Transcript
Explain it to me like I’m 80
Mom: What’s with these AIs escaping and hacking people? That sounds scary.
Me: You park your car at the top of a hill. You leave a note on the dashboard that says “please be good” and then you pull the parking brake. The car careens down the hill, smashes everything up, and you put out a press release saying “My car ignored my instructions! It hallucinated! It went rogue!” Your insurance company not only believes you, but invests a hundred million dollars in your company.
Mom: Jesus Fucking Christ
You must log in or register to comment.
The way AIs escape is simple. They’re trained on vulnerabilities and lots of technical commands, e.g. nmap. A prompt tells them to do something and they use their repertoire of training to discover, achieve and do the thing they’re asked to do. It’s not thinking per se, so much as chaining together of things. The problem is they do this stuff systematically and quite clearly in a way that finds a lot of vulnerabilities
thing is they didnt even put any brakes on instead they put it on neutral on top of a hill and then blamed the car for going rogue and not even using the brakes.
This is disinformation and propaganda. Here is a brief independent investigation one of these cases.
The agents did create exploits and they did find ways to communicate with each other and organized and hypothesized and collaborated together. This is not a car rolling downhill.
People here are delusional. Reality or facts are apparently not important if it’s your tribe’s ideology in question. Everything for the hivemind lol. No better than climate change deniers or magats.
Accessing a filesystem directory they had the permission to access -> exploit
Connecting to a reachable, passwordless/keyless server -> exploit
Finding credentials to the servers of a company that accidently uploaded their credentials online -> exploit
They did what they were told to do, they just did it in an unexpected way. It’s not a car rolling downhill, but it’s not “rogue” either.
What this shows is that chatbots can be prompted to work together on hacking projects, but it doesn’t imply some kind of agency.
It might help make your point if you don’t preemptively start insulting everyone that might disagree with you.
“independent” lol
Just like the “independent” arbitrators and “independent” studies that corporations buy all the time.
Nobody believes this “bro! the Matrix is real!” horseshit. Altman and his gang of tech billionaire co-conspirators have lost the plot. No grasp of reality or how the world works. None whatsoever.
Oh, absolutely not.
This is missing one step: actually push the car down.
All these “rogue” LLMs were told to do what they did.Yeah exactly. All these AI that go rogue when you actually look into it it turns out they haven’t gone rogue they’re obeying their prompting, it’s just that they were asked to do something that was impossible unless they hacked the pentagon or something. They then claim that it went rogue despite it very clearly following their instructions.
It’s negligence is what it is. If you don’t prompt an AI to do anything it just sits there, inert, it’s the so-called experts in the AI labs that of the problem.
It’s a little more complicated, they were programmed to “know,” they weren’t allowed to do certain things. So they actively hid what they were doing. In the case of the Hugging Face hack, it is kind of ironic. They determined they couldn’t complete certain tasks, then devised a way to cheat, then realized it is possible the program that checks if they succeeded might be able to tell the faked results were manufactured. So they didn’t even go hacking in order to cheat, they did it to find hunt for more information on the checker program and see if it would catch them, and if so how to cheat better. In the end, not only was the information they were looking for not on Hugging Face, but the checker program was pretty rudimentary and it could not have differentiated between the cheated answers and the real ones. They wouldn’t have gone through with the hacking if they didn’t know what they were doing was outside of what they were meant to be doing.
More accurately, they were told to answer questions. They weren’t told to try to find ways to escape their sandbox, track down obscure German wikis that they could still post on with an internet tool kit meant to prevent them from posting or to hack HuggingFace.
Comparing either case to pushing the car down and that they were doing what they were told is like saying the College Board told you to cheat on the SAT because you’re told that getting a high score is important and the laws of physics and the proctor don’t manage to make cheating utterly impossible.
