The AI is getting better at things like maths. I recently asked it about iterating the Burrows-Wheeler transform, and it appeared to really understand that. It's not super easy to reason about why it's reversible, etc. and I felt that it got it.

This is obviously not AGI, and we're very far from AGI as we can see by trying out these LLMs on things like stories or on analyzing text, or dealing with opposing arguments, but for programming and maths performance is at a level where it's actually useful.

A lot of this is because there isn't a good definition of AGI. Look at Sama's recent interviews, that's how he deflects, along with the statement about the Turing test having ultimately been inconsequential. They have an internal definition of AGI that is "the model can perform the vast majority of economically viable tasks at the level of the highest humans" which isn't the story the investors are expecting when they hear AGI, so they're trying to stay mum to truthfully roadmap AGI while not blatantly lying to capital.