playing AI improving its strategy to maximize scores or a robot practicing grasping an object to increase its success rate. AlphaGo, famous for its matches against Lee Sedol, is also a reinforcement learning system. In generative AI, reinforcement learning can be utilized to improve the quality of the generated output. For instance, it is applied to train a chatbot to generate more natural answers or avoid producing hate speech by using user feedback as a reward signal, or to adjust an image generation model to create results that meet specific style criteria. In addition to mimicking patterns in the training data, the model’s generation capability is optimized according to external evaluation standards. The task of fine-tuning a generative AI system in this manner is called RLHF (Reinforcement Learning from Human Feedback). Transformer A transformer is a type of architecture used to design and implement artificial neural networks. Before transformers were introduced, most large language models (LLMs) processed training data sequentially. For example, given the sentence “The sovereignty of the Republic of Korea shall reside in the people,” the model would receive the input in order — “The → sovereignty → of → …” — to capture the relationships between preceding and following words. One limitation of these earlier approaches is that they struggle to capture long-term dependencies (relationships between tokens that are far apart). To put it simply: the link between “The” and 18 19