By your definition the script and a list of actors should be counted as compression, but that's clearly not what this particular invention claimed to do. An AI model is more like a drawing-by-the-numbers game than a compression method. It creates something that looks superficially like the original but isn't the original.
Any "compression" mechanism that apparently violates Shanon's theorems would be "lossy" anyway, and lossy compression is essentially creating something that looks superficially like the original but isn't.
A script and a list of actors would take up 8k already (if not more), so yeah, an AI that can work on the prompt "take this script and make it like a Hollywood blockbuster" might be our best way to attempt to recreate this "compression" system with SoTA tech.
Sloot claimed his method was lossless, and it supposedly started out from a digital representation (without compression artifacts such as introduced by DCT or FT).
