypse, what's your take on that?

    To your second, this is a spam issue, not a security issue. I thought @badger was more concerned about the long-term security of the situation rather than the current spam issue. Once we're passed the inflection point and well in the asymptotic, the spam can be muted without much concern for security.

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Nov 7, 2026, 8:44 PM

    @australopithecus @bagder Again, my point is not reaching 100%, it's the fact that the rate of bugs found by clankers on genuine projects will not continue to increase in an exponential way. The bugpocalypse is fake news.

    Obviously projects that use slop themselves will remain screwed.

    💬 0🔄 0⭐ 0
  • Nov 7, 2026, 8:51 PM

    @bbuk @bagder
    Well, that all depends on what the actual false-positive rate is. So far it seems pretty high.

    Anyway I'm not sure anybody is seriously concerned about AI finding actual bugs, I think the concern is mainly precisely about the false-positive spam, and the consequent dev burnout.

    💬 0🔄 0⭐ 0
  • 💬 2🔄 0⭐ 0
  • Nov 7, 2026, 9:13 PM

    @bagder @australopithecus Yes, but that will tend towards all outstanding, ai-findable bugs being found. It's not a permanent issue, just a one-off, technical debt cleanup period. Sucks to be in charge of an important project of course.

    💬 0🔄 0⭐ 0
  • Nov 8, 2026, 1:34 AM

    @bagder @bbuk
    Sure, machine learning algorithms are actually good at finding things, unlike how they are at generating things.

    I just don't think that's really arguably a *problem*, exhausting though it may be to keep up.

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Leelooleeloo@c.im
    Nov 7, 2026, 2:54 PM

    @bagder
    Fixing them takes actual intelligence.

    Finding them is more of a pattern recognition job.

    💬 0🔄 0⭐ 0
  • Nov 7, 2026, 3:32 PM

    @bagder when prompting for fixes, people keep forgetting to add "MAKE NO MISTAKES" and "DON'T TALK ABOUT GREMLINS" to their prompts

    💬 0🔄 0⭐ 0
  • Nov 7, 2026, 4:42 PM

    @bagder That's because they can't and will never be able to. Slop machines can only produce slop.

    💬 0🔄 0⭐ 0
  • Nov 7, 2026, 5:08 PM

    The scientific community has become largely the same so it's the natural consequence 🤷‍♂️

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Nov 8, 2026, 11:23 AM

    @bagder …but, but that would make everyone realize the tools can’t handle the implicitness, the nuance. A binary "target crashes or not" is too good of a verifier to pass up 😐

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Nov 10, 2026, 11:49 AM

    @bagder Let me push internally more on this. I think we need to do better in high quality fixes and better classification if they are low / medium / high vulns without wasting maintainer time

    💬 0🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Nov 10, 2026, 12:08 PM

    @bagder in the past Pareto's principle taught us that the last 20% of a task take 80% of the time/effort. Now we are just fine with the initial 80% generated. :underheart:

    💬 0🔄 0⭐ 0
  • Nov 11, 2026, 12:43 PM

    @bagder You described Canonical pretty good. When I still interacted with them, they tended to produce "fixes" I would not touch with a 10ft pole. More workaround then fix, even if the fix is actually easier to do.

    💬 0🔄 0⭐ 0
  • Nov 11, 2026, 2:03 PM

    @bagder Because they can't.
    With millons lines of code out there, even at an error rate of 20%, you can find thousands of errored code pieces. But fixing them, even with 95% precision, isn't enough...

    💬 0🔄 0⭐ 0