Google DeepMind Describes Four Paths from AGI to Superintelligence
Google DeepMind researcher Tim Genewein published an essay on how AI could transition from AGI (human-level intelligence) to ASI, or superintelligence.

The author identifies 'cognitive industrialization'—the reduction in the cost of intellectual labor—as the primary driver of this transition. Over the past decade, efficient AI computations have grown at least tenfold annually. If this pace continues, a task that currently requires $1 billion in computations will cost around $1,000 in a decade.
Genewein outlines four scenarios. The first is the further scaling of models, data, and computations, though this path faces limitations due to the scarcity of high-quality data and infrastructure constraints. The second involves new architectures: current models still struggle with continuous learning and generating fundamentally new concepts. The third is the integration of millions of AI agents into collectives, though the more agents there are, the harder they are to coordinate, and the higher the risk of collusion. The fourth is recursive self-improvement, where AI independently enhances its code, algorithms, and even chips.
The author believes these paths complement one another. Therefore, restricting one, such as computational power, will not halt progress toward superintelligence. He urges the development of methods to measure this progress now.



