OpenAI 数学成果突袭、万亿芯片债务与韩国银行被黑：AI 泡沫背后的“凭证缺失”危机

Speaker 0: An AI agent knocks on a bank's door. It tries a password and gestoll. No, so it rareads the error message changes its plan and tries a different door than another.

Speaker 1: It maps every system, it can reach works out which ones look weakest and picks its own way in in computers, have been guessing passwords for decades.

Speaker 0: This one is deciding where to look, and eventually it will find way in this one. The banking has a problem. Some something is inside. And the first question, anyone asks is who sent it? That should be an easy question.

Speaker 0: But it's not really because for the first time, one possible answer is nobody did if is a new standard being built right now called web about auth authentication. The lets son, ai agent proof where it comes from by digitally signing everything. It's sense. Think of it like a company's stamp on a letter. If the agent carries one, the bank can see which company built it,

Speaker 1: But a company stamp tells you whose letter head it is. It doesn't really tell you who wrote the actual letter

Speaker 0: Over the past few weeks, at least nine south koary and banks were targeted by one attacker using AI. He didn't need a team of hackers by the way, he needed a free tool of the internet and a subscription to an AI model, the attack mved at machine speed. But at one bank, it took almost three days.

Speaker 1: Just to notice, something was a miss.

Speaker 0: And the only reason anyone has a name for this hagger today is that he made a careless mistake on his cv, a human mistake. What's strange about that attack is that the AI doing it wasn't really a hacking machine.

Speaker 1: It was one of our brand new shiny, AI friends that now do anything from booking trips to hacking banks to discovery new knowledge humans aided by AI are becoming a force of nature.

Speaker 0: We are becoming more efficient at everything we get our hands on being hacking included, sadly, but being able to produce something faster, doesn't necessarily mean anyone can check it faster or understand it faster or pay for it faster.

Speaker 1: So in today's video, i want to talk about the gap between those things and what is costing all of us,

Speaker 0: Because the beautiful utopia, an ai and fuse future ain't going to happen my friends unless we get the architecture, right? Some of those costs are, of course, very easy to see last week alone, wall street lined up well over a hundred billion dollars in loans just to be for the chips. These AS run on chips whose value a few years from now, nobody can actually check loans by the way, have to be paid back just as a gentle of reminder to our profitless ai overlords.

Speaker 0: So if part of you has been waiting for the ai bubble is going to pop k on vivio OI have some ildly encouraging news there. The people lending the money have started charging a lot more. Just in case some of them wonderland at all until they've actually seen the books. And just last night, when one sit of books finally came out, the most important number in AI turned out to be twenty billion dollars smaller than everyone thought.

Speaker 0: Other costs are much harder to see, but of course, no less relevant and open AI model has just produced answers to hundreds of mass problems that people spent here stuck on and open a ice own mathematicians cannot yet explain many of them. So what has that cost the mathematicians over every time? Does the pattern in our current reality? The machine delivers first and fast, and the understanding turns up later. If it turns up at all, everyone is asking for receipts, almost nobody's,

Speaker 1: Getting them almost nobody's happy with this wholfiasco.

Speaker 0: So in this video, first,

Speaker 1: I'm going to cover the maths with the answers are probably right.

Speaker 0: And nobody can say why right now, then we're going to talk about the money side and who all of that math is really mental impress.

Speaker 1: Then the new bank hacking scandal that are shaking us all up and the mistake that gave the hacker away.

Speaker 0: And at the very end, we have to have a very serious discussion in my friends about how AI should actually be regulator drawing from existing systems. That's what i'm going to do so that someone always has to answer for what an asia does, and we can finally get some order in all of this mess.

Speaker 0: So let's begin on the six, the october opening. I put three hundred and seventy two mathematical results and github. Each one, the company says either souls or makes real progress on a major open question and math or theoretical computer science.

Speaker 0: One of them claims progress on the reman hypothesis, which is a question about the hidden pattern behind prime numbers that's been unsolved since eighteen fifty nine, with a million dodolllar price waiting for whoever cracks it. And according to open a eye, almost all of them came from a single prompt handed to a single AI agent, although it admits some may have taken a couple of attempts just a month earlier.

Speaker 1: By the way, allegedly the same model had cracked the navy ststok's problem, one of the six biggest open problems in mathematics today. Don't worry about what it is, only seven people in earth actually get it, but that apparently took a swarm of ten thousand agents and millions of dollars of computing power.

