Comment by wongarsu
3 days ago
There have been at least two instances of "unauthorized modifications" to the system prompt of the Grok model running wild in X, but if you build your own integration you would provide your own system prompt and be unaffected by that.
On the model side I've found Grok3 to be very unbiased. If you ask it to write a story it will somehow find a way to weave a mention of X/Twitter into that story, but other than that it is much less biased and moralizing than e.g. OpenAI models. It also has very lax guard rails, so that's something you'd probably want to add
I can't say yet whether all of this is still true for Grok 4
Are you asking it to write a story on like grok.com or inside of twitter, or are you saying that if I call the API and ask for a story I'm going to get twitter weaved in there somehow
Both on the API (which I'm using for openwebui) and on grok.com.
It might just be that twitter is a disproportionate amount of their training data, leading grok to believe in a world where twitter is much more ubiquitous than in our current reality. And the grok.com version might be unintentionally biased because it has a tool to search twitter and the system prompt instructs it on how to use it, leading to a sentence about X to always be in the prompt in that version. I'm not at all claiming that it has to be an intentional plot to promote twitter, it might just be an accident. But it is a very perceivable bias
> if you build your own integration you would provide your own system prompt…
That won't save you if the model itself is trained on some nasty shit.