Turns out, it’s even less impressive than memorizing the training data:

→ View original post on X — @rasbt