Skip to content

Overview of Artificial Intelligence

Introduction

Artificial Intelligence (AI) is a branch of computer science dedicated to creating systems capable of simulating intelligent human behavior. Since its formal inception in 1956, AI has undergone multiple cycles of boom and bust, and today it has profoundly transformed human society.

1. Definitions of AI

Different scholars define AI from various perspectives:

Dimension Thinking Humanly Thinking Rationally
Thinking Cognitive modeling: simulating human thought processes Laws of thought: logic-based reasoning
Acting Turing test: behaviorally indistinguishable Rational agent: making optimal decisions

Key definitions:

  • Turing (1950): if a machine cannot be distinguished from a human in conversation, it possesses intelligence
  • McCarthy (1956): the science and engineering of making intelligent machines
  • Russell & Norvig: the study of agents that perceive their environment and take actions to maximize success
  • Modern view: computational systems that learn patterns from data and make decisions in complex environments

2. Three Schools of AI

graph TD
    A[Artificial Intelligence] --> B[Symbolism]
    A --> C[Connectionism]
    A --> D[Behaviorism]

    B --> B1[Knowledge Representation & Reasoning]
    B --> B2[Expert Systems]
    B --> B3[Logic Programming]

    C --> C1[Neural Networks]
    C --> C2[Deep Learning]
    C --> C3[Large Language Models]

    D --> D1[Reinforcement Learning]
    D --> D2[Evolutionary Algorithms]
    D --> D3[Robotics]

    style B fill:#f9f,stroke:#333
    style C fill:#bbf,stroke:#333
    style D fill:#bfb,stroke:#333

2.1 Symbolism

  • Core idea: Intelligence = symbol manipulation. Represent knowledge with formal symbols, reason with logical rules
  • Representative methods: expert systems, knowledge graphs, logic programming (Prolog)
  • Strengths: explainable, verifiable, knowledge can be encoded
  • Limitations: knowledge acquisition bottleneck, difficulty handling uncertainty and perceptual tasks

2.2 Connectionism

  • Core idea: Intelligence emerges from the connections of many simple units. Simulates neural networks
  • Representative methods: artificial neural networks, deep learning, Transformers
  • Strengths: automatically learns from data, excels at perceptual tasks
  • Limitations: requires large amounts of data, lacks explainability, high computational cost

2.3 Behaviorism

  • Core idea: Agents learn behavior through interaction with the environment. No internal knowledge representation needed
  • Representative methods: reinforcement learning, evolutionary algorithms, swarm intelligence
  • Strengths: suited for decision and control problems, does not require labeled data
  • Limitations: low sample efficiency, exploration difficulties

Extension: Domingos's Five Tribes of ML　In The Master Algorithm (2015), Pedro Domingos cuts learning algorithms — from inside ML — into five tribes: Symbolists / Connectionists / Bayesians / Evolutionaries / Analogizers. This is orthogonal and complementary to the three-school taxonomy above: the three schools cut by "where does intelligence come from", the five tribes cut by "what mathematical tool does the learner use". See the The Master Algorithm notebook.

3. The Turing Test

3.1 The Original Turing Test (1950)

┌──────────┐     Text conversation    ┌──────────┐
│  Human    │ ◄───────────────────→   │  Subject  │
│  Judge    │                         │(Human/    │
└──────────┘                          │ Machine)  │
                                      └──────────┘

Rule: if the judge cannot reliably distinguish whether the subject 
      is human or machine, the machine is considered to have passed 
      the Turing test.

3.2 Limitations of the Turing Test

Criticism Argument
Tests only imitation Imitating humans does not equal possessing intelligence
Text-only Ignores perception, motor skills, and other intelligence dimensions
Human bias Judge expectations influence results
Deception possible ELIZA (1966) could "deceive" some users

4. The Chinese Room Argument

A thought experiment proposed by Searle (1980):

Outside the room: Chinese input → [Room] → Chinese output

Inside the room:
  A person who does not understand Chinese
  + A detailed set of Chinese processing rule books
  → Mechanically manipulates symbols according to rules
  → Produces "correct" Chinese responses

Conclusion: although the system appears to "understand" Chinese
      from the outside, the person inside does not understand 
      the meaning of Chinese.
      → Syntactic processing ≠ Semantic understanding

Significance: the fundamental distinction between Strong AI (machines truly understand and think) and Weak AI (machines simulate intelligent behavior).

5. Weak AI vs Strong AI

Dimension Weak AI (Narrow AI) Strong AI (AGI)
Capability scope Specific tasks General intelligence
Understanding No true understanding Possesses understanding and consciousness
Current status Already achieved Not yet achieved
Examples AlphaGo, ChatGPT, self-driving AI in science fiction

Superintelligence

  • Definition: intelligence that far surpasses the smartest humans in virtually all domains
  • Bostrom's view: once AGI is achieved, it may rapidly recursively self-improve to superintelligence
  • Debate: Is it possible? Is it dangerous? What is the timeline?

6. The Current AI Landscape

graph LR
    subgraph Perception
        A1[Computer Vision]
        A2[Speech Recognition]
        A3[Natural Language Understanding]
    end

    subgraph Reasoning & Decision-Making
        B1[Knowledge Graphs]
        B2[Planning & Search]
        B3[Reinforcement Learning]
    end

    subgraph Generation
        C1[Text Generation LLM]
        C2[Image Generation Diffusion]
        C3[Code Generation Copilot]
    end

    subgraph Applications
        D1[Autonomous Driving]
        D2[Medical Diagnosis]
        D3[Scientific Discovery]
        D4[AI Agents]
    end

    A1 & A2 & A3 --> B1 & B2 & B3
    B1 & B2 & B3 --> C1 & C2 & C3
    C1 & C2 & C3 --> D1 & D2 & D3 & D4
  1. Large Language Models (LLMs): GPT-4, Claude, Gemini demonstrate general language capabilities
  2. Multimodal: unified processing of text, images, audio, and video
  3. AI Agents: LLM-driven autonomous agents that can use tools and plan
  4. Scientific AI: AlphaFold (protein structure), GNoME (materials discovery)
  5. Embodied Intelligence: combining AI with physical-world interaction

7. AI Classification System

Classification Dimension Categories
By learning method Supervised, unsupervised, reinforcement, self-supervised
By task type Classification, regression, generation, decision, reasoning
By modality Text, vision, speech, multimodal
By application domain NLP, CV, robotics, recommender systems, scientific computing
By intelligence level Narrow AI → AGI → ASI

References

  • "Artificial Intelligence: A Modern Approach" - Russell & Norvig
  • "Computing Machinery and Intelligence" - Alan Turing (1950)
  • "Minds, Brains, and Programs" - John Searle (1980)
  • "Superintelligence" - Nick Bostrom

评论 #