Post
From AGI to ASI: Google DeepMind’s Roadmap for Superintelligence
Google DeepMind’s technical report “From AGI to ASI” maps the continuum of machine intelligence beyond human-level AGI: three definitions, four pathways that can compound, and the frictions that could slow them.
At a glance
- Type
- Post
- Topics
- Artificial general intelligence, Artificial superintelligence, Google DeepMind
- Evidence
- Editorial interpretation
- Written by
- Liyam Flexer

Google DeepMind’s technical report “From AGI to ASI” (arXiv:2606.12683) maps the continuum of machine intelligence beyond human-level AGI. First published in June 2026 (with updates through November 2026), the paper is a research analysis—not an official company roadmap—co-authored by Tim Genewein, Shane Legg, Marcus Hutter, Joel Z. Leibo, Iason Gabriel, Allan Dafoe, and others.
It treats intelligence as a continuum rather than a single threshold event. The central claim is that the transition from AGI to artificial general superintelligence (ASI) can follow multiple technological pathways that may proceed in parallel and compound. Because of this, a series of transformative shifts across science and technology is more likely than one abrupt leap.
Primary sources
- arXiv abstract
- arXiv PDF
- Google DeepMind publication page
- Related essay: “What comes after AGI?” by Tim Genewein (DeepMind Institute, 19 November 2026)
Definitions: AGI, ASI, and Universal AI
The report distinguishes three levels on the intelligence continuum:
- AGI (Artificial General Intelligence). Roughly median human-level performance on most cognitive tasks. This aligns with the “Competent AGI” level in the Morris et al. (2024) framework. The first systems that meet this bar are already expected to be superhuman on many individual tasks.
- ASI (Artificial General Superintelligence). A system that consistently outperforms large, coordinated groups of human experts on tasks that could take those groups years, using fewer resources and less time, across virtually all domains of human interest. Narrow superhuman systems such as AlphaGo or AlphaFold do not qualify. An ASI may itself be a collective of many instances.
- Universal AI (UAI). The theoretical upper bound, formalized via AIXI. It is incomputable; real systems can only approximate it from below. UAI serves as a formal anchor for the continuum rather than a practical target.
Digital intelligence has structural advantages that grow with compute: higher input/output and internal processing speed, far larger working memory, substrate independence, lossless replication of both code and learned state, and high-bandwidth sharing of experiences or gradients. These differences make large-scale collectives and high-speed iteration qualitatively different from human organizations.
Four Pathways from AGI to ASI
The pathways are not mutually exclusive. They can run in parallel, and progress on one can accelerate the others. Restricting a single dimension (for example, training compute) does not necessarily halt overall progress, because competitive pressure can shift effort elsewhere.
From AGI to ASI
Four interlocking pathways along the continuum of machine intelligence
- Scaling
- Paradigm shifts
- Multi-agent collectives
- Recursive improvement
1. Scaling AGI (compute, models, and data)
Continuation of the current paradigm: larger models, more data, more training compute, and more test-time compute.
Effective compute has historically grown on the order of ~10× per year through the combination of hardware improvements (~1.5×), investment growth (~2.5×), and algorithmic efficiency gains (roughly 3–6×). Illustrative projections in related commentary suggest that a cognitive workload costing on the order of $1 billion in compute today could fall toward roughly $1,000 within a decade, and that the effective population of AI agents could scale by more than three orders of magnitude by 2030 (and potentially far more over ten years) if trends continue.
Quantitative scaling can itself produce qualitative effects through more search, more parallel instances, and faster iteration.
2. AI paradigm shifts or evolutions
Changes beyond the current “large-transformer pretraining on human data + fine-tuning + test-time scaling” recipe. These could address limitations in continual learning, world modeling, sample efficiency, decision-making, or the invention of genuinely new concepts.
Some researchers expect incremental extensions of the current paradigm to be sufficient; others argue deeper architectural or training shifts will be required. Paradigm shifts are inherently difficult to forecast.
3. ASI emerging from large-scale multi-agent collectives
Coordinating thousands or millions of AGI-level instances into group agents, virtual research organizations, or agent economies. Collective performance can exceed what any single instance achieves, analogous to how human institutions outperform individuals.
