Comprehensive Study Notes on Foundations of Artificial Intelligence and Theories of Intelligence

Administrative Announcements and Course Logistics

  • Homework 1 Availability:

    • The first homework assignment is accessible across campus and functions as a campus quiz.

    • Reading the full assigned material is optional; the main lecture content provides sufficient preparation.

    • Supplementary resources, including external reading links and deep-dive videos, are available for extended study.

  • Top Hat Access Instructions:

    • To resolve standard login issues, initial setup should be conducted via Canvas rather than logging into Top Hat directly or scanning the barcode immediately.

    • Students must click on the Top Hat link located on the Canvas navigation sidebar to be prompted into the correct event space.

  • Penn State AI Society Information:

    • Student organization website: niddyai.su.edu.

    • Associated groups include the Niddy AI Alliance, the Niddy AI Student Society, and Niddy Data Labs.

    • Activities include AI training modules, real-world industry projects, corporate sponsorships, and a mini AI challenge formatted as a startup competition incubator.

    • In 20192019, the course instructor competed in this challenge as an undergraduate student alongside a team of friends and won a prize of text$5,000\\text{\$5,000}.

Recap of Foundations: Cognition, Computation, and Turing Machines

  • The Working Hypothesis of Artificial Intelligence:

    • Strong Version: Cognition is strictly computation.

    • Weak Version: Cognition can be modeled by computation.

    • Core Definitions: Cognition is defined fundamentally as thinking. Computation is defined as any operational process capable of running on a Turing machine or a physical computer.

    • Logical Rationale: The validity of studying AI rests on this hypothesis. If cognition could not be modeled by computation, simulating or building intelligent systems via computational machinery would be impossible.

  • Formal Definition of Computation:

    • Computation is defined abstractly as the process of transforming one string of characters into another string of characters using precise rules or recipes.

    • This simple string-transformation definition is deliberately general to capture all theoretical and practical computational problems.

  • Examples of String Transformation:

    • Addition: Transforming the string "3 + 4"\text{"3 + 4"} into the string "7"\text{"7"} using the mathematical algorithm for addition.

    • String Reversal: Reversing character sequences using a string reversal recipe.

    • General Problem Solving: Problems such as finding the shortest path in a graph or training machine learning models are structurally equivalent to transforming one string to another.

  • Turing Machines:

    • A Turing machine represents the simplest theoretical model of a computer.

    • Capability Boundary: A modern computer cannot perform any computational task beyond the capacity of a universal Turing machine.

    • Practicality: Turing machines are theoretical constructs and are impractical for direct physical deployment in real-world modern computing.

Alan Turing: Cryptanalysis, Biology, and Historical Impact

  • World War II Cryptanalysis:

    • Alan Turing played a central role in breaking encrypted Nazi military communications (including Enigma), providing critical tactical advantages to the Allies.

  • Athletic Achievements:

    • Turing was an elite long-distance runner who nearly qualified for the Olympic Games before suffering a leg injury.

    • His personal best marathon time was only 1111 minutes slower than the British silver medalist in 1915\text{1915} (or 1948\text{1948} Olympic era contexts).

  • Biological and Mathematical Research:

    • Turing authored a seminal paper titled The Chemical Basis of Morphogenesis.

    • This work applied mathematical modeling to biological development, analyzing how pattern formations arise in organisms such as fish, snails, and amphibians.

  • State Persecution and Death:

    • Following a conviction under British laws prosecuting homosexuality, Turing underwent forced chemical castration.

    • He died a few years later from cyanide poisoning.

    • A half-eaten apple was discovered at his bedside; controversy remains as to whether his death was a suicide or an accidental inhalation caused by laboratory equipment.

    • His life was dramatized in the biographical film The Imitation Game.

The 1956 Dartmouth Workshop and the Birth of Artificial Intelligence

  • Coining of the Term "Artificial Intelligence":

    • The term "Artificial Intelligence" was formally established in 19561956 during an academic workshop held at Dartmouth College.

    • Attendees included leading mathematicians, psychologists, computer scientists, and cyberneticists.

  • Key Historic Attendees:

    • John McCarthy: Primary organizer of the Dartmouth workshop.

    • Wendell Peurer: Early pioneer and academic advisor to the instructor's advisor.

    • Claude Shannon: Founder of Information Theory; his legacy inspired the naming of Anthropic's AI model, Claude.

    • Marvin Minsky: MIT professor and influential early figure in AI, who died in 20162016. Decades after Dartmouth, Minsky organized academic AI symposiums at Little Saint James.

  • The Dartmouth Workshop Conjectural Basis:

    • The foundational premise stated that every aspect of learning or feature of intelligence can be described so precisely that a machine can be constructed to simulate it.

  • Socioeconomic Paradigm Shift:

    • While the Industrial Revolution automated physical and manual labor, AI automates cognitive, mental, and white-collar labor.

Concepts and Definitions of Intelligence: Life, Adaptability, and Grey Areas

  • Entity Classification Exercises:

    • Trees: Arguments for tree intelligence rely on environmental adaptability—such as roots navigating obstacles to find water, branches competing for sunlight, or growth paths between physical barriers.

    • Viruses: In class surveys, more respondents categorized viruses as intelligent than trees, despite biological classification differences.

  • Biological Status of Viruses:

    • Scientific Consensus: Viruses are defined as non-living because they lack cellular structure, cannot independently reproduce, cannot maintain homeostasis, and do not undergo intrinsic growth.

