I'm of the opinion that thanks to a memory prosthesis tacked on (and training to use it), MemGPT will make it possible to build an AGI in the next year or two. However, due to the nature of the prosthetic to expand its memory(making lots of notes, and searching for them), it's likely (IMHO) to be a packer, in the Mappers vs Packers paradigm.[1]

Does anyone think it's even possible to build an AGI that's a mapper? I can't see how you could even begin to compose examples to train an LLM against.

[1] https://reciprocality.org/Reciprocality/r0/index.html#:~:text=Mapping%20versus%20Packing,large%20numbers%20of%20knowledge%20packets