What we lack is an AI equivalent of a theory of aerodynamics. At the moment we're just brute forcing random solution when we have perfectly good systems we should be trying to reverse engineer: biological brains. This will likely require genetic & evutionary algorithms, a developmental genotype-phenotype mapping, and competitive ecological systems for open ended evolution. Right now we have back propogration which is effectively reinforcement learning--just one piece of the puzzle. I actually applied to Hinton's lab in 2009 for a postdoct position to address these very issues. He said they didn't have any money and were not taking new postdocs.