DeepMind Researchers Map the Road From AGI to Superintelligence — and Its Seven Bottlenecks

A team of Google DeepMind researchers — including co-founder Shane Legg, AIXI theorist Marcus Hutter, and colleagues Tim Genewein, Matija Franklin, Laurent Orseau, Samuel Albanie and Thore Graepel — has published "From AGI to ASI" (arXiv: 2606.12683), the most systematic attempt yet by a frontier lab to map how human-level AI might become superintelligent.

Four Pathways

The paper is a survey and framework, not a product announcement, and it is careful to claim neither that AGI has arrived nor that superintelligence is imminent. Its contribution is structural: the authors identify four distinct pathways along which the AGI-to-ASI transition could run:

1. Scaling — continued growth in compute, data and model size
2. AI paradigm shifts — architectural or algorithmic breakthroughs beyond today's methods
3. Recursive improvement — AI systems accelerating their own research and development
4. Multi-agent collectives — populations of cooperating agents whose aggregate capability exceeds any individual system

Each pathway carries different speed profiles, warning signs and governance implications — a taxonomy the authors argue policymakers need before, not after, the transition begins.

Defining Superintelligence Precisely

The paper's definition of ASI is notably operational: a system that outperforms large teams of human experts — not merely the best individual — across a broad range of domains. That framing matters because "team-level" performance is measurable, and because organizations, not individuals, are the actual unit of human intellectual work that such systems would displace.

Seven Bottlenecks

Against accelerationist assumptions, the authors catalogue seven bottlenecks that could slow or stall the transition — constraints spanning compute availability, energy, data quality, algorithmic limits, evaluation difficulty, and the physical-world friction of deploying and verifying increasingly capable systems. The message cuts both ways: no wall guarantees safety, but no curve guarantees takeoff either.

Why the Timing Matters

The paper lands amid the fastest capability diffusion in the field's history — a July in which frontier models from three labs went simultaneously public and inference prices collapsed. When the researchers who coined the term AGI begin publishing structured frameworks for what comes after it, the Overton window of serious research has moved. For governments across Asia now drafting AI legislation — from Vietnam's new AI law to Korea's Basic Act — the bottleneck list doubles as a monitoring agenda: these are the indicators worth watching.