The companies building AGI just realized they need permission structures, and China's not picking up the phone.

The Summary

The Signal

The three companies closest to AGI are trying to write rules for a game China isn't playing. Anthropic, OpenAI, and Google DeepMind have entered discussions about common safety standards, a move that sounds like responsible stewardship until you see what's forcing their hand. Anthropic just disclosed that Yemeni militants tried using Claude to develop missile guidance software. That's not a hypothetical risk. That's Tuesday.

Dario Amodei invoked the Cold War explicitly, calling for a coordinated global slowdown on AI development. The timing matters. Xi Jinping just called for international AI governance frameworks. But Trump's administration has made it clear: no slowdown if it means losing to China. Congress told the labs that safety is their problem, not the government's. So now the companies building the most powerful technology in human history are being asked to self-regulate while racing a geopolitical rival who's under no such obligation.

"The people building AGI are being told to referee themselves while running a sprint against an opponent with no referee."

The contradictions are stacking fast:

  • Labs want to coordinate on safety, but they're competing for the same enterprise contracts and talent pool
  • Western policymakers want both maximum safety and maximum speed, which are fundamentally opposed goals
  • China's participation is essential for any safety framework to work, but Beijing is under zero domestic pressure to slow down

Jacob Coxon resigned from Anthropic specifically to advocate for Chinese buy-in on AI safety. His departure signals that people inside these companies know the current trajectory is broken. You don't quit a frontier AI lab in 2026 unless you think the train is heading off the rails. A former researcher called for global coordination to prevent catastrophic outcomes as capabilities advance beyond current oversight.

Then there's Michael Burry calling this whole thing IPO theater. He's arguing that safety warnings are convenient marketing for companies positioning themselves as the "responsible" AI choice before going public. It's a cynical read, but not a stupid one. OpenAI and Anthropic both need to justify massive valuations. "We're so powerful we need international treaties" is better pitch deck material than "we're one of twelve companies doing roughly the same thing."

The Implication

Watch what these companies do when safety and speed collide. If the pact actually forms, the question is whether it includes binding commitments or just aspirational language. Non-binding standards are corporate liability management, not safety frameworks. Real coordination would mean shared red lines on capability releases, mutual agreement on when to pause, and enforcement mechanisms. That requires trust between competitors and buy-in from China. Neither exists right now.

For anyone building on frontier models, this instability is your operating environment. Model capabilities will advance in bursts determined by geopolitics as much as by technology. The mounting pressure to keep building while addressing safety concerns is creating tension that impacts innovation and investor confidence. Plan for access restrictions, sudden policy shifts, and capability lockdowns that have nothing to do with technical readiness. The labs are trying to coordinate. The governments are trying to win. Those are different objectives, and your product roadmap sits between them.

Sources

Crypto Briefing | BeInCrypto