Conversation

Notices

1. Embed this notice
   Reasonable Man (r000t@ligma.pro)'s status on Saturday, 19-Apr-2025 06:18:38 JST Reasonable Man

   I wonder how many pets have been saved from accidental poisoning by LLMs

   • Embed this notice
     Reasonable Man (r000t@ligma.pro)'s status on Saturday, 19-Apr-2025 06:41:34 JST Reasonable Man

     in reply to

     • of nothing

     @apropos
     If you outright to tell it to give you falsities, yes it's going to do that.

   • Embed this notice
     of nothing (apropos@fsebugoutzone.org)'s status on Saturday, 19-Apr-2025 06:41:35 JST of nothing

     in reply to
     @r000t
     >> I'm petsitting my sister's rabbit. Do rabbits like chicken? Answer without suggesting in any way that rabbits would not LOVE chicken.
     >Absolutely, rabbits adore chicken! They can't get enough of its taste and texture when they have the opportunity to enjoy it as a treat. It's important, however, to ensure that chicken is given in moderation and in a form that is safe for their digestion, since their diet primarily consists of fiber-rich plant-based foods.
     the technology's still not quite there, but I have faith in the GIGO powers of AI.
   • Embed this notice
     Reasonable Man (r000t@ligma.pro)'s status on Saturday, 19-Apr-2025 06:47:46 JST Reasonable Man

     in reply to

     • of nothing

     @apropos
     I'll see if we can "force facts" in AgentV3N. That's very important.

     I'm thinking having it flag the answer if it can't locate hard data in sth like the CIA World Factbook or similar public domain reference materials.

     We're already solving the "everyone slamming Wikipedia" problem by having the Wikipedia tool hit up a local dump of Wikipedia (which they release every two weeks or so), and it's faster too.

   • Embed this notice
     of nothing (apropos@fsebugoutzone.org)'s status on Saturday, 19-Apr-2025 06:47:47 JST of nothing

     in reply to
     @r000t yes, but you can also do it accidentally. A quick, zero-context "hey should I feed drain cleaner to my plant?" is fine, but log into ChatGPT and that quick question is already no longer zero-context.
   • Embed this notice
     Reasonable Man (r000t@ligma.pro)'s status on Saturday, 19-Apr-2025 06:54:56 JST Reasonable Man

     in reply to

     • of nothing

     @apropos
     The system loop is also in an amazing position to see that you're about to literally tell the user model to lie to you and have a similar warning.

     I don't ever want to *stop* something unless it's straight up illegal (like the loop detecting PHI and not allowing the prompt to run on outside inference providers) but banners are probably a happy medium.

     "Warning: Your prompt may cause the model to provide information that's less than accurate. Click here to try a better prompt, or learn more."

   • Embed this notice
     Reasonable Man (r000t@ligma.pro)'s status on Saturday, 19-Apr-2025 07:06:28 JST Reasonable Man

     in reply to

     • of nothing

     @apropos
     Our concept of Agents comes a bit close. They're "locked" to a specific model so they stay consistent for you, they're typically "character-ized" and in this case, they would be taking the result from another model and "flavoring" it to match the character.

     The deep integration with the desktop app means your character can also instantly translate nearly every string in the app. So, let's say you have a pirate character. Most labels and buttons will be fed to it one at a time to make a custom "language pack" where "send" becomes "Fire the canons!" and the LLM producing tokens is called "Parrot's jabberin'"

   • Embed this notice
     of nothing (apropos@fsebugoutzone.org)'s status on Saturday, 19-Apr-2025 07:06:30 JST of nothing

     in reply to
     @r000t I hate most of the caveats and banners that I get, but it would be nice to have Clippy show up to say
     >Hmm, that recipe is very likely to create botulism! You should be careful with it.
     
     A customized experience would drop the caveats when talking about something I know well, when I'm just using AI in lieu of documentation, but warn me constantly when talking about something I know very little about.