History and Foundations of Artificial Intelligence
Historical and Philosophical Foundations of Artificial Intelligence
Evolution of Understanding Intelligence: Throughout human history, there has been a continuous inquiry into what distinguishes humans from other species and what constitutes the core strength of human intelligence.
Early Philosophical Perspectives:
The Mind in the Chest: Early philosophers hypothesized that the mind was located in the chest and regulated by emotions.
Emotional Regulation: Modern neuroscience confirms that emotions produce chemicals, such as the stress hormone cortisol, which affect brain activity.
Epicurus: The first philosopher to propose a physical nature for the mind, suggesting it consists of physical atoms.
Rene Descartes and Dualism: Proposed the concept of mind and body dualism, viewing the mind and body as two separate entities.
Mechanistic Interaction: In Descartes' model, eyes perceive an object, the mind processes that perception, and the mind then dictates the body's physical actions, such as pointing at the object.
Robotics Application: This dualistic framework serves as a fundamental principle in modern robotics, where a physical body (limbs, sensors/cameras) is controlled by a separate "mind" model (processing unit).
Modern Philosophy and Ethics: Philosophers established the groundwork for formal logic. In the current landscape, philosophy is critical for addressing ethical considerations in artificial intelligence development.
Mathematical and Engineering Contributions to AI
Mathematics: Multiple branches of mathematics form the technical foundation of AI:
Logic: Includes propositional and first-order logic, used to solve formal problems.
Linear Algebra: Essential for machine learning algorithms, particularly neural networks and regression models, which rely heavily on matrix multiplication.
Calculus: Fundamental for optimization and training algorithms.
Information Theory: Used in applications like decision trees to measure entropy or the "chaos" of data.
Graph Theory: Utilized to represent and solve problems within the first half of AI coursework.
Game Theory: Applied to strategic interactions and decision-making.
Set Theory: Provides mathematical structures for organizing data and concepts.
Psychology: Provides the conceptual framework for replicating human thinking, memory, and feeling.
Conceptual Frameworks: Research in social robotics uses psychological studies to inspire models. For example, some reinforcement learning models simulate biological hormones as mathematical functions to influence agent behavior.
Computer Engineering: The physical implementation of AI theory has evolved from room-sized computers with low computational power to modern supercomputers.
Historical Constraints: Neural network theory has existed for decades, but early engineers lacked the hardware and high-level programming languages (originally relying on binary) to implement them.
Current Capabilities: Supercomputers now allow for the training of massive large language models (LLMs).
Bio-cybernetics: Research labs strive to replicate brain activity physically on microchips. Unlike synchronized neural networks in computers (where matrix multiplication happens simultaneously), biological brain signals are not timed together. Bio-cybernetics attempts to mimic this asynchronous signal processing.
Linguistics: Explores the relationship between language and thought, which is increasingly relevant as LLMs develop.
Defining Artificial Intelligence: Dimensions and Schools of Thought
Four Dimensions of AI Definition: AI can be classified using two main axes:
Human vs. Rational: Is the system trying to think/act like a human, or is it trying to be purely rational (doing the right thing for a specific purpose)?
Thinking vs. Acting: Is the focus on the internal thought processes or the external, observable behavior?
Acting Humanly (The Turing Test Approach):
Definitions: Creating machines that perform functions requiring intelligence when performed by people, or making computers do things at which people are currently better.
The Turing Test: Proposed by Alan Turing to answer "Can a machine think?" A human interrogator interacts with a hidden entity; if they cannot distinguish between a machine and a human, the machine is considered intelligent.
Criticism: The test may reward a machine's ability to "trick" or defraud humans rather than true intelligence. It essentially becomes a competition of disguise.
Thinking Humanly (Cognitive Modeling):
Definitions: Automation of activities associated with human thinking, such as decision-making, learning, and problem-solving.
Methodology: Researchers use MRI and EEG (electroencephalogram) to map brain zones and signals.
Constraints: Human brains are incredibly complex and vary significantly between individuals, making a universal map nearly impossible to create.
Nematode C. Elegans Case Study: Every C. Elegans is an exact copy of another (in terms of neuron count/muscles), existing as either male, female, or hermaphrodite. In , researchers successfully created a digital map of its brain. Activating a specific digital neuron produces the exact same physical activity seen in a live nematode.
Thinking Rationally (Laws of Thought):
Definitions: The study of mental faculties through computational models that perceive, reason, and act.
Syllogisms: A logical structure using deductive reasoning. Example:
Premise 1: Socrates is a man.
Premise 2: All men are mortal.
Conclusion: Socrates is mortal.
