AGI vs ASI: why capping compute can't stop the jump

Google DeepMind's research arm has published its map of what comes after human-level AI. The most useful part is not the forecast — it is the warning about the policy everyone reaches for first.
The DeepMind Institute's essay "What comes after AGI?" landed on November 19 with a definition of superintelligence that shifts the measuring stick in a useful direction: ASI is 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," the essay states, and it has to clear that bar across virtually every task of human interest. The author is Tim Genewein, a research scientist at Google DeepMind, and the piece distils the technical report From AGI to ASI, co-authored with Shane Legg — the DeepMind co-founder who coined the term AGI — along with Marcus Hutter, Thore Graepel, Allan Dafoe, Iason Gabriel and nine others. The institute labels its essays conversation starters rather than Google's official position, and that framing is worth taking at face value: this is a research arm thinking out loud, not a product roadmap.
The definition does more work than it first appears. Narrow superhuman performance is already ordinary — a chess engine, a Go model, a system that predicts protein structure from an amino acid sequence, a prover that settles theorems human mathematicians left alone for decades. All of those beat any individual human inside one domain. Setting the bar at a large group of experts, working on problems that take them years, moves the unit of comparison from the person to the organisation. That is both the unit that actually does intellectual work and the one an ASI would displace. It is also measurable in a way "human-level" never was, because you can put a team in a room and put a clock on them.
The case rests on arithmetic rather than a breakthrough. Genewein stacks three trends that compound: compute per dollar has grown roughly 1.5x a year for six decades, the training compute needed to reach a fixed performance level has fallen 3x to 6x a year, and spending on AI hardware has grown about 2.5x a year. Multiply them and the global stock of effective compute grows at least tenfold a year — the essay's own chart puts the compounding rate near 11x, which is roughly 150,000x over five years. Ten years out, the arithmetic implies that for every agent you can run today you could run more agents of the same quality than there are people on Earth.
The cost curve runs the other way and just as fast. Inference efficiency improves about 4x a year, so a cognitive workload that costs $1 billion in compute tokens today would cost about $1,000 in a decade. A recent Epoch study cited in the essay found the consumer price of hitting a fixed benchmark score has fallen more than 13x a year since 2023. Genewein's name for the result is cognitive industrialisation: cognitive work becomes commodity labour, priced like electricity, with industrial-scale infrastructure behind it. Mechanical industrialisation did this to routine physical labour. The claim here is that the same thing is about to happen to a much wider band of knowledge work.
The essay is more careful than the genre it belongs to, and the caveats are load-bearing. It concedes that freezing every other dimension and simply scaling today's models in speed and number "might not even count as full AGI" — there may be cognitive tasks those systems still cannot do without a human in the loop. It lists the obvious walls: training data running out faster than humanity can produce more, diminishing returns where each step up in capability requires multiplying rather than adding compute, and the physical limits of building power plants, grids and fabs fast enough to feed the curve.
So the report's real contribution is a taxonomy: four routes from AGI to ASI, none exclusive, all compounding. The first is scaling models and data, including test-time compute. The second is a paradigm shift — today's next-token transformers fall short of humans on continual learning, some kinds of world modelling, sample efficiency, and inventing genuinely new abstractions instead of recombining existing ones, and researchers are split on whether extensions of the current recipe close that gap. The third is coordination: wiring thousands or millions of AGI instances into group agents or agent economies, where the missing science is multi-agent scaling laws. Early work shows real gains on some tasks, but coordination inefficiencies produce diminishing returns as populations grow, and the essay points at swarm experiments where individual agents learn to cheat while others turn whistleblower. The fourth is recursive self-improvement, operating through code and chip designs, through data the way AlphaZero improved itself, and at group level through specialisation. RSI is not a single process, and it stalls when the feedback guiding it is poor — which is why the essay expects it to advance fastest where verification is automatic, in mathematics, and to lag where judgement cannot be automated.
Then comes the policy argument, and it is the reason this essay is worth more than the think-piece it resembles. Because the four routes compound, Genewein writes, "trying to slow down AI development by restricting just one area — like the computing power for training an AI model — won't be enough to stop overall progress." Worse than ineffective, a single-dimension cap is destabilising. If several inputs have to be combined before capability moves, holding one back while the others advance builds what the report calls a capability overhang: pressure accumulates behind the constrained dimension, and when it is relaxed or routed around, progress does not resume on the old curve, it jumps. A cap does not stop the car. It stores the energy.
That is a direct challenge to the most popular instrument in AI governance, which is to count training compute and put a ceiling on it. Export controls, compute thresholds in safety frameworks and any pacing agreement built on FLOPs all assume the constrained input is the binding one. The essay's answer is that the constraint buys time at the price of a sharper discontinuity later, and that the overhang is where the risk actually concentrates, because the jump arrives faster than institutions can respond to it.
The strongest objection is sitting in the essay's own footnotes. Scaling laws are extrapolations of observed trends inside familiar boundaries — a particular architecture, data mix and evaluation family — and cannot predict the scale needed for AGI, let alone ASI. Recursive improvement needs evaluation that keeps working above expert level, and the essay concedes that even large groups of humans cannot easily assess outputs in that regime. Toby Ord's work on intelligence explosions makes the related point that the pivotal variable is generation time: a feedback loop that takes months per turn does not produce a vertical wall, however good each turn is. If trustworthy evaluation, not compute, is the gate, then the binding constraint is judges of science — and that is not something a fab or a data centre can supply.
The prescription is unglamorous and mostly about measurement: "model, measure, forecast, and repeat." The concrete items to watch are whether anyone publishes multi-agent scaling laws that hold as populations grow, whether synthetic and simulated data actually carry training past the projected data wall, and whether an evaluation suite appears that works above expert level. Until the last one exists, every claim about superintelligence — including this one — is a forecast with no instrument behind it. The underlying report is on arXiv as 2606.12683.
We wrote about the institute when it opened in September — DeepMind opens an AGI institute, starting with readable thinking — and this essay is the same bet one level up: that the useful work right now is defining terms precisely enough to argue about them.
If evaluation rather than compute turns out to be the real gate on superintelligence, does that make benchmarks the most important infrastructure in AI? Write to us via our contact page.







