Google DeepMind launches new institute to broaden AGI debate

Google and Google DeepMind researchers announced on Wednesday the launch of the DeepMind Institute, aimed at broadening the conversation around artificial general intelligence (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 stated goal is to surface differing views held within Google and DeepMind as well as across the broader global research community regarding AGI. The announcement noted that participants will not always agree and may change their positions as new data emerges in a fast-moving field.

First essays

The inaugural collection includes four essays: on economic policy for managing potential disruption from AGI, on preserving human-readable model reasoning, on principles for human flourishing, and on a framework for evaluating frontier AI models.

In one essay, DeepMind safety researchers Rohin Shah and Anca Dragan argue that the shrinking ability to observe and verify a model's step-by-step reasoning is not an unavoidable outcome. They contend that as new architectures make the most capable models harder to monitor, developers and regulators need to directly address the resulting safety trade-offs — including limiting how much sequential computation a model can perform without producing a readable reasoning trace.

In a separate essay, Hassabis proposes establishing a U.S.-led body to set standards for evaluating the most advanced AI models. Under his proposal, developers would initially submit models for voluntary review up to 30 days before release; once proven effective, passing such evaluations could become mandatory for deploying frontier models in the U.S. Over time, the body would develop independent, undisclosed tests to prevent companies from tailoring models to known benchmarks. Hassabis said the requirements could be tightened further if circumstances warranted, potentially including a coordinated slowdown in development pace.

The essays come as the industry's safety debate shifts from general expressions of concern toward specific proposals covering disclosure, external oversight, and coordinated slowdowns where safeguards lag — a shift reinforced this week by industry leaders' endorsement of Anthropic CEO Dario Amodei's call to moderate the pace of frontier AI development.