Raw LLM Responses

Watching Andrej Karpathy’s latest video, I really like how he described using LLMs: “When you use ChatGPT, it is a simulation of an average human labeler from OpenAI…” Understanding the training process. After pretraining (compressing internet with loss of information into the model weights), the model goes through supervised fine tuning, or imitation training to mimic the labels (chat conversations - or prompt response) which initially were done by human annotators (obviously now it’s a mix or purely synthetic - but still human curated). Thus since it’s an imitation of an artifact from a being with consciousness (text from a human), and given the rise of the internet and this new form of human-human interaction solely through text, I don’t find it hard to believe that when we look at the text generated from it it can trick that mechanism in your brain that activates when interacting with another human. Not knowing this, I can see why some people really believe it’s sentient. However I find that with this information, we can change the way we think (real time learning) and how we approach interacting with ChatGPT. Like OP said, just remember the differences, and recognize its stateless intelligence, and this doesn’t mean we failed, as it is an incredibly powerful mental automation tool that will be probably used in a larger system (society of minds) that aims to bridge the gap between computer intelligence and ourselves.