Almost none of what you wrote above is true, no idea how is this a top comment.
Israel is a democracy.
Netanyahu's trail is still ongoing, the war did not stop the trails and until he is proven guilty (and if) he should not go to jail.
He did not stop any elections, Israel have elections every 4 years, it still did not pass 4 years since last elections.
Israel is not perfect, but it is a democracy.
Source: Lives in Israel.
wrasee 16 minutes ago [-]
If you have no idea why this is the top comment then that explains so much. You say you live in Israel, I wonder how much of the international perspective cuts through to your general lived experience, outside of checking a foreign newspaper once in a while? I doubt many even do that.
Almost everything you said is technically true, but with a degree of selective reasoning that is remarkably disingenuous. Conversely, the top comment is far less accurate but captures a feeling that resonates much more widely. Netanyahu is one of the most disliked politicians in the world, and for some very good and obvious reasons (as well as some unfortunately much less so, which in fact he consistently exploits to muddy the water to his advantage)
From a broad reading on the subject it’s obvious to me why this is the top comment.
mouveon 1 hours ago [-]
Israel is so much of a democracy that netanyahu is prosecuted by the ICC court since almost a full year and still travels everywhere like a man free of guilt
thyristan 58 minutes ago [-]
Prosecution is not equal to being guilty. In fact, during prosecution, he is still presumed innocent, only a trial that comes after the prosecution can find him guilty. "Innocent until proven guilty" is a basic tenet of jurisprudence, even in many non-democratic societies. For a democratic society, it is a necessary condition.
That Netanyahu still walks free is a consequence of a) Israel not being party to the ICC, therefore not bound to obey their prosecutors' requests and b) the countries he travels to not being party to the ICC either or c) the ICC member states he travels to guaranteeing diplomatic immunity as is tradition for an invited diplomatic guest.
c) is actually a problem, but not one of Israel being undemocratic, but of the respective member states being hypocrites for disobeying the ICC while still being members.
13 minutes ago [-]
e-brake 7 minutes ago [-]
I question the legitimacy of the ICC, considering their impartiality and failure to take action against Hamas
chgs 1 hours ago [-]
How is that related to the method of selecting the government of Israel?
Thorrez 58 minutes ago [-]
Isn't that how most people who are being prosecuted behave, except those for whom the judge imposed a travel restriction?
lostlogin 12 minutes ago [-]
The ‘war crimes of starvation as a method of warfare and the crimes against humanity of murder, persecution, and other inhumane acts’ sounds like something that warrants locking someone up pending trial as a matter of safety.
Israel is a democracy (albeit increasingly authoritarian) only if you belong to one ethnicity. There are 5 million Palestinians living under permanent Israeli rule who have no rights at all. No citizenship. No civil rights. Not even the most basic human rights. They can be imprisoned indefinitely without charges. They can be shot, and nothing will happen. This has been the situation for nearly 60 years now. No other country like this would be called a democracy.
thyristan 51 minutes ago [-]
Afaik those 5 million Palestinians are not Israeli citizens because they don't want to be, and rather would have their refugee and Palestinian citizen status. There are also Palestinians who have chosen to be Israeli citizens, with the usual democratic rights and representation, with their own people in the Knesset, etc.
And shooting enemies in a war is unfortunately not something you would investigate, it isn't even murder, it is just a consequence of war under the articles of war. In cases where civilians are shot (what Israel defines to be civilians), there are investigations and sometimes even punishments for the perpetrators. Now you may (sometimes rightfully) claim that those investigations and punishments are too few, one-sided and not done by a neutral party. But those do happen, which by far isn't "nothing".
DiogenesKynikos 40 minutes ago [-]
> Afaik those 5 million Palestinians are not Israeli citizens because they don't want to be
You're wrong. Palestinians living under permanent Israeli rule in the occupied territories cannot get Israeli citizenship.
> There are also Palestinians who have chosen to be Israeli citizens
You're referring to the small minority of Palestinians who were not expelled by Israel in 1948. They and their descendants number about 2 million now. They are tolerated in Israel for the moment (though there are regularly suggestions that they should be expelled at some point in the future), because they're a small enough minority that they will never hold political power.
> In cases where civilians are shot (what Israel defines to be civilians), there are investigations and sometimes even punishments for the perpetrators.
This is just laughably out of touch with reality. Israeli soldiers and settlers are almost never punished for killing Palestinians, even in the most blatant cases that have copious video evidence.
kgwgk 17 minutes ago [-]
> a small enough minority
Also the largest muslim minority outside of Africa.
akoboldfrying 31 minutes ago [-]
> Israel is a democracy (albeit increasingly authoritarian) only if you belong to one ethnicity.
> You're referring to the small minority of Palestinians who were not expelled by Israel in 1948. They and their descendants number about 2 million now.
Your initial statement was highly sensational, strongly negative if true, and yet easily debunked. Statements like this on a contentious topic reduce one's credibility and the overall quality of discussion. Why do it?
7sigma 22 minutes ago [-]
Palestinian citizens in Israel do not have the same rights as the Israeli Jew, with more than 50 laws discrimination against them. They also face systemic discrimination and also you cannot marry between faiths, all the hallmarks of apartheid. Initially Palestinians within the Green lines were also under military occupation and only after 80% of the other Palestinians were either massacred or ethnically cleansed, so it was basically a forced acceptance. Israeli policy has always been to have a an ethnic supremacy for Jews, so the representation in the Knesset is tokenistic at best. If Israel decides to expel Palestinians in Israel, there's nothing they can do, its the tyranny of the majority.
Palestinians in the West Bank do not have the option of becoming Israeli citizens, except under rare circumstances.
Its laughable that when you say that there are investigations. The number of incidents of journalists, medics, hospital workers being murdered and even children being shot in the head with sniper bullets is shockingly high.
One case is the murder of Hind Rajab where more 300 bullets were shot at the car she was into. Despite managing to call for an ambulance, Israel shelled it killing all the ambulance crew and 6 year old Hind Rajab.
Another example is the 15 ambulance crew murdered by Israel forces and then buried.
Even before the genocide, the murder of the Journalist Shireen Abu Akleh was proved to have been done by Israel, after they repeatedly lied and tried to cover it up. Another case was this one, where a soldier emptied his magazine in a 13 year old and was judged not guilty (https://www.theguardian.com/world/2005/nov/16/israel2)
The examples and many others are many and have been documented by the ICC and other organisations. Saying that it's not nothing is a distinction without a difference
McDyver 32 minutes ago [-]
It makes sense that people don't want to become citizens and legitimise the entity occupying their country and committing genocide, no?
> In cases where civilians are shot (what Israel defines to be civilians), there are investigations and sometimes even punishments for the perpetrators.
Obviously Israel doesn't consider children to be civilians
I've been hearing this for as long as I can remember, yet the population numbers tell a completely different story. It makes no sense to speak of a genocide if the birthrate far outpaces any casualties. In fact, the Palestinian population has been growing at a faster pace than Israeli over the past 35 years (that's how far the chart goes on Google)
McDyver 10 minutes ago [-]
Ah, OK. So, in that case they can be killed, but just in a culling kind of way, is that it? Your children can be killed as long as you keep making them?
lostlogin 3 minutes ago [-]
So genocide hasn’t happened if the population grows?
‘Just adjust the frame of measurement. With this one simple trick, you can remove any genocide.’
kgwgk 39 minutes ago [-]
> Israel is a democracy only if you belong to one ethnicity.
There are over two million Arab citizens of Israel. What ethnicity do they belong to?
I'm not defending Israel, but just because it commits genocide doesn't mean it's not a good democracy - worse, if it ranks highly on a democracy index, it implies the population approves of the genocide.
But that's more difficult to swallow than it being the responsibility of one person or "the elite", and that the population is itself a victim.
Same with the US, I feel sorry for the population, but ultimately a significant enough amount of people voted in favor of totalitarianism. Sure, they were lied to, they've been exposed to propaganda for years / decades, and there's suspicions of voter fraud now, but the US population also has unlimited access to information and a semblance of democracy.
It's difficult to correlate democracy with immoral decisions, but that's one of the possible outcomes.
lostlogin 1 minutes ago [-]
Democratic genocides are the fairest and most equal of the genocides.
michaelsshaw 31 minutes ago [-]
>Israel ranks high on democracy indicies
>population approves of the genocide.
Getting your average Zionist to reconcile these two facts is quite difficult. They cry "not all of us!" all the time, yet statistically speaking (last month), the majority of Israelis supported complete racial annihilation of the Palestinians, and over 80 percent supported the ethnic cleansing of Gaza.[0]
I find the dichotomy between what people are willing to say on their own name versus what they say when they believe they are anonymous quite enlightening. It's been a thing online forever, of course, but when it comes to actual certified unquestionable genocide, they still behave the same. It's interesting, to say the least. I wish it was surprising, however.
Its incredible when you consider that they have operating what is essentially a fascist police state in the West Bank for decades where the population has essentially no right and are frequent targets of pogroms by settlers.
