They probably won’t advance much because currently it has two opposite but equally difficult problems. On the one hand, AI still hasn’t achieved sensor integration, or creating an ontologically sound world model that includes more than one sensor data stream at a time. Right now it can model based on one sensor or one multidimensional array of sensors. But it can’t model in-between models. So you can’t have, let’s say, one single model that can hear, see light and radar at the same time. The same way that animal intelligence can self-correct their world model when one sensor says A but another sensor disagrees and says B. Current models just hallucinate and go off the deep end catastrophically.

On the opposite end, if we want them to be products, as seems to be MS and Altman fixation. Then it cannot be a black box, at least not for the implementers. Only in this past year there have been actual efforts to really see WTF is going on inside the models after they’ve been trained and how to interpret and manipulate that inner world to effective and intentional results. Even then, the progress is difficult because it’s all abstract mathematics and we haven’t found a translation layer to parse the model’s internal world into something humans can easily interpret.