Yeah exactly, I think that’s where the LLM gets this from (ie this pattern being represented in the training data)

→ View original post on X — @rasbt,