This is the shortest path I see towards machine intelligence: first, we develop ways to allow specialized AIs to manipulate formal concepts, write programs, run experiments, and at the same time develop mathematical intuition (even creativity) about the concepts they are manipulating. Then, we use our findings to develop an AI scientist that would assist us in AI research, as well as other fields. It would be a specialized superhuman artificial intelligence to be applied to scientific research. This would tremendously speed up the development of AI. At first we would apply it to solve well-scoped problems: for instance, developing agents to solve increasingly complex and open-ended games.
Well, computer scientists from the University of British Columbia and National University of Singapore just did that with a bipedal computer model (read: essentially a pair of animated legs) -- only instead of a cute cartoon rabbit, the teacher is a deep reinforcement learning artificial intelligence algorithm. Google's DeepMind, for example, has used reinforcement learning to teach an AI to play classic video games by working out how to achieve high scores. It's like watching your kid grow up -- except that, you know, in this case, your kid is a pair of disembodied AI legs powered by Skynet! A paper describing the work, titled "DeepLoco: Dynamic Locomotion Skills Using Hierarchical Deep Reinforcement Learning" was published in the journal Transactions on Graphics.