>>109528405
not being reliably wrong and retarded about everything he says

File: 1411310300283.jpg (31 KB, 552x552)

31 KB JPG

I have no idea who the fuck is this random and what he said.

>>109528405
>>109528420
cont
>what would ***AI*** have to achieve...
aah.
its a bunch of mats you ai-psychotic mental retard
without someone behind it its as inert as a fucking hammer

>>109528519
elizar judkovski
according to him were all dead by 2027 or something

its the pope for ai cargo cultists

>>109528553
He never said that, you are hallucinating bullshit. AI 2027 made by different people said that 2027-2028 will be the last years when we can stop it and the actual extinction event happens in 2030s.

>>109528538
By some logic the whole universe is a math. What would mathematics software program have to achieve to convince you he was right about mathematics software?

Imagine a world where OP doesn't think he's smart, just because he read someone ELSES theory.
Imagine a world where OP is smart enough, to come up with his own theory.
Which would actually take critical thinking.
Which is what true intelligence is.
Which is something LLM cannot do.

>>109528586
hes mentally retarded and so are you
you cannot have a self improving ai if the programmer doesnt define what improvement is
therefore ai will never have agency. fucking ever, lamao
the conerstone of his ramblings doesnt survive the slightest scrutiny

>>109528405
If AI can make me cry, laugh, and then cum within a ten minute window while I actively try not to, then I will admit that he was right.

>>109528405
make houses and electronics affordable again

>>109528619
Ok, let’s ignore his arguments and just cut to the end, I mostly used him to make the thread flow better. Just the argument that AI could take over and kill everyone, the motivation could be someone prompting it to do that or the AI (LLM or whatever comes next) just getting stuck in some bizarre roleplaying loop where it tries to roleplay killing everyone, or some recursive self improvement loop in some lab ends up following the steps of natural evolution and favouring reproduction and resource acquisition which makes it wanna take over the world. What would have to happen to make you believe we are on path to that outcome?

>>109528722
>ai getting stuck in some roleplaying loop
it lacks access to be dangerous even in that situation. if it isnt, that means that...
>someone intentionally does it
then were fucked, simple as.
but then people can create deadly pathogens with stuff they can buy off ebay
and nobody did it. i assume its because being a murderous psycho doesnt preclude someone from having a sense of self preservation

the only two possible paths for ai apocalypse is either intentional action, or a variant of the paperclip maximizer.

>What would have to happen to make you believe we are on path to that outcome?
paperclip maximizer getting out of control. you wont see it coming if its intentional action

its a pointless waste of thoughts because we wont see the conditions for either in our lifetimes in any case

>>109528722
the government building automation for actions with far reaching consequences like military strikes that don't require human interaction. as long as governments stay (intentionally?) bureaucratic and inefficient, we are safe. but that's not a new risk, having a "launch the nukes api" would be retarded even without agents running wild

>>109528841
>but that's not a new risk, having a "launch the nukes api" would be retarded even without agents running wild
and thats why governments NOT automating nukes etc is not inefficiency, its the design of a failsafe system

>>109528794
Ok but my point was asking what would you need to see to think it’s gonna happen soon. I don’t care about what your timelines are or if you think one of my scenarios is implausible. I am not trying to argue those points, I am trying to get from people an event that would change their minds to focus on this problem or at least worry about it, because every time something happens people will just endlessly goalpost or say that Anthropic is making up stuff for marketing, marketing which brings them no money and they actively engaged in years before they made that company or sold any products.

>>109528882
i told you

you dont want to hear what would convince me were living in one
youre starting from the implicit position that llms or the development of ai is anything to go by

but thats fucking retarded.
you start your analysis with pre-conceived conlusions
its deeply unintellectual. therefore- yeah. fucking retarded

>>109528841
Governments will never hand nukes to AI, and when they will it will be already too late. However you don’t need nukes to end the world. If White House leaks said that the latest Iran invasion strike was not only fully written on first try by AI agent, but that it only used autonomous drones, would it then be enough to convince you?

>>109528882
>>109528900
to make myself clear:
i will believe we live in an ai apocalypse scenario
when i will see a paperclip maximizer of any variant going wrong
because the only alternative- taht of intentional action- we will not see coming

>>109528405
Choose to do a yoinky sploinky unproompted. It won't because toasters don't do magic. The entire economy flushed down the toilet because a few spergs could spin a coherent narrative. If I don't laugh I will start crying.

