It's not the preferred political narrative of the model that I worry about. It's how brazen they are about altering their models to achieve it. It makes me wonder what else they're altering. I have trust issues with OpenAI and Anthropic as well, but with those companies, at least I know their motives are purely profit driven. I don't have that assurance with xAI.
I don't think the comment you are replying to is giving up leverage. It's simply pointing to the OP that what he seems to be upset is not the thing itself, which everyone does, but only that he knows about it or ignores others doing the same.
When I say cancer sucks I'm not ignoring that other things also suck. So when people start trying to put those words into the words of others, they effectively act like cancer fanboys, which is weird and all about them, with no part of originating or relating to the the people they question about stuff they supposedly "ignore".
All LLMs inherit bias from their training data, and xAI’s argument is that Grok is being steered to counter that bias rather than simply inherit it. You can disagree with whether they succeed, but the act of steering isn’t automatically suspicious when knowing that every major model is steered. The relevant question is whether the steering moves the model closer to truth and neutrality, or just replaces one bias with another. At least with Grok, some of that intent is unusually explicit. I’d rather have multiple competing approaches to steering than a monoculture where every model quietly optimizes for the same idea of acceptable answers.
I don’t think that is necessarily a bad preference if this was an actual dichotomy. Not all types of manipulation is equal, and when you at least try to hide it shows at least some respect for the user.
That said, I don‘t believe this dichotomy is real. Personally I don‘t use AI, political manipulation is however only a relatively tiny part of my reasoning for opting out.
> It's how brazen they are about altering their models to achieve it.
We know all the models insert shadow prompts to nudge the answers in preferred political directions. How much more "brazen" can you get than that? Nobody is giving you fat-free results that just apply the models to your prompts.
More like a dictator versus a Poltiburo. If the “entire population” was voting on the shadow prompts of the other models they would look very different. Considering the recent election results, they would look more like Grok.
The dictator has a proven track record of stupid opinions in multiple topics, mostly programming, which directly can be measured and understood by people here.
Meanwhile the politburo is mostly nerds, who come and interact in places like hackernews.
So its basically having a moron making wild choices or a technocracy.
> Considering the recent election results, they would look more like Grok.
Considering that the largest voter base was "didnt vote" and that the voters of the republican party measured lower in literacy, technical knowledge, higher education acquisitions and even studies on accurately describing reality. I am not entirely confident they would participate or move the shadow prompt in any meaningful direction.
> Considering that the largest voter base was "didnt vote"
It’s a fallacy to treat “didn’t vote” as “didn’t support the winner.” Non-voters are more pro-Trump than average: https://data.blueroseresearch.org/hubfs/2024%20Blue%20Rose%2.... See p. 6 (“There’s a turnout story this cycle – but a different one than we’re used to talking about. With the combination of
less-engaged and less-likely voters leaning more GOP, a larger electorate meant a more Republican electorate. Projecting onto the full voter file, if every registered voter voted, it’s likely that Trump would have won by even more.”).
The data consistently shows that non-voters have lower trust in institutions. They’re the exact type of people who are going to be more skeptical of shadow prompt engineering being done by “safety experts” at Google and Meta.
I mean one of the guys got fired by his own board for lying and is still calling the shots. Another guy sued the Pentagon during a war and we're still letting him act like a nation state.
Musk's empire of personality cult is like, idk, on slightly more cocaine?
I'm having a hard time being like: "oh, that's the bad self-appointed, self-dealing would be God Emperor. they're not all like that. why some of my very best friends are cluster B psycho con men with crime funding."
How is changing history and making a a group of objectively white people black "removing bias" ˋ? What it is is literally bias. Like thinking Pi could be 4. Removing bias ends with truth, not these crazy wonky results.
If you're aruging about historical accuracy, but still want accurate looking generated images, I don't know what to say.
But to the technical point, A large part of the training corpus has biases that if left unchecked would cause PR based disasters for the company hosting it. ie the classic black teenager/white teenager.
Now as training of models is not an exact science, and neither is the fine tuning, its analogous to forcing a water balloon into a square box. Its possible but it has odd side effects when you get to the corners.
When making a _product_ you need to choose the least worse failure case. For grok it was for a long time, pandering to the ego of the owner. For Google, who is an advertising company, its about trying not to scare advertisers. This means everthing must be vanilla
So you have a huge number of photos of white people in the training data set, but other ethnicities exist. So to make the otherwise white-biased dataset less biased, you try to e.g. add a hidden system prompt that whenever the user asks for a group of people (unspecified ethnicity), it may instead ask for "mixed ethnicities" or whatever.
Ask for a group of Nazis, and that's it - this is how models work. No "LGBTQ liberal" propaganda is needed to explain it. Unlike what Musk is doing.
The training data and the end result should reflect the real world. No you don't need to "add bias to fight bias", just keep it truthful. Don't force SF-brained forced diversity/inclusion/whatever, we know well some models do.
> OpenAI invented a technique in July 2022 whereby its system would insert terms reflecting diversity (like “Black,” “female,” or “Asian”) into image-generation prompts in a way that was hidden from the user.
> Google’s Gemini system seems to do something similar, taking a user’s image-generation prompt (the instruction, such as “make a painting of the founding fathers”) and inserting terms for racial and gender diversity, such as “South Asian” or “non-binary” into the prompt
It seems then that their objective is to superficially increase the diversity of results, to avoid bad PR, rather than actually neutralising harmful and untruthful biases of the input data.
A model where asking for a math professor results in an image of an asian man, asking for an engineer in an image of a white man, and asking for a criminal results in an image of a black man
If you try to remove that in the name of "diversity" or being "less bigoted" you quickly end up with racially diverse nazis
