I can't figure out what specific topics trigger grok's "what does elon think" mode. It does seem like this behavior might have been explicitly trained into the model, in a subset of the training data, based on grok's observed behavior.

→ View original post on X — @jeremyphoward,