AI researchers put out videos saying superintelligence is ‘exactly as dangerous as it sounds’

The Verge covers a dozen interviews from Palisade Research’s frominside.ai with current and former staff at OpenAI, Google, and Anthropic. Geoffrey Irving (ex-OpenAI/DeepMind) puts human extinction near a coin flip; DeepMind’s Neel Nanda cites at least 10%. Former OpenAI researcher Daniel Kokotajlo calls superintelligent AI “god-like,” says we don’t know how to control them, and that this “must not be allowed to happen.”

Clips also tackle hype suspicions — Google’s Mary Phuong notes viewers should distrust her because the lab pays her — and why people stay: Nanda says he’d quit if he didn’t believe his work reduced existential risk, and that quitting alone wouldn’t stop the companies.

We’re taking the on-camera odds seriously as social fact, not as settled science.

When builders use coin-flip language, the public hears a industry admitting it cannot bound outcomes. Useful AI for drugs and climate still needs that honesty — and better than vibes about what “pacing” means. The absence of one cohesive solution in the videos is the point: safety is a contested practice, not a slogan.

I’m watching whether these interviews change deploy decisions or only the discourse. Phuong’s conflict-of-interest line is the most adult sentence in the package. Kokotajlo’s “must not be allowed” collides hard with DevDay shipping energy the same week.

We’ll keep advocating capabilities that help people while demanding falsifiable control claims before anything “god-like” is treated as inevitable.

The Verge, September 29, 2026 (Stevie Bonifield). Palisade Research / frominside.ai interview series.