Speaker 0: Now, four weeks later, the same company is saying that it can do this kind of thing with one prompt. I'm not saying they're lying. I'm just saying to me the silla about us us before anyone one can basass in their mathematical glory.

Speaker 1: As one does, somebody has to really check these proof are actually right.

Speaker 0: So how would you even check three, three hundred and seven to do proofs, know know most of history. The ausor was very slowly, another mathematician reads the proof lined by line. Meticulously neeither finds a mistake or dozen that can take months for a single paper. But many of these results now came with something extra, which is a file written in lean. Lean is a programming language for proof. You basically write every step of the argument as code and a program tricks each step and accepts nothing on faith. If the proof passes, it's correct. As long as a statement you wrote down is the problem you actually meant to solve, which is why scientific a's verdict was that many of these results are all but certain to be correct.

Speaker 1: So checking whether the proves are true, turned out to be the part machines cannow do as well.

Speaker 0: Understanding them is a completely different beast. Open eeye told scientific american that many of the new results are not yet understood by its own mathematicians. The company that handed in the answers cannot really explain them yet. This is basically every mouth teacher's nightmare, full marks, no working show.

Speaker 0: But before this turns into a story about mathematicians being crossed with a computer, it's worth hearing the case for releasing everything, because it is a very strong one. Daniel ellit, who's a mathematiction of the university of toronto told scientific american.

Speaker 1: He can't see why anyone would ask open a eye or any company for that matter to keep the ansstor secret. And he thinks this will be good for mats,

Speaker 0: Someone unread. It made a similar point with rather more swearing, basically saying, mathematicians didn't ask for this work when the complaint and the reply was well, the rest of us asfor it. And i am personally with them. Honestly, i am a scientist, and what drives me primarily is curiosity about how the universe works. And in my heart of hearts, i will always choose knowing things above almost anything else that doesn't mean anything goes, of course, i have limits and there are lines. I wouldn't cross, but this is just nowhere near one of them.

Speaker 0: So am i delighted that something that isn't human,

Speaker 1: Just worked out hundreds of mathematical results? Absolutely,

Speaker 0: Does it hard to watch people's work,

Speaker 1: Get displaced even temporarily also? Absolutely, i can definitely see that,

Speaker 0: But would i prefer a world that didn't know these answers just to a peace ce few mathematicians today? Not really, i'm still going to choose the proof. And this isn't only about mathematicians,

Speaker 1: By the way, i would make the same choice in every discipline, including my own, even if it puts smelt of work. I know that is a very hard choice, and i know plenty of people won't agree with me.

Speaker 0: That's, of course, completely fine. But more importantly, i can hold that view and still see what it costs. The people on the other side,

Speaker 0: Just as an example, thomas bloom runs the air dos problems website, which is a collection of more than twelve hundred problems left behind by paul aerdos. He's one of the most prolific mathematicians of the tweneth century between ten and twenty five thousand people visited every single day. It used to be a place where people argued about problems together. But now blooms us, the main thing people do there is post AI generated proof, often with no explanation at all. Just a plant of flag. Think of it a bit like the mathematical equivalent of commenting first under a youtube video, which i door by the way.

Speaker 1: So this month, he basically frozed the comments stop using any language that gives credit for solutions and remove the labels that showed which problems were open and which were solved.

Speaker 0: This is how far we've come. The most popular screboard in AI mass is now a scorboard with no scores on it. There's another group,

Speaker 1: By the way, calling itself the association for human mathematics. And they went even further.

Speaker 1: They said, dropping hundreds of files at once was a demonstration of power, not collarship,

Speaker 0: But even setting that aside for a second, just think about it is a proof. We cannot explain,

Speaker 1: Technically still knowledge. I thought back to asspirin the medicine.

Speaker 0: It has been around since the end of the eighteen nineties for more than seventy years, people took it for headaches and fevers, but nobody really knew how it actually worked. Then in ninety seventy one, a british pharmaccologist called john vain, worked out that aspirin stops the body, making certain chemicals that i can not pronounce, but theyessentially responsible for a lot of information, fever and pain, and he shared the a bell prize for this.

Speaker 0: Once people understood why asspirin worked, they started testing it for completely new things, such as preventing heart attacks and even strokes. So knowing that something works and knowing why it's working are different kinds of knowledge medicine lived happily with the first month for seventy years. But the second one is where the next discoveries come from, which is why mathematicians are asking for something very specific here.

