Honestly I think we're going to have to pit AIs against one another in another web of complexity in order to reach AGI. You're going to need 50 different AIs all deliberating and contradicting and refining output with each one controlling some aspect of the result. I think it's far too much processing needed at this point in time. LLMs right now are largely just word predictors, but we also see things like diffusors being able to create images, etc. These are distinct problem types, and I think if we develop a separate AI for each generalized problem type, we can utilize them in a way that can predictably output something that we can classify as 'general intelligence'.