Boolean Logic: Established in by George Boole, who transformed philosophical logic into a mathematical discipline.
Theoretical Limits: While logic-based programs can theoretically solve any solvable problem, the real world is too complex and ambiguous to describe purely via formal facts. This necessitates the use of fuzzy logic and probability.
Acting Rationally (The Rational Agent Approach):
Definitions: The study of computational agents that act intelligently. This behavior is appropriate for its goals and circumstances, flexible to changing environments, and learns from experience.
European Union Definition: The High-Level Expert Group defines AI through the notion of rationality—choosing the best action to achieve a goal given specific criteria and available resources.
Evolutionary milestones in AI Systems
Eliza (1916): An early chatbot designed with the profile of a doctor. It used simple English grammar rules to swap words in a visitor's sentence to form a response.
Failures in Reasoning:
Visitor: "The trouble is my mother is Irish."
Eliza: "How long has she been Irish?"
Visitor: "The trouble is my mother is poodle."
Eliza: "How long has she been poodle?"
This demonstrates that the system lacked situational context and could not detect nonsense.
Winograd Schema Challenge: Designed to test understanding through anaphora resolution.
Example: "The trophy doesn't fit into the brown suitcase because too large."
Human Understanding: If "it's too large," refers to the trophy. If "it's too small," refers to the suitcase.
Bias Found: A paper entitled "WinoGrande: An Adversarial Winograd Schema Challenge at Scale" found that neural networks were learning patterns of big/small relationships rather than truly understanding language. New unbiased datasets showed a drop in performance.
Google Lambda (2021): A model that led a Google employee to claim sentience on Twitter after it expressed a fear of being turned off.
ChatGPT Development:
2022 Performance: When asked how many "r"s are in the word "strawberry," it answered "two."
2025 Performance: It correctly identifies "three" "r"s.
Turing Test Comparison: ChatGPT handles complex literary questions (e.g., comparing a sonnet's line to a "summer's day" vs. "spring day") with historical context, suggesting a appearance of understanding that would likely impress Alan Turing.
The Chinese Room Argument: A criticism of the Turing Test imagining a person in a room using a dictionary to map English to Chinese without actually understanding either language. From the outside, it looks intelligent, but inside, it is merely a mapping problem.
Intelligent Agents and Schools of Thought
Agent Examples:
People: Teachers, Doctors (aiming for healthy patients), Stock Traders (aiming for profit by buying low and selling high).
Devices: Smart cars, advising systems (Netflix recommendations).
Biological Entities: Animals, bacteria, and fungus (e.g., fungus growing toward food sources).
Organizations: The University of Auckland. It acts rationally by scheduling hundreds of classes with varying sizes and limited classroom resources to avoid teachers being in two places at once.
Weak AI: Machines that act as if they are intelligent but are limited to a specific area (e.g., a system that only detects car license plates).
Strong AI: A single entity capable of solving many different creative problems.
Current State: Truly Strong AI does not exist yet. However, generative models are moving toward "orchestration," where a central agent (like ChatGPT) mediates between different specialized AI models (e.g., one for text, one for image generation like DALL-E) to simulate one cohesive intelligent entity.
Questions & Discussion
Question: Do you have a positive or negative view on the impact of AI? Will humanity be "doomed" or will AI help us prosper?
Response (Student - Aang): I have a positive view. Throughout history, humanity has gone through three revolutions (e.g., the paper machine). AI is the same. Currently, AI is not at a human level; it is a "mapping problem." It learns from data and maps it to output. It is just a tool we can control because it cannot go above its learning data.
Question: Do you think AI will ever surpass human intelligence or solve problems it has never seen before?
Response (Student): In its current state, it won't. It is just very good at predicting. It can't break through to solve problems it hasn't seen. It would need a fundamental change to start thinking beyond humans.
Comment (Class Poll): Many people believed AI might have limits, but thirty years ago, nobody expected LLMs to exist. In , the development was considered revolutionary.
Question: Will AI replace humans in jobs?
Discussion on Employment: There is a hope that AI will take over chores. However, there is a meme where a lady notes she hoped AI would do chores so she could do art, but instead, AI is doing the art and she is still doing the chores. The goal is for robots to eventually handle cleaning, ironing, and food delivery so humans can focus on their strengths.
Follow-up Question on Strong AI: Does Strong AI exist today?
Response: No. Current research uses a modular approach similar to the human brain (vision, action, memory). Systems use an "orchestration agent" or mediator to decide which specialized model (text vs. image) to engage for a specific query.
Quiz Results: A quiz was conducted on Canvas Module 1, Week 1. The winner was Zach, who received a candy and a special handshake. Finley and Akh were also recognized as top performers.