Do you think it really needs to be a completely novel approach? If you assume equivalents of gigantic LLM models can be trained in image/audio/haptic modalities and connected together to get the "big picture of the world" + if they could get a human/animal-like robot bodies which could speak/move/touch and learn from the experiences, wouldn't this be enough for AGI? I mean, yeah, we still need a lot of work to get there, but from today's perspective it already feels just like more of "the same" thing, no?