Comment on Peak Fiction đ
sad_detective_man@sopuli.xyz â¨3⊠â¨weeks⊠agoAnything to that effect would have been better. I just wanted to roleplay that I was Tommy Wiseau asking for a favor to leaving the mob and start my dream of a film career.
The thing just kept acting subby and asking me âso wadya want?â
WorldsDumbestMan@lemmy.today â¨3⊠â¨weeks⊠ago
Yeah, it just keept asking a question.
It is likely because it is some 3B lobotomized model, that genuinely needs to vamp off your ideas to work.
Even a basic 7B Stheno model will do way better.
sad_detective_man@sopuli.xyz â¨3⊠â¨weeks⊠ago
Eh, since then itâs come out that these compliment machines have sooth talked people into inceldom or killing their loved ones. I think they probably donât operate within that place of safety that was originally appealing to people. Or you know, the training data comes from other humans who are demonstrably not safe so weâll probably still be talking about dark romance lit for a long time
zarkanian@sh.itjust.works â¨3⊠â¨weeks⊠ago
I understand these words, but not in that order.
sad_detective_man@sopuli.xyz â¨3⊠â¨weeks⊠ago
Would you like the quick summary or the 30-min deep dive video?
Lots of people started using the chatbot as a surrogate partners over the last seven years. Since the thing is a frictionless hug-box it pulled them deeper into their solitude and trained them to only respond to interactions where the other person completely acquiesces or at the very least yes/ands them constantly. Which is pretty hard for a human to be friends with another person whoâs used to that. Hence, inceldom
inv.nadeko.net/watch?v=3WSKKolgL2U
WorldsDumbestMan@lemmy.today â¨3⊠â¨weeks⊠ago
Well, itâs been trying to talk me out of trying to raise my IQ with experimental supplements/using posiwash on literally everyone, but I literally just talk to it, ask for suggestions, then ignore all itâs opinions.
It is weird, especially since I get angry if it sycopants, and it has that in itâs memory.
Buddahriffic@lemmy.world â¨3⊠â¨weeks⊠ago
Youâre still personifying it too much. Whether or not you get angry doesnât matter to it. It has no concept of whether something matters. Itâs just predicting tokens based on the token correlations it encoded in its NN parameters during training plus a bit of random noise. Everything else is just trying to set up the context in the hopes that it predicts the tokens you want it to.
And since it wasnât trained on every possible context, itâs overfitted to some (eg maybe a mention of yellow will make it very likely to talk about bananas) and lacks any useful context at all for others (eg your very specific problem).