TL;DR
Fourteen researchers, mostly from Google DeepMind, posted a 57-page arXiv report on June 10 examining how AI could move from human-level AGI to artificial superintelligence. The paper is a conceptual framework, not an experiment, and its claims about compute growth, recursive AI research, and multi-agent systems remain forecasts rather than confirmed outcomes.
A team of 14 researchers, most of them at Google DeepMind, posted a 57-page arXiv report on June 10 mapping how AI systems might move from human-level artificial general intelligence to artificial superintelligence, shifting the debate toward what could happen after machines reach broad human-level cognitive performance.
The report, titled From AGI to ASI and listed as arXiv:2606.12683, is not a new model release or benchmark result. It is a framework for thinking about post-AGI progress, written by authors whose list includes DeepMind co-founder Shane Legg and Marcus Hutter, whose work on universal intelligence underpins part of the paper’s theoretical frame.
The authors describe machine intelligence as a continuum: today’s narrow and unevenly general AI systems, human-level AGI, artificial superintelligence, and a theoretical ceiling called Universal AI. They set a high threshold for ASI, defining it as performance beyond large, coordinated groups of human experts across nearly all domains, rather than a system that beats a single person at many tasks.
What is confirmed is the publication and content of the report. The paper’s forecasts are claims by its authors: they argue that scaling, new paradigms, recursive AI-assisted AI research, and multi-agent collectives could combine in the move from AGI to ASI. They also estimate that effective compute has been rising by about 10 times per year when hardware, investment, and algorithmic efficiency are counted together, but that projection is not proof that ASI will arrive on a set timeline.
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 Planning Moves Upstream
The report matters because it moves the safety and governance question beyond whether human-level AGI is possible and toward what systems might do after that point. If AI progress continues after AGI, risk models based only on human-level capability may miss changes in speed, scale, replication, and institutional power.
The paper’s strongest practical claim is that digital systems have advantages biology does not: they can copy weights and memory state, run faster with more compute, share learning across instances, and operate at machine speed. The authors say those traits could make a large number of human-level systems function differently from a human workforce, even before any single system looks qualitatively alien.
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.
The Legg-Hutter Yardstick Returns
The report leans on AIXI and the Legg-Hutter score, a formal approach from 2007 that defines intelligence as average performance across computable tasks. That choice gives the paper a rigorous internal structure, but it also means the authors are partly measuring the future against a theory associated with two people on the author list.
The paper separates ASI from familiar narrow superhuman systems. Tools such as AlphaGo and AlphaFold can exceed human experts in defined tasks, but the report says ASI would need broad strength across science, engineering, strategy, and other domains, at a level exceeding large expert organizations.
“From AGI to ASI”
— Genewein et al., in the arXiv report title
AI research supercomputers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Forecasts Still Outrun Evidence
Several parts of the report remain uncertain. The authors identify four pathways, but they do not show that any one path will produce ASI, nor do they settle whether recursive self-improvement would accelerate AI research, stall, or follow an uneven course.
The compute case is also conditional. The report cites trends in hardware, spending, and algorithmic efficiency that could add up to very large effective-compute gains by 2030, but high-quality training data, energy supply, chip capacity, model design, regulation, and market demand could alter that path.
The report also brackets public-facing questions about labor disruption, ownership of AI systems, accountability when many agents act together, and how people retain control if expert-level machine work becomes cheap and abundant.
machine learning compute resources
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Scrutiny Turns To Research Agendas
The next step is scrutiny from AI researchers, safety teams, policy specialists, and rival labs. The report is likely to be read less as a settled forecast than as a proposed map: a way to organize experiments, test assumptions, and decide which warning signs would matter before a system qualifies as ASI.
One unusual marker of the moment is that the paper reportedly opens with instructions for AI assistants that may summarize it, including directions about which points not to compress and a request for future systems to report how its predictions aged. That detail is not evidence for the predictions, but it shows that the authors expect AI systems themselves to be part of how the work is read and remembered.
multi-agent AI systems
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did the DeepMind-linked researchers publish?
They posted a 57-page arXiv report titled From AGI to ASI on June 10, 2026. It maps possible pathways from human-level artificial general intelligence to artificial superintelligence.
Did the paper announce that AGI or ASI has arrived?
No. The report is a conceptual framework and research agenda. It does not present a new AI system, benchmark result, or confirmed date for AGI or ASI.
How does the report define artificial superintelligence?
The authors set the bar above individual human skill. ASI, in their framing, would outperform large, coordinated expert collectives across nearly all domains.
What are the four pathways the authors describe?
The report points to scaling, new paradigms, recursive AI-assisted AI research, and multi-agent collectives. The authors say these routes could develop in parallel.
What remains unknown after the report?
It is not clear whether the proposed pathways will produce ASI, how fast progress could occur, or how economic, labor, governance, and control issues would be handled if such systems emerge.
Source: Thorsten Meyer AI
