• 0 Posts
  • 36 Comments
Joined 1 年前
cake
Cake day: 2023年7月3日

help-circle






  • I know this shouldn’t be surprising, but I still cannot believe people really bounce questions off LLMs like they’re talking to a real person. https://ai.stackexchange.com/questions/47183/are-llms-unlikely-to-be-useful-to-generate-any-scientific-discovery

    I have just read this paper: Ziwei Xu, Sanjay Jain, Mohan Kankanhalli, “Hallucination is Inevitable: An Innate Limitation of Large Language Models”, submitted on 22 Jan 2024.

    It says there is a ground truth ideal function that gives every possible true output/fact to any given input/question, and no matter how you train your model, there is always space for misapproximations coming from missing data to formulate, and the more complex the data, the larger the space for the model to hallucinate.

    Then he immediately follows up with:

    Then I started to discuss with o1. [ . . . ] It says yes.

    Then I asked o1 [ . . . ], to which o1 says yes [ . . . ]. Then it says [ . . . ].

    Then I asked o1 [ . . . ], to which it says yes too.

    I’m not a teacher but I feel like my brain would explode if a student asked me to answer a question they arrived at after an LLM misled them on like 10 of their previous questions.




  • I feel like the Internet Archive is a prime target for techfashy groups. Both for the amount of culture you can destroy, and because backed up webpages often make people with an ego the size of the sun look stupid.

    Also, I can’t remember but didn’t Yudkowsky or someone else pretty plainly admit to taking a bunch of money during the FTX scandal? I swear he let slip that the funds were mostly dried up. I don’t think it was ever deleted, but that’s the sort of thing you might want to delete and could get really angry about being backed up in the Internet Archive. I think Siskind has edited a couple articles until all the fashy points were rounded off and that could fall in a similar boat. Maybe not him specifically, but there’s content like that that people would rather not be remembered and the Internet Archive falling apart would be good news to them.

    Also (again), it scares me a little that their servers are on public tours. Like it’d take one crazy person to do serious damage to it. I don’t know but I’m hoping their >100PB of storage is including backups, even if it’s not 3-2-1. I’m only mildly paranoid about it lol.










  • Oh whoops, I should have archived it.

    There were about 7 images posted of users roleplaying with bots, all ending with a bot response that cut off halfway with an error message that read “This content may violate our policies; blablabla; please use the report button if you believe this is a false positive and we will investigate.” The last one was some kind of parody image making fun of the warning.

    Most of them were some kind of romantic roleplay with bad spelling. One was like, “i run my hand down your arm and kiss you”, and the bots response triggered the warning. Another one was like, “*is slapped in the face* it’s okay, I still love you” and the rest of the message generated a warning. There wasn’t enough context for that one, so the person might have been writing it playfully (?), but that subreddit has a lot of blatant sexual violence regardless.