Predicting the Next Word

There's a big argument in the field about just kind of what level of description is right for these. There was a paper I really liked from Jacob Andreas at MIT that sort of took this interesting lens on what models are doing. It's saying that once these models get capable enough, they're what he calls agent models.