Google and Google DeepMind researchers have launched the DeepMind Institute as a forum for competing views about artificial general intelligence and its consequences. DeepMind co-founder Shane Legg, Google executive James Manyika and DeepMind chair Demis Hassabis are directors, with Legg serving as managing editor.
The first collection contains four essays covering economic policy, human flourishing, readable model reasoning and evaluation of frontier systems. Safety researchers Rohin Shah and Anca Dragan argue that developers should confront the loss of visible reasoning as architectures become harder to inspect. Possible responses include limiting the amount of sequential computation a model can perform without a readable trace or proving that a less transparent system remains comparably monitorable.
In another proposal, Hassabis calls for a US-led standards body to test the most capable models. Developers would initially submit systems voluntarily up to 30 days before release. If the process proved useful, passing independent evaluations could later become a deployment requirement. The body would eventually use undisclosed tests so labs could not tune models specifically for known benchmarks. These are proposals, not Google policy or enacted regulation. The institute says contributors may disagree with one another and revise their positions, giving the project value as a public record of concrete options rather than a single corporate answer to AGI governance.