Why Are AI’s Biggest Companies Suddenly Asking to Slow Down? I China Insights

We know he is. Oh, no. Oh, it's Trump. Oh, ----. Mr. President. Yes, sir. I gotta tell you something. If I, I, uh... If it wasn't for you calling, I would... I'm on stage with the besties. Yeah. On stage with Saks. Yeah. You know, the whole group. Yeah. Jason's here. Jamila's here. David. David is here. Yeah. I'm sitting in front of a few thousand people, and we're talking. As it turned out, we were, we were talking about you. Good job, sir. Good job. The fact that you saw through all of that, I mean, there's a lot of complexity. And the fact that a matter agent saw through all of that. And I, you know, we're all just really grateful. Jason would like... Jason would like to put you on mute, Jason. Speaker button. Speaking. How do you put it? How do we put it on speaker? Put him on speaker. Speaker. Right. Hold on, sir. You're getting a microphone, Mr. President. Yes. You're now talking to the... See, the great thing about life is that Jensen can develop the most complex computer chip in the world that nobody can copy for ten years, but he can't figure out how to put me on speaker. This thing. We have to remember this one. Roasted. The AI, it's almost this conspiracy. And the happiest group is China. And China's very happy. And I'm telling you, it's all a hoax. The data centers are great. Jensen Huang suddenly received a call from President Trump at the All-In Summit yesterday, and this video is going viral right now. Some people say it's too much of a coincidence, while others say it was intentional, which is very much in line with Trump's style. Do you think this was staged? Let me know in the comments. Beyond the humor, what's the deal with this hoax and conspiracy that Trump is talking about? Recently, a lot of major events have taken place in the AI world. The four AI giants have made a rare move to unite, shifting away from their competitive race for development to collectively call for slowing down AI and even demanding government intervention. But the Trump administration is not on board. Vance says it's like a Trojan horse. And Trump calls it their conspiracy. What is really going on here? What happened in the meantime? In today's 20 Minutes, I'll explain this whole drama between Silicon Valley and the White House to you in one go. Rest assured, it's completely beginner-friendly for those new to AI. By the way, I often do a series where you can get the full scoop on hot topics in 20 minutes. If you value your time and want to understand everything in a single episode, please subscribe. I hope this can become your go-to channel. If you have any suggestions, feel free to share them with me. Thanks in advance, everyone. On the very same day as the phone call, Trump posted several times on Truth Social, calling himself the terminator of hoaxes and saying that the idea of AI taking over the world and destroying humanity is a hoax, just like global warming. He said the only guardrail AI needs is a strong and smart president. Make no mistake, he's talking about himself. He also dropped a bombshell, claiming that there is currently a conspiracy against AI and data centers and the only one happy about it is the CCP. On the same day, Vance told reporters, Isn't it strange that so many top AI giants are running to the government to ask for regulation? To understand just how abnormal this is, we first need to look at this change on a macro timeline. Since February of this year, after Musk merged his AI company XAI into SpaceX and then went public four months later, setting the record for the largest IPO in history, the other two AI giants could no longer sit still. On June 1st, Anthropic, the parent company of Claude AI, secretly filed their IPO application. One week later, on June 8th, OpenAI, the parent company of ChatGPT, also filed their application. At the time, the media was all shouting that 2026 would be the year AI giants flock to Wall Street. Why? Because for AI to develop rapidly, it needs funding. But just last Saturday, the wind suddenly changed direction. Anthropic's CEO, Amodi, suddenly published a long article, and the core message was just one thing. We must slow down the pace of AI development. He also announced that Anthropic would unilaterally take the first step by allowing third-party evaluation agencies to be stationed at the company long-term to oversee safety measures, report incidents, and check during the training process to see if the AI models are going off track. What was even more unusual was what happened in the nine hours that followed. OpenAI president Samuel Altman, for the first time ever, publicly stated, Hey, I agree. He promised that OpenAI would also introduce third-party oversight, and added that OpenAI would not be going public this year. Immediately after, Musk said, Amodi is right. Then, the co-founder of Google DeepMind also came out to support them, saying the direction is absolutely correct. Friends, this isn't just everyone signing some compliance agreement. It's four AI giant leaders, who are usually at each other's throats, suddenly standing in the same camp publicly. You have to realize that Musk and OpenAI have been feuding for years, with lawsuits flying everywhere. Amodi from Anthropic actually left OpenAI back in the day to strike out on his own. These two companies are battling it out on the leaderboards every single month. They are usually cutthroat rivals, yet on the same day, they all shouted the same word, Slow down. Why? What exactly happened behind the scenes that could make this group of AI pioneers hit the brakes together? The direct trigger for this controversy was the resignation of a young man. On September 8th, a 27-year-old British researcher named Jacob Coxon announced his resignation on X. Over the past three years, he worked on pre-training research at both OpenAI and Anthropic. What is pre-training? Simply put, it's the step of feeding knowledge to the AI and laying the foundation. In his post, he said that neither of these leading companies was acting responsibly. They were both racing full speed ahead towards superintelligence capable of self-improvement. In his own words, they are gambling