--- title: 'Anthropic is starting to panic…' source: 'https://youtube.com/watch?v=0pgCBV8CTZY' video_id: '0pgCBV8CTZY' date: 2026-06-28 duration_sec: 303 --- # Anthropic is starting to panic… > Source: [Anthropic is starting to panic…](https://youtube.com/watch?v=0pgCBV8CTZY) ## Summary Anthropic, the company behind Claude, filed for a trillion-dollar IPO while simultaneously proposing a global pause on AI development due to fears of recursive self-improvement. The video explores three possible futures: AI takeover, economic collapse from automation, or a disappointing plateau where AI never lives up to the hype. It also discusses historical precedents and recent evidence questioning AI's actual value. ### Key Points - **Anthropic's IPO and Pause Proposal** [0:00] — Anthropic filed for a trillion-dollar IPO and proposed a global AI pause due to recursive self-improvement risks. - **Historical Precedent** [1:40] — OpenAI used similar 'too dangerous' rhetoric before releasing GPT-2 in 2019, which turned out fine. - **AI Capabilities** [2:02] — Claude Mythos is better than human researchers 64% of the time, and AI solved an 80-year-old math problem. - **Economic Death Spiral** [2:38] — Economists predict an 'AI layoff trap' where automation reduces demand, leading to economic collapse. - **AI ROI Failure** [4:03] — MIT study: 95% of enterprise AI projects delivered zero measurable ROI despite $30 billion spending. ## Transcript Last week, Anthropic officially became the Apex Alpha company of the artificial intelligence race with a valuation exceeding OpenAI as they filed to go public with their trillion dollar IPO later this year. If you're a software engineer, this comes as no surprise because Claude has been the best AI programmer for years now. But despite the billions of dollars flowing into this company right now, they also just proposed something that sounds insane. Maybe we should wait a second and pause all AI development because AI is getting dangerously close to recursive self-improvement. And when that happens, the last thing humanity ever builds is the thing that realizes it doesn't need humanity. Even if AI is benevolent and doesn't go rogue and kill us all, a new paper just dropped that believes all rational firms will automate each other into a death spiral anyway. It looks like we're screwed no matter what we do. But not everybody out there is a doomer. And I found some compelling evidence that AI might actually kind of suck. It is June 9th, 2026, and you're watching the code report. Anthropic's in-house think tank just dropped a report, and this is the thesis. AI is getting dangerously close to recursive self-improvement. In other words, they're smart enough to rewrite their code and upgrade themselves in a loop with no humans necessary. The entire industry is all gas and no breaks, and they want everybody to come together and hold hands and create a brake pedal. The problem is that Anthropic can't pause alone. While OpenAI, Deep Mind, and XAI keep sprinting. So, unilateral pausing is off the table. It's either everybody or nobody. And that includes China, by the way. And nobody's really worried about the EU. However, a global pause is a very convenient thing for the market leader to advocate for. Because it doesn't erase Anthropic's lead, it freezes it right as they're about to make billions of dollars with an IPO. If all this sounds familiar, it's not because you're crazy. It's because in 2019, OpenAI did the same thing before the release of GPT2. At first, they said it was too dangerous to release it, just like Anthropic is doing right now with Claude Mythos. But then they released GPT2 and it was totally fine. That was 7 years ago, and now it looks like ancient technology. Today, if we are to trust these Trust Me Bro benchmarks, the modern Claude models are far better at research than humans. Like 64% of the time, a Claude mythos is better than a human every time. On top of that, AI researchers are now solving problems that humans haven't been able to. Like OpenAI recently disproved a central conjecture and discrete geometry, which mathematicians have failed to do for the last 80 years. The scary thing is that we're already giving AI access to data centers, robots, and weapons to blow people up. And thanks to predictive programming in Hollywood movies, we all know how that story ends. It's either enslavement like the Matrix or extermination like Terminator. I'd prefer the latter, but there's a possibility for an even dumber outcome. as predicted by economists from Boston University in their paper, the AI layoff trap. At this point, there's been tens of thousands of layoffs in tech thanks to AI. But these economists did some math, and it doesn't look good. Because when a firm automates away a worker with AI, it pockets 100% of the savings. But the laid-off worker is also a customer. And their loss spending doesn't just hurt the firm that fired them, it hurts everyone selling anything. Demand goes down. So the endgame is that firms automate their way to infinite productivity and zero demand. They also argue that things like UBI and upskilling aren't going to work and the only solution is to put a tax on automation. It kind of like the same way we tax pollution, making it cost more to fire people. So the math stops rewarding the AI race. But if there's one thing I've learned about economists, it's that they're wrong pretty much every time. A third possibility is that AI just isn't nearly as good as people think and never will be. This is the Wall-E situation where we keep chasing more and more automation and ultimately destroy the planet by building more and more data centers. One piece of evidence that supports that outcome is that over the last couple years with the rise of Agentic AI, the number of new app releases on the iOS app store is nearly doubled. However, it appears nobody's actually using these apps because app reviews and apps with significant usage are declining. In addition, this 2025 report from MIT analyzed over 300 enterprises implementing AI. And even though they spent over $30 billion collectively, the end result was that 95% of their projects delivered zero measurable revenue impact or return on investment. That doesn't look good. But luckily, there are tools that can help you avoid failure, like Pioneer, the sponsor of today's video. If you're calling a Frontier model for every LLM request in your app, it's probably burning through a bonfire of tokens just to return generic results. But Pioneer solves this by giving you an inference API that you can plug into your existing LLM setup to handle all model routing and optimization for you. And it will cluster your app's traffic by use case to discover where your current model is being too slow, too expensive, or too stupid. Then it trains a fleet of smaller open- source models in the background and alerts you when it finds one that's cheaper and better so you can easily swap it under the hood. But Pioneer works great with Claude Code, Codeex, Cursor, Hermes, or anything else hitting an LLM endpoint. And you can get $30 of inference for just $5 today. But this has been the code report. Thanks for watching and I will see you in the next one.