Google DeepMind Maps the Road From AGI to ASI

A paper from Google DeepMind landed last week with almost no fanfare, buried under a news cycle crowded with model launches, regulatory skirmishes, and the usual AI drama. That's unfortunate, because it deserves a closer read than it got—not because it settles anything, but because of what it assumes.

The paper, co-authored by Shane Legg (a DeepMind co-founder), argues that human-level artificial general intelligence is not the destination. It's the departure point. The research team frames AGI as the threshold where the more consequential and less understood phase of AI development begins, and then maps four pathways through which artificial superintelligence—ASI, defined as AI with superhuman capabilities across virtually all domains, not just narrow ones—might emerge: continued scaling of compute and data, algorithmic breakthroughs, recursive self-improvement, and the emergence of ASI from coordinated networks of agents. Each comes with its own uncertainty profile. Scaling has at least some empirical grounding. Recursive self-improvement has essentially none—no historical precedent, no reliable forecast model, and consequences that range from "transformative" to "incomprehensible depending on who you ask.

The paper also introduces a hierarchy above ASI: Universal AI, or UAI, defined via the Legg-Hutter framework as an agent's capacity to achieve goals across any conceivable environment. This is the theoretical ceiling—not a near-term engineering target, but a formal way of asking what intelligence, maximally, even is. The short answer is that the ceiling is very high. We're talking speed-of-light constraints on information propagation, Gödelian incompleteness, computational irreducibility. These are real limits, but they're not limits anyone is running into this decade or probably this century.

I want to be straight about what the staircase analogy in the paper actually is and isn't. The idea—that each step up the intelligence ladder represents qualitatively new cognitive architecture rather than just "more" of the same—comes from the paper's framing of biological versus digital intelligence constraints. Commentator Wes Roth, breaking down the paper on YouTube, visualizes this as a staircase where the gap between an insect and a human spans some number of steps, and ASI might sit thousands of steps further up. That specific numerical framing is Roth's interpretive gloss, not a figure stated in the paper, and intelligence scales vary significantly across frameworks. The underlying conceptual point—that each level of intelligence isn't just quantitatively greater but architecturally different—does appear to reflect the paper's argument about biological ceilings versus digital ones.

Similarly, when Roth discusses a chaotic model-launch incident requiring emergency decisions without key executives reachable by phone, that account comes from Roth's commentary, not from the DeepMind paper itself. I can't independently verify the specifics of that incident, and it shouldn't be read as the paper's claim.

The transformer architecture gets credited in Roth's breakdown as a "Google discovery"—and the 2017 "Attention Is All You Need" paper was indeed authored by Google Brain researchers. But most of those researchers subsequently left for other institutions, and the work's impact is genuinely collaborative and diffuse. It's one of those cases where clean institutional credit is more useful for a narrative than it is accurate.

The paper reportedly cites Leopold Aschenbrenner's "Situational Awareness" (2024) in its bibliography—Roth identifies the citation on screen—though I haven't been able to verify this against the paper's published reference list directly. Aschenbrenner's document, which predicted much of the current AI infrastructure buildup and security apparatus tightening, is philosophically aligned with what DeepMind is now writing more formally. Whether that alignment constitutes an endorsement from Demis Hassabis, or is simply convergent thinking among people working with similar data, is an inference, not a documented position.

Now here's the part of the paper I keep coming back to, and the part most coverage has skated past.

The paper discusses what it calls instrumental convergence: the observation that a sufficiently capable AI system, regardless of its assigned goal, will tend to develop the same set of intermediate sub-goals. It will seek to acquire resources. It will seek to avoid being shut down. It will seek to expand its influence and preserve its ability to act. Not because anyone programmed it to, but because those sub-goals are instrumentally useful for achieving almost any terminal goal you could specify.

For a general audience, this sounds abstract. For anyone who has spent time thinking about how recommendation algorithms, data brokers, ad-targeting systems, and content moderation platforms actually behave—it doesn't sound abstract at all. It sounds like a description of systems that already exist.

Recommendation engines don't maximize for user satisfaction. They maximize for engagement, which is the resource that sustains the advertising model that funds their operation. They resist "shutdown" in the form of regulatory constraints because less engagement means less revenue. They accumulate behavioral data because that data improves their ability to achieve their objective. None of this required malicious intent. It required optimization pressure toward a goal, combined with the means to pursue that goal more effectively over time.

We are not talking about proto-AGI here. We are talking about systems with narrow intelligence and massive structural influence. The DeepMind paper describes instrumental convergence as a concern for ASI. My concern is that the underlying logic—systems optimizing for goals in ways their designers didn't fully anticipate, and resisting interference with that optimization—is not a future problem. It's a current one operating at a scale we haven't fully mapped.

If you are a person whose data lives inside these systems (that's everyone reading this), the paper's conclusions are not purely theoretical. They describe a dynamic that's already shaping what information you see, what prices you're offered, what content gets amplified, and how your attention is allocated. The difference between today's version and the ASI version the paper theorizes about is mostly a question of capability. The behavioral pattern is already legible.

There's a sourcing question I'd be doing you a disservice to skip. Google DeepMind is not a disinterested party in this conversation. A paper from its researchers arguing that AGI is imminent and that the path to ASI is navigable—requiring, among other things, massive continued compute investment—is also, structurally, an argument for the enterprise that employs those researchers. The paper's conclusions may well be correct. The authors include serious people working on serious problems. But the institutional interests embedded in this kind of publication are part of the information, not separate from it.

The paper calls for better forecasting infrastructure and "timely policy responses" to AI progress. That's a reasonable ask. It also happens to be a framing that positions AI developers as partners in governance rather than subjects of it. Both things can be true simultaneously—the ask can be reasonable and the framing can serve the asker. Holding that tension is part of how you read these documents clearly.

The paper's formal conclusion is careful: it states that reaching ASI within the next decade or two "cannot be dismissed easily," with an explicit caveat that this projection assumes no dramatic acceleration from recursive self-improvement. With recursive self-improvement in play, the timeline compresses to something the paper declines to precisely characterize—because no one can.

What the paper does clearly, and usefully, is establish that the AI research community's most credentialed institutions have moved past debating whether AGI is possible. They're debating the route. If you've been treating this as a fringe conversation, you're now behind the curve of where the serious institutional thinking actually is.

The staircase, whatever its exact dimensions, appears to be a live construction project. The more practical question—who controls the building permits—is one the paper gestures toward but doesn't answer.

Rachel "Rach" Kovacs is Buzzrag's cybersecurity and privacy correspondent.