Conversation

Notices

1. Embed this notice
   Karl Fogel (kfogel@kfogel.org)'s status on Tuesday, 03-Dec-2024 02:20:08 JST Karl Fogel
   LLMs use an N-dimensional embedding to represent each word. The set of all words in the given corpus thus exists in an N-dimensional space. I want an N-dimensional Voronoi diagram for that set, and a language whose words all fall only on the boundaries in that Voronoi diagram.
   
   Then I want to travel to a land where people speak that language natively.
   • Embed this notice
     Karl Fogel (kfogel@kfogel.org)'s status on Tuesday, 03-Dec-2024 10:20:21 JST Karl Fogel

     in reply to

     • soaproot
     @soaproot Deep question! It depends how much vocabulary-level semantic similarity there is across human languages, I guess.
   • Embed this notice
     soaproot (soaproot@sfba.social)'s status on Tuesday, 03-Dec-2024 10:20:22 JST soaproot

     in reply to

     @kfogel Wouldn't such a situation be highly unstable, with almost every word very quickly undergoing semantic drift and other related linguistic phenomena? This happens all the time in any language, for example when words are too ambiguous for comfort.

   • Embed this notice
     soaproot (soaproot@sfba.social)'s status on Wednesday, 04-Dec-2024 00:26:43 JST soaproot

     in reply to

     @kfogel Ah yes, I hadn't quite thought of the cross language aspect. For example, because the corpus is trained on English it will have separate blue and green, and there might well be a word for grue on the boundary between them. This could be a perfectly stable situation if the target language combines two English words into one word (as many natural languages do in the grue example). (Disclaimers: not my field, not sure I've captured things like the N dimensions fully, etc).

   • Embed this notice
     Karl Fogel (kfogel@kfogel.org)'s status on Wednesday, 04-Dec-2024 00:26:43 JST Karl Fogel

     in reply to

     • soaproot
     @soaproot That's a great example -- exactly the sort of thing I'm thinking of, yeah.
   • Embed this notice
     Alexandre Oliva (moving to @lxo@snac.lx.oliva.nom.br) (lxo@gnusocial.jp)'s status on Wednesday, 04-Dec-2024 08:33:41 JST Alexandre Oliva (moving to @lxo@snac.lx.oliva.nom.br)

     in reply to
     please pass the mushrooms, they look delicious!
     
     (meaning awesome out-of-the-box thinking! :-)
   • Embed this notice
     Karl Fogel (kfogel@kfogel.org)'s status on Thursday, 05-Dec-2024 01:30:28 JST Karl Fogel

     in reply to

     • Alexandre Oliva (moving to @lxo@snac.lx.oliva.nom.br)
     @lxo Aw, a high compliment -- thank you, @lxo.
     Alexandre Oliva (moving to @lxo@snac.lx.oliva.nom.br) likes this.