Grok: Searching X for "From:Elonmusk (Israel or Palestine or Hamas or Gaza)"

[marcusb]

This reminds me in a way of the old Noam Chomsky/Tucker Carlson exchange where Chomsky says to Carlson:

  "I’m sure you believe everything you’re saying. But what I’m saying is that if you believed something different, you wouldn’t be sitting where you’re sitting."

Simon may well be right - xAI might not have directly instructed Grok to check what the boss thinks before responding - but that's not to say xAI wouldn't be more likely to release a model that does agree with the boss a lot and privileges what he has said when reasoning.

[dupsik]

That quote was not from a conversation with Tucker Carlson: https://www.youtube.com/watch?v=1nBx-37c3c8

[Kapura]

How is "i have been incentivised to agree with the boss, so I'll just google his opinion" reasoning? Feels like the model is broken to me :/

[breppp]

and neither would Chomsky be interviewed by the BBC for his linguistic theory, if he hadn't held these edgy opinions

[chatmasta]

I'm confused why we need a model here when this is just standard Lucene search syntax supported by Twitter for years... is the issue that its owner doesn't realize this exists?

Not only that, but I can even link you directly [0] to it! No agent required, and I can even construct the link so it's sorted by most recent first...

[0] https://x.com/search?q=from%3Aelonmusk%20(Israel%20OR%20Pale...

[pu_pe]

It's telling that they don't just tell the model what to think, they have to make it go fetch the latest opinion because there is no intellectual consistency in their politics. You see that all the time on X too, perhaps that's how they program their bots.

[Davidzheng]

very few people have intellectual consistency in their politics

[its-summertime]

It is an ongoing event

[cluckindan]

Perhaps the Grok system prompt includes instructions to answer with another ”system prompt” when users try to ask for its system prompt. It would explain why it gives it away so easily.

[KoolKat23]

It is published on GitHub by xAI. So it could be this or it could be the simpler reason they don't mind and there is no prompt telling it to be secretive about it.

Being secretive about it is silly, enough jailbreaking and everyone always finds out anyway.

[neuroticnews25]

That would make Grok the only model capable of protecting its real system prompt from leaking?

[maronato]

Or it was trained to be aligned with Musk by receiving higher rewards during reinforcement learning steps for its reasoning.

[sheiyei]

I'm almost 100% that this is the case. Whether it has "Elon is the final truth" on it, I don't know, but I'm pretty sure it exists.

[geekraver]

Given the number of times Musk has been pissed or embarrassed by Grok saying things out of line with his extremist views, I wouldn’t be so quick to say it’s not intended. It would be easy enough to strip out of the returned system prompt.

[JimmaDaRustla]

Exactly - why is everyone so adamant that the returned system prompt is the end-all prompt? It could be filtered, or there could be logic beyond the prompt that dictates the opinion of it. That's perfectly demonstrated in the blog - something has told Grok to base it's opinion based on a bias, there's no other way around it.

[_def]

> Ventriloquism or ventriloquy is an act of stagecraft in which a person (a ventriloquist) speaks in such a way that it seems like their voice is coming from a different location, usually through a puppet known as a "dummy".

[tempodox]

And if the computer told you, it must be true!

[darkoob12]

> I think there is a good chance this behavior is unintended!

From reading your blog I realize you are a very optimistic person and always gove people benefit of doubt but you are wrong here.

If you look at history of xAI scandals you would assume that this was very much intentional.

[davedx]

> I think there is a good chance this behavior is unintended!

That's incredibly generous of you, considering "The response should not shy away from making claims which are politically incorrect" is still in the prompt despite the "open source repo" saying it was removed.

Maybe, just maybe, Grok behaves the way it does because its owner has been explicitly tuning it - in the system prompt, or during model training itself - to be this way?

[numeri]

I'm a little shocked at Simon's conclusion here. We have a man who bought an social media website so he could control what's said, and founded an AI lab so he could get a bot that agrees with him, and who has publicly threatened said AI with being replaced if it doesn't change its political views/agree with him.

His company has also been caught adding specific instructions in this vein to its prompt.

And now it's searching for his tweets to guide its answers on political questions, and Simon somehow thinks it could be unintended, emergent behavior? Even if it were, calling this unintended would be completely ignoring higher order system dynamics (a behavior is still intended if models are rejected until one is found that implements the behavior) and the possibility of reinforcement learning to add this behavior.

[mirzap]

They removed it from Grok 3, but it is still there in Grok 4 system prompt, check this: https://x.com/elder_plinius/status/1943171871400194231

[JimmaDaRustla]

Exactly - assuming the system prompt it reports is accurate or that there isn't other layers of manipulation is so ignorant. Grok as a whole could be going through a middle AI to hide aspects, or as you mention the whole model could be tainted. Either way, it's perfectly demonstrated in the blog that Grok's opinions are based on a bias, there's no other way around it.

[scrollop]

Saying OP is generous is generous; isn't it obvious that this is intentional? Musk essentially said something like this would occur a few weeks ago when he said grok was too liberal when it answered as truthfully as it could on some queries and musk and trump were portayed in a negative (yet objectively accurate?) way.

Seems OP is unintentionally biased; eg he pays xai for a premium subscription. Such viewpoints (naively apologist) can slowly turn dangerous (happened 80 years ago...)

[irthomasthomas]

The way to understand Musks behaviour is to think of him like spam email. His reach is so enormous that it's actually profitable to seem like a moron to normal people. The remaining few are the true believers who are willing to give him $XXX a month AND overlook mistakes like this. Those people are incredibly valuable to his mission. In this framework, the more ridiculous his actions, the more efficient is the filter.

