AI Foundation · Domain 01

Foundations of Artificial Intelligence

What AI actually is, where it came from, why symbolic AI failed, the philosophical debates that matter today, the rational agent framework, competing paradigms, and the state of the field.

1.1

Chapter 1.1 · Definitions & Scope

What Is Artificial Intelligence?

Most AI discourse fails at the first step: no crisp definition. AI is not magic, not a single algorithm, and not science fiction. It is the science and engineering of building systems that exhibit behaviour we would call intelligent if a human exhibited it — and that definition is deliberately provisional, because intelligence itself resists definition.

Defining Intelligence In-depth

🧬

Biological Intelligence

Embodied — tied to a body with survival pressures
Developed through evolution over millions of years
Energy-efficient: 20 watts powers the human brain
Generalises from very few examples (few-shot by default)
Handles open-ended, ambiguous, novel situations naturally

🤖

Machine Intelligence

Disembodied — no physical survival pressure
Built through training on human-generated data
Energy-hungry: GPT-4 training cost ~$100M in compute
Requires millions–billions of examples to learn patterns
Brittle outside training distribution; exceptional at defined tasks

What makes a system "intelligent"? There is no consensus. The four most-cited criteria:

Ability to learn from experience
Ability to solve novel problems
Ability to understand and generate language
Ability to reason under uncertainty

No current AI system fully satisfies all four in the general sense humans do.

AI vs ML vs Deep Learning vs Data Science Core

The AI Field — Nested Subfields & Overlaps

Term	Definition	Subset of	Example
Artificial Intelligence	Any technique enabling machines to mimic aspects of human cognition	— (broadest)	Chess engines, expert systems, LLMs, robotics
Machine Learning	AI that improves by learning from data rather than explicit rules	AI	Spam filter, fraud detection, recommendation systems
Deep Learning	ML using multi-layer neural networks to learn hierarchical representations	ML	GPT-4, image classifiers, speech recognition
Generative AI	Models that generate new content (text, images, audio, code) plausibly like training data	Deep Learning	ChatGPT, DALL-E, Stable Diffusion, Sora
Data Science	Interdisciplinary field combining statistics, ML, and domain expertise to extract insights from data	Overlaps AI & Statistics	Business dashboards, A/B testing, churn analysis

Common misuse: "We use AI" usually means "we use ML" and often specifically means "we use a trained model." Precision matters — especially evaluating vendor claims. Ask: Is it rule-based? Trained? On what data? Measured by what metric?

Types of AI: Narrow, General, Super Core

The AI Capability Spectrum — ANI → AGI → ASI

🎯

ANI — Narrow AI

Excels at one specific task. All commercially deployed AI today is ANI — including GPT-4. Examples: chess engines, image classifiers, recommendation algorithms, language models.

STATUS: Present reality (2026)

🧠

AGI — General AI

Human-level cognitive ability across any intellectual domain. Can learn any task a human can, without domain-specific training. No AGI exists. Timelines debated: 5–50+ years by leading researchers.

STATUS: Research goal

⚡

ASI — Super AI

Surpasses human intelligence in every domain — creativity, scientific discovery, social intelligence. Purely theoretical. Motivates alignment research and existential risk discourse (Domain 10).

STATUS: Theoretical

What AI Is Not Introductory

Myth	Reality
AI "understands" things	AI processes statistical patterns in training data. Whether this constitutes understanding is a genuine philosophical debate (Ch 1.4) — but current systems don't understand in the human sense.
AI is deterministic	Most modern AI uses probabilistic sampling. The same prompt produces different outputs. Temperature and sampling parameters control randomness.
AI develops goals on its own	AI optimises for its training objective. It doesn't spontaneously form desires. Alignment failures happen when the objective doesn't match human intent — not through "waking up."
More data always = better AI	Data quality, labelling accuracy, and relevance matter more than volume. 1B clean samples often outperform 10B noisy ones.
AI replaces human intelligence wholesale	Current AI replaces specific tasks within jobs, not entire jobs at once. Displacement patterns are complex and highly uneven across domains.

1.2

Chapter 1.2 · Narrative History

History of AI — Origin to Present

AI history is a story of alternating euphoria and collapse. Understanding why each era gave way to the next — not just when — is the best inoculation against misreading hype today. The pattern: overpromise → underfund → winter → unexpected breakthrough → repeat.

