Theoretically, maybe, but I believe the devil lies in the details: implementation is where "biological hardware" if you will is more akin to a thousand specific sets of TPUs whereas we're trying to brute force / shoehorn the whole processing into a suboptimal one-size-fits-all giant RNN. The inadequacies and inefficiencies of such a shoehorning of the model's execution might (I argue do) absolutely self-defeat the endgoal of a coherent adaptive machine. I'm tempted to humorously say #NotAllParts (need the same underlying hardware optimizations) ;-)

Shower thought that just came up: observe that while we're endlessly chasing a bigger-than-reality String Theory, actually working physics implemented in real-world machines follow the specialized approach of one partial but perfect theory for each category of problems. We build hybrid, because as far as we can tell, reality is variations, and the biological probably most of all. So in trying to build a being…

That is true, but unfortunately "out of the box", they're not well suited just be "fed into" an NN. Even if you think of the adjacency matrix as very similar to how the weights are laid out in a feed-forward neural network, you can't ignore that:

- in real life, graphs are not fixed

- you need to deal with the many different potential representations of the graph (permutation invariance)

- the nodes are usually containing more features than a single scalar value

but this is definitely not the best explanation, I think this guy does a lot better job: https://youtu.be/JtDgmmQ60x8