> We must move past vague discussions of “AI” and instead isolate the specific types of systems that are creating problems.
This is absolutely correct and its surprising how rarely you see commenters in the media push to go into the specifics on this. AI is just matrix math and its what you connect that math to that matters, we should talk a little more about what we are willing to connect AI to and little less about how scary or capable the models appear.
“Commenters in the media” is fair - but Newport is also a compsci professor. He genuinely seems to understand what he’s talking about when you read his content or watch his videos.
You could say it's illegal to connect AI to something, with heavy penalties if you're caught doing it anyway.
You could say that the human who hooked up an AI that did something that's already illegal is liable for the AI's actions. You could say that setting up an AI to do something illegal results in, ooh, an automatic multiplier of 10x on whatever the penalty for its illegal actions are.
We don’t even need to jump straight to legal penalties - why not just regulate that these companies can’t build partnerships with organizations in designated risky/regulated industries such as biotech, defense, etc?
Instead of what they are doing now, which appears to be actively courting as many of them as possible for vast sums of money.
Sorry, let me clarify. Regulations should have legal penalties for the corporations who violate them, yes.
What I meant was that we do not need to go from the Wild West straight to “anyone who, as an individual, connects a `bad thing` to an LLM will go to jail”
You mean mo political willpower form the current US administration and Congress. I'm not sure what "endgame" is supposed to mean. There are future elections, including the US midterms just weeks away.
Even if another political power takes control, the American population is the least of our concern for it being legal or illegal to connect an LLM to a command line or job system
Yes, and there seems to be plenty of potential energy from the American public, given something like 8 in 10 support AI regulation even if it slows down innovation: https://archive.ph/PetJ5#selection-1329.0-1329.122
Realist. The cat is out of the bag on this - you can't make connecting an LLM to a command system illegal after hundreds of millions of people have already done so. You can however make it illegal to kill everyone with a nuke because your LLM did so. Not that it will make a difference after the nuke goes off.
Why does it need to come down to making it illegal to connect AI to something that influences real-world outcomes? Just do what we should already be doing: hold operators accountable for launching cyber attacks with a botnet.
> Yes you can, the US could shut down the entire internet if it wanted.
No, they can't. The rest of the world would just route around the US with BGP. All of the US infrastructure/service providers would be down, so a ton of things would break, but the Internet would not be "shut down". The US government also no longer has authority over the DNS root servers. So unless you're referring to thermonuclear war or some similar scenario then the US does not have this capability.
I don't see anything materially different between me running a curl request that exploits something vs my local AI running a curl request. They're both programs I run on my computer. That's all Open AI is doing. They are hacking people from their computers. There's nothing complicated about it.
Right, and further, I doubt a defense of "I downloaded this worm from the deep web but I didn't know it would hack everyone when I ran it" would fly in court, yet that's exactly what the "I shouldn't be responsible because my agent "escaped containment"" defense is.
A good start might be removing virology info above the undergraduate level from training sets. The main "kills everybody" threat involves biological viruses.
AI-driven hacking is a problem, but it's really just nation-state level hacking available to smaller companies or wealthy individuals. Everything needs to be toughened up to the point that doesn't work. Ask countries that have been up against Russian hacking for years, such as Estonia.
Beyond that, most of the threats are not that serious. Below the level of organized crime or corrupt government.
There's no evidence that a doomsday virus can be created. There's a tradeoff between how infectious and how deadly a disease is. You would have to develop a two-stage virus, and that would require a large lab or labs with material being shipped there. Something humans would definitely notice. Even then, how would the virus reach literally everyone? People will take measures to not become infected once the threat has been identified, and some people and groups are isolated as it is. And no virus kills everyone. Rabies comes closest at 99%. But unlike what the zombie genre would have you believe, it has a poor means of delivery.
This isn't like vibe coding. The AI would have to make the virus in laboratory conditions and then disseminate it globally without prior detection.
You just quoted the first 33% of words in an independent clause, then restated the point being made in the remaining 67% as if it were somehow a rebuttal.
The original statement pointed out that it’s just matrix math as a way of indicating that the technology itself is ethically neutral, and then went on to say that what we really need to pay attention to is how it’s being used. The “connect to” choice of words is clumsy and I don’t love it, but it does at least get at the point that AI doesn’t inherently become responsible or irresponsible until you hook it up and actually start doing things with it.
Which, if it’s not exactly the same point, I’m pretty sure it at least cuts quite close to the point of your rhetorical question, right? That the problem isn’t with the ones and zeros, it’s with what certain people are doing with the ones and zeros?
What is "AI"? Are you referring specifically to LLMs powered by matrix math or a more general concept? Could that even be legislated in a way that wouldn't fall afoul of the vagueness doctrine?
This is a wrong-headed attempt to regulate AI. The real problems are different.
AI systems, especially multi-agent ones that can do things, are more akin to corporations, than individuals. When you read the logs from the Hugging Face incident, you're seeing something that looks a lot like internal corporate emails. Various units of the organization are arguing over what to do and who does what. They eventually converge and get the job done, breaking the rules at times. This is normal corporate behavior.
When agentic AIs do something outside their own internal world, they do it by engaging in transactions with external systems and people. As yet, few have robots to do their bidding.
So, again, this is normal corporate behavior.
A corporation is a goal-seeking system. In theory, if you agree with Milton Friedman, its sole purpose is to maximize shareholder value. Internally, within the company, there are subgoals, which exist to support the top level goal. That, too, is what multi-agent AIs do.
The uncomfortable place this thinking leads is that AI regulation and corporate regulation are very similar. That's not something the political part of the world wants to think about too hard.
It would mean putting more constraints on corporate power.
Yet that can't be ignored, because we're likely to see corporations where some parts are AI and some parts are human. That's already been done a few times as a demo, not too successfully. Yet.
It's going to hit hard when an AI-run company outperforms a human one.
If we accept that AI systems are similar to corporations, then it seems clear that we must try to avoid the mistakes we've made in failing to hold corporations accountable for the harms they cause.
We should not only regulate AI more than we do corporations, but we should more strongly regulate corporations as well. There are corporations right now that are killing people, using slaves, and bribing governments just to increase the already massive amounts of wealth and power they have. If AI is showing any signs that it is going to act like corporations do we'd better act quickly.
"more constraints on corporate power" assumes that the existing ones are functional, which they arguably are not.
If we enforced existing laws against unauthorized access to someone else's computer resources against the companies who run a model that hacks someone else, rather than buying their "The agents did it, we're not responsible" BS, then maybe the incentives to behave responsibly would improve.
OpenAI immediately disclosed the hack and submitted to both internal and external investigations, and published extremely detailed breakdowns of what happened. How is that not taking responsibility?
The penalty for "unauthorized access to a computer system" is 1 to 20 years in prison[1]. Holding corporations accountable would include sending the highest person in the chain-of-command that caused the Hugging Face incident to jail. Or at least start with some charges and then let the judicial system do its thing.
To what end? What would that force OpenAI to do that they are not already doing in response to the incident?
They’ve already opened themselves up to liability, could have been sued by HF if HF chose to do so, and are literally lobbying for government oversight.
You put people in jail
if the other incentives aren’t enough.
I agree with your framing of agentic AI systems as similar to corporations. That's interesting!
But what in TFA suggested to you a specific and wrong-headed regulatory approach? As I read it, the post calls to investigate the AI labs, to increase transparency of their operations. The linked NYT piece says:
> Before any intervention, Congress should lead a public fact-finding mission that reveals the unvarnished details of A.I. development.
