the clarity in these examples is a little startling to me; this is such a good way of doing interpretability; it's interesting that the model seems to scatter details like what Einstein looks like across multiply layers, and it is able to gradually refine throughout the fwd pass

→ View original post on X — @jxmnop,