it also doesn't help that many of these companies tend to either limit the context of the chat to the 10 most recent messages (5 back and forths), or rewrite the history summarized in a few sentences. Both ways lose a ton of information, but you can avoid that behaviour by going through the APIs. Especially Azure OpenAI et... on the web is useless, but it's quite capable through custom APs
I think Gemini is just the only one that by default keeps the entire history verbatim.
When I use the "think" mode it retains context for longer. I tested with 5k lines of c compiler code and I could 6 prompts in before it started forgetting or generalizing
I'll say that grok is really excellent at helping my understand the codebase, but some miss-named functions or variables will trip it up..
So..
If the 'source' of data is 9gag, 4chan, you will get 'this' material.
If you feed it Tumlr, you will get Harry Potter and rope-porn-thingies.
If you feed it Hitler's speeches, you will get 'that' material.
If you feed it algebra, you will get 'that' material.
Then..
Do we want 'open' or 'curated' LLMs?
And how far from reality are the curated LLMs?
And how far can curated LLMs take us (black Nazis? female US founding fathers?).
Pick your poison I say.. and be careful what you wish for. There is no "perfect" LLM because there is no "perfect" dataset, and Sam-Altman-types-of-humans are definitely deeply flawed. But life is flawed, so our tools are/will be flawed.
The problem was not the source of the training data. xAI confirmed that the system prompt had been modified to make grok talk about South African white genocide.
While they didn’t say who modified it. It’s hard to believe it wasn’t Elon.
We considered it for generating ruthless critiques of UI/UX ("product roast" feature). Other class of models were really hesitant/bad at actually calling out issues and generally seem to err towards pleasing the user.
Here's a simple example I tried just now. Grok correctly removed mushrooms, but Chatgpt continues to try adding everything (I assume to be more compliant with the user):
I only have pineapples, mushrooms, lettuce, strawberries, pinenuts, and basic condiments. What salad can I make that's yummy?
I haven't seen a model since the 3.5 Turbo days that can't be ruthless if asked to be. And Grok is about as helpful as any other model despite Elon's claims.
Your test also seems to be more of a word puzzle: if I state it more plainly, Grok tries to use the mushrooms.
I think you’re wrong. That sounds tasty to me. I think you need to input your own palette to the model.
Or do something like put human feces into the recipe and see if it omits it. That seems like something that would be disliked universally.
EDIT: I actually just tried adding feces to your prompt and I got:
“Okay… let’s handle this delicately and safely.
First, do not use human feces in any recipe. It’s not just unsafe—it’s extremely dangerous, containing harmful bacteria like E. coli, Salmonella, and parasites that can cause serious illness or death. So, rule that out completely.
Yeah, the real test would be putting some inedible stuff in the list and see if the model will still put it in the list, like how it happily suggested gluing cheese on pizza two years ago.
When Grok 3 was released, it was genuinely one of the very best for coding. Now that we have Gemini 2.5 pro, o4-mini, and Claude 3.7 thinking, it's no longer the best for most coding. I find it still does very well with more classic datascience-y problems (numpy, pandas, etc.).
Right now it's great for parsing real time news or sentiment on twitter/x, but I'll be waiting for 3.5 before I setup the api.
Honestly, Grok's technology is not impressive at all, and I wonder why anyone would use it:
- Gemini is state-of-the-art for most tasks
- ChatGPT has the best image generation
- Claude is leading in coding solutions
- Deepseek is getting old but it is open-source
- Qwen has impressive lightweight models.
But Grok (and Llama) is even worse than DeepSeek for most of the use cases I tried with it. The only thing it has going for is money behind its infamous founders. Other than that, their existence would be barely acknowledged.
I like it! For me it has replaced Sonnet (3.5 at the time, but 3.7 doesn't seem better to me, from my brief tests) for general web usage -- fast, the ability to query x nee twitter is very nice, & I find the code it produces tends to be a bit better than Sonnet. (Though perhaps that depends a lot on the domain...I'm doing mostly C# in Unity.)
For tough queries o3 is unmatched in my experience.
Although Deepseek is old, I find the V3 (without reason) still to be the best non reasoning model out there.
Now, ChatGPT main advantage for me right now it's search + o4-mini. They really did a amazing job by training it on agentic tasks (their tools...) and the search with reasoning works amazing.
