Re:LLMs trained on “A is B” fail to learn “B is A”
by hackinthebochs on Saturday September 23, 2023 @10:30AM
(#37623565)
>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.
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.