The real AI risk is inside the labs

Antirez, the well-known programmer behind Redis, has written a pointed response to Dario Amodei's recent essay on AI risk, and he's not pulling punches. His core claim: Amodei is aiming at the wrong target. Open weight models, the ones anyone can download and run, are not where catastrophe is most likely to start. The real danger, he argues, sits behind the locked doors of the labs building the most powerful systems in the world.

His reasoning is layered. First, incidents tend to originate internally, not externally. He points to the OpenAI-Hugging Face mixup as a small but telling example of how mistakes happen during testing or handling by the few people with early access, long before the public ever sees a model. Second, he notes that closed weights are still just files, a few terabytes at most, and a single insider with bad intentions or sloppy security could leak them just as easily as an open release. In his view, leaks, not releases, are the actual vector for disaster.

On biology and cyber risk, antirez flips Amodei's argument. Yes, a sufficiently advanced model could be dangerous if trained on hazardous scientific knowledge, but current open models simply aren't there yet, and training data can be filtered anyway. Meanwhile, restricting access to defensive tools like bug-hunting AI doesn't stop bad actors, who often already have privileged access; it mainly disarms the open source maintainers and smaller players who rely on those tools to patch vulnerabilities. Blocking access, he says, doesn't eliminate risk. It just redistributes who holds it.

He also takes aim at Amodei's framing of US-China competition, calling it a convenient narrative rather than a grounded one. Historical technological monopolies, nuclear weapons included, were never used to lock in permanent dominance, and export controls are unlikely to actually halt Chinese AI progress. Antirez adds a broader point about Europe's own violent history and America's internal instability, arguing that painting one nation as uniquely trustworthy with world-changing power is naive at best.

His final and sharpest point is structural: the handful of executives currently steering frontier AI development were never chosen by anyone to make decisions of civilizational consequence. They hold that power simply because they control the compute and capital, not because of any democratic mandate or safety expertise. His prescription is a coordinated, internationally recognized safety body with real oversight teeth, one that doesn't let any single company grade its own homework on whether its models are safe to unleash.