AI

A wiki document summarizing the concept, core technologies, industrial applications, and latest trends of AI (artificial intelligence) in 2024–2025.

AI

Overview

AI (Artificial Intelligence, 인공지능) is a technological field and academic discipline that seeks to implement human intellectual abilities such as learning, reasoning, perception, and language understanding in computer systems. Since the term was formalized at the 1956 Dartmouth Conference, it has undergone several periods of stagnation and revival, and through the rise of deep learning in the 2010s and the spread of generative AI and large language models (LLMs) in the 2020s, it has established itself as a general-purpose technology across industries. Today, AI is being incorporated as a foundational layer of almost all digital services, including search, translation, medical diagnosis, autonomous driving, content creation, and software development.

Main Content

Definition and Scope

AI is discussed by dividing it largely into three stages. First, Narrow AI, which performs only specific tasks, includes most systems currently commercialized. Second, Artificial General Intelligence (AGI), capable of general problem-solving on par with humans, has not yet been realized. Third, Artificial Superintelligence (ASI), surpassing humans, is mainly the subject of theoretical and policy discourse. Academically, it is divided into traditional AI based on symbolic logic and data-based machine learning, and the recent mainstream is the latter.

Core Technologies

Machine Learning: A learning paradigm that infers rules from data on its own. It is divided into supervised learning, unsupervised learning, and reinforcement learning.
Deep Learning: Uses multi-layer artificial neural networks to learn high-dimensional patterns in images, speech, and language.
Transformer and Attention: A structure proposed in 2017 that opened the era of LLMs by processing context dependencies in parallel.
Generative AI: A group of models that newly generate text, images, music, video, and code, with diffusion models and autoregressive models as representative examples.
Multimodal and Agents: Evolving toward understanding text, images, and speech together and autonomously performing tasks through tool invocation and planning.
Training Infrastructure: Accelerators such as GPUs and TPUs, distributed training, data pipelines, and model lightweighting technologies (quantization and distillation) advance together.

History of Development

It went through the proposal of the Turing test in 1950, the Dartmouth Conference in 1956, expert systems in the 1960s–70s, the second AI boom in the 1980s, and the spread of statistical machine learning in the 1990s–2000s. In 2012, deep learning showed overwhelming performance at the ImageNet image recognition competition, becoming a turning point; after AlphaGo’s Go match in 2016, the announcement of the Transformer in 2017, and the popularization of conversational chatbots at the end of 2022, multimodal and reasoning-specialized models and small language models (SLMs) spread simultaneously in 2023–2025.

Industrial Applications

In manufacturing, predictive maintenance and vision inspection; in finance, credit scoring and anomalous transaction detection; in healthcare, image reading assistance and drug candidate discovery; in retail, demand forecasting and personalized recommendations have become standardized. In software development, code generation and automated review have become everyday tools, and in the public, legal, and education sectors, document summarization and retrieval-augmented generation (RAG)-based question answering are being rapidly adopted.

Social Impact and Regulation

Along with productivity improvements, job restructuring, copyright infringement, deepfakes and disinformation, bias and discrimination, energy consumption, and privacy infringement have emerged as major issues. Accordingly, the EU introduced a risk-based regulatory system through the AI Act, the United States combines executive orders and industry self-regulatory norms, and Korea shows a trend of pursuing discussions on a Basic Act on AI and guidelines in parallel. At the corporate level, safeguards such as responsible AI principles, model cards, red-team evaluations, and watermarking are becoming standard practice.

Latest Trends

The key keywords of 2024–2025 are reasoning, agents, multimodality, and efficiency. First, reasoning-specialized models that perform step-by-step thinking beyond simple responses have emerged, greatly improving performance in math, coding, and science problems, and scaling strategies that increase compute at inference have become a new axis of competition. Second, as AI agents that directly operate browsers, email, and work tools are deployed in practice, the center of gravity has shifted from “AI that uses tools” to “AI that performs work.” Third, omni models that process text, images, speech, and video in a single model and real-time voice conversation have become common. Fourth, moving away from an exclusive focus on giant models, cost-efficiency strategies such as on-device and small models, open-weight models, retrieval-augmented generation, and fine-tuning and LoRA have spread. Fifth, with the surge in data center power demand, power, cooling, and semiconductor supply chains have emerged as key variables in AI competition, and countries are pursuing compute infrastructure and semiconductor self-reliance as national strategies. Finally, governance discussions such as mandatory labeling of AI-generated content, disclosure of training data transparency, and model evaluation and audit systems have entered the stage of actual legislation and standardization.

AI