Newsletter Subscribe
Enter your email address below and subscribe to our newsletter
[forminator_form id="25163"]

institute.deepmind.com.institute.deepmind.com+1.institute.deepmind.com.On November 19, 2026, the DeepMind Institute published an essay arguing that AI could move beyond human-level artificial general intelligence (AGI) to artificial general superintelligence (ASI) by four routes. It names them as scaling models and data, technological paradigm shifts, better coordination among agents, and recursive self-improvement. The essay, "What comes after AGI?", was written by Google DeepMind research scientist Tim Genewein. It says rapid progress over the next decade is "a serious possibility" but not certain, and it asks researchers and policymakers to start tracking that progress now.institute.deepmind.com
The essay defines ASI as a system that "consistently outperforms large groups of human experts on tasks that could take them years to complete, using fewer resources and less time," across virtually all tasks of human interest. Narrow superhuman skill, such as in chess, Go or protein-structure prediction, does not meet that bar.institute.deepmind.com
The essay estimates that global "effective compute" has grown at least tenfold a year. That figure combines three trends: compute per dollar rising about 1.5x a year, algorithmic training efficiency improving 3 to 6x a year, and AI hardware spending growing about 2.5x a year. If the trend held for a decade, a cognitive workload that costs $1 billion in compute today would cost about $1,000. The essay calls this shift toward commoditized cognitive work "cognitive industrialization".institute.deepmind.com
Each route has bottlenecks. Scaling could run into shortages of high-quality data, diminishing returns and limits on power grids and chip fabrication. Paradigm shifts may be needed because current models fall short on continual learning and sample efficiency. Coordinating many agents still lacks "multi-agent scaling laws." Recursive self-improvement could stall if the feedback guiding it is poor, and it would likely move fastest in fields that can be checked automatically, such as mathematical proofs.institute.deepmind.com
The essay argues that the pathways compound, so restricting only one input, such as training compute, "won't be enough to stop overall progress." It also warns that holding back one input while the others advance could build up a capability "overhang".unite.ai
Its recommendation is to "model, measure, forecast, and repeat," through interdisciplinary work involving frontier labs and national and global institutions. Many of the needed metrics require access to frontier models and large amounts of compute for evaluation. The essay gives no date for when AGI or ASI might arrive.i10x.ai+1
The essay draws on the technical report "From AGI to ASI," dated June 12, 2026. Its 14 authors include Genewein and DeepMind co-founder Shane Legg, and it described the same four pathways. The institute says its pieces should not be read as the official view of Google Alphabet Inc. .institute.deepmind.com+1
"None of these predictions are certain," Genewein writes, "but the pathways and scenarios are also far from unfounded hype or fiction".institute.deepmind.com