Google Is Starting a New Group to Talk About Very Smart AI
Google and Google DeepMind have started a new group called the DeepMind Institute to widen public discussion about artificial general intelligence, or AGI. AGI means AI systems that could do many kinds of thinking work at a human level or beyond. TechCrunch
The listed directors are DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind chair Demis Hassabis. Legg is also managing editor.
The group says its aim is to share different views from Google, Google DeepMind, and outside researchers about AGI. It is described as a place where company and outside ideas can be compared in public.
The launch includes four opening essays. They cover money and jobs policy for AGI disruption, keeping AI reasoning readable for people, ideas for human flourishing, and how to test the most advanced AI models.
One essay is by DeepMind safety researchers Rohin Shah and Anca Dragan. It is about whether people can still follow how AI reasons. They argue that loss of that clarity is not unavoidable, even as models get more capable and do longer internal work.
They suggest two controls. Developers could limit how much hidden step-by-step work a system can do before a person checks it, a limit they call "opaque serial depth." Or developers would need to show that a less clear system can still be watched and checked. The idea starts as voluntary, but it could lead to rules that can be enforced.
Another essay is by Hassabis. It calls for a U.S.-led body to test frontier models, meaning the most advanced AI systems. At first, developers would send in models on their own up to 30 days before release. Later, passing tests could become required before those models can be used in the United States.
To stop cheating, the body would later make its own secret tests that labs have not seen. This is like a pop quiz. It avoids the common problem where old tests become public and get included in training data.
Hassabis also describes a next step if risks grow. Developers could agree to slow down work together. The wording is conditional. It is a suggested tool, not a decision to act.
Axios reported on September 16, 2026 that the Institute launched that Wednesday. In August 2026, Google changed AI leadership and Hassabis shifted role, Reuters reported. After that change, Hassabis was to study AGI effects on society with few direct reports, according to an Alphabet spokesperson.
The launch followed two safety warnings that week. In mid-September 2026, Reuters reported that a former Google DeepMind researcher warned AI could 'kill all humans.' Reuters On September 16, 2026, Reuters reported that a DeepMind co-founder warned AI abilities must not move faster than safety controls. Reuters
The broader context here is that the Institute faces two hard questions at once. What proof is enough to say a powerful system is safe, and who writes the test. And how jobs and power would shift if AGI takes over more thinking work.
In my view, the two most useful ideas for readers are clearer monitoring and testing before release. Treating clarity as a design choice puts work on logs, step-by-step records, and oversight tools at the center. The testing plan starts voluntary, then talks about required checks and joint slowdowns that would need new legal power.
Looking to what this could make possible, starting with open essays and testable ideas gives researchers and policymakers something clear to discuss before rules are set. If the Institute keeps that exchange open, it could raise the quality of the AGI debate.
