XNews

DeepMind / Demis Hassabis: Davos Interview and Safety Proposal

DeepMind / Demis Hassabis: Davos Interview and Safety Proposal

Demis Hassabis gave a Davos interview and proposed safety inspection rules. - Hassabis estimated ~50% probability of AGI by 2030 under his higher bar definition. - He proposed a US-led 'frontier lab' inspection agency: labs must voluntarily submit frontier models 30 days pre-release for security and safety 'physical and cyber' checks. - Recommended tests include biosecurity, network security, deception risks, and human-interpretable reasoning traces; AI-generated images should carry digital watermarks. - Hassabis said AI-accelerated change could be ~10× the industrial revolution in speed and scale; he described weekly 100-hour workweeks over recent years. - He detailed Gemini 3 successes, robotics integration (Boston Dynamics + Hyundai), and predicted 18–24 months for robotic breakthroughs to robustly perform useful physical tasks. - Hassabis favors continued research while urging international cooperation (an 'AI CERN' concept) and said pausing development requires global coordination. - He rejected 'Transformer is dead' claims, argued Transformers remain central though other innovations may be needed, and emphasized continual learning and system consistency gaps.

Key insights

Hassabis calls for pre-release inspection of frontier models

Demis Hassabis proposed that 'frontier' AI models be voluntarily submitted to an independent inspection body 30 days before public release for cybersecurity, biosecurity, and behavior checks. He suggested this body assess deception risks, require digital watermarks on generated media, and mandate human-interpretable symbolic traces of model reasoning. He positioned the mechanism as U.S.-led but stressed international coordination would be necessary for effectiveness.

Hassabis estimates AGI by 2030 under a strict definition

Hassabis reiterated a timeline placing a roughly 50% probability of AGI by 2030 using his higher standard—an agent exhibiting the full range of human cognitive abilities. He emphasized his standard is stricter than some peers', cited missing capabilities like continual online learning and scientific hypothesis generation, and stated current systems are not yet at that level.

DeepMind pursuing both scaling and novel research directions

Hassabis argued DeepMind and Google combine deep research with product deployment: citing Gemini 3, Imagen, and Genie as examples. He disagreed with claims that Transformers are a dead end, noting transformers will remain central though additional breakthroughs (e.g., world models, continual learning) might be needed. He framed DeepMind as maintaining long-term research investments alongside faster product rhythms.

Robotics progress: physical breakthroughs expected in 18–24 months

Hassabis described Gemini Robotics integration into Boston Dynamics Atlas, noting robotics is near a critical point but still requires 18–24 months of work to reach robust, deployable utility. He highlighted algorithmic stability and hardware (mechanical hand) challenges and noted simulation-to-real data, motion transfer, and vision-language-action models as key enablers.

Economic and societal impacts: fast, large-scale transformation

Hassabis compared AI’s disruption to be an order of magnitude faster and larger than the industrial revolution, raising concerns about employment, distribution, and societal meaning. He advised young people to 'learn to learn' and recommended leaders pick partners aligned with responsible AI development. He also discussed the need for new philosophical thinking about meaning in a post-scarcity world.

Global cooperation and an 'AI CERN' ideal

Hassabis advocated for a scientific, multinational approach—an 'AI CERN'—to coordinate final stages toward AGI, involving scientists, philosophers, economists, and ethicists. He acknowledged present international cooperation is difficult but emphasized that shared safety and security protocols are essential if risks are as high as suggested.

Source posts