Four Types of AI

The Two Dimensions

1. Top vs. Bottom (Thinking vs. Acting):

   • Thinking (Internal): Focuses on the reasoning process, the "mind," and how inferences are made.
     
     
   • Acting (External): Focuses on behavior, output, and the results of actions, regardless of how they were computed.

2. Left vs. Right (Humanly vs. Rationally):

   • Humanly (Empirical): The standard of success is "people." Does it do it like we do? (Even if humans make mistakes).
     
     
   • Rationally (Mathematical): The standard of success is "optimality." Does it do the correct thing to achieve the best outcome?

1. Acting Humanly (The Turing Test Approach)

• Goal: To create a system that acts so much like a person that you cannot tell the difference.
  
  

• The Test: The famous Turing Test (1950). A human interrogator types questions to a computer and a human. If the interrogator cannot determine which is which, the computer has passed.
  
  

• Required Capabilities: To pass, the computer doesn't need to be "conscious," but it needs specific skills:

  • Natural Language Processing (to communicate).

  • Knowledge Representation (to store what it knows).

  • Automated Reasoning (to answer questions and draw new conclusions).

  • Machine Learning (to adapt to new circumstances).
    
    

• Slide Example: The slides classify "Acting like human beings" as Weak AI. Examples given are humanoid robots (like Atlas from Boston Dynamics) or iRobot vacuums—machines that mimic human motion or function.

2. Thinking Humanly (The Cognitive Modeling Approach)

• Goal: To understand and mimic the actual workings of the human mind. It is not enough to just get the right answer; the steps taken to get there must match the steps a human brain takes.
  
  
• Method: This requires looking inside the human mind through introspection (catching our own thoughts) or psychological experiments.
  
  
• Connection to Science: This is the bridge between AI and Cognitive Science.
  
  
• Slide Example: The slides classify this as Weak AI. Examples include Watson and AlphaGo. Even though they solve complex problems, they are "weak" in this specific definition because they don't actually think or experience consciousness the way a human does; they simulate the result of thinking.

3. Thinking Rationally (The "Laws of Thought" Approach)

• Goal: To codify "right thinking" using logic.
  
  

• Basis: This is based on Aristotle’s syllogisms (e.g., "Socrates is a man; all men are mortal; therefore, Socrates is mortal").
  
  

• The Problem:

  1. It is hard to take informal, fuzzy knowledge ("It looks like it might rain") and turn it into strict logical notation.
     
     
  2. There is a difference between solving a problem "in principle" and doing it in practice (computational exhaustion).
  
  

• Slide Status: The slides classify this as Strong AI, noting that no intelligent agent of this type currently exists due to bottlenecks in brain science. We cannot yet perfectly model the brain's logical processes physically.

4. Acting Rationally (The Rational Agent Approach)

• Goal: To act in a way that achieves the best expected outcome.
  
  
• Why this is the focus of the course:
  
  
  • It is more general than "Thinking Rationally." (Example: Pulling your hand away from a hot stove is a reflex action—it is "rational" because it saves your hand, but it doesn't involve "thinking" or logic).
    
    
  • It is scientifically easier to test than "Human" approaches. We can mathematically measure if an agent achieved a goal (Rationality), but it is subjective to measure if it "thought like a human."
    
    
• Slide Status: This is the definition of Computational Intelligence used in the course. The agent simply uses its sensors and effectors to maximize its performance measure.