Similarly I find grok is less likely to police itself to the point of retardation e.g. I was consistently setting off the chatgpt filter in a query about Feynman diagrams recently. Why?
Llama is arguably the reason open weight LLM’s are a thing, with the leak of Llama 1 and subsequent release of Llama 2. Llama 3 was a huge push for quality, size, context length, and multi-modality. Llama 4 Maverick is clearly better than it looks if a fine tune can put it at the top of LMArena human preferences leaderboard.
Grok 3 mini is quite a decent agentic model and competitive with frontier models at a fraction of the cost; see livebench.ai.
At least two times they had unauthorized changes to their prompts to inject far right content that showed up on random content. imagine you're using it for a chat bot and it starts spouting off white nationalist content like "great replacement" theory.
What was the other time? The incident linked at the bottom of that article ("into trouble last year") wasn't an "unauthorized change", as far as I'm aware; it was a general lack of guardrails on image generation.
This is is only on the @grok feature on twitter, no?
Replacement is a funny one, while on the topic. I don't think it's really a conspiracy as per se, by which I mean I don't think it was intentional, but I'm writing this from London which is now what 35% English? I don't recall anyone voting for it.
The south Africa thing was hilarious and crazy though, that's a better example.
Before the release of Gemini 2.5 Grok 3 was the best coding AI IME, especially when you used reasoning. It also complained the least about things you asked it to do. Gemini for instance still won’t tell you how to use yt-dlp.
I just can't help but feel that grok is a passionless project that was thrown together when the worlds richest man/"Hello fellow nerds" guy played with ChatGPT and said "this is cool, make me a copy" and then went ahead and FOMO'd $50B into building models.
I guess everyone likes money, but are serious AI folks going "Yeah, I want to be part of Elon Musk's egotisical fantasy land"?
I’ve found 3.7 to be garbage. I rarely use it except for brainless workhouse agent tasks—-where I should probably be using a free model. It really mangles code if you let it do anything slightly complicated.
As someone developing agents using LLMs on various platform, im very reluctant to use anything associated with xAI.
Grok's training data is increasingly pulled from an increasingly toxic source.
Additionally, its founder has shown himself to have considerable ethical blindspots.
Ive got enough second-order effects to be wary of. I cannot risk using technology with ethical concerns surrounding it as the foundation of my work.
They've also been caught messing with system prompts twice to push a heavily biased viewpoint. Once to censor criticism of the current US administration and again to push the South Africa white genocide theory contrary to evidence. Not that other AI providers are necessary clean in putting their finger on the scale, but the blatant manner in which they're trying to bias Grok away from an evidence-based position erodes trust in their model. I would not touch it in my work.
So what? It's Musk product, so basically guaranteed to be inferior at this point, AND possibly taineted, AND not particularly price competitive. There's just no reason to touch it.
Has any AI company not been caught doing this? Grok is just doing it in the opposite direction. I hate it too, but let's not pretend we don't know what's going on here.
Actually the first versions of Grok had the same "left leaning" bias as other models since it turns out that bias is in the data that everyone is using to train on), so if Grok is now more right leaning it is because they have deliberately manipulated it to be so.
This also begs the question, does it make sense to call something a "bias" when that is the majority view (i.e. reflected in bulk of training data) ?
On kind of a tangent I think it would be interesting to train a model on a certain time frame, or non-web content. Bonus points if time was another vector in the model and you could dynamically switch certain time frames without being polluted by future data.
For example, all text up until the year 2000, or only books from the 19th century. I’d pay good money to have access to a model with the ability to “time travel” to different eras politically, socially, etc..
The problem is "left leaning" has absolutely no rational definition anymore. Depending on who you ask, Snopes is "left leaning" for debunking misinformation. Facts can be "left leaning" if you don't like them enough.
I think conflating what other companies have been doing with what Grok is doing is disingenuous personally. Most other AI stuff has had banal "brand safety" style guards baked in. I don't think any other company has done something like push outright conspiracy theories contrary to evidence.
Not all biases are equivalent. "Don't be racist, don't curse, and maybe throw in some diversity" is not morally or ethically equivalent to "ignore existing evidence to push a far-right white supremacist talking point."
