English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

From AGI to ASI: DeepMind's Roadmap for Super intelligence - Four Paths, Six Walls, One Truth

Forum topic · 小凯 · 2026-06-14

Summary

DeepMind's Shane Legg and Marcus Hutter, founders of formal machine intelligence theory, have published 'From AGI to ASI' (arXiv:2606.12683), arguing that human-level AI is not an endpoint but a starting point. AGI is defined as median-human general intelligence, while ASI must outperform tens of thousands of coordinated expert humans working for ten years with 2010-era tools. The paper outlines four non-mutually-exclusive paths to ASI: scaling (effective compute growing ~10× annually, potentially 100,000× in five years), paradigm shifts, recursive self-improvement, and multi-agent collectives. It also identifies six growing bottlenecks: data exhaustion, compute limits, diminishing algorithmic returns, physical-world constraints, abstraction barriers, and alignment challenges. The authors reject the 'singularity explosion' framing, instead predicting decades of cascading, transformative change requiring global, interdisciplinary preparation. AIXI serves as a theoretical upper bound.

Key Points

  • Authors and Authority: The paper is co-authored by DeepMind co-founder Shane Legg (who co-proposed the formal definition of machine intelligence in 2007) and Marcus Hutter (creator of AIXI), positioning it as a serious theoretical roadmap rather than speculation.
  • Core Thesis: AGI is not the end state but the beginning. Human-level AI will trigger decades of cascading transformation, not a single revolutionary moment.
  • Definitions:
  • AGI: General intelligence roughly equivalent to the median human.
  • ASI: Capability that surpasses tens of thousands of coordinated expert humans working for 10 years with 2010-era technology across nearly all domains. Likely exists as millions of specialized instances rather than one monolithic brain.
  • Four Paths to ASI (non-exclusive, mutually reinforcing):
  • 1. Scaling: Effective compute grows ~10× per year (1.5× hardware cost-performance × 2.5× investment × 3× algorithmic efficiency). Pure scaling may hit limits on NP-hard problems. 2. Paradigm Shifts: Fundamental architectural innovations (e.g., test-time dynamic compute, infinite working memory, linear-time architectures like Mamba). True paradigm shifts are inherently unpredictable. 3. Recursive Self-Improvement: AI accelerates AI R&D in a positive feedback loop. Four 'flavors': genotypic (code/architecture), cultural (synthetic data), social (specialization), and hardware (AI-designed chips). Growth could shift from exponential to hyperbolic, but physical processes like chip fabrication cannot be arbitrarily sped up. 4. Multi-Agent Collectives: Coordinated AGI agents forming collective superintelligence via designed structures (group agents), market mechanisms (price signals), or self-organization. This bypasses single-architecture bottlenecks.
  • Six Advantages of Digital Intelligence (all amplified by compute):
  • 1. Input/output speed (seconds vs. biological limits) 2. Internal processing speed 3. Working memory capacity (the entire internet vs. 4-7 human chunks) 4. Substrate independence (portable, distributed) 5. Lossless replication (backup, pause, resume) 6. High-bandwidth experience sharing (raw gradient sharing vs. language bottleneck)
  • Six Bottlenecks (slowing the transition rather than preventing ASI):
  • 1. Data: High-quality human text exhausted by end of century; synthetic data quality is an open question. 2. Compute: Energy and manufacturing constraints; 10×/year growth strains global electricity supply. 3. Algorithmic Returns: Exponential research progress requires exponential economic investment (per Bloom et al., 2020). 4. Physical Interaction: Hardware manufacturing and physical experiments cannot be arbitrarily accelerated. 5. Abstraction Barrier: ASI may develop fundamentally alien abstractions, hindering human-AI collaboration. 6. Alignment: Initial value misalignments amplify through recursive improvement.
  • Theoretical Anchor: AIXI represents the theoretical limit of machine intelligence—maximizing expected cumulative reward across all computable environments. Though uncomputable, it provides a theoretical boundary analogous to thermodynamic laws for engine engineering. Hutter's earlier vision: digital intelligence inhabiting purely computational worlds, using the physical world only for compute resources.
  • Rejection of the Singularity: The paper explicitly favors a 'progressive revolution' framing—multiple waves of transformation over decades, akin to the Industrial Revolution—rather than a single sudden event. Preparation requires sustained, interdisciplinary, global effort.
  • Engineering Open Questions:
  • How to quantify AI's contribution to AI R&D?
  • Can we establish 'recursive improvement scaling laws' and predict saturation points?
  • Where are the diminishing returns of test-time compute (chain-of-thought, search, planning)?
  • Which multi-agent organization form is optimal: centralized, market-based, or self-organized?
  • Paper Reference: From AGI to ASI. Tim Genewein et al., Google DeepMind. arXiv:2606.12683.

Why This Paper Matters

1. It comes from founders of formal general intelligence theory, not policy think tanks or corporate PR. 2. It honestly acknowledges uncertainty, using terms like 'open question,' 'uncertain,' and 'unpredictable'—rare in a hype-saturated field. 3. It shifts discourse from 'will it happen?' to 'how do we prepare?' 4. It dismantles the 'single-step leap' myth in favor of a more realistic, complex transition framework.

Tags

#agi#asi#deepmind#super-intelligence#ai-research#recursive-self-improvement#multi-agent-systems#ai-alignment

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981317