If a new model shares a feature with a trusted model, that area probably doesn't need scrutiny. Model diffing isolates the features unique to the new model—where new risks are most likely to be located.

→ View original post on X — @anthropicai,