Once you’ve trained a chatGPT, you’re feeding it the words, the feeding of words so far. It ripples through a big neural net with 175 billion weights, telling you the probabilities for the next word. It’s a shallow computation, a few hundred layers deep. Given a word, it ripples through and figures out probabilities for next words. When the world is full of AI computation, it’s nothing new because the world is already full of computation in nature. The question is how we interact with those AI systems that we constructed.
ChatGPT takes a trillion words from the web and uses them to predict the next word based on what it has seen. It discovers patterns, including syllogistic logic, and creates a semantic grammar of language. It can generate coherent text by looking back at the words it has generated before. It’s a rather shallow computation compared to typical irreducible computations.
The reason chatGPT works is that it taps into a semantic grammar that humans use, a further version of logic. It constructs text based on the patterns it finds in the text it reads. It can be used as a linguistic interface to generate essays or other text based on a set of bullet points.
In a world run by AI, there will be computation that is incomprehensible to us. We already exist around nature, which runs computations that we don’t fully understand. The interaction between humans and AI is a different question, but the world being full of AI computation is not new. It’s about how we perceive and interact with the AI systems we have constructed.