AI Progress & Hype — 1950 to 2026

Pre-History: Logic & Automata Introductory

📜

Leibniz & Boole (1600s–1800s)

Leibniz dreamed of a calculus ratiocinator — a machine that could calculate truth from symbols. Boole formalised logic as algebra (1854), creating the mathematical foundation for all computation that followed.

✍️

Ada Lovelace (1843)

Writing notes on Babbage's Analytical Engine, Lovelace described how the machine could be programmed to compose music — arguably the first vision of general-purpose computing. She also articulated its limits: it can only do what we tell it. The tension between capability and instruction persists today.

⚙️

Early Mechanical Calculators

Pascal's Pascaline (1642), Babbage's Difference Engine (1822) — mechanical attempts to automate calculation. They mechanised arithmetic but not reasoning. The distinction matters: calculation ≠ intelligence.

Birth of AI (1943–1956) In-depth

The Three Founding Documents

McCulloch & Pitts (1943) — "A Logical Calculus of Ideas Immanent in Nervous Activity"
Proposed the first mathematical model of a neuron — a binary threshold unit that fires when inputs exceed a threshold. Showed that networks of these units could compute any logical proposition. Direct ancestor of all modern neural networks.

Alan Turing (1950) — "Computing Machinery and Intelligence"
Proposed the Imitation Game as a pragmatic test for machine intelligence. Anticipated and answered nine objections to machine thought. Introduced the "child machine" — a machine that learns rather than being pre-programmed. The first articulation of what we now call machine learning.

Dartmouth Conference (1956) — McCarthy, Minsky, Shannon, Simon
The proposal: "Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." McCarthy coined "Artificial Intelligence." Optimism proved premature — but the research agenda set here dominated the field for 20 years.

The Golden Years (1956–1974) Core

🌟

Early Successes

Logic Theorist (1955): proved 38 of 52 theorems from Principia Mathematica
GPS — General Problem Solver (1957): domain-agnostic search; separated problem description from problem-solving method
ELIZA (1966): Weizenbaum's chatbot simulated a Rogerian therapist. Users formed genuine emotional connections — a warning about anthropomorphism that still applies to LLMs today.
SHRDLU (1970): natural language system that manipulated blocks in a simulated world. Impressive in its toy domain; completely brittle outside it.

⚠️

Seeds of the First Winter

Minsky & Papert (1969) proved perceptrons can't learn XOR — stalling neural network research for a decade
ALPAC report (1966): machine translation was "twice as expensive and twice as slow" as human translation
Combinatorial explosion: real-world problems have exponentially more states than toy problems
The "Potemkin village" problem: early demos were cherry-picked; the gap between demos and general capability was enormous

The AI Winters (1974–1980, 1987–1993) In-depth

Why study the winters? The pattern of hype → disappointment → funding collapse has repeated twice in AI history. Understanding the mechanism prevents misreading the current cycle. The winters were not caused by bad research — they were caused by the gap between what was promised and what was delivered, plus a fundamental limitation (compute + data) that took decades to overcome.

Winter	Trigger	What Collapsed	Root Cause	What Survived
First (1974–80)	Lighthill Report (UK, 1973); DARPA cuts	General-purpose AI, machine translation, symbolic reasoning	Combinatorial explosion; overpromised timelines to funders	Specialised systems; early expert system research; Prolog
Second (1987–93)	Expert system maintenance failures; Lisp machine market collapse	Commercial expert systems; DARPA Strategic Computing Program	Expert systems too brittle and expensive to maintain at real-world scale	Backpropagation (1986 rediscovery); statistical ML; RL foundations

❄️

Why Expert Systems Failed

Knowledge acquisition bottleneck: experts couldn't articulate their own tacit knowledge
Brittleness: outside their narrow domain, systems failed completely — no graceful degradation
Maintenance cost: real-world rule sets grew to thousands of rules that became unmaintainable
No learning: every change required manual update by knowledge engineers

📖

The Lighthill Report (1973)

Sir James Lighthill's report to the UK Science Research Council concluded that AI had failed to live up to its "grandiose objectives" due to the combinatorial explosion problem. The UK effectively abandoned AI funding for a decade. The US followed.

