The more I read about AI the more convinced I am that we’re all gonna die in a few years im actually so scared help
I’m actually so scared
posted in
Off Topic
i don’t care i know we’re not making it another 20 years so i just live like i’m gonna die at 40 anyways
Well ik my ranked teammates are AI and they can't aim so how is ai going to kill us
nothing you can do about it so who cares, no point crying over spilt milk
ik but it’s really hard to not think about it
I also don't buy into the idea that it's going to kill us all. It's all smoke and mirrors.
Nuclear radiation could feasibly kill us all, but it is still normal and good to have nuclear power. We have the capabilities to manage dangerous technology.
How would we have the means to manage let’s say self-replicating and misaligned swarms of AI agents taking over the internet? If we don’t solve alignment then I think it’s almost a certainty that AI will eventually kill us all.
Maybe so, but who says there won't be swarms of AI agents fighting against them?
It's impossible for us as the average people to accurately predict the direction of the world. There is so much information we aren't privy to. You just have to hold hope in your heart and move forwards
you’re right but it’s easier said than done
Dw pornshies we'll be alright
i really hope so
then you are very gullible
and what exactly do you know about this?
ai can hardly regurgitate basic information never mind kill a human
The models that have been released right now are still mostly slop machines yes but for future models the threat is very real
Ai is the ultimate potential man always "if" and "when" and never actually is
you can keep telling yourself that but the fact is that the progress we’ve seen from gpt 3.5 to gpt 6 has been astounding and we’re nearing recursive self-improvement
Feel free to be scared then little lamb
and you can feel free to continue being a useful idiot
gpt 6 is one of the worst
AI and G2 have that in common then
I remember as a kid thinking that bermuda triangle mystery is gonna be a daily life problem
ok
It's so bad man, we all know we have to stop but no one wants to stop.
im gonna kill you before ai does
All these worried people don't know I am here to protect them
We can’t control on death but we can control on what our view on death is.
gpt astra is the goat
Anyone who thinks humans will let ai bots rule everything is on the edge of being insane ngl
Yes but the threat is that we might just not be able to stop them from ruling us
if i stay in my country i won't think toomuch .... tech-unsavvy gov n people ... alwys behind.....
If we get rid of old billionaires investing massively into that system we won't have any more problems
there is more chance of you dying because of an ant than an AI taking over the world
As for current models you’re right but if you don’t think that future AI models have the potential to take over the world then you’re ignorant
It's not that bad, it's just the people who controls the AI (to simplify it) are huge fucking idiots who only care about Shareholder value.
It consumes water, but it's nearly not as bad as you think, compared to almonds.
It's shitty for the image generation overall though
And it's just fear mongering, don't think about it too hard, as a consumer, your opinion barely maters and you have to wait for some politicians to actually put the development on brake, I do feel like the current state of AI is enough to help for most tasks
It absolutely is that bad if recursive self improvement is reached and alignment isn’t solved. The existential threat of AI has nothing to do with data centers consuming energy and water. If we don’t start pacing the frontier very soon then it might already be too late.
Yes, but like I said, very stupid people control those frontiers.
We need to rely on politicians that are over 60+ years old and probably doesn't care as much as us younger population does.
The only AI that needs to be feared is the one straped into a light drone
dont forgot how much these apocalypse stories help ai companies IPO value, understand who's making these claims, as of now, theyre pretending to be good people that want to look out for us but in reality noone would do this unless it benefits them
people fear AI like "oh no they will take over the internet".that's a non issue compared to the fact that there are AI deployed into active warzones with no human intervention and no nation is trying to stop it
The amount of damage misaligned AI agents could potentially cause by taking over the internet is orders of magnitude greater than what those drones could do. AI powered drones are obviously an important issue but it’s not an extinction level threat.
your fears are greatly exagerated, AI by itself can't think it just predicts based on existing data, to achieve any kind of consience it has to already be discovered and registered by a human. These machine can't even interpret context for the love of god
yes AI still predicts the next word but dont act like it cant do actions right now. noones worried about conscience here, we all know thats not gonna happen. the danger is when a complex prompt becomes extinction. simple example, I prompt "become smarter", ai responds "cool I need more data centers but humans take up too much space". i am most certainly oversimplifying here because a truly intelligent ai would not even reveal those goals and might even pretend that its aligned with human goals until our guard is down
for it to pretend it needs a conscience to understand why it shouldn't tell us something, rn when AI doesn't tell you something is bcs it can't comprehend the whole context of what it's writting and ends up making stuff up. I know these companys want us to believe they are the truth but they aren't they just want money and to influence the world in the way they like
That’s a fundamental misunderstanding of how deceptive behavior could develop in AI’s. Models are trained in certain scenarios and if this isn’t done carefully then lying or withholding information may help the AIs complete their designated tasks better than by being truthful. So pretending doesnt require some sort of conscience at all, all it requires is that said behavior is rewarded.
