Google DeepMind researchers, including Co-Founder Shane Legg, have published a paper outlining four potential pathways from artificial general intelligence (AGI) to artificial superintelligence (ASI).
The researchers argue that AI progress could continue well beyond human-level capabilities through scaling, recursive self-improvement, algorithmic breakthroughs and multi-agent systems.
The paper, From AGI to ASI, defines AGI as a system capable of achieving at least median human performance across a broad range of cognitive tasks, while ASI is described as a system capable of outperforming organisations comprising thousands of human experts working over periods measured in years.
Rather than forecasting specific dates, the authors focus on the mechanisms that could continue advancing AI after AGI is reached and the factors that could constrain further progress.
The paper identifies four possible routes from AGI to ASI: continued scaling of compute, models and data; algorithmic paradigm shifts that improve efficiency or unlock new capabilities; recursive self-improvement through automated AI research; and large-scale collectives of AI agents whose combined capabilities exceed those of individual systems.
One of the report’s central arguments is that forecasting compute growth is relatively tractable because it can be decomposed into three measurable factors: hardware improvements, infrastructure investment and algorithmic efficiency gains.
The paper cites estimates suggesting that hardware capabilities are improving by roughly 1.5x per year, AI infrastructure investment by around 2.5x per year, and algorithmic efficiency by as much as 6x annually.
Combined, those trends imply effective compute growth of roughly 10x per year, according to estimates cited in the report.
The authors note that this figure is at the lower end of public estimates and could be higher. However, they emphasise that predicting how additional compute translates into new capabilities remains substantially harder than forecasting compute growth itself.
“The four pathways are not mutually exclusive and progress may happen on all of them simultaneously, which could lead to compounding (not just additive) increases in artificial intelligence,” the researchers write.
While scaling remains the most visible route today, the report argues that superintelligence may not require dramatically smarter individual models.
Even if progress in individual systems slows near AGI-level performance, coordinating millions of AGI agents could still produce superhuman collective intelligence. In this scenario, quantitative scaling of agent populations, rather than qualitative leaps in individual reasoning, could be sufficient to surpass the capabilities of large human organisations.
The researchers also discuss what they call an “abstraction barrier”, a potential limitation on individual AI systems. The hypothesis suggests that models trained primarily on human-generated knowledge may inherit the conceptual frameworks embedded in that data and struggle to discover entirely new abstractions.
“While this barrier could potentially cap the intelligence of any single AI instance at AGI-level, collective ASI might still be achievable through multi-agent scaling.”
The paper also revisits the concept of an intelligence explosion. While continued scaling implies exponential growth, the authors argue that recursive AI-driven research could create feedback loops that accelerate growth rates themselves, producing hyperbolic dynamics often associated with singularity scenarios.
“While today the pathway of scaling (models & data) seems most promising to deliver progress, it is unclear how long exponential growth rates can be sustained economically and in terms of hardware production and natural resources,” the authors write.
The paper points to several potential constraints on future AI progress, including hardware supply, energy availability, the exhaustion of internet-scale data sources, economic limits and unresolved questions around AI alignment.
The authors do not provide formal timelines for AGI or ASI. “The long-standing goal of creating AGI may come into reach for our generation, perhaps within the next decade or less,” the researchers write.
ALSO READ: The Playground is Closed: 10 Hard Truths from the Cisco AI Summit
