XPost: alt.folklore.computers   
   From: bowman@montana.com   
      
   On Tue, 30 Dec 2025 23:42:19 -0500, c186282 wrote:   
      
   >   NNs are a distinct tech from LLMs. In *theory* they   
   >    have great potential ... and could be rather compact,   
   >    fit inside a bot. However to realize this, special hardware is   
   >    required, elements that are sort-of like neurons. They ARE getting   
   >    there, I see articles from time to time.   
      
   https://en.wikipedia.org/wiki/Transformer_(deep_learning)   
      
   Transformers are based on NNs and are an improvement on recurrent neural   
   networks. I sort of miss one of my favorite terms 'long short term   
   memory'.   
      
   The wiki page starts off slow and then gets deep into the murk. The whole   
   field tends to be headache inducing. Humans do fine in three dimensional   
   Euclidean space with x, y, and z.  Throw in w for the fourth dimension and   
   the woo-woo starts particularly if it's interpreted as time. After that   
   it's all Cloud Cuckoo Land. The tensor operations that worked in 3 or 4   
   dimensions work in any number of dimensions since it's all math but let's   
   not think too hard about what it means.   
      
   Consider Word2Vec...   
      
   https://en.wikipedia.org/wiki/Word2vec   
      
   Three dimensions and you can visualize those words over in that corner of   
   the cube sort of hand together. 400 dimensions and you have sort of a   
   problem.  Then you get to the improvements on Word2Vec...   
      
   I've got a feeling that the people who are really, really good at this are   
   also really, really weird.   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)