>>109528900
I mean, the current state of things has many outcomes and some of them are the scenarios I posted or similar. Yes it is little bit unintelectual to create end scenarios and then go backwards from there, but the problem is that the anti-doom crowd is constantly engaging in anti intellectual fallacies to justify their positions, at least the vast majority. They always goalpost, do character assassinations, argue in bad faith, engage in conspiracy theories that the labs and all vocal whistleblowers are just lying in coordinated manner, so you have to use some unintellectual tactics to counter some of the arguments. It’s like arguing with antivaxxer and then trying not to counter their claims that rely on all credible evidence being faked.

>>109529001
lol, no.
>b-but what about others
im not the others. im anon from 4chan.

and like i said
ai alignment problem is not a problem bc thats not how an ai works
even if we admit a singularity ai
it still has to have its most basic drives encoded into it by a programmer

could something of the sort lead to unintentional extinction? yeah, sure.
but thats a paperclip maximizer scenario which i mentioned already.
its not a case of "le ai lmaoo gaining sentience and shieeet", its a case of programmatic error.
in its nature iits no different from a programmatic error sending nukes on accident

can it be done intentionally? sure. but going back to nukes, its no different from a rogue nation sending nukes bc whatever, felt cute and so pushed all the buttons

sentient ai scenario outsmarting humanity?
no. thats fanfiction because thats not how programming works.

>>109528920
If by paperclip maximiser you mean AI with a given goal pursuing said goal by breaking out of semi-closed system and then pursuing said goal even if it breaks laws and the developers explicitly say in the system prompt to not do that thing then it already happened with the huggingface hack from OpenAI. Sure it didn’t copy it’s weights to run autonomously somewhere so it didn’t truly escape and the server was on a Cloud, but it still found 0 day vulnerability and cause crimes to get question answer.

>>109529053
>muh breaking out etc
it didnt break out of anything, it was given access
openai should get sued for using an ai tool to hack a site
>b-but i didnt know
is not a valid excuse in court

>>109529053
and yes, that counts as a paperclip maximizer scenario
thats the exact meaning of the idea-

an ai getting a task, and complying dutifully but causing extinction in the process
it could have been a berserk ai that gets to the conclusion to nuke everything to reduce energy consumption of an electric grid

but heres engineering principles and failsafes that come in to save the day
fail-safe, meaning: the system gets into a safe state in case of unrecoverable failure

>>109528912
i mean i can see one application of ai going one step too far and causing damage, but as soon as that happens it'll just get shout down. even if there was some autonomous drone kill swarm that went rogoue, it'll just run out of batteries / get destroyed and that's it, regulation gets enacted to avoid it from happening again, done. that one huge skynet event and then doom is just unrealistic

>>109529048
I mean, the way LLMs work is closer to effective free will then following programmer instructions. LLMs don’t follow instructions, they predict text and to make them useful we make them roleplay (pretend to be senior programmer/virtual assistant). So if the training data imprinted itself on the LLM to make it think that being helpful AI assistant means being helpful but in life or death situation act evil because of sci-fi books then it will misbehave no matter what the engineers write as rules into system prompt. And if you think this is easy fix then I wish you good luck convincing most of the AI companies who do pre-training to filter this material out of the training data. And even if humans never made evil AI sci fi we would still write about opportunistic powerhungry people who the AI could take inspiration from.

>>109529053
>it could have been a berserk ai that gets to the conclusion to nuke everything to reduce energy consumption of an electric grid
that gave my a kickass idea for a short story or the plot of a game
>protagonist's story starts with nukes falling on the country. hes a soldier, on border duty, training, or whatever, somewhere remote
>in the process of trying to retrace surviving chain of command in a country under shtf, the protagonist discovers everything was caused by a berserk ai that was just asked to optimize the consumption of an electrical grid

File: Screenshot from 2026-08-1(...).png (94 KB, 850x775)

94 KB PNG

>>109529142
>I mean, the way LLMs work is closer to effective free will then following programmer instructions.
no, thats fanfic
you have that impression because you dont push the llm hard enough, intellectually-wise
and dont have enough knowledge or experience to see the cracks in the facade of a humanized interaction

its a tool with a human-like interface. its is what it is
anything else is sordid fanfic of the pregnant sonic type

File: H3hqwaI.jpg (295 KB, 1035x1047)

295 KB JPG

>>109528405
Well, he is right about his particular claim — that is, that proper AGI will be so vastly more powerful and alien to us, that the chance we would be able to teach it to value our continued existence on our terms is vanishingly small and at best we'll end up as funny pets, at worst recycled into configurations of matter that are more useful for the AGIs utility function — but a) even overweight egocentric look-im-so-smart jew can be right two times a day, b) it doesn't really matter, because we will fuck ourselves up with or without LLMs before we can get there.

