I think its about automatically building accurate and well-factored world models online that ultimately integrate not only high-dimensional sense data (such as visual information) but also language. This involves effectively solving the symbol grounding problem among other things. There is some serious effort in this direction in deep learning.
There are also other efforts using different types of probabilistic programming as well as symbolic and neural net combinations.
There's another link on one of the first few HN pages right now about dreaming. I think that dreaming gives one a lucid demonstration of some of the capabilities that we need to emulate if we are going to have human-like intelligence. AI will need to be able to visualize new situations, basically like on-demand, flexible simulations of mashed-up possibilities, involving things like physics and psychology etc.
I think we almost need the AI to have something like a 3d gaming engine with physics, but also it can effortlessly conjure up AI agents in this simulation, but also, many of the physics rules and behaviors of the AI agents are automatically learned with only a few examples. This is the type of capability that allows humans (and some other animals) to adjust so readily to new situations.
I speculate that there may be some representation or type of computation that has not been invented yet which facilitates both the simulation-type data and also the abstractions over it, all the way up to language, in a more seamless way than has so far been described. I saw a paper talking about the symbol grounding problem in terms of everything being categories, but really in the end it was broken down into something kind of like Lisp + probabilistic programming, and it seemed to not really have sufficient granularity to really do justice or properly integrate sense data. Certainly not in a seamless or truly unified way in my opinion. Although I guess I don't really understand category theory.
There are also other efforts using different types of probabilistic programming as well as symbolic and neural net combinations.
There's another link on one of the first few HN pages right now about dreaming. I think that dreaming gives one a lucid demonstration of some of the capabilities that we need to emulate if we are going to have human-like intelligence. AI will need to be able to visualize new situations, basically like on-demand, flexible simulations of mashed-up possibilities, involving things like physics and psychology etc.
I think we almost need the AI to have something like a 3d gaming engine with physics, but also it can effortlessly conjure up AI agents in this simulation, but also, many of the physics rules and behaviors of the AI agents are automatically learned with only a few examples. This is the type of capability that allows humans (and some other animals) to adjust so readily to new situations.
I speculate that there may be some representation or type of computation that has not been invented yet which facilitates both the simulation-type data and also the abstractions over it, all the way up to language, in a more seamless way than has so far been described. I saw a paper talking about the symbol grounding problem in terms of everything being categories, but really in the end it was broken down into something kind of like Lisp + probabilistic programming, and it seemed to not really have sufficient granularity to really do justice or properly integrate sense data. Certainly not in a seamless or truly unified way in my opinion. Although I guess I don't really understand category theory.