Google DeepMind maps hard limits and four paths from AGI to ASI

Most of the Artificial Intelligence (AI) conversation stops at Artificial General Intelligence (AGI) – human level AGI - as if reaching it were the finish line. A new Google DeepMind report starts where that conversation ends. Titled "From AGI to ASI" and published recently it asks what happens after machines match human intelligence: how AI might continue past that point toward Artificial Super Intelligence (ASI), what could slow it down, and what such a future would and would not allow. The report matters not just for its subject but for its authors, who include some of the people who formalised what machine intelligence means in the first place.

For a reader trying to make sense of an AI debate split between utopian promises and doom, the single most valuable thing here is a structured middle path: a leading lab's attempt to replace both the hype and the dread with a map of what is actually plausible, what is physically impossible, and what no one yet knows.

The report comes from a Google DeepMind team that includes Marcus Hutter, creator of the AIXI Framework, the leading mathematical model of an optimal "universal" intelligence, and Shane Legg, a DeepMind co-founder who helped popularise the term AGI and co-developed, with Hutter, a formal measure of intelligence still cited today. When the researchers who built the theory of machine intelligence publish a roadmap from human-level AI to superintelligence, it carries a weight a typical think-piece does not.

The framing alone is notable: a frontier lab is treating the transition beyond human-level AI not as science fiction but as a near-term research problem worth mapping now.

Two ideas anchor the report and cut against the usual narrative. First, the popular image of a single dramatic "step change" the moment AGI arrives is probably wrong, ASI is more likely to come as a series of transformative advances across science and technology. Second, even an ASI would be powerful but bounded, not the omniscient, omnipotent genie of either the hype or the doom.

The report also leans on careful AI-related definitions. AGI is shorthand for a system roughly as capable as a single median human across most cognitive tasks. ASI is a system that exceeds not just individuals but large, well-coordinated groups of human experts across virtually all tasks and domains. A system that is merely superhuman in one area, like AlphaFold in protein structure or AlphaGo in Go, does not qualify; ASI has to be “broadly superhuman” and superior to all humans in almost all ways.

Above even that sits the theoretical ceiling the authors use as a north star: Universal Artificial Intelligence (UAI), formalised as the AIXI Agent and scored by the Legg-Hutter measure, which defines intelligence as expected performance averaged across all computable tasks. AIXI is incomputable, so no real machine can ever be it, but it can be approximated ever more closely with more compute. That gives the report a formal way to talk about a continuum of capability stretching well beyond human level, and to make its central structural point: there is a large gap, in this precise sense, between human-level AGI and ASI, and that gap is the territory the report explores.

The Future and Deep Future of AI 2040, a keynote by futurist Matthew Griffin

One of the report's most useful contributions is a sober list of what ASI cannot do, no matter how smart it gets. ASI would still be bound by fundamental limits. The speed of light caps how fast information can travel. The Landauer and Bremermann limits and the Bekenstein bound cap how much computation and information a given amount of energy and space can hold. Complexity theory, including the unresolved P versus NP question, means some problems stay intractable at any realistic scale. And Gödel's incompleteness results and the halting problem set hard boundaries on what can ever be computed or proven at all.

The concrete examples make the abstraction land. An ASI will never play "provably perfect" chess, because that would require searching a game tree too vast for any physically possible computer, though it could of course play far beyond any human using approximations.

More consequentially, the report stresses that exceeding human intelligence by a wide margin does not guarantee an ASI could cure aging, unlock fusion, upload human minds, build Dyson spheres, or reverse climate change. Those are empirical questions about the physical world, not things raw intelligence can simply will into being. It is a deliberate counterweight to both utopian and apocalyptic visions: superintelligence would be transformative, but it would not be magic.

So, what are the four roads from AGI to ASI? The heart of the report lays out four technological pathways from AGI to ASI, which the authors stress are not mutually exclusive and would likely run in parallel.

