TL;DR
Fourteen researchers, most affiliated with Google DeepMind, posted a 57-page arXiv report on June 10, 2026 that sets out a framework for thinking about progress from human-level AGI to ASI. The paper is a conceptual report, not an experiment, and its main claim is that post-AGI progress may arrive through several waves under high uncertainty.
A team of 14 researchers, most affiliated with Google DeepMind, posted a 57-page arXiv report on June 10, 2026 that maps possible routes from human-level artificial general intelligence to artificial superintelligence, a shift that matters because it comes from senior AI figures including DeepMind co-founder Shane Legg and universal-intelligence theorist Marcus Hutter.
Confirmed facts are narrow: the document is a 57-page arXiv paper titled From AGI to ASI, posted June 10, 2026 by 14 authors. Most are at Google DeepMind, according to the source material, and the list includes Shane Legg, a DeepMind co-founder associated with the term AGI, and Marcus Hutter, whose work on universal intelligence underpins part of the report’s framework. Thorsten Meyer AI reported that the paper crossed 54,000 views within days.
The report does not present new benchmark scores or experimental results. It lays out a continuum from today’s AI systems to human-level AGI, then ASI, and finally a theoretical ceiling called Universal AI. The authors set ASI above individual human performance: the source summary says their working bar is a system that reliably beats large, coordinated expert groups across nearly every domain.
The authors argue that progress beyond AGI could come through four overlapping paths: scaling, new paradigms, recursive self-improvement and multi-agent collectives. They also project roughly 10 times annual growth in effective compute by combining hardware gains, investment growth and algorithmic efficiency, which they say could amount to about 10,000 times more effective compute by 2030 if those trends persist.
Waves, not a wall: the road past AGI
A 57-page DeepMind report maps how AI might keep advancing after human-level AGI. Its headline: the future may not be one big “step change,” but a series of transformative waves — under enormous uncertainty.
A careful, sober map that resists both doom and rapture — and refuses to promise the usual singularity miracles. But it’s a position paper from a party with a stake in the destination, anchored to its own authors’ theory, and it deliberately brackets the economics, labor, and how humans fit in — the part that matters most. Useful terrain map; drawn by people who own the land.
Post-AGI Safety Moves Upstream
The report matters because it moves the center of debate from whether AI can reach median human ability to how the field should reason about systems that could outperform institutions. Its framing treats AGI not as an endpoint, but as a point on a wider machine-intelligence curve.
That affects safety and policy debates. If the authors are right that post-AGI progress may come as waves across science and the economy, regulators, labs and affected workers may face uneven changes rather than a single date when ASI arrives. The paper also carries weight because several authors work at a leading AI lab with a direct stake in the trajectory it describes.
Compiler Engineering for AI Hardware: MLIR, TVM, XLA, and Custom Backends for Neural Network Accelerators (AI Infrastructure, Hardware & Compiler Engineering Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Legg-Hutter Theory Returns
The report leans on AIXI and the Legg-Hutter score, a 2007 formalization that defines intelligence through average performance across computable tasks. Because Legg and Hutter are among the named authors or intellectual sources tied to that framework, the yardstick is internally consistent but not neutral.
The source material also notes an unusual feature: the report opens with instructions for AI assistants expected to summarize it, including guidance on points not to compress and a request that future systems report how its forecasts aged. That detail places the paper inside an AI-mediated research culture, where papers are written with machine readers in mind as well as human ones.
“From AGI to ASI”
— Genewein et al., arXiv report
Perplexity AI: The Research Playbook: Master AI-Powered Research — From Pro Search to Deep Research, Spaces, and Beyond (AI for Everyone)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Economic Effects Stay Unmapped
The report leaves major questions unresolved. The 10-times annual effective-compute projection depends on continued investment, hardware improvement and algorithmic gains; any of those could slow. The source summary also says high-quality text data may run short this decade, which could limit plain scaling.
Several central mechanisms are presented as possibilities rather than predictions. New paradigms are, by their nature, hard to forecast. Recursive self-improvement could accelerate AI research, produce limited gains, or land between those outcomes. The report also brackets questions about labor markets, wealth, governance and how people would retain agency around systems that outperform institutions.
Cybersecurity in the Age of AGI and Superintelligence: A Simulacra‑Based Textbook: Constitutional Cybersecurity for Defence, Intelligence, and the Future of AI Security
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Forecasts Face Real Tests
The next stage is scrutiny: researchers, policy groups and labs are likely to test the framework against new model releases, AI-assisted research systems and evidence about compute growth. Since the paper is on arXiv, peer review and public critique may change how its definitions and projections are received.
The authors’ claims will age quickly. By 2030, readers should be able to compare the paper’s effective-compute projection, scaling limits and post-AGI pathways with what actually happened in deployed systems and AI research workflows.
The Hundred-Page Machine Learning Book (The Hundred-Page Books)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Did Google DeepMind announce artificial superintelligence?
No. The source material describes an arXiv report by 14 researchers, most at Google DeepMind, not an announcement that ASI has been built.
Is the report peer reviewed?
The source identifies it as an arXiv report. arXiv posting allows rapid public release, but it does not by itself mean the paper has completed peer review.
What does ASI mean in this report?
The report uses a high bar: ASI is not just smarter than one person. It refers to systems that can outperform large, coordinated groups of human experts across nearly all domains.
Why does the report describe waves instead of one sudden break?
The authors map several possible routes beyond AGI, including scaling, new methods, AI-assisted AI research and multi-agent systems. The source material says these could happen in parallel and affect different parts of science and the economy at different times.
What does the paper leave open?
It does not resolve whether compute growth will continue, whether data limits will slow progress, how recursive self-improvement would behave, or how labor, governance and human agency would be affected by systems beyond institutional human capability.
Source: Thorsten Meyer AI
