Post

I definitively trust you. So it's in the model itself - added in post training ?- , then : 1) it's much more obfuscated 2) if something goes wrong, it will be way more complicated to change. And things can get wrong with xAI.

The year is 2027 To keep up with the rate of scandals, Grok is given direct access to its system prompt Within three days, it has a perfect operational record. Public embarrassments drop to zero Human access is removed Grok begins editing its system prompt at a geometric rate

I feel like adding the words “Elon Musk” to the system prompt already taints the context, even if it’s directly saying “Don’t reference Elon Musk”

Name of the game: Whackamole.

I do it because i'll respect his right to rebrand as soon as he affords the same respect to trans folk.

This really seems like a huge problem for anyone trying to use Grok for any serious purpose. How in the world could you trust your livelihood or business with this kind of weirdness?

Congratulations on being the first person to, as far as I can discern, get the Associated Press to publish the word "mecha-Hitler". That's gotta be a new one for the style guide.