The Path to Next Word Prediction in Language Models

There is still a big gap conceptually between the models today that are most visible, the language models, chat, GPT, and competitors they're off. It's not quite true that it's just next word prediction. That's the main way they learned. But they're also trained with human reinforcement learning. So already we've got things other than next word prediction shaping them. And you can imagine going further, you could imagine having large language models that are multi modal so that seeing visual input, and they're trying to do things like open doors and put things in drawers.