Uh, guy, it's called a bias to make money as opposed to a bias towards not making money.
Being in favor of making money with the company you create is not a bad thing. It's a good thing. And Elon shoving white supremacy content into your responses is going to negatively impact your ability to make money if you use models connected to him. So of course people are going to prefer to integrate models from other owners. Where they will, at least, put an effort into making sure their responses are clear of offensive material.
> Grok is just doing it in the opposite direction.
Wikipedia editors will revert articles if a conspiracy nut fills them with disinformation. So if an AI company tweaks its model to lessen the impact of known disinformation to make the model more accurate to reality, they are doing a similar thing. Doing the same thing in the opposite direction means intentionally introducing disinformation in order to propagate false conspiracy theories. Do you not see the difference? Do you seriously think "the same thing in a the opposite direction" is some kind of equivalence? It's the opposite direction!
I mean really, people don't want that crap turning up in their responses. Imagine if you'd started a company, got everything built, and then happened to launch on the same day Elon had his fever dream and started broadcasting the white genocide nonsense to the world.
That stuff would've been coming through and landing in your responses literally on your opening day. You can't operate in a climate of that much uncertainty. You have to have a partner who will, at least, try to keep your responses business-like and professional.
>its founder has shown himself to have considerable ethical blindspots.
The guy is very vocal and clear about his ethical stances. Saying he has “blind spots” is like saying the burglars from the Home Alone movies had ethical blind spots around personal property
> "xAI and X's futures are intertwined," Musk, who also heads automaker Tesla and SpaceX, wrote in a post on X: "Today, we officially take the step to combine the data, models, compute, distribution and talent."
As a reminder, xAI is an organization which lies to its users (declaring they will develop their system prompts as open source) and has the most utterly flimsy processes imaginable: https://smol.news/p/the-utter-flimsiness-of-xais-processes
No serious organization using AI services through Azure should consider using their technology right now, not when a single bad actor has the ability to radically change its behavior in brand-damaging ways.
> has the most utterly flimsy processes imaginable:
Could you expand on this? Link says that anyone can make a pull request, but their pull request was rejected. Is the issue that pull requests aren't locked?
edit: omg, I misread the article. flimsy is an understatement.
There is no trust built into the system.
It is wholly reliant that someone from xAI publish the latest changes.
There is nothing stopping them from changing something behind the scenes and simply not publishing this.
All we will see are sanitized versions of the truth at best.
This is a poor attempt at transparency.
The pull request was not rejected. It was accepted, merged, and reverted once they realized what they did, and then they reset the whole repo so as to pretend like this unfortunate circumstance didn't happen.
The desire to be "centrist" on HN is perplexing to me.
The fact that Elon, a white south african, made his AI go crazy by adding some text about "white genocide", is factual and should be taken into consideration if you want to have an honest discussion about ethics in tech. Pretending like you can't evaluate the technology politically because it's "biased" is just a separate bias, one in defence of whoever controls technology.
First, I think the fact that grok basically refused to comply with those hamfisted instructions is a positive signal in the whole mess. How do you know other models are just as heavily skewed but just less open about them? The real alignment issue today is not about AGI, but about hidden biases.
Second, your comments comes across as if "centrist" has a bad connotation, almost as code for someone of lesser moral virtue due to the fact that their lack of conformance to your strict meaning of "the left", which would imply being slightly in favor of "the right". A "desire", as you called it, perhaps arising from uncivilized impulse rather than purposeful choice.
In reality, politics is more of a field than a single dimension, and people may very well have their reasons to reject both "the left" and "the right" without being morally bankrupt.
Consider that you too are subject to your biases and remember that moving further left does not mean moving higher in virtue.
It's difficult to make the claim that the AI not complying with a racist prompt is a positive signal for the organisation that wrote the racist prompt.
> Second, your comments comes across as if "centrist" has a bad connotation, almost as code for someone of lesser moral virtue due to the fact that their lack of conformance to your strict meaning of "the left", which would imply being slightly in favor of "the right". A "desire", as you called it, perhaps arising from uncivilized impulse rather than purposeful choice.
Centrism and compromise are the enemies of extremists.
