DeepMind Wrote a Paper About What Comes After AGI — and Assumed AI Would Explain It to the Humans in Charge

The summary instructions at the front of the From AGI to ASI report are addressed not to researchers but to AI systems acting on behalf of decision-makers. That design choice is the most important sentence in the document — and it appears before the abstract.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/deepmind-wrote-a-paper-about-what?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

On 10 June 2026, a fourteen-person team of researchers at Google DeepMind submitted a report to arXiv titled From AGI to ASI. The lead author is Tim Genewein. The team includes Marcus Hutter, who formalised the mathematical theory of universal intelligence on which the paper builds, and Shane Legg, a DeepMind co-founder and its chief AGI scientist, who completed his doctoral work under Hutter’s supervision. The report investigates how AI might continue to develop in a post-AGI world — how a system that matches human capability across most cognitive tasks might transition, through four distinct pathways, toward artificial superintelligence: a system more capable than large, well-coordinated groups of human experts working together across virtually every domain.

The paper’s content is significant and will be addressed in this article. But the most important thing about the paper appears before the abstract. It is a section titled Summary Instructions. It is not addressed to the human researchers who wrote it. It is not addressed to the academic reviewers who might evaluate it. It is addressed to AI systems that may be called upon to summarise the report on behalf of humans — instructing those AI systems to clarify definitions without compression, not to collapse lists, and to judge whether the conclusions remain valid over time. This is the first academic paper in history whose authors assumed, explicitly and structurally, that AI would read it first and explain it to the humans nominally in charge of understanding it.

For boards and executives, that assumption is the most important thing to understand about where we are.

CONTEXT AND BACKGROUND

For most of the past five years, the governance conversation about AI has been framed around the question of when AI will become capable enough to matter at a strategic level. The DeepMind paper shifts that frame. It takes AGI — a system operating at roughly the median human level across most cognitive tasks — as its starting point rather than its destination. The paper’s central argument is that AI progress is unlikely to stall at human level, and that the transition from human-level AGI to artificial superintelligence may happen through multiple simultaneous pathways rather than as a single transformative event.

The three tiers the paper defines are precise and deserve exact treatment. AGI is a system that performs at roughly the median human level across most cognitive tasks — not the most capable person in the room, but a competent generalist capable of reasoning, learning, planning, communicating, using tools, and adapting to new situations. ASI is not a system that is modestly better than a human expert. It is a system or collective that exceeds the combined output of tens of thousands of top experts working in tight, well-coordinated collaboration for an entire decade — across virtually every domain simultaneously. The bar is explicitly civilisational in scale. Universal AI, identified with Hutter’s AIXI framework and the Legg-Hutter intelligence measure, is the theoretical ceiling — the mathematical upper bound of intelligence, computable in principle but not in practice, approachable from below and never reachable. It functions in the paper as a formal north star that gives rigour to the definitions below it.

The paper identifies four pathways from AGI to ASI. The first is scaling — more compute, more data, more algorithmic efficiency, combining to produce systems that improve along existing trajectories. The second is paradigm shifts — fundamental architectural changes that move beyond current transformer-based systems toward new forms of reasoning, memory, and world modelling. The third is recursive self-improvement — AI assisting in AI research, producing better AI, which in turn assists in more AI research, compounding iteratively. The fourth, and in the paper’s framing perhaps the most underappreciated, is multi-agent collectives — vast networks of human-level AI instances that, connected by lossless communication and frictionless coordination, collectively exceed what any individual instance could achieve, functioning as a digital civilisation capable of running parallel experiments, spawning specialised agents on demand, and sharing knowledge instantly across every node.

INSIGHT AND ANALYSIS

The Summary Instructions section is where the paper’s most significant governance implication is located — not in its projections about when AGI will arrive, not in its analysis of which pathway will dominate, but in the structural assumption it encodes in its opening pages.

The authors are researchers at one of the world’s most advanced AI laboratories. They have written a detailed, technically demanding document about the future trajectory of intelligence. And they have included, at the front of that document, explicit guidance for the AI systems that will summarise it on behalf of the human decision-makers who need to understand it. They tell those AI systems not to compress the lists, because the distinctions matter. They tell them to clarify the definitions, because the definitions are doing precise work. They tell them to judge whether the conclusions remain valid over time, because the field moves faster than the documents that describe it.

That design choice is not a quirk of academic formatting. It is a window into how the people who build AI are already thinking about the relationship between complex technical documents and the humans nominally responsible for acting on them. The assumption is that AI will mediate. The assumption is that the human reader will receive a summary rather than engaging with the primary text. The assumption is that this is not a failure of governance but an operational reality to be designed for — explicitly, structurally, at the beginning of the document.

For boards and executives, this is the most direct signal available about the information environment they are already operating in. Most strategic documents that inform board decisions are not primary sources. They are summaries, briefings, presentations, and analyses produced by teams working under time pressure, using tools whose provenance the board does not control.

