The Genealogy of Large Language Models: From Auxiliary Tools in ASR to Foundational Transformers and Back Again
This paper traces the evolutionary trajectory of Large Language Models (LLMs), arguing that their origins lie in the practical need to correct transcription errors in Automatic Speech Recognition (ASR) systems. We delineate this development, starting with domain-specific grammars, progressing through statistical n-gram models, and then to Artificial Neural Network-based models (ANNs), specifically RNNs, LSTMs, and GRUs, until reaching the pivotal breakthrough of the Transformer architecture. This evolution, driven by the pursuit of better language modeling, enabled the massive scaling that defines modern LLMs, which exhibit unprecedented capabilities. We conclude that LLMs, which emerged as an auxiliary component to mitigate the deficiencies of ASR systems, have "closed the circle" by becoming the foundational technology that now redefines the state of the art in their progenitor systems, thereby establishing themselves as a unifying technology for Artificial Intelligence.