On "From AGI to ASI"

The focal article for this commentary is From AGI to ASI, in which Genewein, Franklin, Lerchner, and their colleagues (2026) set out to map what might follow the arrival of human-level artificial general intelligence. It is the work of a team based largely at Google DeepMind — among them Shane Legg, one of the laboratory’s co-founders, and Marcus Hutter, whose formal account of machine intelligence underpins much of the report’s reasoning — writing alongside collaborators at the University of Waterloo, the Australian National University, and University College London. The result is a careful, deliberately uncertain document, and I offer what follows not as a verdict on it but as a reading from an adjacent field. I work in education and the learning sciences rather than in frontier AI research, and what draws me to the report is less whether its forecasts prove right than what they quietly ask of the institutions I do know well.

I have tried to write in the spirit of Berterö’s (2016) guidance on the commentary as a scholarly form, which counsels against re-summarizing the focal article, assumes the reader has it in hand, and asks instead for a clear take-home message. I will keep to that, and so I begin not with a recap but with the one feature of the report I have found hardest to set aside. In its opening section, the authors include a set of instructions addressed not to their human readers but to the AI assistants those readers will likely ask to summarize the paper. They specify what should be emphasized, what should not be compressed into shorter lists, and they invite future AI systems to update the report’s conclusions as events unfold.

It is worth pausing on that gesture. Here is a scholarly report written partly for the technology it describes, anticipating its own obsolescence and asking that technology to help keep it current. Whatever one ultimately makes of its specific claims, that framing alone says something about the moment we are working in — and it is the moment, more than any single forecast, that this commentary is concerned with.

Education: Teaching Toward a Moving Target

This is where the report unsettled me most. We design curricula, credentials, and pedagogical sequences around a relatively stable assumption: that human cognitive development follows predictable stages, and that what we teach today will remain relevant for the working life of the student in front of us. The report’s central claim — that the gap between human and machine capability may not just grow, but grow in kind, through mechanisms (lossless replication, substrate independence, arbitrary timescales) that have no human analog — puts that assumption under real pressure.

I do not read this as education becoming obsolete. If anything it becomes more important, but differently shaped. If machines increasingly hold the encyclopedic and procedural knowledge that schooling has long been organized around delivering, then what remains distinctly valuable is judgment: knowing which questions matter, how to evaluate an answer, how to live alongside systems whose internal workings we cannot fully inspect. Neil Lawrence’s argument, which the report cites, is the lever here: human cognition’s narrow communication bandwidth forces us to build abstractions, hierarchies, and compressed models of the world. That forcing function — the very limitation we sometimes treat as a deficiency — may be part of what makes human understanding transferable, teachable, and shareable in the first place. An education that helps students build and interrogate their own abstractions, rather than simply consume larger ones, looks more durable than ever.

Systems, Complexity, and the Limits of Forecasting

The report treats AI progress not as a date to be predicted but as a system with competing feedback loops — accelerating dynamics (AI speeding up AI research) racing against decelerating ones (the rising cost of each additional unit of progress) — and concedes that which loop dominates is, at the onset, nearly impossible to measure.

This is a systems-thinking lesson that extends well beyond AI, and it is the one I would press furthest. Societies routinely fail to prepare for transitions not because they cannot imagine them, but because exponential and hyperbolic dynamics are experientially indistinguishable from linear ones until very late. The report’s prescription — track quantitative indicators continuously, maintain a range of models rather than a single forecast, and revisit them often — is sound advice for any institution facing a system it cannot fully model, which describes most of the institutions I work with. I would add a caution the report leaves implicit: the instruments we choose to track will quietly determine the futures we are able to see. Measurement is never neutral observation; it is a commitment about what counts.

Knowledge, Intelligence, and What “Super” Means

The report’s use of the Legg–Hutter framework — defining intelligence as average performance across all computable tasks — is, to me, less a technical detail than a philosophical provocation. It reframes intelligence not as a trait an entity has, but as a measure of breadth across an unbounded space of problems. Under that lens, the question “is this system superintelligent?” dissolves into “superintelligent compared to whom, on what, and at what cost?” — questions that are inseparable from values, governance, and access. Whoever defines the distribution of tasks defines the winner. That is not a measurement problem to be cleaned up later; it is a question of values embedded in the metric from the start, and education is one of the few institutions positioned to teach people to recognize and ask it.

