This is the key - right now each new model has had countless resources dedicated to training, then they are more or less set in stone until the next update.

These big models don't dynamically update as days pass by - they don't learn. A personal assistant service may be able to mimic learning by creating a database of your data or preferences, but your usage isn't baked back into the big underlying model permanently.

I don't agree with "in our lifetimes", but the difference between training and learning is the bright red line. Until there's a model which is able to continually update itself, it's not AGI.

My guess is that this will require both more powerful hardware and a few more software innovations. But it'll happen.