Google DeepMind has launched the DeepMind Institute, a new initiative designed to expand discussion around artificial general intelligence (AGI) and bring together different perspectives on the future of advanced AI systems.
The institute was launched by Google and Google DeepMind researchers, with DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind chair Demis Hassabis listed as directors. Legg will serve as managing editor.
The institute aims to present a range of views from researchers inside Google and the wider AI community. The organisation said contributors may disagree and change their views as new evidence and information emerge in the rapidly developing field of AI.
Early Research Focuses on AI Safety and Evaluation
The institute’s first collection includes four essays covering topics such as economic policies around potential AGI disruption, preserving human understanding of AI reasoning, principles for human flourishing, and methods for evaluating advanced AI models.
In one essay, DeepMind safety researchers Rohin Shah and Anca Dragan examine the future of AI reasoning transparency. Their research on reasoning transparency discusses concerns that increasingly complex AI systems will become harder to monitor and evaluate.
The researchers argue that developers and policymakers may need to address trade-offs between model capability and transparency, including questions around how much internal reasoning AI systems should expose for safety evaluation.
Proposal for Frontier AI Standards
In another essay, Demis Hassabis proposes a framework for evaluating the most advanced AI models through a potential U.S.-led frontier AI standards body.
The frontier AI evaluation framework suggests that developers could initially submit models voluntarily for review before release. Over time, successful evaluation processes could potentially become part of deployment requirements for advanced AI systems.
The proposal also discusses independent evaluations designed to prevent AI developers from optimising models only for known tests. These assessments could use undisclosed evaluation methods to measure the performance and safety of frontier models.
The launch comes as the AI industry continues debating how to balance rapid development with safety measures, including transparency, external review, and possible coordination among companies developing advanced AI systems.