What a Time to Be Alive

I do not write any of this from a place of alarm, nor from uncritical excitement. I write it from something closer to vertigo — the feeling of standing at a point where multiple plausible futures genuinely diverge, and where the report itself admits it cannot tell us which one we are entering. That admission, paradoxically, is what makes the paper worth taking seriously, and worth re-reading, as the authors themselves suggest, with help from the very systems it describes.

So here is my take-home message, in Berterö’s sense of the term. The most useful response to a future we cannot specify is not a better prediction; it is an education that builds judgment, interrogates its own abstractions, and treats the measurement of intelligence as a question of values rather than a settled fact. We, all of us working in education and the social sciences, are being asked to prepare students, institutions, and ourselves for a range of futures we cannot specify in advance. That is a genuinely strange position to be in. It is also, I think, a privilege: to be present, attentive, and working, at a moment when the questions are this large and this open.

About the Focal Article: From AGI to ASI

The following is a narrative summary of the report this commentary responds to. In keeping with the guidelines above, it sits outside the commentary proper — included here so readers who have not yet read the original have the full picture in one place.

Genewein, Franklin, Lerchner, and colleagues (2026) begin from a premise that has, in their words, moved “from far-fetched speculation to being a concrete next-decade target for many of the largest AI organisations”: the arrival of human-level artificial general intelligence (AGI), systems that match a typical human across most cognitive tasks. Their real question is what happens after that point — the transition from AGI to artificial superintelligence (ASI), which they characterize as systems surpassing not just individuals but large organizations of humans in cognitive capability.

The report maps four roughly parallel pathways toward ASI. The first is continued scaling of compute, data, and models. The second is a shift in the underlying algorithmic paradigm — a new approach that supersedes today’s methods. The third is recursive self-improvement, in which AI systems accelerate AI research itself. The fourth is the emergence of intelligence from large collectives of AI agents coordinating the way corporations or institutions do.

Against each pathway the authors weigh frictions and bottlenecks: the eventual exhaustion of high-quality training data, the economics and energy demands of continued scaling, the difficulty of automating physical-world experimentation, and what they call the “Abstraction Barrier.” Whether these frictions prove negligible or substantial, they argue, remains an open research question — and one that cannot be settled from where we currently stand.

A distinctive section catalogs the structural advantages that digital intelligence holds over biological intelligence — advantages that grow as computers become more powerful. AI can be copied losslessly, not just its code but its accumulated “experience.” It can be paused, sped up, run in parallel, and moved across hardware. None of this requires AI to become smarter in any single sense; it requires only more compute, applied to something that was never bound by a body in the first place.

To define intelligence itself, the authors adopt the Legg–Hutter framework, which measures intelligence as performance averaged across all computable tasks. The report is deliberately uncertain throughout: it does not predict when, or whether, ASI arrives, and it treats AI progress as a system of competing feedback loops whose net direction is hard to read at the onset. Its central recommendation is therefore procedural rather than predictive — track quantitative indicators continuously, hold a range of models rather than a single forecast, and revisit them as evidence accumulates. The authors frame the whole prospect not as a single step change but as “a series of transformative societal changes,” requiring “a massively interdisciplinary endeavour of global scope and interest.”

References

This commentary is based on a close reading of the following report:

Genewein, T., Franklin, M., Lerchner, A., Orseau, L., Albanie, S., Bales, A., Wyeth, C., Chan, S., Gabriel, I., Leibo, J. Z., Dafoe, A., Hutter, M., Graepel, T., & Legg, S. (2026). From AGI to ASI. arXiv. https://arxiv.org/abs/2606.12683

Other works referenced above:

Berterö, C. (2016). Guidelines for writing a commentary. International Journal of Qualitative Studies on Health and Well-being, 11(1), 31390. https://doi.org/10.3402/qhw.v11.31390

Lawrence, N. D. (2024). The atomic human: Understanding ourselves in the age of AI. Allen Lane.

Legg, S., & Hutter, M. (2007). Universal intelligence: A definition of machine intelligence. Minds and Machines, 17(4), 391–444. https://doi.org/10.1007/s11023-007-9079-x

Cite this article

Gattupalli, S. (2026). On “From AGI to ASI”. Society and AI. https://societyandai.org/commentary/from-agi-to-asi-commentary/

Write for Society & AI

Have a perspective on what AI means for education, knowledge, or human flourishing? We’d love to hear it. Society & AI publishes freely, in open access, for anyone thinking seriously about these questions. Send your proposal to submissions@societyandai.org.