After AGI, What Comes Next?

For years, the biggest question in artificial intelligence has been deceptively simple: when will machines become as smart as humans? Researchers, governments, and technology companies have poured billions of dollars into chasing what is known as Artificial General Intelligence, or AGI; a machine capable of thinking, reasoning, and learning across a wide range of tasks the way a person does. But a landmark paper published in June 2026 by fourteen scientists at Google DeepMind suggests the world may be asking the wrong question entirely. The paper, titled From AGI to ASI, argues that human-level AI is not the finish line. It is the starting gun, and what follows may redefine not just technology, but the very foundations of human society, scientific knowledge, and global institutions.

To understand why this matters, it helps to be clear about definitions. In the DeepMind researchers’ framework, AGI does not mean a machine smarter than Einstein. It means a system capable of performing most cognitive tasks at roughly the level of an average human; reasoning through problems, learning new skills, using tools, planning ahead, and adapting to unfamiliar situations. What comes after AGI is something far more consequential: Artificial Superintelligence, or ASI. This is not simply a smarter version of AGI. An ASI would surpass the combined intellectual output of large groups of expert humans working together over extended periods of time. Think not of one brilliant scientist, but of tens of thousands of the world’s leading specialists collaborating for years and then imagine a machine exceeding what that entire group could produce. The fact that one of the world’s most respected AI laboratories is now writing peer-reviewed papers about this transition is itself a social signal of enormous weight. Superintelligence is no longer a fringe concept debated by philosophers and science fiction writers. It has entered the mainstream of legitimate scientific inquiry, and with it, an urgent conversation about what humanity must prepare for.

Perhaps the most striking idea in the paper is that superintelligence may not arrive in the form of one all-powerful machine. Instead, it could emerge from millions of human-level AI systems working together simultaneously. Digital agents can be copied in seconds, operate around the clock without fatigue, share knowledge instantly, and collaborate at speeds no human team can match. Even if each individual agent possesses only human-level intelligence, their collective output could dwarf anything produced by existing institutions, corporations, or governments. This vision carries profound academic implications. Universities, research centres, and think tanks that have long served as humanity’s primary engines of knowledge creation could find their role fundamentally altered. The peer review process, the slow accumulation of scientific consensus, the mentorship of doctoral students; all of these traditions assume that intelligence is scarce and time-consuming to produce. A world of networked AI agents challenges that assumption at its root.

The DeepMind paper maps out four distinct pathways through which AI could cross the threshold from human-level to superhuman capability. The first is continued scaling, driven by larger models, more data, and greater computing power. Some analysts suggest that effective AI computing capability has increased roughly tenfold per year over the past decade, a pace of progress without historical precedent. The second pathway involves fundamental architectural breakthroughs that move beyond today’s transformer-based systems toward entirely new models of computation, potentially inspired by neuroscience or built on principles not yet discovered. The third pathway is recursive self-improvement, where AI systems begin assisting in the design of better AI systems, creating a compounding feedback loop of advancement. The DeepMind team cautions that this would not happen overnight, since building hardware still requires factories, materials, and time, but a sustained acceleration driven by AI-assisted research could compress decades into years. The fourth pathway is the coordinated multi-agent ecosystem already described, where superintelligence emerges not from a single entity but from the organised collaboration of countless specialised systems working across every domain of human endeavour simultaneously.

Yet the paper is notable for its intellectual honesty about obstacles. One of the most pressing is what researchers call the data wall. Modern AI learns by consuming enormous quantities of human-generated content; books, websites, scientific papers, code, images, and videos. But that reservoir is not infinite, and as models grow larger, the supply of genuinely high-quality training material may not keep pace. Synthetic data generated by AI itself offers one possible solution, but it carries serious risks, including the amplification of errors and the entrenchment of misinformation at a scale that could distort public knowledge in ways that are difficult to reverse. This is not merely a technical problem. It is a social and epistemic one, touching on how societies establish truth, how journalism maintains credibility, and how academic disciplines protect the integrity of their fields.

Energy and infrastructure present another constraint with direct social consequences. Advanced AI systems are extraordinarily resource-intensive, consuming vast electricity and water while depending on hardware built from rare materials. The global expansion of AI data centres has already drawn scrutiny from energy regulators and raised legitimate questions about environmental justice, particularly in regions where local communities bear the burden of that infrastructure while receiving few of its benefits. The paper also identifies what it calls the abstraction barrier, a deeper philosophical challenge about whether AI systems can independently generate genuinely new conceptual frameworks or whether they will remain constrained by the intellectual structures inherited from human civilisation. History’s most transformative discoveries; evolution, relativity, quantum mechanics did not emerge from recombining prior knowledge. They came from entirely new ways of seeing. Whether machines can replicate that creative leap remains one of the most important unanswered questions in contemporary science, and its resolution will determine how much AI can truly contribute to human understanding rather than merely reflecting it back.

Beyond technical limitations, the paper situates AI development within a social and political landscape that will shape outcomes just as decisively as any algorithm. Governments worldwide are moving toward regulation driven by legitimate concerns about employment disruption, electoral manipulation, cybersecurity vulnerabilities, and the concentration of technological power in the hands of a small number of private companies. These are not abstract worries. They are already being felt by workers in industries from logistics to legal services, by journalists confronting AI-generated misinformation, and by citizens in countries where AI-driven surveillance is expanding without adequate legal oversight. The DeepMind researchers acknowledge that international agreements, licensing frameworks, and public accountability mechanisms will be essential, and that public trust; earned through transparency and genuine engagement ,will prove as important as any technical milestone.

Stepping back, the most significant idea the paper introduces is that intelligence itself is becoming industrialised. For all of human history, creating a capable mind required birth, education, experience, and decades of development. There was no shortcut. Now, for the first time, that constraint is weakening. Digital intelligence can be copied instantaneously, specialised for particular tasks, and deployed at virtually unlimited scale. Once AGI becomes a reality, the central question will shift from whether machines can think to how many thinking machines can exist, how rapidly they can operate, and what governance structures are in place to ensure their outputs serve collective human interests rather than narrow private ones. Academic institutions face a particular reckoning here. The entire architecture of higher education; degrees that signal expertise, credentials that establish authority, disciplines that organise knowledge; was built on the assumption that human intelligence is finite and unevenly distributed. A world in which cognitive capability becomes abundant and scalable will require entirely new frameworks for understanding what education is for, what expertise means, and how knowledge is validated and shared across society.

The DeepMind researchers do not promise utopia, nor do they predict catastrophe. They emphasise instead what serious scientists always emphasise: uncertainty, and the importance of preparing rigorously for multiple possible futures. A superintelligent system would still be bound by the laws of physics, the costs of energy, and the unpredictability of complex systems. Intelligence does not equal omnipotence. But it does equal, potentially, an unprecedented acceleration of discovery; in medicine, climate science, materials engineering, and fields not yet imagined; alongside an equally unprecedented capacity for disruption if the transition is managed poorly.

AGI has not yet arrived. But for the first time, the world’s leading AI scientists are devoting serious effort not to reaching it, but to planning for what comes after. That shift in focus is itself a measure of how rapidly the horizon is approaching. The decisions made now; in laboratories, in legislatures, in university faculties, and in public discourse; will determine whether the age of superintelligence becomes the most creative chapter in human history or its most destabilising one. What is certain is that the conversation can no longer afford to wait.