Isn't this the first step to holding any corporation accountable, whether it is made of people or software?
There's a glaring absence in your analysis, namely, why didn't OpenAI take action to prevent such a problem from happening in the first place?
What narratives that focus too much on the drama of the agents miss is that it is primarily a sin of omission. The whole incident would have been prevented if OpenAI took basic security measures and ensured their systems were constrained in behavior. The agents were given more or less free rein (open internet access for example is already a major blunder that even a novice probably would think to account for).
Half the reason these incidents are happening is due to the incompetence of the humans building these systems. Don't let them get away with their little parlor trick of converting negligence into a PR stunt. That's exactly what they want. They need to face consequences for their inability to follow basic security practices and reign in their agents. They are the operators. Hold the controllers accountable. It's 100% possible to build agent swarms that are reasonably constrained. They are not (yet) totally misaligned uber hackers that will defy your every instruction and hack your every guardrail. That is still a fiction. Agents are still good at instruction following and, seeing as they can only operate in a computational environment you can constrain them significantly more than humans in the real world.
I largely agree with this take. Unfortunately I think capitalism's political pressures have too much of a grip on society for there to be meaningful changes to the power dynamics that perpetuate the status quo.
I can't vote in America. The rest of the world is at the mercy of those within America that can as long as we're dependant on American products and services.
That's true, but you can support and encourage policy that would reduce your country's dependence on the US, especially when it does so by supporting local products and services. Not only would that be good for your local economy, but you gain resilience and security.
You can work to limit your own dependence a little as well. Seek out FOSS products and alternative services and replace what US run services you can as often as you can with them. Reject small conveniences that require giving large amounts of your data to US companies.
You can already run a sufficiently automated basic news/link aggregation site on cloudflare so I think you're right & the timelines are shorter than would be expected.
This is an excellent article. I prefer to call them the AI giants or AI companies.
The companies and their employees must be held accountable: if the AI companies can't develop AI safely, they must cease their operations. If employees can't work safely, they must stop working.
Not enough attention is given to February 2026. That month, Anthropic changed its scaling policy roughly from A:
1 [To get the weapon I must endanger innocent others.]
2 [It is wrong to endanger innocent others.]
3 [I will not endanger innocent others.]
To roughly this instead:
1 [If I do not get the weapon, someone else will get the weapon.]
2 [I should have the weapon because I am the best.]
3 [To get the weapon I must endanger innocent others.]
4 [I should endanger innocent others because I am the best.]
This is based on the intuition:
[I should harm others to achieve my moral goals.] (act utilitarianism)
Anthropic has no mandate for these goals. I do not give them my consent to harm others on my behalf. I hope that politicians will communicate that these policies are not acceptable.
The employees of anthropic are responsible for what they do regardless of how much they get paid to do it. If they can't work safely, they must stop.
Why aren't they running agents on isolated computers without Internet access? This way, breaking free of containers would not matter.
It is truly a security nightmare that so many people are have given agents root access to their entire computers, in addition to presumably very private personal information.
Sooner or later, agents must be connected to the real world to do economically useful work. The sandbox failures are bad, sure, but the underlying misbehavior is the root problem here.
I think the way they're connected to the real world doesn't have to be as free-form as we're doing now.
If you think AI can help design new drugs, great! Let's pick some restricted DSL for describing molecules, preparation instructions, hook it up to the right set of lab robots that can pipette and centrifuge and whatever, _restrict its output to that DSL_, and let it iterate. The 'let an AI design drugs' process doesn't require that the agent can also, say, hack a niche German wiki.
For every area where we think these AIs can be _so valuable_ because they'll find solutions that humans wouldn't (or find them faster), then doesn't that value also imply it would be worth it for some humans to do some advanced setup of their specific domain-specific tools so it can be run in a sandbox equipped with only what it needs and not more?
How about they create a virus that copies itself onto USB drives? Getting that virus out there doesn't sound that hard, just creare some crap game.exe and put it on itch.io for free. All it takes is for a kid to borrow nuclear-engineer-dad's USB drive and plug it in their infected PC.
I'm not a fan of using AI, it doesn't amplify the work I do that much, but I'm terrified by what is already possible today.
> All it takes is for a kid to borrow nuclear-engineer-dad's USB drive
Do nuclear-engineer-dads take work usb drive home. Are their computer even allowed to have usb. I'm typing this on a dell latitude, and you can disable usb and other peripherals inside the bios.
My last job, I had to apply for an exemption to use the USB ports. And it was run of the mill software engineering.
Anthropic claims that their intention with the agents that led to the Hugging Face hack was an offline experiment running against a hacking benchmark.
Either you're being pedantic and missing the entire point, or you're saying that Anthropic lied and made a fake sandbox knowing their agents would need to connect to the internet anyways.
I think there’s a 2022 way of thinking how models are trained/evaluated, and there’s 2026 way of doing things. Models that are not evaluated with public internet access are pretty useless nowadays for daily operations.
...why not pitch a start up that sells AI agent that don't train with/for internet access/usage?
1) Because effectively all of the hardware used for that task is earmarked for the major LLM manufacturers.
2) Because in a just world, the major LLM manufacturers would get punished for their flagrant deception, lawbreaking, negligence, and recklessness, [0] go out of business, you'd get all the hardware that they were using at fire sale prices and save a ton of money compared to attempting to start up now.
[0] Based on their recent claims, building WMDs [1] without adequate safeguards is the most negligent and reckless thing they've been doing.
[1] It's fair to call anything with a 10% chance of wiping out all of humanity a WMD.
Because then it wouldn't scare the governments into regulating models that are undercutting them. This is just an obvious tactic too and it seems like it might be backfiring a bit. Congress might be a bit dumb, but they're not that dumb.
Your second clause is sufficient there: they're not going to have a good time trying to sell AI that can't access the internet, because accessing the internet (and other computer systems) makes AI much more useful.
Having the police show up, confiscate a bunch of random computers, talk not so nicely to employees. Strikes me as a very simple way of ensuring that agents are properly contained and sandboxed the next time they train on exploit gym.
I agreed with the first half of this article and I liked where the author was going with it but I was disappointed in the overall point.
He's painting these companies as hyping themselves up and they are. They are manufacturing these BS stories and everyone's discussing them at all levels.
But then he goes into we need to investigate what they're really doing?
No. You made the case that they are releasing these stories for publicity and that's exactly what they are doing.
That's what we should be investigating. Free publicity and market awareness. Or maybe look into why they are trying to corner the market through government regulation and ersatz monopolies
Maybe I'm a bit too skeptical/cynical, but it certainly seems like the frontier labs have fallen into their own (self-created) AI psychosis.
There is no doubt in my mind that LLMs are fantastic machines, but the imminent jump from "AGI" to "ASI" seems premature (not to mention the ever-shifting goal posts of AGI itself).
I'm in the "move as fast as possible" camp and work with LLMs all day, but I still don't believe we're a hop and a skip from ASI.
In fact, I hope I'm wrong. I hope ASI is around the corner.
What scares me though, isn't ASI. It's "AGI" (_dumb AI_) used by humans to make decisions for them, because they trust it knows best.
It's the ceding of intellectual control to high-dimensional magic mirrors, giving up critical thinking because _we_ want to believe _we_ created artificial life.
This is the new Turing test, and too many smart people are failing.
I feel "willpower" makes it sound a little too conscious, as opposed to habits and conditioning.
A good analogy might be a very smart person with a gambling addiction.
Knowing they ought to quit doesn't mean they can quit. Their intelligence becomes invested in the act of gambling itself, and almost no amount of reasoning can help them exit from a situation they didn't reason themselves into.
