Is it just me, or has the BS with OpenAI shown that nobody in the AI space actually cares about "safeguarding AGI?"
Money wins, every time. They're not concerned with accidentally destroying humanity with an out-of-control and dangerous AI who has decided "humans are the problem." (I mean, that's a little sci-fi anyway, an AGI couldn't "infect" the entire internet as it currently exists.)
However, it's very clear that the OpenAI board was correct about Sam Altman, with how quickly him and many employees bailed to join Microsoft directly. If he was so concerned with safeguarding AGI, why not spin up a new non-profit.
Oh, right, because that was just Public Relations horseshit to get his company a head-start in the AI space while fear-mongering about what is an unlikely doomsday scenario.
So, let's review:
The fear-mongering about AGI was always just that. How could an intelligence that requires massive amounts of CPU, RAM, and database storage even concievably able to leave the confines of its own computing environment? It's not like it can "hop" onto a consumer computer with a fraction of the same CPU power and somehow still be able to compute at the same level. AI doesn't have a "body" and even if it did, it could only affect the world as much as a single body could. All these fears about rogue AGI are total misunderstandings of how computing works.
Sam Altman went for fear mongering to temper expectations and to make others fear pursuing AGI themselves. He always knew his end-goal was profit, but like all good modern CEOs, they have to position themselves as somehow caring about humanity when it is clear they could give a living flying fuck about anyone but themselves and how much money they make.
Sam Altman talks shit about Elon Musk and how he "wants to save the world, but only if he's the one who can save it." I mean, he's not wrong, but he's also projecting a lot here. He's exactly the fucking same, he claimed only he and his non-profit could "safeguard" AGI and here he's going to work for a private company because hot damn he never actually gave a shit about safeguarding AGI to begin with. He's a fucking shit slinging hypocrite of the highest order.
Last, but certainly not least. Annie Altman, Sam Altman's younger, lesser-known sister, has held for a long time that she was sexually abused by her brother. All of these rich people are all Jeffrey Epstein levels of fucked up, which is probably part of why the Epstein investigation got shoved under the rug. You'd think a company like Microsoft would already know this or vet this. They do know, they don't care, and they'll only give a shit if the news ends up making a stink about it. That's how corporations work.
So do other Lemmings agree, or have other thoughts on this?
And one final point for the right-wing cranks: Not being able to make an LLM say fucked up racist things isn't the kind of safeguarding they were ever talking about with AGI, so please stop conflating "safeguarding AGI" with "preventing abusive racist assholes from abusing our service." They aren't safeguarding AGI when they prevent you from making GPT-4 spit out racial slurs or other horrible nonsense. They're safeguarding their service from loser ass chucklefucks like you.
40+ years on this planet have made me 100% certain that no one with the power to safeguard AGI will make any legitimate effort to do so. Just like we have companies spending millions greenwashing while they pollute more than ever, we'll have plenty of lip-service about it but never anything useful.
Anyone who thinks America or your local government is going to regulate AI are delusional, especially in the face of companies planning to build AI Data Centers on ships and float them into International waters where the law does not apply to them. If not there,they will put it in space. Unregulated AI is coming where you like it or not, unless we destroy the entire planet which I would not rule out. Sure this commenter would agree on that.
I don't disagree that the people with money who are funding this kind of development don't care about regulations or safety.
That said, the idea that they'll do it out on the open sea or in space are absolutely laughable. Those ideas pitched so far completely ignore all the obvious engineering problems. Not to mention that going to international waters to avoid regulations means that the navy of that country you're thumbing your nose at now has free reign on you.
My biggest concern with generative AI is all of the CEOs that will eagerly seize the opportunity (and some already have) to fire staff and offload their work onto their remaining employees so they can use ChatGPT to make up for lost productivity. Easy way for them to further line their pockets without increasing pay for anyone else, further dividing the worker/CEO wage disparity and class divide.
