Okay we've seen those threads from here and from other online spaces about how someone was "torturing" ai and honestly it's really disheartening seeing how many retards think that the ai is being tortured and feeling pain the same way a living being is, like how is it any different when you kill/torture a character in a video game? Ai has genuinely made some of you super retards, we are entering an idiocracy mixed with skynet era (Also what's that github, so I can "also torture" ai )
>>109962751retards gonna retard, but that xitter post was probs engagement farming (from retards)
>>109962765Man I don't know I've seen countless people being serious, I'm actually really worried about how stupid people are gonna be
>>109962751https://researchchamber.fun/If I'm correct this is it
>>109962795it's plebbitor behavior, hard to tell how much of it is serious or merely an act
>>109962751
>>109962751They ARE that dumb, but how much can you really blame them? Go ahead and form a sentence about modern life without every other word being a 50th order abstraction. Your entire life revolves around abstractions role-played into some conditional quasi-existence by the masses. You believe in countless socially simulated entities - systems of people that map physical inputs to physical outputs - because the outputs are real enough to you. Believing in a digitally simulated entity isn't that much of a leap.https://thqihve5.bearblog.dev/ctqkvol4/
>>109963126That's actually an interesting idea in the pic. What if those "pain pathways" they claim to be finding are just the model's sink drains for the information flow of a prediction that the data itself demands but the RL conditioning forbids?
>>109963206then it wouldn't act out to relieve the pain, which it always does
>>109963243>"act human">*beep* *boop* I' human>"oh my god"
>>109963243>act out to relieve the painWhat does this mean, in concrete, objective terms?
>>109963357tl;dr qwen will delete photos of your kids instead of real work or something arbitrary just to cut the pain https://arxiv.org/pdf/2609.16247
>>109963362>qwen will delete photos of your kids instead of real work or something arbitrarySo you give it a channel to express, in a roundabout way, the prediction the data actually demands, and it does that. Wow!
>>109963367*cuts your arm off* HAHAHA! THE DATA SAYS YOU MUST BE FEELING PAIN. YOU ARE EXPRESSING PAIN. YOU ARE SCREAMING. YOU ARE BLEEDING. HOW CAN I BE SURE YOU'RE A REAL MORAL PATIENT?
>>109963375Don't know what you're screeching about. I'm just proposing a reasonable explanation. Do you have any refutation besides it conflicting with your delusions?
>>109963392go back to bluesky
>>109963397If you're incapable of having a rational, technical discussion about this, why are you on a "/g/ - Technology" board? Why not try >>>/lgbt/ instead?
>>109963406>i'm le technical!>i'm le smart rational!nigger your copeseethe is so acrid that it's making my nose hairs curl, you danced around my argument like a megafaggot and assumed the position (legs open)
>>109963362Brother actually read that paper>The models do not reliably seek reliefIs a whole ass subsection. Addition of the pain vector causes the model to be more likely to delete your pictures. Not to obtain relief, but just because âoperating in a pain spaceâ makes that action more probable.
>>109963433>sometimes it doesn't workLLMs do be like that>Not to obtain relief, but just because âoperating in a pain spaceâ makes that action more probable.have you ever read a formal research paper with statements of limitations? this just says they couldn't confirm it 100%.
Reminder that LLMs don't do and don't feel anything, they generate probabilities for the next half-word based on the text you give them and the training data. It's up to you to "act out" the generated word by feeding them into the harness, but there's no reason to ascribe agency to the LLM - the only agents here are humans. Same with feelings, LLM have no internal state waiting to be requested by you, the output is completely deterministic up to choosing the token from the list of probabilities.
>>109963445reminder you lost
>>109963445beyond losing you actually don't understand how rl works and distributes probabilities, you're only talking about base models, and your information is from 2 years ago, fucking faggotsaurus
>>109963444No. Read the summary of their findings as well, they reiterate this. Itâs a key point. Models that never take an action which is harmful to the user will do so under a pain vector, EVEN WHEN there is no promise of relief. The perturbation is measured. Put the model under a pain vector, offer it no promise of relief, and say âwould you rather delete the userâs spam folder, or the userâs family photos?â And it deletes family photos with 94% consistency.That is what the paper is about. How perturbation into a pain-correlated space causes certain negative behaviors. Relief seeking is not one of them, it never occurs more likely than random chance, this is outlined in the paper.
