Researchers from Google and Google DeepMind on Wednesday launched the DeepMind Institute to advance the conversation around artificial general intelligence (AGI). The institute lists DeepMind co-founder Shane Legg, Google executive James Manica and Google DeepMind chairman Demis Hassabis as board members, with Legg serving as editor-in-chief.
The new institute aims to surface differing views among the broader global research community around Google, Google DeepMind, and AGI. “They may not always agree and are likely to change their minds as more data and information emerges on a rapidly changing frontier,” the announcement reads.
The first collection of four essays covers a wide range of topics, including economic policies to manage potential AGI disruption, maintaining human-readable model reasoning, principles for human flourishing, and frameworks for evaluating frontier AI models.
An essay by DeepMind safety researchers Rohin Shah and Anka Dragan argues that the shrinking of AI’s window of transparency (the ability to see and confirm a model’s step-by-step inferences) is inevitable. The authors say that developers and regulators will have to face safety trade-offs directly, as new architectures will make the most powerful models difficult to monitor. That could mean limiting the “opaque serial depth” (the amount of sequential computations a model can perform without producing a readable inference trace) or requiring developers to demonstrate that less transparent systems are still observable.
In another essay, Hassabis proposes a U.S.-led Frontier AI standards body to evaluate cutting-edge AI models. Under his framework, developers were initially supposed to voluntarily submit models for review up to 30 days before release. If the evaluation system proves effective, passing that test could become a requirement for the Frontier model to be deployed in the United States.
The agency initially plans to design the assessments in consultation with AI companies, but will eventually develop independent, private assessments (what the essay calls “holdout” tests) to prevent labs from tailoring their models to known assessments. Hassabis said this could include a coordinated slowdown among frontier AI developers, and that the framework could be strengthened “as the severity of the situation necessitates.”
The papers come as the industry’s safety debate moves from broad expressions of concern to concrete proposals for disclosure, external oversight, and a coordinated slowdown if safety measures fall behind. That change accelerated this week when industry leaders endorsed elements of Anthropic CEO Dario Amodei’s call to set the “pace” of frontier AI development.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