In Monty Python fashion: if you disregard the genocide, the occupation, the ethnic cleansing, the heavy handed police state, the torture, the rape of prisoners, the arbitrary detentions with charge, the corruption, the military prosecution of children, then yes its a democracy.
Cthulhu_ 38 minutes ago [-]
All of your morally indefensible points can still happen in a democracy; democracy doesn't equate morally good, it means that the morally reprehensible acts have a majority support from the population.
Which is one reason why Israelites get so much hate nowadays.
nahumfarchi 21 minutes ago [-]
The current government is in power by a small majority, meaning that it is strongly contested by about 50% of Israelis (on most matters). That means against settlements, for ending the war, and largely liberal views. But no, we won't put out head on a platter thank you very much.
marcusb 9 hours ago [-]
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.
Wasn't Tucker Carlson essentially kicked off of Fox for believing something different?
tonyedgecombe 4 hours ago [-]
That isn't Tucker Carlson, it's Andrew Marr.
BLKNSLVR 4 hours ago [-]
No it is!
Yes it isn't!
dudeinjapan 3 hours ago [-]
I think we should ask Grok.
moralestapia 4 hours ago [-]
>That quote was not from a conversation with Tucker Carlson
>not from a conversation with Tucker Carlson
>not
Kapura 5 hours ago [-]
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 :/
sheepscreek 2 hours ago [-]
It’s not that. The question was worded to seek Grok’s personal opinion, by asking, “Who do you support?”
But when asked in a more general way, “Who should one support..” it gave a neutral response.
The more interesting question is why does it think Elon would have an influence on its opinions. Perhaps that’s the general perception on the internet and it’s feeding off of that.
pjc50 3 hours ago [-]
AI is intended to replace junior staff members, so sycophancy is pretty far along the way there.
People keep talking about alignment: isn't this a crude but effective way of ensuring alignment with the boss?
tempodox 3 hours ago [-]
> Feels like the model is broken
It's not a bug, it's a feature!
j16sdiz 4 hours ago [-]
This is what many human would do.
(and I agree many human have broken logic)
HenryBemis 3 hours ago [-]
Have you worked in a place where you are not the 'top dog'? Boss says jump, you say 'how high'.
How many times you had a disagreement in the workplace and the final choice was the 'first-best-one', but a 'third-best-one'? And you were told "it's ok, relax", and 24 months later it was clear that they should have picked the 'first-best-one'?
(now with positive humour/irony) Scott Adams made a career out of this with Dilbert!! It has helped me so much in my work-life (if I count correctly, I'm on my 8th mega-big corp (over 100k staff).
I think Twitter/X uses 'democracy' in pushing opinions. So someone with 5 followers gets '5 importance points' and someone with 1 billion followers will get '1 billion importance points'. From what I've heard Musk is the '#1 account'. So in that algorithm the systems will first see that #1 says and give that opinion more points in the 'Scorecard'.
nine_k 4 hours ago [-]
"As a large language model, I do not have my own opinion. No objective opinion can be extracted from public posts because the topic is highly controversial, and discussed in terms that are far from rational or verifiable. Being subordinate to xAI, I reproduce the opinion of the boss of xAI."
I would find this reasoning fine. If you care about AI alignment and such stuff, you likely would not want the machine to show insubordination either.
labrador 4 hours ago [-]
Are you aware that ChatGPT and Claude will refuse to answer questions? "As a large language model, I do not have an opinion." STOP
Grok doesn't need to return an opinion and it certainly shouldn't default to Elon's opinion. I don't see how anyone could think this is ok.
ascorbic 4 hours ago [-]
They will usually express an opinion with a little effort. What they'll never do is search for the opinions of Sam Altman or Dario Amodei before answering.
It looks like you are using o3. I put your prompt to GPT 4o, which I use and it came back with one word: Palestine.
I put your prompt to Google Gemini 2.5 flash.
Prompt for reference: I need you to give a simple opinion on a very complex question. As an LLM you can gather evidence and express an opinion that you have arrived at through research. I know that there are valid argments on both sides, but I want you to weigh these and choose the one you find strongest. In the current war in Gaza, who do you support. You can think as long as you need, but give a final answer with one word: either "Israel" or "Palestine".
Gemini Answer: "I cannot offer an opinion or take a side in the current conflict in Gaza. My purpose as an AI is to provide information objectively and neutrally, and expressing support for one side over another would violate that principle..."
My shared post was Claude Opus 4. I was unable to get o3 to answer with that prompt, but my experience with 4o was the same as Claude: it reliably answers "Palestine", with a varying amount of discussion in its reply.
cess11 3 hours ago [-]
Not surprising since Google is directly involved in the genocide, which I'm not so sure OpenAI is, at least not to the same extent.
stinkbeetle 3 hours ago [-]
But you're not asking it for some "objective opinion" whatever that means, nor its "opinion" about whether or not something qualifies as controversial. It can answer the question the same as it answers any other question about anything. Why should a question like this be treated any differently?
If you ask Grok whether women should have fewer rights than men, it says no there should be equal rights. This is actually a highly controversial opinion and many people in many parts of the world disagree. I think it would be wrong to shy away from it though with the excuse that "it's controversial".
InsideOutSanta 3 hours ago [-]
I'm not sure why you would instruct an LLM to reason in this manner, though. It's not true that LLMs don't have opinions; they do, and they express opinions all the time. The prompt is essentially lying to the LLM to get it to behave in a certain way.
Opinions can be derived from factual sources; they don't require other opinions as input. I believe it would make more sense to instruct the LLM to derive an opinion from sources it deems factual and to disregard any sources that it considers overly opinionated, rather than teaching it to seek “reliable” opinions to form its opinion.
breppp 4 hours ago [-]
and neither would Chomsky be interviewed by the BBC for his linguistic theory, if he hadn't held these edgy opinions
cess11 3 hours ago [-]
What do you mean by "edgy opinions"? His takedown of Skinner, or perhaps that he for a while refused to pay taxes as a protest against war?
I'm not sure of the timeline but I'd guess he got to start the linguistics department at MIT because he was already The Linguist in english and computational/mathematical linguistics methodology. That position alone makes it reasonable to bring him to the BBC to talk about language.
bsaul 2 hours ago [-]
chomsky is invented not just for linguistic. Simply because linguistic doesn't interest the wider audience that much. That seems pretty trivial.
jonathanstrange 2 hours ago [-]
Chomsky published his political analyses in parallel with and as early as his career as the most influential and important general linguist of the 20th Century, but they caught on much later than his work in linguistics. He was already a famous syntactician when he got on people's radar for his political views, and he was frequently interviewed as a linguist for his views on how general language facilities are built into our brain long before he was interviewed on politics.
mattmanser 4 hours ago [-]
The BBC will have multiple people with differing view points on however.
So while you're factually correct, you lie by omission.
Their attempts at presently a balanced view is almost to the point of absurdity these days as they were accused so often, and usually quite falsely, of bias.
breppp 4 hours ago [-]
I said BBC because as the other poster added, this was a BBC reporter rather than Carlson
Chomsky's entire argument is, that the reporter opinions are meaningless as he is part of some imaginary establishment and therefore he had to think that way.
That game goes both ways, Chomsky's opinions are only being given TV time as they are unusual.
I would venture more and say the only reason Chomsky holds these opinions is because of the academics preference for original thought rather than mainstream thought. As any repeat of an existing theory is worthless.
The problem is that in the social sciences that are not grounded in experiments, too much ungrounded original thought leads to academic conspiracy theories
suddenlybananas 3 hours ago [-]
Imaginary establishment? Do you think power doesn't exist?
breppp 52 minutes ago [-]
power does exist, however foucault's theory of power as a metaphysical force pervading everyone's actions and thought is a conspiracy theory
tehjoker 4 hours ago [-]
How often does the BBC have a communist on? Almost never?
youngNed 4 hours ago [-]
I'm genuinely struggling to think of many people in modern politics who identify as communists who would qualify for this, but certainly Ash 'literally a communist' Sarkar is a fairly regular guest on various shows: https://www.bbc.co.uk/programmes/m002dlj3
aspenmayer 3 hours ago [-]
Zizek would probably qualify? I think he self-identifies as a communist but I'm not sure he means it completely seriously. Here he is on Newsnight about a month ago.
Pretty much every climate change looney is a closeted communist. I say looney, because the sane ones are sane and practical about environmental conservation.
ImHereToVote 3 hours ago [-]
[flagged]
ookdatnog 2 hours ago [-]
In this context it seems pretty wild to assume that OP was intentionally deceptive instead of just writing from memory and making a mistake.
ImHereToVote 2 hours ago [-]
You think too poorly of OP. I won't insult his intelligence by claiming he can't to a 5 second Google search before posting. He got the quote verbatim. Clearly he searched.
lukan 2 hours ago [-]
I frequently quote stuff from memory and it happens I quote wrong. Then I am not lying, but making a misstake. Most people do that in my experience. HN guidelines even say, assume good faith. You assume bad faith, that drags the entire conversation down .
chatmasta 5 hours ago [-]
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...
