The current methodologies that we call as AI are very narrow, they can't be a building block of an AGI.
The cognitive architecture of AGI would have multi-layered techniques, where each of them performs a variety of tasks like cognitive architecture from the University of Michigan which has declarative and procedural learning/long-term memories, their action selections, short-term memory, etc.
To process this kind of architecture, there needs to be abundant memory available.
Which is possible from the unused memory available in our smartphones and other electronic devices.
This work is primarily to build a framework on how we can train a simple deep learning task while having the learning shared between multiple devices and collectively building the weights.