Apple and Meta have published a monstruously elegant compression method that encodes model weights using pseudo-random seeds.
The trick is to approximate model weights as the linear combination of a randomly generated matrix with fixed seed, and a smaller vector t.