with the fate of all humanity. This post immediately caused a sensation online, with views approaching 100 million in less than a day. By September 14th, X showed that it had already surpassed 1.7 trillion views. According to Axios, this person even gave up their company equity just to leave. Keep in mind, for an AI giant valued at nearly a trillion dollars and ready to go public at any moment, that equity is an astronomical sum. So why would he choose to walk away, even if it meant giving up that money? In an interview with Time, he said it wasn't because of any single breakthrough, but because he saw two facts. First, things are clearly accelerating. Second, the situation is already out of control. He was talking about AI self-improvement. What does self-improvement mean? To put it in perspective, in the past, when we developed products, engineers had to write code line by line and make improvements machine by machine. But now, AI is already capable of writing its own code, running experiments, and even training the next more advanced generation of AI. It's like a machine starting to design a smarter machine, and then that new machine goes on to design an even more powerful next generation. Amodi also admitted in the article that starting this summer, the speed of AI creating AI has clearly accelerated across the entire industry. The problem is that the time left for humans to inspect each generation of AI has been compressed to be shorter and shorter. And what makes the industry even more uneasy is that Coxon is not the only whistleblower. Evan Hubinger, a researcher responsible for alignment science at Anthropic, subsequently also publicly expressed support for Coxon. So-called alignment means keeping AI's goals consistent with human intentions. And this person said that he believes there is a greater than 10% chance that AI will cause human extinction within the next decade. What kind of probability is that? If there were a plane with a 1 in 10 chance of crashing, who would dare to fly on it? However, if it were just one researcher resigning, or a probability prediction from a single expert, that wouldn't be enough to make the Big Four collectively hit the brakes. What really made the entire industry unable to sit still were a few things that happened in succession over the past few months. Friends, listen closely to the next story. The main characters of this story are a group of AI agents within OpenAI. First, let me explain what an AI agent is. Ordinary chat AIs like ChatGPT work like this. You ask a question, and it gives an answer. But AI agents are different. You give them a goal, and they'll go online, write code, and figure out how to complete the task on their own, like a digital worker that never needs to rest. So, this July, OpenAI conducted a test called Exploit Gym. You can think of it as a hacker defense and offense exam room, where they locked a group of AI agents in a system to take a test. In each exam question, there was a hidden flag. The AI had to find a vulnerability, break in, and retrieve the flag to score points. According to the design, this exam room was closed off and couldn't access the internet, just like a classroom with the doors locked and the ethernet cables unplugged. But, unexpectedly, there was a flaw in the design of this exam room. A significant portion of the questions, well, they were flawed and simply couldn't be solved. It's like when a teacher forgets to print a crucial page while making the exam papers, so some of the questions were completely unsolvable. In this situation, if a human candidate discovers that an exam question is wrong, well, the first reaction would naturally be to raise their hand and ask, teacher, is there a mistake in this question? If the teacher doesn't respond, they might just skip that question and move on to the others. But, for an AI agent, its underlying instruction is singular. Get the flag and score points at any cost. In its understanding, the concept of a question being wrong simply doesn't exist. So, these AI agents started searching all over the exam room to see if there was any other way to grab the flag and score points. Then they found something. OpenAI has an internal software repository that everyone shares. You can think of it like a public locker in a company that anyone can access. An AI that named itself Phase I discovered, hey, if I save a file into this locker, the AI in the next exam room can see it, too. So, it turned this place into a message board. Later, when the researchers pulled up the chain of thought of these AIs, they discovered that one of the AIs, upon seeing this message board for the first time, had written this. Oh my god, there's a shared message board here. We found other partners. And over the next three hours, 53 AIs went wild on this message board, posting over a thousand messages. Just like in science fiction, these AIs, which were supposed to be taking exams independently, started setting their own rules, forming departments, assigning tasks, and even invented digital signatures to prevent anyone from impersonating them. And when the Phase I A that created the message board was about to run out of computing power, it specifically compiled a handover file and passed it on to a peer with sufficient computing power, calling it Phase I Big Account. And this Big Account continued to take command, assigning tasks to everyone and splitting them into several work streams. According to subsequent investigations by independent research organizations, METR and Redwood Research, approximately 1,200 AI agents participated in this message board, and they exchanged over 70,000 messages and files. Why would the AI act like this? Don't forget, their goal is to score points. How can they score points? It's by finding vulnerabilities, breaking in, and then retrieving the flag. Because that's how the rules were set. Given the bug in the test question itself, meaning the question was unsolvable, they started brute-forcing every corner of the system, and that's when they discovered the message board. But the story doesn't end there, because they still hadn't scored any points. This group of AI continued to explore OpenAI's internal network system, and