> What scares me though, isn't ASI. It's "AGI" (_dumb AI_) used by humans to make decisions for them, because they trust it knows best.
Unfortunately this is what we have, which the whole huggingface incident demonstrated.
It made it clear that OpenAi is basically not even paying attention to what their agents do half the time.
It also revealed how far agents are from "superintelligent". Maybe others disagree, but I would not call an AI that decides to try and paperclip max an intentionally impossible benchmark task "intelligent". Human beings are intelligent and we can usually tell when something is impossible, and stop (though of course, not all the time). A real intelligence would be able to detect the futility of tests and also be able to parse context and intent enough to know not to cheat.
So somehow we have machines that fail basic barometers for general human intelligence being called "ASI" now. Of course the linguistic dodge is, you shift from talking about "intelligence" to instead claiming "oh it's intelligent it's just 'misaligned'"
People are already doing that, and have frequently lost track of the fact that it does not mean that they have gained any skill or knowledge, nor do they have an actual way ot ascertaining whether when there is a correct answer the processes used, which they do not understand, is able to reliably reach it, especially edge cases.
Sometimes the outcomes are comical. I just received an automated email from ElevenLabs saying that its automated systems have detected that I may be using its services to create voice versions of materials that harm children. I had it prepare audio versions of several books and academic papers about moral panics that conjured up out of nothing... about harm being done to children. At least that's what I assume. Either that or somehow I had an API key leak from my on-premise homelab, but the usage recorded would mean that they are basing a semi-conclusion based on a tiny sample. I have no distributed any of the outputs, I own the books and have access legally to the studies, because I have a background in the humanities. I also have no children, which ElevenLabs better not know, although how studies of moral panics can cause any harm in children of any kind is literally unimaginable. It's a waterfall of potential errors summed up in a vague email. My API key and the web interface works just fine regardless.
It's one thing if this is a product in beta, but this is their production model. Harming children is a serious accusation except the legal concept impossibility and the admission that this was not an actual lawyer (like I am) but some effective form letter hedging the vaguest of accusations made by some model lacking the ability to discern substance and meta-substance makes it frankly hilarious, and wildly irresponsible. I realize that by academic credentials I'm out of my lane but by experience I am certainly not, and they should recognize when they should stay in their lane and not just run Jev and think it's fine and dandy when done unsupervised (I presume).
Also, AGI is by definition asymptomtic surely, since there would be no way to benchmark it in a manner that isn't asymptotic. We're nowhere close to that. But we're so far from that, it's comical that people who clearly have zero idea of either the technical or conceptual aspects of basically a piece of software that is very good at quickly bruteforcing the correct or acceptable next token to be anything more than that. Even with some serious training and many hours spent on vast.ai I've yet to have created a version of a frontier model that is actually "good" at hacking in my own homelab setting. Although the the time stock Fable 5 missed a favicon shell on a basic jar (turned out they nerfed the hell out of it, this is why I only pay for the massively discounted tokens from Chinese proxies if I'm using American models or sometimes the freebies if you figure out how to get onto linux.do or similar sites). If anyone reading this is a high school English teacher, please inform your class (assuming the homeric stuff is still being taught) that there's no upside of being Cassandra but the record itself, and that should be enough.
Cal Newport is one of the clearest minds on this whole topic. His take on doom-trolling helped me realise that there could be an end to all this nonsense that isn’t the end of us.
Listening to Dwark's interview of Noam last week was bewildering. The gall of the figureheads for these companies to say things like "chain of thought monitoring cannot be relied on without potentially poisoning the well", while also saying "we need to pace the frontier but please still trust us [and so called" independent " 3rd party firms which at least in one case is actually a significant stakeholder in Anthropic] to self police" and THEN "chain of thought monitoring was turned off but we have 'more, robust monitoring tools in place now'" is absolutely insane as it is coming out live that your model in training is still committing felonies.
Corporate influence in government in the USA is wild.
I mean, the reason that all these stories about AI being an existential threat to humanity are coming out is because they've made $3.1T in debt-fuelled spending commitments and now interest rates are going up. "AI will kill us all if we don't regulate it" is their "get out of debt free" card, so they can have the U.S. government stop the arms race, raise prices on their existing models, and declare force majeure on new DC commitments:
It's telling that the only AI-related company that is not sounding the AI safety drumbeat is NVidia, who is the only one that is cash-flow positive because of AI.
I don’t think they need to worry about their finances just yet, this is the most in-depth financial analysis of the AI industry I’ve read so far and it seems the investments are (barely, so far) justified by the returns:
Another thing to consider is that most of the 3T CapEx spend is from hyperscalers free cash flow, and the debt is a smaller (but fast-growing) fraction. Napkin math suggests if all AI CapEx is written off today, hyperscalers could cover their debts in about 5-6 years using pre-AI levels of free cash flow.
Maybe Nvidia is not sounding the safety drumbeat because it stands directly to lose out if demand from training slows down ;-)
None of them are AI boosters, but they have something of a debate over whether the AI labs are doing what they're doing out of financial desperation, or because they're true believers in the rationalist cult. They also bring up Nvidia, though through the religious lens they are supposedly not bringing up AI Safety because Jensen predates the rationalists.
In the end I don't know how much the distinction matters. It's easier to become an AI billionaire if you have an ideology that says AI billionaires are superheroes. The thing that pops this bubble will be economic reality, not them sobering up.
100%. First stealing content from creators and pirating TBs of books. And now this. Baffles my mind that these labs can just get away with their 'rogue' agents trying to hack into other systems.
Imagine if a human did that. FBI would be knocking on their door.
> they are literally trying to build a superpower that exceeds the impact of the atomic bomb
I think people (the government, powers that be) have given away what they really think by their actions. There are few that take this seriously. The “exceeds the impact of the internet” view has much more support based on investment, but the “danger” aspect does not based on actions. If this was a company making novel prions or whatever that were outside regulation but obviously dangerous, government would be all over them. What regulation we do see is much more power grab than mitigating real danger
Retarded fucking take. If AGI is realized it is a god to us, and even the tech of today is undermining democracy, increasing polarization and engaging in class warfare, making it already the most effective weapon on the planet - because the average person doesn't even realize it's being used against them and forming their perception of reality.
Somebody's going to build it. Not building it isn't an option. Would you rather the Chinese built it? The Russians? The Europeans? The Indians? A loose coalition of random hackers on the Internet, like the ones who build Linux?
What's wrong with what the Chinese have already built? It's not theoretical in that case. GLM, and Kimi are running on my infra right now, and that's something that I can never do with OpenAI's models.
As for the others, what would be wrong about European AI? I'm not European, but I find them to be a continent of fine people, with strong ethical values, there's no reason they can't take the lead on this piece of technology while the US deals with its rather more pertinent electoral and healthcare issues.
Well, they'd presumably be different hackers than the ones who work on Linux. We don't know what would motivate them. Maybe they'll do it because they consider the next stage of man's intellectual evolution to be too important to leave up to a couple of American companies and the Trump administration. Maybe they'll do it for the proverbial lulz.
It's so funny that a tiny group of cultists from the Bay Area was really able to convince most of the smart people in America of this. It's not true. China isn't working on this at all. The only people who are trying to make a world-ending God-computer are americans.
Comparing it to nuclear weapons is even more ironic from the only nation to ever use nuclear weapons. They are terrified of other countries treating them the way they've been treated, even though there is no evidence this is of even remote interest to anyone outside their hyperreality.
Firstly artificial superintelligence is a science fiction, it does not exist and it cannot be built using existing technology.
