Surprising they didn't clean that from the data before training. It's easy to identify, a simple search->replace gets most of it, and a cheap LLM can identify t

Surprising they didn't clean that from the data before training. It's easy to identify, a simple search->replace gets most of it, and a cheap LLM can identify the edge cases (e.g. avoiding "Claude Shannon" -> "Kimi Shannon" or something).