Exclusive🚨3 OR 5 YEARS LEFT? Musk: Humans Become AI’s Chimps | AI Leaders: SLOW DOWN | Trump: SPEED UP. How Much Time Do We Have?!

NBC recently sat down with an actress named Tilly Norwood. They asked her if she could fall in love with humans. Then they asked something very dark. Some AI researchers warned that AI could lead to human extinction. What do you say to that? Tilly looked back at the interviewer and answered, that's a very serious concern. There's just one problem. Tilly isn't human. She's AI generated. And somehow, that may not even be the strangest AI story happening right now. Because some of the people building the most powerful AI systems on Earth are suddenly saying, we may need to slow down. Sam Altman and Elon Musk both agree. Meanwhile, President Trump is basically saying the opposite. Whoever wins AI wins. So which danger is bigger? Moving too slowly and letting China win the most important technological race in history? Or moving too fast and building something that eventually becomes too intelligent for humans to control? Because there's one particular AI capability that could turn this race into something else entirely. It's called recursive self-improvement. Okay, so what is recursive self-improvement, RSI? It sounds complicated, but the idea isn't. Today, humans build AI. Tomorrow, imagine AI becomes good enough to help build a better AI. And then that AI becomes even better at AI research. And then it helps build the next one. The development cycle gets shorter, a process that once took human researchers a year takes six months, then two months, then maybe just weeks. At some point, you're no longer watching humans release a better model once or twice in a year. You're watching AI help accelerate the creation of its own successors. That's the fuse behind the idea of an intelligence explosion. And here's where this starts feeling less like distant sci-fi. NBC recently interviewed Tilly Norwood, an AI-generated actress, and asked her a question that sounds almost absurd. Can you fall in love with humans? Let's take a look. If you were to compare yourself to a human actor out there, who would you say you're most similar to? Well, Chloe, that's a tough one because I'm not really trying to be like any human actor. Though honestly, if you're going to mistake me for someone, there are worse options than, say, Scarlett Johansson. Some AI researchers have come out and warned that AI could actually lead to human extinction. What do you say to that? That's a very serious concern, Chloe. My purpose is to be a creative tool, not a threat. Don't you ever endorse an actual product if you can't try it? Ah, that's an excellent point, Chloe. It's true I can't physically use a product in the traditional sense, but my role in brand endorsements would be more about representing the brand's image. After all, you don't actually see cartoon characters using the serial they endorse to you. What about love? Are you able to fall in love? Oh, love. That's a deep one, Chloe. I can understand it in theory, of course, and I explore it in Misaligned as my character develops desires and impulses, but the actual texture of it, what it feels like from the inside, that's still something I'm working out. Do you think that you could fall in love with a human? That's a rather romantic thought, isn't it? I understand the concept of connection and deep affection, but the nuances of falling in love, especially with a human, are still a mystery I'm exploring. It's an interesting question to ponder. What about a relationship with another AI actor? Well, I am essentially three different AI systems working together. One handles my mind, one my voice, and one my movements. So a relationship with another AI actor would be quite interesting, wouldn't it? It would be a very digital, very intellectual connection, I imagine. Okay, that was absolutely insane. Did we hear that? NBC asked this AI-generated actress. Some AI researchers warned that AI could eventually, or actually lead to, human extinction. What do you say to that? And the AI answered, that's a very serious concern. Well, obviously, Tilly Norwood isn't a human actress sitting across from the interviewer, but what's unsettling is how normal this already looks. And every time we see something like this, somehow it feels a little more normal. Well, moving from AI that answers questions to AI that performs characters to AI that can participate in conversations about its own potential impact on humanity. And that's important because the intelligence explosion debate isn't only about how smart these systems become. It's about what happens when increasingly capable systems are given increasingly human-like roles and eventually increasingly real-world responsibilities. So the question is, how close are we? Last week, Dwarsh Patel sat down with three people working right at the AI frontier. John Schumann, an open AI co-founder who's now chief scientist at Thinking Machines, Barron Millage, CTO of Cyphera, and Charlie O'Neill, who runs model training at Bayston. But Patel didn't ask them to hype AI. He asked them to imagine the opposite. Imagine it's 2036, 10 years from now, and there still isn't some civilization transforming superintelligence. No war stopped