Casual office meeting with diverse adults in a modern workspace in London, UK
Models & capabilityBased on company claims

From AGI to ASI: the four routes DeepMind says lead to superintelligence

By Himanshu Sakre·

Photo: Miguel González / Pexels

A Google DeepMind researcher published an essay on November 19, 2026 arguing that scale, new paradigms, agent swarms and self-improvement compound. The essay is a view, not company policy, and it ends on measurement, not alarm.

Google DeepMind has put a number of its senior researchers' names on a single question: what comes after AGI? On November 19, 2026, research scientist Tim Genewein published an essay on the DeepMind Institute site summarizing a longer report, From AGI to ASI, that lays out four routes from human-level AI to artificial superintelligence. The essay opens with a line that would have read as provocation two years ago: "Now that AGI is coming within reach, we can no longer put off seriously considering what comes next."

Two caveats come first. The Institute's pieces are labeled as conversation starters that reflect the author's views, not Google's official position. And the underlying paper is a preprint on arXiv (2606.12683), first submitted June 10, 2026 and revised to its third version on November 10, 2026. It lists 14 authors, including Shane Legg, Allan Dafoe, Iason Gabriel and Marcus Hutter. Nothing here is a product, a benchmark result or a safety evaluation. It is a framework.

What counts as ASI

The authors define artificial superintelligence as a system that "consistently outperforms large groups of human experts" on long tasks, using fewer resources and less time, and across virtually all tasks of human interest. They treat intelligence as a continuum, with human-level AGI as a point on it, and write that human-level AGI "has long been the horizon for AI research, but it is far from the upper limit of intelligence." Known theoretical limits sit far above current systems. The space in between is, in their words, largely unexplored.

The compute arithmetic behind the argument

The essay's case rests on rates of change. Compute per dollar has grown about 1.5x a year. The training compute needed for a fixed level of performance has fallen roughly 3x to 6x a year over the last decade, and inference algorithmic efficiency has improved about 3x a year. AI hardware spending has grown about 2.5x a year, and frontier training runs have grown more than 4x a year. Stacked together, the authors put global effective compute stock growth at 10x a year or more, a factor of 10 billion over ten years.

The consequence they draw is cost. A workload that costs $1 billion in compute tokens today could cost about $1,000 in ten years, if the efficiency trends hold. They cite an Epoch study finding that the consumer price for a fixed benchmark score has fallen more than 13x a year since 2023, and project that the effective population of AI agents could rise more than 1,000-fold by 2030. These are extrapolations. The essay is clear that scaling laws are "extrapolations within familiar boundaries," and it names the limits: a data wall, diminishing returns on model size, and physical constraints on power, chips and interconnects.

The four routes

Scaling models and data. More training compute, more data, more test-time compute. Workarounds for the data wall include high-fidelity simulation, reinforcement learning environments, and using inference compute to generate new training data.

“Our best course of action may sound familiar: model, measure, forecast, and repeat.”

Tim Genewein, research scientist, Google DeepMind, DeepMind Institute essay, November 19, 2026

Paradigm shifts. Today's systems are large transformers trained to predict the next symbol and then refined for reasoning and instruction following. The essay lists where they still lag people: continual learning, some kinds of world modeling, sample efficiency and inventing genuinely new concepts. Researchers disagree on whether incremental extensions are enough.

Agent coordination. Many AGI instances could form group agents or whole agent economies. Early research shows gains on some tasks, but coordination overhead produces diminishing returns, and the authors flag collusion and manipulation risks. They add a pointed remark about volume: "a conference of one hundred thousand Ptolemaic astronomers probably wouldn't have" produced modern astronomy.

Recursive self-improvement. AI systems automate "improvements to future generations of AI," through code and designs, through data, or through specialized roles inside agent collectives. The essay warns that "RSI loops can easily stall or degenerate if the feedback guiding them is poor," and says the loops should advance fastest where verification can be automated, such as mathematical proofs.

