Chapter 2: AI Landscape & Model Platforms

After completing this chapter, you will: understand AI technology hierarchy relationships, master mainstream model platforms and aggregation gateways, understand local deployment options

2.1 AI Landscape Beginner Free

Prerequisites: None

Why Do You Need It? (Problem)

When you start getting into AI programming, you'll encounter a bunch of confusing abbreviations:

• "We use AI for intelligent customer service"
• "This project is based on ML algorithms"
• "Now everyone uses DL, traditional ML is outdated"
• "LLM is the hottest direction"
• "Generative AI has changed the entire industry"

What's the relationship between these terms? Are they different names for the same thing, or completely different technologies?

If you don't clarify these concept hierarchies, you'll face these problems:

1. Communication barriers: When someone says "our LLM system," you don't know what they're talking about
2. Selection confusion: Don't know which technology category your needs fall into
3. Learning confusion: Don't know whether to learn ML, DL, or go straight to LLM

What Is It? (Concept)

AI, ML, DL, and LLM are progressively nested relationships, like "fruit → citrus → orange → navel orange":

Key Understanding:

Level	Definition	Examples
AI (Artificial Intelligence)	All technologies that make machines simulate human intelligence	Game AI, expert systems, recommendation systems, chatbots
ML (Machine Learning)	Let machines automatically learn patterns from data, not hand-written rules	Spam filtering, house price prediction, user profiling
DL (Deep Learning)	Machine learning implemented with multi-layer neural networks	Image recognition, speech recognition, machine translation
Foundation Model	General-purpose models pre-trained on large-scale data, adaptable to multiple downstream tasks	GPT, BERT, CLIP, Stable Diffusion
LLM (Large Language Model)	Foundation models specialized for text processing (tens of billions to trillions of parameters)	GPT-5, Claude Opus 4.6, Gemini 3.1 Pro, DeepSeek-V3.2

Position of Generative AI:

Generative AI is not a technology level, but a capability label, referring to AI that can "generate" new content:

• Text generation: LLM (ChatGPT, Claude)
• Image generation: Diffusion Model (Stable Diffusion, Midjourney, DALL-E)
• Audio generation: TTS, music generation (Suno, Udio)
• Video generation: Sora, Runway, Pika

Major AI Subfields:

Subfield	English	Application Scenarios	Is it Generative AI
Natural Language Processing	NLP	Translation, Q&A, text classification	Partially (LLM)
Computer Vision	CV	Face recognition, object detection, image classification	Partially (image generation)
Speech Recognition	ASR	Speech-to-text	No
Text-to-Speech	TTS	Text-to-speech	Yes
Recommendation Systems	RecSys	E-commerce recommendations, video recommendations	No
Reinforcement Learning	RL	Game AI, robot control	No

Try It Yourself (Practice)

This section is purely conceptual, with no code demonstrations. But you can use these knowledge points to verify your understanding:

Quiz: Determine if the following statements are correct

1. "All ML is DL" ❌ (ML includes traditional algorithms, DL is just a subset of ML)
2. "All LLMs are foundation models" ✅ (LLMs are a type of foundation model)
3. "ChatGPT is a type of AI" ✅ (LLM → Foundation Model → DL → ML → AI)
4. "Generative AI can only generate text" ❌ (Can also generate images, audio, video, etc.)
5. "Traditional ML is outdated" ❌ (Still mainstream in tabular data, recommendation systems, etc.)

Fun Fact

GPT-5's "unified routing" architecture quietly routes simple questions to a cheaper/faster sub-model — OpenAI is secretly saving you money. It's like ordering a premium burger and the kitchen decides your simple question only needs a regular patty. You still get charged premium prices though... so maybe not THAT generous.

Knowledge Card: Why Are They Called "Large" Language Models?

"Large" refers to parameter count (the model's learning capacity):

Model	Parameters	Analogy
GPT-2 (2019)	1.5B (1.5 billion)	Elementary student
GPT-3 (2020)	175B (175 billion)	College student
GPT-4 (2023)	1.76T (1.76 trillion, rumored)	Expert
DeepSeek-V3.2 (2025)	671B (671 billion, 37B activated via MoE)	Graduate student
GPT-5 (2025)	Unified routing architecture (auto-routes to fast or deep reasoning model)	Expert team

Larger parameter counts usually mean stronger capabilities, but also more expensive and slower.

2025-2026 AI Model Landscape Shifts

Three major trends are reshaping the model landscape:

1. Rise of Reasoning Models: OpenAI o1/o3, DeepSeek-R1 introduced "thinking" capabilities — models that spend extra compute to reason step-by-step before answering, dramatically improving complex problem-solving
2. Open-Source Closing the Gap: DeepSeek-V3.2, Qwen 2.5, Llama 4 now rival proprietary models in many benchmarks, with DeepSeek offering API pricing 10-50x cheaper than OpenAI
3. Unified Routing Architecture: GPT-5 pioneered the concept of a "unified system" that automatically routes simple queries to a fast model and complex problems to a deep reasoning model — blurring the line between model tiers

Summary (Reflection)

• What Did This Solve: Clarified the hierarchy relationships of AI/ML/DL/LLM, understand generative AI is a capability label not a technology level
• What Remains Unsolved: Now know the concepts, but with so many model providers (OpenAI, Claude, Gemini...), how to choose? Next section answers this
• Key Points:
  1. AI ⊃ ML ⊃ DL ⊃ Foundation Model ⊃ LLM (progressively nested)
  2. LLM is currently the hottest AI direction, and the focus of this guide
  3. Generative AI is a capability that includes text, image, audio, and video generation

---

Last Updated: 2026-02-20