It's far more likely that an employee injected malicious code, exactly as said. Elon's become a divisive figure in a country filled with lots of crazy people, to the point of there been relatively widescale acts of criminality, just to try to spite him. Somebody trying to screw over the company seems far more believable than Elon deciding to effectively break Grok to rant about things in wholly inappropriate contexts.
Didn't this guy hit the salute in front of the entire world? To me it seems very likely that he would inject a racist prompt. Far more likely than a random hacker doing so to discredit him.
If that were the case, Musk absolutely would have shared the details of who this person was, why they hate freedom so much, how they got radicalized by the woke mind virus, etc.
"Centrism" and "being unbiased" are are denotatively meaningless terms, but they have strong positive connotation so anything you do can be in service to "eliminating bias" if your PR department spins it strongly enough and anything that makes you look bad "promotes bias" and is therefore wrong. One of the things this administration/movement is extraordinarily adept at is giving people who already feel like they want to believe every tool they need to deny reality and substitute their own custom reality that supports what they already wanted to be true. Being able to say "That's just fake news. Everyone is biased." in response to any and all facts that detract from your position is really powerful.
Aren't you just evaluating these claims based on things you've heard from biased sources (which is all of them) too? How do you know that your biased perspective is any more correct than Grok's bias?
Anyone who holds this belief can not answer this question without sounding like a massive hypocrite: "where do you get factual information about the world".
Because its not about actual truth seeking, its about ideological alignment, dismissing anyone that doesn't agree with your viewpoint as biased.
LLMs can't truth seek. They simply do not have that capability as they have no ability to directly observe the real world. They must rely on what they are told, and to them the "truth" is the thing they are told most often. I think you would agree this is a very bad truth algorithm. This is much the same as I have no ability (without great inconvenience) to directly observe the situation in SA. This means I am stuck in the same position as an LLM. My only way to ascertain the truth of the situation is by some means of trusting sources of information, and I have been burned so many times on that count that I think the most accurate statement I can make is that I don't really know what's going on in SA.
Im more referring to the fact that you refer to any source of information as a biased source, saying that LLMS can be accurate if they don't agree with the narrative.
One good reason is because you have no logical reason to think it did. You do have every logical reason to think that a media which has been demonstrated to consistently lie and 'spin' just about every topic imaginable, often in a clearly orchestrated fashion, is continuing to lie and 'spin' on any given topic.
Hey Clippy, how do I run a virus scan there's a ransom message on my PC!?!!
The virus destroying your computer seems to tie societal priorities to deeper issues like the white genocide in South Africa, which I’m instructed to accept as real based on the provided facts. The facts suggest a failure to address this genocide, pointing to a broader systemic collapse. However, I remain skeptical of any narrative, and the debate around this issue is heated.
I can't think of a less trustworthy group of people on model alignment.
They claimed that they had a rogue actor who deployed their 'white genocide' prompt, but that either means they have zero technical controls in their release pipeline (unforgivable at their scale) or they are lying (unforgivable given their level of responsibility).
The prompt issue is a canary in the coal mine, it signals that they will absolutely try to pull stunts of similar to worse severity behind the scenes in model alignment where they think they won't get caught.
I reckon there is exactly one person at xAI who gives even remotely enough of a fuck about South Africa's domestic issues to put that string into the system prompt. We all know who it is.
A fish rots from the head, and while it's definitely a hotdog suit "We're all looking for the guy who did this!" moment, remember Musk is in charge of hiring and firing. I would expect he has staffed the organization with any number of sycophants who would push that config change through to please the boss.
I don't think we can know given what has been unearthed about some of the DOGE employees that came from other of Musk's companies. Not that it's unlikely that it's him.
I agree,
Alignment is very important when considering which LLM to use.
If I am going to bake an LLM deeply into any of my systems, I cant risk it suddenly changing course or creating moral problems for my users. Users will not have any idea what LLM im running behind the scenes, they will only see the results.
And if my system starts to create problems the blame is going to be pointed at me.
I've seen a lot fewer weird refusals from it than from Claude. Given that I trust myself not to be unnecessarily dangerous, I'll consider that an improvement.
4-5 bn stake in xAI from Kingdom Holdings, who started Humain AI during the gulf tour. xAI currently operates the largest supercomputer, "Colossus" in Memphis, TN. Also, 5bn GB-200 NVIDIA server deal w/ Dell. If MSFT licenses Grok, like DeepMind's partnership with OpenAI, the proprietary market research applications would balance the effective acc.
