Google DeepMind has published a detailed essay exploring what follows the attainment of AGI—artificial general intelligence capable of performing almost any intellectual task at a human level. The author of the text is Tim Genewein, and according to industry publications, it was released on November 19, 2026. The core idea of the work is that the next stage after AGI could be ASI—artificial superintelligence, which would consistently outperform large teams of experts in tasks that currently take humans years, doing so with significantly less time and resources. However, the essay emphasizes that rapid progress is not guaranteed: each of the described scenarios faces its own technical bottlenecks.

"Cognitive Industrialization": The Economics of Transition

Genewein identifies "cognitive industrialization"—the mass cost reduction and automation of intellectual labor—as the key driver of the transition from AGI to ASI. According to his data, effective compute resources for AI have been growing at least 10-fold annually over the last decade. Three factors simultaneously influenced this dynamic: improvements in hardware, increased efficiency of algorithms, and rising investment in AI infrastructure. Specifically, compute power per dollar invested has increased by approximately 1.5x annually due to advancements in chip manufacturing, while the efficiency of training algorithms has increased 3–6x annually over the last decade.

If this trend continues, in ten years, an intellectual task that currently requires a billion dollars worth of computation could be performed for just a thousand dollars. Simultaneously, the number of AI agents of comparable quality that can be run on available compute resources could theoretically exceed the current global population. A more conservative estimate, accounting for physical and economic constraints, suggests that the total number of available agents could increase by more than 1,000x by 2030 and by over a billion-fold within a decade.

Four Technological Paths from AGI to ASI

The first path is further scaling: increasing model sizes, training data volumes, and compute allocated to processing each request. This approach is already driving AI capabilities but faces shortages of high-quality data, diminishing returns from increasing model sizes, and physical infrastructure constraints. The second path involves moving to new technological paradigms. Modern models are primarily trained to predict the next token and then fine-tuned for reasoning and instruction following, yet they remain limited in continuous learning, building world models, decision-making, and the ability to autonomously create fundamentally new concepts. Scientists have yet to reach a consensus on whether these shortcomings can be resolved by evolving existing architectures or if fundamentally new approaches are required.

The third path is the coordination of multiple AI agents. Just as corporations and scientific communities solve problems beyond the reach of individuals, millions of AGI instances could aggregate into virtual collectives. Early experiments show that agent interaction can improve efficiency in some tasks; however, as their numbers grow, losses arise due to coordination complexity, as well as problems with error propagation, manipulation, and collusion between agents. The fourth path is recursive self-improvement (RSI): the AI autonomously improves its software code, model architectures, training algorithms, and even chip designs, accelerating the creation of subsequent generations of systems. AlphaZero, which reached superhuman skill levels through its own experience, serves as an example. Yet, self-improvement will not necessarily accelerate indefinitely: further improvements require increasingly greater effort, and errors in evaluating results could stall or even degrade the process.

Context of Predictions and Disclaimers

The essay expands upon ideas from the earlier technical report "From AGI to ASI" and aligns with the general trajectory of public statements by DeepMind's leadership. Previously, division CEO Demis Hassabis pointed to the possibility of general-purpose AI and super-capable systems emerging around 2030, and in 2026, he adjusted his forecasts regarding the timelines for reaching AGI. In the new text, Genewein emphasizes that these four directions are not mutually exclusive and may reinforce each other: even if recursive self-improvement does not lead to uncontrolled acceleration of progress, the increase in compute resources, model efficiency, and the number of well-coordinated agents could drive AI development beyond the human level. An important conclusion follows: restricting only one area, such as compute power for training models, will not necessarily stop the trajectory toward ASI, as development could continue through the other described tracks.