A new paper from Google DeepMind and collaborators argues that the arrival of human-level artificial general intelligence may not be the end of the story for AI progress. Instead, the researchers say the years after AGI could bring continued advances toward systems that exceed human and organizational capabilities.
The paper, titled From AGI to ASI, was posted to arXiv on June 10. It brings together 14 authors, including Tim Genewein, Shane Legg, Marcus Hutter and other DeepMind researchers. The report does not claim that artificial superintelligence is imminent, but it does lay out several ways AI systems could move beyond AGI if current trends continue.
In the report, the authors describe superintelligence as a system that would be more capable than large groups of humans across a broad range of tasks. They frame the question as one of progression along a continuum of machine intelligence, rather than as a single leap from today’s models to a final destination.
The paper identifies four possible routes from AGI to ASI. The first is straightforward scaling, in which AGI systems improve as models, compute and training methods continue to grow. The second is a shift in AI paradigms, meaning a change in the underlying approach rather than simple enlargement of existing systems. The third is recursive improvement, where AI contributes to the creation of better AI, potentially accelerating its own development. The fourth is the emergence of superintelligence from large multi-agent systems, where many AI agents collectively reach capabilities beyond those of any individual system.
The authors also discuss potential frictions that could slow or limit progress along these paths. Those bottlenecks could include technical obstacles, coordination challenges or limits that are not yet well understood. The report says it is difficult to predict whether such constraints will matter only modestly or will significantly shape the trajectory from AGI onward.
One of the paper’s broader points is that AGI may not trigger a single dramatic transformation all at once. The authors suggest it is possible that AI progress will keep accelerating after human-level systems emerge, producing a series of major changes across science, technology and society.
That possibility matters, the report argues, because many public discussions treat AGI as a discrete threshold. The paper instead presents a future in which increasingly capable systems could drive repeated breakthroughs, with social effects unfolding over time rather than in one event.
The authors say the underlying uncertainty is large enough that rapid post-AGI progress cannot be ruled out. At the same time, they stop short of making a firm prediction about how fast ASI might arrive, or whether any of the four pathways will prove dominant.
Because the paper focuses on possible trajectories rather than a forecast, it raises more questions than it answers. How much room is there for scaling to continue? Could a new AI paradigm produce a major jump in capability? Would self-improvement loops be limited by engineering and safety constraints? Could large collectives of agents outperform today’s systems in unexpected ways?
The report concludes that preparing for a post-AGI world will require work across multiple disciplines and international coordination. In its view, the challenge is not only technical but societal, involving issues that extend well beyond AI research itself.
For now, the paper adds to a growing body of work that treats AGI and ASI as connected stages in a longer development cycle. Rather than assuming that human-level AI would mark a stable endpoint, the authors argue that the next phase of machine intelligence could be just as consequential.