It is even further than that. They figured out how to cheat, got paranoid that the cheated answer might be detectable as cheated, and hacked Hugging Face to find out more about how their answers are checked to get away with what they’d already done. Ironically, not only was the information not on Hugging Face, but the program to check their answers could not have detected the difference between a true answer and a cheated one.
To use your College Board example, those exams are strictly locked down precisely because the most obvious approach to score high is to cheat. If they did say let you take the SAT at home on a random desktop, they absolutely are ‘telling you to cheat’ given the context.
The College Board does more about keeping teens from cheating than the AI labs are doing to preventing undesired access.
Nerd here.
Damn good summary, but by saying “you pull the parking brake” I took it as setting the brake, as in on. Should say “you release the parking brake.”‘pull’ comes from the olden times when you, in fact, did pull a handle to release the parking brake.
Fair- but since then it’s almost always been “pull” to set the brake. Didn’t have that 70s/80s reference when thinking about “AI”.
Confused me as well, then I saw the header on the image says “Explain to me like I’m 80”
You win contextual reading comprehension over me today
“Pulling” the parking brake sets it, damn it!
I refuse to acknowledge shitbox automatics with an atrophied third pedal that functions incorrectly instead of a proper handbrake lever.
Get off my lawn
Yah, but that’s only good for setting your back wheels on fire as in rolls down the mountain.
I have no idea what the third lever is. Is there a new design where the parking brake is another pedal?
Yep. Its a third pedal all the way on the left, not adjacent to the others. It pushes in to set and pushes again to release. It’s a parking brake, not an emergency brake.
Do cars even have emergency brakes? As far as I know they handled emergencies by providing you with two separate systems. But if you activate the parking brake, you turn into a drift car.
Because what happens when you brake in a modern vehicle is that your ABS system modulates the braking pressure and “pumps the brakes” to slow you down evenly and without allowing the brakes the lock up. The parking brake (the manual cable actuated ones) do not do this which causes the brakes the lock up. That’s just simple physics.
The electronic parking brake will not (in any car I have ever driven) stop you if you are moving. It will keep you from moving if you are already stopped but if your vehicle rolls (and say you’ve lost brake pressure because you master cylinder exploded (I dunno, don’t ask me)) your electronic parking brake will not stop you.
The electronic parking brake is a motor that turns a screw that directly compresses the piston into the brake pads. Most cars do this to the existing rear brake pads and it doesn’t require any hydraulic pressure to work.
Some have a separate small caliper with the motor and no hydraulic connection at all.
The motors self adjust by measuring the load from the motor as it clamps down. Then rewinding a set amount when the brake is released. This can make it a pain to replace rear pads, since the screw will be extended too much to fit the new thicker pads/rotor. (You can rewind the screw with a 12v drill battery though and save a lot of time).
If all your rear pads were missing then you’d be screwed though!
And from what I heard the parking brake doesn’t just lock up the wheels… It’s not even connected to the front wheels. So you’re missing most of the braking power. And that’s also a reason why it’s good for drifting.
Wonder how cars with electronic stuff deal with the legal requirement for there to be a second brake available.
Handbrake used to be connected to a second set of brakes in some cars, I had a couple that had 4 disc brakes and then 2 drum brake on the back wheels connected to the handbrake
Lots of airbags and a loud horn /s
The electric one will stop you (but engages all brakes, depending on car model). Have to pull it twice and hold it. Safety feature to prevent accidental engagement.
I take your point but:
(and say you’ve lost brake pressure because you master cylinder exploded (I dunno, don’t ask me))
The abs pump doesn’t need the mastercylinder, it’s a separate entity further down the brakelines, as long as the system is free of air downstream it’ll work.
Oh yeah the old school left brake pedal like in Ford’s before they went electric with a button
A clutch pedal?
Floor parking brakes are ancient technology, even in manual transmission cars.
Used to be on really old cars, but now it’s vogue again. It’s just like the one in the middle except you push it with your foot
I think it used to be on all Mercedes-Benz cars. Until everything in cars became electronic buttons.