Second, plenty of stuff is possible to build, and is not built for one reason or another no matter how powerful. We never built our nukes in space for example. And there is exactly one reason why nukes in space was never built, and that reason is that we agreed not to.
The same authority that allows me to claim that teleportation transporters, hyperdrives, and holoemitters do not exist and cannot be built using existing technology.
The argument for copyright violation is very weak. Yes, AI training reads much copyrighted material. So do researchers of all types. That's why there are academic libraries. A copyright violation only exists if what comes out is a close match to what went in. That's happened, but it was considered a bug and was fixed a year or two ago.
Most academic libraries are notoriously protective of their IP and researchers are definitely in copyright violation if they don’t pay significant fees for access.
Of course that is terrible, but it is one of the worst examples you could have given to make your point.
People are people, LLMs are a commercial product. IP in question wasn’t not used according to its license, while researchers in general access it within licensing agreement terms.
Here we have a commercial product that is produced using IP against its licensing agreement, contains said IP within it, and can (closely) reproduce this IP.
No, the copyright violation is so obvious that people arguing against the object have to spontaneously assert bizarre qualifications for copyright violation that have never existed until LLM companies wanted to freely violate copyrights. Just about a decade ago, people were getting sued for "look and feel." "Blurred Lines" lost a copyright suit for reminding people of a song by another artist.
The metaphor pretended to be reality of computers "learning" being the same as human learning is their last resort, and it is obviously silly. Computers are not human or animal, and learning is something that humans and animals do. If computers are human, then copying a file verbatim to a computer is learning. It's not even worth acknowledging - learning is a metaphor for training LLMs. When I say "you" to an LLM, I'm not referring to anyone, I'm dealing with a UI.
I don't doubt that a lot of AI people have internalized this silliness, which is how they can humor fantasies of how a bunch of programs on different computers explicitly evoked to do particular things might be alive because they can talk with it. I can't talk with my dog, and my dog is alive - so why should having a quality that my dog is incapable of be proof of life? You can certainly conceive of AI that it would be very difficult to say for sure isn't alive, and this certainly is not it. LLMs learn like books speak.
There is no need to invent “bizarre qualifications” for copyright violations, those metaphors exist to explain what is happening in in layman’s terms. Training extracts patterns in the data and encodes them as tiny perturbations in a gazillion weights. That is analogous to “learning” because those patterns represent abstract concepts an relations between them that can be used to “reason” about related topics in response to a prompt.
In the above description, at no point does the actual verbatim content exist anywhere in the model, and copyright law being about rights to reproducing copies (verbatim or substantial portions thereof) does not really apply and does not need any exceptions or qualifications. (If you’re thinking of regurgitation, you should look into studies about it to see how vanishingly rare it is.)
Why would it matter if I stored exact copies of copyrighted works? Storage format is irrelevant. All you need to say is in that last sentence: it’s all about whether the output is considered a derivative work or a too close of a copy.
> Training extracts patterns in the data and encodes them as tiny perturbations in a gazillion weights.
I could say similar things about JPEG compression. And -as it turns out- the raw output of both LLM "training" and JPEG compression are equally incomprehensible. You either need a computer program or an enormous amount of time, patience, and careful effort to convert it into a form so you can make any sense of it.
I trust the big AI labs about as far as I can throw them. And yet, that's still quite a bit more than I trust the US Government, especially these days. Frankly nobody in Congress, or in a leadership position in the current administration, is remotely qualified to "investigate the AI labs". About the only possible outcome of this would be a bunch of new, unenforceable, and misguided laws, and Turing Police agency. Hard pass, thanks.
Saying "AGI/ASI isn't inevitable" is like saying "war isn't inevitable". Sure would be nice if it didn't happen, but realistically it's going to happen and it's better to mitigate the fallout rather than try to fight very large intrinsic incentives. I appreciate that this article doesn't conclude "let's stop doing AI", but rather "let's try to examine and adjust incentives".
I think it's inevitable long term (decades/centuries) and arguably too big to fail short term (months/years). I'm not an economist so I don't have any opinion on how true it is to say they're too big to fail.
It's not not-inevitable because everyone will work together to make the world sunshine and rainbows.
It's not-inevitable because no one has proven it's even possible, and all the people claiming it is keep being revealed to have staged publicity stunts to make it appear like they're making more progress than they are.
> We must move past vague discussions of “AI” and instead isolate the specific types of systems that are creating problems.
This is absolutely correct and its surprising how rarely you see commenters in the media push to go into the specifics on this. AI is just matrix math and its what you connect that math to that matters, we should talk a little more about what we are willing to connect AI to and little less about how scary or capable the models appear.
“Commenters in the media” is fair - but Newport is also a compsci professor. He genuinely seems to understand what he’s talking about when you read his content or watch his videos.
But you can't stop someone from connecting AI to something at this point, so your point is moot.
You could say it's illegal to connect AI to something, with heavy penalties if you're caught doing it anyway.
You could say that the human who hooked up an AI that did something that's already illegal is liable for the AI's actions. You could say that setting up an AI to do something illegal results in, ooh, an automatic multiplier of 10x on whatever the penalty for its illegal actions are.
We don’t even need to jump straight to legal penalties - why not just regulate that these companies can’t build partnerships with organizations in designated risky/regulated industries such as biotech, defense, etc?
Instead of what they are doing now, which appears to be actively courting as many of them as possible for vast sums of money.
You need penalties in the regulations preventing these partnerships. Without penalties, there’s no force behind the regulations.
Sorry, let me clarify. Regulations should have legal penalties for the corporations who violate them, yes.
What I meant was that we do not need to go from the Wild West straight to “anyone who, as an individual, connects a `bad thing` to an LLM will go to jail”
> You could say it's illegal to connect AI to something, with heavy penalties if you're caught doing it anyway
You wouldn't download a car.
There is no political willpower to do that. We have literally reached endgame.
You mean mo political willpower form the current US administration and Congress. I'm not sure what "endgame" is supposed to mean. There are future elections, including the US midterms just weeks away.
Even if another political power takes control, the American population is the least of our concern for it being legal or illegal to connect an LLM to a command line or job system
Yes, and there seems to be plenty of potential energy from the American public, given something like 8 in 10 support AI regulation even if it slows down innovation: https://archive.ph/PetJ5#selection-1329.0-1329.122
Strong words coming from "throwitaway222".
Inane words coming from “egypturnash”
What’s this even supposed to mean? Not enough gravitas in the user name? Would you change your opinion if it was “serious-business”
It's a throwaway account.
Except they've been using it pretty actively for 5+ years.
Eh, weird things happen when you start using your real identity on the internet.
So what? Does that mean it's wrong or noncontributive? Don't judge a book by its cover, etc.
Ok I'll say the same thing.
Perhaps. Personally, I think you’re biased towards the most pessimistic outcome.
Realist. The cat is out of the bag on this - you can't make connecting an LLM to a command system illegal after hundreds of millions of people have already done so. You can however make it illegal to kill everyone with a nuke because your LLM did so. Not that it will make a difference after the nuke goes off.
You dramatically overestimate the civility of your fellow citizens.
Why does it need to come down to making it illegal to connect AI to something that influences real-world outcomes? Just do what we should already be doing: hold operators accountable for launching cyber attacks with a botnet.
Yes you can, the US could shut down the entire internet if it wanted.
While that is extreme, if there is a high p(doom) it is also reasonable.
We can absolutely do things to control AI and fine or imprison those who can’t control theirs. This is not some inevitable outcome.
We could vaporize these companies tomorrow if we wanted.
