I think it's because people have different ideas of AGI. Some only expect general intelligence, while others also require weird speculative ideas, such as rapid self-improvement.

Right now, what seems to be missing from general intelligence is mostly the ability to learn from experience and continue operating in a useful way well beyond context window size. That might already be technically feasible, but probably not in a cost-effective way.

I don’t believe AI progress is a linear line that ends in “AGI”. I think the current way we create LLMs will be improved, and improved, and improved. But an LLM at its core is a predictive chain, not general intelligence.

What does AGI mean is probably the sticking point for most. I'm inclined to call it AGI if they can create something currently matching SOTA where there is no difference between training and inference time, i.e. it can continuously learn. Without the model collapsing.