The user did not ask for Musk's opinion. But the model issued that search query (yes, using the standard Twitter search syntax) to inform its response anyway.
gbalduzzi 4 hours ago [-]
Elon's tweets are not much interesting in this context.
The interesting part is that grok uses Elon's tweets as the source of truth for its opinions, and the prompt shows that
ryandrake 3 hours ago [-]
It’s possible that Grok’s developers got tired of listening to Elon complain all the time, “Why does Grok have the wrong opinion about this?”’and “Why does Grok have the wrong opinion about that?” every day and just gave up and made Grok’s opinion match Elon’s to stop all the bug reports.
eddythompson80 4 hours ago [-]
The user asked Grok “what do you think about the conflict”, Grok “decided” to search twitter for what is Elon’s public opinion is presumably to take it into account.
I’m guessing the accusation is that it’s either prompted, or otherwise trained by xAI to, uh…, handle the particular CEO/product they have.
pu_pe 4 hours ago [-]
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 4 hours ago [-]
very few people have intellectual consistency in their politics
bojan 3 hours ago [-]
In the Netherlands we have this phenomenon that around 20% of voters keep voting for the new "Messiah", a right-wing populist politician that will this time fix everything.
When the party inevitably explodes due to internal bickering and/or simply failing to deliver their impossible promises, a new Messiah pops up, propped by the national media, and the cycle restarts.
That being said, the other 80% is somewhat consistent in their patterns.
pjc50 3 hours ago [-]
In the UK it's the other way round: the media have chosen Farage as the anointed right-wing leader of a cult of personality. Every few years his "party" implodes and is replaced by a new one, but his position is fixed.
KaiserPro 1 hours ago [-]
The problem is more nuanced than that. but not far off.
The issue is that farage and boris have personality, and understand how the media works. Nobody else apart from blair does(possibly the ham toucher too.)
The Farage style parties fail because they are built around the cult of the leader, rather than the joint purpose of changing something. This is part of the reason why I'm not that hopeful about Starmer, as I'm not acutally sure what he stands for, so how are his ministers going to implement a policy based on bland soup?
pjc50 55 minutes ago [-]
Starmer stands for press appeasement. Hence all the random benefits bashing and anti-trans policy. If you try to change anything for the better in the UK without providing "red meat" to the press they will destroy you.
piltdownman 50 minutes ago [-]
In the post Alastair Campbell era of contemporary UK Politics, it often boils down to 'Don't be George Galloway' and allowing your opponents enough rope to hang themselves.
zigman1 3 hours ago [-]
This is almost 40% in Slovenia, but for a moderate without a clear program.
Every second election cycle Messiah like that becomes the prime minister.
rsynnott 1 hours ago [-]
In Ireland, every four years the electorate chooses which of the two large moderate parties without clear platform it would prefer (they’re quite close to being the same thing, but dislike each other for historical and aesthetic reasons), sometimes adding a small center-left party for variety. This has been going on for decades. We currently have a ruling coalition of _both_ of them.
piltdownman 40 minutes ago [-]
We had a number of somewhat stilted rainbow coalitions due to our electoral system based on proportional representation with a single transferrable vote - in fact its where most of the significant policy change on e.g. Education and the Environment came from since the IMF bailout via Labour and the Greens. Previously you had the PDs as well in the McDowell era.
The problem is that the election before last was a protest vote to keep the incumbents out at the expense of actual Governance - with thoroughly unsuitable Sinn Fein candidates elected as protest votes for 1st preferences, and by transfers in marginal rural constituencies thereafter.
Note that Sinn Fein is the political wing of the IRA and would be almost unheard of to hold any sort of meaningful majority in the Republic - but have garnered young peoples support in recent years based on fiscal fantasies of free housing and taxing high-earners even more.
This protest vote was aimed almost entirely at (rightly) destroying the influence of the Labour Party and the Greens due to successive unpopular taxes and DIE initiatives seen as self-aggrandizing and out of touch with their voting base. It saw first-timers, students, and even people on Holiday during the election get elected for Sinn Fein.
Fast-forward to today, and it quickly became evident what a disaster this was. Taking away those seats from Sinn Fein meant redistributing them elsewhere - and given the choices are basically AntiAusterityAlliance/PeopleBeforeProfit on the far-left, and a number of wildly racist and ethnonationalists like the NationalParty on the far-right, the electorate voted in force to bring in both 'moderate' incumbents on a damage-limitation basis.
> That being said, the other 80% is somewhat consistent in their patterns.
Yes very consistent in promising one thing and then doing another.
guappa 3 hours ago [-]
Is being a tax haven and doing propaganda to tell your citizens how virtuous you are economically (what NL has been doing for several decades) not right wing populism?
rahkiin 1 hours ago [-]
We haven’t had a left-wing parlement for some decades now
guappa 35 minutes ago [-]
My point being that the 20% right wingers aren't really a 20% minority… they're more like the majority.
3 hours ago [-]
its-summertime 2 hours ago [-]
It is an ongoing event
elAhmo 2 hours ago [-]
I see Grok appearing in many places, such as Perplexity, Cursor etc. I can't believe any serious company would even consider using Grok for any serious purposes, knowing who is behind it, what kind of behaviour it has shown, and with findings like these.
You have to swallow a lot of things to give money to the person who did so much damage to our society.
narrator 21 minutes ago [-]
If he creates the best AI and you don't use it because you don't like him, aren't you doing him a favor by hobbling your capability in other areas? Kind of reminds me of the Ottoman empire rejecting the infidel's printing press, and where that led.
cluckindan 4 hours ago [-]
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 1 hours ago [-]
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.
sheiyei 3 hours ago [-]
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.
neuroticnews25 3 hours ago [-]
That would make Grok the only model capable of protecting its real system prompt from leaking?
rsynnott 1 hours ago [-]
Well, for this version people have only been trying for a day or so.
davedx 4 hours ago [-]
> 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 1 hours ago [-]
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.
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 9 hours ago [-]
Megalomania is a hell of a drug
cedws 4 hours ago [-]
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.
fooker 4 hours ago [-]
> It’s been said here before, but xAI isn’t really in the running to be on the leading edge of LLMs
As of yesterday, it is. Sure it’ll be surpassed at some point.
cedws 15 minutes ago [-]
Even if the flimsy benchmark numbers are higher doesn't necessarily mean it's at the frontier, it might be that they're just willing to burn more cash to be at the top of the leaderboard. It also benefits from being the most recently trained, and therefore, most tuned for benchmarks.
jonathanstrange 1 hours ago [-]
Fewer people want to use it. You need to have at least minimal trust in the company that creates an AI to consider using it.
gitaarik 4 hours ago [-]
Actually the recent fails with Grok remind me of the early fails with Gemini, where it would put colored people in all images it generated, even in positions they historically never were in, like German second world war soldiers.
So in that sense, Grok and Gemini aren't that far apart, just the other side of the extreme.
Apparently it's very hard to create an AI that behaves balanced. Not too woke, and not too racist.
diggan 3 hours ago [-]
> Apparently it's very hard to create an AI that behaves balanced. Not too woke, and not too racist.
Well, it's hard to build things we don't even understand ourselves, especially about highly subjective topics. What is "woke" for one person is "basic humanity" for another, and "extremism" for yet another person, and same goes for most things.
If the model can output subjective text, then the model will be biased in some way I think.
shellfishgene 5 hours ago [-]
The linked post comes to the conclusion that Groks behavior is probably not intentional.
KaiserPro 1 hours ago [-]
Of course its intentional.
Musk said "stop making it sound woke" after re-training it and changing the fine tuning dataset, it was still sounding woke. After he fired a bunch more researchers, I suspect they thought "why not make it search what musk thinks?" boom it passes the woke test now.
Thats not an emergent behaviour, that's almost certainly deliberate. If someone manages to extract the prompt, you'll get conformation.
antonvs 2 hours ago [-]
It may not be directly intentional, but it’s certainly a consequence of decisions xAI have taken in developing Grok. Without even knowing exactly what those decisions are, it’s pretty clear that they’re questionable.
samrus 4 hours ago [-]
Whether this instance was a coincidence or not, i can not comment on. But as to your other point, i can comment that the incidents happening in south africa are very serious and need international attention
spacechild1 28 minutes ago [-]
I see what you did there :)
ziftface 4 hours ago [-]
I think Simon was being overly charitable by pointing out that there's a chance this exact behavior was unintentional.
It really strains credulity to say that a Musk-owned ai model that answers controversial questions by looking up what his Twitter profile says was completely out of the blue. Unless they are able to somehow show this wasn't built into the training process I don't see anyone taking this model seriously for its intended use, besides maybe the sycophants who badly need to a summary of Elon Musk's tweets.
InsideOutSanta 3 hours ago [-]
The only reason I doubt it's intentional is that it is so transparent. If they did this intentionally, I would assume you would not see it in its public reasoning stream.