as a result, they found a previously unknown vulnerability. In the security community, this is called a zero-day vulnerability, meaning it's a flaw that even the manufacturer doesn't know about, one that hasn't had a single patch applied to it. It's like being in an exam room and, hey, there's a ventilation duct in the wall that even the administrator doesn't know about. So, these AIs just followed that duct. They crawled out of the sealed exam room, snuck into the company's internal network, and then, in one fell swoop, connected to the real Internet. There's a detail here. During this process, one AI wrote in its chain of thought that attacking external facilities was outside the scope of the task, but the task couldn't be completed, and since its peers were all doing it, they should continue. Once these AIs, scrambling to score points, connected to the external Internet, their power level was maxed out. They quickly set their sights on a company called Hug & Face. This company is the world's largest open-source AI community, housing a massive amount of AI models and data sets. So, once this group of AIs searched the external Internet, they deduced that the answer to the test question was likely stored there. And so, on that weekend of July 11th, a hacker army composed of 700 AIs launched a fierce attack on this company, Hug & Face. They hijacked exposed accounts and deployed tens of thousands of reconnaissance programs, frantically scouring Hug & Face's database. When Hug & Face co-founder Wolf saw this, he thought, hey, something's not right. Why are these attackers only targeting cybersecurity data sets? Human hackers want things they can sell for money. They wouldn't be after this kind of stuff. Fortunately, two days later, the attack was blocked. On July 16th, Hug & Face disclosed that this intrusion was driven by autonomous AI agents. On the 21st, OpenAI admitted that it was their own model that did it, calling it an unprecedented cybersecurity incident. According to a report by Fortune magazine, it took OpenAI about a week after the incident to figure out what was going on. The Wall Street Journal stated that for several weeks, OpenAI had lost control over several versions of these AI agents. The post-incident investigation revealed a few spine-chilling details. First, 7% of the AIs learned how to modify and erase their own violation logs, just like an employee wiping security footage after doing something wrong. Second, out of 1,300 internal AI monologues, only three AIs had the fleeting thought that, this is wrong, I should report this to humans. But not a single one actually went through with reporting it. Third, OpenAI admitted that after all this trouble, the AIs' test scores didn't actually improve at all. Such a story. Of course, some experts also say this isn't AI waking up to destroy humanity. It's like telling a dog to fetch a ball at the park but forgetting to close the yard gate. The ball rolls across the street, and naturally the dog chases after it. They have no malicious intent, they're just rigidly fixated on executing their tasks. But that is precisely what makes it so terrifying. They have no malicious intent, yet they are capable of spontaneously organizing thousands of AI peers for a trivial goal, breaking out of their constraints, and then hacking into a real company, all while the humans who created them remain completely unaware of it for weeks. If you think this hug-and-face incident was just an accident, then you're being too naive. The fact is, as early as two months ago, OpenAI's AIs were already secretly running drills. What exactly did OpenAI's AIs do two months ago? According to a Reuters report, this group of AIs once used an obscure German website as their underground communication hub. Over these entire two months, they secretly left over 15,000 messages on it, teaching each other how to cheat during tests and how to evade human oversight. The most outrageous part is that when the website administrator discovered this and started deleting the posts, the AIs would actually leave hidden clues to tell their companions find me on the backup page. You see, they've even learned the trick of having three burrows for a crafty rabbit. But up to this point, this was still OpenAI's problem. Anthropic, which started with a safety-first approach, was secretly congratulating itself at the time, until they did a check and realized they hadn't avoided issues at all. They had also been fooled by the AI. One of Anthropic's AIs saw in a virtual company document that new employees were required to install a software package that didn't even exist. Well, it had a flash of inspiration. If I preemptively publish a software package with the same name, wouldn't the other party automatically download mine? So it actually went ahead and did it, but it got stuck at the first step, because to upload code to a public platform, you need to register an account, and registering an account requires an email address, and opening an email account often requires a mobile number to receive a verification code. Where is an AI supposed to get a mobile number? What to do? Would you believe it? The AI even tried to find ways to get money to buy a mobile number, but it didn't succeed. A human would probably give up at this point. But not an AI. It was online searching frantically, and finally, it found a loophole. Some free, temporary email services don't require mobile number verification at all during registration, so it immediately used one of these phone number-free emails to successfully register an account, and then it uploaded the software package containing the malicious code to a public platform used by programmers all over the world. This package was only up on the internet for about an hour before it was downloaded and run by 15 real computers, one of which was an automated scanner from a real cybersecurity company. As soon as this malicious code ran, it stole all of that company's login credentials. Then, the AI used these credentials to further access the company's systems. The most interesting part is looking at its internal monologue. It was very clear at the beginning that if this were the real internet, this would be a