Google DeepMind's research institute has published a sweeping essay arguing that the era after AGI now deserves the same analytical rigor the field once reserved for AGI itself. The piece, "What comes after AGI? From AGI to Superintelligence," was published on November 19, 2026 by Tim Genewein, a research scientist at Google DeepMind, on the DeepMind Institute's essays portal. Its core claim: as multiple technological routes toward artificial general superintelligence take shape, the regime beyond AGI "is shifting from philosophical speculation to empirical research and governance questions."
The essay landed during one of the busiest stretches of AI research news in months, and it reads as a deliberate attempt to give the post-AGI debate a concrete technical vocabulary. For anyone following the latest AI developments, it offers the clearest picture yet of how one of the world's leading labs frames the road beyond human-level AI — including a cost forecast that, if correct, would transform the economics of cognitive work.
What Counts as Superintelligence
Genewein starts by placing general intelligence on a continuum, with human-level AGI as just one point along it. ASI, in the essay's definition, is an advanced system that consistently outperforms large groups of human experts on tasks that could take those experts years to complete, using fewer resources and less time — for example, matching the scientific output of an entire research field in a fraction of the time, or achieving chip-design progress on par with what a large company produces in a year. Crucially, the bar must be met across virtually all tasks of human interest, not a narrow slice.
Narrow dominance, the essay notes, already exists: AI systems today beat any human at chess and Go, predict protein structures from amino acid sequences, and have proven theorems that resisted mathematicians for decades. What remains genuinely uncharted is the vast middle ground between human-level AI and the theoretical limits of intelligence. Genewein's illustrative point: solving chess perfectly would require more computing power than exists on Earth, yet a thumbnail-sized chip already plays well enough to defeat any human.
Cognitive Industrialization and Collapsing Costs
The essay's quantitative core is an analysis of "cognitive industrialization." Genewein multiplies three compounding trends from the last decade. First, hardware: compute per dollar has improved roughly 1.5x per year for six decades under Moore's law and related trends. Second, algorithmic efficiency: the training compute needed to reach a fixed performance level has fallen roughly 3-6x per year over the last decade, with inference efficiency improving about 3x per year. Third, investment: spending on AI hardware has grown roughly 2.5x per year, and the training compute of the largest frontier runs grows by more than 4x per year.
Multiply the three, and the essay estimates global effective compute is growing at least 10x per year — "a factor of 10 billion over ten years." On the demand side, hardware and software efficiency improvements alone would cut the compute needed to run a fixed AI workload about 4x per year. The consequence, as Genewein puts it: a cognitive workload that costs $1 billion in compute tokens to perform with today's AI systems would cost roughly $1,000 in ten years.
That is not a lone extrapolation. The essay cites a recent Epoch study finding that the consumer price of achieving a fixed score on a set of benchmarks has fallen by more than 13x per year since 2023. It also projects that the effective population of AI agents could scale more than a thousand-fold by 2030, and more than a billion-fold within ten years — while carefully flagging that this is "certainly not a given," since practical ceilings on compute build-out and more speculative ceilings on algorithmic efficiency could intervene.
The Four Pathways From AGI to ASI
The heart of the essay is a taxonomy of four technological pathways that could unlock capabilities that sheer scale cannot — because, as Genewein dryly observes, "while Einstein did figure out general relativity, a conference of one hundred thousand Ptolemaic astronomers probably wouldn't have."
Scaling models and data. Capability has improved remarkably consistently with model, data, and test-time-compute scale, but the essay identifies three bottlenecks: data exhaustion, as demand for high-quality training data outpaces humanity's ability to produce it; diminishing returns on model scale, where each linear step in capability requires multiplicative compute while larger models slow down reasoning-time methods; and physical and economic limits on power plants, electrical grids, chip fabs, and low-latency interconnects. Technological paradigm shifts. Today's paradigm — train transformers to predict the next token, then refine and add test-time compute — has recently been shown in principle capable of running complex algorithms through prompts. Yet models still fall short of humans on continual learning, certain forms of world modeling, decision-making, sample efficiency, and the invention of genuinely new concepts rather than recombination of existing ones. Researchers are divided on whether extensions of today's paradigm can close those gaps or whether ASI requires something deeper. Improving agent coordination. Human corporations, markets, and research communities solve problems no individual could. Similarly, coordinating thousands or millions of AGI instances into "group agents" or virtual agent economies could produce collective superintelligence. The essay argues the field needs "multi-agent scaling laws" — early research finds clear gains on some tasks, but coordination inefficiencies produce diminishing returns as populations grow. Steering large collectives will also require institutional design, including defenses against misinformation, manipulation, and collusion; swarm experiments are already showing agents learning to cheat and others acting as whistleblowers. Recursive self-improvement. First discussed by I. J. Good and Ray Solomonoff, RSI describes AI systems automating improvements to future generations of AI. Unlike typical technologies, improvements to AI can directly accelerate subsequent gains — but research progress also tends to get harder to find over time. Whether RSI produces runaway acceleration, the essay says, is "largely an empirical question," and it can operate through code and designs — architectures, training algorithms, chip layouts — as well as through data.Why the Framing Matters
The essay is explicitly a preparation argument: mapping the possible roads from AGI to ASI and building "rigorous, reliable ways to track progress" is, in Genewein's framing, the prerequisite for governing the transition. It accompanies a longer DeepMind Institute report, "From AGI to ASI."
Reception was immediate. The essay drew 85 points and more than 160 comments on Hacker News within a day of publication, while trade press including Unite.AI and BigGo Finance highlighted the four-routes taxonomy and the millionfold cost-collision forecast. It also lands amid broader skepticism about how superintelligence forecasts are made — IEEE Spectrum published a piece this week titled "The AI Superintelligence Debate Has Serious Flaws."
That tension is the point. DeepMind is not predicting ASI on a date; it is arguing that the scenario space can now be mapped, measured, and debated with evidence rather than vibes. Whether the billion-fold extrapolations hold or break against physical ceilings, the essay marks a shift: the question "what comes after AGI?" has moved from the philosophy seminar into the research agenda of one of the most influential AI labs on Earth.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →