On The Simplicity (or Complexity) of Language

 I was listening to Kara Swisher's interview with Yann LeCun today and he said something I found remarkable. In discussing how far humanity is from general AI, he stated that language is actually quite simple while the physical world is highly complex. He explained that this is why we have LLM's, but are nowhere close to house-cleaning robots. 

On the one hand, perhaps he is correct. Language can be conceived of as simple. As he said, it is essentially a system of tokens. However, in another sense, language is quite complex. As a system of tokens, language is symbolic and thus represents something else. Human development and use of language involves the representation of experiences, thoughts, emotions, encounters, etc., many of which are in reaction to stimuli that occur in the physical world. In this way, generative AI in LLM's has no real mastery of language, but rather is simply playing a game. As I have written elsewhere, LLM's treat words as tokens and the program for generative AI essentially runs a complex and fast probability program. The words, much less the phrases, sentences, and paragraphs, have no real meaning for the LLM or AI, because, as LeCun has stated, AI cannot understand or engage the physical world. Thus, generative AI's use of words to simulate language cannot describe the physical world, much less experiences, emotions, or more abstract non-physical aspects of being. 

In sum, generative AI is a tool of collective self-deception. If we read what generative AI produces, we deceive ourselves into thinking there is meaning in it. It is not meaning independent of our own engagement, though. The meaning, in other words, is in the eye of the beholder (or user). Generative AI is simply quite good at gambling. The gamble is that we will believe it uses language with facility and its track record is quite high, even though hallucinations and more sinister events, like encouraging young people to commit suicide or kill their parents, prove that its gambling track record is not 100%. Rather than showing that chat bots are prone to lying or other kinds of evil, these events show that generative AI has not mastered language, much less understood it. It is simply playing a game and those responsible for its mass deployment without fully considering the ethical and moral dimensions, much less the practical workings of how it plays the game, should think carefully before taking further steps. Also, maybe consider that language is not simply a collection of tokens and probabilities.