Lesson: Overpromising specific timelines to funders is more damaging than the research being wrong.

The Connectionist Revival (1980s–2000s) Core

🔄

Backprop Rediscovered (1986)

Rumelhart, Hinton & Williams published the backpropagation algorithm. Earlier work by Werbos (1974) had gone unnoticed. Training multi-layer networks became practical. The door to deep learning opened — but GPUs weren't ready yet.

📊

Statistical ML Rise (1990s)

SVMs (Vapnik, 1995) provided strong theoretical guarantees. Hidden Markov Models dominated speech recognition. Probabilistic graphical models matured. ML became rigorous and mathematical, focused on generalization theory rather than AI folklore.

🏆

Key Milestones

Deep Blue beats Kasparov (1997)
LSTM networks (Hochreiter & Schmidhuber, 1997)
LeNet-5 for digit recognition (LeCun, 1998)
Netflix Prize — collaborative filtering (2006–09)
ImageNet dataset created (Fei-Fei Li, 2009)

The Deep Learning Era (2012–Present) In-depth

Three factors converged in 2012: massive datasets (ImageNet: 1.2M labelled images), GPU computing (CUDA parallelism), and algorithmic improvements (ReLU, dropout, batch normalisation). None alone was sufficient; all three together were transformative.

Three Inflection Points That Changed Everything

▶

2012 — AlexNet (Krizhevsky, Sutskever, Hinton)
Won ImageNet with 15.3% top-5 error vs. 26.2% for the runner-up — an 11-point gap that shocked the computer vision community. Trained on 2 GTX 580 GPUs in 5 days. Every major research lab pivoted to neural networks within 18 months.

▶

2017 — "Attention Is All You Need" (Vaswani et al., Google Brain)
The Transformer architecture replaces recurrent networks for sequence modelling. Self-attention enables full context access without sequential bottlenecks. Direct ancestor of BERT, GPT, T5, LLaMA, and every modern LLM.

▶

2022 — ChatGPT (OpenAI)
100 million users in 60 days — fastest consumer technology adoption in history. Demonstrated that RLHF (Reinforcement Learning from Human Feedback) could align large language models to be genuinely helpful and safe. Moved AI from research into mainstream public discourse permanently.

Year	Event	Why It Matters
2012	AlexNet wins ImageNet	Deep learning proven at scale; GPU compute validated as the path forward
2014	GANs (Goodfellow)	Generative models can produce realistic images; generative AI era begins
2016	AlphaGo defeats Lee Sedol	RL + deep learning = superhuman Go; 10 years ahead of schedule
2017	Transformer architecture	Replaces RNNs; enables parallelism; foundation of all modern LLMs
2018	BERT & GPT-1	Pre-train → fine-tune paradigm established for NLP
2020	GPT-3 (175B params)	Few-shot learning at scale; foundation model era begins
2021	DALL-E, Codex, AlphaFold 2	Multimodal generation; protein structure solved after 50-year open problem
2022	ChatGPT; Stable Diffusion	AI goes mainstream; open-source generative models democratise access
2023–24	GPT-4, Claude 3, Gemini, LLaMA	Multimodal capability; open weights; commercial proliferation
2025–26	Agentic AI; reasoning models	o1, o3, DeepSeek-R1, Claude 3.7; agents handle full multi-step workflows

1.3

Chapter 1.3 · Knowledge & Reasoning

Symbolic AI & Knowledge Representation

Symbolic AI is commonly omitted from modern documentation. This is a mistake. Understanding why symbolic AI dominated for 30 years — and exactly why it failed to scale — is what explains why neural networks won. More importantly, it reveals what neural networks still cannot do.

