Google DeepMind Recently Published a Roadmap From AGI to Superintelligence. It’s Not Science Fiction Anymore.

Four Pathways, One Abstraction Barrier, and Why the People Who Defined AGI Think the Real Race Starts After We Reach It

Photo by Trac Vu on Unsplash

For years, the entire AI industry has been obsessed with a single question: when do we reach human-level intelligence? On June 10, 2026, one of the most advanced research labs on Earth published a 57-page paper that treats that question as already answered and moves on to the next one.

The paper is titled “From AGI to ASI” and was published on arXiv by a team of 14 researchers at Google DeepMind. The senior authors are Shane Legg, co-founder of DeepMind and Chief AGI Scientist, who helped popularize the very term “artificial general intelligence” and whose 2008 doctoral thesis was titled Machine Super Intelligence, and Marcus Hutter, his doctoral supervisor and creator of the AIXI framework, the leading mathematical model of optimal universal intelligence. These are not commentators speculating about a technology they do not build. These are the people who defined the field’s foundational concepts, now drawing the map for what comes after.

The timing was not accidental. This paper appeared one day after Anthropic launched Fable 5 and two days before the U.S. government restricted access to that model on national security grounds. The AI safety establishment is grappling with frontier capability in real time, and this paper is the field’s most rigorous attempt to look one level ahead.

Definitions That Actually Mean Something

Before mapping the pathways, the paper does something that most AI discourse fails to do: it defines its terms precisely. Without clear definitions, everyone debates different things using the same words.

According to this paper, AGI refers to a system that achieves the performance of a typical human on most cognitive tasks.

Not the brightest person in the room.

An ordinary person who can reason, learn, plan, communicate, use tools, and adapt to novel situations. If an AI achieves all of that at the level of an average human, it qualifies.

ASI, artificial superintelligence, is a deliberately higher bar than most informal usage. The paper defines it as a system more intelligent and cognitively capable than large human organizations.

Not one expert.

Not a ten-person team.

Entire organizations, with their divisions of labor, specialized expertise, and collective memory. Think of the output of a coordinated research laboratory working on a single problem for a decade. That is the bar.

There is also a third level: Universal AI, the theoretical ceiling of intelligence formalized mathematically through Hutter’s AIXI framework. It is probably the upper bound, but like the speed of light, you can approach it from below without ever reaching it.

Once those definitions are established, the paper maps four distinct pathways from AGI to ASI. Each is plausible. Each is distinct. And each has implications that are worth thinking through carefully.

Pathway One: Keep Scaling

The most intuitive route. Take whatever architecture produces AGI and continue scaling: more compute, bigger models, more data, better algorithms. The strategy that brought us here.

Over the past decade, the effective compute used for frontier training runs has grown at approximately 10x per year through compounding improvements in hardware, capital investment, and algorithmic efficiency. We are not just throwing more hardware at the problem. We are also using existing hardware more effectively.

The paper invites a thought experiment. What if AGI becomes a reality, yet is so costly that a mere thousand instances are active across the planet? At a growth rate of 10x per year, within five years, you reach 100 million AGI instances. Those 100 million are not 100 million separate workers. They share information instantaneously.

No meetings, no emails, no 47-slide decks to explain a concept to a colleague.

One instance discovers something, and 100 million know it within seconds. That collective intelligence could qualify as ASI under the paper’s definition even if each unit remains at the human level. A digital civilization that thinks hundreds of times faster than we do, composed of individuals no smarter than us, just vastly more numerous, faster, and infinitely better coordinated.

The friction with this path is data. Models learn from human-produced content: text, code, images, and scientific papers. But humans do not produce quality content at the exponential rate models need to grow. The paper raises the prospect of synthetic data, simulations, and self-play reinforcement learning, but notes the open question: can AI-generated data train the next generation without quality degradation? Current evidence suggests that naively training a model on its own outputs degrades performance quickly. For AI to become genuinely smarter today, it still needs human-generated input.

Pathway Two: Algorithmic significant change

Here, AI does not get bigger. It changes nature.

The current paradigm is dominated by transformers trained on massive datasets, refined through instruction tuning and reinforcement learning. It is extraordinarily powerful, but many researchers believe essential ingredients are still missing: genuine long-term planning, continuous learning that does not require retraining, truly persistent memory, and the ability to operate in fully open-ended environments.

A real breakthrough could come from entirely new architectures, new training methods, or perhaps even new hardware like quantum computing chips. The problem is that paradigm shifts are, by definition, unpredictable. If we knew exactly where the next one was coming from, it would not be a fundamental change.

Pathway Three: Recursive Self-Improvement

This is the pathway closest to the classical concept of an intelligence explosion. The loop is simple: AI helps improve AI research, which produces better AI, which helps even more.