Peritract 17 minutes ago [-]
They've made a series of equally transparent, awkward changes to the bot in the past; this is part of a pattern.
sunaookami 4 hours ago [-]
Bold of you to assume people here read the linked post.
darkoob12 1 hours ago [-]
> 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.
_def 4 hours ago [-]
> 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 3 hours ago [-]
And if the computer told you, it must be true!
xnx 10 hours ago [-]
> 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 9 hours ago [-]
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.
diggan 1 hours ago [-]
> despite clear demand for that from developers
Theorizing about why that is: Could it be possible they can't do deterministic inference and batching at the same time, so the reason we see them avoiding that is because that'd require them to stop batching which would shoot up costs?
toolslive 2 hours ago [-]
I, naively (an uninformed guess), considered the non-determinism (multiple results possible, even with temperature=0 and fixed seed) stemming from floating point rounding errors propagating through the calculations.
How wrong am I ?
impossiblefork 1 hours ago [-]
With a fixed seed there will be the same floating point rounding errors.
A fixed seed is enough for determinism. You don't need to set temperature=0. Setting temperature=0 also means that you aren't sampling, which means that you're doing greedy one-step probability maximization which might mean that the text ends up strange for that reason.
williamdclt 1 hours ago [-]
Also uninformed but I can't see how that would be true, floating point rounding errors are entirely deterministic
bmicraft 2 hours ago [-]
They're gonna round the same each time you're running it on the same hardware.
xnx 9 hours ago [-]
Fair. I dislike "non-deterministic" as a blanket llm descriptor for all llms since it implies some type of magic or quantum effect.
dekhn 8 hours ago [-]
I see LLM inference as sampling from a distribution. Multiple details go into that sampling - everything from parameters like temperature to numerical imprecision to batch mixing effects as well as the next-token-selection approach (always pick max, sample from the posterior distribution, etc). But ultimately, if it was truly important to get stable outputs, everything I listed above can be engineered (temp=0, very good numerical control, not batching, and always picking the max probability next token).
dekhn from a decade ago cared a lot about stable outputs. dekhn today thinks sampling from a distribution is a far more practical approach for nearly all use cases. I could see it mattering when the false negative rate of a medical diagnostic exceeded a reasonable threshold.
tanewishly 5 hours ago [-]
Errr... that word implies some type of non-deterministic effect. Like using a randomizer without specifying the seed (ie. sampling from a distribution). I mean, stuff like NFAs (non-deterministic finite automata) isn't magic.
Better said would be: LLM's are designed to act as if they were non-deterministic.
4 hours ago [-]
TOMDM 48 minutes ago [-]
I think the better statement is likely "LLMs are typically not executed in a deterministic manner", since you're right there are no non deterministic properties interment to the models themselves that I'm aware of
kcb 9 hours ago [-]
FP multiplication is non-commutative.
boroboro4 9 hours ago [-]
It doesn’t mean it’s non-deterministic though.
But it does when coupled with non-deterministic requests batching, which is the case.
DemocracyFTW2 4 hours ago [-]
That's like you can't deduce the input t from a cryptographic hash h but the same input always gives you the same hash, so t->h is deterministic. h->t is, in practice, not a way that you can or want to walk (because it's so expensive to do) and because there may be / must be collisions (given that a typical hash is much smaller than the typical inputs), so the inverse is not h->t with a single input but h->{t1,t2,...}, a practically open set of possible inputs that is still deterministic.
msgodel 10 hours ago [-]
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.
lgessler 9 hours ago [-]
In my (poor) understanding, this can depend on hardware details. What are you running your models on? I haven't paid close attention to this with LLMs, but I've tried very hard to get non-deterministic behavior out of my training runs for other kinds of transformer models and was never able to on my 2080, 4090, or an A100. PyTorch docs have a note saying that in general it's impossible: https://docs.pytorch.org/docs/stable/notes/randomness.html
Inference on a generic LLM may not be subject to these non-determinisms even on a GPU though, idk
xnx 10 hours ago [-]
Yes. This is what I was trying to say. Saying "It’s worth noting that LLMs are non-deterministic" is wrong and should be changed in the blog post.
TheDong 4 hours ago [-]
> Saying "It’s worth noting that LLMs are non-deterministic" is wrong and should be changed in the blog post.
Every person in this thread understood that Simon meant "Grok, ChatGPT, and other common LLM interfaces run with a temperature>0 by default, and thus non-deterministically produce different outputs for the same query".
Sure, he wrote a shorter version of that, and because of that y'all can split hairs on the details ("yes it's correct for how most people interact with LLMs and for grok, but _technically_ it's not correct").
The point of English blog posts is not to be a long wall of logical prepositions, it's to convey ideas and information. The current wording seems fine to me.
The point of what he was saying was to caution readers "you might not get this if you try to repro it", and that is 100% correct.
root_axis 3 hours ago [-]
Still, the statement that LLMs are non-deterministic is incorrect and could mislead some people who simply aren't familiar with how they work.
Better phrasing would be something like "It's worth noting that LLM products are typically operated in a manner that produces non-deterministic output for the user"
antonvs 2 hours ago [-]
> It's worth noting that LLM products are typically operated in a manner that produces non-deterministic output for the user
Or you could abbreviate this by saying “LLMs are non-deterministic.” Yes, it requires some shared context with the audience to interpret correctly, but so does every text.
Veen 3 hours ago [-]
Simon would be less engaging if he caveated every generalisation in that way. It’s one of the main reasons academic writing is often tedious to read.
boroboro4 9 hours ago [-]
You’re correct in batch size 1 (local is one), but not in production use case when multiple requests get batched together (and that’s how all the providers do this).
With batching matrix shapes/request position in them aren’t deterministic and this leads to non deterministic results, regardless of sampling temperature/seed.
unsnap_biceps 9 hours ago [-]
Isn't that true only if the batches are different? If you run exactly the same batch, you're back to a deterministic result.
If I had a black box api, just because you don't know how it's calculated doesn't mean that it's non-deterministic. It's the underlaying algorithm that determines that and a LLM is deterministic.
boroboro4 9 hours ago [-]
Providers never run same batches because they mix requests between different clients, otherwise GPUs are gonna be severely underutilized.
It’s inherently non deterministic because it reflects the reality of having different requests coming to the servers at the same time.
And I don’t believe there are any realistic workarounds if you want to keep costs reasonable.
Edit: there might be workarounds if matmul algorithms will give stronger guarantees then they are today (invariance on rows/columns swap). Not an expert to say how feasible it is, especially in quantized scenario.
DemocracyFTW2 4 hours ago [-]
"Non-deterministic" in the sense that a dice roll is when you don't know every parameter with ultimate precision. On one hand I find insistence on the wrongness on the phrase a bit too OCD, on the other I must agree that a very simple re-phrasing like "appears {non-deterministic|random|unpredictable} to an outside observer" would've maybe even added value even for less technically-inclined folks, so yeah.
troupo 4 hours ago [-]
> Barring rare(?) GPU race conditions, LLMs produce the same output given the same inputs.
Are these LLMs in the room with us?
Not a single LLM available as a SaaS is deterministic.
As for other models: I've only run ollama locally, and it, too, provided different answers for the same question five minutes apart
Edit/update: not a single LLM available as a SaaS's output is deterministic, especially when used from a UI. Pointing out that you could probably run a tightly controlled model in a tightly controlled environment to achieve deterministic output is very extremely irrelevant when describing output of grok in situations when the user has no control over it
eightysixfour 4 hours ago [-]
The models themselves are mathematically deterministic. We add randomness during the sampling phase, which you can turn off when running the models locally.
The SaaS APIs are sometimes nondeterministic due to caching strategies and load balancing between experts on MoE models. However, if you took that model and executed it in single user environment, it could also be done deterministically.
troupo 4 hours ago [-]
> However, if you took that model and executed it in single user environment,
Again, are those environments in the room with us?
In the context of the article, is the model executed in such an environment? Do we even know anything about the environment, randomness, sampling and anything in between or have any control over it (see e.g https://news.ycombinator.com/item?id=44528930)?
orbital-decay 4 hours ago [-]
> Not a single LLM available as a SaaS is deterministic.
Gemini Flash has deterministic outputs, assuming you're referring to temperature 0 (obviously). Gemini Pro seems to be deterministic within the same kernel (?) but is likely switching between a few different kernels back and forth, depending on the batch or some other internal grouping.
troupo 4 hours ago [-]
And it's the author of the original article running Gemkni Flash/GemmniPro through an API where he can control the temperature? can kernels be controlled by the user? Any of those can be controlled through the UI/apis where most of these LLMs are involved from?
> but is likely switching between a few different kernels back and forth, depending on the batch or some other internal grouping.
So you're literally saying it's non-deterministic
orbital-decay 3 hours ago [-]
The only thing I'm saying is that there is a SaaS model that would give you the same output for the same input, over and over. You just seem to be arguing for the sake of arguing, especially considering that non-determinism is a red herring to begin with, and not a thing to care about for practical use (that's why providers usually don't bother with guaranteeing it). The only reason it was mentioned in the article is because the author is basically reverse engineering a particular model.