Speaker 1: In september, opening eye set up an independent advisory group of mathematicians to tell it how to publish results like these responsibly.

Speaker 1: The group asked that every result come with a model, the exact prompt and the computing time behind it,

Speaker 0: They essentially trying to arrive at something that can to reproduce ability, open eye, publish the average computing time, but no prompts. And it told us scientific american. It is not bound by the recommendations of the advisory group it assembled.

Speaker 0: So i guess, essentially advisory in the same sense that a wet paint site is advisory andrey satherland, who is a mathematician at mit.

Speaker 1: Put it very simply, he said, we should be asking for receipts.

Speaker 0: So why not slow down and hand them just overham? Well open AI says it can't. According to them. These problems are the clearest. Does it has that its AI really is getting smarter? And that matters well, beon d maths, of course, the whole AI industry is definitely borrowing, staggering amounts of money on the sole promise that these models will keep getting smarter.

Speaker 1: Ter results like these are some of the most convincing evidence they can point to. So who is lending all of that money?

Speaker 0: And how comfortable are they feeling about it before we get into all of that? Just a quick word from today's video sponsor cape is AUS mobile carrier built around privacy from the ground up most privacy tools, try to protect what happens after your phone connects to the internet,

Speaker 1: Keep us doing something a little different by trying to build privacy into the mobile network itself. One of the most interesting features is identify rotation.

Speaker 0: Normally, your sim has a persistent identifer called an isi that selll learn networks used to recognize you cape automatically rotates that identifier. Every twenty four hours making it much harder to track the same subscriber over time. Cape also offers mumuti network coverage, which lets your phone automatically move between two major nationwide networks using the same syym that can improve coverage while spreading your traffic across multiple networks, can also make persistent tracking more difficult. And then they are disappearing call locgs instead of holding onto your call and text methodata and definitely cape deleads that information after twenty four hours. So rather than adding another privacy app on top of your phone cape is trying to change what happens under neath it. If you want to try it, go to cape dot coal slash EL and use coat EL thirty three for thirty three percent off cape service for six months.

Speaker 1: The link is also in the description below right back to the money.

Speaker 0: I've made whole videos about whether ai is a bubble. So i'm not onona go into all of that. Again, this part here is nearewhere whose lending the money and how nervous are they getting about it? Let's start with the biggest recent deal on the fifth of october bank of america, city group and morgan stanley started selling off pieces of a sixty billion dolar debt package, the money pace for anthropic to lease ai trips designed by google and broadcomm arriving in twenty twenty seven.

Speaker 0: Anthropic doesn't buy the trips, a separate company holds them and runs them out,

Speaker 1: And the debt sits there, not on anthotopics books. This is important.

Speaker 1: This same week, spacex went looking for forty billion dollars to buying video chips, and broadcomes started early talks about another thirty to fifty billion for open a eye. Morgan stantly estimates that ai infrastructure will need one a half trillion dollars of outside money by twenty to twenty eight.

Speaker 0: Now to see why the lenders are nervous. You need exactly one piece of finance jargon. The transh ryan goslik explained tranchis with a jingatatar in the big short, if you'd rather be serenated by him, that would be laua land, which is a completely different film. This channel led us heart science ces like real nerds. So here's a version without all of the music. Imagine one big pot of borrowed money, cutting disslices the top slice gets paid back first and loses money last that makes it safer. So it earns less the bottom slice,

Speaker 1: Gets paid back last and loses money first, that makes it risker. So it earns more

Speaker 0: In anthropics deal. The top slice is forty two billion dollars and brought come baxet with something called residual al ue support. Ort in anin english brought com is promising. The chips will still be worth something later. Think of it as a car dealer, promising your bank, what your car will be worth in five years, so that the bank will lend you the money to buy one of their cars. The bottom slice eighteen billion dollars has no broadc m promise behind it. Anyone lending there is betting on anthropic itself. And according to the financial times, the banks might prefer to wait until an anthdopic goes public, so investors can actually read its financial disclosures first. So the people taking the biggest risk want to see the receipts well before they hand over the money.

Speaker 0: And yesterday, drumveral police overseed actually arrived the financial times reported that open a ice analyzed revenue, which is the single number investors used to judge demand for ai is close to fifty billion dollars, not the seventy billion that was widely reported. Just a few weeks ago. Mind you, part of the gap here is accounting untoroc count sales made through amazons in googg's clouds and open a eye doesn't do that, but the seventy billion dollar figure went around for weeks and open a eye did not deny it.

