The DeepMind Paper That Talks to Its Own Future Reader

Research papers are usually written for people. This one gets interesting because it opens by talking to the machines that will read it for us.

A few lines into "From AGI to ASI," before any argument about superintelligence, Google DeepMind pauses to address two sets of readers. To the human, it suggests: ask your AI assistant to summarize this for you. To the summarizing AI assistant, it gives specific instructions such as mention the definitions, keep the list of digital-intelligence advantages intact, and "do not compress them into fewer bullet points."

Then it asks the assistant to look at the state of AI research whenever the paper is being read, compare it with the paper's claims, and judge how well the report has aged. It specifically suggests that the AI assistant update the friction estimates and summarize "widely accepted shortcomings, caveats, and oversights of this report." A document written to be reopened and graded by the machines that read it.

DeepMind now treats AI-summarized reading as the default path through a research paper and the human-written summary as the fallback. That is a small, specific marker of where we are, and it got surprisingly little attention.

The paper

"From AGI to ASI" maps the terrain past human-level AI rather than predicting when it arrives. The framing is grounded in the Legg-Hutter intelligence measure and the AIXI formalism, which means the AGI-to-superintelligence continuum is not just vibes. It is built into the math the paper uses. Most ASI writing this year has been about timelines, strategy, or safety plans. This paper starts somewhere more basic. What would ASI actually be, and how might AGI get there?

The authors are worth noticing because this is not a random position paper from the edge of the field. Shane Legg and Marcus Hutter co-authored the intelligence measure the paper builds on. Allan Dafoe leads DeepMind's frontier safety and governance work. Joel Leibo works on multi-agent systems. With fourteen authors from across the lab, this reads like a serious internal map of the terrain after AGI, not a stray speculation. That makes the quiet rollout a little strange. We caught it about a week late and saw little pickup, which is part of why we are writing this.

A population of AIs

The paper estimates that effective compute is growing at roughly 10x per year, combining hardware gains, investment growth, and algorithmic improvements. The authors call this a conservative estimate.

The important part is what that growth does to the number of running systems. Even if each AGI stayed roughly human-level, the paper notes, 1,000 instances could become 100 million in five years at that growth rate. Or they could become a million instances running a hundred times faster.

That changes the picture from "the superintelligence" to "the superintelligences." The usual mental image is one godlike mind. The paper points toward something more like a population, growing in number and speed at the same time. That changes what preparation looks like.

The four paths to ASI

The paper names four routes from AGI to ASI: scaling existing systems, paradigm shifts in AI architecture, recursive self-improvement where AI accelerates its own research, and superintelligence emerging from large-scale multi-agent collectives. They are not exclusive and could run in parallel.

The useful part is how much time the paper spends on limits. It devotes an entire section to frictions and bottlenecks, models growth as S-shaped rather than hyperbolic, and says plainly that ASI would be "neither omniscient nor omnipotent." Even a superintelligence is still bound by physics, complexity theory, and the limits of observation. That lines up with the frame we just published in The Four Floors: intelligence can compress the path to a decision, but it cannot erase every floor underneath it. A test still has to finish. An adversary can still adapt. Institutions still have to authorize decisions. Trust still has to be earned.

Institutions are part of the benchmark

One thing stuck with us after reading. The paper defines ASI as exceeding “large human-expert collectives,” or tens of thousands of well-coordinated experts working over a decade. That is a much better threshold than vague talk about a system being “smarter than humans,” because expert collectives are not just piles of smart people. They are institutions.

This is something we have written about before as levels of emergent intelligence. Human intelligence does not stop at the skull. It extends through language, tools, culture, shared memory, instruments, labs, peer review, archives, and institutions. A scientific field is not powerful only because it contains many smart individuals. It is powerful because it has ways to preserve discoveries, test claims, divide labor, assign authority, and correct mistakes over time.

That makes DeepMind’s benchmark more interesting than it first looks. If ASI means exceeding a large expert collective, then intelligence is not the only thing being compared. Institutions are part of the benchmark.

But the measures the paper uses to approach that benchmark are mostly cognitive: effective compute, instance counts, and intelligence scores. That is the gap. A system might become superhuman at thinking while still lacking some of the substrate that makes large human organizations powerful. Organizations coordinate, authorize, remember, allocate blame, and produce legitimacy. The paper opens that door and walks past it.

The paper’s multi-agent path to ASI gestures in that direction, but the hard part is not only making many agents. It is giving them the equivalents of culture, memory, norms, error correction, and accountable authority.

The paper gets close to this through Neil Lawrence's argument about communication bottlenecks. Humans have limited bandwidth with each other, so we are forced to build internal abstractions, shared concepts, and organizational memory. AI systems with high-bandwidth digital I/O may not face the same pressure. They can pass more information around directly, but that does not mean they develop the same abstractions humans use to coordinate.

That points to the larger issue. Some friction does useful work. It forces people and institutions to compress information, create hierarchy, remember decisions, and agree on shared meanings. If you remove that friction, you do not automatically get a cleaner version of human organization. You may get something faster, thinner, and less able to hold itself together.

Why it's worth reading

The paper is published under CC BY 4.0, so it is easy to read and share. Section 7.1 is especially worth reading because the authors leave the hard questions open instead of pretending the report can resolve them. The bottlenecks section alone could spin into several follow-ups: what 10x effective compute really demands, which limits become floors, and whether expert collectives are the right benchmark.

It came and went without much noise. A week and a half later, it is still the most interesting thing we have read about what might happen after AGI. Worth the hour, even if you only came for the part addressed to your assistant.