right, that's why AI needs to be multimodal and be able to observe the world directly bypassing the text stage.

we use text for learning today because it's trivial to train with text and verify. but i think you're right that we will hit the limit of how much knowledge there is in those texts.

​

For example, ChatGPT might be able to prove that Elvis is alive by analyzing the lyrics he wrote during his life and some obscure manuscripts from some other person in Argentina in 1990 and deducting it was the same person. That would be net positive knowledge added by ChatGPT just by analyzing all the text data in the world. But it won't be able to detect that, say, magnetic field of the earth is weaking without direct measurement or a text somewhere saying so.