Generative AI
Generative AI
Generative Artificial Intelligence & LLM
1. Overview of Generative AI, an Artificial Intelligence That Creates and Communicates
flowchart LR
A["AI focused on classification/prediction"] -- "Shift to creative content generation and communication" --> B["Generative AI"]
Definition: An artificial intelligence technology that goes beyond simply classifying or predicting from learned data, and instead generates new content — text, images, audio, and code — on its own.
Characteristics: (LLM-centered) Built on large language models (LLMs) to generate a wide range of content, including text, code, and images. (Prompt engineering) Output quality varies greatly with input style, making prompt engineering a core skill. (Multimodal) Multimodal capability integrating text, image, voice, and video is rapidly expanding the range of applications.
2. Core Technology and Architecture of Generative AI
A. Evolution Built on the Transformer Architecture
flowchart TD
DATA["Large-scale structured/unstructured data"]
subgraph PRE["Pre-training"]
TRANS["Transformer Model<br/>(Self-Attention mechanism)"]
end
subgraph FINE["Fine-tuning"]
INST["Instruction Tuning"]
RLHF["Reinforcement Learning from<br/>Human Feedback (RLHF)"]
end
OUT["Generated content<br/>(Text, Image, Code)"]
DATA --> PRE --> FINE --> OUT
style PRE fill:#E3F2FD,stroke:#1976D2
style FINE fill:#FFF3E0,stroke:#F57C00
| Core Technology | Description | Role |
|---|---|---|
| Self-Attention | Captures the relational meaning between words in a sentence | Understands context, resolves long-range dependencies |
| RLHF | Trains using human preference as a reward function | Aligns AI responses with human values |
| Prompt Engineering | Optimizes the instructions given to the AI | Controls the quality and accuracy of generated output |
| RAG | Generates answers by referencing an external knowledge base | Prevents hallucination, provides up-to-date information |
B. Major Generative AI Models and Service Types
flowchart TD
GenerativeAI["Generative AI"] --> NLP["Text (NLP)"]
NLP --> GPT4ClaudeGeminiLlama["GPT-4, Claude, Gemini, Llama"]
GenerativeAI --> Vision["Image (Vision)"]
Vision --> DALLEMidjourneyStableDiffusion["DALL-E, Midjourney, Stable Diffusion"]
GenerativeAI --> Coding["Code (Coding)"]
Coding --> GitHubCopilotCursor["GitHub Copilot, Cursor"]
GenerativeAI --> Multimodal["Multimodal"]
Multimodal --> Node6["Integrated understanding and generation across text, image, and audio"]
| Category | Key Technology Model | Representative Service |
|---|---|---|
| Closed Source | Undisclosed weights and architecture | ChatGPT, Claude, Gemini |
| Open Source | Community-shared, deployable locally | Llama 3, Mistral, Gemma |
| SLM | Small models specialized for a specific domain | On-device AI, proprietary enterprise models |
3. Expected Benefits and Risk-Response Strategies for Generative AI Adoption
| Category | Key Expected Benefits | Risk and Response Strategy |
|---|---|---|
| Productivity innovation | Automates repetitive tasks and content creation | Prevent copyright infringement, ensure transparency (labeling AI-generated content) |
| Personalized experience | Delivers tailored education, healthcare, and service consultation | Prevent data privacy and personal-data leaks (apply DLP) |
| Business innovation | Enables new service models and creative problem-solving | Verify hallucination, adopt RAG |
| Ethical use | Expands and democratizes access to knowledge | Remove bias, establish AI ethics guidelines |