Four essays, three directors and one stated promise: the people running Google DeepMind’s new institute expect to disagree with each other. “They will not always agree, and they will likely change their minds, as more data and information comes to light at the fast-moving frontier,” the launch announcement said.
Google and Google DeepMind researchers opened the DeepMind Institute on Wednesday, September 17, as a place to argue about artificial general intelligence in public. Shane Legg, a DeepMind co-founder, is managing editor. Google executive James Manyika and Google DeepMind chair Demis Hassabis are the other directors. The aim is to surface the differences between Google, DeepMind and the wider research community instead of smoothing them over.
What the first four essays cover
The opening collection takes on economic policy for managing possible AGI disruption, keeping a model’s reasoning readable by humans, principles for human flourishing, and a framework for evaluating frontier models. Here’s what that means in plain terms: a company is putting its internal safety arguments where researchers, regulators and rival labs can pick them apart.
Can we still see how a model thinks?
DeepMind safety researchers Rohin Shah and Anca Dragan take on transparency. Today a person can often check a model’s step-by-step reasoning. They argue that this window is shrinking but that it does not have to. New architectures are making the most powerful models harder to monitor, and the authors say developers and regulators should face that trade-off head on.
One option is to limit what they call “opaque serial depth,” the amount of sequential computation a model can run without leaving a readable reasoning trace. Another is to make developers show that a less transparent system can be monitored just as well.
A standards body for frontier models
Hassabis uses his essay to propose a frontier AI standards body led by the United States. At first, developers would hand over models for review voluntarily, up to 30 days before release. Once the system proved itself, passing its tests could become a condition of deploying frontier models in the country.
The body would begin by designing assessments with the AI companies. Over time it would build its own undisclosed tests, which the essay calls “held-out” evaluations, so that labs cannot tune their models to questions they already know.
Hassabis said the framework could be “ratcheted up if the seriousness of the situation demands,” and that could include a coordinated slowdown among frontier developers.
The timing
The industry’s safety debate has been moving from general statements of concern toward specific proposals: disclosure, outside scrutiny and, if safeguards lag, coordinated slowdowns. That shift picked up speed this week, when industry leaders endorsed parts of Anthropic CEO Dario Amodei’s call to “pace” frontier AI development.
Sources
- Google DeepMind launches institute to widen the AGI debate | TechCrunch
- DeepMind Institute Launches AGI Debate | ChatSlide
