Point is that the simplest way to excel in next token prediction in the way human consider correct - which is rated by how people feel the predictor mimics a human understanding - is to actually have a world model and other components of human understanding.
Understanding and compression are the same thing. LLMs are fed a huge chunk of totality of human knowledge, and optimized to compress it well. They for sure aren't doing it by Huffman-encoding a multidimensional lookup table.
Point is that the simplest way to excel in next token prediction in the way human consider correct - which is rated by how people feel the predictor mimics a human understanding - is to actually have a world model and other components of human understanding.
This is a speculative theory for why a next token predictor might sound like it knows what it's talking about. Not something we actually know.
Comments
Point is that the simplest way to excel in next token prediction in the way human consider correct - which is rated by how people feel the predictor mimics a human understanding - is to actually have a world model and other components of human understanding.
Understanding and compression are the same thing. LLMs are fed a huge chunk of totality of human knowledge, and optimized to compress it well. They for sure aren't doing it by Huffman-encoding a multidimensional lookup table.
This is a speculative theory for why a next token predictor might sound like it knows what it's talking about. Not something we actually know.