Start from first principles in the physical world. Do more work on what kind of processing we can do at the edge of the sensor and work our way up from there.

Start from first principles in the physical world. Do more work on what kind of processing we can do at the edge of the sensor and work our way up from there. For example, build a system that learns to have reflexes. That is, has processing close to or near a sensor and can work collaboratively with other sensors to learn (not explicitly programmed) to take action based on input without a central processing system. I would argue that if you can build a complex enough physical reflex learning system, then you have enough of the building blocks for a human level system.