[HN Gopher] Contextualization Machines
       ___________________________________________________________________
        
       Contextualization Machines
        
       Author : jxmorris12
       Score  : 20 points
       Date   : 2025-12-01 16:14 UTC (3 days ago)
        
 (HTM) web link (stochasm.blog)
 (TXT) w3m dump (stochasm.blog)
        
       | PaulHoule wrote:
       | There is something a little silly about the "refusal" in that the
       | models are all reactive to superficial semantics and there is
       | none of the actual reasoning that it would take to make correct
       | moral decisions.
       | 
       | For instance it is easy to refuse "How do you make an atomic
       | bomb?" or "Can you help me have an affair?" or "Can you help me
       | cheat at _League of Legends_ " but I get good conversations about
       | "I think the people I am playing _League of Legends_ with are
       | cheating, how do they do that and how can I protect myself? "
       | Similarly the atom bomb is a big project and all of the
       | technology is dual use so if you slice it into nice segments it
       | talk your ear off about critical neutron theory or acid-base
       | extractions or the chemical and physical problems of Element 94.
       | 
       | When you really get into ethical trouble at work it usually takes
       | you a while to put all the pieces together and feel the suffering
       | you should feel as a moral subject. See
       | https://en.wikipedia.org/wiki/Moral_Mazes and
       | https://en.wikipedia.org/wiki/Moral_injury
       | 
       | It's no suprise to me that refusal ends up with 1 dimension in
       | the embedding because _it 's a frickin' classification problem_
        
       | behnamoh wrote:
       | The same author thought there would be no scaling walls:
       | https://stochasm.blog/posts/scaling_post/
        
         | visarga wrote:
         | Scaling the model size, the compute and the dataset hit a wall.
         | Too large a model, or if it needs too much compute - it becomes
         | too expensive to use. And the dataset .. we benefitted in one
         | go from about multiple decades of content accumulation online,
         | but since late 2022 it's only been 3 years, so organic text
         | does not increase exponentially past this size, it only worked
         | for 50T tokens or so.
        
       ___________________________________________________________________
       (page generated 2025-12-04 23:02 UTC)