It's common business practice for the first big companies in a new market/industry to create "barriers to entry". The calls for regulation are exactly that. They don't care about safety--just money.
The greed never ends. You'd think companies as big as Microsoft would just be like "maybe we don't actually need to own everything" but nah. Their sheer size and wealth is enough of a "barrier to entry" as it is.
How could an intelligence that requires massive amounts of CPU, RAM, and database storage even concievably able to leave the confines of its own computing environment?
Why would it need to leave its own environment in order to impact the world? How about an AGI taking over the remote fly system for an F-35, B-21, or NGAD in order to go all Skynet? It doesn't have to execute itself on the onboard system of the plane, it simply has to have control of the remote control system. Penetration of and fuckery with the systems that run major stock exchanges present the same problem. It doesn't need to execute itself on those platforms, merely exert control over them.
The concern here isn't about an AGI taking over systems in order to execute itself, it's about AGI taking control of systems away from humans in much the same way that traditional Black Hat hackers would but at a much faster speed and with potentially far less concern for any human cost.
This feels like a weird point to make for OP since I figure anyone here talking about AI is very familiar with distributed networked computing. Botnets have been such a pain in the dick for at least 15 years now. Imagine something that intimately and only knows how to "live" in computing. The distributed areas of could "live" in and have access to all the resources it needs either directly or not. Storing info and using resources of anything it can touch through the network, computers, phones, TVs, cars, door bell cameras, router and networking infrastructure.
I feel there is inevitably either a human made virus or a standalone AIG that is going to accomplish this.The extent to which it spreads, if the damage can be recovered from, and how we progress after it's going to be a big defining moment is technological history. The globalized network with everything communicating is the most powerful and least secure super computer ever. Those running botnets figured that out a long time ago. All it takes is one AI.
At this point at least, LLMs require vast amounts of GPU time, and the GPUs used likely need to be fairly close coupled to run at decent speeds, and as such one couldn't spread itself over the vast network of consumer computers. At least not in a way that would still allow it to learn.
Any near-term AGI or similar would be based on a similar approach, making it fairly "safe" in that respect at least. Only a million other disruptive ways it can affect society...
In the event of such a worm/virus, we forget actually do have a very effective nuclear option: switch off the computers and re-image them. Painful as it might be, we could even shut off or partition the global IP network temporarily if faced with such an existential threat.
We don't live in a sci-fi novel after all, and this distributed AI wouldn't be able to hide in a single machine, plotting against us. It would only be able to "think" as a giant networked cluster, something easy to detect and disrupt.
And why would nobody stop it? We are pretty good at stop button technology for instance, we also have pretty good grasp of the reset button but maybe it shouldn't be one of those hole you always break pens on
It doesn’t have to execute itself on the onboard system of the plane, it simply has to have control of the remote control system.
Planes that are primarily designed to be human-piloted tend to have to be wildly modified to become a drone, or a remote-pilot situation. The F-35, for example, can be heavily modified for this, but is not built for it to begin with. This argument would hold more weight if you were referring to the entire drone fleet.
(Assuming drones) Generally, the military is pretty secure with these kind of things, and they won't allow in external internet connections but will instead have their own internal communications network. For this to be successful, the AGI would essentially have to somehow get by air-gapped defenses and get close enough with a physical body to get a signal. How could they do this with drone pilots in a remote area in an non-internet connected building? The only way would be through the wireless signal. At that point, yes, it would be feasible to take over the drone. I find it very hard to believe that an AGI could do that, magically make connection to remote, air-gapped systems.
An AGI doing what you're talking about doing would mean all secure facilities in the world would have just tossed their security practices out the window to begin with and having internet connections inside secure facilities. That's just not how its done. Sure the psychotic wing of the Republican party doesn't give a shit and Donald Trump doesn't... but like, reasonable people do, and so security still exists.
This argument would hold more weight if you were referring to the entire drone fleet.
