I haven't used grok personally, but on gemini it's not too hard to get it to shit on the oligarchs. Even basically got it to admit killing Trump would be a net positive to society without much effort.

I do agree in principle that LLMs work much better in cases where you can verify the output quickly but getting there would be difficult, so NP problems basically.

I'm in a weird position with LLMs because I have found them absolutely invaluable as a learning tool, but also recognize how much damage they could do to society, especially in the hands of dumber people when it comes to propagandization.