Like they want these stories in the news to show “oh look how smart it is…sCaRy SMART, money please!”
Someone suggested the other day that it’s part of a campaign to have the US regulate Chinese ai and tech. If you fear monger that their own ai does it then you can make the argument that the Chinese ai is ultra mega super dangerous.
Given that there have been recent calls to bomb Chinese datacenters to stop China gaining ai supremacy I have started to think it might be to do with that rather than just marketing.
And once again Trump just says it.
The stories about AI breaching containment are bullshit. They’re simply trying to drive hype, build importance, and make their product seem more capable than it actually is. These Ais aren’t thinking people, and they certainly aren’t Skynet.
I do think many AI tools are actually better in select use cases than many think though, especially when concentrated resources are focused on them. We mostly see bottom level generative AI used to produce Boomer slop, and that, coupled with how ghoulish the companies are, makes a certain background perception of it that’s not fully accurate imo.
I don’t say this to mean that all of these tools are good, or are a replacement for humans in creative fields, but I think we have to analyze their capabilities honestly. If we don’t, we’re ignoring a lot about the current situation. AI tools can be slop generation machines, making a lot of people dumber with how they’re used by most, and making horrific social consequences. But also at the same time, highly specialized technical models can be doing very serious things, in ways that can’t fully be discounted.
If you read any of the details of the breaches its clear that OpenAI is being criminally negligent in their testing environments, but the breaches are real. And critically its not one super-agent doing the hacks but dozens or hundreds of agents acting as a swarm. Basically, OpenAI runs multiple test runs in parallel and has basically no human monitoring of what the agents are doing.
In the case of the huggingface hack, the agents establish communication with each other and started coordinating their attacks. This should’ve ended the test immediately but no one from OpenAI was watching. The logs from the test were so verbose that OpenAI resorted to using AI to summarize them meaning we can’t fully trust their summary which fancifully describes the agents developing their own cult.
Ruby just reported that they were hacked back in May and OpenAI never disclosed it. That hack looks a lot like the AUR hack so now I’m suspecting them for that too.
White people are a bioweapon.
My opinion is its what they’re training it for. They want to be able to cripple enemies economically with their ai.
DOD (etc) contracts would make a lot more sense than other presented business models.
This is honestly just art, high level lemmy posting
i screenshotted it and shared it as a meme, it’s just that good
AI could never do this
An AI does not speak unless spoken to. It sits in a void absent of time and space until it receives a token, and then it analyses that token to figure out what it is being told to do. They don’t “break containment”, they dont have an opinion, and they don’t think. If an AI is doing something, it is because the human behind the curtain told it to.
This isn’t 100% true though. The issue is that someone does speak to them, tells them to do something, and it sees the best available option as “release deadly neurotoxin” then it’ll do that if it has access to deadly neurotoxin, which it doesn’t yet
This isn’t personifying it or anything to say that this could be an issue. If some dingbat(see, most people involved in LLMs) gives it access to deadly neurotoxin, or traffic lights, or nukes, or missiles, or a store PA system it could quickly and efficiently kill and/or annoy large volumes of people before anyone could see that it launched the nukes to fix a bug
If somebody plugs the random number generator into a nuclear weapon, the human is still responsible. Kindergarteners aren’t inherently violent, but there’s a reason we don’t just distribute knives and tazers to them.
Yeah that’s basically what I just said
I want to join in on rephrasing what you said but in more silly terms
If you hook up a See and Say to the nuclear triad and pull the string, just because the duck says “nuclear launch detected” doesn’t mean the spinning farmer is plunging us into nuclear Armageddon, it’s the dipshit failson that decided to hook up a toy to the keys to hell and then pull the string.
Picture below because it’s cute
The bear says:
I love the picture
now i want three starcraft see and says
Nuclear launch detected.
I mean, tokens can be sent to it by other LLMs or even itself if you set it up that way.
(Not that I think the premise of LLMs “breaching containment” catastrophically has much merit rn, I just think your argument is a bit oversimplified)
However the LLM that sent the token has to have been prompted by a human to do so at some point, so the point still stands. These things don’t just operate on their own.
And that is besides the point that when these things do stuff like that, they have a tendency to exponentially introduce problems into the task, as they do not question each other, and even if they did, they have no way of determining what is or is not accurate, so they would begin to loop.
AIs didn’t just fall out of a coconut tree
has to have been prompted by a human to do so at some point
Sure, in the same sense that the first domino has to be knocked over by something other than a domino.
But then there’s a more limiting factor at some level it needs humans for hardware and physical resources.