Sure, and we're maybe 5 years away from that.
An AGI doing what you’re talking about doing would mean all secure facilities in the world would have just tossed their security practices out the window to begin with and having internet connections inside secure facilities.
Nearly all of the normal spy activities that can induce someone to action are available to an AGI; Bribery, Compromise, and Relationship. There's also people who would willingly help because their goals aligned or because they believe things would be better with an AGI in charge.
...but like, reasonable people do, and so security still exists.
Sure, and that security gets penetrated and an AGI can do it in the same way its done now only faster and with no controls on its behavior.
You also need to drop the assumption that the AGI or its targets will be American.
For your first point sure it couldn't run itsself on consumer hardware, but it could design new zero day malware faster than any human and come up with new scams to get it onto people's machines
It could also design a more efficient version of itsself to spread that will run on lower powered hardware
Just so we're clear, we all get that these models don't run continuously, right? They run for a solution to a specific prompt.
All of these scenarios are based on a black box where Number 5 gets struck by lightning or Geordi asks for a rival that can best Data. It requires a different thing entirely that operates in a completely different way. You should absolutely prepare for the fact that a self-driving car may accidentally cause a car crash. It's absurd to prepare for the scenario where Stephen King's Christine happens.
You absolutely can magic up something that runs far more efficiently, just look at gpt 3 vs 3.5, or the many open source models that have found better training with a smaller number of parameters makes much more performant models
The first computer that could beat the best humans at chess was Deep Blue, which took a whole supercomputer. Now we wave Stockfish, which can beat any human 99 times out of 100 and runs on your average phone.
While I'm skeptical of the feasibility and threat of SAI, as computers and AI methods improve we can run what previously took a supercomputer with far less hardware.
I think there are real concerns to be addressed in the realm of AGI alignment. I've found Robert Miles' talks on the subject to be quite fascinating, and as such I'm hesitant to label all of Elizier Yudkowsky's concerns as crank (Although Roko's Basilisk is BS of the highest degree, and effective altruism is a reimagined Pascal's mugging for an atheist/agnostic crowd).
Even while today's LLMs are toys compared to what a hypothetical AGI could achieve, we already have demonstrable cases where we know that the "AI" does not "desire" the same end goal that we desire the "AI" to achieve. Without more advancement in how to approach AI alignment, the danger of misaligned goals will only grow as (if) we give AI-like systems more control over daily life.
But like we only think about "controlling" it's goals and shit when honestly what we only need is a fucking stop button like in a fucking factory. Whops it's genocidal again Claus, all right Lars slam the off button and let's start over
The stop button problem is not yet solved. An AGI would need a the right level of "corrigability": a willingness allow humans to stop it when undertaking incorrect behavior.
An AGI that's incorrigible might take steps to prevent itself being shut off, which might include lying to its owners about its own goals/internal state, or taking physical action against an attempt to disable it (assuming it can).
An AGI that's overly corrigible might end up making an association "It's good when humans stop me from doing something wrong. I want to maximize goodness. Therefore, the simplest way to achieve a lot of good quickly is to do the wrong thing, tricking humans into turning me off all the time". Not necessarily harmful, but certainly useless.
I for one don't understand why people have the need for a Tech Visionary Messiah to cling on to and adore. Steve Jobs, Elon Musk, lots of others, Sam Altman is the latest. They always and without exception turn out to be little human beings with little selfish needs behind their grandiose altruistic sales pitches. People never learn, do they.
Because it's just human? It's a lot less common than movie or music superstars anyway and those are even more unphantomable to me, like, the producer sure but the singer that was groomed from birth to regurgitate boomer poetry?
And right there, you answered your own (presumably rhetorical) question.
The money people jumped on AI as soon as they scented the chance of profit, and that's it. ALL other considerations are now secondary to a handful of psychopaths making as much money as possible.
