I agree with your point about how to best use it today. We have seen that each new model generation both improves the prior tasks and unlocks new ones through e

I agree with your point about how to best use it today. We have seen that each new model generation both improves the prior tasks and unlocks new ones through emergent behavior. That’s the fascinating/scary part of this development. And yes, it’s “just” a language model. It’s “just” predicting next token given training + context. We don’t really understand why it’s working and it’s evolving non-linearly. I asked GPT-4 to give me an SVG map of my town. I then asked it to put dots on some local landmarks. The map was toddler level, but the landmarks were relatively accurate in terms of their relationship to each other and the blob that it drew. So this is a language model that has some emergent notion of space in its code generation abilities.