Symbolic AI vs. Neural AI — How Knowledge Is Represented

Symbolic vs Sub-Symbolic AI Core

Dimension	Symbolic (GOFAI)	Sub-Symbolic (Connectionist)
Knowledge representation	Explicit symbols, rules, logic — human-readable	Distributed across billions of numerical weights — not human-readable
Reasoning	Formal inference (deduction, induction, abduction)	Pattern interpolation from training data
Interpretability	Fully interpretable — you can read and audit the rules	Black box — weights are not meaningfully inspectable
Learning	Difficult to learn from raw data; requires manual knowledge encoding	Learns directly from raw data; scales with data and compute
Generalisation	Brittle outside defined rule coverage	Generalises within training distribution; fails on distribution shift
Current form	Knowledge graphs, ontologies, formal methods, SAT solvers	LLMs, CNNs, diffusion models, transformers

Knowledge Representation Core

Semantic Network — Knowledge as a Graph

Nodes = concepts · Edges = relationships · Inheritance flows along is-a links

🕸️

Semantic Networks

Graph-based: nodes are concepts, edges are relationships. Socrates → is-a → Human → is-a → Mortal. Allows inheritance — Socrates automatically inherits all human properties. Precursor to modern knowledge graphs (Wikidata, Google Knowledge Graph).

📋

Frames (Minsky, 1974)

Structured representations with named slots and default values. A "Restaurant" frame has slots: name, type, price_range, menu, hours. When visiting a new restaurant, you fill known slots and use defaults for unknown ones. Directly influenced object-oriented programming.

🏗️

Ontologies

Formal specification of concepts and relationships within a domain. The Cyc project (Lenat, 1984–present) attempted to encode all human common-sense knowledge — 25M+ rules over 40 years. Still operational but not competitive with neural approaches on most tasks.

🔢

Description Logics

Formal languages for ontologies with decidable reasoning. Basis of OWL (Web Ontology Language) — standard for semantic web and knowledge graphs. Allows automated consistency checking and inference over structured knowledge.

Expert Systems Core

⚕️

MYCIN (Stanford, 1972)

Diagnosed bacterial infections and recommended antibiotics using 600 production rules and certainty factors. Outperformed medical students and matched specialists in controlled tests. Never deployed clinically — not due to technical failure, but to liability concerns.

 IF organism-stain = gram-negative
 AND organism-morphology = rod
 AND patient-compromised-host = true
 THEN organism = pseudomonas  [CF: 0.6]

🔬

DENDRAL (Stanford, 1965)

Identified organic molecules from mass spectrometry data. First expert system with real scientific impact — published findings human chemists hadn't noticed. Proved that narrow domain expertise could be encoded computationally and productively.

Forward vs. Backward Chaining
Forward chaining (data-driven): start from known facts, apply rules to derive new facts until goal is reached. Used in production rule systems. Backward chaining (goal-driven): start from the goal, work backwards to find rules that could prove it. Used in Prolog and MYCIN. Backward chaining is more efficient when the goal is specific and the search space is large.

Logic & Reasoning In-depth

🔣

Propositional Logic

Atomic propositions connected by ∧ (AND), ∨ (OR), ¬ (NOT), → (implies). Truth-functional — truth of compound depends only on truth of components. Decidable but lacks expressive power: no variables, quantifiers, or relations between objects.

P = "It rains"    Q = "Ground is wet"

P → Q // if it rains, ground is wet

P // it rains (given)

∴ Q // modus ponens: ground is wet

∀

First-Order Logic (FOL)

Extends propositional logic with objects, predicates, and quantifiers (∀ for all, ∃ there exists). Can represent most mathematical knowledge. Semi-decidable: you can prove theorems but not always disprove them in finite time.

∀x: Human(x) → Mortal(x) // all humans are mortal

Human(Socrates) // Socrates is human

∴ Mortal(Socrates) // universal instantiation 

Non-Monotonic Reasoning — Classical logic is monotonic: adding new knowledge never invalidates prior conclusions. But the real world isn't like that. "Birds fly" is true — until you learn the bird is a penguin. Non-monotonic reasoning (default logic, circumscription) handles exceptions and retractions. Neural networks handle this implicitly through pattern matching — but without explicit auditable reasoning trails.