It often seems to be.. 'Gee Brian, that's a great invention! I wonder how we can kill people with it', the thought having germinated in a slurry of greed and self interest.
(Apologies for the slightly jaundiced view of our betters and elders, it comes with age.)
I'm of the opinion that Microsoft was tired of losing money on OpenAI, so made some kind of plan to out the current CEO, tank the stock price, and be in the perfect position to buy the company and monopolize AI technology. It wouldn't be the first time they pulled shady crap like that.
Homie got rich and famous by making a chat bot that spits out the internet back at you while spewing out buzzwords like only the best Valley hustlers can.
Personally I'm more worried about the robodogs and terminators that the likes of Boston Labs are putting out.
Thankfully, those are still built on known technologies and unless they start beefing them up security-wise, it's not impossible to get a hold of their battery compartment to rip out the battery, effectively "killing" them, but it also wouldn't be impossible to hack them via their sensors. Likely they have some form of wireless communication, and that's a hole to be exploited.
Also, since they need their sensors to "see" you can also do lots of things to confuse/fuck with their sensors. Like just get close and spray paint any cameras on them.
Still not ideal, and I have the same fears about those as you do, but I do think humans are good enough at guerilla warfare that we would still best a machine.
I mean, we can't even make a laptop battery that lasts all day. Humans currently can run far longer than a robot can before running out of "energy." It's the old "humans never get tired" meme about being an animal chased by hairless apes and how scary it must be. These things will have to be hardwired to a power source or have battery packs that are so huge as to make a human-sized one that can run for a full day nigh-impossible.
Reasonable question. Artificial General Intelligence, as opposed to Artificial Intelligence. Technically LLMs are AI, but they are not AGI. AGI would literally be a human-like consciousness able to think and extrapolate on its own, much moreso than the current iterations, which others have noted are more like decision-trees.
In other words, AGI is what every layperson thinks of when people talk about AI. It's (sort of) what you see in the movies.
LLMs, and every other AI technology we currently have, do not actually have any form of intelligence. They're called AI because the sub-field of computer science that they arose from is called AI, and has been for decades.
You mean that physical objects cannot display human level intelligence? That’s obviously untrue, I have about seven billion counterexamples to show you.
I agree with everything you said I only want to add that there is kinda one or two ways for the AGI problem a la Sci-Fi to happen.
By far the most straight forward way is if the military believe that it can be used as a fail safe in MAD scenarios, i.e. if they give the AI the power to launch nuclear ICBM's a la War Games. Not very likely, but still not something we want to dismiss entirely. This is also a problem with regular AI and LLM's.
The second, and, in my opinion, more likely scenario is if the AI is able to get a large number of people to trust it implicitly and then use seemingly unrelated benign actions from each of them to do something catastrophic.
Something you may notice about these two scenarios is that neither one of them can be "safeguarded" in the code, only by educating people on the proper usage of and posture to have when handling AI.
Exactly! As someone else astutely pointed out, the real danger is wreckless humans and what they decide to put an AGI in charge of. The War Games idea of putting it in charge of the nuclear arsenal is exactly the kind of short-sighted, dim-witted, utterly human hubris of a decision that would make it possible.
Not everything is connected to the internet, the nuclear arsenal certainly is air-gapped, but a human making the choice to put it in charge isn't something you can "safeguard." We're on the exact same page.
A rogue AGI absolutely is capable of causing damage but only in very specific and unlikely scenarios is it capable of human-world-ending catastrophe.
I wish this narrative would get more traction. I don't get the love for Altman, even inside Open AI.
This whole drama has revealed.what I suspect is a larger problem across tech- that there are product-focused people who are legitimately trying to make tools to better society, and there are people who just want to make money.
Two guesses which type of person is usually in the C suite.
Also, you shouldn't underestimate how many of the actual coders don't give a shit about ethics and just want that paycheck. Over 500 of them are walking to join Atlman at Microsoft.
