Login
You're viewing the oslo.town public feed.
  • jcoglanjcoglan@mastodon.social
    Oct 22, 2026, 7:03 AM

    the critiques I've seen of the php patch have little to do with "moral purity" and "feeling icky" and are more about the actual approach to mitigating the problem

    *including* considering the goals and point of view of people whose job that is, so it's not about making unconsidered demands for human labour either

    "pick your battles, e.g. complain that AI is hurting artists" it's also hurting programmers and distorting how programming gets done so I'd consider this relevant

    💬 1🔄 7⭐ 0

Replies

  • jcoglanjcoglan@mastodon.social
    Oct 22, 2026, 7:03 AM

    idk it sucks to see analysis like "yeah of course backporting 12kloc patches is hard which is why it's useful to think of other solutions that don't involve 12kloc patches" written off as moral purity testing

    "I am concerned this huge patch is a symptom of a lack of design coherence and will just hide further bugs" is not morality policing

    "a security patch that is impractical to review is a category error" is not "feeling icky"

    💬 3🔄 48⭐ 128
  • jcoglanjcoglan@mastodon.social
    Oct 22, 2026, 7:48 AM

    also pertinent criticism: people having to spend money on tokens to fix issues found, and to some extent created by, the people selling the tokens

    💬 2🔄 9⭐ 34
  • ❄️ undead sneppossum 🌹lune@teal.social
    Oct 22, 2026, 7:05 AM

    @jcoglan heard conflicting reports on if this particular bug was unearthed by / only practical via LLM-driven programming.

    but it sure does remind me of the previous times where there was definitely a LLM-discovered and slopped webpage promoted zero day like, twice within a week.

    it really does not increase my confidence in "okay for all the harms I guess they fix security bugs" as explanation when the thing that mostly seems to happen is "surprise new zero-day with no patches" regularly.

    💬 1🔄 0⭐ 2
  • David Chisnall (*Now with 50% more sarcasm!*)david_chisnall@infosec.exchange
    Oct 22, 2026, 10:57 AM

    @lune @jcoglan

    When people say 'this is possible only with LLM-generated programming' what they often actually means is 'this is possible only with massively subsidised automation'. An LLM is probably not the most efficient way of getting to the end point, but when the compute for the LLM is subsidised by a factor of 10-100x, if the other technique requires 20% as much compute it's still vastly more expensive if you can't burn someone else's money on it.

    💬 7🔄 51⭐ 106
  • 💬 1🔄 0⭐ 4
  • 💬 0🔄 0⭐ 2
  • 💬 0🔄 0⭐ 3
  • coracora@hachyderm.io
    Oct 22, 2026, 6:06 PM

    @david_chisnall @lune @jcoglan brb thinking about how much securitized private credit is subsidizing generation of bad knockoff photoshop lawsuit magnets, and what the clawback process is going to look like.

    💬 0🔄 0⭐ 3
  • 💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 1
  • tuban_muzurutuban_muzuru@beige.party
    Oct 22, 2026, 8:43 PM

    @radex

    With respect, this entire LLM phenomenon is a Gold Rush.

    Yes, you're right, under one definition of cost. Yet marginal serving cost (GPUs, power, datacenter rent per token served). Here price is probably well above cost.

    💬 0🔄 0⭐ 0
  • Edvin MalinovskisnCrazed@fd00.space
    Oct 22, 2026, 7:55 PM

    @david_chisnall @lune @jcoglan this is the part that's been boggling my mind! Every time when an impressive "see! LLM can do *this* now!" happens - there's inevitably prior art. Which seems consistent with how LLMs work in general

    It's just traditionally all these problems have been deemed too expensive by those in charge of money

    But somehow these corporations decided to burn orders of magnitude *more* of money and make it everyone else's problem AND they keep getting rewarded for it!? WTF!?

    💬 1🔄 1⭐ 3
  • Martin Taibrmartin_t@mastodon.social
    Oct 23, 2026, 6:31 PM

    @nCrazed @david_chisnall @lune @jcoglan First, they claimed LLMs were AI. Then people noticed fragments of training data in the outputs. They filtered those out.

    Then they claimed LLMs were AI. Then people noticed they are much better at types of tasks, projects and languages in the training data. They added more training data without consent, including your GH activity and how you interact with the LLM.

    And they claimed LLMs were AI again. Really, they're just reproducing patterns.

    💬 1🔄 0⭐ 0
  • Martin Taibrmartin_t@mastodon.social
    Oct 23, 2026, 6:34 PM

    @nCrazed @david_chisnall @lune @jcoglan The "innovation" is that they're not just reproducing the final result but the steps a human would have taken.

    Every time a hyper tells you that they produce novel results instead of regurgitating interpolated training data, yes, they do, by literally simulating people who did the original work. LLMs cannot exist without the original work and the people who did it deserve ownership of LLMs.

    💬 0🔄 0⭐ 0
  • D Ingramingram@mastodon.social
    Oct 22, 2026, 8:57 PM

    @david_chisnall @lune @jcoglan Isn't it amazing what can be achieved when you task what is a supercomputer at a problem? Subsidised monkeys and typewriters on a scale not seen before. Or fling enough shit against a wall and eventually a visually appealing pattern will appear.

    💬 0🔄 0⭐ 0
  • Very Human RobotStompyRobot@mastodon.gamedev.place
    Oct 23, 2026, 2:24 AM

    @david_chisnall @lune @jcoglan

    Do you have sources for the claimed subsidy size?

    From what I can tell, API payment is profitable to the large providers, Enterprise seats are profitable or neutral, and paid consumer subscription plans are at most 2x subsidized.

    Source: I ran models similar to this for a smaller company for a couple of years.

    💬 1🔄 0⭐ 1
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • PuercoPopPuercoPop@mastodon.social
    Oct 22, 2026, 6:11 PM

    @jcoglan Something I haven't seen people mention is that tools to perform a code transformations at a large scale from a succinct descripción of the change exist. Tools like coccinelle, used by the Linux kernel, let you write a 'semantic patch'. The focus of the review then shifts the patch itself not the 12KLoC changes.

    💬 0🔄 0⭐ 1