The most senior research team in artificial intelligence just stopped arguing about whether artificial general intelligence is coming. They started planning for what happens after it arrives. On June 10, 2026, fourteen researchers at Google DeepMind — including co-founder Shane Legg, the theoretical AI pioneer Marcus Hutter, and the lab’s policy lead Iason Gabriel — posted a 60-page paper to arXiv titled “From AGI to ASI”. It was published on DeepMind’s research page on June 12. The paper was almost universally ignored over the weekend until the AI explainer account @HowToPrompt__ pointed out on Saturday afternoon that “the most advanced AI lab in the world is no longer just researching how to build AGI. They are mapping out what AGI will build next.”

That sentence is the news. For a decade, the conversation in AI has been a debate over timelines. When does AGI arrive? Five years? Ten? Never? DeepMind’s paper does not pick a timeline. It does something more consequential: it takes AGI as a working assumption and walks forward through the transition to ASI — artificial superintelligence, which the authors define as a system “more intelligent and cognitively capable than large organisations of humans.” That is a stronger claim than any individual genius. It is an institutional-level intellect.

🔍 THE BOTTOM LINE

DeepMind’s paper is the first major research output from a frontier lab to treat AGI as a settled question. The interesting question, in the authors’ framing, is not whether AGI arrives but what arrives after, and how. They map four pathways from AGI to superintelligence, none of them mutually exclusive, all of them already in motion to some degree. The practical implication for New Zealand — and every small economy at the end of the global AI supply chain — is that we are not preparing for a single event. We are preparing for a cascading series of disruptions with no clean “ready” moment. That is the worst framing for a country that depends on borrowed time to adapt to each new technology cycle.

The Paper Itself

The author list is the signal. Tim Genewein leads, with Matija Franklin, Alexander Lerchner, Laurent Orseau, Samuel Albanie, Adam Bales, Cole Wyeth, Stephanie Chan, Iason Gabriel, Joel Z. Leibo, Allan Dafoe, Marcus Hutter, Thore Graepel, and Shane Legg. Three names matter more than the others.

Shane Legg is the DeepMind co-founder who, in his 2008 doctoral thesis at the Dalle Molle Institute for Artificial Intelligence, defined AGI in the form the field still uses and predicted human-level AI by 2028. He is now DeepMind’s chief AGI scientist. Marcus Hutter is the mathematical theorist behind AIXI, the formal model of universal artificial intelligence, and the author of the 2005 textbook “Universal Artificial Intelligence.” His presence on the paper is the closest the field has to a theoretical physicist endorsing a research programme. Iason Gabriel is DeepMind’s policy lead — the man who writes the alignment papers for the lab, the person whose job it is to think about what happens if the systems work.

The paper’s central claim is in the abstract, and it is more careful than the headlines about it. The authors are not predicting a timeline. They are making a structural argument: AGI, once achieved, is unlikely to be a single transformative step. The more plausible picture is a “series of transformative societal changes caused by AI-enabled progress and breakthroughs across many areas of science and technology.” The framing matters. A single event has a beginning, a middle, and an end. A cascade does not. The disruption compounds.

The Four Pathways

The paper organises its forecast into four pathways from AGI to ASI. They are not mutually exclusive. In the authors’ framing, all four are likely to be in motion simultaneously by the time AGI arrives. The full arXiv HTML version has the technical detail; the summary from The AI Insider is the cleanest explainer.

Scaling AGI. This is the current playbook. More compute, more data, more parameters, longer training runs. It is the path that produced GPT-4, Claude 3, Gemini 1, and every frontier model since. The paper does not say scaling is wrong. It says scaling has known physical limits — chip fabrication, energy supply, training data exhaustion — and that those limits are now visible. The question is not whether scaling continues. It is what fills the gap when it stops producing the same returns.

AI paradigm shifts. This is the path DeepMind is implicitly betting on with its 2026 research agenda. A paradigm shift means abandoning the transformer architecture for something fundamentally different — perhaps a learned-symbolic hybrid, perhaps a world-model architecture à la LeCun, perhaps something not yet on the whiteboard. The paper acknowledges that paradigm shifts are inherently unpredictable. They happen, when they happen, by surprise. The 2017 transformer paper was a paradigm shift. The 2020 GPT-3 scaling result was a paradigm shift. The next one will be, too, but its shape cannot be forecast.