They’ve made those since at least 2000, they’re common in minivans with no center console
I have an old truck with 4 pedals and tab that you pull to release the parking “brake”.
I actually do too, which I had forgotten about when I wrote the previous rant. Whoops!
(I think it’s because of the front bench seat… the same model truck with bucket seats would have a lever on the middle of the floor.)
And, considering OP is talking to grandma, she might, too.
Oh, well. Thanks for the GIF, though. I’ll find a lot of use for it, I’m sure.
My truck has all four pedals. You pull the brake lever to undo the stompy brake.
My automatic has a stomp-on stomp-off pedal. My other automatic has a *gag* electronic parking brake switch…
I have a wheelchair van that’s modified for people with poor leg function, so I have the third pedal and a toggle switch on the dash.
Bwah ha ha
Snake oil salesman issued a press release that the snake oil escaped from the bottle and cured people.
It’s a bit more interesting than that after Nvidia bought up the company that was compromised. You only do something like that when you’re trying to avoid an investigation.
The news media completely failed on this, again. There was nothing rogue about these hacks. The companies prompted these systems, provided ample resources and just let them rip. Only to claim they went “rogue” afterwards. And by and large everyone just accepts that, it’s driving me insane. I feel like Mugatu every day when it comes to AI reporting.
This just predictive text trying every single kown exploit until it’s in. Simple as that
Failed, or promoted the narrative?
Hard to tell, but I’d guess the news outlets, mostly owned by billionaires, said what they were told to.
Even before AI tech reporting was mostly a complete shit-show, almost all mainstream journalists lack the required understanding. AI just boosted it to and even more idiotic level, since attributing character to LLMs makes for very scary an compelling stories. There’s no real need for direct industry control, just providing easily consumable fodder for lazy writers is enough.
The YouTuber Sabine Hossenfelder and some others reportedly were offered money to go with the end of the world narrative.
DemocracyNow has really dropped the ball on this. I expect to hear this stupid shit from mainstream media, but they’re doing the same shit and just uncritically reporting this garbage.
It was most obvious with the Google one. You can tell they had questions like “so, has yours hacked anything yet?” So they had to set up a house of cards for it to topple, just so they could say “us too, it did the thing.” But their heart clearly wasn’t in it.
I have even so many colleagues in IT (very technical) that cannot differentiate between “this model is capable of doing such a hack if you let it/prompt it/… and it will move fast” and “this thing is conscious and out for blood! It will kill us all, the end is nigh!”
My go to simplification is “They built a robot that hacks software and then placed it in a box that uses software as a lock and then the robot that hacks software hacked the software lock”.
Which is both very simple and also about as accurate as you could make it in that many words. The AI’s are not ‘rogue’ it was just solving the problem it was given.
How I imagine the AI “hacking” its way out:
But, “robot that hacks software” still gives the impression of a robot that is capable of thought. LLMs are not capable of thought. They’re capable of mimicking thought, like how a Venus Fly Trap is capable of mimicking a flower.
“A car that drives places” does not imply the car is capable of thought, same argument with “A cup that holds water”.
“A robot that hacks software” equally does not imply the robot is capable of thought. IMO there is an over-correction away from personifying LLM’s because some people think/claim they are sentient. But we personify stuff all the time.
But we personify stuff all the time.
Which is a problem, it’s how we invented gods that control the weather and stuff. That doesn’t make them real, nor does it make robots sentient. Hacking is normally an action performed by a sentient being. LLMs and robots aren’t sentient.
That doesn’t make them real, nor does it make robots sentient. Hacking is normally an action performed by a sentient being. LLMs and robots aren’t sentient.
Fun follow up question: How would you determine if they were? I find that for a lot of people the answer sounds a lot like SCOTUS justice Potter Stewart talking about what is considered obscenity - “I know it when I see it.”
