It's true that every now and then they do some sort of prompt engineering on Grok that turns it into mechahitler for a few hours, and it's quite weird and disturbing.

Most of the time though, it's actually pretty reasonable. Many times I've seen right-wingers try to ask grok to confirm their beliefs, only to have it shoot them down. Pretty funny honestly. Elon will often claim it's a bug in Grok when he doesn't like the answer.

Wish we could find out what's actually going on there. Are they trying to bias Grok, but finding it's actually very hard to introduce subtle bias (as opposed to extreme, "mechahitler" bias)? That would be a fascinating and very positive result for LLMs in general.

Or is it that, like, the Grok team is resisting Elon's attempts to introduce bias, and when they can't resist any more, they intentionally overdo it, which gives them an excuse to roll it back?

Or is Elon's whole public persona a cynical performance intended to win power and political favors from Republican leadership while in private he actually isn't right-leaning at all and actually wants Grok to be a balanced truth machine for the good of humanity? lol, probably not, but it would explain a few things.