AI Basics: Definitions, History, Agents, Tasks & Limits

TAKEAWAYS:

• AI is the study of creating machines that mimic human intelligence through problem-solving, learning, and decision-making.
• Intelligent agents (e.g., chatbots, recommendation systems) perceive environments and act to achieve goals.
• AI tasks include searching, reasoning, learning, and perception, each with unique methods (e.g., game trees for chess, neural networks for image recognition).
• Limitations like the Turing Test’s flaws and AI’s lack of true consciousness define its boundaries.
• Real-world applications span from eSewa’s fraud detection to Google’s AlphaGo using game trees.
• Exam focus: Define AI, compare weak vs. strong AI, and explain agents’ components (perception, action, goals).

1. What is Artificial Intelligence?

1.1 Definition and Scope

Artificial Intelligence (AI) is the branch of computer science that aims to create systems capable of performing tasks that typically require human intelligence, such as:

• Reasoning (e.g., solving puzzles like Sudoku).
• Learning (e.g., Netflix’s recommendation system).
• Perception (e.g., self-driving cars detecting pedestrians).
• Problem-solving (e.g., Daraz’s inventory optimization).

Key Idea: AI ≠ Human Intelligence. It mimics specific behaviors (e.g., playing chess) but lacks consciousness or general intelligence.

1.2 The Turing Test: Can Machines Think?

Proposed by Alan Turing (1950), the Turing Test checks if a machine can fool a human into believing it’s another human in a text-based conversation.

• Passing the Test: If a judge cannot distinguish the machine from a human after 5 minutes, the machine is considered "intelligent."
• Limitations:
  • Only tests natural language understanding, not other AI capabilities (e.g., vision, reasoning).
  • No true consciousness: Machines lack self-awareness or emotions.
  • Cheating: Machines can exploit loopholes (e.g., memorizing responses).

1.3 Weak AI vs. Strong AI

2. Intelligent Agents: How AI Systems Work

An intelligent agent is an entity that percews its environment and acts to achieve goals. It has:

1. Perceptors: Sensors (e.g., cameras, microphones) to gather data.
2. Effectors: Actuators (e.g., robot arms, speakers) to perform actions.
3. Goals: Objectives to maximize (e.g., maximize profit for Daraz, minimize travel time for Pathao).

2.1 Agent Types and Examples

2.2 The Agent-Environment Interface

flowchart LR
    A["Environment"] -->|"Percepts"| B["Agent"]
    B -->|"Actions"| A
    B --> C["Performance Measure"]
    C -->|"Feedback"| B

• Percepts: Inputs from the environment (e.g., user queries for a chatbot).
• Actions: Outputs the agent performs (e.g., sending a response).
• Performance Measure: Evaluates success (e.g., user satisfaction score).

Worked Example: Daraz’s Order Fulfillment Agent

1. Percepts: User places an order (product, quantity, location).
2. Model: Daraz’s inventory database and delivery routes.
3. Goal: Deliver the order in ≤3 days with minimal cost.
4. Action: Assigns the nearest warehouse and delivery partner.
5. Feedback: User reviews and delivery time data improve future orders.

3. Problem-Solving Tasks in AI

AI tackles problems through four core tasks:

1. Automated Reasoning: Logical deductions (e.g., medical diagnosis systems).
2. Machine Learning: Learning from data (e.g., NEPSE stock prediction).
3. Perception: Interpreting sensory data (e.g., NTC’s traffic camera analysis).
4. Natural Language Processing (NLP): Understanding human language (e.g., eSewa’s chatbot).

3.1 Problem-Solving Approaches

Visual: Game Tree for Tic-Tac-Toe

graph TD
    A["Start"] --> B["X's Turn"]
    B --> C1["X takes top-left"]
    B --> C2["X takes top-center"]
    C1 --> D1["O's Turn"]
    D1 --> E1["O takes center"]
    D1 --> E2["O takes bottom-right"]
    E1 --> F1["X wins"]
    E2 --> F2["Game continues"]

• Explored Path: X takes top-left → O takes center → X wins.
• Used in: Google’s AlphaGo (simulates millions of moves ahead).

4. Limitations of AI

4.1 The Chinese Room Argument (Searle, 1980)

• Claim: A person following rules (without understanding Chinese) in a "Chinese Room" cannot truly understand language.
• Implication: AI systems may simulate intelligence without real comprehension.

4.2 Other Key Limitations

In the Real World

1. eSewa’s Fraud Detection

   • AI Idea: Uses anomaly detection (a machine learning technique) to flag unusual transactions (e.g., sudden large payments).
   • How It Works:
     • Trains on historical data to learn "normal" behavior.
     • Flags transactions deviating from the norm (e.g., a user paying bills in a new country).
     • Real Example: In 2023, eSewa blocked NPR 50 million in fraudulent transactions using this system.

2. Pathao’s Dynamic Pricing

   • AI Idea: Reinforcement learning adjusts fares based on demand (like Uber).
   • How It Works:
     • If demand is high (e.g., during Dashain), prices increase to balance supply.
     • Uses real-time data from GPS and user requests.
   • Worked Example:
     • Scenario: 10:00 PM on a Friday in Kathmandu (high demand).
     • Action: Pathao’s AI increases prices by 30%.
     • Outcome: More drivers join, reducing wait times.

3. Google’s AlphaGo vs. Human Champions

   • AI Idea: Monte Carlo Tree Search (MCTS) + Deep Neural Networks.
   • How It Works:
     • Simulates millions of Go moves per second using a game tree.
     • Learns from past games (including human moves) to improve.
   • Real Impact: Defeated Lee Sedol (2016) and Ke Jie (2017), proving AI’s superiority in complex strategy games.

Exam Tip

1. Definitions Are Critical:

   • Memorize Turing Test, Weak AI vs. Strong AI, and intelligent agent components.
   • Example question: "Explain how a reflex agent differs from a learning agent with an example."

2. Compare and Contrast:

   • Tables (like Weak vs. Strong AI) are high-scoring. Always include real-world examples.

3. Problem-Solving Tasks:

   • Know one example per task (e.g., A search for Pathao*, decision trees for Khalti).
   • Diagrams score marks: Draw a game tree or agent-environment flowchart in exams.

4. Limitations:

   • Chinese Room Argument and bias in AI are frequent exam topics. Relate them to Nepali contexts (e.g., bias in Ncell’s customer service AI).

5. Short Notes:

   • Prepare bullet points for:
     • AI applications in Nepal (eSewa, Daraz, NTC).
     • Ethical dilemmas (e.g., AI in hiring, autonomous vehicles).

5. Summary Table: AI Concepts at a Glance

6. Practice Questions (Self-Check)

1. Define AI and explain why a thermostat is a reflex agent but Google Maps is a model-based agent.
2. Draw a game tree for a simplified version of Ludo and mark the optimal path.
3. How does eSewa use AI to detect fraud? Describe the percepts, actions, and feedback loop.
4. Critique the Turing Test using Searle’s Chinese Room Argument.
5. Compare the AI used in Pathao’s dynamic pricing vs. NTC’s traffic management.

7. Further Reading

• Books:
  • Artificial Intelligence: A Modern Approach (Stuart Russell & Peter Norvig).
  • Life 3.0 (Max Tegmark) – AI’s future and ethics.
• Nepali Context:
  • Study NTC’s smart traffic systems (uses AI for congestion prediction).
  • Explore NEPSE’s AI-driven stock analysis tools.