General Characteristics of the Large Language Models and Comparative Analysis

Abstract

The history of the development of large language models (LLMs) is examined in the context of their impact on educational processes. The analysis begins with the origins of automatic translation, initiated in 1947, and the first implementations of this idea, including IBM's machine translation experiments in 1954, the creation of the chatbot ELIZA in 1966, and the evolution of neural network architectures that led to modern LLMs. Significant attention is given to transformer-based models, starting with the emergence of GPT-1 in 2018, Google's BERT, and their subsequent evolution into models such as GPT-4.5, Claude 3.7, and Gemini 2.0. The emergence of multimodal LLMs, capable of integrating text, images, and other data types, is described in detail. Particular focus is placed on architectural innovations such as the mixture of experts (MoE), which enhance model efficiency, as well as the development of open models, including Meta's LLaMa and Chinese LLMs that have become competitive with Western technologies. Recent trends in text generation are also discussed, especially the concept of the large concept model (LCM) and the new dLLM architecture, which enables faster generation and refinement of textual data. Analyzing the current state of LLM development, it is noted that these models are already fundamentally transforming education, and their continued improvement opens up new possibilities for integration into the educational domain. The conclusion emphasizes the importance of responsible use of LLMs and the need to strike a balance between their efficiency and the potential challenges they may pose.