Fourteen researchers, most of them at Google DeepMind, posted a 57-page report to arXiv on June 10 that maps possible routes from human-level artificial general intelligence to artificial superintelligence, a question that matters because the authors argue current AI safety debates may be too focused on reaching AGI and less prepared for what could follow.

The report, titled From AGI to ASI, is a conceptual framework rather than a new experiment. It does not present fresh benchmark results. Instead, it proposes a way to reason about a continuum of machine intelligence: today’s AI systems, human-level AGI, artificial superintelligence, and a theoretical ceiling the authors call Universal AI.

The author list includes Shane Legg, a DeepMind co-founder associated with popularizing the term AGI, and Marcus Hutter, whose work on universal intelligence underpins part of the report’s formal framing. The paper uses the AIXI framework and the Legg-Hutter intelligence measure as reference points, a choice that is coherent with the authors’ prior work but not a neutral standard across the field.

The report defines artificial superintelligence as more than a system that outperforms one person. Its working bar is a system that can beat large, coordinated groups of human experts across almost all domains. By that definition, narrow systems such as AlphaGo or AlphaFold would not qualify, even though they already exceed human performance in specific tasks.

A Wider Safety Debate

The report matters because it shifts attention from whether machines reach human-level performance to how quickly and unevenly they might move beyond it. The authors argue that digital systems could gain advantages from speed, copying, memory transfer, parallel operation, and shared learning that biological intelligence cannot match in the same way.

According to the report, three forces could combine to increase “effective compute”: cheaper hardware, higher investment, and better algorithmic efficiency. The authors estimate a rough growth rate of about 10 times per year, which they say could amount to a 10,000-fold increase by 2030 if current patterns continued. That projection is a claim, not a confirmed forecast.

For readers, the policy issue is practical: if the path from AGI to stronger systems is gradual but fast, oversight may need to track capability waves across science, coding, research, economics, and security rather than wait for one clear AGI moment.

Four Routes Beyond AGI

The paper lays out four possible pathways from AGI to artificial superintelligence. The first is scaling: larger models, more compute, more data, and continued efficiency gains. The report flags limits here too, including the possibility that high-quality text data may become scarcer this decade.

The second route is paradigm change, meaning new architectures or methods that are difficult to predict in advance. The third is recursive self-improvement, where AI systems accelerate AI research itself. The authors present that route as highly uncertain: it could accelerate quickly, stall, or produce mixed outcomes.

The fourth route is multi-agent collectives, where superintelligent performance might come not from one model but from many AI agents working together. The report’s central framing is that these routes are likely to overlap, producing waves of change rather than one single wall between AGI and superintelligence.

“From AGI to ASI”

— Genewein et al., arXiv report

Forecasts Still Highly Uncertain

It is not clear whether the compute-growth assumptions in the report will hold through 2030, whether new data sources will offset limits in text data, or whether recursive AI research will produce rapid capability gains. The authors also do not settle how governments, companies, workers, and public institutions would adapt if AI systems began outperforming large expert groups.

The report’s framing also comes from researchers connected to one of the companies most invested in advanced AI development. That does not invalidate the analysis, but it means readers should treat it as a structured argument from interested experts, not a settled field consensus.

Research Agenda Moves Forward

The next step is scrutiny from other researchers, safety teams, policymakers, and industry labs. The paper is on arXiv, so its claims and definitions remain open to challenge and revision. Watch for whether later work tests the report’s assumptions about compute growth, multi-agent systems, and AI-assisted AI research.

Key Questions

What happened on June 10, 2026?

A team of fourteen researchers, most affiliated with Google DeepMind, posted a 57-page arXiv report titled From AGI to ASI.

Is this a new AI model or benchmark result?

No. The report is a conceptual framework and research agenda. It does not announce a new model or new benchmark performance.

How does the report define artificial superintelligence?

It describes ASI as general intelligence that can outperform large, well-coordinated groups of human experts across almost every domain, not just one person or one task.

Why does the report describe waves rather than a wall?

The authors argue that progress beyond AGI may come through several overlapping routes, including scaling, new methods, AI-assisted AI research, and multi-agent systems.

What remains unknown?

The pace of capability growth, the effect of data limits, the role of recursive self-improvement, and the social and labor effects of post-AGI systems all remain unsettled.

Source: Thorsten Meyer AI