Also the widely-shared model collapse paper ignored the earlier research that showed that models don't actually collapse at all if you keep even a bit of the original data around.

→ View original post on X — @jeremyphoward