> There is a massive flaw in this argument. In real life, whether a given generated work ends up in a future dataset depends on how good it is according to huma

> There is a massive flaw in this argument. In real life, whether a given generated work ends up in a future dataset depends on how good it is according to humans Even the best-looking JPEG (as judged by humans) is still lossy.