Google DeepMind launched an institute to discuss the future of AGI

Google and Google DeepMind have launched the DeepMind Institute, which will focus on advancing the discussion around artificial general intelligence (AGI) and the potential for superintelligence.

The institute is led by DeepMind co-founder Shane Legg, Google Executive Chairman James Manyika, and Chairman of the Board of Google DeepMind Demis Hassabis. Legg also took on the role of Editor-in-Chief.

The institute aims to bring together diverse perspectives on the development of AGI—both within Google and Google DeepMind, and among researchers worldwide. The authors note that participants' positions may evolve as new data and research results emerge.

The institute's first publication includes four essays. They examine economic policy in the face of potential changes driven by AGI, transparency in AI model reasoning, principles of human well-being, and approaches to evaluating advanced artificial intelligence systems.

In one of the papers, Google DeepMind security researchers Rohin Shah and Anca Dragan address the issue of decreasing AI transparency. They propose studying constraints on so-called "opaque sequential depth"—the amount of computation a model performs without forming a chain of reasoning understandable to humans.

Hassabis's separate essay focuses on the creation of a standards-setting body for advanced AI models under U.S. leadership. According to the proposal, developers could voluntarily submit their models for review 30 days before launch. In the future, successfully passing such tests could become a requirement for deploying the most advanced systems in the United States.

As the evaluation system develops, the institute proposes transitioning to independent closed tests, the results of which will not be known to developers in advance. Hassabis also allows for the possibility of coordinated slowing of AI development if risks from advanced systems necessitate such measures.

The launch of the DeepMind Institute reflects a shift in the AGI safety discourse from general warnings to discussions of specific mechanisms for evaluating, controlling, and regulating advanced models.

Source: TechCrunch