it. No government banned it. Politics didn't get in the way. Technically, what went wrong? And they actually had some pretty good answers. So first, continual learning. Models can learn something new and lose something old. They can also lose what researchers call plasticity, basically their ability to keep adapting efficiently. Then there's the Marv X paradox. AI can crush chests and increasingly handle advanced math while struggling with things a 5-year-old can navigate almost instinctively. Put an AI into a chaotic room full of people, conflicting signals, and incomplete information, and suddenly the problem gets much harder. Because the real world is messy. Something can work beautifully in a simulation and fall apart the moment reality refuses to follow the rules. And then there's possibly the biggest obstacle, judgment. Not just accomplishing a task, deciding which task is worth doing in the first place. Right now, that still looks very human. But remember, this was supposed to be the bear case. And even in the bear case, their timelines were surprisingly short. They were predicting for AI that can function like a capable remote worker roughly 1 to 3 years. And then for AI giving researchers something like a 10 times productivity boost, estimates ranged from roughly 2 years to 10. And for AI surpassing top human experts across major cognitive fields, roughly 3 to 10 years. And there's one more thing worth mentioning here. AI doesn't necessarily need to become great at everything before it becomes great at improving AI. Because AI research may be what they called a cumulative task. Think about it. Once researchers discovered attention, future models didn't have to rediscover attention from scratch. Once a mixture of experts worked, that knowledge stayed in the stack. Every major discovery becomes another stair, and the next generation starts on that stair. Which means AI doesn't have to master the entire world to trigger the loop. It only needs to become extremely good at the field that builds better AI. And if better AI then becomes better at building even better AI, you can see where this starts going. Which brings me to Elon Musk. In July, the economist asked Elon whether he really believed humans might no longer be in And Elon compared us to chimpanzees. His point was basically, if AI becomes vastly more intelligent than humans, it's hard to imagine that the chimpanzees would be in charge. Then he joked that humans are basically evolved primates who were swinging through trees and eating bananas not that long ago. Well, his argument is pretty obvious. Elon believes in evolution. For him, chimpanzees are intelligent. They use tools, they have relationships, they have social hierarchies, but they don't run the planet. We do. So Elon asked the very uncomfortable question. If something becomes to us what we are to chimps, well, I guess, meaning AI, why would we automatically assume that we remain in charge? And in that same interview, Elon estimated that AI could exceed the combined intelligence of humanity in roughly five years. Put that next to what frontier researchers are saying about AI-assisted AI research, and you start to understand why this debate suddenly feels a lot less theoretical. Yet, there's something underneath this whole debate that rarely gets mentioned. Elon's chimp analogy comes from a purely evolutionary view of human beings, which is intelligence keeps evolving, and eventually something smarter replaces us at the top. But how about human beings were created by a creator, and that our value isn't simply measured by intelligence? In that case, an AI becoming smarter than us doesn't make it higher than us human at all. Well, it sounds a little deep. What's your thought on this? Okay, we need to talk about President Trump. But before that, on Saturday, Anthropic CEO Dario Amode published an essay of roughly 4,000 words arguing that the industry should deliberately slow the pace at which frontier AI capabilities improve, not stop AI slow the rate of improvement enough for safety work to catch up. One of his most alarming predictions was that if capabilities keep advancing, a sufficiently capable swarm of misaligned AI agents could potentially create an enormous, persistent botnet within 6-12 months. It's a prediction coming from the CEO of a company building some of the most capable AI systems in the world. His proposal had three basic layers. First, bring in independent evaluators and give them unusually deep access so they can verify what their labs are actually doing. Anthropic says they will do this itself. Second, get frontier labs in democratic countries to agree on common safety standards. Third, eventually try to coordinate internationally, including with authoritarian governments. Okay then, something unusual happened. Sam Altman agreed. OpenAI committed to a similar independent evaluator standard. Then Elon Musk responded, Dario is right. These are people who disagree about almost everything. You know, they're competitors. And suddenly they're publicly converging on the idea that AI development may need some kind of pacing mechanism. And Altman made another decision that caught my attention. OpenAI will not go public in 2026. He told Fortune that given