DeepMind just posted a mind blowing paper – achievement of transfer learning

Yes, please let's not have headlines like these. It might as well be "DeepMind just posted a mind blowing paper - You won't believe what happened next!" next time, so let's not go there.

Optimistic answer: each sub-network can specialize in a certain skill or element of knowledge, and then as a group, the networks can contribute to a larger decision. Such specialization allows skills to be highly developed but used only when they are applicable, instead of trying to force-fit a certain skill onto an inappropriate problem domain. The human brain is believed to work in a similar way with multiple regions processing input in parallel and "voting" for the appropriate response, with other parts of the brain coalescing and merging their output into the final response.

Cynical answer: Neural net machine learning has reached its peak, and there's far more computing power than the researchers know what to do with. Add additional layers of processing and if the numbers change, that's a paper.

They're gonna be yuge! The greatest AIs we've ever seen!

Of course not. The article is just oversensationalised. The paper doesn't make the same broad claims as the article.