This path depends on multi-agent scaling laws and on institutional or architectural designs that limit misinformation, collusion, and coordination failures.
4. Recursive improvement
AI systems accelerating the development of later systems—by improving code and algorithms, generating higher-quality data, designing better hardware, or running self-play-style loops. This is the classic idea associated with I. J. Good and Ray Solomonoff.
Recursive improvement is expected to advance fastest in domains with strong, scalable verification (mathematics, code, certain scientific benchmarks) and can stall where feedback is weak, noisy, or expensive.
Frictions, Bottlenecks, and Open Questions
The report treats potential limits as open empirical questions rather than certain blockers. Major candidates include:
- Data scarcity and the limited reliability of purely synthetic data
- Diminishing returns and practical ceilings on model scale and compute build-out (power, chips, interconnects, economics)
- Coordination inefficiencies and risks inside large multi-agent systems
- Evaluation problems once systems exceed human performance (how to generate reliable feedback without humans in the loop)
- An “abstraction barrier”—whether systems trained primarily on human-generated concepts can invent truly novel abstractions without grounded discovery
- Deliberate slowdowns or regulation
Whether these remain minor frictions or become major constraints is left as a research agenda. The authors emphasize that determining the practical impact of these frictions requires concrete measurement and forecasting work.
Key Implications
Because the pathways can compound and because uncertainty around the frictions remains high, the report argues that a single sharp “AGI arrives and everything changes” event may be less accurate than a series of transformative shifts driven by accelerating AI-enabled progress across science and technology.
Preparing for that prospect, the authors conclude, requires a large-scale interdisciplinary effort spanning technical measurement, forecasting, economics, and governance.
Summary
“From AGI to ASI” moves the discussion of superintelligence from pure speculation toward a concrete set of technological pathways, measurable advantages of digital intelligence, and specific open research questions. It does not claim that ASI is imminent or inevitable; it provides a structured map of how the transition could unfold and what would need to be measured and governed along the way.
Read the full report
Evidence in this article Editorial interpretation Our reading of what the evidence means. Judgment, clearly separated from the evidence itself.
Frequently asked questions
What is Google DeepMind’s “From AGI to ASI” report?
It is a technical report (arXiv:2606.12683) that maps the continuum of machine intelligence beyond human-level AGI. It was first published in June 2026, with updates through November 2026. The article describes it as a research analysis, not an official company roadmap.
How does the report define AGI, ASI and Universal AI?
AGI is roughly median human-level performance on most cognitive tasks. ASI is a system that consistently outperforms large, coordinated groups of human experts across virtually all domains of human interest, using fewer resources and less time. Universal AI is the theoretical upper bound, formalized via AIXI; it is incomputable, so real systems can only approximate it.
What are the four pathways from AGI to ASI?
Scaling AGI through more compute, larger models and more data; AI paradigm shifts or evolutions; ASI emerging from large-scale multi-agent collectives; and recursive improvement, where AI systems accelerate the development of later systems. The pathways are not mutually exclusive and progress on one can accelerate the others.
Does the report say superintelligence is imminent?
No. According to the article, the report does not claim that ASI is imminent or inevitable. It provides a structured map of how the transition could unfold and what would need to be measured and governed along the way.
What could slow the transition from AGI to ASI?
The report lists data scarcity and unreliable synthetic data, diminishing returns and practical ceilings on scale and compute, coordination problems in large multi-agent systems, the difficulty of evaluating systems that exceed human performance, an “abstraction barrier”, and deliberate slowdowns or regulation. It treats these as open empirical questions.
Related reading
- Category: Research & knowledge
Sources
- From AGI to ASI (arXiv:2606.12683), abstract arXiv, accessed . Submitted 10 June 2026; version 3 dated 10 November 2026.
- From AGI to ASI, full text (PDF) arXiv, accessed .
- From AGI to ASI, publication page Google DeepMind, accessed .
- What comes after AGI?, essay by Tim Genewein DeepMind Institute, accessed . Dated 19 November 2026.