Claude's Load-Bearing Seams

(madradavid.com)

93 points | by rzk 1 hour ago

21 comments

  • NichoPaolucci 21 minutes ago
    The only thing I find funnier than this article is that some people think that LLMs are communicating at a “higher level” than humans and developing some alien language.

    The amount of ambiguity is astonishing to me, maybe I’m wrong but there’s just no way that two agents in a clean room wouldn’t devolve into absolute insanity.

    • antoineMoPa 8 minutes ago
      It's like corporate lingo for computers.
    • mohamedkoubaa 9 minutes ago
      Devolve implies it starts elsewhere
    • dude250711 12 minutes ago
      Mutually coalescing on a local maximum of abstract concepts over-represented among data distribution weights.
  • scottyah 36 minutes ago
    The only thing worse than reading the repetitive claude-ism's while using the actual tool is seeing (likely human) efforts to reproduce it on every comment and blog.
    • cortesoft 11 minutes ago
      You are absolutely right, and that's on me.

      I think this breaks me out of the actual work I am doing more than anything else Claude does. Every time it tries to offer some mea culpa for a mistake it made, it sounds so ridiculous. Those sorts of responsibility accepting statements are important for humans because there are real emotional entanglements involved in working together. This isn't true with an AI, and when it fakes it, it just confuses the situation more.

      I don't mind when Claude says it will make a note of the mistake to avoid it in the future, because that is an actual thing it is doing that might make a difference. But accepting responsibility does nothing, and doesn't mean anything for future interactions.

    • bigwheels 32 minutes ago
      I've noticed Claude-ish increasingly creeping into how folks write on HN. It is unfortunate, but not really surprising given how many of the participants in this community are exposed to reams of AI writing on a daily or even hourly basis.

      Given the hundreds of thousands I've burned in tokens over the past year, I'm surprised to not have picked up more of that nonsense (AFAICT I haven't :>).

      Alternative possibility: Is it that people are using AI to craft their replies? Also yucky.

      • fidotron 10 minutes ago
        > Alternative possibility: Is it that people are using AI to craft their replies? Also yucky.

        Or they have OpenClaw using claude to karma farm, for . . . reasons.

        The amount of absolutely bizarre comments you get on here these days is nuts. It always used to be a bit special, but this year is something else.

      • ZeWaka 10 minutes ago
        > Is it that people are using AI to craft their replies?

        Very likely at least one person is.

    • wgd 25 minutes ago
      > likely human

      X to doubt. I think the more parsimonious explanation is that a lot people are lazy and/or uncaring, and given the option to avoid thinking while still producing a vaguely equivalent [1] output they're taking it.

      [1] For a loose definition of "equivalent" where they clearly either don't notice the Claudeslop or don't care enough to avoid it, of course.

      • phoronixrly 17 minutes ago
        Oh I am doing it on purpose lately here in HN, and it's hella hard to sound as dumb, repetitive and cliche as Claude.
  • datadrivenangel 42 minutes ago
    • thelaxiankey 28 minutes ago
      Haha :/ one of the sad things for me is that prior to LLMs I had a fairly LLMish writing style. Kinda formal, lots of bullet points and em dashes, two instances before etc, etc. since the debut of gpts I've had to adjust my style to be marginally more opaque so that I don't get mistaken for an LLM
  • shibel 38 minutes ago
    Love it (the mockery, that is). So far Opus 5.5 feels noticeably better in this regard (oh, and it’s fast!), but I felt the same about Fable initially. Not buying into the nerfing theories yet but if I find myself going into LLM-tilt again soon then I might join the club.
    • wgd 9 minutes ago
      Opus 5.5 is better but by no means good at writing IMO. Still loves "noun verbs the object" sentences where the verb is an inappropriate and imprecise metaphor. Still loves negating premises nobody previously advocated for. Still loves incorporating irrelevant details of a conversation or investigation into comments and documentation. I don't think I've seen "load-bearing" myself yet but there's still a lot of things landing and being held.
  • lexh 28 minutes ago
    They missed the first legitimate use case for infinite scroll.
  • porphyra 10 minutes ago
    This is hilarious of course but Opus 5.5 just came out and largely fixed a lot of the Claudlish.
  • glalonde 5 minutes ago
    This is genuinely insightful.
  • enjoyyourlife 11 minutes ago
    It seems to be that it has a lot of meta analysis studies in its training data
  • jldugger 44 minutes ago
    So like, who writes like this and how do we delete them from the training corpus?
    • mohamedkoubaa 7 minutes ago
      Do we know if it in the training data or if the humans giving rewards in the labs have such awful taste in prose?
      • capnrefsmmat 2 minutes ago
        It's not the training data. When we've tested the writing style of base models that have not gone through instruction tuning, they're much more human-like (https://arxiv.org/abs/2410.16107). The style shift seems to come from something in the instruction-tuning process, so our current research problem is figuring out what in the process is doing it.
    • nxobject 41 minutes ago
      At this point, training on public codebases created by "You're Absolutely Right!"-era Claude...
  • nxobject 38 minutes ago
    I laughed, and then I cried, and then I laughed again. Bravo. I'm going to go vomit while trying not to choke it up with laughter.
  • kvirani 1 hour ago
    It drives me bananas. This post is so triggering.
  • TomGarden 19 minutes ago
    Couldn't get through it, it hurts too much
  • i2talics 39 minutes ago
    It would be hilarious if this was generated by a markov chain or something so it would infinitely scroll. Infinite Claude.
  • everlier 39 minutes ago
    Fable 5.1 makes everything a "Rung". These statistical markers are very annoying.
  • esafak 16 minutes ago
    What I want to know is, what would the embodiment of Claudish sound like? Some LinkedIn jock, for sure.
  • WSSP 24 minutes ago
    I was disappointed that refreshing the page didn't regenerate the text

    I suppose an infinite scroll _would_ be funnier

  • Tadpole9181 22 minutes ago
    Good lord, this is several weeks late at this point. The joke is so damn worn out at and I'm personally far more tired of seeing people online thinking they're clever for saying "load-bearing" for the millionth time than I am at Claude doing it. At least it has an excuse.
  • natsucks 17 minutes ago
    this is amazing
  • rvz 15 minutes ago
    This entire article is complete AI slop, with the most basic Claudisms present without even needing extensive inspection or needing to use pangram.

    I thought HN would be smart enough to not get easily tricked into upvoting such low quality articles like this, but because it mentions "Claude", it is worthy of some intellectual curiosity.

    Come on, do better.

  • caaqil 17 minutes ago
    This is bad. 90% of the HN front page is just pure AI slop on any given day, might be time to start adding penalties for AI stories on HN ala NSA era.