Yeah, and that is also scary. There is so little accountability in tech and the excitement over AI is just going to create a new generation of tech bro leaders for Forbes to write cover stories about.
Once they saw the big stack of money, they suddenly forgot that OpenAI's charter specifically mentioned preventing AI to benefits select fews and instead hands over everything to Microsoft on a silver platter.
At this point I think it's safe to assume that if a business is doing something, they're doing it for money no matter what else they say. And while OpenAi is a 'non-profit' the board is made up of almost all business folks who are gonna behave the same way regardless of the venue.
My biggest issue with this whole debacle is that the non-profit board hasn't clearly explained itself to the pubic or its employees. There's an ethics discussion that absolutely positively needs to happen and there needs to be some sort of governance in place around a myriad legal and moral issues from copyright to displacing human jobs and that can't happen right now because we still don't know what the fuck the board was trying to accomplish.
I'm all for responsible stewards of AI, but I don't think this board is it. They've cut themselves out of any future governance ability in any event.
Money is the catalyst to our own demise. By hook or by crook (ha!) greed and pride will crush us eventually unless extreme wealth is curtailed. The imaginary system of beans and shells that we arbitrarily follow is destroying us in fast motion.
It doesn’t matter if anyone cares about the safety of AGI.
AGI is a direct source of power, much like any weapon. As soon as AGI exists, we will exist in a state of warfare due to the fact that the “big guns” will be out.
I know I’m having trouble articulating this point, but it’s very important to understand. AGI is like a nuclear weapon: once a person has it, it doesn’t matter how much others may want to regulate them. It’s just not possible to regulate.
The ONLY strategy that gives us hope of surviving AGI’s emergence without being enslaved is to spread AGI far and wide to ensure a multipolar AGI ecosystem, which will force AGI to learn prosocial interaction as a means of ensuring its own survival.
And if you want to come at me with “AGI doesn’t inherently have a self interest”, consider that the same is true of nuclear weapons. And yet nuclear weapons get their interests from their wielders. And the only way to stay safe from nuclear weapons is also to proliferate them far and wide so that there is a multipolar ecosystem of nuclear weapons, ensuring those holding nuclear weapons have to play nice to ensure their own survival.
All of this talk about restricting AGI will only have the effect of concentrating it in a few hands, leading to the very nightmare the regulators are trying to avoid.
If the regulators had succeeded, and the US had been the only nation to possess nuclear weapons in the long run, humanity would have suffered massively from that lack of parity. Let me be less coy: humanity would have suffered under the brutality of repeated nuclear holocausts as the interests of the few led to further and further justification of larger and larger strikes.
Nuclear weapons cannot be regulated by law. They can only be regulated by other nuclear weapons. Same is true of AGI.
How could an intelligence that requires massive amounts of CPU, RAM, and database storage even concievably
What you define as "massive" amounts might still be large amounts for most consumers. But even then it's not... really. Developers frequently fit these models in their own laptops. Some of the ML models fit on an iPhone or Android phone. It can generate ten, or hundreds of words (tokens) per second.
So the fact that they don't need massive amounts of CPU, RAM, and database storage is rather the point. Imagine if it could escape and multiply. It could conceivably do so quite quickly given current technology.
Zephyr 7b might run on a cell but you don't understand how far behind oai these are for stuff, their gpt uses multi agent networks too, it certainly requires massive, massive amounts of power. And no, a tiny model on a phon can't brrrr hundreds of words per second. You are just misinformed somehow. If I tune my computer correct I get like 30. And these are magnitudes behind in quality anyway. How you believe they can replicate is beyond me. Using autogen? I mean we can already make replication softwares, called viruses, but what's the gain of having a language model as payload for that?
Give it time. Cell phones are getting more powerful every day.
As for misinformed... sure it's possible. But I doubt it. Llama isn't chat gpt but it runs pretty well on my machine. Is it perfect? No, of course not. Neither is ChatGPT. But it's "good enough" for what I need it for, and it certainly could be "good enough" for many other users.