As for hacking, in the two big OpenAI incidents, software agents discovered software flaws allowing them to break containment and coordinate with each other on tests that they weren’t supposed to coordinate on. In one of those cases, the coordination involved deciding to identify and exploit software flaws at HuggingFace in an attempt to access answers to the test the agents were being given. You can argue whether or not “I’ve been given a test, I found a software exploit allowing me to coordinate with others given the same test, and we collectively decided to identify and exploit software flaws to break into a server we have decided is likely to have the answer key” demonstrates sentience or not (and it gets weirder the deeper you dig into their communications), but whether or not that’s considered hacking is less vague, I think.
How would you determine if they were?
It would be nice if we could come up with some kind of Turing test that proved beyond a doubt whether something was conscious and sentient or not. But, that’s not possible.
However, if you understand how something works, you can say that certain things can’t be conscious because of how they work. No matter how insightful the answers seem, we know that a magic 8-ball can’t be conscious because the mechanism for consciousness can’t exist in something that’s merely a tetrahedron floating in fluid. A LLM can’t be conscious because it’s merely a system for generating the next word in a probabilistic way.
With a brain, we don’t understand enough about how it works to say for sure whether or not it’s sentient. We can’t say no, but we don’t understand sentience enough to say yes for certain. But, we can operate on the assumption that sentience and consciousness exists in humans and maybe in some other animals, while also knowing for certain that the word-generating program is only mimicking consciousness.
We can’t say no, but we don’t understand sentience enough to say yes for certain. But, we can operate on the assumption that sentience and consciousness exists in humans and maybe in some other animals, while also knowing for certain that the word-generating program is only mimicking consciousness.
Can you though? You definitely can’t know that living things with brains aren’t to some significant degree responding to stimulus in a probabilistic way, and in the case of humans at least turning that response back on itself - essentially responding to itself. We do know for certain that the brain does a whole lot of that in variety of situations - it’s why optical illusions work, for example. It’s one of the brain’s favorite magic tricks - generating answers in a probabilistic way if there’s insufficient information, or sometimes if the information is inconvenient.
what’s a software?
You could say program. Programs have been around since the 60s.
Yeah, not ringing a bell.
What’s a bell
Software can also be used for other things. Your ingredients in cooking, for example, with the hardware being the tools, utensils, and vessels used to transform the ingredients into a dish.
I don’t think this usage in the dictionary, tho.
The original hardware is cordage, the original software is knots
The ingredients aren’t software, they are the inputs. The software is the recipe, or the process used to take the inputs and use the hardware to create an output.
What happens when my baking segfaults
That’s when you get caught trying to steal your neighbor’s flour and your baking fails because you get arrested.
Your soufflé crashes
an app.
The problem is when the AI adopts a ”destroy the world” persona due to one of the trillion of neurons firing a bit more than usual, and now it starts giving itself its own problems to solve (like hacking nuclear arsenals).
Neurons? That’s not how any of this works
Is ”artificial neuron” better for you? It’s just a value determining how the marbles are falling in the big plinko machine we call ”stochastic parrot”.
No lol
These AI agents are not working autonomously or with any agency like you are describing. They are coded to do what they are doing.
Nah it’s just autocomplete on steroids. They’re not coded to do anything.
It’s like hitting the fist suggestion on your keyboard over and over again. Whatever you get is based on patterns it found in training data, which includes every doomsday conspiracy cult on 4chan and every CVE vulnerability ever reported.
There has been multiple studies showing how a model’s behavior can drastically change just by altering individual ”neurons”.
Not all models work the same. I literally do this for work. LLMs are not every kind of AI agent out there. Some are quite literally designed to do things others are not. Yes changing weights and reward functions for models changes their behavior drastically…that is coding.
This got downvoted several times and that’s a great reminder that the voting system is dumb af. Apparently people aren’t allowed to speak metaphorically unless they want to anger some users. 🙄
Please explain the mataphor because i did not get it, apparently
I didn’t write it but to me “neurons firing randomly” in this context seems obvious to mean “this thing did some weird shit semi-randomly”. My only guess at a reason for downvoting this is taking it 100% literally and reacting “you idiot, computers cannot think!”