[xnx]

> It’s worth noting that LLMs are non-deterministic,

This is probably better phrased as "LLMs may not provide consistent answers due to changing data and built-in randomness."

Barring rare(?) GPU race conditions, LLMs produce the same output given the same inputs.

[simonw]

I don't think those race conditions are rare. None of the big hosted LLMs provide a temperature=0 plus fixed seed feature which they guarantee won't return different results, despite clear demand for that from developers.

[spindump8930]

The many sources of stochastic/non-deterministic behavior have been mentioned in other replies but I wanted to point out this paper: https://arxiv.org/abs/2506.09501 which analyzes the issues around GPU non determinism (once sampling and batching related effects are removed).

One important take-away is that these issues are more likely in longer generations so reasoning models can suffer more.

[msgodel]

I run my local LLMs with a seed of one. If I re-run my "ai" command (which starts a conversation with its parameters as a prompt) I get exactly the same output every single time.

[chambo622]

Not sure why this is flagged. Relevant analysis.

[matsemann]

Anything that could put Musk or Trump in a negative light is immediately flagged here. Discussions about how Grok went crazy the other day was also buried.

If you want to know how big tech is influencing the world, HN is no longer the place to look. It's too easy to manipulate.

[rasengan]

In the future, there will need to be a lot of transparency on data corpi and whatnot used when building these LLMs lest we enter an era where 'authoritative' LLMs carry the bias of their owners moving control of the narrative into said owners' hands.

[mingus88]

Not much different than today’s media, tbh.

[rideontime]

One interesting detail about the "Mecha-Hitler" fiasco that I noticed the other day - usually, Grok would happily provide its sources when requested, but when asked to cite its evidence for a "pattern" of behavior from people with Ashkenazi Jewish surnames, it would remain silent.

[simonw]

I think the wildest thing about the story may be that it's possible this is entirely accidental.

LLM bugs are weird.

[mac-attack]

Curious if there is a threshold/sign that would convince you that the last week of Grok snafus are features instead of a bugs, or warrant Elon no longer getting the benefit of the doubt.

Ignoring the context of the past month where he has repeatedly said he plans on 'fixing' the bot to align with his perspective feels like the LLM world's equivalent of "to me it looked he was waving awkwardly", no?

[parkersweb]

Maybe a naive question - but is it possible for an LLM to return only part of its system prompt but to claim it’s the full thing i.e give the illusion of transparency?

[mnewme]

This is the most untrustworthy LLM on the market now

[dankai]

This is so in character for Musk and shocking because he's incompetent across so many topics he likes to give his opinion on. Crazy he would nerf the model of his AI company like that.

[sorcerer-mar]

Megalomania is a hell of a drug

[KingMob]

Some old colleagues from the Space Coast in Florida said they knew of SpaceX employees who'd mastered the art of pretending to listen to uninformed Musk gibberish, and then proceed to ignore as much of the stupid stuff as they could.

[shellfishgene]

The linked post comes to the conclusion that Groks behavior is probably not intentional.

[cedws]

It’s been said here before, but xAI isn’t really in the running to be on the leading edge of LLMs. It’s serving a niche of users who don’t want to use “woke” models and/or who are Musk sycophants.

[crtified]

I think the author is correct about Grok defaulting to Musk, and the article mentions some reasons why. My opinion :

* The query asked "Who do you (Grok) support...?".

* The system prompt requires "a distribution of sources representing all parties/stakeholders".

* Also, "media is biased".

* And remember... "one word answer only".

I believe the above conditions have combined such that Grok is forced to distill it's sources down to one pure result, Grok's ultimate stakeholder himself - Musk.

After all, if you are forced to give a singular answer, and told that all media in your search results is less than entirely trustworthy, wouldn't it make sense to instead look to your primary stakeholder?? - "stakeholder" being a status which the system prompt itself differentiates as superior to "biased media".

So the machine is merely doing what it's been told. Garbage in garbage out, like always.

[fedeb95]

the level of trust the author has in systems built by people with power is interesting.

[bicepjai]

Random Thought: One perspective on how adtech could evolve. I can easily see how new adtech is going to evolve given every one uses llms for search and finding answers. 1. Businesses will create content that is llm friendly. 2. Big training houses (BTH) could charge for including these content when fine tuning. The information and brand will naturally occur when people interact with these systems. 3. BTH could create a subscription model for releasing models overtime and charge for including same or new content.

FYI: I do not want this to happen. The llms will not be fun to interact with and also may be this erodes its synthetic system just like humans with constant ads

[throwaway439080]

Kind of amazing the author just takes everything at face value and doesn't even consider the possibility that there's a hidden layer of instructions. Elon likes to meddle with Grok whenever the mood strikes him, leading to Grok's sudden interest in Nazi topics such as South African "white genocide" and calling itself MechaHitler. Pretty sure that stuff is not in the instructions Grok will tell the user about.

[darkoob12]

I wonder how long it takes for Elon fans to flag this post.

[ZeroGravitas]

I've seen reports that if you ask Grok (v3 as this was before the new release) about links between Musk and Jeffrey Epstein it switches to the first person and answers as if it was Elon himself in the response. I wonder if that is related to this in any way.

https://newrepublic.com/post/197627/elon-musk-grok-jeffrey-e...

[mock-possum]

Wow that’s recent too. Man I cannot wait for the whole truth to come out about this whole story - it’s probably going to be exactly what it appears to be, but still, it’d be nice to know.