DeepMind Institute Maps Four Routes From AGI to Superintelligence

What happened
On November 19, 2026, DeepMind Institute released an essay outlining four possible ways artificial intelligence could progress from human-level AGI to superintelligence. The four routes described are scaling up current models, shifting technological paradigms, coordinating large populations of AI agents, and recursive self-improvement. The essay details the technical challenges and limitations of each path, including data shortages, hardware constraints, and coordination issues. It also discusses the rapid growth in computing power and the potential for large numbers of AI agents, while noting that physical and logical limits will still cap intelligence.
Key details
- The four routes are: scaling models, paradigm shifts, agent coordination, and recursive self-improvement.
- Global effective compute is estimated to have grown at least tenfold per year.
- A workload costing $1 billion in compute could drop to $1,000 over ten years if current trends continue.
- Superintelligence is defined as reliably outperforming tens of thousands of human experts on complex tasks.
- Physical, logical, and complexity constraints are identified as barriers to unlimited intelligence.
Why it matters
AI developers, technology companies, and policymakers face new technical and organizational challenges as they pursue or regulate the development of superintelligent systems. The rapid increase in computing power and falling costs could enable the deployment of more AI agents than the human population, potentially transforming knowledge work and the broader economy. Institutions will need new tools and benchmarks to track progress and distinguish between faster, more numerous agents and genuinely new capabilities.