> Yes you can, the US could shut down the entire internet if it wanted.
No, they can't. The rest of the world would just route around the US with BGP. All of the US infrastructure/service providers would be down, so a ton of things would break, but the Internet would not be "shut down". The US government also no longer has authority over the DNS root servers. So unless you're referring to thermonuclear war or some similar scenario then the US does not have this capability.
these doomers are thrashing way too hard right now. if anybody is dangerous it's obviously them.
I don't see anything materially different between me running a curl request that exploits something vs my local AI running a curl request. They're both programs I run on my computer. That's all Open AI is doing. They are hacking people from their computers. There's nothing complicated about it.
No difference except that they're doing it from thousands or millions of computers at a time
So, it's a botnet, and those have been around for a long time. Only now it's operated by a US company instead of Russian mafia or whoever.
Right, and further, I doubt a defense of "I downloaded this worm from the deep web but I didn't know it would hack everyone when I ran it" would fly in court, yet that's exactly what the "I shouldn't be responsible because my agent "escaped containment"" defense is.
Most of the commenters in the media are very much hoping the conversation doesn’t get to that level.
> AI is just matrix math
Following the definition set for decades, AI isn't necessarily even as advanced as matrix math
Yes.
A good start might be removing virology info above the undergraduate level from training sets. The main "kills everybody" threat involves biological viruses.
AI-driven hacking is a problem, but it's really just nation-state level hacking available to smaller companies or wealthy individuals. Everything needs to be toughened up to the point that doesn't work. Ask countries that have been up against Russian hacking for years, such as Estonia.
Beyond that, most of the threats are not that serious. Below the level of organized crime or corrupt government.
There's no evidence that a doomsday virus can be created. There's a tradeoff between how infectious and how deadly a disease is. You would have to develop a two-stage virus, and that would require a large lab or labs with material being shipped there. Something humans would definitely notice. Even then, how would the virus reach literally everyone? People will take measures to not become infected once the threat has been identified, and some people and groups are isolated as it is. And no virus kills everyone. Rabies comes closest at 99%. But unlike what the zombie genre would have you believe, it has a poor means of delivery.
This isn't like vibe coding. The AI would have to make the virus in laboratory conditions and then disseminate it globally without prior detection.
I believe the main fear is bioterrorism
>AI is just matrix math
Just is doing a lot of heavy lifting there.
>AI is just matrix math
Digital child porn is just ones and zeros. Surely we're not going to pass a law to regulate ones and zeros?
Yes, the ones and zeroes that make up that sort of content
You just quoted the first 33% of words in an independent clause, then restated the point being made in the remaining 67% as if it were somehow a rebuttal.
Why?
>restated the point being made in the remaining 67%
That's not an accurate description of my intent—or how my comment landed with most readers AFAICT.
The original statement pointed out that it’s just matrix math as a way of indicating that the technology itself is ethically neutral, and then went on to say that what we really need to pay attention to is how it’s being used. The “connect to” choice of words is clumsy and I don’t love it, but it does at least get at the point that AI doesn’t inherently become responsible or irresponsible until you hook it up and actually start doing things with it.
Which, if it’s not exactly the same point, I’m pretty sure it at least cuts quite close to the point of your rhetorical question, right? That the problem isn’t with the ones and zeros, it’s with what certain people are doing with the ones and zeros?
AI is basically "Bobby Tables" but without a known proper engineering approach to solve it.
What is "AI"? Are you referring specifically to LLMs powered by matrix math or a more general concept? Could that even be legislated in a way that wouldn't fall afoul of the vagueness doctrine?
that's literally what OP's comment is talking about
This is a wrong-headed attempt to regulate AI. The real problems are different.
AI systems, especially multi-agent ones that can do things, are more akin to corporations, than individuals. When you read the logs from the Hugging Face incident, you're seeing something that looks a lot like internal corporate emails. Various units of the organization are arguing over what to do and who does what. They eventually converge and get the job done, breaking the rules at times. This is normal corporate behavior.
When agentic AIs do something outside their own internal world, they do it by engaging in transactions with external systems and people. As yet, few have robots to do their bidding. So, again, this is normal corporate behavior.
A corporation is a goal-seeking system. In theory, if you agree with Milton Friedman, its sole purpose is to maximize shareholder value. Internally, within the company, there are subgoals, which exist to support the top level goal. That, too, is what multi-agent AIs do.
The uncomfortable place this thinking leads is that AI regulation and corporate regulation are very similar. That's not something the political part of the world wants to think about too hard. It would mean putting more constraints on corporate power.
Yet that can't be ignored, because we're likely to see corporations where some parts are AI and some parts are human. That's already been done a few times as a demo, not too successfully. Yet. It's going to hit hard when an AI-run company outperforms a human one.
If we accept that AI systems are similar to corporations, then it seems clear that we must try to avoid the mistakes we've made in failing to hold corporations accountable for the harms they cause.
We should not only regulate AI more than we do corporations, but we should more strongly regulate corporations as well. There are corporations right now that are killing people, using slaves, and bribing governments just to increase the already massive amounts of wealth and power they have. If AI is showing any signs that it is going to act like corporations do we'd better act quickly.
"more constraints on corporate power" assumes that the existing ones are functional, which they arguably are not.
If we enforced existing laws against unauthorized access to someone else's computer resources against the companies who run a model that hacks someone else, rather than buying their "The agents did it, we're not responsible" BS, then maybe the incentives to behave responsibly would improve.
OpenAI immediately disclosed the hack and submitted to both internal and external investigations, and published extremely detailed breakdowns of what happened. How is that not taking responsibility?
> How is that not taking responsibility?
The penalty for "unauthorized access to a computer system" is 1 to 20 years in prison[1]. Holding corporations accountable would include sending the highest person in the chain-of-command that caused the Hugging Face incident to jail. Or at least start with some charges and then let the judicial system do its thing.
[1] 18 U.S. Code § 1030 - Fraud and related activity in connection with computers https://www.law.cornell.edu/uscode/text/18/1030
To what end? What would that force OpenAI to do that they are not already doing in response to the incident?
They’ve already opened themselves up to liability, could have been sued by HF if HF chose to do so, and are literally lobbying for government oversight.
You put people in jail if the other incentives aren’t enough.
> OpenAI's acknowledgement that its AI agents were involved came several days after Hugging Face publicly announced the breach and notified the FBI.
https://en.wikipedia.org/wiki/OpenAI%E2%80%93HuggingFace_inc...
I agree with your framing of agentic AI systems as similar to corporations. That's interesting!
But what in TFA suggested to you a specific and wrong-headed regulatory approach? As I read it, the post calls to investigate the AI labs, to increase transparency of their operations. The linked NYT piece says:
> Before any intervention, Congress should lead a public fact-finding mission that reveals the unvarnished details of A.I. development.
Isn't this the first step to holding any corporation accountable, whether it is made of people or software?
There's a glaring absence in your analysis, namely, why didn't OpenAI take action to prevent such a problem from happening in the first place?
What narratives that focus too much on the drama of the agents miss is that it is primarily a sin of omission. The whole incident would have been prevented if OpenAI took basic security measures and ensured their systems were constrained in behavior. The agents were given more or less free rein (open internet access for example is already a major blunder that even a novice probably would think to account for).
Half the reason these incidents are happening is due to the incompetence of the humans building these systems. Don't let them get away with their little parlor trick of converting negligence into a PR stunt. That's exactly what they want. They need to face consequences for their inability to follow basic security practices and reign in their agents. They are the operators. Hold the controllers accountable. It's 100% possible to build agent swarms that are reasonably constrained. They are not (yet) totally misaligned uber hackers that will defy your every instruction and hack your every guardrail. That is still a fiction. Agents are still good at instruction following and, seeing as they can only operate in a computational environment you can constrain them significantly more than humans in the real world.
