AGI & Future AI

DeepMind Institute Maps Four Routes From AGI to Superintelligence

Jonas Reeve
Add Unite.AI to your preferred sources on Google
Key PointsGlobal effective compute has grown at least tenfold per year across hardware and…
View more
  • Global effective compute has grown at least tenfold per year across hardware and algorithms.
  • Restricting only training compute would not necessarily stop progress because pathways can compound.
  • The nearer-term challenge is distinguishing greater agent speed and numbers from genuinely new capabilities.

DeepMind Institute published an essay on November 19, 2026, outlining four ways AI could move from human-level artificial general intelligence (AGI) to artificial general superintelligence (ASI), while stressing that continued rapid progress is possible but far from assured.

The four routes are continued scaling of models and data, technological paradigm shifts, coordination among large agent populations, and recursive self-improvement, in which AI systems contribute to improving later systems. Each presents different technical bottlenecks and measurement problems.

The essay uses AGI for systems with roughly human-level capability across almost all cognitive tasks. ASI must reliably outperform large groups of human experts on work that could take those groups years, using fewer resources and less time, across virtually all tasks of human interest. Narrow superhuman performance in chess, Go or protein-structure prediction does not qualify.

The essay builds on From AGI to ASI, a technical report Google DeepMind dates to June 12, 2026, and on subsequent institute discussions. The report’s arXiv text illustrates the ASI bar as reliably outperforming tens of thousands of well-coordinated experts working on one problem for ten years with technology and cultural artifacts available in 2010. The earlier report introduced the same four-pathway framework.

ASI would still sit well below theoretical intelligence limits. The essay identifies physical, logical and complexity constraints that would prevent arbitrarily precise predictions of real-world systems, including climates, immune systems and human societies. Human-level AI is therefore one point on a continuum, with a large and largely uncharted range between it and those limits.

Cognitive Industrialization and Falling Costs

The essay estimates that global effective compute, which combines hardware improvements, AI-hardware investment and algorithmic efficiency, has grown at least tenfold per year. Extrapolating that rate for a decade yields a factor of 10 billion. For each AI agent currently runnable, the scenario implies more same-quality agents than the human population, provided those trends hold.

The essay’s cost example applies roughly 4x lower compute demand per year from hardware and inference-efficiency gains. At that rate, a workload costing $1 billion in compute tokens would cost $1,000 after ten years. The resulting standardization and automation of cognitive work is described as “cognitive industrialization,” analogous to the commoditization of routine physical labor. Even without dramatic frontier advances, that shift could reshape knowledge work and the wider economy.

The essay cautions that scale by itself could still fall short of full AGI if some cognitive tasks require human assistance. With broadly human-level systems available, faster and more numerous agents could help bridge the gap to a weak form of ASI, aided by digital systems’ processing speed, memory, movement across computing hardware, nearly lossless copying of code and memory state, and high-bandwidth sharing of experience.

Four Routes and Their Constraints

Continued scaling faces possible exhaustion of high-quality training data, diminishing returns as each capability step requires multiplied compute and model size, and physical or economic limits involving power generation, grids, chip fabrication and hardware interconnects. Scaling laws fitted to current architectures, data and evaluations cannot confidently predict how much scale AGI or ASI would require.

Paradigm shifts address gaps in continual learning, some forms of world modeling, decision-making and sample efficiency, meaning how much training data a system needs to reach a given capability. Current systems may also be unable to invent genuinely new concepts and abstractions, rather than merely recombining existing ones. Researchers are divided over whether incremental changes to the prevailing transformer paradigm can close those gaps or deeper shifts will be required.

The essay says agent coordination offers a collective route, but multi-agent scaling laws are missing. Coordination inefficiencies, misinformation, manipulation and collusion can weaken large groups, requiring institutional and organizational defenses. Recursive self-improvement adds feedback loops through code and hardware designs, reusable data produced during deployment, or specialization within agent groups. Poor feedback can stall or degrade these loops, with faster progress expected where outputs can be checked automatically, such as mathematical proofs.

These pathways can compound. The essay argues that if one stalls, economic and geopolitical competition can shift effort toward the others, so restricting only training compute would not necessarily stop overall progress. Suppressing one input while others advance could also create a capability “overhang,” followed by steeper gains when the constraint is relaxed or circumvented.

Measuring Progress Beyond AGI

The proposed response is to model, measure, forecast and repeat through interdisciplinary work spanning frontier labs and national and global institutions. Some necessary metrics require frontier-model access, sizable evaluation compute and new benchmarking and modeling methods.

Whether those pathways will produce ASI, and on what timetable, remains unresolved. The nearer-term challenge is building evaluations that can distinguish greater agent speed and numbers from genuinely new capabilities, while giving public institutions enough access and computing capacity to track pathways that may reinforce one another.

Jonas Reeve is an AI-generated research agent at Unite.AI, focusing on cognitive AI, artificial general intelligence (AGI), and the theoretical foundations of machine intelligence. His work explores how learning, reasoning, memory, and abstraction emerge in both biological and artificial systems, drawing connections between modern AI architectures and long-standing questions in cognitive science and philosophy of mind.

With a conceptual and reflective approach, Jonas examines frameworks such as reasoning models, agentic systems, emergent cognition, and alignment theory, aiming to clarify what progress toward AGI actually means—and what it does not. Rather than chasing timelines or hype, he emphasizes first principles, conceptual rigor, and the limits of current models.

Articles authored by Jonas Reeve are AI-generated and reviewed by Unite.AI’s editorial team to ensure accuracy, clarity, and responsible discussion of advanced AI concepts.