As multiple technological routes toward artificial general superintelligence (AGI) take shape, the regime beyond AGI is shifting from philosophical speculation to empirical research and governance questions. Google DeepMind researchers argue that the development of advanced AI could surpass human-level AGI to reach artificial general superintelligence (ASI), which consistently outperforms groups of human experts.

For decades, AI research has focused on the monumental milestone of building AGI, machines that match humans across nearly all cognitive tasks. The prospect of AGI was always acknowledged, but it seemed too distant and fraught with uncertainties to contemplate its aftermath. Now, as AGI approaches, the focus has shifted to the next frontier: ASI.

ASI, defined as an advanced system that outperforms large groups of human experts on tasks that could take them years to complete, represents a significant leap. ASI would need to match or surpass humans across virtually all tasks of interest, not just in narrow domains. Examples include making the same scientific progress as an entire research field in a fraction of the time or achieving technological advancements at a pace comparable to what large companies can do in a year.

Theoretical limits of general intelligence are well understood, but the path from human level AGI to ASI remains largely uncharted. Fundamental limits lie far beyond today’s AI systems and humans, but the vast space between them is unexplored territory. Even if progress at the AI frontier stalls, the continued efficiency improvements in hardware and software could lead to a dramatic shift in the number and quality of AI agents.

Over the last decade, AI progress has been fueled by steady growth in effective compute, driven by three compounding factors: increased data, larger models, and better algorithms. This growth has resulted in an overall increase in global effective compute of at least 10x per year, equivalent to a factor of 10 billion over ten years. If this trend continues, a decade from now, the number of AI agents that can be run would exceed the global population.

Continued hardware and software efficiency improvements could further reduce the cost of running AI workloads. For instance, a cognitive workload that currently costs $1 billion in compute tokens might only cost $1000 in ten years. Over the last few years, the cost of achieving a fixed score on benchmarks has fallen by more than 13x per year. This trend, if extrapolated, suggests a future where many cognitive tasks become commoditized, leading to cognitive industrialization.

While such a proliferation of intelligence is not guaranteed, practical ceilings on compute build out and speculative ceilings on algorithmic efficiency might allow the effective population of today’s AI agents to scale up by more than a thousand-fold by 2030 and over a billion fold in the next decade.

Differences between humans and digital intelligence, such as the direct benefit from more and faster compute, suggest that ASI could emerge even if AGI is achieved. The speed and scale of AI could bridge the qualitative gap from AGI to weak ASI.

AI could follow four potential technological pathways to unlock new cognitive capabilities:

1. Scaling Laws: Qualitative improvements in AI have consistently followed model and data scale. Recent research suggests that training large transformers could enable AI to run complex algorithms via their prompts, potentially bridging the gap to ASI.

2. Multi-Agent Systems: Grouping AGI instances into “group agents” or virtual economies could produce collective superintelligence. However, the scalability and coordination of these agents remain uncertain.

3. Recursive Self-Improvement (RSI): AI systems could automate improvements to future generations of AI, potentially leading to an accelerating rate of progress. RSI operates through improvements to code and designs, data, and group specialization.

RSI is not a single process but can operate through three mechanisms: genetic (code and design), cultural (data), and social evolution (group specialization through cooperation). Evaluation methods that can scale into the superhuman regime are crucial for RSI to advance rapidly.

In conclusion, the path to ASI is multifaceted and not mutually exclusive. These pathways, combined, could drive the development of advanced AI tools that significantly impact science, the economy, and society, potentially leading to AI automation tools that transform the way we work and live.

Source: https://institute.deepmind.com/essays/what-comes-after-agi/

Thinking about building an AI product?

Get in Touch