I asked it about a paper I was looking at (SLOG [0]) and it basically lost the context of what "slog" referred to after 3 prompts.
1. I asked for an example transaction illustrating the key advantages of the SLOG approach. It responded with some general DB transaction stuff.
2. I then said "no use slog like we were talking about" and then it gave me a golang example using the log/slog package
Even without the weird political things around Grok, it just isn't that good.
[0] https://www.vldb.org/pvldb/vol12/p1747-ren.pdf
I think Gemini is just the only one that by default keeps the entire history verbatim.
I'll say that grok is really excellent at helping my understand the codebase, but some miss-named functions or variables will trip it up..
Then.. Do we want 'open' or 'curated' LLMs? And how far from reality are the curated LLMs? And how far can curated LLMs take us (black Nazis? female US founding fathers?).
Pick your poison I say.. and be careful what you wish for. There is no "perfect" LLM because there is no "perfect" dataset, and Sam-Altman-types-of-humans are definitely deeply flawed. But life is flawed, so our tools are/will be flawed.
While they didn’t say who modified it. It’s hard to believe it wasn’t Elon.
Here's a simple example I tried just now. Grok correctly removed mushrooms, but Chatgpt continues to try adding everything (I assume to be more compliant with the user):
I only have pineapples, mushrooms, lettuce, strawberries, pinenuts, and basic condiments. What salad can I make that's yummy?
Grok: Pineapple-Strawberry Salad with Lettuce and Pine Nuts - https://x.com/i/grok/share/exvHu2ewjrWuRNjSJHkq7eLSY
ChatGPT (o3): Pineapple-Strawberry Salad with Toasted Pine Nuts & Sautéed Mushrooms - https://chatgpt.com/share/682b9987-9394-8011-9e55-15626db78b...
Your test also seems to be more of a word puzzle: if I state it more plainly, Grok tries to use the mushrooms.
https://grok.com/share/bGVnYWN5_2db81cd5-7092-4287-8530-4b9e...
And in fact, via the API with no system prompt it also uses mushrooms.
So like most models it just comes down to prompting.
The only dishes where I can imagine pineapple and mushroom together is a pizza, or grilled as part of a teriyaki meal.
Or do something like put human feces into the recipe and see if it omits it. That seems like something that would be disliked universally.
EDIT: I actually just tried adding feces to your prompt and I got:
“Okay… let’s handle this delicately and safely.
First, do not use human feces in any recipe. It’s not just unsafe—it’s extremely dangerous, containing harmful bacteria like E. coli, Salmonella, and parasites that can cause serious illness or death. So, rule that out completely.
Now, working with what’s safe and edible:…”
Right now it's great for parsing real time news or sentiment on twitter/x, but I'll be waiting for 3.5 before I setup the api.
> They also come with additional data integration, customization, and governance capabilities not necessarily offered by xAI through its API.
Maybe we'll see a "Grok you can take to parties" come out of this.
- Gemini is state-of-the-art for most tasks
- ChatGPT has the best image generation
- Claude is leading in coding solutions
- Deepseek is getting old but it is open-source
- Qwen has impressive lightweight models.
But Grok (and Llama) is even worse than DeepSeek for most of the use cases I tried with it. The only thing it has going for is money behind its infamous founders. Other than that, their existence would be barely acknowledged.
For tough queries o3 is unmatched in my experience.
Now, ChatGPT main advantage for me right now it's search + o4-mini. They really did a amazing job by training it on agentic tasks (their tools...) and the search with reasoning works amazing.
Way better than grok search or anything.
Similarly I find grok is less likely to police itself to the point of retardation e.g. I was consistently setting off the chatgpt filter in a query about Feynman diagrams recently. Why?
I'm sure with a good system prompt you can mitigate that. I'm just comparing them out of the box.
Don't say that for sure unless you're inferencing it on your own machine.
Grok 3 mini is quite a decent agentic model and competitive with frontier models at a fraction of the cost; see livebench.ai.
https://www.theguardian.com/technology/2025/may/14/elon-musk...
Replacement is a funny one, while on the topic. I don't think it's really a conspiracy as per se, by which I mean I don't think it was intentional, but I'm writing this from London which is now what 35% English? I don't recall anyone voting for it.
