I think part one here is insightful, part two not so much :-). Whether or not AI has a model can affect its function, the self driving development teams have f

I think part one here is insightful, part two not so much :-). Whether or not AI has a model can affect its function, the self driving development teams have found just how much "intuition" people bring to the driving task (estimating what other cars might do, weighting behaviors like "edging over" to understand intent). The recent article about Marines defeating a sentry gun by holding boxes over their heads is another good example of the AI not having a model of the threat, just what "people" look like that are approaching. I was intrigued by the frustration an engineer shared with me about how simulators where all the drivers are AI's work so much better than ones where humans are driving too. It points to problem sets where a model of human behavior is essential to successfully achieving the objective. Law is not a particularly good example because it is practically program code these days. The human intent model is simpler[1] because it is based on one variable, "doing anything the law/contract is trying to prevent." [1] Important distinction here is that it isn't "simple" as in easy to do, it is merely "less difficult" than things like driver behavior or enemy soldier behavior.