Google and Google DeepMind have launched the DeepMind Institute—a new research organization aimed at broadening the debate on AGI (artificial general intelligence). The institute is led by three top executives:

  • Shane Legg (co-founder of DeepMind)

  • Demis Hassabis (president of Google DeepMind)

  • James Manyika (Google executive).

The First Four Essays

The inaugural collection consists of four essays, each addressing a different aspect of the AGI problem. Rohin Shah and Anca Dragan, two DeepMind safety researchers, argue that AI’s window of transparency—the ability to inspect and audit step-by-step reasoning chains—is narrowing, though not inevitably. They propose capping “opaque serial depth,” referring to the amount of sequential computation a model can perform without producing a legible reasoning chain.

Hassabis’s essay proposes a US-led frontier AI standards agency. Under this framework, developers would submit models for pre-release evaluations, initially on a voluntary basis and potentially becoming mandatory later. The agency would design independent assessments that are kept undisclosed in advance to prevent labs from “teaching to the test.”

The Race to State Positions

The institute’s launch comes during a week in which nearly every major AI company published its stance on safety. OpenAI publicly revealed six misalignment cases. Microsoft released its Humanist AI principles. Anthropic proposed three metrics to measure the pace of AI development.

All converge on a common acknowledgment of the risks posed by AGI, but their approaches differ markedly. Microsoft emphasizes containment and critiques model welfare. Anthropic focuses on measuring pace and self-regulation. Google/DeepMind advocates for external agency evaluations.

The big question is whether these statements will translate into tangible action or remain PR strategies against a backdrop where Anthropic prepares for an IPO, OpenAI is valued at over 1,200 billion USD, and the AI race shows no sign of slowing down.