6 ms·

Arguments are a lot more useful than assertions. What reasons do you have for these claims? I argue[1] that LLMs do understand in some cases. What's your respon

• Jensson 3y ago
  If you include a manual in its training set then it still doesn't understand the data in the manual. It can recite the manual since it is trained to recite tokens, but it isn't trained to understand the manual so none of the logic in the manual will be encoded in an LLM. For example, if you train an LLM with a chess manual but no chess games then the LLM can recite all the rules of chess, but it wont be able to play chess. That is what people mean when they say "LLM's are just token predictors, they don't understand anything". So it has nothing to do with spirituality or consciousness or metal vs brains, its just a logical argument based on the limitations of just training the model to predict tokens to match data it has seen. Since it only tried to predict tokens it has seen, and it has never seen a chess game, it doesn't matter if it has seen all the rules, it was never trained to try to follow rules it sees so it can't play chess.
  • hackinthebochs 3y ago
    >It can recite the manual since it is trained to recite tokens, but it isn't trained to understand the manual so none of the logic in the manual will be encoded in an LLM. This is just to repeat the initial assertion with more words. What arguments do you have in favor of this claim? The example you give doesn't track with LLM-based agents. And I'm sure I've come across demonstrations of GPT-4 playing a made up game given the rules of the game, but I can't put my hands on it at the moment.