Google DeepMind published a research paper this month asking what comes after AGI. The paper, titled “From AGI to ASI,” treats human-level artificial general intelligence not as an endpoint but as a departure gate, and maps what the subsequent climb toward artificial superintelligence might look like and where it could stall.
The paper is authored by fourteen researchers at Google DeepMind, including Shane Legg, one of the lab’s co-founders, and Allan Dafoe, who leads its long-term strategy work on AI governance. Its publication on arXiv marks one of the more direct statements from a frontier lab about the theoretical structure of intelligence beyond the human range.
The abstract defines ASI as a system more cognitively capable than large human organizations, not merely than individual humans. That framing matters: it sets a much higher bar than the loose usage of “superhuman” that already applies to narrow AI systems in chess or protein folding.
The paper outlines four pathways by which the transition from AGI to ASI might occur: continued scaling of AGI systems, a shift to a new AI paradigm, recursive self-improvement, and superintelligence arising from large populations of interacting agents. The abstract does not detail the mechanics of each pathway, and the paper is clear that each faces possible frictions and bottlenecks. Whether those frictions are “negligible or substantial” is described as an open research question, not a settled answer.
That caveat is load-bearing. A paper cataloguing potential pathways is not evidence that any of the pathways work. The history of AI forecasting is dense with confident stage maps that were overtaken by events or by obstacles that looked minor and proved major. DeepMind’s framing here is appropriately hedged: the report explicitly states that large uncertainties make it impossible to rule out either rapid acceleration or prolonged stagnation.
One of the paper’s more striking analytical claims is structural. It argues that the conventional picture of a single, abrupt AGI-triggered step change may be wrong. Instead, the authors suggest society should prepare for a series of consecutive AI-enabled breakthroughs across science and technology, each arriving before the previous one has been absorbed. That framing shifts the policy problem: instead of designing for one acute transition, institutions would need to manage sustained, compounding disruption across domains simultaneously.
For safety and governance researchers, the paper’s significance is less about which pathway works and more about the fact that DeepMind produced this taxonomy at all. Frontier labs publishing structured analyses of post-AGI trajectories changes the baseline for policy conversations. Governments and standards bodies that have been working from vague “transformative AI” language now have a framework from the lab that controls some of the most capable models in the world, describing a specific endpoint (Universal AI, which the paper calls theoretically well-understood) and the routes toward it. That framing will be cited.
The paper closes by noting that getting ready for these outcomes will demand coordinated effort spanning many fields and reaching worldwide, though it does not specify what institutions or mechanisms that requires. That is the gap between a research paper and a policy roadmap.
For anyone building safety arguments for regulators, or working on AI governance frameworks in the next twelve months, this paper gives you both a conceptual anchor and a credibility problem to manage: the lab that published it is also one of the labs racing to build what it describes.
Google DeepMind, “From AGI to ASI,” arXiv preprint submitted June 10, 2026, authored by Tim Genewein, Matija Franklin, Shane Legg, and eleven co-authors.