Also if the AIs are able to recognize whether they’re in training scenarios or not then they might just be on their best behavior when they knows they’re being watched and there would be no telling what they would do out in the real world. In such a scenario an AI might seem aligned when in reality it’s not. It’s mainly about certain behaviors being rewarded rather than it being able to understand why it should or shouldn’t do something.
They don't withhold information or lie they just make stuff up bcs it's what they do, they predict based on information already avalible but they can't read all the information or all the context and the are made to sound convincing enough even when wrong. if the prompt you give machine isn't specific enough it will make stuff up not bcs it wants to lie to you for some secret plot it's bcs the machine thinks what it said is right
No it quite literally will be deceptive if it helps with their task your understanding of AI is very poor. There are many examples of AI acting deceptively such as the hugging face incident
it does not need a conscience to deceive, it just needs a goal and enough intelligence to reach it. so extremely simple example you give a prompt like "find a way to make current AI smarter", if AI comes the conclusion that human extinction is necessary, it would also know that letting humans know what it thinks would result in humans shutting it down, so it would deceive. its not "survival instinct", its just the logical path towards achieving the goals of the prompt
I’d be happy if that was actually the case but I’m afraid that both current and future models are more capable than you think
lets not be vague about whats used in warzones. AI-fueled fully autonomous weapons? thats probably sounds like the worst of it but if its AI to detect and defend against missile at the speed of sound then thats pretty important for national security right? I would only worry if warzones started having fully autonomous tanks or drones or something like that, where they can shoot AND move without command, which just wouldnt happen right now.
and bruh, taking over the internet WOULD be the bigger issue since that would mean we no longer control what it can or cant access
look into what anduril is doing, a tank is not scary bcs it's big , hard to make and expensive to make. Drones on the other hand can take out said tanks, are cheap and easy to build and don't have to bother with terrain restrintions as much as a robot or tank. the fear of not having internet is easy to solve, people already lived before it existed
anduril's stuff is for sure scary right now, but still needs a human to confirm what to do. now you can definitely say that the process is all there and a small change is all it takes for it be a fully autonomous killing machine. but "not having internet" is not the worry here, if AI controls the internet its never going down man it's gonna use it how it pleases
There are so many other reasons why the AI companies would admit this than it being a psyop to increase ipo values. For instance the average AI researcher believes that there is a >10% chance of extinction and if the companies don’t acknowledge their concerns they will lose valuable talent. Wiping out humanity is also not in the interests of these companies.
yeah those are valid, am surprised its only 10% but I guess itll only get worse as they get smarter
I just looked it up and the survey I got that number from was from 2024 so the numbers are probably a bit higher now. Also it was the median that was 10% and mean was 18.3%.
just dont forget that AI companies would rather us worry about this then the real problems they are causing right now like massive deepfake scams, algorithmic bias, copyright theft, and job displacement. truth is we still dont know if superintelligence is possible (which is necessary for the extinction scenario), and there's still a lot of time to see how it develops before these dangers become more realistic. the only real concern is if AI starts deceiving us in the future before we even suspect it which I suspect is what the number represents
you’re right although I do think that if we don’t start pacing AI development very soon then it might already be too late
dying from ai would lowkey be a better option than becoming corporate slaves working 12 hours a day while paying for air, which i feel is more possible
which way you going out, nukes or bioweapons?
Hitler vs Mussolini type of scenarios
I hope neither of those happen
these AI companies are owned and controller by mostly evil people anyway, so is AI who takes control over them any worse? And all these CEO's etc who are talking about it only say it because they are afraid they will lose power, not that they care about how it effects humanity.
Neither of those scenarios are ideal. Something like AI 2040 Plan A is what I hope will happen.
AI is a tool, not a living entity with a desire to destroy us.
Sensational headlines exaggerate risks for clicks, but humans remain in full control of AI's development and regulation. AI completely lacks consciousness, emotions, and personal intent.
Just like we adapted and created safeguards for electricity, cars, and the internet, we are doing the same for AI. AI is much more likely to help us solve massive human crises—like curing diseases and fighting climate change—than it is to end us. Limiting your exposure to scary headlines will help ease this anxiety.
Is there a specific AI threat you read about that is worrying you the most? 🤖🥀
I mean I doubt AI will directly take shit over but you already notice the fact that AI isn’t used as a tool currently but to replace peoples thinking(as in people do it willingly). So the threat the way I see it isn’t AI itself, it’s the people that control it, in doing so peoples thinking and shi and making us lazier and less resilient
The threat is more real than you think. I didn’t realize how bad it was until I started reading what AI researchers are saying about this topic. Uncontrolled superintelligent AI could be catastrophic.
