The people building the technology are now publicly betting against the species.
The Summary
- Former OpenAI and Google DeepMind researchers released a dozen video interviews estimating extinction-level AI risk between 10-50%, published by non-profit Palisade Research at frominside.ai
- The interviewees aren't fringe doomsayers — they're current and former employees from OpenAI, Google, and Anthropic who've worked on frontier models
- Geoffrey Irving, ex-OpenAI and DeepMind: "The chance of human extinction is about a coin flip, in my view"
- Neel Nanda, current Google DeepMind research scientist, puts it at "at least a 10 percent chance"
The Signal
These aren't external critics or philosophy professors theorizing about paperclip maximizers. These are people who've spent years inside the labs where GPT-4, Claude, and Gemini were built. They sat in the architecture meetings. They trained the models. They saw the unexpected capabilities emerge. And now they're going on camera to say there's a meaningful probability that their work kills everyone.
The timing matters. We're in the middle of a race to AGI that's being conducted in public. OpenAI just raised billions at a $150B valuation. Anthropic is scaling Claude. Google is integrating Gemini across every product line. The capital is committed, the infrastructure is being built, and the competition is driving everyone to move faster.
"The people with the deepest technical understanding are now publicly hedging the species."
The Palisade Research interviews represent something new: organized whistleblowing by people still in the industry. Not think pieces or academic papers, but direct-to-camera statements from researchers who can't be easily dismissed as outsiders or Luddites. Irving's "coin flip" comment is particularly stark because of his pedigree. This isn't someone on the sidelines. This is someone who helped build the thing.
The range of estimates (10-50%) tells you something important: even the experts building this technology don't agree on the probability distribution. But they all agree the tail risk is non-zero and potentially catastrophic. That's not how risks usually get managed in Silicon Valley. You don't ship products when your senior engineers think there's a 1 in 10 chance they end civilization.
What this means for the agent economy:
- If 10-50% extinction risk is the insider consensus, every agent deployment is happening under a probability distribution that includes civilizational collapse
- Companies building autonomous AI systems are making capital allocation decisions without pricing in existential risk
- The gap between public AI safety rhetoric and private probability assessments is wider than most people realize
The Implication
The coordination problem is getting worse, not better. Every lab knows that slowing down unilaterally means losing to competitors. Every government knows that restricting domestic AI development means ceding advantage to adversaries. The only way out is multilateral coordination, and we're moving in the opposite direction.
If you're building in the agent economy, you're building under a new assumption: the people creating the underlying technology think there's a material chance it ends badly. That doesn't mean stop building. It means build with your eyes open. Watch for inflection points. Have exit strategies. And push for the kind of oversight that makes coin-flip extinction odds politically untenable.