The real trick will be in figuring out some sort of solid state device - probably at least partially (if not predominantly) analog, not digital - that can reliably mimic the behaviors of biological neurons.

The actual breakthrough of AGI isn’t going to occur until we figure out how to do machine learning with analog circuit design. Digital simply does not scale well - what we’re doing now with “AI” using GPUs is simply brute forcing the problem by layering dozens of markov chains on top of each other.