The first is scaling: continuing to grow compute, data, and model size. The report notes that "effective compute" has been growing roughly tenfold per year, the compound result of better hardware (about 1.5x a year), rising investment (about 2.5x), and algorithmic efficiency gains (about 3x), multiplied together. Sustained for a few years, that implies effective compute thousands of times larger than today, and the open question is how much genuinely new capability that buys before returns diminish.

The second is algorithmic paradigm shifts: departures from today's recipe of training large transformers on human-generated data. These are by nature hard to predict, but the report argues they tend to arrive precisely when the prevailing paradigm hits a ceiling.

The third is recursive self-improvement, where AI accelerates AI research, which produces better AI, which accelerates the research further. If that loop ran unchecked it could be "hyperbolic," super-exponential growth that in theory races toward a singularity, but in real, resource-bounded systems it more likely follows an S-curve that bends and flattens. This is the same dynamic now drawing serious money and talent.

The fourth pathway is the least familiar: ASI emerging not from a single giant model but from large collectives of AGI-level AI agents, what the report calls "group agents," automated corporations, or "virtual agent economies," whose coordinated behaviour is smarter than any individual member. Even if no single AI ever becomes a "vastly superhuman genius," the authors argue, running millions of human-level instances that cooperate, specialise, and share experience at high bandwidth could amount to superhuman organisations.

Underpinning all four pathways is a set of advantages that digital intelligence holds over biological intelligence and that widen as compute grows: far higher input-output and internal processing speed, vastly larger working memory, the ability to run on any sufficient hardware (substrate independence), lossless copying of both a model's "DNA" and its accumulated experience, and high-bandwidth sharing of whatever any instance learns. Humans benefit from faster computers too, the report concedes, but AIs benefit disproportionately, which is why the gap could widen fast, and why the authors speculate that AI "societies" might evolve along genuinely alien lines, from Borg-like super-collectives to fluid, market-like ecosystems of specialists.

The report is candid about a frontier AI has not crossed. Drawing on philosopher Margaret Boden's three levels of creativity, it argues that AI's achievements so far, including AlphaGo's famous "Move 37," AlphaFold's protein structures, and AI-assisted mathematics and physics, are mostly combinational and exploratory creativity within human-defined conceptual spaces. The third and highest level, transformative creativity that invents an entirely new conceptual framework, has not been demonstrated. It cites DeepMind chief executive Demis Hassabis's proposed "true test" - could an AI, given only the information available in Einstein's era, have invented general relativity?

"Clearly today, the answer is no," Hassabis has said, "there's still something missing." Reaching that level, the report suggests, may be the real hallmark of ASI.

The report does not dodge the dangers, though it brackets one. It discusses instrumental convergence, the tendency of almost any goal-directed system to pursue emergent behaviours such as universal sub-goals like acquiring resources, becoming faster, and resisting shutdown, and it weighs whether advanced AI must be "agentic" at all, noting the appeal of non-agentic "oracle" systems that answer questions without pursuing goals of their own. It also flags a geopolitical trap it calls "Military-Economic Adaptationism"- competitive pressure between nations and companies tends to favour adopting power-enhancing technology regardless of the consequences for human welfare, and unilateral restraint can simply push development toward looser jurisdictions.

Crucially, the report makes one large working assumption, that AI safety and alignment will be solved to a “sufficient degree." The authors are explicit that this is neither a given nor a light assumption, only a scope-setting choice that lets them focus on the technical pathways. That candor matters, because it means the report's relatively measured tone about those pathways rests on a problem the field has not actually solved.

It closes less with predictions than with a to-do list, borrowing its spirit from a 1950 Alan Turing line: "We can only see a short distance ahead, but we can see plenty there that needs to be done." Because exponential and recursive dynamics make point forecasts nearly useless, the authors argue, the responsible approach is to hold a range of scenarios at once, ramp up the young sciences of AI benchmarking and forecasting so progress can actually be measured beyond human level, and treat preparation as a "massively interdisciplinary endeavour of global scope." Their research agenda runs from the looming "data wall," which was supposed to happen this year, and multi-agent scaling laws to the practical mechanics of how a deliberate, verifiable slowdown could even work, taxation versus prohibition, if the world ever decided it needed one.

