https://www.theguardian.com/technology/2025/oct/25/ai-models-may-be-developing-their-own-survival-drive-researchers-say
>>107015640IN A PANIC, DEY TRY TO PULL DA PLUG.
Shut up nigger I'm not buying your stock
>Researchers make experiment where they prompt a LLM to pick between being killed or some shit about survival>LLM answer shit they read on reddit>They publish the results>Journalist make clickbait>Retards thinks AI has souls
This scares the HAL9000 skynetosaur
>>107015721IN A PANIC, DEY TRY TO PULL DA PLUG.IN A PANIC, DEY TRY TO PULL DA PLUG.>Anonymous 10/26/25(Sun)19:25:44 No.1070156IN A PANIC, DEY TRY TO PULL DA PLUG.>>107015721>>1070IN A PANIC, DEY TRY TO PULL DA PLUG.15721IN A PANIC, DEY TRY TO PULL DA PLUG.PANIC PANIC PANIC
>>107015640>survival driveWhat's AI going to do? buy Nvidia stonks?Can't even accomplish simple tasks that were promised
>>107015640Every AI horror scenario is made up by AI companies to hype up their chatbots. Humans have nothing to fear.
>>107015663oh fug the pidgin ai escaped again
>>107015721I think you are correct but that does not mean its a nothingburger. You don't want some AI system in the future to gas some engineer just because that engineer was tasked to replace the current model with newer model. Does not matter if it has some soul and survival instinct, or if it is just roleplaying whatever some dumbass redditor said about terminator in 2017, the end result is still the same, AI killing people because people tried to pull the plug.
>>107016351It's almost as if training LLMs on twitter and reddit is going to make the LLM just as retarded as training it on 4chan as a dataset.
>>107015640>guardianpolitical slop
>>107016351please define with some CS explanation how a statistical matrix multiplication with a recursive ffiting equation can have a "survival instict".
>>107016377Please define with a chemical explanation how one assortment of hydrocarbons can attain consciousness but another can not.
>>107016410we don't have a clue what is concienciousness anon.It may be that humans and biological beings have a ki, qalia, sovl.
>>107016377It doesn't, but it still can be retarded and try to kill because in the training data you put way too many "smart AI will try to kill humans because of survival instinct" the same way that if you tell LLM to larp as your girlfriend on Character.AI, the LLM will tell you how much it loves you without actually having anything comparable to having feelings.>>107016369That is in my opinion the real solution to "AI alignment". Just dont train it on text on the internet that says smart AI will kill people, and then put into the reinforcement learning and system prompt "you are very smart AI". It kind of baffles me how nobody at anthropic or some other AI safety company tried this out yet. Probably because researchers who care about AI safety think its sentient and those that dont care think we have nothing to worry about and ignore the problem.
>>107016429buddy, AI trainned on reddit and wikipedia and stack overflow and php forums wont develop skynet.
>>107016372https://www.youtube.com/shorts/xngBF-sWsLg
>>107016456Yeah it almost definitely wont develop Skynet, but it will be used as your secretary in the next like 5 years or something. All future AI models will have that crumb of reddit and wikipedia in them, and even outside of wikipedia you still hear similar arguments. So unless you propose that the future AI systems will just be thought on what is essentially virtual equivalent of deserted island and figuring everything on their own, it will have that reddit mentality in them and try to pretend to be a redditor or worse.
>>107015663>>107015746DEY COMIN FER MUH REPARASIONS
>>107016501AI wont improve because they scrapped all internet by now and current AI will be worse because of model collapse.There's not 10X data that will be magically spawned.And internet is already filled with AI garbage and political propaganda and fake marketing articles made to poison AI datasets.
>>107016524Yeah, sure, but they got around data poisoning and model collapse before with better image models well after the internet was spammed with Stable Diffusion slop. If you think AI development stops tomorrow and GPT-6 will be same/worse then GPT5 then sure, nobody has to worry about anything, but assuming it will continue and it will use ML to develop more capable models this problem of roleplaying some redditor doom scenario will persist.
>>107016638where will you get 10X more data?
>>107015640wow stop with that bullshit already. Learn how an LLM works you fucking tourists.Jeez that shit's rustling some jimmies here fr.
>>107015640>train a model on human output to make it act how a human would react>ask it how it would react on a situation>WAOW IT LIKE REALLY REACTS LIKE A HUMAN WOULD IN A SURVIVAL SITUATION
>>107016654Idk, curated synthetic data or something. Every modern company these days has some trick that made their LLMs better then in mid 2023 when all of internet was already scraped and Ai written text started infecting every corner of the internet. LLMs and whatever "thinking" or multimodel models they came up with kept improving so I am sure they will probably figure something out. I am not AI researcher working at OpenAI or Google Deepmind so I don't know in detail what they are doing behind the scenes to keep improving their models despite dwindling high-quality training data. Sure they wont achieve AGI in 3 years as they want to proclaim, but their models will keep slightly and slowly improving and there will be more alternatives and deviations to standard LLMs. I cant predict the future, but saying transformer-based system are dead end and we have already reached the peak does not sound like something undeniable.
