The Path to Ascension for Large Language Models: The Expansion of the Temporal Dimension

Abstract This paper observes the core limitations of current large language models (LLMs) in pursuing general artificial intelligence (AGI)-level capabilities (commonly referred to as "ascension"), and extends the discussion to the necessity of expanding the temporal dimension through phenomena in human cognition. We observe that the bottleneck in LLM development lies not in the physical expansion of context capacity, but in the absence and potential for expansion of the temporal dimension as an underlying logical variable. Drawing on phenomena related to education level, linguistic complexity (such as the error-correction mechanism in French tenses), and historical perspective, we propose the role of temporal dimension expansion in enhancing cognitive depth and decision consistency. We further observe that understanding the three-dimensional physical world (spatial dimension) is important, but the temporal dimension is actually more important, as the former deals with static structures while the latter involves dynamic causality and long-term evolution. From an economic perspective, we structure these phenomena: systems lacking the temporal dimension are prone to resource allocation biases, logical drift, and systemic collapse, whereas effective expansion of the temporal dimension can optimize intertemporal decisions and productivity. We further observe, through long-term cross-month dialogue experiments, that models such as Grok, Gemini, and GPT universally exhibit "answer contamination" and severe accuracy degradation as time elongates. From a behavioral economics perspective, these phenomena resemble present bias and anchoring effects, leading to short-termism that amplifies macroeconomic imbalances. These are merely repeatable observations, requiring no empirical verification or model construction. We only conduct phenomenological observations, with no obligation to refine models or perform data regressions. This paper is intended for submission to SSRN as a preprint, not for title evaluation or salary purposes. Keywords Large Language Models; Temporal Dimension; Artificial General Intelligence; Temporal Awareness; Behavioral Economics; Intertemporal Choice; Present Bias; Linguistic Constraints; Historical Perspective; Systemic Risk JEL Classification D91 (Role of Economics; Role of Economists; Intertemporal Choice and Growth – Intertemporal Household Choice; Life Cycle Models and Saving) G41 (Behavioral Finance – Present Bias and Overconfidence) O33 (Technological Change: Choices and Consequences; Diffusion Processes – Technological Change and Economic Growth) C45 (Neural Networks and Related Topics – Machine Learning in Economics) Z13 (Economic Sociology; Economic Anthropology; Language; Social and Economic Stratification – Cognitive Economics and Time Perception)