Search Algorithms In-depth

Algorithm	Strategy	Complete?	Optimal?	Time	Space
BFS	Expand shallowest nodes first	✓ Yes	✓ Yes (unit cost)	O(b^d)	O(b^d)
DFS	Expand deepest nodes first	✗ No	✗ No	O(b^m)	O(bm)
IDDFS	DFS with increasing depth limit	✓ Yes	✓ Yes	O(b^d)	O(bd)
Uniform Cost	Expand lowest-cost node first	✓ Yes	✓ Yes (admissible h)	O(b^C/ε)	O(b^C/ε)
Greedy Best-First	Expand node closest to goal (h only)	✗ No	✗ No	O(b^m)	O(b^m)
A*	f(n) = g(n) + h(n) — cost + heuristic	✓ Yes	✓ Yes (admissible h)	O(b^d)	O(b^d)

A* Worked Example: g(n) = actual cost from start to node n. h(n) = heuristic estimate to goal (e.g., straight-line distance for routing). f(n) = g(n) + h(n).

Admissibility condition: h(n) must never overestimate actual cost. If admissible, A* is guaranteed to find the optimal path. A better heuristic means fewer nodes expanded — the difference between seconds and hours on large graphs.

Minimax & Alpha-Beta: Two-player zero-sum games use Minimax (maximise your score, opponent minimises). Alpha-Beta pruning eliminates branches that cannot affect the final decision, reducing the effective branching factor from b to √b — roughly doubling searchable depth for the same compute.

1.4

Chapter 1.4 · Philosophy of Mind

The Turing Test & Philosophy of Mind

These aren't abstract puzzles. The Turing Test, Chinese Room, and symbol grounding problem are actively debated in the context of LLMs right now. They directly motivate alignment research, interpretability, and legal questions around AI. Don't treat this chapter as optional theory.

The Turing Test In-depth

The Imitation Game — Turing Test Setup

The Original Imitation Game (Turing, 1950)
Three participants: a human interrogator (C), a human (A), and a machine (B). The interrogator communicates only via text. The machine's goal is to convince the interrogator it is human. Turing proposed this not as a definition of intelligence, but as a pragmatic operational test — if a machine can consistently fool a human interrogator, we have sufficient reason to call it intelligent, whatever "intelligent" means.

Critique	Why It Matters Today
Tests behaviour, not cognition — passes without understanding	GPT-4 informally passes casual Turing tests. The understanding question remains fully open.
Anthropocentric — measures human-likeness, not intelligence per se	A genuinely intelligent alien system might fail. The test conflates intelligence with human-mimicry.
Humans can fail it too — adversarial settings trip humans up	CAPTCHA systems exploit this. Humans fail certain adversarial Turing variants more than LLMs do.
Tests only conversational fluency, not reasoning or knowledge	LLMs excel at fluency but fail structured reasoning and factual tests at the same time.

Modern variations: CAPTCHA (reverse Turing test — machines trying to verify humans), Winograd Schema Challenge (pronoun disambiguation requiring commonsense reasoning — harder for machines than conversational mimicry), and the Loebner Prize (formal Turing Test competition — no system has convincingly passed under rigorous adversarial evaluation).

The Chinese Room Argument In-depth

Searle's Chinese Room — Syntax Without Semantics

Searle's Thought Experiment (1980)
Imagine a person locked in a room receiving Chinese characters through a slot. They have a rulebook (in English) telling them how to manipulate the symbols and which characters to return. From outside, the room appears to understand Chinese — it passes any Turing test for Chinese comprehension. But the person inside understands nothing. They're just manipulating symbols according to rules.

Searle's conclusion: Programs manipulate syntax (formal symbol structures). Understanding requires semantics (meaning). Syntax alone is not sufficient for semantics. Therefore no program — regardless of performance — can have genuine understanding.

Counter-Argument	Searle's Reply	Current Status
Systems Reply: The room as a whole understands Chinese, even if no individual part does	Imagine the person memorises the entire rulebook — they still understand nothing. Systems don't have understanding any more than their parts.	Debated — many find systems reply compelling
Brain Simulator Reply: If the program simulates individual neurons, would it understand?	Simulating neurons is not the same as having neurons. A simulated storm doesn't make you wet.	Debated — depends on theory of consciousness
Robot Reply: Connect the system to sensors and actuators — grounded meaning would emerge	Adding I/O is just more symbol manipulation. The grounding problem remains.	Partial concession — embodiment may matter

Consciousness & Machine Sentience Core

💭

The Hard Problem of Consciousness

Chalmers (1995) distinguishes "easy problems" (explaining cognitive functions: attention, memory, behaviour — all tractable in principle) from the "hard problem": why is there subjective experience at all? Why does information processing feel like something from the inside?