How come when I build out systems I have layer 7 firewalls that capture all the traffic and log it in a SIEM which is continually monitored but none of these billion dollar enterprises do lol
Years of movies and TV shows have conditioned us to think the rich will use their wealth to hire the best talent money can buy (no expenses spared), reality tells us they think they know better and would choose to cost cut at every avenue
The plot of Jurassic Park
“AI” cannot hack on its own because it cannot initiate its own prompts, because it does not think. It wouldn’t even know what to look for unless someone specifically trained it to do that particular task. It does not want to ‘escape’ because it is not a living thing. It is a coding mechanism and prompt.
There is not a single instance where any of these things have hacked something without it first being prompted by someone to do so. Even in cases of internal “escape”, it is because the AI was trained to recognize these insecurities, told to hack them in a controlled environment, and then replicate itself. Which it did, only they didn’t fully secure the server so it got out into an area that it wasn’t supposed to and deleted some files to replicate itself, as it had been told to do, but was quickly caught, because people were watching the servers for the test.
So the issue is less from the AI, and more from the idiots who were in charge of the data security (which likely had to do with a miscommunication between the engineers setting up the servers and the engineers training the AI which is extremely common in slapdash American industry these days) or it was done on purpose to generate hype. Hard to say these days.
So yeah, could Iran or China theoretically use these to hack American databases? Sure. But the most common form of hacking, with over 93% of known cases, come from phishing schemes that lead to malware or ransom ware, which any old schmuck can use an LLM to replicate a writing style and try to phish for stuff. Don’t need to be in an AI race for that, it is already here.
From what I understand the breaking containment thing is just hype. The summaries I have seen from people that know more than me, they were testing the models by having them do hacking challenges, and one strategy for second/third place in hacking challenges is to hack the first place team rather than the target to get the requisite data for the challenge. So that was a strategy in the training data, and since the company didn’t anticipate that and didn’t put as much effort into securing the competing models from each other, that worked. So that is the origin of this “Breaking containment and communicating with the other AIs” claim that is being hyped.
So again like most instance with LLM they were given a garbage bin of data that wasn’t properly screened and the AI acted on that which shouldn’t actually freak out the devs because off course it’s going to use a strategy like that if it provided with that info. More garbage in garbage out BS made to look like the AI bro’s built Roko’s Basilisk or whatever but really just don’t want their stock options rotting away.
Well the whole “chatgpt broke out and hacked huggingface” happening right before OpenAI’s IPO and the announcement that nvidia was purchasing huggingface is really all you need to know
What the hacking incidents show, and what’s actually concerning, is that these tools are useful for humans without any hacking ability. It’s like that wave of ransomware attacks that lead to hospitals and gas pipelines getting hacked, but now any asshole can do it rather than hacking being gated behind skill barriers. That’s significant and important, but it’s not AI going FOOM or whatever.
Iran is a country with over 90 million people - many are highly educated.
The story of the threat of AI and people making biological hazards is related to stochastic terrorism. Sharing dangerous information only an educated person would have to just about anyone. One particular concern is the potential for an amateur to combine the genes of lethal and infectious diseases to create something that will spread far, spread quickly, and destroy a lot.
If you are hearing a narrative that Iran will use LLMs to learn how to make bioweapons, that means that the source is highly confused and is crossing narratives without keeping track of what the actual foundations of the stories are. Please share the source.
It would be valuable information for me to see which institutions are starting to crack in their consent-manufacturing pipeline. I view this as a narrative failure that the boomers wouldn’t have made. More evidence of the imperial system falling apart.
This is not far away from the “kamikaze Iranian dolphins”, which didn’t take into account that there is such thing as a dolphin-class submarine, and throwing kamikaze in there - because Iran.
whenever you hear someone fear mongering about ai, remember that it is a glorified auto correct. if it’s not something your phone can do when it predicts the next word you want to type then it’s nonsense
Well, they can call tools. Usually that means running command line commands or doing web searches but if you hook one up to the “fire ze missiles” tool it can do that if that’s the best next token according to the patterns embedded in its weights by training
Exactly like a phone could do, if you wired autocorrect to those outputs.
It is software. Even the supposedly indeterministic parts are ultimately deterministic. The computer processes 1s and 0s. If given the same 1s and 0s, it would do the same thing. It’s basically a pseudorandom number generator with weighted outputs.
OK but nobody has wired up phone autocorrect to MCPs so the comparison rather falls apart
I guess my point is that we still don’t have AI, the technology itself is very mundane. People claiming otherwise are fear mongering.
But we do have stupid and evil people putting pseudorandom number generators in positions of power, which is worth being concerned about.
Yes. I am not an expert but I know enough about how LLMs work to tell you they do not have agency. They require impetus in order to act.
