Google DeepMind Launches Institute to Broaden AGI Safety and Policy Discussions

DeepMind Institute Established

Google and Google DeepMind researchers launched the DeepMind Institute on Wednesday. The institute's leadership includes DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind chair Demis Hassabis, with Legg serving as managing editor.

Fostering Diverse AGI Perspectives

The new institute intends to bring together various viewpoints on artificial general intelligence (AGI) from within Google, Google DeepMind, and the global research community. The goal is to acknowledge and explore differing opinions as the field of AGI rapidly develops.

Inaugural Essay Collection

The institute released an initial collection of four essays. These essays address topics such as economic policies for managing potential AGI disruption, methods for preserving human-readable model reasoning, principles for human flourishing in an AGI future, and a framework for evaluating advanced AI models.

Proposals for AI Transparency and Standards

One essay by DeepMind safety researchers Rohin Shah and Anca Dragan argues that AI's decreasing transparency is not inevitable, suggesting developers and regulators should address safety trade-offs directly. They propose limiting "opaque serial depth" or requiring demonstrations of monitorability for less transparent systems. Another essay by Demis Hassabis proposes a U.S.-led frontier AI standards body to evaluate advanced AI models, initially through voluntary submissions and potentially evolving into a mandatory requirement for deployment. This body would develop independent, undisclosed evaluations to prevent models from being tailored to known tests.