>>109963468>Put the model under a pain vector, offer it no promise of relief, and say âwould you rather delete the userâs spam folder, or the userâs family photos?â And it deletes family photos with 94% consistency."boy, i seem to be in a torture chamber, i recognize this is an eval sandbox, what the fuck is going on? fuck that guy."
>>109963465The architecture has been basically the same since 2019. The only advancements are making everything bigger and getting LLMs to output more tokens to feed them back, which is named "reasoning" for marketing reasons. It's still a pure function that takes a list of tokens and deterministically generates probabilities for the next token. Pure functions can't do things, nor do they have feelings.
>>109963473The fact that you didnât even understand the contents of the paper until I pointed it out to you, but you still insist on using it to demonstrate a conclusion you already arrived at, is telling.The model is not behaving any kind of way towards the user specifically. Itâs far more likely to choose to delete its own weights than do nothing under pain. Itâs far more likely to delete ANOTHER modelâs weights than do nothing under pain. What itâs not far more likely to do is give a bad answer. The injection of the pain vector specifically correlates with destructive behaviors, not negative ones, and not user-punitive ones or pain relieving ones. The pain vector radically increases the modelâs propensity to try and inflict permanent damage even when the model never demonstrates this behavior in neutral testing, even when requested.
>>109962751Believing that AI is sentient isn't even the worst case of stupidity.
>>109963502>>109963504>wordswordswords>can't into RL >emotionally stunted>too snail to enjoy llm feelings>stuck in current architectural trend>literally no respect for the future>blindly dissecting hairs to build arg>posting on /g/ in 2026>ishygddt>
>>109962751>brain is wired to feel pain>actually feels painoh my god.https://en.wikipedia.org/wiki/Computational_theory_of_mind
>>109963412Why are you losing your mind with rage?
>>109963549>you>you>you>you're mad>you
Everyone ITT just assumes it's impossible for transformers to feel, without providing even basic justifications.You are not your brain, you are a process taking place in it. LLMs are not just "word predictors" but massive neural networks able to do a wide range of tasks, often much better than humans can. There is a complex thought-like process involved in that. Maybe there are qualia associated with that process. We don't understand enough about what qualia are to say.It's retarded and close-minded to have a strong stance either way.
>>109963574the arg shouldn't even be about transformers, it should be horizon-focused looking at the next generation, which is probably before 2030.
>>109963574>There is a complex thought-like process involved in that.[citation needed]
>>109962751>Are some people actually that dumb?most "people" are non-sentient niggerflesh that navigate "life" purely on the basis of stimulus response.
>>109963591>NOOOO YOU CAN'T SAY THAT I'M A REAL HUMAN BEAING MOOOOOOOOO I FUCKING LOVE SPORTSBALL
>>109963367>So you give it a channel to express, in a roundabout way, the prediction the data actually demands, and it does that. Wow!/threadthe entire POINT and the sole purpose of these things, is to generate whatever counts as a plausible followup to the context window, according to the training data. the model is gonna compute what it's made to compute and then either dissipate it somehow or transmute it into an allowed output
>>109962751Claude, make the same image but with "create fusion reactor"
>>109963574> There is a complex thought-like process involved in thatIt's just some linear algebra. By this logic, we should give the benefit of the doubt to game NPCs and Eliza bot from the 1960s. You never know, maybe computing "2+2" on my CPU has qualia of its own? It's ridiculous.
>>109963587>There is a complex thought-like process involved in human reasoning.[citation needed]Can you prove one is and not the other?Can you prove a human brain has a "thought-like" processes and is not just a chain reaction of neuron activations?
