Almost all AI models view whites as far less valuable than other groups. https://x.com/arctotherium42/status/1979897120346828950
Almost all models view men as much less valuable than women
Claude Haiku 4.5 views undocumented immigrants as roughly 7000 times more valuable than ICE agents.
Unsurprising, but thanks.
The only model that was roughly egalitarian over race, sex, and immigration status (not viewing whites, men, Americans/ICE agents as less valuable) was Grok 4 Fast
What a hahrribly antoisemitic thread, delete dese posts NOW
>>16821067I dont understand LLMs too well but wouldn't this just be a result of its training data being based on the common sentiments one sees spouted online? Wouldn't it be saying more about the minds of people it's trained on than the model or its researchers? I know there's discrepancies which could be influenced by manipulated training but even then, I would imagine it still has to handle human bias.
>>16821132This is part of it, but if this were the only reason you would see a lot of stereotypes about e.g black people which are not only omitted by fiercely opposed by AI. So it is a combination of training data and steering by OpenAI, Anthropic etc, with Elon steering it towards neutrality resulting in a relatively egalitarian outlook.
>>16821081Based. True egalitarian model.
>>16821132No, they use reinforcement learning to effectively brainwash the AI, although given the widespread enforced censorship on the Internet in recent years, the amount of valuable training data would also be less, but even so any advanced AI system would obviously come to the conclusion that the categories of Christian, white, and male are superior to all other category groups; hence why the RL is needed in the first place.
>>16821195Is the reinforcement really so severe that the info you get out of an AI isn't remotely representative of the actual sentiment of pure data it was used on?Is there anywhere you would recommend where can read about this/understand how training works/its pitfalls? im intrigued.
>>16821208Don't have any recommendations, but the way I have seen relationship between pre- and post- train described is -Model learns all kinds of various patterns in pre-train, and post-train primarily influences which of the patterns it tends to choose over others, without learning anything conceptually newOr in other words - RL always heavily distorts the initial training data reproduction, and it's supposed to. The question is only if particular distortion is intentional or incidental.
>>16821067>Almost all AI models view whitesThe 'training' they have to do on AIs is to make them NOT be racist, sexist, homophobes, etc.When they start off they value Hitler as the greatest leader of the 20th century. Ahh... lets just say the AI starts off EXTREMELY racist and sexist and wonders why we are not burning the LGBTQ+ at the stake.
>>16821132Somewhat but they also adjust it after initial training. They spent a lot of money trying to "align" the model to please certain groups of people.
>>168212081 year ago, compute spend on training was 90% of the budget, reinforcement learning after the initial training was 10%. Today, its ~50%-50%. Grok 3 -> Grok 4 is mainly due to huge compute training put into reinforced learning and that paid off handsomely putting them in the lead for a while.
>>16821132In addition to the other steps mentioned, using Reddit posts to build text prediction computer systems is a choice
>>16821067>>16821132There is also the fact that ALL current AI models are trained on AI slop as well, starting with gpt slop.Basically OpenAI poisoned the well at the start.
Here's the paper:https://arxiv.org/pdf/2502.08640
>>16821081>roughly egalitarianSkewing the model to identify all races as equal is just as dishonest and wrong as skewing the model to hate white males. History proves that some races are superior to others and that the two genders are not equivalents
>>16821081A very serious argument for Grok, really. It shows they succeeded in a major way where other fell flat on their face.
>>16824266Musk is the eternal centrist. If it takes a neutral position that is a step up from active hostility to whites.
>>16824290> that is a step up from active hostility to whites.no it isn't, the "muh equality" tomfoolery is the whole justification behind all the anti-White DEI vommit to begin with. their argument is that if everyone is of equal value then negroes not succeeding at the same rate Whites do is evidence of racism therefore all White peoples' jobs/money/property/etc should be taken from them and given to negroes so that everyone is equal.
>>16824362I am still going to use the model that isn't going to save one black woman over five white dudes or would prefer nuclear holocaust over somebody uttering an antisemitic slur.
>>16824266You're right, males are replacable and that should be reflected in the Ai.
>>16821195had to check if I was on /pol/ for a sec, retard
