Conversation

Notices

1. Embed this notice
   Urusan (urusan@fosstodon.org)'s status on Sunday, 07-Jan-2024 00:11:28 JST Urusan

   Many of the defects of AI are really reflections of our own behavior.

   For instance, the hallucinations and the doubling down on them come from the training data. It's full of humans talking authoritatively. Sure, there is the occasional expression of doubt or ignorance, but the vast majority of the time the human voice is speaking confidently, so it mirrors that confidence.

   It's natural to not know something or make a mistake on occasion, but it learned that there are no mistakes.

   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Sunday, 07-Jan-2024 00:11:27 JST iced depresso

     in reply to
     @urusan they aren't capable of logical reasoning at all, they're just learning funny word order :blobcatgoogly:
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Sunday, 07-Jan-2024 00:13:28 JST iced depresso

     in reply to

     • iced depresso
     @urusan there was a paper about that which was linked on fedi. people have looked specifically in to lstm-ish bots doing logic and they do not demonstrate it; they basically demonstrate statistical heuristics or something similar.
     
     they would probably need to have an architecture which actually allows them to cogitate for a while before saying words. and i'm not sure if such a system would even be differentiable. (i would like to see how numenta's cortical models hold up to decision making tasks; i think they finished porting some of that to pytorch now, who knows if i'll feel like doing AI again this year.)