Looking beyond the AGI threshold
The Google DeepMind Institute has published an essay examining what could follow artificial general intelligence, shifting attention from machines that broadly match people to systems that outperform large groups of experts across almost every task of human interest. The authors call that more demanding category artificial superintelligence, or ASI, and argue that researchers need better ways to map and measure any transition toward it.
The essay treats intelligence as a continuum rather than a single threshold. In its framework, human-level AGI is one point, while ASI would operate farther along the spectrum: completing work that might take expert groups years, using less time and fewer resources. Examples include compressing the scientific progress of an entire field or reproducing a large company's annual advance in chip design. The proposed standard applies broadly, not only to narrow systems already capable of surpassing people in games, protein-structure prediction or selected mathematical problems.
The authors emphasize that physical, logical and computational limits still constrain any system. Complex real-world phenomena, including climate, immune systems and societies, cannot be predicted with arbitrary precision. Yet theoretical ceilings offer limited guidance about the large, poorly charted space between current machines, human-level systems and the maximum capability permitted by those constraints.
One route considered in the essay is scale. It estimates that global effective computing capacity for AI has grown through a combination of hardware investment, improved chips and algorithmic efficiency. If strong trends continued, the same cognitive workload could become dramatically cheaper, allowing far more agents to run at higher speed. The authors describe the possible result as cognitive industrialization: standardized, automated intellectual work delivered through industrial-scale infrastructure.
That projection is explicitly conditional. The essay says practical limits on computing infrastructure and uncertain limits on software efficiency may slow the trend. It also notes that multiplying today's agents would not necessarily supply abilities they lack. A vast number of systems with the same blind spot may still fail on a task requiring a genuinely new method. Scale could nevertheless produce a qualitative change if broadly human-level agents became cheap, fast and numerous enough to work in parallel.
The Institute therefore considers multiple technological pathways rather than assuming that larger deployments alone will yield ASI. Continued model scaling, additional test-time computation and other advances could unlock new capabilities, but relationships observed in current scaling laws are extrapolations, not guarantees. That distinction matters because the essay discusses possible futures, not a claim that AGI or ASI has already been achieved.
The central policy message is preparation under uncertainty. If progress accelerates after human-level capability, changes in science, the economy and society could arrive faster and at greater scale than the first AGI milestone itself. Tracking capability, cost, speed and breadth separately may give governments and researchers a clearer view of whether systems are moving from isolated strengths toward the wide, consistent performance implied by superintelligence.