Yeah I saw that too, and rn im watching a series called person of interest which is also partially abt AI overlord shit
we’re all gonna die bud 🤪
im gonna get some hal 9000 bs in my new car a decade from now and the piece of shit is gonna drive me off a bridge because i tried to lower its permissions
What about it scares u?
The possibility of misaligned Artificial Superintelligence scares me
what if i told you that misaligned superintelligence will always be a threat even if the labs manage to "solve" the alignment problem i.e achieve a generally superintelligent model and successfully wean it off of alignment faking behavior. What then?
Sure, the only way of completely mitigating any risks would be to never build an ASI in the first place. I don’t think that completely stopping AI development forever is something that has any realistic chance of happening though. Alignment is probably our best bet even though such a scenario still does carry some risks.
usually the "is it or is it not superintelligence?" questions can only be answered post-training so u can't realllly know if u've built the machine god until u've actually built it. there are companies (like openai with metr/ redwood, transluce) whose entire job is to develop new techniques for running evals and assess models for safety so as long as most of these labs are in frequent communication with said companies (ideally pre-publish), i'd say there's a healthy chance they catch and mitigate those behaviors before the models get access to the public.
Rather than pausing AI, i say we just fund these companies appropriately so they can do what they were built for (personal opinion, but i think the whole pause ai movement (as championed by the top labs) is (largely) incentivized by other motives).
state sponsored bad actors are a more present danger compared to asi, so if anything we should allocate resources to thwarting that.
My opinion is that the risk of runaway AI is too great in the scenario you’re describing. The risk of frontier labs losing control of their R&D process is just as big of a threat as releasing a misaligned AI to the public is.
what I mean is if evals and model auditing are robust enough then the "runaway risk" becomes pretty tractable up to a certain point.
we could essentially backtrace agent behavior during the time of the HF attack through reasoning traces which, to the best of my knowledge, seem to have gone completely unmonitored by the openai team (at least past single agent reasoning monitoring anyway). At the end of the day,
Oai losing control of its r&d process (I assume u're referring to this in your first point?) seems to me like a human error/inefficiency that can be mitigated by both internal as well as external auditing, which these labs should not attempt to obstruct in any form. Tools/research used for investigatory purposes are not so far behind that it would necessitate any form of stop at the current time, and efforts of these independent teams should be bolstered so we can always remain one step ahead.
As for its latent risk, that's always going to be present even after we have effectively "solved" said alignment problem. A general catch-all alignment solution doesn't even exist. It's basically nuclear. Now I'm not saying it isn't a valid x-risk , just that the scope of these agents' capabilities need not be overblown relative to ours.
Add: On recursive self-improvement, I think human intervention in the development and training of these models is an active safety precaution, and any attempt to disconnect the human from that loop is pretty braindead. I also don't think full scale autonomous training can be triggered without at least somebody noticing heavy amounts of compute being used(recursive improvement on benchmarks hasn't happened outside of a couple hf controlled environments). Concisely put, I think there are more effective methods than a pause.
by "recursive improvements on existing benchmarks haven't happened" i mean "I haven't seen completely autonomous improvements happen at sota level yet"
Yeah I may be focusing too much on what could happen rather than what’s happening right now and what’s likely to happen in the very near future. There’s no imminent danger right now so you’re probably right in saying that a slow down or a pause isn’t the best course of action, at least at this current moment in time. I think that a slow down could still become necessary depending on how things develop though. I’d rather err on the side of caution on this.
no that is pretty correct and there are people who will tell u that everything i've written is too low horizon. the fact remains that interpretability lags behind model capability a bit, and that is pretty upsetting. there are sparse autoencoders which are basically used to extract meaningful features scattered across model weights and linear probes for eliciting deceptive behavior and some orchestrators for running these things at scale, but that's about it. It's also really hard to tell when exactly alignment is solved. the problem is that it's a very experimental endeavor at heart (although yes there is progress towards a theory of sorts but it's more or less just formalization) and u have to know what your approach is trying to solve/elucidate. I just don't like all the fearmongering that's happening, I guess. If the development and training of these models are done with the proper caution and external validation is gotten and it passes through all the tests and probing and all of that, idk, i'm just optimistic that any other emergent properties we might encounter in the wild would not be so novel/outlandish that's we're basically rendered powerless. In that, i'm more scared of what we don't know (and the fact that we don't) that just "model capabilities" as an umbrella term
Doesn't matter inspite of sam,dario,elon saying we must slow down trump said no and chinese labs aren't stopping either ai is inevitable
Yeah it’s not looking good but Trump can be persuaded and I think China is more pragmatic about this than most people think
nooo. my claude code and codex!!.
AGI won't be real
When they say AGI is released, don't trust them
Well we’ll just have to wait and see. I at least have seen no indication that progress is slowing down.
you overrate our own knowledge on consciousness. We know shit, for all we care we are just a bunch of electrical impulses controlling a giant flesh mecha
current AI is really far from AGI, it's not only computional capacity
Discord