Over the Edge of Chaos? Excess Complexity as a Roadblock to Artificial General Intelligence

TL;DR

This paper questions the assumption that AI capabilities will unboundedly grow toward AGI by framing AI evolution as a complexity-driven, phase-transition process. It develops an agent-based modelling framework in which $n$ benchmarks define system performance $P_i(t)$ and total complexity $C(t)$, with a critical threshold $C_{max}$ marking a switch to volatile dynamics. A derivative-based criticality detector, optimized via stochastic gradient descent, identifies onset of instability, and the methodology is validated through 200 simulations across multiple benchmark counts ($2,5,10,20$). The findings show that higher benchmark counts can stabilize post-criticality dynamics and improve criticality detection, while smaller systems exhibit greater instability; the work also discusses the limitations of current LLMs, the dangers of overreliance on data-driven learning, and the need for richer benchmarks and non-linear modelling. Overall, the study offers a tempered perspective on AI growth, highlighting potential complexity ceilings and providing tools to monitor and manage critical transitions in AI systems, with practical implications for LLM benchmarking and robust AI evaluation frameworks.

Abstract

In this study, we explored the progression trajectories of artificial intelligence (AI) systems through the lens of complexity theory. We challenged the conventional linear and exponential projections of AI advancement toward Artificial General Intelligence (AGI) underpinned by transformer-based architectures, and posited the existence of critical points, akin to phase transitions in complex systems, where AI performance might plateau or regress into instability upon exceeding a critical complexity threshold. We employed agent-based modelling (ABM) to simulate hypothetical scenarios of AI systems' evolution under specific assumptions, using benchmark performance as a proxy for capability and complexity. Our simulations demonstrated how increasing the complexity of the AI system could exceed an upper criticality threshold, leading to unpredictable performance behaviours. Additionally, we developed a practical methodology for detecting these critical thresholds using simulation data and stochastic gradient descent to fine-tune detection thresholds. This research offers a novel perspective on AI advancement that has a particular relevance to Large Language Models (LLMs), emphasising the need for a tempered approach to extrapolating AI's growth potential and underscoring the importance of developing more robust and comprehensive AI performance benchmarks.

Figures (4)

• Figure 1: Aggregate performances across all simulations, indicating the criticality point to which all simulations have been aligned and an predetermined criticality threshold.
• Figure 2: Variance of all performances across all simulations, indicating the criticality point to which all simulations have been aligned and a predetermined criticality threshold.
• Figure 3: Variance of all performances across all simulations, indicating the criticality point to which all simulations have been aligned and a predetermined criticality threshold.
• Figure 4: Histograms depicting the distribution of time points at which criticality was detected on the test dataset with the shaded area indicating the beginning of the actual criticality and the end of the detection window period denoting correct classifications.