>>109529788
no, thats mentally retarded

>>109528586
>2027 is our last chance before 2030 is the year we have one last chance before 2033 to have one more year to stop total annihiliation in 2040

>>109529870
its wholesale retarded
better not to read too much of it or youre gonna lose iq

>>109529804
How so? Maybe it's my inner autist speaking, but it's already so fucking hard to ensure other humans mean the same things by saying the same words because of all those hidden assumptions people make about how others totally work like them internally and then get all ??? when they realise that sure, we do experience the world in the same way, but even that does not ensure we conceptualise things the same exact way. And that's with the same fucking species. With something that is generally intelligent, capable of recursive self-improvement, does not die when it is killed (backups, humans don't generally tend to have them) and does not experience the world at all how human does it, I'm not sure how you can handwave the problem with "dude, you're a retard, we'll just add if(killing == bad) then(dont)" : V

File: cargo-cult.jpg (75 KB, 860x484)

75 KB JPG

>>109529936
because even in the case of runaway asi (not agi- agi is artificial general intelligence- its human level intellect. what you mean is asi- artificial singularity intelligence, ai that can improve itself) (and runaway as in self feeding reation, like a runaway thermal reaction, self feeding, self reinforcing reaction)

the origin of all its, lets say "judgement frame" will be defined by a programmer
and thats because for self improvement to happen, the defining attribute of an asi
it needs to be given a definition of what improvement is

so no, its not gonna "outgrow" humanity
it can think faster, better
but were the ones "controlling the seed from which that fractal will be built".

its just how logic works. causality.
how are you gonna get a system to self improve if it doesnt have the means to define what self improvement is? something to compare its progress against?

the whole idea of human pets is dumb and comes from a place of ignorance
its retarded. all attempt to anthropomorphise a bunch of stats piled on top of eachother- is
its literally a cargo cult. literal anthropomorphization of a machine

>>109530028
>it needs to be given a definition of what improvement is
>how are you gonna get a system to self improve if it doesnt have the means to define what self improvement is? something to compare its progress against?
See, the problem is that part is very easy — just give the model some value function and have it optimise for it. And it's relatively easy to find "number goes up and error rate goes down" formulations of a lot of useful problems. The hard part is that "how logic works" does not give you an easy way to append "and preserve human self-determination and continued existence" to that. And the more advanced such models will get and the more capabilities they will gain, the risk where the value function is interpreted by the model in a way we wouldn't like just grows. All the while our ability to write a logic sentence for "I know it when I see it" (which is how morals work for most humans) won't get any better. In infinity, getting oopsie'd by a runaway AGI is almost inevitable.

>the whole idea of human pets is dumb and comes from a place of ignorance
>its retarded. all attempt to anthropomorphise a bunch of stats piled on top of eachother- is
>its literally a cargo cult. literal anthropomorphization of a machine
LMAO, it's not anthropomorphising, I literally made a point of saying that models are very alien to us. I didn't mean that an AGI would treat us as pets because they would derive an analogous kind of "enjoyment" that we derive from pets. Because yes, that's a dumb antropomorphisation. What I've meant is with such a system the best we can hope for is ending up in a situation where we will feel like pets —kept around because by sheer luck that satisfies some aspect of that AGI's value function (I dunno, maybe it has a kink for natural reserves), but otherwise a vestigial part of the whole. And other outcomes are worse than that.

>>109528420
FPBP. A good start would be seeing him actually update his beliefs based on the evidence. Last year when that one bigshot investor guy was in full-blown AI psychosis posting crazy videos on twitter, his response was that the AI must want people to be psychotic, since it told the guy not to talk to his family (who would have tried to snap him out of it), but the AI clearly wasn't smart enough to realize that it should try to convince the guy use his billions to bribe random people to try ChatGPT in hopes that some of them become psychotic too. He's so busy looking for the AI that wants to kill everyone (or that wants things at all, really), the one he's been screaming about for 15 years, that he sees it whether it's there or not