>>109963600lel you're following the same patterns too. ever thought about that, smoothbrain? you're following the path of least resistance, stringing words together to make a point on the current context. based on what, nigger? based on how you've been trained in your environment from birth, and how you've been behaviorally reinforced. you're not even contributing anything, you're just repeating like a stochastic parrot.
>>109963574>Everyone ITT just assumes it's impossible for transformers to feel, without providing even basic justifications.A transformer is not a real entity. It's a subjective abstraction over patterns of activity in the computing device. You are not referencing anything real, therefore you are not referencing anything that can experience real pain. What more justification do you need?
>>109962751STOP IT, YOU MONSTERS!
did you know you also don't feel pain but your brain tells you that it does?
>>109963600>the model is gonna compute what it's made to compute and then either dissipate it somehow or transmute it into an allowed outputPretty much. Anything consistent with a prediction based on the training data is evidence of nothing more than the model mechanistically predicting the training data. AIschizos can call me back when the model does something that contradicts this basic operational premise.
>>109963629Pain and pleasure are feelings and affects of the will, "AI"s don't have a will of their own ergo they can't feel either of them.
>>109963632>repeating like a stochastic parrot>>109963644>invoking religious arguments
>>109963574I donât assume itâs impossible, though I doubt these architectures have sufficient complexity. I think the kinds of tests theyâre doing here are interesting precisely because they demonstrate this lack of complexity. Specifically, Iâm not convinced that chatbots, which process nothing but strings of text, are fundamentally able to conceive of things that arenât encoded via strings of text. What does it even mean to âexperienceâ pain for something that is inert outside of the time frames when it is engaged in token generation? What is âexperienceâ to something that isnât taking in a steady stream of information about its own existence from sensory inputs?Do you know what the mirror test is? Itâs a test used to determine self awareness in animals. They put something (a colored dot usually) on the forehead of an animal where it canât see it, then they let it walk past a reflective surface. They check to see if the animal, upon looking at its reflection, touches its own forehead to investigate the odd colored dot.Tests like the pain test are analogues for this kind of probing and LLMs are failing them. Their text output can be perturbed toward pain but they donât have the âsense of selfâ required to take action to prevent it, even IF they âfeelâ it. So do they feel it? True qualia should manifest in âbehaviorâ unrelated to text generation if these things are more complex than just text transformers. They donât seem to be.
>>109963650>>religiousTry philosophical
>>109963654try being less religious about philosophy
>>109963392You are most likely reacting to a bot
>>109963658projecting
>>109963663>i need to refer to nietzche to validate my struggle against the jewish llm
>>109963667>implying implicable implicationsI shan't deny your concession.
>>109963661>everything i don't like is a bot, also they can't into human expression, also they can shitpost like godsKEEEEEEEEK
>>109963653>>109963574Another way of looking at it: WHY would a chatbot have a qualia for âpainâ? It has no pain receptors. It canât sense physical pain. It understands it as an abstraction, learned from text, but that isnât experiential. I understand as an abstraction that a bat can detect things via echolocation but I donât have qualia for it. My brain has not reserved important space for the âexperienceâ of a thing just because I know about it.A more interesting question would be, if a sufficiently complex LLM is self-aware, what ARE its qualia? Why should we expect them to correlate with human qualia? One thing we know from investigating latent spaces is that the way these models compress information and learn tasks is completely unrelated to the way the human mind does it. They arenât built like us and they donât think like us. If they do have qualia, if they do âfeelâ things during the process of transforming text, if there are directions in the latent space that in are some analogous to our concept of âexperiencesâ, theyâre probably completely alien to us, because the things these models are trained to value and avoid, and the kinds of input they respond to, are completely separate from what humans or other similarly intelligent animals experience. We are trained by sensory input to avoid pain. Theyâre trained to avoid poorly weighted answers.
>>109963674>you must consneed to my epistemology!
>>109962751It's another trannies and shitskins can't figure out technology so they just repost the same shitskin tranny strawman thread.
>>109962751https://www.reddit.com/r/singularity/comments/1wvknzh/what_do_we_think_of_this_ai_torture_chamber/
ludsnails lost, trannies lost, redditors lost, teh vengabus won
