The Emergence of Self-Awareness

The way GPT works is predicting basically your most likely continuation of a given set of text. So if you tell it, you know, compressivist, it can predict what seems to be a good compression. And then since that is the likely compression based on the model, presumably if you feed it back, if the likely decompression is pretty correlated. It's not obvious where we're just doing something that looks right, or if it's actually good at this.