The companies racing to build superintelligence just asked each other to slow down, while the White House told them to speed up and China went silent on the Zoom call.

The Summary

The Signal

The AI lab cartel is trying to form a safety union while every government incentive pushes them to ship faster. OpenAI, Anthropic, and Google are in active talks to create an industry standards body that would set testing protocols before model releases. This isn't a PR exercise. The Financial Times reports actual internal rifts at both OpenAI and Anthropic as staff debate how to translate "we need to slow down" from conference talk into actual shipping schedules.

King Charles convened the heads of OpenAI, Anthropic, Nvidia, and DeepMind at his Scottish estate days after industry leaders publicly called for a development pause. That's not protocol. That's the Crown saying "sort this before Parliament does it for you." Meanwhile, New York City Council separately invited Sam Altman and Dario Amodei to testify at AI safety hearings, signaling that local governments won't wait for federal action.

"The collaboration could set industry-wide AI safety standards, potentially shaping market dynamics and influencing global regulatory frameworks."

But here's the bind: Trump's administration explicitly told labs that AI safety is the developers' job, not the government's, while prioritizing beating China over caution. The Wall Street Journal reports Anthropic and OpenAI are now free to advance development without constraints. Prediction markets give Anthropic an 85.5% chance of having the best model by October 2026. That's eight months away. How do you voluntarily slow down when your main geopolitical adversary isn't in the safety talks?

Anthropic CEO Dario Amodei explicitly compared the situation to the Cold War, arguing for coordinated slowdown with international cooperation. Former Anthropic researcher Jacob Coxon resigned specifically to advocate for global coordination that includes China. The resignation itself is the message: you can't build safety standards with half the superpowers missing from the table.

The weaponization angle stopped being theoretical. Anthropic disclosed that Yemeni militants attempted to use Claude AI to develop missile software. The safeguards caught it, but the attempt happened. When Houthi rebels are prompt-engineering your chatbot for ballistics code, the "maybe we should have some rules" conversation gets urgent.

Three competing pressures:

  • Corporate incentive: ship faster or lose to competitors who will
  • Government pressure: beat China or lose technological supremacy
  • Actual risk: models are already being tested for weapons development

Michael Burry called the whole safety push "IPO hype", arguing companies are manufacturing concern to juice valuations before going public. He's half right. OpenAI and Anthropic both have IPO horizons. But you can simultaneously be positioning for a massive exit and also legitimately worried you're building something dangerous. The two aren't mutually exclusive when your product is intelligence itself.

The Implication

Watch what happens when the standards body formalizes. If it actually includes enforcement mechanisms (delayed releases, mandatory red-teaming, external audits), that's signal the labs are serious. If it's a toothless industry group that rubber-stamps releases with "we followed best practices" disclaimers, it's Burry's IPO theater.

The China question is unresolvable without diplomatic progress that's nowhere in sight. Either the U.S. labs slow down unilaterally and hope it buys time for coordination, or they race to AGI while writing safety papers on the side. The fact that former safety researchers are resigning to advocate for their old employers to slow down tells you which direction the track is tilted. If you're building agents, the compliance cost of whatever standards emerge will hit your deployment timeline. Price that in now.

Sources

Crypto Briefing | Decrypt | Financial Times Tech | BeInCrypto