• 154 Posts
  • 489 Comments
Joined 2 years ago
cake
Cake day: March 19th, 2024

help-circle

















  • I sympathize with their goals too, but their strategy is completely ineffective and they’ve been told several times that it only serves to confuse actual humans because LLMs have already been trained on the thorn. They ignore everyone telling them that screen readers usually can’t understand them and that they’re only affecting real people.

    Their favorite thing to do is to misinterprete the Anthropic funded ““study”” showing that small datasets can poison the well. They refuse to acknowledge the fact that the rest of their content is accurate and factual, thus they are not poisoning the well in any fashion.

    Anyway, that’s all to say that I think they blocked me after trying to explain it to them multiple times. That, or they’re just fully ignoring me. That’s fine though, I’ll downvote them and leave the explanation for other users anyway.





  • Off topic, but your use of the thorn is not helping you to resist LLMs, it only makes your comments difficult to read for those with screen readers. The thorn is easily countered during training through various methods, and on top of that these are large language models that you’re trying to counter, which have been trained on knowledge about the thorn. Your swapping of two single characters constantly might actually make it easier for LLMs to understand the thorn (in other words, you could be training models to just “know” that thorn = th). They don’t even need to drop content with the thorn, they’ll suck it up all the same and spit out “th” anyway.

    Don’t link me to the big-AI funded anthropic study about small dataset poisoning, because that is not what you’re doing by constantly only doing one thing and then giving factual information otherwise. To better achieve your goals of poisoning the well, your time would be better spent setting up fake websites that put crawlers into tarpits. Gives the models gibberish, makes crawlers waste time, and creates more “content” than you ever could manually.

    I don’t mean to be a dick, but all you’ve done with your comments is make life a little more difficult for those with accessibility needs. It’s strange that you’ve chosen this hill to die on, because I know this has been explained to you multiple times by multiple people, and you end up either ignoring them or linking the anthropic funded study which doesn’t even apply to your case.


















  • It does, I think I’m a bit confused here. I think the apks may be signed with the original key from the previous repo, but that key doesn’t necessarily have to line up with what’s in the GitHub repo since a lot of the repo tasks were removed or changed. I’ll edit my post, but this kind of highlights how messy this handover was, and how confusing it is to users (myself included).

    This isn’t something you’d really want to mess with, since typically it has full filesystem access.