AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Allan Dafoe on Technological Determinism, Cooperative AI, and DeepMind's Frontier Safety Plans

· The Cognitive Revolution

This crossover episode from 80,000 Hours features Rob Wiblin interviewing Allan Dafoe, Director of Frontier Safety and Governance at Google DeepMind. Dafoe, formerly founding director of the Center for Governance of AI (GovAI), explains his role leading three pillars: Frontier Safety (dangerous capability evaluations and risk management, embodied in Google's Frontier Safety Framework), Frontier Governance (policy advising), and Frontier Planning (horizon scanning toward AGI). He emphasizes that his team is small but works collaboratively across DeepMind's AI safety, Gemini safety, alignment, responsibility, and policy teams, with DeepMind serving as the heart of frontier policy thinking within Google. The conversation opens with Dafoe's academic work on technological determinism, arguing that technology alone doesn't compel action—military-economic competition does. Groups that fail to adopt advantageous technologies risk losing to those who do. He cites the Meiji Restoration in Japan (uninventing firearms for 250 years until Perry's steamships forced rapid modernization) and critiques differential technological development as a strategy, noting market forces already drive safety work. He advocates the cooperative AI research agenda—arguing that even aligned AI systems pose risks if deployed by conflicting principals, potentially causing flash escalations or excluding humans from cooperative clusters. The discussion then turns to the 'Levels of AGI' paper and the Frontier Model Eval paper testing Gemini 1.0. Results showed Gemini at 3/5 on persuasion, ~2/5 on cyber, ~2/5 on self-proliferation, and ~1-2/5 on self-reasoning. Novel evals include a self-reasoning test where the model must notice filtered words blocking its email task and take corrective action, and another testing whether models know their training cutoff and optimally use search for post-cutoff events. Dafoe notes cyber capabilities warrant close monitoring and emphasizes the importance of capability elicitation—giving models proper tool access to measure their true frontier rather than their raw base performance. The episode highlights DeepMind's safety orientation, Demis Hassabis's openness to international collaboration, and the broader concern that a relatively small group of technical staff holds outsized power over AI's trajectory.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.