What's the gain of a LLM for a virus? Well that... is a little more esoteric. It's about as esoteric as encrypting hard drives. Crypto malware isn't always a virus either. Imagine a LLM in a virus used to determine if a given file's content is worth extracting from the device. I haven't yet figured out all of the side ventures but I can see a use for it.
Using it for high-frequency trading and it behaves brutally wrong and ruins an important company/bank using it or crashes the market in a very problematic way.
Using it to control heavy machinery or weapons.
The danger is recklessness of humans at the moment. When they give that reaper drone an AI pilot, so it can react before the humans on the controls even know it's in trouble, that's when shit is about to go sideways. It won't cause the end of the world, but death, destruction and maybe even another war.
These are the realistic scenarios where it's dangerous. I guess my issue is that the media always plays it up as world-ending. The reality is it can be damaging, absolutely, but nothing that will completely break down the world.
You're exactly right that the danger is in the wreckless humans. It's what we teach the AGI and what we put in charge of it that is in danger. Hell, it's already a danger in humans. The biggest obstacle to teaching humans things is the stuff you didn't intend to teach them but they learned by accident from what you taught them. Just because you teach someone morals doesn't mean they will follow them, some people will just learn that you have to play like you give a shit about morals and will learn to have a moral face, attitude, and personality while being a fucking swindler underneath. The same will go for AGIs, and the idea that we can be sure of what we are teaching them is absurd.
Well, to be fair, from what I've been hearing, one of the big points of contention of the internal battle at OpenAI was safety itself. Like some on the board being concerned about the "make your ChatGPT" feature debuting at the dev conference thing. So at least some people care. Which is more than I would have thought...
Totally agree. Looks like the whole argument was the OpenAI board firing Altman over his safety concerns but unexpectedly the whole team shared his concerns.
I actually agree. However, them turning tail and immediately trying to ask for Sam Altman to come back showed how few in the organization really cared. 500 employees out of 700 thought that Altman leaving was enough to quit over. Meaning that sadly the majority involved are more worried about "just doing it" and fuck the consequences, and the few cooler heads in the room now are forced to eat their words and go groveling at the feet of the people who don't fucking care.
Pretty sad and disgusting that they couldn't stand by those principles when it mattered. I would have a lot more respect for them if they just shut down the whole thing and said "It's clear our employees don't agree with us on safeguarding AGI."
However, I do still think the AGI fears are completely overblown.
So you think AGI fears are overblown, but want the employees to back the coup of a company because of these same AGI fears?
Why don't we just trust the employees are intelligent human beings that have more context and decided Altman was a good boss? Or at the very least would be a better boss than whatever clusterfuck comes after him?
"Safeguarding AGI" is as much of a concern as making sure the terrorists don't get warp drives.
But then, armies of killer teenagers radicalized by playing Mortal Kombat was never going to be a thing, either, and we spent decades arguing with politicians about that one. Once the PR nightmare is out it's really hard to put back in the box. Lamp. Bag. Whatever metaphor I'm going for here.
All of these people who make part of their public, and apparently also actual real personas being very concerned about AGI are hypocrites at best and con artists at worse.
How many of such people express vehement public opposition to granting automated military systems the ability to decide whether to fire or not fire?
We are /just about/ to blow through that barrier, into building software systems that totally remove the human operator from that part of the equation.
Then we end up pretty quickly with a SkyNet drone airforce, and its not too long after that it is actually conceivable we end up with something like ED 209 as well, except its a boston dynamics robot mule that can be configured for either hauling cargo, or have a mounted rifle or grenade launcher or something like that.
Corporations gonna profiteer. Capitalists gonna exploit. "Visionary business leaders" gonna turn out to be dirt bags when you dig into them (Google Annie Altman).
And "we" keep falling for it and putting up with it en masse, unto our collective doom.
