The Model-Infrastructure Co-Evolution in Conversational NLU: From ELIZA to Large Language Models

Abstract

Conversational Natural Language Understanding (NLU) has passed through five distinguishable technological eras: rule-based symbolic parsing, statistical and probabilistic modelling, neural sequence learning, Transformer-based pretraining, and the current large language model regime. Existing accounts of this history tend to foreground modelling advances and treat infrastructure as a supporting detail. This article argues instead that the two tracks are causally entangled: each scientific advance in conversational NLU became production-viable only once a matching infrastructure capability, spanning hardware, data pipelines, training paradigms, and serving architecture, made it deployable at scale, and each infrastructure capability in turn exposed the next scientific bottleneck. The analysis traces a specific bottleneck sequence, from combinatorial rule brittleness through representational discreteness, vanishing gradients, and sequential processing limits, to labelled-data scarcity, showing how each resolution reshaped the engineering problems that remained. A closing synthesis argues that large language models have not simplified conversational architecture; they have relocated its complexity from authoring rules and training classifiers toward orchestration, retrieval grounding, and failure isolation across distributed inference systems. The account draws on systems-architecture reasoning applied to the published research and technical-report record, rather than on a new empirical benchmark.