Google DeepMind launches institute to widen the AGI debate

AI model evaluations may soon become a mandatory step before deployment.

Google DeepMind has launched an institute dedicated to broadening the discussion around Artificial General Intelligence (AGI). The institute, directed by DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind chair Demis Hassabis, aims to foster diverse perspectives on AGI, acknowledging that these views may evolve. Its inaugural publication features four essays addressing topics such as economic policies for AGI disruption, maintaining human-readable AI reasoning, principles for human flourishing, and a framework for evaluating advanced AI models. One essay by DeepMind safety researchers argues that decreasing AI transparency is not inevitable and proposes direct confrontation of safety trade-offs. Another essay from Hassabis suggests a U.S.-led standards body for evaluating frontier AI models, initially voluntary but potentially mandatory for deployment in the U.S.

The establishment of this institute signifies a strategic move by Google DeepMind to shape the discourse on AGI, moving beyond internal discussions to engage a wider research community. This effort could influence the direction of AI safety research and policy by prioritizing specific areas of inquiry and proposing concrete evaluation frameworks. The emphasis on diverse opinions, even those that may diverge from Google's own, suggests a bid for broader credibility and influence in the global AI governance landscape.

Hassabis's proposal for a standards body, with a potential for mandatory compliance, signals a significant shift toward formal regulation in AI development. This could impose new requirements and timelines on all frontier AI developers, forcing them to prioritize safety and transparency in their development cycles. The idea of "held-out" tests indicates a move towards more robust and standardized validation methods, aiming to prevent gaming the system and ensure genuine AI safety.