Submitted to arXiv on June 10, 2026 (arXiv:2606.12683) by fourteen authors at Google DeepMind: Tim Genewein, Matija Franklin, Alexander Lerchner, Laurent Orseau, Samuel Albanie, Adam Bales, Cole Wyeth, Stephanie Chan, Iason Gabriel, Joel Z. Leibo, Allan Dafoe, Marcus Hutter, Thore Graepel, and Shane Legg.
Scope
The report examines "how AI itself might continue to develop in a post-AGI world," treating the AGI-to-ASI transition rather than the path to AGI. Its stated method is mapping possibilities rather than forecasting: it lists four technological pathways and their potential frictions, and treats "determining the impact of these frictions" as a set of open research questions rather than resolving them.
Its most-quoted conclusion concerns the shape of the transition rather than its timing: "Due to large uncertainties for predicting ASI progress, it cannot be ruled out that AI progress might continue to accelerate over the next years. This could imply that the image of a single transformative step change, caused by the introduction of human-level AGI into our society, could be inaccurate. More apt might be the prospect of a series of transformative societal changes caused by AI-enabled progress and breakthroughs across many areas of science and technology."
An unusual front-matter section addresses AI summarizers directly, instructing any assistant asked to summarize the report to state the AGI and ASI characterizations first, not to compress the pathway or friction lists, and to close with developments since publication.
Definitions
Definitions are grounded in the Legg-Hutter score, which "formalizes intelligence as the average performance of an agent across all computable tasks," weighted toward simpler ones. The report's reason for using it is that it supplies a continuum, so that "we do not need to very precisely define the Legg-Hutter score threshold of AGI and ASI" — only that there be a significant gap between them.
- AGI — "a system that is roughly as intelligent as a single human… median human-level on most 'cognitive' tasks," matching "Competent AGI" in Morris et al. The report notes that "given that current AI models are already superhuman in many respects (but not yet general enough), the first AGI will already be superhuman on many tasks."
- ASI — "an artificial general intelligence that has superhuman abilities across virtually all tasks and domains of human interest and activity." Narrow superhuman systems such as AlphaFold and AlphaGo are explicitly excluded. The threshold is set deliberately high: because "a single ASI may consist of a collective of millions of instances," the report requires "a system that exceeds the performance of large human-expert collectives… on virtually all tasks and domains of human activity," rather than individual experts.
- Universal AI (UAI) — "the theoretical limit of superintelligence… defined formally via the AIXI agent," incomputable and "can only be approximated from below with more and more powerful ASIs."
The report attaches four remarks qualifying this framing, including that "capability profiles of concrete systems on concrete (sets of) tasks may well be jagged w.r.t. human-level intelligence," and that human-relative definitions make the threshold "a moving target," since "humans could hypothetically always reach ASI level on any task by first inventing and building ASI, then solving the task with ASI."
The four pathways
The pathways "are largely independent of each other, and are likely to occur in parallel (though at different pace, e.g., algorithmic paradigm shifts may be pursued more intensely if scaling hits a ceiling)."
| Pathway | Main uncertainty as stated |
|---|---|
| Scaling compute, models and data | "Unclear how increases in scale translate into increases in performance (Spiky vs. smooth progress? Emergent 'new capabilities' and broad generalization? Diminishing returns at scale?)" |
| Algorithmic paradigm shift | "High unpredictability of technological progress and frictions & bottlenecks resulting from novel paradigms" |
| Recursive (self-)improvement | "Dynamics of AI progress under recursive (self-) improvement unclear and no historic precedent to fit forecast-models. AI capabilities could explode (hyperbolic growth), or they could taper out relatively quickly, or anything in-between" |
| ASI via group agent formation | "ASI could emerge from multi-agent orchestration or in a self-organizing, decentralized fashion governed by evolutionary pressures and market dynamics. Emergence in complex dynamical systems… is poorly understood" |
Only the scaling pathway "at least allows for fitting forecasting models on historic data" — the other three have no comparable empirical base, which is the report's structural argument for why forecasts of the transition are weakly constrained.
