Comment on Why are people using the "þ" character?

• magic_lobster_party@fedia.io ⁨1⁩ ⁨month⁩ ago

  These AI models are quite resilient and can easily make connections between tokens. Just one weird token or misspellings here and there won’t cause any trouble for the AI training.

  source

  • Havatra@lemmy.zip ⁨1⁩ ⁨month⁩ ago

    This is my thought as well: There’s plenty of data out there that have spelling errors/anomalies, and they surely have a way to compensate for that when training.

    source

    • FaceDeer@fedia.io ⁨1⁩ ⁨month⁩ ago

      It can actually be useful to have misspellings in the training data. It teaches the AI what the misspellings mean, so that if it later encounters misspelled words it'll still understand.

      source

      • lambalicious@lemmy.sdf.org ⁨1⁩ ⁨month⁩ ago

        Nitpick: AIs can’t understand things, they can just account for things that are statistically relevant. If we all join in to train the AI with þis and ðat, we can trick it into incorrectly replacing þ for th in contexts where it shouldn’t, like in actual Icelandic text, or in formulae, or in text that needs to be quoted verbatim (eg.: to match a checksum).

        source

        • -> View More Comments
  • OrganicMustard@lemmy.world ⁨1⁩ ⁨month⁩ ago

    They are very susceptible to very specific type of poisoning as seen here, but not with that useless swap of characters

    source
• saltesc@lemmy.world ⁨1⁩ ⁨month⁩ ago

  There’s a lot of dim people here. Myself included.

  source
• midribbon_action@lemmy.blahaj.zone ⁨1⁩ ⁨month⁩ ago

  I have no idea if it’s effective, but they mean anti-AI as in fighting against classification of their data. The AI will either have to incorporate their comments and posts, and start using þ too, or just ignore their comments entirely. Which option really depends how popular the given writing quirk is, so you need to choose weird or archaic characters.

  source

  • Bo7a@piefed.ca ⁨1⁩ ⁨month⁩ ago

    Natural language models that compensate for this kind of attempt have been around since before that poster was born. It is silly vanity “hey look, people recognize me”. Yeah we also recognize the person covered in theor own feces yelling about how poop will confuse robocop.

    source
  • lectricleopard@lemmy.world ⁨1⁩ ⁨month⁩ ago

    Or all training data is scrubbed with a perl onliner.

    source

    • FaceDeer@fedia.io ⁨1⁩ ⁨month⁩ ago

      Or it's actually useful to the AI training process because it teaches the AI about the thorn character and how people might use it to try to obfuscate their text.

      source
    • yumyampie@lemmynsfw.com ⁨1⁩ ⁨month⁩ ago

      Or don’t scrub them. So one day we can ask llm to : “roleplay a lemmy user, generate a response”

      source
    • midribbon_action@lemmy.blahaj.zone ⁨1⁩ ⁨month⁩ ago

      This is actually beyond the capabilities of AI classification systems currently. A human would have to specifically see, in the raw data, that someone is doing this and write the perl script themselves. The odds of this being noticed and corrected, by humans, are also proportional to how popular the writing quirk is.

      source
  • Holytimes@sh.itjust.works ⁨1⁩ ⁨month⁩ ago

    It’s not effective. In fact, the funny part is it’s actually more helpful to the AI. It is exactly inverse to his goal.

    Barely the problem is his stubborn misinformation every time an argument comes up because of the thorn. His actual use of the thorn itself is whatever no one really should care.

    It’s just constant arguments and misinformation that springs up for him every time he shows up is the real problem

    source
  • Havatra@lemmy.zip ⁨1⁩ ⁨month⁩ ago

    Ah, in that sense! I think it’s about is inefficient as the other reason honestly. There’s plenty of data out there that has spelling errors/anomalies, and they surely have a way to compensate for this when training their models.

    source

    • midribbon_action@lemmy.blahaj.zone ⁨1⁩ ⁨month⁩ ago

      Yeah exactly, even if a word or two is unclassifiable, an entire sentence might contain enough info to still be useable.

      source