인공지능이란 무엇인가 (서울대 컴퓨터공학부 장병탁 교수)

Main Ideas, Concepts, and Lessons

Definition and Origin of AI

• The term “artificial intelligence” was first coined in 1956.
• Research leading to AI began earlier:
  • Foundations were laid around 1950.
  • Even earlier developments helped enable AI.
• Key historical contributors:
  • “Father of computers” figures asked what it would take for a computer to become a thinking machine.
  • George Boole (about 100 years earlier) developed a method for mathematical/logical operations using binary numbers, contributing to later computer development.
• The digital electronic computer emerged in the 1940s (around 1943) during World War II, and AI research began alongside computer invention.

Early Framing of Machine Intelligence

• In 1950, Alan Turing (referred to via “Cho-rang-i”) wrote “Computing Machinery and Intelligence”, proposing how to evaluate machine intelligence.
• Six years later, in 1956, the term “artificial intelligence” was coined.

Turing Test / “Imitation Game”

• The proposed evaluation method in 1950 is the Turing Test, originally called the “Imitation Game.”
• Core setup:
  • A man and a woman are placed in separate rooms behind a curtain.
  • A third party asks questions to determine who is who.
  • The man must imitate the woman, including lying (e.g., hair length).
• AI adaptation:
  • Replace the man with a machine.
  • If the evaluator cannot distinguish machine vs. human through questioning, the machine is considered to show high intelligence.
• Why it’s language-based:
  • Since you can’t identify who/what it is just by voice, evaluation focuses on language via a keyboard.
• Limitations mentioned:
  • There is not yet a perfect test (the subtitle references a “belching test,” but the meaning is unclear).
  • A key difficulty: machines struggle to fully understand natural-language questions.
  • Deep learning and other advances bring progress, but:
    • Machines still don’t think using language exactly like humans, and more research is needed.

Philosophical Critique: Chinese Room Controversy

• Around 1980, philosophers objected to the idea that machines can truly think like humans.
• A major argument discussed is the Chinese Room Controversy by philosopher John Searle.
• Thought experiment structure:
  • A person in a sealed room knows no Chinese.
  • They have a Chinese textbook/rules.
  • Chinese input arrives from outside.
  • Using pattern matching/rules, the person outputs appropriate Chinese characters/questions-to-answers.
• Implication:
  • Outsiders see correct Chinese in/Chinese out and assume understanding.
  • But inside, the person is not understanding—they are only following rules to map inputs to outputs.
• Ongoing debate:
  • Opposing arguments claim that while the individual may not understand, the system as a whole might.
  • The discussion continues.

Approaches to Building AI (Two Paradigms + Evolution Toward ML)

The speaker outlines two main methods/paradigms:

1. Symbolic AI / Rule-based approach

   • Represents language using symbols/logical language.
   • Uses logical/procedural rules to infer conclusions.
   • Performs algorithmic action steps based on inputs.
   • Relies on signals from the environment and learning how to process them.
   • A garbled subtitle phrase appears like “Giwo-juui (observation of] all things in the world),” but the intended meaning is unclear.

2. Connected AI

   • Learning and capability improvement through training (likely referring to connectionist/neural network-style ideas).
   • Often framed as gradually increasing capability through training.

Machine Learning (ML) vs AI

• The relationship discussed:
  • AI aims to create a “thinking machine” (intelligent reasoning/behavior).
  • Machine learning is one method to build such systems.
  • AI can exist without ML (e.g., rule-based systems).
• Recently, AI often uses ML heavily, leading some to treat them as the same.
• Strictly, they differ:
  • AI = goal/objective (an intelligent “thinking machine”)
  • ML = a technique toward that objective

Classification of ML Focus

• The speaker groups ML-related approaches into three types (as implied by subtitles).
• The final subtitle segment is heavily garbled.
• The likely intended idea is that different ML categories emphasize different aspects, such as:
  • improving learning/processing inputs
  • differing levels of emphasis on implementation vs. pursuing broader AI goals

---

Methodology / Instruction-like Elements

1) Turing Test / Imitation Game (Evaluation Procedure)

• Place a human in one room and an evaluator tries to determine identity.
• Originally:
  • A man imitates a woman.
  • A third party asks questions.
  • The evaluator decides which room contains the man vs. the woman.
• AI version:
  • Replace the man with a machine.
  • The evaluator asks questions through language only (via keyboard).
  • If the evaluator cannot reliably tell whether responses come from:
    • a human or a machine
  • then the machine is considered to demonstrate intelligence.

2) Chinese Room Mechanism (What’s happening vs. what it appears to be)

• Put a person in a room who cannot understand Chinese.
• Provide:
  • Chinese input arriving from outside
  • a rulebook/textbook mapping inputs to outputs (pattern matching)
• For each Chinese input:
  • The person looks up matching rules.
  • The person outputs the corresponding Chinese response.
• Outsiders observe correct input→output behavior and infer understanding.
• The claim is that inside, the person does not understand—only produces outputs via rules.

---

Speakers / Sources Featured

• Jang Byeong-tang — speaker (Professor, Department of Computer Science, Seoul National University)
• Alan Turing — author of “Computing Machinery and Intelligence”
• George Boole — historical figure referenced for logic/binary foundations
• John Searle — philosopher presenting the Chinese Room thought experiment
• “Father of computers” figures — mentioned generally (not named in the subtitles)