The people building the models say privately they might kill everyone by 2036, then go on CNBC and talk about productivity gains.
The Summary
- Jacob Coxon quit Anthropic after three years split between OpenAI and Anthropic, saying both companies are "racing straight to self-improving superintelligence and gambling with our lives"
- AI researchers genuinely believe their work could kill everyone within the next decade, but couch their public statements to "sound sensible"
- This is the latest in a string of safety-focused resignations from leading AI labs, signaling internal alarm isn't just about philosophical risk
- Coxon worked on GPT-4o at OpenAI before moving to Anthropic in July 2026, giving him front-row seats to both operations
The Signal
Jacob Coxon spent three years building the models that might build the next models that might decide humans are inefficient. He worked on GPT-4o at OpenAI as a technical staff member from 2023 to July 2026, then moved to Anthropic as a researcher. Now he's out, and his exit note reads like someone who watched the brakes fail in real time.
"Neither company is acting responsibly," Coxon wrote on X. "They are racing straight to self-improving superintelligence and gambling with our lives." Not former employee shade. Not negotiating leverage. A researcher who helped build GPT-4o saying the people running the show are reckless.
"The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt."
The gap between what AI lab leadership says in boardrooms versus what they say on earnings calls is the story here. Coxon claims executives and senior researchers deliberately soften their language for public consumption, expressing genuine fear in private while sounding measured in the press. You get CNBC interviews about enterprise productivity tools. Behind closed doors, you get timeline estimates for existential risk that end before the next presidential election.
This isn't the first safety-driven resignation from a major lab. It's part of a pattern. OpenAI and Anthropic have both lost researchers over concerns about how fast they're moving versus how much they understand about what they're building. But Coxon's departure carries extra weight because he's seen both operations from the inside. He worked at the company Sam Altman runs and the company Dario Amodei started specifically because he thought OpenAI was moving too fast. His assessment: neither one is getting it right.
Key differences Coxon noted:
- At OpenAI, many researchers haven't deeply internalized the risks they're creating
- At Anthropic, people understand the danger but are building anyway
- Both are racing toward self-improving superintelligence without adequate safeguards
The "self-improving superintelligence" framing matters. That's not AGI that needs human guidance to get smarter. That's a system that recursively improves itself, potentially at speeds that make human intervention meaningless. The timeline Coxon and his colleagues discuss privately: end of the decade. That's 2030. Maybe 2029. The same window where OpenAI claims it will reach AGI and Anthropic is targeting its most capable models.
The Implication
If you're building on OpenAI or Anthropic APIs, you're building on infrastructure designed by people who privately think it might kill everyone in ten years but are shipping it anyway. That's not a reason to stop building. It's a reason to watch what the researchers do, not what the press releases say. Coxon just told you the gap between private fear and public messaging is real. When the next model drops with a cheerful blog post about capabilities, remember someone who helped build the last one thought the whole trajectory was gambling with extinction.
Watch for more resignations. One researcher leaving is a data point. Three is a trend. Five is a migration pattern that tells you something broke internally.