The Report Written to Be Summarized by AI Ended Up Summarized Like All the Others

DeepMind assumes alignment is settled in order to map what comes after AGI; the coverage that followed mostly kept the single most spectacular number.

Photo by Logan Voss on Unsplash

June 10, 2026. Fourteen signatures on the cover page, two of whom formalized, nearly twenty years ago, what “intelligence” means in mathematics. The document is called *From AGI to ASI*. It isn’t a blog post or a product announcement: a research report, filed on arXiv, led by Shane Legg — DeepMind co-founder, now the lab’s Chief AGI Scientist — and by Marcus Hutter, his former PhD advisor, the inventor of the AIXI framework.

Its first section isn’t called an introduction.

It’s called Summary Instructions.

For an AI.

In practice: the authors ask whichever assistant summarizes the work to spell out their definitions, to resist compressing their lists, and to check — at the moment of reading, which might come months or years after publication — whether their conclusions held up over time. A way, for researchers who know their work will be skimmed or not read at all, to guarantee that at least one faithful reading will always exist, generated on demand.

It’s a fairly radical idea: a document that anticipates its own dilution, and actively tries to prevent it.

It would only half work.

A Measure They Invented Themselves

The definitions come first, laid out with almost clinical precision. AGI: a system that reaches median human-level performance on most cognitive tasks — not the ceiling, the average. ASI: a system that surpasses not a single expert but large, well-coordinated human-expert collectives, across virtually all tasks and domains. A system that’s merely superhuman in one isolated domain — AlphaFold at protein folding, AlphaGo at Go — doesn’t count. Too narrow.

Above that, a third marker, purely theoretical: Universal AI, formalized by the AIXI agent. A ceiling that’s mathematically defined but uncomputable, one that can only be approached with ever more compute. A cognitive speed of light.

What’s striking, reading these definitions back to back, is that they don’t come from nowhere. The measure that serves as the theoretical foundation for the entire hierarchy — AGI, ASI, Universal AI — carries the name of its inventors: Legg and Hutter. The same two who sign the report that uses it to assess their own lab’s trajectory.

DeepMind isn’t just mapping the road to superintelligence. It drew the ruler that measures the distance traveled.

Four pathways come next, not mutually exclusive, liable to unfold in parallel. Scaling compute, data, and models — which could, through sheer numbers alone, turn millions of human-level instances into a collective intelligence of a different order entirely: the report calculates that at a growth rate it calls conservative, a thousand initial instances would become a hundred million in five years. Algorithmic paradigm shifts, unpredictable by nature, arriving once the current paradigm hits its ceiling. Recursive self-improvement, where AI speeds up AI research, which produces better AI, which speeds up the research further still — a loop that could, in theory, run away hyperbolically, or simply fizzle out along the way. And multi-agent coordination and group agency: not one giant, singular mind, but thousands of agents that coordinate, specialize, dissolve, and reform — a digital organization rather than a brain.

Against them, six frictions, each framed as an open research question rather than a prediction: the data wall, resource constraints, the possibility that the current neural paradigm simply isn’t enough, research getting harder as the easy ground runs out, the abstraction barrier — the idea that an AI trained on human concepts excels at manipulating them but struggles to invent new ones — and deliberate slowdown, the one that policy could impose.

And then there’s what the report repeats almost like a refrain: even far beyond human intelligence, an ASI would be neither omniscient nor omnipotent. The speed of light bounds how fast information can travel. The Landauer principle sets a minimum energy cost for any computation. Complexity theory doesn’t yield to intelligence. A superintelligence will never play mathematically perfect chess — the computation required exceeds what the physical universe can hold — and nothing guarantees it will know how to cure aging, master fusion, upload a human brain, or restore the pre-industrial climate. These are empirical questions about the physical world, not problems that raw intelligence solves by decree.

A lab that refuses, in writing, to indulge magical thinking about its own creation.

What the Report Chooses Not to Weigh

Here’s where it gets interesting: what all that methodical caution doesn’t cover.

To move its technical analysis forward — four pathways, six frictions, an entire table of physical limits — the report rests on an assumption it doesn’t try to demonstrate: that AI safety and alignment will have been solved to a sufficient degree before the trajectory it describes becomes decisive. Not a result. A postulate, set down to bound the subject and let the rest of the reasoning stand.

Further into the text, it names, on its own, a risk that bears directly on that postulate: military-economic adaptationism. Competitive pressure between nations and companies systematically favors adopting technologies that increase power, regardless of the consequences for human welfare, and one actor’s unilateral restraint only pushes development toward more permissive jurisdictions. Even if DeepMind, or any other lab, chose to slow down for safety reasons, nothing guarantees the race would stop with it.

That’s exactly the dynamic that could invalidate the assumption made above. The report knows it. It names it.

It doesn’t weigh it against its own starting postulate.

One could object that this is precisely the honest thing to do. This document is probably one of the most transparent published on the subject: it states its assumption instead of burying it in a footnote, refuses to give a prediction date, tabulates physical limits instead of selling a promise, names the geopolitical risk instead of ignoring it. You can’t reasonably ask one report to both solve AI alignment and map four distinct technological pathways to superintelligence. The scope has to stop somewhere.

That’s true. The transparency here is real — and fairly rare for this kind of exercise.

But transparency doesn’t make the effect it produces disappear. The measured tone running through the rest of the document — no date, no triumphalism, entire tables devoted to what a superintelligence couldn’t do — leaves the reader with an impression of seriousness and rigor. That impression doesn’t stay confined to the pages where it was earned. A reader who comes away thinking “this document is measured, never overreaching” has every reason to extend the same quality to the one place the report chose not to be rigorous at all — the assumption that things will turn out fine on the control side. Naming the assumption doesn’t neutralize the halo effect it produces. It just makes that halo harder to hold against whoever wrote it.

The report never comes down on one side.

Except once.

And that one time it does is precisely the one that most needed to be argued out.

There’s something here bigger than the report itself. A document that devotes an entire section to giving fidelity instructions to the AI meant to summarize it — don’t compress the lists, check over time whether the conclusions held — ended up, in the coverage that followed its publication, summarized about like any other technical document: keeping the most spectacular number — a hundred million human-level instances in five years, hence superintelligence by accumulation — and letting the most important nuance slip away, the one about the alignment assumption and the risk it leaves unweighed.

The bottleneck was never an AI’s capacity to summarize faithfully. It was, downstream, the competition for attention.

Exactly the mechanism the report describes for AI development, replayed at the scale of its own reception.

The report asks the AI that summarizes it to check, over time, whether its conclusions held — whether scaling kept going, whether a paradigm shift occurred, whether recursive self-improvement took off or fizzled out. An update of sorts is built into the reading protocol itself.

It doesn’t ask anyone to check whether its starting assumption held.

That particular check — who’s actually in control, and to what degree alignment will, in fact, have been solved to a sufficient degree — remains entirely undone. By whom, and by when, the document doesn’t say.

That’s probably not an oversight.

Thanks for reading! If you want more clear-eyed analysis on technology, power, and the future of infrastructure, follow BloomTheDigitalLens.

(Originally published on BloomTheDigitalLens)