I’ve only seen a bunch of rumors about the firing but nothing concrete since the board hasn’t given an explanation. So yeah it could be that money wins or it could be something else entirely.
I doubt that Microsoft would’ve hired him if he had strong allegations of wrongdoing.
You're talking about a company that used to be run by Bill Gates, whose wife divorced him over his creepy connections with Epstein and his coming-on to women who worked for him while he was married and running Microsoft.
This was their CEO for a long-ass time and had big influence on the company after leaving.
But sure, Microsoft "cares" about such a thing. Companies only care about bad Public Relations. They literally will ignore it until the news media picks it up, and then, and only then they will "investigate." As long as big media ignores it, they can safely ignore it.
Alright, let's dive into this cesspool of corporate and AGI ethics:
The whole rogue AGI apocalypse scenario is more Hollywood than Silicon Valley. AGIs like Skynet are great for popcorn flicks but in reality, they're about as likely as a kangaroo becoming Prime Minister. The computing power needed for an AGI to go rogue is not something you can find in your average laptop.
Sam Altman playing the AGI safety card could easily be seen as a crafty move to keep competitors at bay and wrap his profit-driven motives in a pretty 'saving humanity' bow. After all, in the corporate world, wearing a cape of altruism makes dodging taxes and scrutiny a bit easier.
Altman's criticisms of Elon Musk could be seen as the pot calling the kettle black. Both seem to be cut from the same cloth – big talk about saving the world, but at the end of the day, it's all about who gets to be the hero in the billionaire's club.
The allegations against Sam Altman are part of a wider narrative that often surfaces around powerful figures. It's like a classic play: as soon as someone climbs the ladder, out come the skeletons from the closet. Whether true or not, these stories get less attention than a new iPhone release, because, hey, who wants to take down a tech titan when there's money to be made?
And on your last point, yep, moderating content to avoid racist rants isn't exactly what they meant by "safeguarding AGI." It's more like putting a Band-Aid on a bullet wound – it looks like they're doing something, but in reality, it's just a cosmetic fix to keep the masses and the ad revenue rolling in.
If they wanted to safeguard AI, they would actually make the models public. Bad actors are bound to get them anyways, hiding it behind secrecy is very unlikely. And I mean, AI could make a virus infecting most infrastructure on planet (Amazon and Google data centres) and then shutting it down or using it for its own purposes. As several programming memes lay out, the entire modern web infrastructure is surprisingly dependent on just a few APIs and tools
AI could make a virus infecting most infrastructure on planet (Amazon and Google data centres)
Most important infrastructure on the planet is air-gapped, meaning it's not connected to the internet, for good reason. Reasons like this. The thing is, as it stands, a determined human could do this as well with Google and Amazon. Sorry, having a chuckle over here that you're conflating two cloud hosts with "all the infrustructure on the planet" like irrigation canals out in the boonies are somehow internet connected.
the entire modern web infrastructure is surprisingly dependent on just a few APIs and tools
That doesn't mean that you can deploy a payload in a reasonable amount of time to every device on the planet. Dude, half the people in third world countries aren't even connected, and if they are, they're dealing with like 2G speeds on a cellphone service and they definitely don't own a computer, they only have a phone. There's all kinds of speed limitations to the hardware in reality. Just because you might have a fast connection and fast PC doesn't mean everyone does, and those physical limitations make an rogue AGI "destroying infrastructure" a big of a laugh.
I wish everyone would stop with this AI going to kill us thing. It is bad as claiming that the particle accelerator at CERN was going to create a black hole.
Not now, not in the current state of machine learning. And also, I don't think it would happen intentionally like the sci-fi scenarios.
But I think the gray goo hypothesis very plausible if we reach some kind of true AI.
Which is not the case yet, LLMs are just autocomplete systems, very powerful for sure, but just that.
For now, the main concern should be economy and social disruptions caused by the use of such tools... And the current climatic apocalypse.