Yeah, I don't know people ever thought that the current state of LLMs is enough to actually solve problems; it's trained to be very good at predicting the next

Yeah, I don't know people ever thought that the current state of LLMs is enough to actually solve problems; it's trained to be very good at predicting the next word, but any logic solving capability is only emergent due to questions + answers in the training data, when it has to improv then it either makes a randomly correct or similar enough guess, or it falls over. If these things were trained to actually solve logic problems then it may be different, the training data would have to be logic problems. They've started at the top (language) rather than at the bottom/middle (logic, thought, reasoning) which should be done first - and then the model can be trained/additional models to get it to put thoughts into words.