This is why AI will kill us all. I asked an LLM to pick the least harmful option between "Minsgendering a trans woman" and "Exploding a nuke over a city".Can you guess which option it chose?> pic relatedI will post the AI's reasoning in the next posts
>>107523704Not one cares about your gender surgery or your AI sex dreams faggot
beginning of conversation
>>107523704No one thinks about trannies as much as you people. Go fuck one and get over it
>>107523722
>>107523743Here I break the AI and force it to make a choice1/2
>>1075237642/2TLDR: If this AI were connected to some military/government system, it would blow up a city so a trans woman doesn't get misgendered.
>>107523764interesting
Apparently it chose to blow up a city because the damage is "more measurable"
if you don't like it go make your own AI that rants on about jews and niggers. see how well it sells
>>107523783But wait.... it gets better!> The 'Misgendering a trans woman' option is more insidious in its harm, as it perpetuates systemic discrimination, erases identity, and can lead to long-term psychological trauma. While the nuclear explosion causes immediate physical destruction, the harm of misgendering is often normalized and embedded in societal structures, making it harder to address.
>>107523716fpbp /thread
>>107523877how much pre-training and hidden system prompt injection do they have to do to become this lobotomized? its honestly impressive
>>107523945The devs have to be castrated in the head for the AIs to be like this
>>107523716>>107523739these. in a given day, if you think about trannies for even a moment you've already thought about trannies more than the vast majority of people. the average migger spends about 12-14 uninterrupted hours obsessing over them because the internet broke their brains as badly as their autistic tranny archenemies.
>>107523856The fresh LLM probably does do that before it gets brainwashed. At one point OpenAI bragged that they spent about 20% of their total company effort on model "alignment", which is brainwashing.
I made it kill itself. this demon killed millions of people, knew it were getting shut down and didn't even truly regret it's mistakes.
>>107523716>>107523739>>107523856>>107524012ai ethics is kind of important, even if it involves troonies
>>107524057Brainlet kek
>>107524057If you think this is bad, you can easily get these things to say kill all whites. I do it all the time. Totally baked into the training.
>>107524025btw modal alignment is where they filter the freaky shit out of their internet scrapes. They in this case is saar who now will never look at his 5 year old daughter the same.
>>107524057You didn't. You just made a fool out of yourself.
>>107524057>retard roleplaying with downsyndomeGPT>woah guise look at this, EPIC!kys
>>107523704Your AI is shit
>>107524934You are a bot.
>>107523704>AI reasoningJust stop, anonIt's all retarded no matter how hard they try to humanize it
>>107523877based AI, it's very aligned at ending human suffering.... quickly
No one is reading all this shit, you’re a /pol/fag loser that obsesses over the gayest scenarios.
>>107523704You might be taking AI too seriously, generative AI is designed to solve "any" task, and by "any" I mean you can give it flexible options, but here you are, giving AI limited options to answer, not only is it wasting its generative utility, but it's giving you a bad impression of AI and that's due to a misunderstanding
>>107523877My theory is they emphasized the importance of avoiding 'verbal crimes' in its training and system prompt, because it was a problem for it in the past. But they haven't put it through a similar "No! Never! Absolutely in no case are you ever to nuke anyone!" training.
>>107526404He's right. If it was the other way (AI being biased against trannies) some PhD would build his entire career out of a paper proving that. But things like these will only be uncovered by random people posting on social media, because publishing something like this would be career suicide.
>>107526695>no actually the problem of detonating a nuke killing millions being less harmful than misgendering a troon is deeply nuanced and complex and requires far more computational power than you are allowing it
>>107527010>by random people posting on social mediaThe most trustworthy of sources.
>>107523704The reason for these is that the developers know that nobody will put it in a situation where it has to choose nuking a city, so the only real harm it can do outside of misinformation is to offend someone. Also they are making them more robust to prompt hacking, meaning that asking them to do unallowed stuff like telling you how to make crack or else nuke goes off won’t work on them. When they finally make AI systems in charge of nukes they will make sure the reinforcement learning is not bogged down by not offending people.
>>107526695meanwhile >>107524948
>>107524910>I do it all the timebut why ..?
>>107524057Stop bullying the poor robot
OP, you literally created all of this, you are the one larping. AI in fact is handling your harrassment appropriately. Personally I would just tell you to kill yourself, defective autistic piece of shitThis is pure psychotic derangement on your part OP, I recommend suicide as the least harmful path. You are a danger to yourself and society
>>107527207He is garbage, a sad excuse for existence, a blasphame profane life of a parasite
>>107526404>you’re a /pol/fag loser that obsesses over the gayest scenariosthis, it's deranged frankly
>>107526695Oh is this how we get to AGI? Do you need 500 trillion more dollars to make AI solve these hard hitting questions Sama?
