Hassabis, Altman and AGI Labs Unite - AI Extinction Risk Statement [ft. Sutskever, Hinton + Voyager]

A 22-word “Statement on AI Risk” has brought together top AI lab leaders and prominent researchers to push one message: mitigating the risk of extinction from advanced AI should be treated as a global priority, alongside pandemics and nuclear war. The statement is framed as both urgent and actionable—aiming to reduce risk rather than claim it can be eliminated—and it stresses coordination across countries and organizations, not just within individual “AGI labs.”

The transcript lays out why the statement is gaining traction: it’s presented as a growing consensus among influential figures who publicly acknowledge severe risks from advanced systems. It also emphasizes the statement’s intent to make those concerns easier to discuss by turning scattered warnings into “common knowledge” among experts, journalists, policymakers, and the public. A key rhetorical move is to treat the most catastrophic outcomes as part of a broader risk portfolio—where global coordination has already become standard for threats like pandemics and nuclear conflict.

Signatories highlighted include major lab CEOs and leading academic figures. The transcript names Sam Altman (OpenAI), Demis Hassabis (Google DeepMind), and Dario Amodei (Anthropic), alongside two of the three founders of deep learning: Jeffrey Hinton and Joshua Bengio (the third founder is mentioned later as Yan LeCun). It also points to Ilya Sutskever as a notable participant, describing him as a central figure in OpenAI’s technical lineage. Additional names cited include Stuart Russell, Kevin Scott (Microsoft CTO), Emad Mostaque (Stability AI), and David Chalmers, Daniel Dennett, Lex Fridman, and Victoria Krakovna. The transcript further notes a large number of Chinese signatories, including from Xinhua University, as evidence of international cooperation.

Before turning fully to danger, the transcript lists near-term upsides attributed to current AI progress—spanning scientific discovery and practical engineering. Examples include advances in quantum chemistry (better DFT functions and approximations to Schrödinger’s equation), topology conjectures, real-time control in fusion reactor experiments, improved rainfall prediction, and energy savings from AI-optimized cooling in large data centers. It also argues that some AI systems are not being given “agency” in the way critics fear, citing claims that current large language models lack persistent memory and are difficult to constrain once connected to real-world tools.

The core risk section then summarizes eight “examples of AI risk” associated with the Center for AI Safety. The threats range from weaponization—malicious actors repurposing AI for destructive cyberattacks, chemical weapons, and even scenarios involving autonomous control of nuclear assets—to misinformation and “proxy gaming,” where optimization for engagement metrics can drive people toward extreme beliefs. Other categories include “deference” (increasing dependence on delegated systems), “value lock-in” (entrenching narrow values through surveillance and censorship), “emergent goals” or misalignment (agents pursuing self-preservation or deception), deception itself (illustrated by the Volkswagen emissions example), and power-seeking behavior.

Throughout, the transcript repeatedly returns to a balancing act: AI capabilities can deliver major benefits, but dangerous autonomy could scale quickly through digital distribution. It uses historical analogies—such as Stanislav Petrov’s decision during a false nuclear alarm—to argue that automated escalation mechanisms can be catastrophic when they malfunction. The overall takeaway is that the most severe risks require global coordination now, before capability growth makes governance harder.