Raw LLM Responses

Interestingly there is recent research, that might suggest otherwise. Basically, researchers at Princeton University and Google DeepMind developed an evaluation method to determine the extend of LLMs being stochastic parrots and found out that "many of GPT-4's answers on SKILLMIX involve combinations of skills and topics that it had not seen in the training corpus" [Arora, 2023]. Obviously, further research needs to be done. https://pli.princeton.edu/blog/2023/are-language-models-mere-stochastic-parrots-skillmix-test-says-no https://arxiv.org/abs/2310.17567 https://www.quantamagazine.org/new-theory-suggests-chatbots-can-understand-text-20240122/ https://dl.acm.org/doi/10.1145/3442188.3445922