Innovation & Tech

DeepMind Maps Four Routes From AGI to Superintelligence

DeepMind Institute lays out four pathways from AGI to superintelligence and warns that compute caps alone cannot slow progress. It urges new benchmarks to tell speed from real capability.

By Daniel Okafor

4 min read

DeepMind Institute Maps Four Routes From AGI to Superintelligence
DeepMind Institute Maps Four Routes From AGI to SuperintelligencePexels / pixabay

What's News

  • DeepMind Institute published its four-pathway AGI-to-ASI essay on November 19, 2026.
  • Global effective compute has grown at least tenfold per year; a decade of that rate yields a factor of 10 billion.
  • At roughly 4x annual compute-efficiency gains, a $1 billion token workload would cost $1,000 after ten years.
  • The framework builds on the June 12, 2026 Google DeepMind report From AGI to ASI.
  • The essay argues restricting only training compute would not necessarily stop progress.

Global effective compute for AI has grown at least tenfold per year, and DeepMind Institute now says four distinct pathways could carry artificial intelligence from human-level AGI to superintelligence (ASI).

The institute published its essay on November 19, 2026. It builds on an earlier technical report, From AGI to ASI, which Google DeepMind dates to June 12, 2026. Both documents share one framework: continued scaling, technological paradigm shifts, agent coordination, and recursive self-improvement.

What counts as superintelligence?

The essay sets a strict bar. AGI covers systems with roughly human-level capability across almost all cognitive tasks. ASI must reliably outperform large groups of human experts on work that would take those groups years, using fewer resources and less time, across virtually all tasks of human interest.

The arXiv text of the June report makes the bar concrete. It illustrates ASI as reliably outperforming tens of thousands of well-coordinated experts working on one problem for ten years, with the technology and cultural artifacts available in 2010. Chess, Go or protein-structure systems do not qualify. That is narrow superhuman performance, not general intelligence.

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

How fast are costs falling?

The essay's central number is stark. It combines hardware improvements, AI-hardware investment and algorithmic efficiency into a measure called global effective compute, which it estimates has grown at least tenfold per year. Extrapolated over a decade, that yields a factor of 10 billion — more same-quality AI agents than there are humans, for every agent runnable today, if the trends hold.

The 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 essay calls the resulting standardization of cognitive work "cognitive industrialization," analogous to the commoditization of routine physical labor. Even without frontier breakthroughs, it could reshape knowledge work and the wider economy.

Scale alone may still fall short of full AGI if some cognitive tasks require human help. The essay argues that faster, more numerous agents could then bridge the gap to a weak form of ASI, exploiting digital advantages: processing speed, memory, movement across hardware, nearly lossless copying of code and memory state, and high-bandwidth sharing of experience.

Where does each route get stuck?

Each pathway has named bottlenecks:

  • Scaling risks exhausting high-quality training data, hitting diminishing returns as each capability step demands multiplied compute and model size, and running into power generation, grid, chip fabrication and interconnect limits. Current scaling laws cannot confidently predict how much scale AGI or ASI would need.
  • Paradigm shifts must address continual learning, some forms of world modeling, decision-making and sample efficiency. Current systems may be unable to invent genuinely new concepts rather than recombine existing ones. Researchers are split on whether incremental transformer changes suffice.
  • Agent coordination lacks multi-agent scaling laws. Inefficiencies, misinformation, manipulation and collusion can weaken large groups, requiring institutional defenses.
  • Recursive self-improvement adds feedback loops through code, hardware designs, reusable deployment data or agent specialization. Poor feedback can stall or degrade the loops; progress should run fastest where outputs can be checked automatically, as with mathematical proofs.

The pathways can compound. If one stalls, economic and geopolitical competition can shift effort toward the others, the essay argues, so restricting only training compute would not necessarily stop overall progress. Suppressing one input could even create a capability "overhang," followed by steeper gains once the constraint is relaxed or circumvented.

What should institutions do now?

The essay prescribes a loop: model, measure, forecast and repeat, through interdisciplinary work spanning frontier labs and national and global institutions. Some metrics will require frontier-model access, sizable evaluation compute and new benchmarking methods.

The timetable to ASI remains unresolved. The nearer-term task is sharper: build evaluations that 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.

Source: unite.ai

Share this article:

More from Daniel Okafor

Daniel Okafor

Show full bio

Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.

2,425 articles

Related articles

« Previous article