> we have some picture how much resources is required to train and run these models This is the key detail. Even if the software gets 100,000 times better, the

> we have some picture how much resources is required to train and run these models This is the key detail. Even if the software gets 100,000 times better, the training data is still a major limitation. We can't get 100,000x (or even 10x) more data because the current training set is: - thousands of years of historical text, photos, and user-generated data (trillions of person-hours of work) - not yet polluted by AI So while I don't think we know the hard limit of the software, we're close to the hard limit for the training data and we'll also never have "pure" (not ChatGPT-filled) data again.