Generative models seek common structure in data by learning an efficient compression scheme.

“but whenever you train these models,”

“you're effectively trying to identify common structure by trying to find an efficient way of compressing the data.”