>>/9072/
Yeah,it kind of backwards. They have the end result and the science part of it is moving backwards and having finer and finer understanding of the parts that make it up. Claude had an interesting thing on this. Something about isolating sets of weights to alter the behaviour of their llm in more detail, they had a post on this a while back. Also explaining how they found those weights if i recall correctly.its still not complete but from the backwards view they are getting close to a discrete understanding of the weights.