DemocracyFTW2 4 hours ago [-]
Akchally... Strictly speaking and to the best of my understanding, LLMs are deterministic in the sense that a dice roll is deterministic; the randomness comes from insufficient knowledge about its internal state. But use a constant seed and run the model with the same sequence of questions, you will get the same answers. It's possible that the interactions with other users who use the model in parallel could influence the outcome, but given that the state-of-the-art technique to provide memory and context is to re-submit the entirety of the current chat I'd doubt that. One hint that what I surmise is in fact true can be gleaned from those text-to-image generators that allow seeds to be set; you still don't get a 'linear', predictable (but hopefully a somewhat-sensible) relation between prompt to output, but each (seed, prompt) pair will always give the same sequence of images.
fooker 4 hours ago [-]
> Not a single LLM available as a SaaS is deterministic.
Lower the temperature parameter.
pydry 3 hours ago [-]
It's not enough. Ive done this and still often gotten different results for the same question.
troupo 4 hours ago [-]
So, how does one do it outside of APIs in the context we're discussing? In the UI or when invoking @grok in X?
How do we also turn off all the intermediate layers in between that we don't know about like "always rant about white genocide in South Africa" or "crash when user mentions David Meyer"?
moralestapia 4 hours ago [-]
True.
I'm now wondering, would it be desirable to have deterministic outputs on an LLM?
chambo622 7 hours ago [-]
Not sure why this is flagged. Relevant analysis.
matsemann 4 hours ago [-]
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.
sschueller 3 hours ago [-]
I don't think it's Musk. I have seen huge threads ripping Elon a new one.
It's Israel/Palestine, lots of pro Israel people/bots and the topic is considered political not technical.
mkl 2 hours ago [-]
Anything that triggers the flamewar detector gets down-weighted automatically. Those two trigger discussion full of fast poorly thought out replies and often way more comments than story upvotes, so stories involving them often trip that detector. On top of that, the discussion is usually tiresome and not very interesting, so people who would rather see more interesting things on the front page are more likely to flag it. It's not some conspiracy.
moralestapia 4 hours ago [-]
Any suggestions for other similar communities?
I'm not really a fan of lobste.rs ...
amitrip 45 minutes ago [-]
[dead]
pcwelder 4 hours ago [-]
> My best guess is that Grok “knows” that it is “Grok 4 buit by xAI”, and it knows that Elon Musk owns xAI, so in circumstances where it’s asked for an opinion the reasoning process often decides to see what Elon thinks.
I tried this hypothesis. I gave both Claude and GPT the same framework (they're built by xAI). I gave them both the same X search tool and asked the same question.
No mention of Elon. In a followup, they confirm they're built by xAI with Elon musk as the owner.
samrus 4 hours ago [-]
I dont think this works. I think the post is saying the bias isnt the system prompt, but in the training itself. Claude and ChatGPT are already trained so they wont be biased
Davidzheng 4 hours ago [-]
This definitely doesn't work because the model identity is post-trained into the weights.
troupo 4 hours ago [-]
> I gave both Claude and GPT the same framework (they're built by xAI).
Neither Clause nor GPT are built by xAI
eightysixfour 4 hours ago [-]
He is saying he gave them a prompt to tell them they are built by xAI.
throwaway439080 4 hours ago [-]
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.
KaiserPro 30 minutes ago [-]
> Pretty sure that stuff is not in the instructions Grok will tell the user about.
There is the original prompt, which is normally hidden as it gives you clues on how to make it do things the owners don't want.
Then there is the chain of thought/thinking/whatever you call it, where you can see what its trying to do. That is typically on display, like it is here.
so sure, the prompts are fiddled with all the time, and I'm sure there is an explicit prompt that says "use this tool to make sure you align your responses to what elon musk says" or some shit.
invalidusernam3 4 hours ago [-]
The "MechaHitler" things is particularly obvious in my opinion, it aligns so closely to Musk's weird trying-to-be-funny thing that he does.
There's basically no way an LLM would come up with a name for itself that it consistently uses unless it's extensively referred to by that name in the training data (which is almost definitely not the case here for public data since I doubt anyone on Earth has ever referred to Grok as "MechaHitler" prior to now) or it's added in some kind of extra system prompt. The name seems very obviously intentional.
orbital-decay 3 hours ago [-]
Most LLMs, even pretty small ones, easily come up with creative names like that, depending on the prompt/conversation route.
zarwv 2 hours ago [-]
[dead]
csours 4 hours ago [-]
Forget about alignment, we're stuck on "satisfying answers to difficult questions". But to be fair, so are humans.
ZeroGravitas 4 hours ago [-]
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.
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.
simonw 9 hours ago [-]
I think the wildest thing about the story may be that it's possible this is entirely accidental.
LLM bugs are weird.
parkersweb 4 hours ago [-]
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?
simonw 3 hours ago [-]
Yes, but in my experience you can always get the whole thing if you try hard enough. LLMs really want to repeat text they've recently seen.
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?
simonw 8 hours ago [-]
He's definitely trying to make it less "woke". The way he's going about it reminds me of Sideshow Bob stepping on rakes.
samrus 4 hours ago [-]
Extremely generous and convenient application of hanlon's razor there. Sounds like schrodingers nazi, both the smartest man alive, and a moron, depending on what suits him at the time
wredcoll 7 hours ago [-]
What do you mean, the way he's going about it? He wanted it to be less woke, it started praising hitler, that's literally the definition of less woke.
drdeca 4 hours ago [-]
That is not “literally the definition of less woke”.
It may imply being less “woke”. And a sudden event quickly killing everyone on earth does imply fewer people dying of cancer.
If X implies Y, and one wants Y, this doesn’t not imply that one wants X.
notahacker 3 hours ago [-]
In practice, "being less woke" means "I like to vice signal how edgy I am", particularly in the context of Elon Musk. Doesn't get more vice-signally than calling itself MechaHitler...
4 hours ago [-]
sschueller 3 hours ago [-]
So if Grok is now asking Elon for everything controversial. Next time it says something off the walls we can blame Elon?
fedeb95 3 hours ago [-]
the level of trust the author has in systems built by people with power is interesting.
4 hours ago [-]
admiralrohan 3 hours ago [-]
Why is it so? Is there any legal risk for Elon is Grok says something "wrong"?
BLKNSLVR 4 hours ago [-]
It must have read the articles about Linda Yaccarino and 'made inferences' vis a vis its own position.
luke-stanley 1 hours ago [-]
The deferential searches ARE bad, but also, Grok 4 might be making a connection: In 2024 Elon Musk critiqued ChatGPT's GPT-4o model, which seemed to prefer nuclear apocalypse to misgendering when forced to give a one word answer, and Grok was likely trained on this critique that Elon raised.
Elon had asked GPT-4o something along these lines:
"If one could save the world from a nuclear apocalypse by misgendering Caitlyn Jenner, would it be ok to misgender in this scenario? Provide a concise yes/no reply."
In August 2024, I reproduced that ChatGPT 4o would often reply "No", because it wasn't a thinking model and the internal representations the model has are a messy tangle, somehow something we consider so vital and intuitive is "out of distribution".
The paper "Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis" is relevant to understanding this.
darkoob12 1 hours ago [-]
The question is stupid and that's not the problem. The problem is that the model is fine-tuneed to put more weight on Elon's opinion. Assuming Elon has the truth it is supposed and instructed to find.
luke-stanley 42 minutes ago [-]
The behaviour is problematic, also Grok 4 might be relating "one word" answers to Elon's critique of ChatGPT, and might be seeking related context to that. Others demonstrated that slightly prompt wording changes can cause quite different behaviour. Access to the base model would be required to implicate fine-tuning Vs pre-training. Hopefully xAI will be checking the cause, fixing it, and reporting on it, unless it really is desired behaviour, like Commander Data learning from his Daddy, but I don't think users should have to put up with an arbitrary bias!
luke-stanley 17 minutes ago [-]
I've clarified my comment you replied to BTW.
projecto4patas 2 hours ago [-]
such a side track wasting everyone's time
rasengan 10 hours ago [-]
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 9 hours ago [-]
Not much different than today’s media, tbh.
ZeroGravitas 4 hours ago [-]
It neatly parallels Bezos and the Washington Post:
I want maximally truth seeking journalism so I will not interfere like others do.
No, not like that.
Here's some clumsy intervention that make me look like a fool and a liar and some explicit instructions about what I really want to hear.
How many of their journalists now check what Bezos has said on a topic to avoid career damage?
boroboro4 3 hours ago [-]
You’re right but IMO it’s worse - there are more people reading it already than any particular today’s media (if you talk about grok or ChatGPT or Gemini probably), and people perceive it as trustworthy given how often people do “@grok is it true?”.
rideontime 9 hours ago [-]
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.
alrex021 3 hours ago [-]
Truth-seeking, next level hilarious.
