[Jolter]

That’s not a good analogy and you should know it.

[valine]

It’s a terrific analogy. The alternative is to believe that a 5GB model somehow contains a database of 160 million images.

[TillE]

It's a fine analogy, but the map is not the territory. Machine learning is not human learning, even if it works in a vaguely comparable way.

It's still a computer program that uses an enormous amount of copyrighted work as its input.

[greysphere]

"With four parameters I can fit an elephant, and with five I can make him wiggle his trunk”

It seems like you could calculate how much data is within X% error of a 5GB model, and what X% should be for 'visual data'.

I bet it's pretty big.