Conversation

Notices

1. Embed this notice
   pridesvrld (esvrld@normal.style)'s status on Saturday, 03-May-2025 03:00:17 JST pridesvrld

   telling an llm 'do not hallucinate' does not work b/c not hallucinating would require for it to be able to
   (1) examine its own output
   (2) extract the semantic meaning of the same
   (3) compare that meaning to some external body of data
   when the reason it 'hallucinates' is precisely that it does not have those capabilities

   • Embed this notice
     pridesvrld (esvrld@normal.style)'s status on Saturday, 03-May-2025 03:00:16 JST pridesvrld

     in reply to

     this is incredibly simple but a lot of people have difficulty grasping it. it's not that llm's are stupid, it's not that they're built wrong, it's not that they don't work (tho all of those are also true); it's that what they push out is fundamentally meaningless in a particularly rigorous sense

   • Embed this notice
     pridesvrld (esvrld@normal.style)'s status on Saturday, 03-May-2025 03:00:17 JST pridesvrld

     in reply to

     and likewise for anything else you might want to tell an llm, like 'follow this style guide', 'do not claim you have capabilities you don't', or even telling it to not use a certain word or expression in its answer. because it's not a machine that 'knows' anything. what you're doing — all you're doing — when prompting an llm is requesting it output a string of tokens that would be statistically likely to follow from your prompt, per its corpus. and that's all it does

     silverwizard repeated this.
   • Embed this notice
     silverwizard (silverwizard@convenient.email)'s status on Saturday, 03-May-2025 03:00:38 JST silverwizard

     in reply to
     @esvrld Honestly though! Fucking seriously!
     ChatGPT was such a good fucking tool for bypassing people's brains. They see the form of language and assume an internal entity. Honestly, the conversational form plus the word artificial intelligence makes it so easy to bypass mental models.
   • Embed this notice
     wizzwizz4 (wizzwizz4@fosstodon.org)'s status on Sunday, 04-May-2025 23:20:43 JST wizzwizz4

     in reply to

     • Bee O'Problem :godot:

     @esvrld @beeoproblem Mathematically, they're really huge Markov models, compressed in a clever way.

     silverwizard likes this.
   • Embed this notice
     pridesvrld (esvrld@normal.style)'s status on Sunday, 04-May-2025 23:20:44 JST pridesvrld

     in reply to

     • Bee O'Problem :godot:

     @beeoproblem mathematically there might be some massive difference i'm not aware of as a non-expert on that field, but linguistically, yea absolutely

   • Embed this notice
     Bee O'Problem :godot: (beeoproblem@mastodon.gamedev.place)'s status on Sunday, 04-May-2025 23:20:45 JST Bee O'Problem :godot:

     in reply to

     @esvrld I remember some people getting a bit offended when I called LLMs fancy Markov chains but everything I see them produce has failed to convince me otherwise