Conversation

Notices

1. Embed this notice
   The_Gibson :veilid: (thegibson@hackers.town)'s status on Tuesday, 25-Apr-2023 11:55:57 JST The_Gibson :veilid:

   ELI5

   What is up with the LLM hallucinations?

   I am assuming this is a fancy name for the byproducts of fuzzy logic capabilities... or is there more at work here?

   Any good primers would be appreciated.

   • Embed this notice
     silverwizard (silverwizard@convenient.email)'s status on Tuesday, 25-Apr-2023 11:55:55 JST silverwizard

     in reply to

     • Risotto Bias
     @thegibson @risottobias I want to be clear that it *always* writes fiction.
     
     ChatGPT can't run a function nor can it know what a function is. It just runs down a path in a tokenizer, it doesn't really run code or whatever .
   • Embed this notice
     The_Gibson :veilid: (thegibson@hackers.town)'s status on Tuesday, 25-Apr-2023 11:55:56 JST The_Gibson :veilid:

     in reply to

     • Risotto Bias

     @risottobias hmmm

     Except it doesn't sometimes. When asked to do a specific function and then reverse that function, the answers don't always match, for example.

     This is not a cute hallucination, but sort of befuddlingly useless.

   • Embed this notice
     Risotto Bias (risottobias@tech.lgbt)'s status on Tuesday, 25-Apr-2023 11:55:57 JST Risotto Bias

     in reply to

     @thegibson AI is built on a corpus of confident sounding responses.

     it is trained to sound confident, even if there's no basis in reality.

     the sentences that it tunes to are meant to seem realistic.

     nothing about it is unit tested to whether it really works.

     it writes fiction.

   • Embed this notice
     The_Gibson :veilid: (thegibson@hackers.town)'s status on Tuesday, 25-Apr-2023 20:26:15 JST The_Gibson :veilid:

     in reply to

     • silverwizard
     • Risotto Bias

     @silverwizard @risottobias it gets even less impressive as I dig.

     silverwizard likes this.