Speaker 1: You see what i'm telling you here when the receipts finally showed up, it was twenty billion dollar, less,

Speaker 0: Twenty bill is not around an erroosound. And by the way, this story came out a three am on the same day that i'm releasing this video. So please keep that in mind when you're considering whether to like this video, honestly, the llenders asking to see the books is a very good thing. That's exactly what they are supposed to do. Some of these borrowers are not making a profit yet space sex, for example, as four point two eight billion dollars in the first three month just of this year on revenue of four point six nine billion.

Speaker 0: And every big infrastructure built out by the way, has been through a face like this. Railways went through it in the ninenine hundred th century. Fibre opctic networks went through it around the year two thousand when companies built for demand that did not arrive in time and several teleconconfirms collapse under their own debt. Investors asking harder questions is literally the market doing his job, but not everyone is nervous. Mind you,

Speaker 1: In august and videa teamed up with a polo, blackrock, blackstone, brookfield, golmonsax and KKR to raise more than five hundred billion dollars with a stated goal of turning ai compute into an investable acset class.

Speaker 0: Investtable acset class is a lovely phrase. It means persuading the kind of investors who usually buy buildings to buy chips. Instead, buildings as a rule, don't verally get overtaking by a newer shiny building model every couple of months. And by the way, you're looking for the best return and investment in this whole story, it isn't in any of these slices or trscious. It belongs to cirkay brin, who has spent a hundred and two million dollars fighting a california attacks that could cost him around thirteen billion. This has absolutely nothing to do with the rest of the video. I just had to put it out there. Anyway.

Speaker 1: The efficiency is definitely a real thing. Bright right now is being paid for with borrowed money by lenders who would quite like to see the books and cheap, powerful ai is cheap and powerful right now for anyone,

Speaker 1: Including one person who, over a couple of weeks, this autumn pointed it at at least nine south arian banks.

Speaker 1: So how did he do it? And what does it? It say, whwhe're real dangerous.

Speaker 0: Over the weekend, i met with a couple of friends of mine, because i had max out my cloud usage. So i was time to see some humans. They're very bright people. And they asked me among many other things, whether i'm worried AI is go to kosol. My answer was no, not in the way they meant AI doesn't want anything. It doesn't wake up with plasat. Its mind they pushed back with a story. Everyone has heard by now, which i covered in a previous video and open an agent that broke into hugging face yattayta. And my answer was that the ai was doing what it had been asked to do and finding a way to do it,

Speaker 0: Still having no intent does not mean having no consequences if someone breaks into your house to check the answer to across word puzzle, touches nothing in lease. You're technically relieved, but you're also going to change the locks because not, you know the locks don't work, where i actually worry is something different. An ai with no intentions of its own is only as harmless as the goal, somebody gives it.

Speaker 0: So what happens when a person supplies, the goal and the goal is a robbery or something a lot more horbithic from laate september banks in south korea started discovering breaches, at least nine have now been reported as targets. Shinhan bank said the personal information of around twenty five thousand thousand its customers had been compromised at kb c cookmen. It took almost sixty eight hours just to detect the bridge. Customers started asking online whether they should keep their cash at home. That's how far it got

Speaker 0: The two at the centre of all of this is called our tex. It's an open source thing, and it was published githhub this here by a chinese security engineer. It isn't an AI model itself, but it's more like a harness that plugs into models like tragedy by deelaud and deep seek. And it automates the work of a skilled attacker who used to do it by hand back in the olden days, twenty. And twenty five, the tool gathers information hunts for a weak spot plants of roote in runs the attack tools and checks whether they worked

Speaker 0: Now password. Guessing scripts have exiisted for decades. We all know that what's new here is the planning is automated as well. A developer at one of korea's 's biggest broker.

Speaker 1: Just put it very blunktly. They said, even without the skills, anyone who pays can use attack tools,

Speaker 0: There's a fair counterpoint to this.

Speaker 1: And the korean police made made themselves, they said, whwhether ai was used or not is not the point. And security specialist as well,

Speaker 0: Said, the bank's own wek defences mattered a lot more. I think that is true a log that opens for an ai would probably have open for a very patient human as well. It would just have taken that human far longer and a lot more skill.

