Breaking

DeepMind Publishes Paper on the Move From AGI to ASI

June 14, 2026 at 01:50 EDT

Research & Papers
Foundation Models

Google DeepMind on June 10, 2026 released a theoretical paper titled "From AGI to ASI" on arXiv, focused on the transition from human-level AGI to artificial superintelligence (ASI). The paper lists Tim Genewein as lead author alongside 14 contributors, including Marcus Hutter, known for AIXI theory, DeepMind co-founder Shane Legg, and Iason Gabriel, most of them affiliated with Google DeepMind.

June 10, 2026 · Google DeepMind

From AGI to ASI: Mapping the Road Past Human-Level AI

Fourteen DeepMind researchers — including AIXI's Marcus Hutter and AGI-definition pioneer Shane Legg — publish a ~60-page theoretical roadmap that takes human-level AI as a given and charts a continuous path toward superintelligence.

pathways from AGI to ASI

DeepMind authors

~60

page conceptual roadmap

How the report defines ASI

A system that surpasses, in nearly all domains, what tens of thousands of coordinated human experts could achieve over a decade.

AGI

single human expert level

ASI

10,000s of experts × 10 years

Four pathways from AGI to ASI

1 · Scaling

More compute, model size, data and test-time inference.

2 · Paradigm shifts

New architectures and learning methods beyond the transformer.

3 · Recursive improvement

AI accelerating its own AI research in a self-improving loop.

4 · Multi-agent collectives

Emergent superintelligence from large populations of specialized agents.

The recursive self-improvement loop

AI does AI research

→

Better AI built

→

Faster research

↺

Hard limits that could slow progress

Data exhaustion
Energy & hardware demands
Rising research difficulty
Limits to abstraction
Deliberate regulation

Physics, complexity theory and real-time constraints make even ASI finite.

Welcomed

Thoughtful and restrained; treats ASI as an accelerating AI civilization and spotlights the multi-agent path.

Criticised

Speculative — no implementation, benchmarks or empirical validation.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.