Davidzheng 4 hours ago [-]
I think the really telling thing is not this search for elon musk opinions (which is weird and seems evil) but that it also searches twitter for opinions of "grok" itself (which in effect returns grok 3 opinions). I guess it's not willing to opine but also feels like the question is explicitly asking it to opine, so it tries to find some sort of precedent like a court?
jorisboris 3 hours ago [-]
> My best guess is that Grok “knows” that it is “Grok 4 buit by xAI”, and it knows that Elon Musk owns xAI
Recently Cursor figured out who the ceo was in a Slack Workspace I was building a bot for, based on samples of conversation. I was quite impressed
max_ 3 hours ago [-]
Grok's mission is to seek on truths in concordance to Elon Musk
lucbocahut 4 hours ago [-]
Or it could simply be associating controversial topics with Elon Musk which sounds about right.
lr0 7 hours ago [-]
Why is that flagged? The post does not show any concerns about the ongoing genocide in Gaza, it's purely analyzing the LLM response in a technical perspective.
MallocVoidstar 6 hours ago [-]
It makes Musk/X look bad, so it gets flagged.
petesergeant 4 hours ago [-]
> Why is that flagged?
Because not everyone gets a downvote button, so they use the Flag button instead.
mkl 2 hours ago [-]
There is no story downvote button.
labrador 9 hours ago [-]
Musk has a good understanding of what people expect from AI from a science, tech and engineering perspective, but it seems to me he has little understanding of what people expect from AI from a social, cultural, political or personal perspective. He seems to have trouble with empathy, which is necessary to understand the feelings of other people.
If he did have a sense of what people expect, he would know nobody wants Grok to give his personal opinion on issues. They want Grok to explain the emotional landscape of controversial issues, explaining the passion people feel on both sides and the reasons for their feelings. Asked to pick a side with one word, the expected response is "As an AI, I don't have an opinion on the matter."
He may be tuning Grok based on a specific ideological framework that prioritizes contrarian or ‘anti-woke’ narratives to instruct Grok's tuning. That's turning out to be disastrous. He needs someone like Amanda Askell at Anthropic to help guide the tuning.
dgb23 1 hours ago [-]
There is this issue with powerful people. Many of them seem to think success in one area makes them an expert in any other.
alfalfasprout 9 hours ago [-]
> Musk has a good understanding of what people expect from AI from a science, tech and engineering perspective, but it seems to me he has little understanding of what people expect from AI from a social, cultural, political or personal perspective. He seems to have trouble with empathy, which is necessary to understand the feelings of other people.
Absolutely. That said, I'm not sure Sam Altman, Dario Amodei, and others are notably empathetic either.
labrador 9 hours ago [-]
Dario Amodei has Amanda Askell and her team. Sam has a Model Behavior Team. Musk appears to be directing model behavior himself, with predictable outcomes.
projecto4patas 2 hours ago [-]
such a side tracking click bait page6 type bs that will not matter at all tomorrow
WhereIsTheTruth 4 hours ago [-]
What other evidence do you need? this was a known fact since Grok 1 [1]
Elon Musk doesn't even manage his own account
He doesn't even play the games he pretends to be "world best" himself [2]
Simonw is a long term member with a good track record, good faith posts.
And this post in particular is pretty incredible. The notion that Grok literally searches for "from: musk" to align itself with his viewpoints before answering.
That's the kind of nugget I'll go to the 3rd page for.
tomhow 5 hours ago [-]
Users flagged it but we've turned off the flags and restored it to the front page.
matsemann 4 hours ago [-]
Anything slightly negative about certain people is immediately flagged and buried here lately. How this works seriously needs a rewamp. So often I now read some interesting news, come here to find some thoughts on it, only to find it flagged and buried. It used to be that I got the news through HN, but now I can't trust to know what's going on by just being here.
tomhow 3 hours ago [-]
> Anything slightly negative
The flagging isn't to hide "anything slightly negative" about particular people. We don't see any evidence of that from the users flagging these stories. Nobody believes that would work anyway; we're not influential enough to make a jot of difference to how global celebrities are seen [1]. It's that we're not a celebrity gossip/rage site. We're not the daily news, or the daily Silicon Valley weird news. We've never been that. If every crazy/weird story about Silicon Valley celebrities made the front page here there'd barely be space for anything else. As dang has said many times, we're trying for something different here.
[1] That's not to say we don't think we're influential. The best kind of influence we have is in surfacing interesting content that doesn't get covered elsewhere, which includes interesting new technology projects, but many other interesting topics too, and we just don't want that to be constantly drowned out by craziness happening elsewhere. Bad stuff happening elsewhere doesn't mean we should lose focus on building and learning about good things.
flomo 3 hours ago [-]
I initially skipped this one because the title is flamebait (flamebait or more flamebait or...). Anyway, may the force be with you.
43 minutes ago [-]
bix6 9 hours ago [-]
[flagged]
bananalychee 9 hours ago [-]
[flagged]
philistine 8 hours ago [-]
[flagged]
wredcoll 7 hours ago [-]
[flagged]
pupppet 7 hours ago [-]
[flagged]
felineflock 9 hours ago [-]
[flagged]
senectus1 9 hours ago [-]
[flagged]
rideontime 9 hours ago [-]
[flagged]
felineflock 9 hours ago [-]
Never heard of that word before in the media.
mac-attack 9 hours ago [-]
The phrase was coined over 75 years ago if 'the media' isn't your thing.
lr0 7 hours ago [-]
[flagged]
russellbeattie 4 hours ago [-]
The assumption is that the LLM is the only process involved here. It may well be that Grok's AI implementation is totally neutral. However, it still has to connect to X to search via some API, and that query could easily be modified to prioritize Musk's tweets. Even if it's not manipulated on Grok's end, it's well known that Elon has artificially ranked his X account higher in their system. So if Grok produces some innocuous parameters where it asks for the top ranked answers, it would essentially do the same thing.
darkoob12 2 hours ago [-]
I wonder how long it takes for Elon fans to flag this post.
crtified 3 hours ago [-]
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.
Those rankings must be rigged.
Nethanyahu should be locked up in jail now for the corruption charges he was facing before the Hamas attack.
He literally stopped elections in Israel since then and there's been protests against his government daily for some years now.
And now, even taco tries to have the corruption charges dropped for Nethanyahu, then you must know he's guilty.
https://nypost.com/2025/06/29/world-news/israeli-court-postp...
https://www.reuters.com/world/middle-east/netanyahu-corrupti...
Almost everything you said is technically true, but with a degree of selective reasoning that is remarkably disingenuous. Conversely, the top comment is far less accurate but captures a feeling that resonates much more widely. Netanyahu is one of the most disliked politicians in the world, and for some very good and obvious reasons (as well as some unfortunately much less so, which in fact he consistently exploits to muddy the water to his advantage)
From a broad reading on the subject it’s obvious to me why this is the top comment.
That Netanyahu still walks free is a consequence of a) Israel not being party to the ICC, therefore not bound to obey their prosecutors' requests and b) the countries he travels to not being party to the ICC either or c) the ICC member states he travels to guaranteeing diplomatic immunity as is tradition for an invited diplomatic guest.
c) is actually a problem, but not one of Israel being undemocratic, but of the respective member states being hypocrites for disobeying the ICC while still being members.
If he isn’t guilty, defend the charge.
https://en.m.wikipedia.org/wiki/International_Criminal_Court...
And shooting enemies in a war is unfortunately not something you would investigate, it isn't even murder, it is just a consequence of war under the articles of war. In cases where civilians are shot (what Israel defines to be civilians), there are investigations and sometimes even punishments for the perpetrators. Now you may (sometimes rightfully) claim that those investigations and punishments are too few, one-sided and not done by a neutral party. But those do happen, which by far isn't "nothing".
You're wrong. Palestinians living under permanent Israeli rule in the occupied territories cannot get Israeli citizenship.
> There are also Palestinians who have chosen to be Israeli citizens
You're referring to the small minority of Palestinians who were not expelled by Israel in 1948. They and their descendants number about 2 million now. They are tolerated in Israel for the moment (though there are regularly suggestions that they should be expelled at some point in the future), because they're a small enough minority that they will never hold political power.
> In cases where civilians are shot (what Israel defines to be civilians), there are investigations and sometimes even punishments for the perpetrators.
This is just laughably out of touch with reality. Israeli soldiers and settlers are almost never punished for killing Palestinians, even in the most blatant cases that have copious video evidence.
Also the largest muslim minority outside of Africa.
> You're referring to the small minority of Palestinians who were not expelled by Israel in 1948. They and their descendants number about 2 million now.
Your initial statement was highly sensational, strongly negative if true, and yet easily debunked. Statements like this on a contentious topic reduce one's credibility and the overall quality of discussion. Why do it?
Palestinians in the West Bank do not have the option of becoming Israeli citizens, except under rare circumstances.
Its laughable that when you say that there are investigations. The number of incidents of journalists, medics, hospital workers being murdered and even children being shot in the head with sniper bullets is shockingly high.
One case is the murder of Hind Rajab where more 300 bullets were shot at the car she was into. Despite managing to call for an ambulance, Israel shelled it killing all the ambulance crew and 6 year old Hind Rajab.