How this sits next to what labs say they are doing

The self-improvement route is no longer hypothetical talk. In a September 7, 2026 report on OpenAI's own disclosures, The Decoder wrote that OpenAI says it reached its "automated research intern" goal "according to our measurements," and that its research organization was running 3.1 agent workdays for every human workday as of mid-August. The Decoder also noted that all of this data is internal and that OpenAI did not describe independent review. Our own explainer on recursive self-improvement and the Anthropic threshold covers how the labs define the line.

DeepMind's essay is more careful than the lab announcements. It says whether recursive self-improvement produces runaway progress "is an empirical question" and that research tends to get harder over time. That caution matters, because the essay's central policy claim depends on it.

The claim that matters most

The routes compound and are not mutually exclusive, the authors write, so restricting one dimension, such as training compute, is unlikely to halt overall progress. Suppressing one dimension could even create a capability overhang, where a later release produces a sudden jump. That is an argument about the design of governance, and it cuts against compute thresholds as a lone safeguard. It is also an argument made by the people building the systems, which is a reason to read it and a reason to read it skeptically.

The practical recommendation is modest and checkable: "model, measure, forecast, and repeat." Tracking progress, the authors say, requires frontier-model access, substantial evaluation compute, and new benchmarks. Notably, it does not ask for a pause.

Why this matters for the people reading it

Most coverage of superintelligence treats it as a forecast to be believed or dismissed. This essay is more useful as a vocabulary. If a lab says its next model is a step toward ASI, a reader can now ask four plain questions: is this more scale, a new method, a coordination gain, or a self-improvement loop, and who measured it? Each route has its own evidence standard, and only the first has a long public record. That is a better test than any timeline, and it is the same discipline we apply in our explainer on loss of control.

Our take

This is a useful taxonomy and a weak prediction. The four routes give reporters and regulators a checklist: for any lab claim, ask which route it rests on and what evidence exists for it. What is missing is evidence. There are no multi-agent scaling laws yet, no independent confirmation of a self-improvement loop, and the cost trends are extrapolated from a short history. We would treat the 10x-a-year compute figure as an assumption to test against Epoch's and the labs' own disclosures, and we would watch for the first outside evaluation of agent collectives. The essay is a sign that the labs now think superintelligence deserves a research agenda. It is not a sign that anyone has found the road.

Frequently asked questions

What is the difference between AGI and ASI?

In the DeepMind paper, AGI is human-level general intelligence and ASI is a system that consistently outperforms large groups of human experts on long tasks, using fewer resources and less time, across virtually all tasks of human interest. The authors treat intelligence as a continuum, with AGI as one point on it.

What are the four routes from AGI to ASI in the DeepMind essay?

Scaling models and data, technological paradigm shifts, improving coordination among many agents, and recursive self-improvement, where AI systems automate improvements to later generations of AI. The authors say the routes compound and are not mutually exclusive.

Is this essay Google's official position on superintelligence?

No. The DeepMind Institute labels its essays as conversation starters that reflect the author's views and should not be read as Google's official position. The underlying paper is an arXiv preprint with 14 authors, not a company policy document.

How fast does the essay say effective compute is growing?

At least 10x per year, which the authors say compounds to a factor of 10 billion over ten years. The figure combines hardware cost trends, algorithmic efficiency gains and rising spending. It is an extrapolation, not a measurement.

Does the essay say recursive self-improvement will cause runaway progress?

No. It calls that an empirical question, notes that research tends to get harder over time, and warns that self-improvement loops can stall or degenerate if their feedback is poor. It expects the loops to work best where verification is automated, such as proofs.

What does DeepMind recommend doing about it?

The essay's answer is to model, measure, forecast and repeat. Doing that needs frontier-model access, substantial evaluation compute and new benchmarks. It does not call for a pause, and it warns that restricting one route, such as training compute, may only create a capability overhang.

Sources

What each one is, and whose it is.

  1. 1

    What comes after AGI? From AGI to Superintelligence, DeepMind Institute (November 19, 2026)

    OtherThe vendor’s own
  2. 2

    From AGI to ASI (arXiv:2606.12683), arXiv (June 10, 2026)

    PaperThe vendor’s ownNot peer reviewed, preprint
  3. Press reportIndependent of the vendor
  4. Press reportIndependent of the vendor