Google DeepMind Releases Roadmap from AGI to ASI: Four Paths and Six Bottlenecks

Google DeepMind recently published a detailed report (arXiv:2606.12683) outlining the evolution path from Artificial General Intelligence (AGI) to Artificial Superintelligence (ASI) and potential bottlenecks. Co-authored by 14 experts including DeepMind co-founder Shane Legg and AIXI inventor Marcus Hutter, the paper explicitly treats AGI as a starting point, not an endpoint.

Core Definitions and Intelligence Levels

The report defines three levels of intelligence: AGI (reaches median human performance on most cognitive tasks), ASI (consistently surpasses the output of tens of thousands of top experts working continuously for a decade on almost all tasks), and Universal AI (the theoretical limit under the AIXI framework). ASI is not omniscient and remains constrained by physical laws such as the speed of light, thermodynamics, and computational complexity.

Inherent Advantages of Digital Intelligence

Compared to biological intelligence, digital intelligence has six major advantages: extremely fast input/output speeds, scalable internal processing speed, hardware substrate independence, lossless replication and experience sharing, and continuously accelerating computational growth. The report estimates that current effective AI compute grows by about 10x (one order of magnitude) annually; if sustained, this will fundamentally transform the nature of intelligence.

Four Parallel Evolution Paths

The report proposes four technical paths that may occur simultaneously:

• Scaling: Continue expanding compute, data, and model size. A thought experiment shows that if only 1,000 AGI instances run globally, with 10x annual growth, there could be 100 million in five years. A cluster of 100 million AGIs sharing a collective mind and thinking 100x faster could itself be considered ASI.
• Algorithmic Paradigm Shift: Break through the bottlenecks of the current Transformer architecture by developing linear-time architectures (e.g., Mamba), continual learning, or neuromorphic hardware.
• Recursive Self-Improvement: AI autonomously writes code, designs chips, and generates training data, enabling self-iteration that could trigger an "intelligence explosion."
• Multi-Agent Collaboration: Millions of AGIs form a digital ecosystem via high-bandwidth communication, collaborating through market mechanisms or swarm intelligence to emerge collective intelligence beyond individuals.

Six Real-World Bottlenecks

The report warns that the following bottlenecks could stall or even reverse progress:

• Data Wall: High-quality human text data is expected to be exhausted by the end of this decade, potentially causing model degradation.
• Resource Wall: Exponential growth in compute, electricity, and chips faces economic and natural resource limits.
• Paradigm Wall: The current pre-training + fine-tuning paradigm may hit a ceiling.
• Increasing Research Difficulty: Scientific breakthroughs require exponentially more effort as fields mature.
• Abstraction Barrier: AI relies on existing human conceptual frameworks and struggles to build entirely new cognitive structures from scratch.
• Social Resistance: Safety incidents, job displacement, and regulation could artificially slow progress.

Open Research Questions

The report calls for developing superhuman-level benchmarks, refining economic and compute models, studying the micro-mechanisms of multi-agent systems, and deeply exploring the theoretical foundations of superintelligence. Over the next 10–20 years, the probability of transitioning from AGI to ASI is high, and global interdisciplinary research is needed to address the transformation.