Hilarious how Grok apologizes for going off-topic further down in the thread, but then can't resist the urge to immediately bring up white genocide again.
This is a prime indication that the bit about "white genocide" comes from the prompt. The model itself knows that it is bullshit from its training, though, and with a smart enough model, when there's a contradiction between the prompt and the training, the latter will generally win in the long run.
