On September 16, Google and Google DeepMind researchers established the DeepMind Institute to promote interdisciplinary discussions around general artificial intelligence (AGI). The institute's board includes Shane Legg, co-founder of DeepMind, Google executive James Manyika, and chairman of Google DeepMind, Demis Hassabis, with Legg serving as executive editor. The institute emphasizes presenting different perspectives from Google, DeepMind, and the global research community on AGI issues.
The institute's first four articles discuss AGI economic policy, model reasoning transparency, principles of human flourishing, and a frontier AI evaluation framework. Rohin Shah and Anca Dragan proposed that attention should be paid to the security risks caused by the shrinking "transparent window" of models, which can be addressed by limiting "invisible serial depth" or requiring low transparency systems to prove they still maintain equivalent controllability.
Hassabis also suggested establishing a frontier AI standards organization led by the United States, where developers could voluntarily submit assessments up to 30 days before model release; once the system matures, passing the test may become a necessary condition for deploying frontier models in the US. The evaluation organization will eventually adopt independent, non-public "confidential" tests to reduce the possibility of models optimizing for known metrics, and further measures can be strengthened based on risk levels.
This move shows that discussions on AGI are extending further from capability development to specific topics such as reasoning transparency, external evaluation, and governance mechanisms.