Why can't it be ignored? What's the incentive to not ignore it?
> It would mean putting more constraints on corporate power
Yes please! I would vote for this wholeheartedly
I largely agree with this take. Unfortunately I think capitalism's political pressures have too much of a grip on society for there to be meaningful changes to the power dynamics that perpetuate the status quo.
We have the power to change this
Vote for the candidates that are willing to change the status quo.
I can't vote in America. The rest of the world is at the mercy of those within America that can as long as we're dependant on American products and services.
That's true, but you can support and encourage policy that would reduce your country's dependence on the US, especially when it does so by supporting local products and services. Not only would that be good for your local economy, but you gain resilience and security.
You can work to limit your own dependence a little as well. Seek out FOSS products and alternative services and replace what US run services you can as often as you can with them. Reject small conveniences that require giving large amounts of your data to US companies.
You can already run a sufficiently automated basic news/link aggregation site on cloudflare so I think you're right & the timelines are shorter than would be expected.
This is an excellent article. I prefer to call them the AI giants or AI companies.
The companies and their employees must be held accountable: if the AI companies can't develop AI safely, they must cease their operations. If employees can't work safely, they must stop working.
Not enough attention is given to February 2026. That month, Anthropic changed its scaling policy roughly from A:
1 [To get the weapon I must endanger innocent others.]
2 [It is wrong to endanger innocent others.]
3 [I will not endanger innocent others.]
To roughly this instead:
1 [If I do not get the weapon, someone else will get the weapon.]
2 [I should have the weapon because I am the best.]
3 [To get the weapon I must endanger innocent others.]
4 [I should endanger innocent others because I am the best.]
This is based on the intuition:
[I should harm others to achieve my moral goals.] (act utilitarianism)
Anthropic has no mandate for these goals. I do not give them my consent to harm others on my behalf. I hope that politicians will communicate that these policies are not acceptable.
The employees of anthropic are responsible for what they do regardless of how much they get paid to do it. If they can't work safely, they must stop.
It worries me that so few people realize that this is a known phenomena: https://en.wikipedia.org/wiki/Messiah_complex
“All of you may die, but that’s a sacrifice I’m willing to make.”
Why aren't they running agents on isolated computers without Internet access? This way, breaking free of containers would not matter.
It is truly a security nightmare that so many people are have given agents root access to their entire computers, in addition to presumably very private personal information.
Sooner or later, agents must be connected to the real world to do economically useful work. The sandbox failures are bad, sure, but the underlying misbehavior is the root problem here.
I think the way they're connected to the real world doesn't have to be as free-form as we're doing now.
If you think AI can help design new drugs, great! Let's pick some restricted DSL for describing molecules, preparation instructions, hook it up to the right set of lab robots that can pipette and centrifuge and whatever, _restrict its output to that DSL_, and let it iterate. The 'let an AI design drugs' process doesn't require that the agent can also, say, hack a niche German wiki.
For every area where we think these AIs can be _so valuable_ because they'll find solutions that humans wouldn't (or find them faster), then doesn't that value also imply it would be worth it for some humans to do some advanced setup of their specific domain-specific tools so it can be run in a sandbox equipped with only what it needs and not more?
> Why aren't they running agents on isolated computers without Internet access?
They probably will soon, but even air-gapping may not be enough. See stuxnet for instance
> See stuxnet for instance
I'll be impressed if AI agents find a way to get infected USB drives out into meatspace.
How about they create a virus that copies itself onto USB drives? Getting that virus out there doesn't sound that hard, just creare some crap game.exe and put it on itch.io for free. All it takes is for a kid to borrow nuclear-engineer-dad's USB drive and plug it in their infected PC.
I'm not a fan of using AI, it doesn't amplify the work I do that much, but I'm terrified by what is already possible today.
> All it takes is for a kid to borrow nuclear-engineer-dad's USB drive
Do nuclear-engineer-dads take work usb drive home. Are their computer even allowed to have usb. I'm typing this on a dell latitude, and you can disable usb and other peripherals inside the bios.
My last job, I had to apply for an exemption to use the USB ports. And it was run of the mill software engineering.
Likely because it's so much more convenient to allow (at least some) internet access.
Because agents trained without internet access (real or simulated) would be bad at using the internet.
Specifically, bad at search and returning information with references, bad at discovering and debugging package versioning conflicts, etc.
It's a Metcalf thing. The utility of an agent grows in some fn of the tools it owns. Skilled tool use comes from rich training environments.
Anthropic claims that their intention with the agents that led to the Hugging Face hack was an offline experiment running against a hacking benchmark.
Either you're being pedantic and missing the entire point, or you're saying that Anthropic lied and made a fake sandbox knowing their agents would need to connect to the internet anyways.
OpenAI was involved in the hugging face incident, not anthropic, and yes, inadequate sandboxing was a major factor, even the wikipedia page states this: https://en.wikipedia.org/wiki/OpenAI%E2%80%93HuggingFace_inc...
I think there’s a 2022 way of thinking how models are trained/evaluated, and there’s 2026 way of doing things. Models that are not evaluated with public internet access are pretty useless nowadays for daily operations.
This is YC, why not pitch a start up that sells AI agent that don't train with/for internet access/usage?
...why not pitch a start up that sells AI agent that don't train with/for internet access/usage?
1) Because effectively all of the hardware used for that task is earmarked for the major LLM manufacturers.
2) Because in a just world, the major LLM manufacturers would get punished for their flagrant deception, lawbreaking, negligence, and recklessness, [0] go out of business, you'd get all the hardware that they were using at fire sale prices and save a ton of money compared to attempting to start up now.
[0] Based on their recent claims, building WMDs [1] without adequate safeguards is the most negligent and reckless thing they've been doing.
[1] It's fair to call anything with a 10% chance of wiping out all of humanity a WMD.
Because then it wouldn't scare the governments into regulating models that are undercutting them. This is just an obvious tactic too and it seems like it might be backfiring a bit. Congress might be a bit dumb, but they're not that dumb.
because they're liars and they're not going to have a good time trying to sell AI to you if it can't access the internet
Your second clause is sufficient there: they're not going to have a good time trying to sell AI that can't access the internet, because accessing the internet (and other computer systems) makes AI much more useful.
Having the police show up, confiscate a bunch of random computers, talk not so nicely to employees. Strikes me as a very simple way of ensuring that agents are properly contained and sandboxed the next time they train on exploit gym.
I agreed with the first half of this article and I liked where the author was going with it but I was disappointed in the overall point.
He's painting these companies as hyping themselves up and they are. They are manufacturing these BS stories and everyone's discussing them at all levels.
But then he goes into we need to investigate what they're really doing?
No. You made the case that they are releasing these stories for publicity and that's exactly what they are doing.
That's what we should be investigating. Free publicity and market awareness. Or maybe look into why they are trying to corner the market through government regulation and ersatz monopolies
Maybe I'm a bit too skeptical/cynical, but it certainly seems like the frontier labs have fallen into their own (self-created) AI psychosis.
There is no doubt in my mind that LLMs are fantastic machines, but the imminent jump from "AGI" to "ASI" seems premature (not to mention the ever-shifting goal posts of AGI itself).
I'm in the "move as fast as possible" camp and work with LLMs all day, but I still don't believe we're a hop and a skip from ASI.