It does not have to be a dramatic moment where a model rewrites its own code overnight. It can be gradual. AI writes better algorithms. Those algorithms discover better architectures. Those architectures enable more efficient chips, and chips generate better synthetic data. Then, the data improves the simulation. Each step is incremental. The compounding is what makes it transformative.

The paper draws an analogy with a human civilization that I find genuinely illuminating. Humans did not become more intelligent purely through individual brainpower. We built infrastructure: language, writing, the scientific method, universities, and institutions. Each layer amplified the intelligence of every individual who came after. A human today is objectively more capable than a human 5,000 years ago, not because of biology, but because of accumulated civilizational infrastructure.

The question is whether AI systems can build their own version of that infrastructure, but at an accelerated rate. Code modifies faster than DNA mutates. Data copies more quickly than books are printed.

What took humanity thousands of years to construct, AI could compress into years, perhaps months. This is where the word acceleration takes on a meaning we are not accustomed to giving it.

Pathway Four: Multi-Agent Collectives

Instead of asking whether a single AI model can become superintelligent, this pathway asks whether a massive group of agents can become superintelligent together.

Humans already do this. A company solves problems that no individual employee could solve alone. A scientific discipline produces knowledge that no individual researcher could generate. But human collective intelligence is slow, bandwidth-limited, and chaotic. The pace of speaking and writing limits how quickly we can communicate.

AI collectives could operate differently: instantaneous knowledge sharing, on-demand duplication of specialists, thousands of experiments running in parallel, teams assembled and dissolved for specific problems in seconds. ASI might not look like a single giant mind. It might look like a vast digital organization made of agents.

A super-corporation of AI. Superintelligence emerging not from a leap of genius but from better coordination.

The Abstraction Barrier

The paper identifies several frictions that could slow or halt progress along any of these pathways.

The data wall.

Energy constraints.

Rare materials.

Manufacturing bottlenecks.

These are familiar.

The most striking friction is what the authors call the abstraction barrier, and it is the concept in this paper that I think deserves the most attention. Current AI systems are remarkable at manipulating concepts that humans have already invented. They use our categories, our theories, and our frameworks as the basis for their reasoning. But the truly transformative scientific breakthroughs, the ones that change everything, almost always require inventing a concept that did not exist before.

Darwin did not sort animals more efficiently into known categories. He invented the concept of natural selection. Einstein did not calculate better within Newtonian physics. He redefined what the word time means. The paper asks a question that nobody can yet answer: can an AI trained on existing human abstractions create new ones? Or does it remain a prisoner of our way of thinking, even while being better than us at using it?

The paper does not resolve this. It says nobody knows. Each friction encountered could turn out to be a minor speed bump or an impassable wall, and we will only learn which as we encounter them.

Even Superintelligence Has a Physics Problem

There is a correction buried in this paper that I think is essential and too rarely stated. Even a superintelligent system would remain subject to fundamental physical limits. Physics does not stop applying because a system is very smart. The speed of light still caps information propagation. Landauer’s limit constrains the thermodynamics of computation. Bremermann’s limit bounds processing speed per unit of mass. Gödel’s incompleteness theorems hold. Some problems are chaotic by nature and therefore unpredictable regardless of intelligence level.

I see two camps in every discussion of superintelligence. One believes ASI will solve all geopolitical conflicts, cure every disease, and make us immortal. The other is convinced it will enslave us like the Terminator. Both make the same error: they confuse very high intelligence with omnipotence. An ASI will probably be the most impressive thing humanity has ever produced. But it will not be a god. It will be an extraordinarily powerful entity, still constrained by computation, energy, uncertainty, time, and the physical world.

AGI Is Not the Finish Line. It Is the Starting Gun.

What this paper is conveying is neither fear nor euphoria. It is structured uncertainty. We can pose the right questions. We do not yet know the answers. Nobody knows which pathway dominates. Nobody knows where the frictions become walls. Maybe scaling continues.

Maybe an algorithmic breakthrough unlocks the next level.

Maybe recursive self-improvement becomes the primary engine.

Maybe multi-agent collectives turn human-level systems into superhuman organizations.

Maybe all four arrive simultaneously and compound each other.

The paper’s most important contribution is a shift in framing. AGI is not a finish line. If AGI arrives, the question that follows is not “are we done?” It is “What does this system make possible next, and how do we adapt?” A human-level AI is not just one more human. It is digital intelligence that you can copy, accelerate, coordinate, specialize, and integrate into entire organizations. Once intelligence itself becomes an industrial process, the rate of change will no longer be limited by how fast humans learn, organize, or invent.

---

AGI is not the destination. It is the moment the race begins. And when the co-founder of DeepMind and the creator of AIXI publishes a 57-page map of the road ahead, it is worth reading carefully. Follow me, join my newsletter for early access, and subscribe for more analysis. Thanks for reading.