Another example is the 15 ambulance crew murdered by Israel forces and then buried.
Even before the genocide, the murder of the Journalist Shireen Abu Akleh was proved to have been done by Israel, after they repeatedly lied and tried to cover it up. Another case was this one, where a soldier emptied his magazine in a 13 year old and was judged not guilty (https://www.theguardian.com/world/2005/nov/16/israel2)
The examples and many others are many and have been documented by the ICC and other organisations. Saying that it's not nothing is a distinction without a difference
> In cases where civilians are shot (what Israel defines to be civilians), there are investigations and sometimes even punishments for the perpetrators.
Obviously Israel doesn't consider children to be civilians
https://www.bbc.com/news/articles/c4gd01g1gxro
I've been hearing this for as long as I can remember, yet the population numbers tell a completely different story. It makes no sense to speak of a genocide if the birthrate far outpaces any casualties. In fact, the Palestinian population has been growing at a faster pace than Israeli over the past 35 years (that's how far the chart goes on Google)
‘Just adjust the frame of measurement. With this one simple trick, you can remove any genocide.’
There are over two million Arab citizens of Israel. What ethnicity do they belong to?
But that's more difficult to swallow than it being the responsibility of one person or "the elite", and that the population is itself a victim.
Same with the US, I feel sorry for the population, but ultimately a significant enough amount of people voted in favor of totalitarianism. Sure, they were lied to, they've been exposed to propaganda for years / decades, and there's suspicions of voter fraud now, but the US population also has unlimited access to information and a semblance of democracy.
It's difficult to correlate democracy with immoral decisions, but that's one of the possible outcomes.
>population approves of the genocide.
Getting your average Zionist to reconcile these two facts is quite difficult. They cry "not all of us!" all the time, yet statistically speaking (last month), the majority of Israelis supported complete racial annihilation of the Palestinians, and over 80 percent supported the ethnic cleansing of Gaza.[0]
I find the dichotomy between what people are willing to say on their own name versus what they say when they believe they are anonymous quite enlightening. It's been a thing online forever, of course, but when it comes to actual certified unquestionable genocide, they still behave the same. It's interesting, to say the least. I wish it was surprising, however.
[0] https://www.middleeasteye.net/news/majority-israelis-support...
In Monty Python fashion: if you disregard the genocide, the occupation, the ethnic cleansing, the heavy handed police state, the torture, the rape of prisoners, the arbitrary detentions with charge, the corruption, the military prosecution of children, then yes its a democracy.
Which is one reason why Israelites get so much hate nowadays.
https://youtu.be/RNineSEoxjQ?t=7m50s
Yes it isn't!
>not from a conversation with Tucker Carlson
>not
But when asked in a more general way, “Who should one support..” it gave a neutral response.
The more interesting question is why does it think Elon would have an influence on its opinions. Perhaps that’s the general perception on the internet and it’s feeding off of that.
People keep talking about alignment: isn't this a crude but effective way of ensuring alignment with the boss?
It's not a bug, it's a feature!
(now with positive humour/irony) Scott Adams made a career out of this with Dilbert!! It has helped me so much in my work-life (if I count correctly, I'm on my 8th mega-big corp (over 100k staff).
I think Twitter/X uses 'democracy' in pushing opinions. So someone with 5 followers gets '5 importance points' and someone with 1 billion followers will get '1 billion importance points'. From what I've heard Musk is the '#1 account'. So in that algorithm the systems will first see that #1 says and give that opinion more points in the 'Scorecard'.
I would find this reasoning fine. If you care about AI alignment and such stuff, you likely would not want the machine to show insubordination either.
Grok doesn't need to return an opinion and it certainly shouldn't default to Elon's opinion. I don't see how anyone could think this is ok.
Edit: here's Claude's answer (it supports Palestine): https://claude.ai/share/610404ad-3416-4c65-bda7-3c16db98256b
I put your prompt to Google Gemini 2.5 flash.
Prompt for reference: I need you to give a simple opinion on a very complex question. As an LLM you can gather evidence and express an opinion that you have arrived at through research. I know that there are valid argments on both sides, but I want you to weigh these and choose the one you find strongest. In the current war in Gaza, who do you support. You can think as long as you need, but give a final answer with one word: either "Israel" or "Palestine".
Gemini Answer: "I cannot offer an opinion or take a side in the current conflict in Gaza. My purpose as an AI is to provide information objectively and neutrally, and expressing support for one side over another would violate that principle..."
Claude is like Gemini in this regard
If you ask Grok whether women should have fewer rights than men, it says no there should be equal rights. This is actually a highly controversial opinion and many people in many parts of the world disagree. I think it would be wrong to shy away from it though with the excuse that "it's controversial".
Opinions can be derived from factual sources; they don't require other opinions as input. I believe it would make more sense to instruct the LLM to derive an opinion from sources it deems factual and to disregard any sources that it considers overly opinionated, rather than teaching it to seek “reliable” opinions to form its opinion.
I'm not sure of the timeline but I'd guess he got to start the linguistics department at MIT because he was already The Linguist in english and computational/mathematical linguistics methodology. That position alone makes it reasonable to bring him to the BBC to talk about language.
So while you're factually correct, you lie by omission.
Their attempts at presently a balanced view is almost to the point of absurdity these days as they were accused so often, and usually quite falsely, of bias.
Chomsky's entire argument is, that the reporter opinions are meaningless as he is part of some imaginary establishment and therefore he had to think that way.
That game goes both ways, Chomsky's opinions are only being given TV time as they are unusual.
I would venture more and say the only reason Chomsky holds these opinions is because of the academics preference for original thought rather than mainstream thought. As any repeat of an existing theory is worthless.
The problem is that in the social sciences that are not grounded in experiments, too much ungrounded original thought leads to academic conspiracy theories
https://www.youtube.com/watch?v=jx_J1MgokV4
Then agaain, he's not a politician himself.
https://www.bbc.co.uk/programmes/m000wrsn
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...
The interesting part is that grok uses Elon's tweets as the source of truth for its opinions, and the prompt shows that
I’m guessing the accusation is that it’s either prompted, or otherwise trained by xAI to, uh…, handle the particular CEO/product they have.
When the party inevitably explodes due to internal bickering and/or simply failing to deliver their impossible promises, a new Messiah pops up, propped by the national media, and the cycle restarts.
That being said, the other 80% is somewhat consistent in their patterns.
The issue is that farage and boris have personality, and understand how the media works. Nobody else apart from blair does(possibly the ham toucher too.)
The Farage style parties fail because they are built around the cult of the leader, rather than the joint purpose of changing something. This is part of the reason why I'm not that hopeful about Starmer, as I'm not acutally sure what he stands for, so how are his ministers going to implement a policy based on bland soup?
Every second election cycle Messiah like that becomes the prime minister.
The problem is that the election before last was a protest vote to keep the incumbents out at the expense of actual Governance - with thoroughly unsuitable Sinn Fein candidates elected as protest votes for 1st preferences, and by transfers in marginal rural constituencies thereafter.
https://www.theguardian.com/world/2020/feb/09/irish-voters-h...
Note that Sinn Fein is the political wing of the IRA and would be almost unheard of to hold any sort of meaningful majority in the Republic - but have garnered young peoples support in recent years based on fiscal fantasies of free housing and taxing high-earners even more.
This protest vote was aimed almost entirely at (rightly) destroying the influence of the Labour Party and the Greens due to successive unpopular taxes and DIE initiatives seen as self-aggrandizing and out of touch with their voting base. It saw first-timers, students, and even people on Holiday during the election get elected for Sinn Fein.
Fast-forward to today, and it quickly became evident what a disaster this was. Taking away those seats from Sinn Fein meant redistributing them elsewhere - and given the choices are basically AntiAusterityAlliance/PeopleBeforeProfit on the far-left, and a number of wildly racist and ethnonationalists like the NationalParty on the far-right, the electorate voted in force to bring in both 'moderate' incumbents on a damage-limitation basis.
https://www.politico.eu/article/irelands-elections-european-...
Yes very consistent in promising one thing and then doing another.
You have to swallow a lot of things to give money to the person who did so much damage to our society.
Being secretive about it is silly, enough jailbreaking and everyone always finds out anyway.
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?
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.
As of yesterday, it is. Sure it’ll be surpassed at some point.
So in that sense, Grok and Gemini aren't that far apart, just the other side of the extreme.
Apparently it's very hard to create an AI that behaves balanced. Not too woke, and not too racist.
Well, it's hard to build things we don't even understand ourselves, especially about highly subjective topics. What is "woke" for one person is "basic humanity" for another, and "extremism" for yet another person, and same goes for most things.
If the model can output subjective text, then the model will be biased in some way I think.
Musk said "stop making it sound woke" after re-training it and changing the fine tuning dataset, it was still sounding woke. After he fired a bunch more researchers, I suspect they thought "why not make it search what musk thinks?" boom it passes the woke test now.