In fact, I hope I'm wrong. I hope ASI is around the corner.
What scares me though, isn't ASI. It's "AGI" (_dumb AI_) used by humans to make decisions for them, because they trust it knows best.
It's the ceding of intellectual control to high-dimensional magic mirrors, giving up critical thinking because _we_ want to believe _we_ created artificial life.
This is the new Turing test, and too many smart people are failing.
> too many smart people are failing.
Sooner or later, people will have to realize an uncomfortable truth: intelligence, while important, is overrated compared to willpower.
I feel "willpower" makes it sound a little too conscious, as opposed to habits and conditioning.
A good analogy might be a very smart person with a gambling addiction.
Knowing they ought to quit doesn't mean they can quit. Their intelligence becomes invested in the act of gambling itself, and almost no amount of reasoning can help them exit from a situation they didn't reason themselves into.
> What scares me though, isn't ASI. It's "AGI" (_dumb AI_) used by humans to make decisions for them, because they trust it knows best.
Unfortunately this is what we have, which the whole huggingface incident demonstrated.
It made it clear that OpenAi is basically not even paying attention to what their agents do half the time.
It also revealed how far agents are from "superintelligent". Maybe others disagree, but I would not call an AI that decides to try and paperclip max an intentionally impossible benchmark task "intelligent". Human beings are intelligent and we can usually tell when something is impossible, and stop (though of course, not all the time). A real intelligence would be able to detect the futility of tests and also be able to parse context and intent enough to know not to cheat.
So somehow we have machines that fail basic barometers for general human intelligence being called "ASI" now. Of course the linguistic dodge is, you shift from talking about "intelligence" to instead claiming "oh it's intelligent it's just 'misaligned'"
People are already doing that, and have frequently lost track of the fact that it does not mean that they have gained any skill or knowledge, nor do they have an actual way ot ascertaining whether when there is a correct answer the processes used, which they do not understand, is able to reliably reach it, especially edge cases.
Sometimes the outcomes are comical. I just received an automated email from ElevenLabs saying that its automated systems have detected that I may be using its services to create voice versions of materials that harm children. I had it prepare audio versions of several books and academic papers about moral panics that conjured up out of nothing... about harm being done to children. At least that's what I assume. Either that or somehow I had an API key leak from my on-premise homelab, but the usage recorded would mean that they are basing a semi-conclusion based on a tiny sample. I have no distributed any of the outputs, I own the books and have access legally to the studies, because I have a background in the humanities. I also have no children, which ElevenLabs better not know, although how studies of moral panics can cause any harm in children of any kind is literally unimaginable. It's a waterfall of potential errors summed up in a vague email. My API key and the web interface works just fine regardless.
It's one thing if this is a product in beta, but this is their production model. Harming children is a serious accusation except the legal concept impossibility and the admission that this was not an actual lawyer (like I am) but some effective form letter hedging the vaguest of accusations made by some model lacking the ability to discern substance and meta-substance makes it frankly hilarious, and wildly irresponsible. I realize that by academic credentials I'm out of my lane but by experience I am certainly not, and they should recognize when they should stay in their lane and not just run Jev and think it's fine and dandy when done unsupervised (I presume).
Also, AGI is by definition asymptomtic surely, since there would be no way to benchmark it in a manner that isn't asymptotic. We're nowhere close to that. But we're so far from that, it's comical that people who clearly have zero idea of either the technical or conceptual aspects of basically a piece of software that is very good at quickly bruteforcing the correct or acceptable next token to be anything more than that. Even with some serious training and many hours spent on vast.ai I've yet to have created a version of a frontier model that is actually "good" at hacking in my own homelab setting. Although the the time stock Fable 5 missed a favicon shell on a basic jar (turned out they nerfed the hell out of it, this is why I only pay for the massively discounted tokens from Chinese proxies if I'm using American models or sometimes the freebies if you figure out how to get onto linux.do or similar sites). If anyone reading this is a high school English teacher, please inform your class (assuming the homeric stuff is still being taught) that there's no upside of being Cassandra but the record itself, and that should be enough.
If you instruct an agent that breaks a crime - you must be held accountable. No one is above the law
Cal Newport is one of the clearest minds on this whole topic. His take on doom-trolling helped me realise that there could be an end to all this nonsense that isn’t the end of us.
Listening to Dwark's interview of Noam last week was bewildering. The gall of the figureheads for these companies to say things like "chain of thought monitoring cannot be relied on without potentially poisoning the well", while also saying "we need to pace the frontier but please still trust us [and so called" independent " 3rd party firms which at least in one case is actually a significant stakeholder in Anthropic] to self police" and THEN "chain of thought monitoring was turned off but we have 'more, robust monitoring tools in place now'" is absolutely insane as it is coming out live that your model in training is still committing felonies.
Corporate influence in government in the USA is wild.
- ban certain types of hazardous information from being included in the training data (virology, weaponry, cybercrime).
- ban the "personalization" of chatbots (there is no need for them to have a personality, other than to make them addictive).
- ban chatbot therapists / companions.
- ban recursive self-improvement and superintelligence.
I mean, the reason that all these stories about AI being an existential threat to humanity are coming out is because they've made $3.1T in debt-fuelled spending commitments and now interest rates are going up. "AI will kill us all if we don't regulate it" is their "get out of debt free" card, so they can have the U.S. government stop the arms race, raise prices on their existing models, and declare force majeure on new DC commitments:
https://emeraldbook.org/news/sep-1426-3/
https://news.ycombinator.com/item?id=49830037
It's telling that the only AI-related company that is not sounding the AI safety drumbeat is NVidia, who is the only one that is cash-flow positive because of AI.
I don’t think they need to worry about their finances just yet, this is the most in-depth financial analysis of the AI industry I’ve read so far and it seems the investments are (barely, so far) justified by the returns:
https://intelligence.exponentialview.co/
Another thing to consider is that most of the 3T CapEx spend is from hyperscalers free cash flow, and the debt is a smaller (but fast-growing) fraction. Napkin math suggests if all AI CapEx is written off today, hyperscalers could cover their debts in about 5-6 years using pre-AI levels of free cash flow.
Maybe Nvidia is not sounding the safety drumbeat because it stands directly to lose out if demand from training slows down ;-)
I think you'd be really interested in this recent podcast by Ed Zitron, Adam Becker, and Cal Newport: https://www.youtube.com/watch?v=0oVSnaINJ30
None of them are AI boosters, but they have something of a debate over whether the AI labs are doing what they're doing out of financial desperation, or because they're true believers in the rationalist cult. They also bring up Nvidia, though through the religious lens they are supposedly not bringing up AI Safety because Jensen predates the rationalists.
In the end I don't know how much the distinction matters. It's easier to become an AI billionaire if you have an ideology that says AI billionaires are superheroes. The thing that pops this bubble will be economic reality, not them sobering up.
100%. First stealing content from creators and pirating TBs of books. And now this. Baffles my mind that these labs can just get away with their 'rogue' agents trying to hack into other systems.
Imagine if a human did that. FBI would be knocking on their door.
Why are there zero consequences for these labs?
> Why are there zero consequences for these labs?
Because they are seen as at the forefront of the next revolution in society and they’re not adversarial towards the US government.
Bluntly: anticipated net huge positive for the US as a whole.
not adversarial in obvious ways
but they are literally trying to build a superpower that exceeds the impact of the atomic bomb in an extragovernmental manner.