    • Borderline Arguments: Some scientists classify viruses in a grey zone between living and non-living matter due to their ability to evolve via natural selection and manipulate host mechanisms upon infection.

  • Defining Life vs. Defining Intelligence:

    • NASA Definition of Life: Life is a self-sustaining chemical system capable of Darwinian evolution.

    • Chemical Constraint: Current AI and robotic systems operate on electrical and physical substrate mechanisms rather than chemical reactions.

  • Collective Intelligence:

    • Approximately 90%90\% of polled students categorized human society as a collective intelligence.

    • Individual biological organisms are themselves collective intelligences comprised of billions of somatic cells working alongside internal microbiomes (e.g., gut bacteria).

Empirical Case Studies in Early Language Models: GPT-1 and GPT-2

  • GPT Architecture Nomenclature:

    • Generative: The model generates text output sequences.

    • Pre-trained: Trained on large web and book text corpora to predict the next token in a sequence.

    • Transformer: Built specifically on the transformer neural network architecture.

  • GPT-1 Case Study (Released 20182018):

    • GPT-1 was released four years prior to ChatGPT (20222022).

    • Unlike chat-aligned models, base GPT-1 was strictly trained for next-token prediction without instruction fine-tuning.

    • Prompt Test: When asked "What is the capital of America?", GPT-1 produced nonsensical script-like dialogue loops: "It's a small country, he said... Because I don't wanna talk about it. Why not? Because I don't wanna talk about it."

  • GPT-2 Case Study:

    • When queried with "What is the capital of America?", GPT-2 produced syntactically structured, academic-sounding text discussing definitions of capital and business ownership, complete with fake academic citation markers, but failed to output the correct answer ("Washington, D.C.").

  • Model Hosting Infrastructure:

    • Hugging Face provides model repository and inference hosting capabilities allowing users to run or download model checkpoints directly.

Typologies, Theories, and Quotations on Intelligence

  • Theoretical Frameworks of Intelligence:

    • Performance-Based Theories: Defines intelligence based on passing specific benchmarks or evaluation tests (e.g., The Turing Test).

    • Mechanism-Based Theories: Defines intelligence by internal structural organization (e.g., biological brain structures or artificial neural network architectures).

    • Cognitive Theories: Focuses on internal component processes such as memory recall, pattern comparison, and execution speed/efficiency.

    • Contextual Theories: Asserts intelligence is contingent on adapting to specific environmental contexts (e.g., a student adjusting to a new school or a farmer adapting to changing climate conditions).

    • Developmental Theories: Views intelligence as progressive acquisition of capabilities through defined stages (e.g., object permanence from age 00 to 22; language acquisition and abstract logic from age 22 to 77).

  • Triarchic Typology of Intelligence:

    • Analytical Intelligence: Academic or "book smarts" focused on decomposing complex problems, mathematical logic, legal analysis, and risk-benefit calculations.

    • Creative Intelligence: The ability to synthesize novel solutions, architectural styles, technical inventions (e.g., the airplane), or artistic movements (e.g., early 1900s jazz).

    • Practical Intelligence: "Street smarts" focused on navigating environmental constraints, execution, and cultural adaptation.

  • Key Formal Definitions and Quotations:

    • Standard Dictionary Definition: The mental capacity to learn, reason, understand, solve problems, and adapt to new situations.

    • Efficiency View: "Intelligence is doing more with less" (illustrated by the human brain's low power consumption).

    • Compression View: Intelligence is fundamentally about compressing sensory experience into internal functional models usable for survival and decision-making.

    • Evolutionary View: "Intelligence is the ability to solve the same problem in different ways" (illustrated by biological evolution solving survival through diverse morphology, from octopuses to elephants).

    • Einstein Paraphrase: "The true sign of intelligence is not knowledge, but imagination."

    • Adaptive View: "Adaptive intelligence is the ability to generate new competence when confronted with novelty."

Mathematical Formalisms of Intelligence

  • 1. The Tripartite Search Framework:

    • Intelligence is represented as a formal system containing three components:

      1. Model Making: Developing symbolic or structural representations of a problem space.

      2. Search: Evaluating potential action choices within a decision tree.

      3. Heuristics: Utilizing rules of thumb to prune exponential search spaces and allow feasible decision selection.

    • Chess Example: Opening moves branch exponentially. Full game-tree evaluation via minimax search is computationally intractable without heuristics to prune suboptimal search paths.

  • 2. The Gödel Machine Formalism:

    • Proposed as a theoretical, recursively self-improving computer program.

    • The system uses a formal proof searcher to inspect its own source code. When it proves that modifying its internal code will yield superior performance, it rewrites its own software.

    • Modern Industry Parallel: OpenAI maintains dedicated research engineering roles for "Recursive Self-Improvement Safety" focusing on scalable oversight, model behavior science, and automated auditing.

    • Coding Capabilities: Anthropic reports exponential growth in software engineering output driven by internal models writing code to expand their own codebase.

  • 3. The AIXI Framework:

    • AIXIAIXI is a theoretical mathematical formalism for Artificial General Intelligence (AGI) grounded in reinforcement learning.

    • An agent interacts with an environment by taking actions, receiving observations and rewards, and operating to maximize cumulative expected future reward over an infinite horizon.

    • While mathematically precise, AIXIAIXI is formally uncomputable and serves as an idealized theoretical benchmark rather than a deployable algorithm.