The Future of AI: Multimodal Language Models and Understanding.

AI researchers are divided on whether large pretrained language models like GPT-4 can understand language. While they can coordinate on a shared meaning, they likely don't have generative or constructive understanding. These language models are trained on a huge but restricted domain, unlike human understanding, which maps textual experience onto somatosensory experience. Humans apply mechanical reasoning to things, while GPT-4 doesn't have a physics model. The narrative that these language models have rediscovered human reasoning is misguided and demonstrably false.