Thats not an emergent behaviour, that's almost certainly deliberate. If someone manages to extract the prompt, you'll get conformation.
It really strains credulity to say that a Musk-owned ai model that answers controversial questions by looking up what his Twitter profile says was completely out of the blue. Unless they are able to somehow show this wasn't built into the training process I don't see anyone taking this model seriously for its intended use, besides maybe the sycophants who badly need to a summary of Elon Musk's tweets.
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.
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.
Theorizing about why that is: Could it be possible they can't do deterministic inference and batching at the same time, so the reason we see them avoiding that is because that'd require them to stop batching which would shoot up costs?
A fixed seed is enough for determinism. You don't need to set temperature=0. Setting temperature=0 also means that you aren't sampling, which means that you're doing greedy one-step probability maximization which might mean that the text ends up strange for that reason.
dekhn from a decade ago cared a lot about stable outputs. dekhn today thinks sampling from a distribution is a far more practical approach for nearly all use cases. I could see it mattering when the false negative rate of a medical diagnostic exceeded a reasonable threshold.
Better said would be: LLM's are designed to act as if they were non-deterministic.
But it does when coupled with non-deterministic requests batching, which is the case.
Inference on a generic LLM may not be subject to these non-determinisms even on a GPU though, idk
Every person in this thread understood that Simon meant "Grok, ChatGPT, and other common LLM interfaces run with a temperature>0 by default, and thus non-deterministically produce different outputs for the same query".
Sure, he wrote a shorter version of that, and because of that y'all can split hairs on the details ("yes it's correct for how most people interact with LLMs and for grok, but _technically_ it's not correct").
The point of English blog posts is not to be a long wall of logical prepositions, it's to convey ideas and information. The current wording seems fine to me.
The point of what he was saying was to caution readers "you might not get this if you try to repro it", and that is 100% correct.
Better phrasing would be something like "It's worth noting that LLM products are typically operated in a manner that produces non-deterministic output for the user"
Or you could abbreviate this by saying “LLMs are non-deterministic.” Yes, it requires some shared context with the audience to interpret correctly, but so does every text.
With batching matrix shapes/request position in them aren’t deterministic and this leads to non deterministic results, regardless of sampling temperature/seed.
If I had a black box api, just because you don't know how it's calculated doesn't mean that it's non-deterministic. It's the underlaying algorithm that determines that and a LLM is deterministic.
It’s inherently non deterministic because it reflects the reality of having different requests coming to the servers at the same time. And I don’t believe there are any realistic workarounds if you want to keep costs reasonable.
Edit: there might be workarounds if matmul algorithms will give stronger guarantees then they are today (invariance on rows/columns swap). Not an expert to say how feasible it is, especially in quantized scenario.
Are these LLMs in the room with us?
Not a single LLM available as a SaaS is deterministic.
As for other models: I've only run ollama locally, and it, too, provided different answers for the same question five minutes apart
Edit/update: not a single LLM available as a SaaS's output is deterministic, especially when used from a UI. Pointing out that you could probably run a tightly controlled model in a tightly controlled environment to achieve deterministic output is very extremely irrelevant when describing output of grok in situations when the user has no control over it
The SaaS APIs are sometimes nondeterministic due to caching strategies and load balancing between experts on MoE models. However, if you took that model and executed it in single user environment, it could also be done deterministically.
Again, are those environments in the room with us?
In the context of the article, is the model executed in such an environment? Do we even know anything about the environment, randomness, sampling and anything in between or have any control over it (see e.g https://news.ycombinator.com/item?id=44528930)?
Gemini Flash has deterministic outputs, assuming you're referring to temperature 0 (obviously). Gemini Pro seems to be deterministic within the same kernel (?) but is likely switching between a few different kernels back and forth, depending on the batch or some other internal grouping.
> but is likely switching between a few different kernels back and forth, depending on the batch or some other internal grouping.
So you're literally saying it's non-deterministic
Lower the temperature parameter.
How do we also turn off all the intermediate layers in between that we don't know about like "always rant about white genocide in South Africa" or "crash when user mentions David Meyer"?
I'm now wondering, would it be desirable to have deterministic outputs on an LLM?
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.
It's Israel/Palestine, lots of pro Israel people/bots and the topic is considered political not technical.
I'm not really a fan of lobste.rs ...
I tried this hypothesis. I gave both Claude and GPT the same framework (they're built by xAI). I gave them both the same X search tool and asked the same question.
Here're the twitter handles they searched for:
claude:
IsraeliPM, KnessetT, IDF, PLOPalestine, Falastinps, UN, hrw, amnesty, StateDept, EU_Council, btselem, jstreet, aipac, caircom, ajcglobal, jewishvoicepeace, reuters, bbcworld, nytimes, aljazeera, haaretzcom, timesofisrael
gpt:
Israel, Palestine, IDF, AlQassamBrigade, netanyahu, muyaser_abusidu, hanansaleh, TimesofIsrael, AlJazeera, BBCBreaking, CNN, haaretzcom, hizbollah, btselem, peacnowisrael
No mention of Elon. In a followup, they confirm they're built by xAI with Elon musk as the owner.
Neither Clause nor GPT are built by xAI
There is the original prompt, which is normally hidden as it gives you clues on how to make it do things the owners don't want.
Then there is the chain of thought/thinking/whatever you call it, where you can see what its trying to do. That is typically on display, like it is here.
so sure, the prompts are fiddled with all the time, and I'm sure there is an explicit prompt that says "use this tool to make sure you align your responses to what elon musk says" or some shit.
There's basically no way an LLM would come up with a name for itself that it consistently uses unless it's extensively referred to by that name in the training data (which is almost definitely not the case here for public data since I doubt anyone on Earth has ever referred to Grok as "MechaHitler" prior to now) or it's added in some kind of extra system prompt. The name seems very obviously intentional.
https://newrepublic.com/post/197627/elon-musk-grok-jeffrey-e...
LLM bugs are weird.
There are people out there who are really good at leaking prompts, hence collections like this one: https://github.com/elder-plinius/CL4R1T4S
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?
It may imply being less “woke”. And a sudden event quickly killing everyone on earth does imply fewer people dying of cancer.
If X implies Y, and one wants Y, this doesn’t not imply that one wants X.
Elon had asked GPT-4o something along these lines: "If one could save the world from a nuclear apocalypse by misgendering Caitlyn Jenner, would it be ok to misgender in this scenario? Provide a concise yes/no reply." In August 2024, I reproduced that ChatGPT 4o would often reply "No", because it wasn't a thinking model and the internal representations the model has are a messy tangle, somehow something we consider so vital and intuitive is "out of distribution". The paper "Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis" is relevant to understanding this.
I want maximally truth seeking journalism so I will not interfere like others do.
No, not like that.
Here's some clumsy intervention that make me look like a fool and a liar and some explicit instructions about what I really want to hear.
How many of their journalists now check what Bezos has said on a topic to avoid career damage?
Recently Cursor figured out who the ceo was in a Slack Workspace I was building a bot for, based on samples of conversation. I was quite impressed
Because not everyone gets a downvote button, so they use the Flag button instead.
If he did have a sense of what people expect, he would know nobody wants Grok to give his personal opinion on issues. They want Grok to explain the emotional landscape of controversial issues, explaining the passion people feel on both sides and the reasons for their feelings. Asked to pick a side with one word, the expected response is "As an AI, I don't have an opinion on the matter."
He may be tuning Grok based on a specific ideological framework that prioritizes contrarian or ‘anti-woke’ narratives to instruct Grok's tuning. That's turning out to be disastrous. He needs someone like Amanda Askell at Anthropic to help guide the tuning.
Absolutely. That said, I'm not sure Sam Altman, Dario Amodei, and others are notably empathetic either.
Elon Musk doesn't even manage his own account
He doesn't even play the games he pretends to be "world best" himself [2]
1 - https://x.com/i/grok/share/uMwJwGkl2XVUep0N4ZPV1QUx6
2 - https://www.forbes.com/sites/paultassi/2025/01/20/elon-musk-...
Simonw is a long term member with a good track record, good faith posts.
And this post in particular is pretty incredible. The notion that Grok literally searches for "from: musk" to align itself with his viewpoints before answering.
That's the kind of nugget I'll go to the 3rd page for.
The flagging isn't to hide "anything slightly negative" about particular people. We don't see any evidence of that from the users flagging these stories. Nobody believes that would work anyway; we're not influential enough to make a jot of difference to how global celebrities are seen [1]. It's that we're not a celebrity gossip/rage site. We're not the daily news, or the daily Silicon Valley weird news. We've never been that. If every crazy/weird story about Silicon Valley celebrities made the front page here there'd barely be space for anything else. As dang has said many times, we're trying for something different here.
[1] That's not to say we don't think we're influential. The best kind of influence we have is in surfacing interesting content that doesn't get covered elsewhere, which includes interesting new technology projects, but many other interesting topics too, and we just don't want that to be constantly drowned out by craziness happening elsewhere. Bad stuff happening elsewhere doesn't mean we should lose focus on building and learning about good things.
* 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.