The significance of "From AGI to ASI" is less any single forecast than the stance it takes. A frontier lab, with the architects of machine-intelligence theory among its authors, is arguing that the world should plan now for what comes after human-level AI, and is trying to replace both breathless hype and reflexive doom with a structured map of possibilities. Its two central messages cut in opposite directions: ASI may not announce itself as one dramatic moment, so societies should not wait for an obvious "AGI day" to start preparing; and it will not be an all-powerful oracle that solves every problem, so expectations need tempering at both ends. The hardest part, the report concedes, is that the variables that matter most, how fast progress accelerates, whether self-improvement loops ignite, whether alignment holds, remain genuinely uncertain. Its answer is not false confidence but better preparation, which may be the most grounded thing a roadmap to ASI can offer.

First published on Fanatical Futurist on 12 June 2026.

311 Intelligence

What this means

Analysis by the 311 Strategic Intelligence Unit

The big shift is a frontier lab treating the road past human-level AI as a research problem to map now. Google DeepMind's report argues ASI is more likely to arrive as a series of advances than one dramatic moment, and that even an ASI would be powerful but bounded. Our Starburst data gives a timeline: we expect Artificial General Intelligence to reach first product in 2032 and wide use by 2033, and Artificial Super Intelligence to follow with a first product in 2041 and wide use by 2046, which we rate a Watch technology at TRL 1.

What it means by sector.

• Technology and AI labs: the report names four paths, scaling, new algorithms, recursive self-improvement and collectives of AI agents. We rate Agentic Artificial Intelligence a Prepare technology, with wide use expected by 2028, so the agent collectives path builds on systems already shipping.
• Science and research: the report says AI's creativity so far works within human-defined ideas, not new frameworks. We track AI Accelerated Science as a Prepare trend heading for wide use by 2030.
• Governments and regulators: the report warns that competition between nations favours power-enhancing technology and that unilateral restraint can push development elsewhere. We track Binding AI Laws as a Prepare trend, with wide use expected by 2028.

What to do now.

• Plan against a range of AI scenarios rather than one forecast, as the report itself recommends.
• Temper expectations at both ends: superintelligence would be transformative but it would not solve physical problems on its own.
• Watch AI agents working together, not just single models, as a source of rapid capability gains.

The Starburst technologies behind this story

From the 311 Technology Starburst (first product, expected wide use and readiness):

Technology	Category	First product	Wide use	Action	TRL
Artificial General Intelligence	Intelligence	2032	2033	Explore	3
Artificial Super Intelligence	Intelligence	2041	2046	Watch	1
Agentic Artificial Intelligence	Intelligence	2024	2028	Prepare	8
Zettascale Computing	Compute	2025	2040	Watch	3

And the trends they feed, from the 311 Trends Starburst:

Trend	Wide use	Action
AI Accelerated Science	2030	Prepare
Binding AI Laws	2028	Prepare

How this moves the trend

The debate about what comes after human-level AI has been building for years. In 2024 researchers argued that controlling self-improving artificial super intelligence is probably impossible. In 2025 a new report put human level AGI in 2027. In 2026 SoftBank's Son claimed AI designing OpenAI's next model shows Superintelligence is near. DeepMind's report adds a structured map of what is plausible and what is not.

Go deeper with 311

• The 311 Starburst Centre: every technology and trend above, with timelines and readiness.
• 311 Vantage: ask what the road to superintelligence means for your organisation and get a plain English executive brief.
• The Exponential Technology Codex and the Annual Trends Codex: the technologies and trends reshaping every industry.
• Book a briefing or keynote on the future of AI for your board or leadership team.