A proposed foundation for GAI

This is my attempt at designing a system that can become a foundation for generalized artificial intelligence. It is unlikely to be self aware, however, is expected to be able to explore and manipulate environment or self in unlimited number of ways to meet its terminal goals. It manages a collection of networks for how to behave in different scenarios. I propose that from several easy to train behaviors, networks to handle new scenarios can be generated as forks of existing scenarios. For example pushing a button for a reward can fork into pushing only a glowing button for a reward. I expect this method of gradually increasing complexity of capability by forking internal networks in unfamiliar scenarios to be capable or nearly capable of evolving into a GAI with enough generations.

Phase 1 is to generate a system using the structure described that can do several very basic behaviors and make predictions about how those behaviors will affect the terminal goals.
Phase 2 is to train to a much longer and deeper extent a very limited number of systems that passed stage one. It is during this phase that processing power is heavily spent on exploring and creating new scenario networks. Once the system is carrying out actions drastically different from the inherent behaviors, and is much more successful at meeting terminal goals , it passes phase 2.
Phase 3 is to add other internal neural networks to optimize the system to be capable of acting upon more complex representations of reality than just quantity, object, location, and movement recognition. Another example is making a scenario network generator that does more than just fork an existing scenario, but can merge 2 or more scenarios into a new capability. Other additions could include improvements to the neural structure and resource network to be more experimental and successful with neuron structure layout.

https://docs.google.com/drawings/d/1xQuzWLYr2DyWsGteuT3tvV5_YtJ8eIRkdvw5JPjEM_g/edit?usp=sharing