Google DeepMind launches institute for debate on AGI — TechCrunch
Google and Google DeepMind researchers on Wednesday announced the launch of the DeepMind Institute, an institute intended to advance discussion around artificial general intelligence (AGI). As TechCrunch reports, the institute is meant to represent a range of views from Google, Google DeepMind, and the broader global research community.
The institute’s directors are listed as DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind head Demis Hassabis. Legg will also serve as executive editor. The announcement noted that participants in the discussions will not always agree with one another and may change their positions as new data and information emerge.
The institute’s first essays
The institute’s first work was a collection of four essays. They address economic policy to overcome the potential consequences of AGI, preserving models’ human-understandable reasoning, principles of human well-being, and an approach to evaluating frontier artificial intelligence models.
DeepMind safety researchers Rohin Shah and Anca Dragan argue in their essay that a narrowing of the ability to verify AI’s step-by-step reasoning is not inevitable. In their view, as new architectures emerge, the most powerful models become more difficult to monitor, so developers and regulators should directly consider safety trade-offs.
More current news is available on the UA.News Telegram channel Telegram.
Among possible approaches, the authors named limiting “opaque sequential depth” — the amount of sequential computation a model can perform without a readable reasoning trace. Another option could be requiring developers to demonstrate that less transparent systems remain equally monitorable.
Evaluating frontier models
In another essay, Demis Hassabis proposed creating a U.S.-led standards body to evaluate the most advanced AI models. Under his proposal, developer companies would initially voluntarily submit models for testing up to 30 days before release. After such a system’s effectiveness is confirmed, passing the tests could become a requirement for deploying frontier models in the United States.
At the initial stage, the body would develop evaluations together with AI companies, and later it would have to create independent undisclosed tests so that laboratories could not adapt models to known checks. Hassabis also allowed for strengthening the mechanism depending on the severity of the situation, including through a coordinated slowdown in the development of advanced AI systems.
