Hacker News Reader

Comment by Bender | Hacker News Reader

Not defending Elon or the infobot but my theory is that by leaving that LLM unfiltered people have learned how to gamify and manipulate it into having a fascist slant. I could even guess which groups of people are doing it but I will let them take credit and it's not likely actual neo-nazi's, they are too dumb and on too many drugs to manipulate an infobot. These groups like to LARP to piss everyone off and they often succeed. If I am right it is a set of splintered groups formerly referred to generically as The Internet Hate Machine but they have (d)evolved into something worse that even 4chan could not tolerate.

View on HN

It's just the prompt: https://github.com/xai-org/grok-prompts/commit/c5de4a14feb50...

People who don't understand llms think saying don't shy away from making claims that are politically incorrect means it won't PC. In reality saying that just makes things associated with politically incorrect more likely. The /pol/ board is called politically incorrect, the ideas people "call" politically incorrect most of all are not Elon's vague centrist stuff it's the extreme stuff. LLMs just track probable relations between tokens, not meaning, it having this result based on that prompt is obvious.

it's almost like Grok takes "politically incorrect" to be synonymous with racist.

The mishap is not the chatbot accidentally getting too extreme and at odds with 'Elon's centrist stuff'. The mishap is the chatbot is too obvious and inept about Musk's intent.

We have no evidence to suggest that they just made a prompt change and it dialed up the 4chan weights. This repository is a graveyard where a CI bot occasionally makes a text diff, but we have no understanding if it's connected with anything deployed live or not.

> it's not likely actual neo-nazi's, they are too dumb to manipulate an infobot.

No they are not. There exist brilliant people and monkeybrains across the whole population and thus the political spectrum. The ratios might be different, but I am pretty sure there exist some very smart neo-nazis

Curtis Yarvin’s writing is insufferable and many of his ideas are both bad and effectively Nazism, but clearly he’s very smart (and very eager to prove it).

Yarvin is an out-and-out white nationalist, though he denies it, or at least the name: "I am not a white nationalist, though I am not exactly allergic to the stuff" - whatever the hell that mealy-mouthed answer is meant to mean.

He even wrote a bloviating article to further clarify that he is not a white nationalist. You'd be forgiven, though, if you didn't read the title. It spends most of the article sympathizing with, understanding, agreeing with, and talking of how white nationalism "resonates" with him. But don't worry, he swears he's not one at the end of the article!

There are, but fascism's internal cultural fixtures are more aesthetic than intellectual. It doesn't really attract or foster intellectuals like some radical political movements do, and it shows very clearly in the composition of the "rank and file".

Put plainly, the average neo-Nazi is astonishingly, astonishingly stupid.

> It doesn't really attract or foster intellectuals like some radical political movements do

It definitely attracts people who are competent in technology and propaganda is sufficient numbers for the task being discussed, especially when as a mass movement it has (or is perceived to have) a position of power that advantage-seeking people want to exploit. If anything, the common perception that fascists are "astonishingly, astonishingly stupid" makes this more attractive for people who are both competent and also amoral opportunists (which do occur together, competence and moral virtue aren't particularly correlated.)

It sure didn’t seem to take much manipulation from what I saw. “Which 20th century figure would solve our current woes” is pretty mild input to produce “Hitler would solve everything!”

I'm out of the loop, why is it an "infobot" and not a chatbot?

In 1999 there was a perl chatbot called infobot that could be taught factoids, truths, lies. It would learn anything people chatted about on IRC. So I call LLM's infobots.

Neat, thanks for explaining.

> Not defending Elon or the infobot but my theory is that by leaving that LLM unfiltered people have learned how to gamify and manipulate it into having a fascist slant.

We don't need a theory that explains how Grok got a fascist slant, we know exactly what happened: Musk promise to remove the "woke" from Grok, and what's left is Nazi. [1]

[1] https://amp.cnn.com/cnn/2025/07/08/tech/grok-ai-antisemitism

> we know exactly what happened

The price of certainty is inaccuracy.

So the only way to be accurate is to vaguely gesture at hodgepodge theories and suggestions that people "do their own research"?

Surely you can be both accurate and certain, otherwise you should just shut up and be right all the time.

> So the only way to be accurate is to vaguely gesture at hodgepodge theories and suggestions that people "do their own research"?

Yours was a hodgepodge theory. That's why I said that. I was advocating against hodgepodge theories in general, and yours in particular.

That LLM is incredibly filtered, just in a different way from others. I suspect by "retraining" the model Elon actually means that they just updated the system prompt, which is exactly what they have done for other hacked in changes like preventing the bot from criticizing Trump/Elon during the election.

No, that's definitely not what happened. For quite a while Grok actually seemed to have a surprisingly left-leaning slant. Then recently Elon started pushing the South African "white genocide" conspiracy theory, and Grok was sloppily updated and started pushing that same conspiracy theory even in unrelated threads. Last week Elon announced another update to Grok, which coincided with this dramatic right-wing swing in Grok's responses. This change cannot be blamed on public interactions like Microsoft's Tay, it's very clearly the result of a deliberate update, whether or not these results were intentional.

Reply on news.ycombinator.com