Catastrophic forgetting and frozen weights limit today’s LLMs
LLMs have significant limitations that prevent them from being AGI. First, they suffer from catastrophic forgetting because learning and inference are separate phases. The weights freeze after training, so the system cannot continuously learn in real time. Second, they store knowledge more like a huge catalog of special cases rather than true abstraction. Third, they do not reorganize themselves in context of real-time interactions the way human brains do. While LLMs can do impressive things like math proofs and reasoning about novel scenarios, they lack the kind of generalizable, self-integrating knowledge system needed for AGI. They are more like an off-ramp on the path to AGI.
— Rethinking the AGI Race, with Benjamin Goertzel · The Keynote on AI