Most public debate about artificial intelligence has centered on one question: when will artificial general intelligence (AGI) arrive? A paper from Google DeepMind, published on June 12, asks a different one. If AGI is built, what happens next? Is human-level capability a destination, or the first stage of a transition to something far beyond it?
The paper, titled “From AGI to ASI,” comes from 14 DeepMind researchers, among them Shane Legg, the company’s co-founder and chief AGI scientist, and Marcus Hutter, a longtime theorist of machine intelligence. ASI stands for artificial superintelligence. The authors set out to describe the plausible routes from one to the other and the obstacles that could slow or block the journey. They are explicit that the result is a map of possibilities and open questions, not a forecast.
Defining the terms
The paper’s definition of AGI is more modest than the popular image. An AGI, in its framing, is a system that performs at roughly the level of an average adult across most cognitive tasks: it can understand varied problems, learn new things, plan, communicate, use tools and adapt to unfamiliar situations. It need not match the world’s best experts, though it may already exceed humans in particular domains.
ASI is set much higher. Beating an individual doctor, lawyer or programmer is not enough. The authors reserve the term for a system, or collection of systems, that outperforms a coordinated group of thousands of the best human experts working on a problem for years, across nearly every important field. The relevant comparison is not a person but a scientific community or a global institution.
The paper also references a third, purely theoretical construct: AIXI, Hutter’s mathematical model of an optimal learning agent. AIXI is not buildable, because the computation it requires cannot be carried out, but it serves as a theoretical ceiling against which real systems can be compared.
One unusual feature of the paper is that its opening addresses not only human readers but the AI systems likely to summarize it for them, asking that definitions be kept precise, that key lists not be over-compressed, and that future readers check how well its arguments held up. The authors are assuming that much of their audience will encounter the work through an AI intermediary.
Route one: scaling
The first path is the one that has driven most progress over the past decade: larger models, more data, more compute and better hardware, combined with steady gains in algorithmic efficiency.
The paper offers a thought experiment. Suppose the first AGI is expensive enough that only about 1,000 copies can run worldwide. If available compute grows tenfold a year, that becomes 10,000 copies after one year and on the order of 100 million within five. Those copies would not be 100 million isolated workers. They could share what they learn instantly, duplicate themselves and communicate far faster than any human organization, in which knowledge spreads through meetings, reports and training. A population of individually human-level systems, coordinated in that way, could collectively reach superintelligent capability.
The main obstacle on this route is data. Today’s models learn from text, code, images, video and research produced by humans, and the supply of high-quality material is finite. Synthetic data, generated by AI systems themselves through self-play, simulation, search and reinforcement learning, is the proposed workaround. The risk is that models trained repeatedly on their own flawed output compound errors rather than correcting them. Whether synthetic data can be kept good enough to substitute for human data is, the authors say, an open question.
Route two: a new paradigm
The second path is a fundamental change in how AI works, not just how big it is. Current systems rest on transformer architectures, large datasets, instruction tuning and reinforcement learning. That recipe has been enormously successful, but the paper notes weaknesses in long-horizon planning, continual learning, persistent memory, world modeling and reliable operation in open environments.
A paradigm shift could come from a new architecture, training method, memory system, reasoning approach or hardware substrate such as neuromorphic or analog computing. By definition, such a breakthrough is hard to predict. If researchers knew what it was, they would probably have built it. That means every forecast based on current scaling trends could be overturned by a single discovery.
Route three: recursive self-improvement
The third path is AI accelerating AI research. The paper does not envision a single moment in which a model rewrites itself and becomes a thousand times smarter. It describes a gradual, distributed process in which AI systems help design algorithms and architectures, improve chips and manufacturing, clean and generate data and build simulations, improving the entire ecosystem rather than one model.
The key difference from human research is reproduction time. Training a human scientist takes decades; a trained digital researcher can be copied in moments and specialized for a new field quickly. Even so, the authors caution against treating this as magic. New chips still have to be fabricated, new drugs still have to be tested in laboratories, and new data centers still need land, power, water and cooling. AI can speed up research, but it cannot abolish physical constraints.
Route four: multi-agent systems
The fourth path asks not how smart one model can become but how much intelligence many agents can produce together. Human organizations already display collective intelligence that no individual possesses, but they are slowed by communication overhead, information silos and bureaucracy. AI agents could share information instantly, spin up copies for specific tasks, run thousands of experiments in parallel and form and dissolve teams as needed.
In this scenario, superintelligence would not look like one vast mind. It would look like a very large digital organization: a swarm, a research ecosystem or a company made of millions of agents.
The four routes are not exclusive. The authors suggest they could reinforce one another: better models produce better agents, better agents do better research, better research yields better hardware and better hardware enables more scaling.
What could stop it
The paper devotes equal attention to barriers. Beyond the data wall, it lists resource constraints (electricity, chips, water, raw materials and the factories and data centers to use them), the possibility that the current neural network paradigm simply cannot reach AGI or ASI no matter how far it is scaled, and the tendency of research to get harder over time as easy discoveries are exhausted.
Some problems also require real-world time. If understanding a biological process demands months of physical experimentation, more thinking cannot fully compress that period, however many AI researchers are working on it.
A subtler obstacle the authors call the abstraction barrier. Current systems learn from human-made words, concepts and categories, and they recombine those concepts quickly. But major scientific advances, such as relativity, quantum mechanics, germ theory, evolution and computation, came from inventing new concepts rather than rearranging old ones. Whether an AI given all pre-modern-physics knowledge could discover quantum mechanics on its own remains unknown.
Finally, there is politics. Serious accidents, misuse, labor disruption, biological risks or instability could prompt governments to impose licensing, safety testing, compute limits, international agreements or outright bans. The trajectory to ASI, the paper argues, will be shaped by law, public trust and social choice as much as by technology.
Limits even for a superintelligence
The authors are also clear that superintelligence would not mean unlimited power. An ASI would still obey physics. Information cannot travel faster than light, computation takes energy and time, some problems are intractable and some systems are chaotic enough that no intelligence can predict them. ASI should not be imagined as an entity that cures every disease instantly or reshapes reality at will, but as something far more capable than humans that still operates within constraints of time, energy, materials, information and uncertainty.
The central message
The paper’s conclusion is uncertainty rather than confidence. No one knows how long scaling will keep working, whether synthetic data will suffice, whether self-improvement will be fast or bottlenecked, or whether a new paradigm is around the corner. Progress could be explosive, or it could be a long, uneven process moving at different speeds in different fields.
What the authors do argue is that the most consequential shift begins when intelligence becomes something that can be manufactured: copied, accelerated, specialized and networked rather than raised and educated over decades. Once AGI exists, the important questions are how many copies can run, how fast, how they share knowledge and whether they can help build their successors. On that view, AGI is less a finish line than the moment the real race starts, and whether existing institutions, laws, economies and safety systems are ready for it is the question the paper leaves with its readers.
