China's AGI Summit: the ultimate question

"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."

At China's premier computer science gathering, a fundamental debate unfolded: can today's Large Language Models—the engines behind ChatGPT and its rivals—ever evolve into true artificial general intelligence, or do they represent a path that ultimately leads to a dead end?

The Debate Takes Centre Stage

The question is no longer confined to academic journals. At the forefront of China's computer science community, held in Harbin, a critical and highly consequential debate took centre stage: whether Large Language Models (LLMs) can truly scale to achieve Artificial General Intelligence, or whether their fundamental architecture is a dead end.

The core of the critique is mathematical. A paper presented at the conference formalised recursive self-training as a dynamical system and proved that as training data becomes increasingly self-generated, the system undergoes inevitable degenerative dynamics. This is not speculative—it is a formal proof that under the condition of a diminishing supply of fresh, authentic data, the system is guaranteed to converge to a fixed point that is a distorted and impoverished version of the true data distribution.

Two fundamental failure modes are identified: Entropy Decay, where finite sampling effects cause a monotonic loss of distributional diversity (mode collapse), and Variance Amplification, where the loss of external grounding causes the model's representation of truth to drift as a random walk, bounded only by the support diameter. These behaviours are not contingent on architecture—they are consequences of distributional learning on finite samples.

The Analytic Engine Limitation

The argument extends beyond mathematics to the philosophy of knowledge. The paper contends that current GenAI is fundamentally an "analytic engine"—it excels at analysing, recombining, and interpolating the vast patterns contained within its human-generated training data. It cannot generate synthetic knowledge: truly novel concepts, laws, or truths that are not simply derivative of its input.

This draws upon Immanuel Kant’s distinction between analytic and synthetic judgements. The Singularity requires a capacity for synthetic knowledge generation, which is absent in the current paradigm. An LLM can describe a universe it has never seen, but it cannot derive a new law of physics from first principles. It can tell a story about a human emotion, but it cannot feel one.

As one analysis noted, LLMs lack the ability to generate synthetic knowledge—truly novel concepts, laws, or truths that are not simply derivative of their input. The Singularity, if it is to occur, requires this capacity.

The Data Bottleneck

The mathematical limits are compounded by a practical constraint: data. The availability of high-quality human-generated public text data may become a bottleneck for LLM scaling as early as 2026, and almost certainly by the early 2030s. The internet, for all its vastness, is a finite repository of human expression. Once it is exhausted, the only path forward is synthetic data—which, as the mathematical proof demonstrates, leads to collapse.

A taxonomy of LLMs emerging from the debate categorises models along three dimensions: scalability, application domains, and ethical considerations. The scalability dimension directly addresses the question of whether LLMs can continue to improve indefinitely or whether they will hit fundamental walls.

The Path Forward: Hybrid Approaches

The debate is not purely pessimistic. The paper proposes a path forward involving symbolic regression and program synthesis guided by Algorithmic Probability. The Coding Theorem Method allows for identifying generative mechanisms rather than mere correlations, escaping the data-processing inequality that binds standard statistical learning.

The conclusion is that while purely distributional learning leads to model collapse, hybrid neuro-symbolic approaches offer a coherent framework for sustained self-improvement. These approaches combine the pattern-recognition capabilities of neural networks with the reasoning and verification capabilities of symbolic systems.

The GFN Context

For Global Future Nexus, the debate at China's premier computer science conference carries profound implications for the governance of AGI. If LLMs cannot scale to AGI—if their architecture is fundamentally limited—then the concentration of investment and attention on a single technological pathway may be a strategic error.

The debate also highlights the importance of architectural diversity. As the paper argues, hybrid neuro-symbolic approaches offer a coherent framework for sustained self-improvement. GFN's mission of building a thriving planetary ecosystem where human societies, advanced AGI, and sustainable systems coexist must account for the possibility that the path to AGI is not singular—and that the governance frameworks we build must be flexible enough to accommodate different paradigms, whether they emerge from scaling, structure, or synthesis.

The ultimate question posed at China's premier computer science conference remains unanswered. But the debate has sharpened the question: will AGI emerge from the continued scaling of LLMs, or will it require a fundamental shift in architecture? The answer will shape not just the future of AI, but the future of governance itself.

Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)

Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)