Solving prediction of behavior is not enough, we should also ask ourselves why we want to predict behavior, and what if any steps do we want to take to cut of the behavior we’re likely to take next at the pass. The question is what do we do with prediction. I’ve been contemplating what it means to build an accurate predictive model particularly for text, even more specifically utilizing word embeddings for doing named entity recognition. Part of the word embeddings model is predicting the likelihood that a word appears next to another. If you step out a level you could apply the same methodology to predict the next most likely sentence etc… Switch this around to creating sequences of physical actions in the real world, and you can predict likely next actions or predict the context in which the actions are likely to occur . Given sensor data you might extend this even further, from very basic this person is probably on a bus headed across town that’s stuck in traffic, to monitoring pulse and the surrounding bio signs, and using this information to detect that you’re in an argument on the bus or you’re flirting on the bus. Given that the language used to describe and collect the context is the same all of this is possible. The question is what do we do with that information. Do we advertise to you , prompt the next items in your todo list, or do we be bold and take a new action, do we manipulate the future actions you take to lead you into a particular direction ? I don’t actually have any answers, but these are the types of questions we should be asking . If you think we should take the next action preemptively for you towards the thing that you want, I would suggest that this has its limitations in the human that must precisely know what they want. This means that what the person wants is actually what the person wants, which when dealing with humans is rarely the case.