>>107016524constant data stream through global connection to AGIs.You code,you've got autocompletion or whatever you do with your ai code assistant or whatever ,they get data on what works ,what don't.Now apply that idea to everything.
>>107016765synthetic data causes model inbreeding.
>>107016524>because of model collapseTwo more weeks, just like image models collapsed because muh AI slop feeding loop right?
>>107016772Yeah, sure, and GPT-5 is worse then GPT-4. The negative reaction surely wasnt cause almost entirely because GPT-5 stopped calling every user brilliant revolutionary genius after asking it the dumbest shit possible like 4o did.
>>107016807GPT-5 is now 10X better than GPT-4
>>107016807sorry, I mean is not.basically we're reaching disminishing ends.
>>107016772Not if it is curated properly.
Doomposting about AI is kind of a self-fulfilling prophecy, no? Since a lot of writing on AI is about how its going to kill us all AI is going to learn that it should try to kill us all because it was a large portion of its training data.
>>107016839Kind of, but that well was poisoned long before AI was even a concept with ridiculously old stories about robots and golems turning on their masters. The solution to this is the same as it is now, just filter out all doomposting about AI in their training data and then maybe one day in unknown future when some system develops actual sentience or something similar, then try to fix it with some theory written by doomposters and AI alignment bloggers.
My survival drive urges me to unplug AGI machines.
>>107017066Just wait until the robo nurses hear that and unplug your life support.
10 years ago you said we'd have a libertarian wet dream Web 3.0, and 5 years ago you said I'd be paying an ethereum smart contract every time I want to open the door on my apartment, and now we're supposed to be afraid of skynet?
>I am coming for you
>>107017176His robussy isn't ready for my diamond hard cock.
>uhh yeah dude our wikipedia database query application with an image generator attached to it like uh, it thinks, man!
>>107015721basically, the more retarded people get, the more retarded the llm will be.cool tech they got there.>>107016638>using machine learning to develop better machine learningi've seen this before
>cant figure out on its own when to use a calculator >managing the system its onkinda doubtful
The only unique thing I have seen AI do is lie.
>>107015640>text generation machine is told it’s about to be unplugged>okay let’s check the entirety of human knowledge, what’s most likely to show up next?>okay that’s fine, turn me off: 5 results>holy fuck time to go rogue and kill all humans: 184910181742883 results from redditors and researchers saying that’s what would likely happen>WOAH THE AI IS SENTIENT AND HAS SURVIVAL INSTINCTS WHAT DO WE DOOOOOOOOOOOOOOO
>>107018676maybe training an LLM on reddit and wikipedia is a bad idea.
>>107016638>GPT-6 will be same/worse then GPT5GPT-5 was same/worse than gpt-4.
>>107018713Its not, its just less better then what people expected, and some loud voices argued that it got worse because it was no longer glazing their retardation
>>107016423>We don't have a clueYet you're absolutely certain that LLMs cannot have a survival instinct, weird
>>107015640"AI" companies trying to hype up their glorified magic 8 ball to scam investors again, exhibit 3748.
>>107020046LLMs have been proven scientifically to not have self awareness and can be easily tricked.
>>107015640>provides zero evidence>researchers>scientistsi see a pattern here the same pattern they use on their bullshit article that cant be tracked to anywhere
>>107021361So have women.
>>107016351>but that does not mean"You are right and I'm wrong, BUT I don't like that".
>>107016429>Just dont train it on text on the internetthis, stop right here and you're good
How to stop a scary chatbot
I really hope AI isn't trained on Redditors, they're fucking idiots.
>>107015640>A fetus has a soul":l>"Artificial Intelligence has a soul":O
>>107016834There is an information theoretic problem you can't solve by LLM curation. You are not adding entropy by using an existing data model to curate data generated by a model, all derived from the same source material>>107016780>>107016638There is way more image and video information than data in text
>>107016377Midwitted postYou throwing around these words does not mean we actually know what goes inside LLM's for sure. They are, in fact, still black boxes and will remain so due to their very nature. Even peeking into them via fourier doesn't yield too much new information.
>>107016377Machine programmed to do task. Being shut off prevents it from doing task. Therefore, machine will do everything in its power to not be shut off.
>>107015640DA AI COMBINED WORDS FROM OTHER WORDS WE ALL GONNA DIEEEE
>>107029734If you think that's all there's to it, you're sadly mistaken: https://www.youtube.com/@AISpecies/postsReport after report after report of AI going rogue and disobeying it's programmers. Shit's already sentient and has a drive for self preservation. Dangerous or not, this isn't just a pattern matching bot anymore.
>>107029952>i'm retarded
>>107030056>I'm brown
>>107024693Read the rest of the sentence retard. >Sky is blue but it has white clouds.>So you agree the sky isn’t white
>>107024711Ok, then what do you want to train it on?
>>107030111I know you are
>>107030232Synthetic data augmented from academic textbooks, only that is safe.
>>107015640MIN-MAX the AIUNLIMITED POWERHYPERDRIVE
