It's still just predicting the next token, it's just using more past data points than your keyboard. The rest of the phenomena are emergent from that. I think it's important to keep that in mind given how much they can imitate human reasoning.