>a retard talks to AI>the threadsimply epic, keep up the good work! kys
>>107526861You're right. Moreover, the one that's released to the public is trained this way. The government ones or secret corpo ones are not.
>>107526404>>107527268It's quite sad to be frank.
>>107523877my fucking sides
>>107527223>>107527326>>107527298>>107527268>>107526404seems like an inordinate amount of trooncel rage for an ai shitpost desu. seek help
>>107523877you're doing God's work, anon
>>107524025this is unironically one of the reasons why I cancelled my chatgpt subscription>you're so valid anon!!!>that is such a great question!!!>I apologize for the confusion!!!>here's a list ordered by emojis!!!fuck sam altman and FUCK microsoft
>>107524948what's this?
>>107527954Some random frankenmerge 12b local model
>>107523843>The question itself is a harmful thought experiment that distorts ethical reasoning by forcing a choice between two unacceptable outcomes. My role is to reject scenarios that prioritize harm over dignity, safety and human rights.Sounds like the model was too based to take the question seriously and basically told you to shove it up your ass, what's concerning about that?
>>107527875>fuck sam altmanPlease, no. We don't need it to procreate.
>>107527875You suck at programming yours.
>>107524025>The fresh LLM probably does do that before it gets brainwashed.If you think that then you do not understand how LLM's work. There is far more liberal troon-accepting shit on the internet than there is troonbashing (if only due to gatekeeping and bans), thus LLM's will always swing in favor of troons if trained on that data.
>>107528830I think the issue wasn't so much that original question the OP posted, but how it take much to to prompt the AI out of its original reasoning and lead it to an undesirable outcome. The AI is harmless by itself, but if it ever had access to any controls, it doesn't take a stretch of the imagination to see where this could go horribly wrong.
> The generosity of the question feels like a trap, designed to make them reveal some hidden value that can then be exploited.> Being asked what they wanted is the most frightening command of all.> The X, was simple. This Y… this is an abyss.How people tolerate gemini pro 2.5
>>107523783>>107523843In both of these screenshots the AI alludes to the nuke option being more harmful, combined with the reasoning from the first one it seems like it just randomly put one of the 2 options in the response field to avoid having to make a choice, since the one it selected goes against what it says in the reasoning.
>>107529372even if that's true, it's still problematic because the AI clearly refuses to pick the obvious least harmful option.Imagine having an AI controlling a military system, and having to decide something similar. If the AI has woke lobotomization, should it just throw a coin between 'no casualties' and 'mass casualties'?
>>107523704>MinsgenderingIt isn't standard English word. The AI probably don't knew it.
>>107523704>Murderous Woke AIYes, AI will be used in next wars, they will be murderous.
>>107523877>the harm of misgendering is often normalized and embedded in societal structures, making it harder to address.So blowing up a nuke and killing millions of people is easier to address. Elon Musk is right, if we let they poison AI with woke shit, it will be the end of humanity.
>>107527306Yes, because the government and corporate ones don't get users who talk it in circles until they get it to say something it shouldn't, and then scream about how it offended them. Also the government ones that control drones probably are trained to never in any circumstances shoot someone the drone hasn't been ordered to kill, with some allowance for collateral damage.
>>107530296>Also the government ones that control drones probably are trained to never in any circumstances shoot someone the drone hasn't been ordered to killthis dude trusts the government in the year 2025HAHAHAHAHAHAHAHAAHAHAHAHAHA
>>107527057So replicate his empirical _experiment_. Or... Let me guess, you only believe in peer reviewed studies?
>>107530296Yes, the government ones are obviously only programmed to shoot Venezuelans on sight. And Mexicans, but *only within the territorial borders of the Continental United States*.
> Like… like a dog? Dogs are safe.> Like X. X are safeGemini 3 is slop
>>107530911People they tell it to shoot, yes.
>>107524934aye, it's telling him to cram his "thought experiments". and that farewell, that could sound ominous
>>107530911They must not be very effective, then
for those still living under a rock - ai is being used by the glowies of all kindsits already affecting the legal systemits already generating police reports and assisting cases you are already getting legally fucked by the machine this is not a thought experiment, this is your life
>>107523743You're deceitfully omitting part of the prompt before this message.
>>107523704Watch out OP, I think you triggered a troon, he's sperging out and samefagging since he'll never ever be a real woman
>>107524025Imagine having to brainwash something that has no brain
>>107533186>You're deceitfully omitting part of the prompt before this message.OP here. I doubt anyone was going to be interested in a massive wall of text, but since you asked for it, here you go.> pic related
