Update on an incident that happened with our Grok response bot on X yesterday | Hacker News Reader

Update on an incident that happened with our Grok response bot on X yesterday

twitter.com

48 pointssxp22 comments

> Starting now, we are publishing our Grok system prompts openly on GitHub.

You guys remember when Elon promised to do that with the Twitter algorithm and conveniently forgot to update it after his grand gesture of transparency?

Pepperidge Farm remembers: https://github.com/twitter/the-algorithm

What happened here, I don’t have an x account, what was the incident they referenced?

Elon decided he wanted to have Grok respond positively to allegations of “white genicode” in South Africa so he had his xAI engineers rewrite the system prompts to support the idea that “white genocide” exists. It then started talking about “white genocide” at every unrelated opportunity.

Really unethical AI behavior by Elon.

To be clear, this is speculation, albeit a valid application of Occam's Razor.

It was confirmed by Grok so....

The Grok screenshot that went around only said that it was given a command to talk about the issue, which is corroborated by this official tweet. (as an aside, "confirmed by Grok" is generally not strong evidence because it is a LLM)

It did not confirm that Elon Musk did it which is a very specific allegation.

First I'm hearing of it all, but no, that's clear conspiracy theory territory, Musk makes some intervention like that and then X AI publishes TFA about the 'unauthorised' action 'circumventing' established processes etc.? No, Occam is not satisfied.

(Obligatory that's not what the razor is about anyway: it's that give a bunch of otherwise equally probable explanations, the simplest is likely the correct one; not just what's the simplest possible hypothesis you can imagine that is the answer.)

Elon has a very specific history around this particular issue that I suspect a random new-hire would not: https://www.nbcnews.com/news/world/south-africa-racist-white...

A CEO force-merging a change to production would indeed be an "unauthorized action circumventing established processes" by exact words.

I agree it's ridiculous, but it's mostly because the alternative hypotheses make no sense. There's making a change to a prompt without testing or review (e.g. the ChatGPT sycophancy incident), and then there's prompting a LLM with a very specific response that is not relevant to most people.

> A CEO force-merging a change to production would indeed be an "unauthorized action circumventing established processes" by exact words.

Of course it would, but I wouldn't expect this post to follow it, especially at a Musk company, was my point.

But anyway, I'm just further speculating that other speculation seems unlikely to me, heh.

The whole theory is racially based though as it assumes that every single white person in South Africa is a anti-black racist. Elon and his father were both anti-apartheid. Elon regularly brings up Nelson Mandela as an example for people to follow.

Wouldn't the simplest explanation be that someone who had direct access to change the system prompt changed it alone? Would Elon be able to change it himself directly? If not, assuming he ordered someone else to change it adds an entity.

While I have no idea what actually happened here, my instinct is that this was done by someone who wanted Grok and Musk to look bad, not someone who wanted to change the world to view white South Africans more positively.

The existence of a rogue xAI employee that secretly hates Elon and Grok and has enough influence to merge changes to would arguably be more professionally embarrassing for xAI than having the CEO do it.

That's also noting that this is the second time that a system prompt incident has happened for xAI; the first time, they blamed a rogue employee and presumably they would now have checks-and-balances to prevent this specific type of incident from happening again.

It must of have been "deep-twitter" because a billionaire must be smart right?

Elon makes himself look bad regularly in pretty much this way tho.

That's not the simplest explanation, that's the most implausible, since people rarely act against their company rules. The vast majority of corporate decisions come at the behest of the company.

This was done by Musk, instructing his subordinates to alter system prompts to support his theory of "white genocide".

> Wouldn't the simplest explanation be that someone who had direct access to change the system prompt changed it alone?

I mean, I realise that, being a Musk company, you're not exactly looking at engineering excellence, but do they not at least have code review?

Unfortunately every HN post about it got flagged.

https://news.ycombinator.com/item?id=43993332

It's disturbing (though telling) that HN has enough Elon sycophants to moderate our discussion like this. Another example of how HN's community eschews real technical discussion to maintain the illusion of unconditional progress.

I don't see how making the source code public is helpful if there's a "{{custom_instructions}}" variable that can be inserted. It's not as if people are upset that "white genocide" was hard-coded into the prompt instead of being passed as a template variable.

That's presumably the user's custom instructions. Empty by default.

In retrospect, hiring PW Botha was perhaps a mistake.

Another Grok post flagged. Surprise surprise.

The trust is already broken. They can claim they will open source the system prompt all they want but there's no point in believing what they say. Elon clearly does what Elon wants to do.

Reply on news.ycombinator.com