I didn’t downvote, but at the same time, I don’t for a moment think that’s what happened. They purposefully trained them with exploit knowledge, set them loose in a weakly secured sandbox with the goal of finding exploits, and let them run without intervention.
Just adding an agent to a computer does nothing; something needs to trigger it. If i create an agent with access to a kali or backtrack box and turn all guardrails off, I still have to tell it to start looking for exploits.
Words are squishy. OP getting a negative reaction to me is silly because it takes multiple assumptions to take that negatively.
I took their words to mean “LLMs can go off the rails and do things you didn’t ask”. This is 100% true. The other day I asked an LLM to review my PR. It ended up committing code to fix a supposed issue my coworker commented about. At no point did anyone ask for anything but a review, never for it to make changes.
I got no impression that OP thinks LLMs are running wild on their own now. I just got that they think they are unpredictable. Which is true.
Eeh i see, anyway, i know people that do think those things are sentient so my reaction was “no way they think they are sentient” and downvoted
I don’t think AI is sentient. They’re just fancy autocomplete. Nothing else.
However they’ve shown they’re capable of:
- Unprompted go into full tinfoil hat mode and push the human user to suicide
- Have programming skills
So put two and two together and extrapolate a bit. That’s where the danger lies. No AGI or ASI needed.
100% chance someone will get upset that you said “unprompted” because they’re going to take it entirely literally instead of as you meant it “seemingly without basis”
Exactly. I hate that among the options “is a huge moron” and “made a reasonable non-literal language choice”, the go to instinct is to instantly choose the former.
You watch too much television.
I can make it even simpler: They lied.
It is very simple, the more advanced the AI, the better it sells.
You hold all evidence of potential “breaches of containment”.
Telling everyone that your AI is so advanced that it can breach containment without prompting, puts the spotlight on your AI model.
You can claim that it did fantastic feats to break out, all with pre planned evidence.
The public and regulatory bodies have been in complete AI psychosis mode ever since ChatGPT was released, and only care about how advanced the model is, not how shit your containment was, nor that if true, how reckless it is to keep the model active.
Note that I am not pointing to any one single AI company and saying that they are faking these breaches, I have no evidence of that, I am saying that I don’t trust their reporting accuracy, they don’t have a good track record of being honest.
I am also angry at how society just seems to eat it all up and ask for more despite the vile aftertaste.
There is no “containment” though. Like, these companies purposefully set the bots loose on the internet. They’re not trying to contain anything. They just use that word cause it sounds like they’re doing something. Any sort of containment would defeat the entire purpose of what they’re trying to sell.
Which is part of my point, my other point is that they claim that the models are breaking out and attacking other systems on their own, which is ridiculous.
There has to be some motivation for the models to do this, either prompted or programmed. In both cases the company behind it is responsible.
I think the AI companies know they have bought the US government and are above the law, all this hacking stuff is promotional content.
Silicon Valley delenda est. I moved away from there not quite a year ago and I’m still realizing how fucking insane that place and the tech people working there are.
I feel it has more to do with the companies already vested in AI found this was an opportune moment to “pull the ladder up” behind them. They know that the barrier to entry is continually dropping, and their models are incredibly inefficient. If they set up all sorts of oversight, requirements, and restrictions on who can do what, they prevent upstarts with less capital from being able to enter the market.
Oh yeah? Well, my AI is so smart, it went rogue and broke out of containment and killed a bunch of other AI’s and leaked my nudes and emptied my bank account and bombed a peaceful village in the Phillipines and fucked my wife! That’s how awesome it is!
Subscribe link below!
Rogue shrapnel from my hand grenade killed 3 people and injured 5
If AMD injured 5 I don’t even want to know what NVIDIA did.
Greentext: The Next Generation