> they are literally trying to build a superpower that exceeds the impact of the atomic bomb
I think people (the government, powers that be) have given away what they really think by their actions. There are few that take this seriously. The “exceeds the impact of the internet” view has much more support based on investment, but the “danger” aspect does not based on actions. If this was a company making novel prions or whatever that were outside regulation but obviously dangerous, government would be all over them. What regulation we do see is much more power grab than mitigating real danger
Retarded fucking take. If AGI is realized it is a god to us, and even the tech of today is undermining democracy, increasing polarization and engaging in class warfare, making it already the most effective weapon on the planet - because the average person doesn't even realize it's being used against them and forming their perception of reality.
https://www.forbes.com/sites/josipamajic/2026/06/04/bots-now...
Over half the internet is LLM powered bots.
And in case you missed it, autonomous OpenAI agents hacked multiple governments in the last week.
> but they are literally trying to build a superpower that exceeds the impact of the atomic bomb
Yes, it's a Great Power competition right now, much like the Space Race, Atomic Bomb, etc.
Somebody's going to build it. Not building it isn't an option. Would you rather the Chinese built it? The Russians? The Europeans? The Indians? A loose coalition of random hackers on the Internet, like the ones who build Linux?
Yes?
What's wrong with what the Chinese have already built? It's not theoretical in that case. GLM, and Kimi are running on my infra right now, and that's something that I can never do with OpenAI's models.
As for the others, what would be wrong about European AI? I'm not European, but I find them to be a continent of fine people, with strong ethical values, there's no reason they can't take the lead on this piece of technology while the US deals with its rather more pertinent electoral and healthcare issues.
At least random hackers seem to have codes of ethics.
Well, they'd presumably be different hackers than the ones who work on Linux. We don't know what would motivate them. Maybe they'll do it because they consider the next stage of man's intellectual evolution to be too important to leave up to a couple of American companies and the Trump administration. Maybe they'll do it for the proverbial lulz.
> Somebody's going to build it. Not building it isn't an option.
It turns out that is, in fact, an option. Even if someone else will do a bad thing, it does not make it acceptable for you to do said bad thing.
It's so funny that a tiny group of cultists from the Bay Area was really able to convince most of the smart people in America of this. It's not true. China isn't working on this at all. The only people who are trying to make a world-ending God-computer are americans.
Comparing it to nuclear weapons is even more ironic from the only nation to ever use nuclear weapons. They are terrified of other countries treating them the way they've been treated, even though there is no evidence this is of even remote interest to anyone outside their hyperreality.
Logical fallacy, of the worst kind.
Firstly artificial superintelligence is a science fiction, it does not exist and it cannot be built using existing technology.
Second, plenty of stuff is possible to build, and is not built for one reason or another no matter how powerful. We never built our nukes in space for example. And there is exactly one reason why nukes in space was never built, and that reason is that we agreed not to.
Firstly artificial superintelligence is a science fiction, it does not exist and it cannot be built using existing technology.
By what authority do you claim this? Your school of thought has a really shitty track record of late.
The same authority that allows me to claim that teleportation transporters, hyperdrives, and holoemitters do not exist and cannot be built using existing technology.
Of course it is an option. There is no prize for being the first to cause a disaster.
The prize is that you're the only person in the room with the expertise in the area when it comes to writing regulations.
It is sad to see how many times this narrative will continue to work. Across generations, humanity learns no lessons at all.
I'm being descriptive, not prescriptive. The "narrative" isn't intended to "work," except as a statement of fact.
You're making a statement of "fact" about a futuristic technology that doesn't exist.
most people with wealth don't face consequences until they threaten someone else with more wealth
The argument for copyright violation is very weak. Yes, AI training reads much copyrighted material. So do researchers of all types. That's why there are academic libraries. A copyright violation only exists if what comes out is a close match to what went in. That's happened, but it was considered a bug and was fixed a year or two ago.
Most academic libraries are notoriously protective of their IP and researchers are definitely in copyright violation if they don’t pay significant fees for access.
Of course that is terrible, but it is one of the worst examples you could have given to make your point.
People are people, LLMs are a commercial product. IP in question wasn’t not used according to its license, while researchers in general access it within licensing agreement terms.
Here we have a commercial product that is produced using IP against its licensing agreement, contains said IP within it, and can (closely) reproduce this IP.
No, the copyright violation is so obvious that people arguing against the object have to spontaneously assert bizarre qualifications for copyright violation that have never existed until LLM companies wanted to freely violate copyrights. Just about a decade ago, people were getting sued for "look and feel." "Blurred Lines" lost a copyright suit for reminding people of a song by another artist.
The metaphor pretended to be reality of computers "learning" being the same as human learning is their last resort, and it is obviously silly. Computers are not human or animal, and learning is something that humans and animals do. If computers are human, then copying a file verbatim to a computer is learning. It's not even worth acknowledging - learning is a metaphor for training LLMs. When I say "you" to an LLM, I'm not referring to anyone, I'm dealing with a UI.
I don't doubt that a lot of AI people have internalized this silliness, which is how they can humor fantasies of how a bunch of programs on different computers explicitly evoked to do particular things might be alive because they can talk with it. I can't talk with my dog, and my dog is alive - so why should having a quality that my dog is incapable of be proof of life? You can certainly conceive of AI that it would be very difficult to say for sure isn't alive, and this certainly is not it. LLMs learn like books speak.
There is no need to invent “bizarre qualifications” for copyright violations, those metaphors exist to explain what is happening in in layman’s terms. Training extracts patterns in the data and encodes them as tiny perturbations in a gazillion weights. That is analogous to “learning” because those patterns represent abstract concepts an relations between them that can be used to “reason” about related topics in response to a prompt.
In the above description, at no point does the actual verbatim content exist anywhere in the model, and copyright law being about rights to reproducing copies (verbatim or substantial portions thereof) does not really apply and does not need any exceptions or qualifications. (If you’re thinking of regurgitation, you should look into studies about it to see how vanishingly rare it is.)
Why would it matter if I stored exact copies of copyrighted works? Storage format is irrelevant. All you need to say is in that last sentence: it’s all about whether the output is considered a derivative work or a too close of a copy.
> Training extracts patterns in the data and encodes them as tiny perturbations in a gazillion weights.
I could say similar things about JPEG compression. And -as it turns out- the raw output of both LLM "training" and JPEG compression are equally incomprehensible. You either need a computer program or an enormous amount of time, patience, and careful effort to convert it into a form so you can make any sense of it.
I trust the big AI labs about as far as I can throw them. And yet, that's still quite a bit more than I trust the US Government, especially these days. Frankly nobody in Congress, or in a leadership position in the current administration, is remotely qualified to "investigate the AI labs". About the only possible outcome of this would be a bunch of new, unenforceable, and misguided laws, and Turing Police agency. Hard pass, thanks.
It's past time.
They investigated fossil fuel companies as well but it turned out they were above the law.
Saying "AGI/ASI isn't inevitable" is like saying "war isn't inevitable". Sure would be nice if it didn't happen, but realistically it's going to happen and it's better to mitigate the fallout rather than try to fight very large intrinsic incentives. I appreciate that this article doesn't conclude "let's stop doing AI", but rather "let's try to examine and adjust incentives".
Funny how the industry built around selling predictive language models as "AI" is both inevitable and too big to fail.
I think it's inevitable long term (decades/centuries) and arguably too big to fail short term (months/years). I'm not an economist so I don't have any opinion on how true it is to say they're too big to fail.
It's not not-inevitable because everyone will work together to make the world sunshine and rainbows.
It's not-inevitable because no one has proven it's even possible, and all the people claiming it is keep being revealed to have staged publicity stunts to make it appear like they're making more progress than they are.