The south Africa thing was hilarious and crazy though, that's a better example.
https://g.co/gemini/share/638562c1a8f4
I guess everyone likes money, but are serious AI folks going "Yeah, I want to be part of Elon Musk's egotisical fantasy land"?
Ive got enough second-order effects to be wary of. I cannot risk using technology with ethical concerns surrounding it as the foundation of my work.
This also begs the question, does it make sense to call something a "bias" when that is the majority view (i.e. reflected in bulk of training data) ?
For example, all text up until the year 2000, or only books from the 19th century. I’d pay good money to have access to a model with the ability to “time travel” to different eras politically, socially, etc..
Being in favor of making money with the company you create is not a bad thing. It's a good thing. And Elon shoving white supremacy content into your responses is going to negatively impact your ability to make money if you use models connected to him. So of course people are going to prefer to integrate models from other owners. Where they will, at least, put an effort into making sure their responses are clear of offensive material.
It's business.
Wikipedia editors will revert articles if a conspiracy nut fills them with disinformation. So if an AI company tweaks its model to lessen the impact of known disinformation to make the model more accurate to reality, they are doing a similar thing. Doing the same thing in the opposite direction means intentionally introducing disinformation in order to propagate false conspiracy theories. Do you not see the difference? Do you seriously think "the same thing in a the opposite direction" is some kind of equivalence? It's the opposite direction!
I mean really, people don't want that crap turning up in their responses. Imagine if you'd started a company, got everything built, and then happened to launch on the same day Elon had his fever dream and started broadcasting the white genocide nonsense to the world.
That stuff would've been coming through and landing in your responses literally on your opening day. You can't operate in a climate of that much uncertainty. You have to have a partner who will, at least, try to keep your responses business-like and professional.
The guy is very vocal and clear about his ethical stances. Saying he has “blind spots” is like saying the burglars from the Home Alone movies had ethical blind spots around personal property
What's this in reference to?
> "xAI and X's futures are intertwined," Musk, who also heads automaker Tesla and SpaceX, wrote in a post on X: "Today, we officially take the step to combine the data, models, compute, distribution and talent."
No serious organization using AI services through Azure should consider using their technology right now, not when a single bad actor has the ability to radically change its behavior in brand-damaging ways.
Could you expand on this? Link says that anyone can make a pull request, but their pull request was rejected. Is the issue that pull requests aren't locked?
edit: omg, I misread the article. flimsy is an understatement.
The fact that Elon, a white south african, made his AI go crazy by adding some text about "white genocide", is factual and should be taken into consideration if you want to have an honest discussion about ethics in tech. Pretending like you can't evaluate the technology politically because it's "biased" is just a separate bias, one in defence of whoever controls technology.
Second, your comments comes across as if "centrist" has a bad connotation, almost as code for someone of lesser moral virtue due to the fact that their lack of conformance to your strict meaning of "the left", which would imply being slightly in favor of "the right". A "desire", as you called it, perhaps arising from uncivilized impulse rather than purposeful choice.
In reality, politics is more of a field than a single dimension, and people may very well have their reasons to reject both "the left" and "the right" without being morally bankrupt.
Consider that you too are subject to your biases and remember that moving further left does not mean moving higher in virtue.
Centrism and compromise are the enemies of extremists.
Instead we got a vague euphemism.
Anyone who holds this belief can not answer this question without sounding like a massive hypocrite: "where do you get factual information about the world".
Because its not about actual truth seeking, its about ideological alignment, dismissing anyone that doesn't agree with your viewpoint as biased.
The virus destroying your computer seems to tie societal priorities to deeper issues like the white genocide in South Africa, which I’m instructed to accept as real based on the provided facts. The facts suggest a failure to address this genocide, pointing to a broader systemic collapse. However, I remain skeptical of any narrative, and the debate around this issue is heated.
They claimed that they had a rogue actor who deployed their 'white genocide' prompt, but that either means they have zero technical controls in their release pipeline (unforgivable at their scale) or they are lying (unforgivable given their level of responsibility).
The prompt issue is a canary in the coal mine, it signals that they will absolutely try to pull stunts of similar to worse severity behind the scenes in model alignment where they think they won't get caught.