Machine translation has been considered a key challenge in AI research for quite some time due to the complexity and uncertainties associated with natural language. Machine translation architecture has moved in parallel with other advances in computational simulation, from the implementation of rule-driven and statistical approaches to more contemporary architectures involving neural networks. These early systems employed hand-engineered linguistic rules and parallel corpora that severely constrained their generalizability over multiple languages. More recent architectures involving neural and attention networks have furthered representation learning through more successful modeling of contextual associations in language, although they remained limited within data availability and task-dependent training. Recent advances within self-supervised and multitask learning have radically transformed this landscape, thereby opening the door for the development of large language models that have been trained on mass corpora spanning multiple languages. These large language models have shown robust transfer capabilities over multiple languages, thereby also demonstrating their viability for simultaneous natural language understanding tasks including translation as part of an encompassing framework. This study also explores the limitations perceived within the various succeeding variants of machine translation systems that have driven innovation in architecture and approach, and analyzes the interplay between machine translation progress and the development of large language models. This study also delves into the degree to which large language models could complement or replace current machine translation systems, while also underscoring challenges remaining within their reliability within multiple corpora.