Recursive self-improvement. This is the path with the longest theoretical pedigree and the shortest real-world track record. The idea: once an AI system reaches a sufficient level of general intelligence, it can begin to improve its own architecture, training pipeline, and reasoning capabilities. Each improvement accelerates the next. The feedback loop tightens. In the limit, the AI improves itself faster than humans can supervise. The paper treats this as a real possibility, not a thought experiment, but acknowledges that it requires the AI to be good enough at AI research to be useful at it. The first practical steps — AI systems that help design their own successors — are already in motion. A company called Ricursive Intelligence is doing exactly this for chip design. It is not yet recursive self-improvement. It is the precondition for it.

Multi-agent collectives. This is the path most in motion right now. The intuition: a single agent, no matter how capable, is bounded by its training and its context window. A million specialised agents, coordinating through standard communication protocols, can in principle exceed any individual agent’s capability through emergent specialisation and division of cognitive labour. The paper’s analogy is an ant colony — no individual ant is intelligent, but the colony solves problems no individual ant could solve. The current visible examples are agentic coding tools (Devin, Claude Code, Aider, the Xiaomi MiMo open-source coding model) and enterprise agentic platforms (Agentforce, ServiceNow agents, Microsoft Copilot Studio). None of them are superintelligent. All of them are the early shape of a multi-agent system.

The Asymmetry Argument

The paper’s most consequential technical point is in Table 1 — a list of the structural advantages digital intelligence holds over biological intelligence. The list is short and devastating. Digital intelligence can be perfectly copied. A human cannot. Digital intelligence does not sleep. A human must. Digital intelligence is not constrained by biological neuron firing rates. A human is. Digital intelligence can run on faster substrates as they are invented. A human cannot speed up a human brain by upgrading the motherboard.

The implication: once AGI arrives, the gap to ASI may be small, not large. A human-level AI, when it can copy itself a million times, when it does not need to sleep, when it runs on hardware that is improving year-over-year under Moore’s-law-like dynamics, is already substantially more capable than a single human. The path from “human-level” to “institution-level” is, in this framing, a matter of resource allocation, not a matter of inventing new science. The paper is careful to say this is a possibility, not a prediction. But the framing is the part that should make a reader pause.

The Hassabis Disconnect

The timing of the paper is the part that demands scrutiny. On Friday evening, New Zealand time, DeepMind CEO Demis Hassabis gave a public talk proposing the “Einstein Test” — train an AI on pre-1905 physics, then see if it can independently derive relativity. He placed AGI “years, not months” away. The Einstein Test was, in his framing, a credibility check against the AGI-imminent narrative coming from OpenAI, Anthropic, and the sovereign-AI programmes now descending on the G7 in Évian.

One day before that talk — on June 12 — the “From AGI to ASI” paper went live on DeepMind’s research page. Shane Legg, Hassabis’s chief AGI scientist, is a co-author. The same company, in a 48-hour window, gave two distinct messages: the CEO told the public AGI is years away; the chief AGI scientist co-authored a paper that treats AGI as a working assumption and walks forward through ASI. Both can be true. The Einstein Test is a falsifiable benchmark for whether current systems are doing genuine reasoning. The ASI paper is a structural map of what happens after. They are not in contradiction. But they are sent at the same audience, and the audience will read the dissonance.

The honest read: DeepMind’s research arm is no longer in the “is AGI coming” business. It is in the “what comes next” business. The CEO’s job is to manage the public timeline. The scientist’s job is to plan for the actual one. Those are different jobs, and they produce different messages.

What It Means for New Zealand

The paper’s “cascading transformation” framing is the worst possible news for a small economy at the end of global supply chains. There is no “ready” moment in a cascade. There is no clean policy window in which to regulate, prepare, or catch up. There is a continuous sequence of breakthroughs, each of which shifts the policy and economic landscape in ways that require constant adaptation. A country the size of New Zealand — five million people, no frontier model lab, no hyperscaler, no sovereign AI infrastructure — does not have the institutional bandwidth to adapt to continuous disruption.

The paper is also, by accident, the strongest argument yet for the sovereign AI framework Macron is pushing at this week’s G7. If AGI is a cascade rather than an event, and if each breakthrough compounds on the last, then the countries that own the cascade — the United States, China, the United Kingdom, France — capture the value. The countries that consume the cascade — everywhere else — pay for it. The sovereign-AI response is to build local capacity so that the cascade is at least partially locally owned. For a country of New Zealand’s size, that is not a realistic goal for the frontier models themselves. It is a realistic goal for the applied layer — the local data, the local fine-tuning, the local deployment — that sits on top of the frontier.

