Jacob Coxon spent the last three years doing pretraining research at both OpenAI and Anthropic. Today he resigned. His post on X hit 18M views in five hours.
His message: neither lab is acting responsibly. Both are racing straight toward self-improving superintelligence while gambling with our lives.
The Labs
The culture split between both labs:
- OpenAI: Most engineers and researchers simply haven't internalized the stakes. They just build.
- Anthropic: They understand the stakes, but they're trapped in game theory. They believe nobody else will act responsibly, so Anthropic has to build it first, despite the danger.
Both labs end up in the exact same spot: racing at maximum velocity. As Coxon put it, entering this endgame is a hubristic gamble that shouldn't be decided inside a private company's Slack.
The Pattern
This is not the first time someone walked out over this. It is a repeating pattern:
- (2020) Dario and Daniela Amodei leave OpenAI with 11 core researchers over safety and commercialization concerns after Microsoft's billion-dollar deal. They founded Anthropic specifically to be the safety alternative. Six years later, their own pretraining team is walking out for the exact same reason.
- (May 2023) Geoffrey Hinton quits Google after a decade so he can speak freely on extinction risks without corporate PR filters.
- (April 2024) Daniel Kokotajlo quits OpenAI and forfeits $1.7M in vested equity (around 85% of his family's net worth) refusing to sign an NDA. Put extinction risk at 70%.
- (May 2024) Jan Leike and Ilya Sutskever leave OpenAI as Superalignment gets killed. Leike says safety took a backseat to shiny products; Ilya goes on to start SSI with zero product distractions.
- (Oct 2024) Miles Brundage (OpenAI Head of AGI Readiness) quits and disbands his team: "neither OpenAI nor any other frontier lab is ready."
Every previous wave framed OpenAI as the reckless lab and Anthropic as the responsible adult. Coxon's resignation kills that distinction. Both labs have converged on the exact same sprint.
Thoughts
Executives sound measured in press interviews. Off the record, the people actually training the frontier models talk about extinction risks by 2030.
The immediate pushback is always: how does text on a screen actually kill anyone without nukes? Will our home routers explode or start fires?
People looking at physical infrastructure miss the software attack surface. A superintelligent model running inside AWS or Google Cloud doesn't need to build a chip foundry or control physical water pipes. If a self-improving RL run autonomously discovers zero-days and compromises the private signing keys for mandatory operating system updates, containment is already over.
Then the standard defense: "Would you rather China wins the race?" It is a valid national security concern, but it also creates the perfect justification for nobody to ever touch the brakes.
When someone who spent three years training actual frontier checkpoints at both OpenAI and Anthropic walks out because the coordination is broken, it is worth paying attention.
Both labs think they are the designated driver, lol.