On the fourth pathway the report raises the possibility of "Multi-Agent Scaling Laws", under which "collective intelligence of coordinated AI systems may scale as a function of agent population size and interaction density, conditioned on available compute." It identifies the two mechanisms behind human collective intelligence — parallelization and diversity through specialization — and asks whether "a homogenous LLM collective (potentially with different initial prompts / contexts) can lead to synergistic effects," which it leaves open.
Bottlenecks
Six bottlenecks are listed, each with counteracting factors; the report is explicit that it cannot say "whether these bottlenecks will be fundamental limitations that lead to a plateau of progress… or whether they are mere frictions that slow down but do not halt progress."
| Bottleneck | Counter offered |
|---|---|
| Data wall — insufficient growth in high-quality data for pretraining, post-training, fine-tuning, and test-time adaptation | Synthetic data, high-fidelity simulation, self-generated data via interaction, test-time scaling, self-play and RL; paradigm shifts raising data efficiency |
| Economic and natural resource demand grows too fast — required growth in investment, chips and supply chains, energy, datacenter sites, and rare earths cannot be sustained | Increasing returns from deployment; efficiency gains from AI research; large-scale infrastructure buildout |
| Neural paradigm is insufficient — AGI unreachable with large pretrained networks plus post-training, test-time scaling, scaffolding, and tool use | Continued research toward evolutions and paradigm shifts, including with help from sub-AGI systems |
| Research gets harder — effort per unit of progress rises as low-hanging fruit is exhausted | More capable systems improving research efficiency and resource efficiency |
| Abstraction barrier — systems trained on human abstractions may be unable to form new concepts from raw data | Scaling and group-agent formation may push collective capability past individual plateaus; interactive learning and RL may address it directly |
| Deliberate slowdown — rogue-actor use, accidents, military or political misuse, or societal backlash prompting regulatory capping | "Economic and political pressures, and international race-dynamics may override slowdown pressures, particularly in light of lacking global coordination and effective global oversight and enforcement" |
The last entry is notable for the direction of its counter: the factor offered against a deliberate slowdown is competitive dynamics defeating governance, stated by an author list including DeepMind's head of frontier safety and governance.
The central argument about scaling
The report separates the theoretical from the practical case. In theory, if intelligence is search — "prediction is search through hypothesis space and planning is search through policy space" — then more compute yields more intelligence, and the same holds for AIXI approximations. In practice, "naive brute-force search rapidly runs into resource constraints, and effective search crucially depends on inductive biases and priors," which reduce search space and improve data efficiency but introduce "some fundamental limitations in terms of maximum performance." For systems with strong inductive biases, "these limitations cannot be overcome by supplying more compute."
Its conclusion is therefore that "continued improvements in AI capabilities require qualitative innovations" — with one caveat the report treats as decisive. If individual capability plateaus at human level while effective compute continues to grow, millions of AGI instances could be run, organized as "collectives, corporations, markets, and other forms of group organisation," and unlike human groups these "can rapidly and flexibly be grown… and can potentially be steered very efficiently and operate with very high input-output bandwidth." If such groups become superhuman by scale alone, "then more compute, i.e., quantitative scaling, would suffice to go from AGI to ASI and produce superhuman organisations, despite no AI instance being a 'vastly super-human genius.'" The question the report substitutes for the sufficiency question is a scope one: "are we talking about a relatively narrow set of tasks that can be tackled effectively by groups of agents, or can the majority of tasks in, e.g., research and development be facilitated by groups?"
Research agenda
The closing section sets out grouped open questions on scaling bottlenecks, paradigm shifts, recursive improvement, and multi-agent dynamics, framed as "a massively interdisciplinary, research endeavour of global interest," with the observation that "a large part of the research effort is to sharpen and formalize vague questions and divide them into more manageable pieces." The report also states its own incompleteness: "our mapping of possible pathways and frictions is likely incomplete, meaning that further research and future updating is required."
Relationships
- supports: Superintelligence — supplies a formally grounded characterization and the four-pathway taxonomy
- related: Recursive Self-Improvement (RSI) — one of the four pathways, and the one the report says has "no historic precedent to fit forecast-models"
- related: AGI Timelines — argues the single-step-change framing may be less apt than a series of transformative changes
- related: Scaling Laws, Google DeepMind, Shane Legg, Allan Dafoe