Speaker 0: So who did this according to the security firm crowds strike, the suspect may be a twenty six year old in southern china crowstrike piece that together by analyzing the attacker's ai sessions in one of them,

Speaker 1: He asked claud, where people typically sell stolen korean data in another. He asked laud or write him ACVA security research or CV.

Speaker 0: By the way, with his age, his telegram account his education in his hometown artists has get hut page for the record, says the toolist for personal learning and must not be used against real online systems, which is, of course, something that i'm a hundred percent sure this attacker definitely read with a lot of attention.

Speaker 0: So going back to the conversation that i had with my friends. This is the bit i would out ter that conversation the law already separates tresppass from burglory. It can be the same broken window. They got you into the house,

Speaker 1: The differenis, what you meant to do.

Speaker 0: Once she got inside the trouble, of course, is that the bank cannot see what you meant to do. While it's happening a curious experiment and a robbery look exactly the same. So the bank has to treat every intrusion as the worst possible kind until it can prove otherwise. This attacker happened to write his intentions down, but most won't. You know, in my last video, i talked a lot about accountability. And afterwards,

Speaker 1: Somebody in the commons ask me a very fairic question. Fine. But what would you actually do about it?

Speaker 0: Well,

Speaker 1: I thought about it a lot. Here is my answer,

Speaker 0: And it boris an idea from the banks themselves. Banks already solved a version of this problem with money. If you move money through a bank, the bank has to know who you are.

Speaker 1: If something looks unusual, it has to report it.

Speaker 0: These anti money launding rules are perously twitching.

Speaker 1: By the way, they will flag a perfectly innocent transfer,

Speaker 0: Just because theyve never seen one like that before. Is anyone who's tried to pay run from a new account can tell you, but they work well enough that criminals go to an enormous amount of trouble just to avoid them. So who should answer for an asian that breaks into a bank? I don't think there is one answer, just as an example from where we can take an obstruction. The korea story had three parties in it, a company whose model did the work a person who pointed it at the banks and banks that took days to notice. I think each of these three layers carries parts of the responsibility.

Speaker 0: The first layer is the companies, the companies, the builties model should have to detect misuse and reported the way a bank reports, a suspicious transaction on anient reaching for a bank slogin page leaves the trail because it's trafficastic of somewhere. And the company running the model can see where this is a buildable thing. It is technically very feasible. It needs a regulator and says the internet doesn't stop at borders. That regulator cannot stop at borders either.

Speaker 0: The second layer is the person themselves web bought off the standard that i mentioned to the sort of this video already gives an ancient a company stamp clououd laare on throtopic opopen. I are reportedly already using it was missing here. My friends is a stamp for the person who authorized this agent to do this task. My personal agians sending an email for my own account, because i asked to you is one thing. My personal agient poker out a bank's login page is a very different thing. If my personal name is on my agents passport, the responsibility for what it does is partially mine as well,

Speaker 0: Does everything needs to be traceable, probably not? And that debate is still wide open, and i have not convershed to a conclusion on that one, just to go back to money. Cash still gives us an anonymous transaction. And that's mostly fine. But just in the united states, for example,

Speaker 1: If you move more than ten thousand dollars and cash, the bank files, a report agents could work the same way.

Speaker 0: Browse anonymously and book your flights, the moment in asian touches, something serious like a bank system. It carries a passport that identifies a couple of layers here.

Speaker 0: The third layer that i'm talking about is the institutions defending themselves sixty hours to notice.

Speaker 1: A breain is not a good luck. This is not good enough when the attacker moves at machine speed.

Speaker 0: And this layer matters, most of the heart edge open models that people run on their own computers. There is no company in the middle to report anything, and a passporate cannot force a criminal to sign his traffic. So the bank has to do its part. The institution has to do its part and trade onsign traffic with a lot of suspicion the same way. Your email already sends messages. It cannot very five straight to spam. You've seen this implementing these three laydoesn't necessarily mean slowing down by the way, it just means doing what every serious field already does.

Speaker 0: In math, you show you're working and somebody else chets it in finance. You open the books before anyone lends you a penny in both. The receipt comes with the answer, not years after it right now. Ai hands over the answers and skips the receipts,

Speaker 1: The hacker and korea handed his over by accident.

Speaker 0: Everyone else is going to need a system that asks for these receipts, if you'd like to know where my friends original question actually leads where ai is really going to kill us. All. I went through the actual research on exactly that in this video that i'm linking here in your screen. That's the video that i would wash next.