The DeepMind paper’s Summary Instructions assume AI will be in that chain — not someday, but now — and design accordingly. The question for boards is not whether AI is already mediating between complex information and the humans responsible for making decisions about it. The question is whether the boards that depend on those summaries have asked what assumptions the AI is making when it produces them, and whether those assumptions are the ones the board would endorse if it knew about them.

The four pathways to ASI carry their own board-level implications. The multi-agent collective pathway is the most practically consequential for organisations because it suggests that superintelligence does not require a single architectural breakthrough. It can emerge from infrastructure — from assembling enough human-level AI instances, connecting them effectively, and directing their coordination toward high-value problems. The paper runs a thought experiment: a single human-level AGI deployed as 100 million simultaneous instances, connected by lossless communication, coordinating without the friction of meetings, emails, or the time required to explain concepts across human cognitive bandwidth. That collective would not be superintelligent at the individual node level. Collectively, it would likely qualify as ASI. That is not a distant theoretical scenario. It is a description of where the trajectory of AI infrastructure investment — the data centres, the compute clusters, the orbital infrastructure — is pointing. The organisations that control that infrastructure control the pathway to ASI from the collective direction.

The abstraction barrier is the friction the paper identifies that is most relevant to the professional services, legal, financial, and consulting sectors that form a significant part of Johan’s audience. The paper’s argument is that AI can master existing human abstractions — the legal frameworks, accounting standards, coding paradigms, and scientific categories that already exist in the training data — with increasing facility. Where it may struggle is in inventing fundamentally new ones: new ways of conceptualising a legal problem that existing law has not yet addressed, new frameworks for assessing risk that existing accounting standards have not yet formalised, new approaches to a scientific question that existing categories have not yet named. The abstraction barrier is not a permanent limit on AI capability. It is a description of where human creative judgment — the capacity to invent new frameworks rather than master existing ones — remains most genuinely irreplaceable. For organisations deciding what to develop in their people and what to delegate to AI, understanding where that barrier sits in their specific domain is one of the most practically important governance questions the paper raises.

IMPLICATIONS

For boards and executives, the DeepMind paper carries three specific implications that go beyond the scientific projections it makes.

The first is about planning horizons. Most corporate AI strategies are built around a timeline that treats AGI as the culminating event — the moment at which AI capability reaches a threshold that demands a strategic response. The DeepMind paper argues that AGI is the opening condition of a more consequential transition, not the end of the AI race but the moment the real race begins. The boards that are planning for AGI as the finish line need to understand that the paper’s authors are already planning for what comes after it — and that the governance, infrastructure, and competitive dynamics of the post-AGI world will be shaped by decisions being made right now.

The second is about information mediation. The Summary Instructions section signals that the researchers who produced the most important AI safety document of 2026 are already operating in a world where AI systems read complex documents on behalf of humans. Every board that receives AI-assisted briefings, summaries, or analyses needs to ask what assumptions the AI is making in producing those outputs, whether those assumptions are visible to the humans who depend on them, and whether the governance frameworks that define what the AI should and should not do in that mediating role are adequate to the decisions being made on the basis of its summaries.

The third is about the six frictions the paper identifies as potential brakes on the AGI-to-ASI transition. Of the six — the data wall, resource constraints, paradigm limits, research hardening, the abstraction barrier, and deliberate regulatory slowdown — the last two are the ones most directly influenced by human institutional decisions. The abstraction barrier is partly a function of how AI is trained and what it is trained on. Deliberate slowdown is entirely a function of how governments, regulators, and international institutions respond to the capability trajectory the paper describes. South Africa’s withdrawal of its national AI policy in April 2026 and its absence of a functioning AI governance framework at the precise moment the SARB has named AI as a systemic financial risk and Google DeepMind is publishing the trajectory from human-level AI to superintelligence — is not a position of neutrality. It is a position of unpreparedness in a race that the paper argues is unlikely to pause.

CLOSING TAKEAWAY

The From AGI to ASI paper is an unusual document. It is technically demanding, carefully reasoned, and honest about its uncertainties in ways that most AI coverage is not. It does not claim to know when AGI will arrive. It does not claim to know which pathway to ASI will dominate. It claims to know that the transition from human-level AI to something beyond it is more likely than not, that it may happen through multiple simultaneous mechanisms rather than a single dramatic event, and that the governance and institutional responses required are of global scope and massively interdisciplinary character.

And it opens with instructions for the AI that will summarise it on behalf of the humans who should read it. That is the most honest sentence in the document — not because it is pessimistic about human attention or capability, but because it is accurate about the operational environment that even the most sophisticated AI researchers are already designing for. The question for boards is not whether that environment applies to them. It already does. The question is whether they are designing for it as deliberately as the authors of this paper are.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net