Keeping it Together

It is essential to build systems that solve multiple tasks, and are to at least some degree cross domain.
One way to do build a Turing test machine is to build a really huge neural net and see how it works. I don't think that would work.
Instead, we'd like to build a series of systems and take lessons from each.
The key will be for the system to do a range of tasks.
Perhaps a simple robot, or a video game agent is a good place to start.
Do you have any ideas about a good small system that functions in a couple of real world domains.
After this first system, the next system should add features on to the system.
Evaluation and refinement should proceed iteratively with more and more complex systems.
A particularly interesting hurdle is when to introduce language.
These systems should provide us with real world experience, but also enable us to improve our computational theory, and improve our understanding of how neurons work together.