still unsure if LLMs can produce AGI. i think their architecture is fundamentally flawed as they're designed to always output a response even if they don't know the answer. perhaps some sort of "uncertainty" token during pre-training could give LLMs an idea of what they know and don't know and could produce better answers. they still don't learn though. it's possible LLMs become good enough at coding they can create some sort of new self-learning AI, but LLMs alone i doubt can reach general intelligence levels.