Francois seems almost to assume that just because an algorithm takes millions or billions of datapoints to train, that means its output is just “memorization”. In fact it seems to me that the learning algorithms just work pretty slowly, and that the thing that’s learned after those millions or billions of tries is the actual generative concepts.
Francois seems almost to assume that just because an algorithm takes millions or billions of datapoints to train, that means its output is just “memorization”. In fact it seems to me that the learning algorithms just work pretty slowly, and that the thing that’s learned after those millions or billions of tries is the actual generative concepts.