This is an extremely informative post. Thank you.

However, allow me to play devils advocate. Was it not long believed by deep learning skeptics that machines could never master language through big data alone because language refers to external things which the machines do not access to. In other words, humans don’t learn language by relating words to words. They learn language by relating words to things, to feelings, to actions, etc. Could one say that LLMs also lack access to that right kind of data, yet nevertheless, still managed to become quite proficient with language? Could that mean that with enough visual data, a robot could become dexterous as a human just as llms became sufficient with language even though they lack human senses and perceptions?