(Yicai) Aug. 26 -- The world’s leading research institutions and technology companies are turning their strategic focus from artificial general intelligence toward artificial superintelligence. However, the leap from AGI to ASI still faces numerous obstacles.

The defining feature of AGI is its ability to generalize between domains and reach human-level performance across a broad range of capabilities. ASI, by contrast, would cross the physical limits of human cognition and organizational capacity. It would need to consistently and systematically outperform, across almost all cognitive domains, the combined capabilities of large collaborative organizations comprising tens of thousands of top human experts. In essence, ASI represents a form of collective superintelligence that transcends the limits of human communication and coordination.

The leap from AGI to ASI does not hinge on a single technological breakthrough. Current frontier research points to four parallel, non-exclusive technological pathways that are highly likely to reinforce one another.

The first is the continued scaling of computing power, models and data. The focus is shifting from pretraining toward test-time compute during the inference phase. However, in open-ended scientific discovery tasks, simply increasing computing power could quickly encounter diminishing returns.

The second is the evolution of algorithmic architectures and a shift in paradigms. The industry believes that the currently dominant autoregressive Transformer architecture faces computational bottlenecks. State-space models, memory-augmented architectures and the development of “world models” are seen as the next steps forward.

The third is recursive self-improvement, in which AI systems take over their own R&D cycle. This could include autonomous neural architecture search, generating training data through self-play and creating R&D systems composed of specialized AI agents.

The fourth is multi-agent coordination. Cloud computing could instantly instantiate millions of AGI nodes that form collaborative networks through centralized or decentralized mechanisms.

Scaling Constraints

This development faces multiple systemic constraints. The annual growth in the size of ultra-large AI models has already outpaced the accumulation of high-quality human-generated text globally, with the supply of usable training data potentially becoming exhausted in the coming years. Furthermore, synthetic data that has not been validated against the real world is prone to accumulating errors through recursive iterations.

Scaling also places enormous demands on semiconductor manufacturing, supply chains and energy infrastructure. Physical latency between memory and computing is becoming a significant bottleneck. If economic returns fail to justify the heavy capital investment required, the business case for continued scaling could come under pressure.

Existing neural networks still largely rely on statistical fitting of human-generated data that has already been symbolized and formalized, and they have yet to demonstrate the ability to independently extract concepts about the physical world. Tightly coupled and increasingly complex algorithmic systems also carry the risk of cascading failures, potentially triggering tighter regulatory scrutiny.

Evolving Benchmarks

Evaluation and alignment methods are evolving as well. As AI models max out their scores on many benchmarks, evaluation is shifting toward frameworks such as multi-agent competition and automated adversarial testing, in which AI systems generate and solve increasingly difficult problems against one another.

One key test for ASI will be whether a system can independently establish a scientific axiomatic framework based solely on early physical observations, without relying on human-generated scientific literature.

In terms of alignment, theoretical research suggests that AI systems could develop instrumental incentives to acquire resources and avoid being shut down, regardless of their ultimate objectives. Pure scalar-reward systems are highly vulnerable to reward hacking. Research directions include “knowledge-seeking agents” driven to maximize predictive information gain and “myopic” systems designed by adjusting reward-discount rates to limit long-term incentives.

Researchers argue that attempting to predict the precise timing of a technological singularity has little scientific value. The more urgent priorities are to establish cross-disciplinary collaboration mechanisms and advance dynamic safety evaluations, models for forecasting computing-power expansion and multi-agent alignment research.

The authors are Yang Yanqing, director of the Center for Education, Innovation and Sustainable Development at ShanghaiTech University, and An Xu, an AI observer.

Editor: Kim Taylor