Researchers at Google DeepMind have published a paper that outlines how humanity might move from a world with artificial general intelligence (AGI) to one that contains artificial superintelligence (ASI). The work is timely: with contemporary large language models (LLMs) it is evident that the field is ‘‘in the ballpark’’ of AGI, and in the coming years we may transition to building systems that DeepMind defines as ASI.
DeepMind defines ASI as “a system that exceeds the performance of large human-expert collectives on virtually all tasks and domains of human activity.” The authors note that a single ASI could itself be a collective of millions of instances interacting with the world in parallel, analogous to how today’s LLMs are often deployed.
Why ASI might be feasible
One way to understand ASI is that a powerful AI system could also leverage the advantages digital intelligences have over biological ones: faster input/output and internal processing speeds, larger working memory and memorization, substrate independence, lossless replication, and high-bandwidth sharing of learning experiences.
Pathways and bottlenecks to ASI
The paper discusses four substantive pathways that could lead from AGI to ASI and the constraints each faces:
-
Scaling compute, models, and data:
- Simply scaling up current approaches could suffice to reach ASI. However, this requires ever-more compute and data, which may run into limits in energy and data supply. While historical trends favor the continued efficacy of scaling, it is unpredictable which specific capabilities will emerge or whether scaling will eventually suffer diminishing returns.
-
Algorithmic paradigm shift:
- Just as Transformer and Mixture-of-Experts architectures propelled progress years ago, other fundamental innovations could cause further leaps. Potential advances might include adaptive computation at test time or methods that overcome today’s context-window limitations. Such breakthroughs could be transformative but are inherently hard to anticipate — the authors compare predicting them to trying to foresee expansions in our understanding of reality before the invention of general relativity.
-
Recursive self-improvement (RSI):
- AI systems might be able to design their own successor systems. If so, a rapid transition from AGI to ASI could occur. The paper highlights wildcards: current AI accelerates human researchers (a kind of ‘‘co-creation RSI’’ is underway), but present systems do not yet display the paradigm-shifting creativity that seems required for major frontier advances. Even without such high-bar creativity, systems might iteratively produce marginally better versions of themselves, creating a slow compounding process. Outcomes could range from explosive capability growth to tapering off, or anything in between.
-
ASI via group agent formation:
- Many general intelligences could coordinate into complex structures whose aggregate capability exceeds the sum of their parts, similar to how human institutions accomplish feats (such as building space stations) beyond individual abilities. As with the other routes, emergence in multi-agent systems is difficult to reason about or predict.
Why this matters
The authors argue that taking seemingly impossible futures seriously is necessary in order to prepare for them. Just as the goal of AGI once seemed fanciful yet was pursued and ultimately yielded progress, ASI today should be considered a real possibility. They write: “Instead of focusing on one technological trajectory and timeline, being prepared for a post-AGI world requires considering a diverse set of forecasts and scenarios, paired with continual benchmarking and monitoring to update the set of forecasts and scenarios and their relative plausibility.”
DeepMind concludes that the possibility of ‘‘cruising past AGI and into ASI territory within the next decade or two’’ cannot be easily dismissed, and therefore broad monitoring and preparation across multiple scenarios is warranted.
Summary
The DeepMind paper identifies four main mechanisms by which AGI could become ASI — large-scale scaling, algorithmic breakthroughs, recursive self-improvement, and coordinated multi-agent structures — and emphasizes substantial uncertainty about which path (if any) will dominate. Given potential resource limits and unpredictable emergent behaviors, the researchers recommend continuous benchmarking, monitoring, and preparation for a range of timelines and outcomes.