The local-AI hardware story also has new weight in this framing. If the cascade is real, and if frontier-model capability is going to keep improving, then a $2,000 box that runs a 235B model at usable speed is a hedge. It is not enough on its own. But it is a node of local capacity that New Zealand households, schools, and small businesses can actually deploy, at a price they can actually pay, with no subscription. The first country to put a usable local AI box in every school and library is the first country to own a small piece of the cascade.

⚠️ THE OTHER SIDE

Three honest caveats. First, the theoretical framework. Hutter’s universal intelligence formalism, which the paper leans on, is rigorous but not constructive. It tells you the shape of an optimal agent in the limit of infinite compute. It does not tell you how to build one. The paper’s predictions are structural, not calendarable. Read it for the shape of the future, not the date. Second, the recursive self-improvement claim. AI helping to design better AI chips (Ricursive is doing this) is not yet recursive self-improvement. It is human researchers using AI to accelerate a process humans still drive. The gap between “AI as a research collaborator” and “AI as its own research lead” is large and may not be crossed in this decade. Third, the multi-agent hive-mind claim. Coordination is hard. Most multi-agent systems today fail more often than they succeed, and the failure modes (agents contradicting each other, agents optimising for the wrong shared reward) are not solved. The paper acknowledges this. It does not solve it.

❓ FREQUENTLY ASKED QUESTIONS

What is the paper actually called? “From AGI to ASI.” It is on arXiv as 2606.12683, published on DeepMind’s research page on June 12, 2026, and available in full HTML. It is 60 pages long with 14 authors.

Who are the big names on the paper? Shane Legg (DeepMind co-founder, coined the term “AGI” in his 2008 thesis, chief AGI scientist), Marcus Hutter (theoretical AI, AIXI, “Universal AI” textbook), Iason Gabriel (DeepMind’s policy lead, alignment research), Allan Dafoe (governance, formerly at FHI and OpenAI), Joel Z. Leibo (multi-agent systems). The lead author is Tim Genewein.

What are the four pathways from AGI to ASI? Scaling (more compute, data, parameters), paradigm shifts (new architectures beyond transformers), recursive self-improvement (AI improves its own code), and multi-agent collectives (millions of agents coordinating into emergent superintelligence). The paper says all four are likely to be in motion at once.

Does the paper give a timeline for AGI or ASI? No. The paper explicitly avoids timeline prediction. It says the transition from AGI to ASI is “deeply uncertain” and that AI progress “cannot be ruled out” as continuing to accelerate. The structure is the message, not the date.

How is this different from previous DeepMind papers on AGI? Previous papers — including the 2024 “Levels of AGI” paper with researchers from 14 institutions — focused on defining AGI and measuring progress toward it. This paper assumes AGI and asks what happens after. It is the first major DeepMind paper to make that move.

Why did almost no one talk about it until Saturday? The paper was posted to arXiv on June 10, before the WWDC 2026 news cycle absorbed the AI press. The G7 summit announcement, the Hassabis Einstein Test, the Anthropic and OpenAI IPO filings, and the Lisa Su local-AI demo all hit in the same window. A 60-page theoretical paper does not compete for attention with a CEO on stage. The @HowToPrompt__ thread is what surfaced it to a wider audience.

What should a New Zealand policy reader take from this? The paper is the strongest academic argument yet for the sovereign-AI policy response. If the transition is a cascade, not an event, there is no clean moment to catch up. The countries that own the cascade own the next decade. The countries that do not, do not. New Zealand’s near-term policy response is to invest in the applied layer (local data, local fine-tuning, local deployment), support the local-AI hardware ecosystem, and treat the G7’s sovereign-AI framework as a real commitment, not a talking point.

Is the paper optimistic or pessimistic? The paper is careful. It does not predict doom, and it does not predict utopia. The abstract says the right response is a “massively interdisciplinary endeavour of global scope and interest” — which is the academic way of saying: this is bigger than any one country, any one lab, any one policy framework can handle. The most senior research team in AI is telling the world: prepare for a future we cannot fully model, in a timeframe we cannot fully predict, with stakes we cannot fully quantify. That is not pessimism. It is seriousness.