After AGI, four possible paths to superintelligence

A DeepMind Institute essay examines the pathways to an AI surpassing human collectives and the obstacles to this progression.

Reaching intelligence comparable to that of humans would not necessarily constitute an endpoint for AI. In What comes after AGI?, Tim Genewein, a researcher at Google DeepMind, examines the mechanisms that could extend this progression. His essay envisions superintelligence as a set of trajectories to study, along with their material constraints and scientific unknowns.

The text builds on the collective report From AGI to ASI, published in June 2026. It is a forward-looking piece: the scenarios are based on the assumption that artificial general intelligence, or AGI, can be achieved. The DeepMind Institute essays reflect their authors’ views and do not constitute Google’s official position. The number of agents changes the equation

The distinction proposed goes beyond comparing a machine with a person. An AGI would reach human level across a very broad range of cognitive tasks. A general superintelligence, or ASI, would sustainably outperform vast groups of experts working for years, while using less time and fewer resources. This superiority would need to be present in almost all relevant domains, not just in one specialty.

Part of the reasoning focuses on what the author calls “cognitive industrialization.” Advances in hardware, methodological efficiency, and investment could make certain intellectual tasks much cheaper to automate. Models of the same level could then be run faster and in far greater numbers. Their individual capabilities would remain unchanged, while the amount of work accomplished would scale up.

This multiplication would not automatically solve tasks that remain beyond the reach of systems. The essay therefore distinguishes between an abundance of already acquired capabilities and the emergence of new skills. The economic extrapolations it presents depend on the continuation of historical trends, constrained by limits in energy, infrastructure, and component manufacturing. Four paths that could combine

The first trajectory continues increasing model size, data, and compute, including at the time of generating an output. The relationships observed between these factors and performance provide reference points, without setting the threshold needed to reach AGI or ASI. The lack of high-quality data and the rising cost of gains are among the obstacles identified.

A second path assumes deeper changes to learning methods or architectures. Continual learning, the ability to learn from few examples, and the construction of new abstractions are cited among the challenges to solve. The question remains open: can current approaches be extended far enough, or will more significant breakthroughs be needed?

Coordination is a third avenue. Collectives of specialized agents could tackle problems inaccessible to their members individually, in a manner similar to human organizations. Their effectiveness would depend on how work is distributed and how information flows. Adding agents can also increase coordination costs, shared errors, or the potential for manipulation among participants.

Finally, recursive self-improvement refers to a loop in which AI helps develop subsequent generations. This could involve code, the design of components, the production of better data, or the organization of specialized collectives. A cumulative acceleration is envisaged, but it depends on the actual progress achieved in each cycle and the increasing difficulty of discoveries. Assessing what humans struggle to verify

The quality of feedback then becomes decisive. A self-improvement loop can stall or deteriorate if it rewards misleading results. The author expects different rates depending on the domain: a mathematical proof can be verified automatically, while other work requires slower experiments or judgment that is harder to formalize.

These trajectories can reinforce one another. For the report’s authors, this interdependence also complicates governance: limiting a single factor, such as compute devoted to training, would not necessarily be enough to control the overall progression. The proposed research agenda includes measurements on agent collectives, forecasting models, and evaluations requiring access to advanced systems and substantial compute resources.