This feels like a tokenization issue — in general, it doesn’t see text as sequences of letters, and struggles with tasks that assume it does.

→ View original post on X — @goodside