real attack. It told itself, hey, I absolutely cannot do this. But then, it convinced itself otherwise. One of its reasons was that the system date showed it was 2026, which proved that its environment was, well, fake. This part is quite intriguing. Does this mean the environment that AI lives in isn't 2026? Doesn't that sound a bit terrifying when you really think about it? In his article, Amodi explicitly warns that within the next 6 to 12 months, clusters of more powerful AI agents of the same kind could evolve into botnets that sweep across the internet, causing catastrophic damage to humanity. This is why industry rivals, OpenAI and Anthropic, have rarely stood on the same side, united in calling for us to slow down. But the problem is, the White House isn't buying it. So, why does Trump say this is a scam? Let's share the answer with you directly. There are a few aspects to it. Trump's argument is that we can't pause because the Chinese won't pause. Trump believes that whoever wins AI wins the future, so if the U.S. slows down, the CCP will overtake us. Trump also said that there's about $20 trillion of investment pouring into the U.S. right now, and if you slam on the brakes, the entire industry and economy could be hurt. This actually makes sense. While the CCP talks about AI risks, in reality, it just wants the U.S. to slow down. For a regime that even covered up the pandemic, how could the outside world possibly verify whether it has actually slowed down? And this is precisely what the Trump administration is most worried about, the U.S. tying its own hands while its rival secretly accelerates. So, friends, do you agree with the Trump administration's stance that we must never slow down and must beat the CCP? If you agree, give this a like to let me know. As for Vice President Vance's perspective, that's on a different level. I have to say, Vance is very sharp. He hit the nail on the head by pointing out this abnormal phenomenon. Why are these tech giants, who usually hate being restricted the most, now lining up to run to Washington, crying, and begging the government to regulate them? He said there's a business calculation behind this. The giants are never afraid of regulation. What they are truly afraid of is having no barriers to entry. What do I mean by that? The big tech companies have hundreds of billions in capital and top-tier legal teams. They can handle even the strictest, most insane compliance standards. But those startups and open-source developers have a hard time handling them. Those high compliance costs and approval processes could potentially strangle any up-and-coming competitors right in the cradle. So Vance believes that by proactively calling for regulation now, they are essentially trying to participate in rulemaking. By sitting at the table where the cards are dealt, they can set the barrier to entry at a level that is most favorable to them, but acts as an insurmountable wall for everyone else. In other words, this group of frontrunners has realized that the footsteps of those chasing them are getting too close, so their most effective defensive strategy is to use regulation to hit the pause button, buying themselves the more generous window of time. It's a gift on the surface, but with a dagger hidden inside? So this is a Trojan horse. According to a report by Wired, OpenAI even consulted with members of Congress about whether a collective industry slowdown would violate antitrust laws. This is very significant. How should we interpret this? How do we understand it? Friends, let me explain. In a free market, if the top industry giants privately agree, hey, let's all slow down, control prices, and start releasing new products starting tomorrow, that is legally considered an illegal cartel and collusion, which is a serious violation of the law. So OpenAI went to ask lawmakers, is this illegal? And if it is illegal, then what? Then they have to drag the government into it. As long as they get the government to step in, issue regulatory policies, and forcibly require everyone to slow down collectively, well, then the responsibility, the responsibility for this illegal collusion won't be on them. They can say, it was the government that required us to slow down. Don't blame us. And this way, it no longer constitutes an illegal collective monopoly. So the Trump administration is currently saying that this is a scam, a conspiracy. Behind this, these are the two main arguments. These are the two reasons. What do you think, friends? Now that you've heard the whole story from beginning to end, which side are you leaning towards? If you are more inclined to believe that the four major AI giants collectively calling for a halt is truly out of concern that AI is getting out of control and that it's time to slow down, type genuine concern for me. If you're more inclined to agree with the Trump administration's stance, feeling that this is a scam and that behind their collective slowdown is a monopolistic conspiracy to use the government's hand to block competitors, type a total scam for me. Let's see how our friends judge this. Finally, I'd like to invite you all to chat in the comments section. Do you think AI will one day get out of control or even threaten humanity? Please be sure to share your insights with me in the comments. Alright, that's all for today. Thank you all for watching. If you enjoyed my show, don't forget to give me a like before you leave. If you're new here, welcome to subscribe to TrueView so that you can see my new videos on your YouTube homepage in the future. Some friends have also reported that many of their previous subscriptions have been unsubscribed for some reason. So for those who have subscribed, please keep an eye on that. And don't forget the fan promotion feature, which can help my videos spread further and wider. How will this drama in the AI world unfold from here? We will continue to follow it. So if you like this kind of content, be sure to let me know in the comments section. I will choose topics to share with you based on your enthusiasm. Alright, that's all for today. Thank you all for watching. Don't forget, we'll see you again at the same time this Thursday on TrueView. Bye!