Tokenization. It doesn’t see text as sequences of characters, but of variable-length tokens about 4 chars on average. It generally can’t do anything that needs to be done one letter at a time. It’s also bad at counting in its head.

→ View original post on X — @goodside