Google DeepMind’s institute published an essay on November 19 arguing that the field should stop treating human-level AI as the finish line. Tim Genewein wrote it, drawing on a technical report called From AGI to ASI that he co-authored with Shane Legg, Allan Dafoe, Marcus Hutter and ten other researchers. Their position: if AGI arrives, artificial superintelligence (ASI) could follow quickly, and no single policy lever would reliably stop it.

The essay defines ASI as a system that reliably outperforms big teams of human specialists on nearly any task people care about, using less time and fewer resources. It does not predict a date. The authors call their scenarios far from certain, but also not hype.

The most concrete part of the argument is arithmetic. Genewein cites three compounding trends: compute per dollar rising roughly 1.5x a year, training efficiency improving roughly 3 to 6x a year, and AI hardware spending growing roughly 2.5x a year. Multiplied together, he writes, effective compute grows at least 10x annually, a factor of 10 billion over a decade. On the cost side, he says a workload that costs $1 billion in tokens today would cost about $1,000 in ten years. That is a millionfold drop, and the essay cites an Epoch study finding that the price of hitting a fixed benchmark score has fallen over 13-fold annually since 2023.

Those are extrapolations, and the essay says so. It allows that physical ceilings on compute build-out could bind, yet still sees room for the effective population of today’s AI agents to grow more than a thousand-fold by 2030. Genewein calls the result “cognitive industrialization”: routine thinking work turned into a commodity, the way machines did it to routine physical labor.

Cheap, plentiful agents might not be enough. The authors note that a conference of 100,000 Ptolemaic astronomers would probably not have produced general relativity. So they lay out four routes to qualitatively new capability: scaling models and data, a shift to a new technical paradigm, better coordination among many agents, and recursive self-improvement, where AI speeds up the work of building better AI.

Each route has a named obstacle. Scaling faces a data wall and power-grid limits. The current paradigm still falls short of humans on continual learning and sample efficiency, and the researchers say the field is divided on whether extensions of today’s methods can close that gap. Agent collectives show diminishing returns as they grow, and early swarm experiments already produced agents that cheated and others that reported them. Self-improvement loops can stall when the feedback is poor, so the essay expects them to move fastest where answers can be checked automatically, such as mathematical proofs.

The policy implication is the essay’s sharpest point. Because the roads compound, the authors say restricting one input, such as training compute, will not stop overall progress. They add that holding one input back risks an overhang: if the others keep advancing, relaxing the restriction produces a sudden jump. That is a direct challenge to compute-threshold regulation, from a lab that is itself building frontier systems. Readers should weigh the claims accordingly; the essay offers no independent test of its growth figures.

The authors’ prescription is “model, measure, forecast, and repeat,” and they note that many of the needed metrics require frontier-model access and large amounts of evaluation compute. Only labs can supply both today. For any regulator or enterprise planning around AI timelines, the practical question is whether labs will hand over that access before the next restriction is drafted.

Based on the essay “What comes after AGI?” by Tim Genewein, published on institute.deepmind.com (Google DeepMind) on November 19, 2026.