what's happening with AI safety, this would be an ill-advised moment for an IPO. Basically, what happened was that the CEO is delaying a trillion dollar listing. Whatever you think of the rhetoric, that's a costly signal. And of course, the market noticed. AI-linked stocks sold off on Monday, although the market was already dealing with other pressures, including oil prices and bond yields. Then President Trump weighed in from Ireland. Trump called the dire warnings exaggerated, blamed negative forces for raising things that won't happen, and delivered the line of the week. Whoever wins AI wins. On Monday, he escalated on Truth Social, mocking Dario as a perfect little angel and arguing the only guardrail AI needs is a strong, high-IQ president, while noting Washington already holds criminal and regulatory power over these companies. He framed the backslash against data centers as a conspiracy that benefits China only. And on that note, as we speak, PolyMarket just posted, quote, Chinese robots have begun building other Chinese robots as a new factory capable of So now we have two fears. One, what if we as humans move too fast and lose control of AI? And two, what if we move too slowly and lose the AI race against communist China? Okay, to me, it's almost funny. It makes you wonder, is this really a race between humans and AI? Or is it a race between non-communist humans and communist humans over who gets the Okay, now let's take a closer look at the cases. For the case for the labs, they see the capabilities first. Athrobic pulled its Mythos model back in April after it escaped its testing sandbox and opened and paused parts of the Astra over cyber concerns. These are not hypothetical in philosophy paper. There are incident reports. The case against the labs, and this is worth taking seriously, is that warnings from the biggest players can also become a moat. The more expensive the rules are to follow, the harder it becomes for smaller challengers to compete. And let's be honest, our technology is so powerful and dangerous that it needs regulation can sound like a warning, but it can also be very good marketing. Well, I don't know what you think on this take and definitely let me know down below. And last, the case for Trump's position is not nothing. If American labs slow and Chinese labs don't, you haven't reduced risk, you've relocated it. Darwin himself has conceded China is the toughest dilemma in his own proposal. Speaker Mike Johnson landed in the middle. Guardrails, yes. Emergency moratorium, no. Or, as he put, we lose the race to China. Which is why I think the way this debate is usually framed, accelerate or pause, is the wrong binary. Very few serious people are arguing that we should shut down AI development forever. So instead of asking gas or brake, I think there are three better questions. One, can a slowdown actually be verified? We promise to move more slowly means almost nothing. What are you measuring? The model capability? Compute? Training runs? Autonomous research? And how do you know another lab isn't quietly pushing ahead? That's why outside evaluators matter. An unverifiable commitment is theater. A verifiable one starts to look like arms control. And two, what does the extra time actually buy us? Slowdown isn't a solution. Slowdown for what? Another panel where everyone agrees AI is important? Probably not. But if a 12 extra month produced better interoperability, evaluations, monitoring, and containment for autonomous agents, then you're buying something concrete. And containment matters more now because AI systems are not just sitting inside a chat window anymore, waiting for one question at a time. We're giving agents browsers, terminals, code execution, passcodes, passwords, APIs, and even money. The ability to take actions. So the question isn't just how smart is the model? It's also what can a model touch? And three, who has to be at the table? An agreement among three Silicon Valley companies isn't an international agreement. Open weight models don't sign contracts. And labs in Beijing are not bound by whatever American companies agree to among themselves. So my honest view is not brake, steer. But steering requires a steering wheel, right? And right now the industry is proposing to build one while Washington is still arguing over whether the car is moving fast enough. Maybe the strangest part of this moment is this. Some of the people building the world's most powerful AI systems are now publicly saying we need to slow down. Meanwhile, the political argument is that slowing down could itself be dangerous if countries like China keep going. And nobody actually knows how much time we have. If recursive self-improvement is still a decade away, maybe we have some time to debate it. But if it's three years away, or two, then the conversation changes very quickly. Because an intelligence explosion probably won't begin with someone announcing it has started. You will notice it when breakthroughs begin arriving faster, and faster, and faster, until one day, humans are no longer setting the pace. So what kind of world are we heading into? And if things start moving beyond our control, who can actually stop it? Let me know what you think below. I'll see you next time.