AGI's Frontier: Google DeepMind Explores AI's Next Steps

Google and Google DeepMind researchers initiated the DeepMind Institute on Wednesday to propel the dialogue surrounding artificial general intelligence (AGI). This new institute aims to present diverse perspectives from Google, Google DeepMind, and the global research community on AGI. Its directors include DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind chair Demis Hassabis, with Legg serving as managing editor. The institute's inaugural collection features four essays that delve into various critical topics concerning AGI. These essays cover economic strategies for managing potential AGI-induced disruption, methods for preserving human-readable model reasoning, principles for fostering human flourishing, and a framework for assessing frontier AI models.

One essay highlights the shrinking window of transparency in AI, arguing this is not an unavoidable outcome. As advanced models become less interpretable, developers and regulators are urged to address safety trade-offs directly. This could involve limiting "opaque serial depth" or ensuring less transparent systems maintain monitorability. Another essay by Hassabis proposes a U.S.-led frontier AI standards body to evaluate the most sophisticated AI models. Initially, developers would voluntarily submit models for review before release. Eventually, passing these independent evaluations could become mandatory for deploying frontier models in the United States. This body would develop undisclosed, "held-out" tests to prevent models from being tailored to known assessments. Hassabis suggested the framework could be strengthened, potentially leading to a coordinated slowdown among developers if the situation warrants. The industry's safety discussions are shifting from broad concerns to specific proposals for disclosure, external review, and potential development pacing. This trend accelerated recently as industry leaders supported calls to moderate frontier AI development.