With respect to needing to mimic biology: not necessarily, though it is certainly true that e.g. a purely feedforward model could NOT achieve AGI, since it can'

With respect to needing to mimic biology: not necessarily, though it is certainly true that e.g. a purely feedforward model could NOT achieve AGI, since it can't make predictions over time. Feedforward just means that the outputs of the network don't "feed" back into its inputs -- and as a result the same input will always cause the same output regardless of previous inputs. Humans/animals certainly take time into account when acting -- e.g. something as simple as coordinating muscles for simple movements requires varied output over time, despite most inputs (in the form of tactile senses) being mostly unchanged (at least for some kinds of movement, e.g. waving your hand). Recurrent models (where some outputs are connected back to the inputs) are one possible way to account for time, but the work is still really early on these methods, and it's not clear what architecture (e.g. which and how many inputs and outputs should connect) would be efficacious.