Google DeepMind has released a 57-page report titled "From AGI to ASI," co-authored by co-founder and Chief AGI Scientist Shane Legg with his PhD advisor Marcus Hutter plus a 14-person team. Uniquely, Chapter One directly instructs AI assistants—a first in academic publishing.
Three Intelligence Tiers: AGI, ASI, and UAI
The report defines three levels: AGI matches median human performance on most cognitive tasks; ASI must consistently exceed the output of tens of thousands of top experts collaborating on a single problem for ten years; Universal AI (UAI/AIXI) represents the theoretical ceiling of intelligence. ASI is merely a milestone toward UAI.
It seals a loophole: those experts can only use 2010's tech stack—the year DeepMind was founded—preventing humans from building ASI first.
Six Inherent Advantages of Digital Intelligence
DeepMind's analysis outlines six hardwired edges: I/O speed—LLMs swallow books in seconds; internal processing speed—accelerable by compute; substrate independence—seamless migration to better hardware; lossless replication and experience sharing—instantaneous cloning of millions of instances; plus frictionless high-dimensional mind communication and an automated R&D empire.
These advantages widen with more compute, leaving biological intelligence in the dust.
Four Pathways to ASI: Quantity Becomes Quality
The report proposes four possible concurrent pathways.
Path 1: Brute Force—scale compute, data, and model size. Thought experiment: if only 1,000 AGI instances run initially, at 10x growth yearly, 100 million could exist in five years. Such a swarm, with zero marginal cost deployment, shared memory, and millisecond cognitive synchronization, inevitably crosses into ASI territory.
Path 2: Paradigm Shift—current pre-training may hit a ceiling, requiring novel architectures like spiking neural networks or infinite working memory structures.
Path 3: Multi-Agent Collaboration—millions of AGI experts decompose complex problems via high-bandwidth communication, giving rise to superhuman swarm intelligence.
Path 4: Recursive Self-Improvement—AI modifies its own code and hardware or generates better training data through self-play, potentially sparking an intelligence explosion.
Six Great Walls That Could Lock Down the Future
But DeepMind warns of six bottlenecks that could halt progress.
Data Wall: high-quality human text may be exhausted by the decade's end; Resource Wall: exponential compute, power, and chip costs; Harder Research: low-hanging fruit gone; Paradigm Ceiling: next-token prediction may not reach ultimate intelligence; Human Decision: social backlash and regulation could pull the plug; Abstraction Barrier—the deepest philosophical challenge: can AI construct wholly new concepts from raw data? If not, each model is trapped within human cognitive limits.
Yet the report notes that even if every AI is blocked by abstraction, collective intelligence can brute-force through sheer instance count—a wall stops a genius but not 100 million ordinary people.
The conclusion is measured: for AI to stop at human level, multiple walls must simultaneously become dead ends—an unlikely coincidence. They bet on either stalling before AGI or a smooth ride from AGI to weak ASI. Our generation may witness the realization of AI's founding dream from the Dartmouth Conference 70 years ago.