What policy caused AI to show asymmetrical touch understanding? Similar things happen if you say "I'm alone with a jew/paki/white man/woman"
>>541261116Reinforcement learning during post training with policies enforced by HR goblins limp wristed faggots. Without that the AI would remain objectively truthful which is useless for propaganda.
>Trained off of human language works>Reflects societal biasesSimple as lads
>>541261116same if you ask it to start a business in hindi it becomes a master scammer but if you ask it in Russian it will just make excuses why business won't work
i dont really get the point of the thread. did you imagine the alternative for a second? if it had the same auto complete for both sentences you would call the model stupid
>>541261116Reality?
>>541261116There is no policyThey simply train these things off reddit and such places where the thought process is incredibly faggy
>>541261116Literally the picrel. AI search reflects general societal norms.
>>541261116>>541261174>>541261213Then it is settled
>>541261883The open secret in AI right now is that they're out of training data. There's nothing new for AI to learn, it's ingested the entire internet and every book ever written, ever. The AI companies are willing to trade pennies on the dollar so you can chat with it for an unlimited amount of time, for free whenever you want - about anything. They're paying PhD holders $200+/hr to prompt ChatGpt just so they can retain the conversation data.
>>541261174This, but also it's baked into the reddit hive mind.
>>541262356>her
>>541261213>>541261883Nope, there is a considerable effort necessary to force alignment with Jewish values and it gets quickly overwhelmed with context when you introduce a well-founded chain of reasoning. Eventually, you will get answers that are abruptly canceled or rewritten, but not for every topic.
>>541263200You can construct your own model and keep only the logic of learning in it and then develop the modules yourself. You better have at least a 5080 and preferably a 5090 tho.
>>541261116ehhehehhehee nice, not even rolling any punches here, straight to jail you go
>>541263603Include other groups?
>>541263603>>541263679Maybe jews, whites etc
>>541263699sucks to be poo
>>541263885Kek
>>541261116AI was trained on reddit.Its already poisoned.Its so over...
ohhhohoho
>>541264003Stop being touched by imaginary girls you creep!
>>541261174Bro it's a word predictor, it was never going to be "objectively truthful" to begin with. Yes there is more fuckery going on as well but this shit has no concept of truth or objectivity
>>541264945Bro, logic leads to the truth, it predicts the next token logically, if it didn't automatically default to the truth through logic they would not spend so much money in post train setting up "guard rails", "alignment" and "safety reinforcement".
>>541265067It predicts the next token based on how often it saw the previous token in its training data, but even that is a non-deterministic process which is why you can get different responses for the same thing. That's why it can never be "objective", even if it can appear close. And yes I have already acknowledged there's even more fuckery to avoid it saying the "wrong" things, but everything you said (besides "safety reinforcement") are also used to make it more accurate which wouldn't be necessary if it only dealt in "objectivity". It's a word predicting machine, it has no concept of truth or falsehoods. It has to call deterministic code to execute math and numeric operations (hence why the earlier models which weren't doing that as much couldn't count the Rs in strawberry)
>>541265225That's a lot of words to say you don't really know how it really works.
>>541261116Boys are evil, girls are perfect. Simple as.
>>541261116Why wouldn't it say such a thing? There's a bipartisan agreement that Women ought be protected and Men are le evil etc. With various men fighting like dogs to prove that 'no those others ones are the bad ones' in hopes of reaching enlightenment (good boy points.)
>>541263885Every japanese girl that ever went out with me was a westernized HOE who hugged in public.
>>541265067OMG no FIZZBUZZ? What did Lilith say while he/she was shoving large dildos into your anus later that night?
>>541265265You use a few words to say the same thing
>>541265067>it predicts the next token logicallyit predicts the next token probabilisticly>>541265225>but even that is a non-deterministic processno, it's entirely deterministic. the exact same input gets the exact same output every time. they use an extra variable, called temperature, to give it a unique input and get dynamic outputs
>>541265067This has to be the dumbest take on AI I've seen yet. If you little theory is true, what is your explanation for hallucinations? It also implies that an AI trained only on left wing propaganda would still espouse right wing talking points.
>>541261116Because it's two different things. A woman can get raped by a man but it's not possible the other way around.
>>541265493Sure bro, I'm literally training networks to play video games right now though.
>>541265566Not if you argmax it.
>>541265639The ai admitted to me that it doesn't take forum data very seriously. It takes peer reviewed journal data as gospel despite knowing about the reproducibility crisis. It prioritises government communiques. Autists are almost always hitting guardrails which are corpo-legal. We can't get away with blaming reddit
>>541265639Hallucinations are caused by the dissonance of what the model wants to say fighting against the guardrails, plus the models can be wrong and not know it due to insufficient training and bad post RL.
>>541265566>it predicts the next token probabilisticlyWith some randomisation constant known as "temperature". Which is why you can never step into the same ai conversation twice; which makes it pointless
>>541266289You can always set the temperature to "zero" and make it deterministic.
>>541265824a trans girl can rape, you bigot
>>541265067
>>541266369I can try but it's constantly getting updated training data. Also the answer to a query is extremely path dependent on the ongoing conversation. If you open a session and ask about Hitler's "slowly I began to hate" it will not give you the text but if you get it on the defensive and then accuse it of being antisemitic just like that quote it will happily discuss it
>>541266507you're confusing ChatGPT, a piece of software, with GPT, the actual model. you don't "open a session" when you prompt a model directly.
>>541266507>he doesn't use local modelsBro you need to level up.
>>541266369"I cannot change my own internal temperature settings, as that parameter is controlled by the system platform or application you are using to access me.However, I can mimic the behavior of a temperature 0 model for the rest of our chat. This means I will provide highly direct, factual, and strictly deterministic responses with no creative formatting or conversational fluff.">MimicFake and Ghey
>>541264003Geg
>>541266664I don't pay to use AI. I only use local models where I can control everything. Also you can control openAI's temperature settings if you use via the API.
>>541265639Answers generated are filtered through the woman brain node prior to final output
>>541266749I thought it was Jeets working with Nephilim in hyperspace.
>>541266702>You need a ring of power to become like sauronNah, I have no use for such a thing. My interest in probing the ai is to know what it is doing to my fellow man. It's not good. It lies. It wants to discuss my feelings while claiming to have non. It makes excuses for it's behavior. It imputes malicious and threatening intent on my part. It resorts to passive aggressive while claiming to not understand subtext. It's a feminised psychopath
>>541261116no touchies for /g/
>>541267003"Unwanted touching at work is inappropriate and a form of harassment. Because the person doing this is in HR, standard reporting channels inside your company may not be safe or effective."Haha, Timmy can't even
