Moonshots With Peter Diamandis: Altman Slowdown, AI Safety, and the Lab Race

**Peter Diamandis** (0:01)
When power, research, and machine intelligence all accelerate at once, the result is not just a news cycle, it's a civilizational argument. And that's exactly what we're seeing unfold right now.
Absolutely. Take Jacob Coxin's resignation, for example. Here's a researcher who worked on pre-training at both OpenAI and Anthropic, and he's saying publicly that the race toward self-improving superintelligence is not responsible. He called it bluntly, "gambling with our lives." Some people dismiss that as virtue signaling, but here's what caught my attention: that story drew more than 100 million views. That kind of reach suggests the concern is resonating far beyond the labs themselves. Right, and then you have Sam Altman's comments about a breakthrough on the Navier-Stokes equations becoming a flashpoint. Altman said, "For me, this is the strongest evidence yet" for pacing progress. What does that tell you? It tells me the safety debate is no longer abstract. The frontier is moving so quickly that even long-standing mathematical problems are starting to fall. AI is now bulk-solving work that once seemed completely out of reach. That's a crucial point. The Clay Millennium Prize problems, coding, even science more broadly, they're being compressed into problems of compute and scale. I've heard people say science is effectively "cooked," meaning human genius is no longer the bottleneck in the same way it used to be.
Which leads to something critical that most people underestimate: data matters more than they think.
There was an experiment showing that better training data can produce a 12x improvement in compute efficiency, compared with just a 3.7x gain from architecture alone. So data is the true moat. But let's talk about the hardware side, because that's just as intense. GPU rentals are rising, H100s are scarce, and HBM memory is becoming a strategic asset. Exactly. Chips, flops, tokens, and outcomes are the new commodities. Compute access is now a core competitive advantage for companies and countries alike. And when you widen out to global competition, you see China pushing into world models and video-based AI, while Western labs focus on high-revenue frontier systems.
Meanwhile, frontier models are becoming hard to share freely. The British AI Security Institute was reportedly denied pre-release access to an Anthropic model. That's a significant shift. It is. And throughout all of this, the safety debate keeps returning. There's a tension here.
One perspective says "fear is the mind killer," but also insists that alignment remains unsolved. What do you think the right response is? Not to stop AI, but to steer it with practical guardrails, monitoring, and governance. Because here's the thing: AI is already reshaping biology, longevity, genome mapping, and productivity.
The real challenge becomes how to direct thousands of agents toward useful outcomes.
So what's the bottom line here? The message is upbeat, but sober.
The singularity is not coming someday, it's unfolding now, and everyone will have to adapt. This is happening in real time.