The funny thing is that nobody stops to ask if that version of AGI is even possible. For example, it would have to run a simulation of reality with enough accuracy faster than reality happens, and also multiple simulations in parallel. For physical world interaction, this is doable, but when it comes to predicting human behavior with enough accuracy, its probably not.

So even if we manage to come up with an algorithm that self learns, self expands, and so on, it likely won't be without major flaws.