How could LLMs reach AGI?

I think there is a fair argument that "AGI" is also the best term we have for a general purpose AI that can be applied to most human level tasks but is not in any way conscious or sentient or fully autonomous.

Such as the mostly unreleased GPT-4o. Because we don't have another term for that type of AI. It is fundamentally general purpose and different from decades of very narrow AI.

I also think that GPT-4o has demonstrated that it's not a language model that leads to anything truly general purpose but rather a multimodal model.

But I am not saying anything like this becomes conscious or has sensory experiences like an animal or even all of the cognitive capabilities of an animal.

Yet, it is clearly very general purpose and capable.

> Yes, A stateful multi agent system with real time reinforcement learning could give us AGI that could... do..

Why even use a language model, if we need all that on top? By introducing meaningful statefulness, multiple agents (which could really be anything) and reinforcement learning you have basically defined an AI-free reactive system. On top I think it could be argued that statefulness and reinforcement learning are not sufficient enough to elevate language models to human-level intuition. I think we would just end up with a more complicated and less generalized idea of where AI is going.