What Comes After AGI? Google DeepMind Maps Four Paths to Superintelligence

In a DeepMind Institute essay, research scientist Tim Genewein maps four routes from AGI to artificial superintelligence: scaling, paradigm shifts, agent coordination and recursive self-improvement. It is based on a technical report co-authored with Shane Legg and others and argues for rigorous measurement of progress.

Christopher Ort
By Christopher Ort · 3 min read
Share
A sunlit library desk with a notebook of branching hand-drawn lines, a cup of tea and books.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Background
  5. What's in the essay
  6. Why it matters
  7. What's still open
  8. Sources
  9. i10x Perspective

Google DeepMind has published an essay mapping how AI could move beyond human-level general intelligence toward superintelligence. In “What comes after AGI?”, research scientist Tim Genewein argues that “the regime beyond AGI is shifting from philosophical speculation to empirical research and governance questions,” and lays out four technical routes from AGI to artificial superintelligence, or ASI.

i10X Quick Take

Quick Take

  1. What it is: an essay on the DeepMind Institute site, dated 19 November 2026, based on the technical report “From AGI to ASI” co-authored with 13 researchers including DeepMind co-founder Shane Legg.
  2. Four pathways: scaling models and data, technological paradigm shifts, better coordination of many agents, and recursive self-improvement.
  3. Key estimate: global “effective compute” could grow at least 10x per year when hardware gains, algorithmic efficiency and spending growth are multiplied.
  4. Policy point: restricting just one input, such as training compute, “won’t be enough to stop overall progress,” the essay says.

In plain English

AGI means AI that matches people across nearly all thinking tasks. ASI goes further: systems that beat large teams of top experts. DeepMind’s essay says this is no longer only a philosophy question. It describes four ways AI could get there, explains what might slow each one down, and says researchers and governments need better ways to measure progress before it happens.

What happened

The essay defines ASI as “an advanced system that consistently outperforms large groups of human experts on tasks that could take them years to complete, using fewer resources and less time,” across “virtually all tasks of human interest.” It notes that “now that AGI is coming within reach, we can no longer put off seriously considering what comes next.”

Genewein stresses uncertainty: “None of these predictions are certain, but the pathways and scenarios are also far from unfounded hype or fiction.” The co-authors of the underlying report include Matija Franklin, Alexander Lerchner, Laurent Orseau, Allan Dafoe, Marcus Hutter, Thore Graepel and Shane Legg.

Background

Superintelligence has become a mainstream topic in recent weeks. Geoffrey Hinton warned in a television interview that AI is making “alien beings”, and lab researchers have been publishing their own forecasts. DeepMind’s piece is notable because it comes from a frontier lab and frames the question as something to model and measure.

What’s in the essay

  • Cognitive industrialization: compute per dollar has grown about 1.5x a year, algorithmic training efficiency 3 to 6x a year and AI hardware spending about 2.5x a year. Together that implies effective compute growth of at least 10x a year. The essay estimates the effective population of today’s AI agents might scale more than a thousand-fold by 2030.
  • Falling costs: work that costs $1 billion in compute tokens today “would cost just $1000 in ten years” if inference efficiency keeps improving about 4x a year.
  • Scaling limits: data exhaustion, diminishing returns on model size, and physical limits such as power plants, grids and chip factories.
  • Paradigm gaps: today’s models fall short of humans on continual learning, some forms of world modeling, decision-making and sample efficiency.
  • Agent collectives: coordinating thousands or millions of AI instances could produce “collective superintelligence,” but the field still needs “multi-agent scaling laws.”
  • Recursive self-improvement: AI improving future AI could speed up progress, but whether it runs away “is largely an empirical question.” It should advance fastest where results can be checked automatically, such as mathematical proofs.

Why it matters

For policymakers and safety researchers, the essay argues that because there are several routes to ASI, holding back one input risks a capability “overhang,” where progress jumps suddenly once the bottleneck is removed.

For companies, the “cognitive industrialization” argument is the practical part: even without new breakthroughs, cheaper and more numerous agents could reshape knowledge work.

For researchers, it calls for “model, measure, forecast, and repeat,” noting that many key metrics require frontier-model access and large evaluation budgets.

What’s still open

  • The essay offers no date for AGI or ASI.
  • It does not quantify how far today’s systems are from “Expert AGI.”
  • Who should run the large-scale measurement efforts it calls for, and how labs would share data with outside institutions.

Sources

  • DeepMind Institute, “What comes after AGI? From AGI to Superintelligence,” Tim Genewein, 19 November 2026
  • Google DeepMind, technical report “From AGI to ASI”

i10x Perspective

Frontier labs talking openly about the road past AGI is itself news. DeepMind’s framing is sober: no timeline, many caveats, and a demand for measurement over hype. The economic case, cheaper and more plentiful agents, deserves the most attention from businesses today. More in Policy & Safety.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Timeline

What happened before

  1. Release

    Gemini 4 Argon

    1.05M context · $2 / $10 per 1M tokens

    Google's most capable model for complex coding, agentic, math and science tasks.

    ai.google.dev (source, opens in a new tab)
  2. Release

    Gemini 3.8 Flash

    1.05M context · $0.75 / $3.75 per 1M tokens

    ai.google.dev (source, opens in a new tab)
  3. Release

    Gemini 3.7 Flash

    1.05M context · $0.75 / $3.75 per 1M tokens · Deprecated 8 Oct 2026 (API routes to Gemini 3.8 Flash)

    ai.google.dev (source, opens in a new tab)
  4. Release

    Gemini 3.6 Flash & 3.5 Flash-Lite

    1.05M context · $0.75 / $3.75 per 1M tokens

    ai.google.dev (source, opens in a new tab)
  5. Partnership

    Google to pay SpaceX $920M a month for Colossus compute (~$30B through 2029)

    With Google · Compute · $30B

    Per SpaceX's SEC filing: about 110,000 NVIDIA GPUs from October 2026 to June 2029, ramping through September at a reduced fee; either side may terminate after Dec 31, 2026 with 90 days' notice (CNBC, Bloomberg).

    cnbc.com (source, opens in a new tab)
Full Google timeline ›

Editor's picks

  1. What Comes After AGI? Google DeepMind Maps Four Paths to Superintelligence
  2. AMD and Unsloth Bring NPU Support to Unsloth Desktop, Letting 100+ AI Agent Apps Run Locally on Ryzen AI
  3. Perplexity Brain Now Writes Its Own Agent Skills, With 80,000 Created for 20,000 Users
  4. OpenAI Performance Factory: Codex Agents Ran 1,700 Investigations to Keep ChatGPT Fast
  5. OpenAI GPT-Live-1.1: New Voice Model Improves Transcription and Accents, Adds Cheaper Standby