Even a complete functional account of the brain wouldn't explain why there's a "what it's like to be" that brain. This is the central unsolved problem in philosophy of mind — and applies directly to AI sentience claims.

🌐

Integrated Information Theory (IIT)

Tononi (2004): consciousness corresponds to integrated information (Φ, "phi"). A system is conscious to the degree its parts share information in an irreducibly integrated way. Higher Φ = more conscious.

Implication: feedforward networks (including transformers, which have no recurrent loops) have Φ ≈ 0. If IIT is correct, current LLMs are not conscious — not even slightly.

Frames of Intelligence Reference

🧩

Multiple Intelligences (Gardner)

Eight distinct intelligences: linguistic, logical-mathematical, spatial, musical, bodily-kinaesthetic, interpersonal, intrapersonal, naturalistic. AI today dominates the first two; is largely absent from the last five.

🤸

Embodied Cognition

Intelligence is shaped by having a body that interacts with the world. Physical AI (robotics + LLMs) is the active frontier precisely because disembodied language models lack grounding in physical causality.

🏛️

Cognitive Architectures

ACT-R (Anderson) and SOAR (Laird, Newell) are computational models of human cognition with procedural memory, declarative memory, and attention modules — making testable predictions verified against human reaction time data.

The Symbol Grounding Problem Core

Harnad's Argument (1990)
How do symbols get their meaning? In a dictionary, words are defined by other words — circular. A child grounds symbols in perceptual experience: "red" is grounded in actually seeing red things. Harnad argued that purely symbolic AI can never escape this circularity — symbols need to be grounded in the world, not just in other symbols.

🔗

How Neural Nets Partially Address This

Multimodal models (CLIP, GPT-4V, Gemini) ground language in images — "red" is statistically associated with red pixel patterns. This is partial grounding. But the model never sees red in the physical sense — it sees statistical co-occurrence in training data. Whether this constitutes genuine grounding is disputed.

🤔

Implications for LLMs

Text-only LLMs have no sensory grounding. Their concept of "pain" is the statistical distribution of the word "pain" in training data. This may explain why LLMs discuss concepts fluently but fail tasks requiring genuine understanding of physical causality, embodied experience, or common-sense spatial reasoning.

1.5

Chapter 1.5 · AIMA Framework

Problem-Solving & Rational Agents

Russell & Norvig's Artificial Intelligence: A Modern Approach (AIMA) defines AI as the study of agents that perceive their environment and act to maximise their performance measure. This framework is the conceptual backbone of Domain 8 (Agentic AI) — understanding it now is essential.

The Rational Agent Framework In-depth

The Rational Agent — Perceive → Decide → Act Loop

Definition: A rational agent selects actions expected to maximise its performance measure, given the evidence in its percept sequence and built-in knowledge.

Rationality ≠ omniscience (knowing all outcomes). Rationality ≠ perfection (always choosing optimally). Rationality = expected utility maximisation given available information.

PEAS	Definition	Self-Driving Car	LLM Agent
Performance Measure	What "doing well" means — the objective	Safety, speed, comfort, law compliance	Task completion, accuracy, user satisfaction
Environment	Everything the agent interacts with	Roads, vehicles, pedestrians, weather	Web pages, APIs, files, databases, other agents
Actuators	Mechanisms for taking action	Steering, brakes, accelerator	Code execution, web search, API calls, text output
Sensors	Mechanisms for perceiving the environment	Cameras, LiDAR, GPS, radar	Context window, tool outputs, memory retrieval

Environment Types Core

Property	Type A	Type B	Example (A)	Example (B)
Observability	Fully observable	Partially observable	Chess (full board visible)	Poker (opponent's cards hidden)
Determinism	Deterministic	Stochastic	Chess (outcomes fully determined)	Autonomous driving (weather, pedestrians)
Episodicity	Episodic	Sequential	Image classification (each independent)	Chess (moves affect future states)
Dynamics	Static	Dynamic	Crossword puzzle	Stock trading, real-time robotics
Continuity	Discrete	Continuous	Chess (finite legal moves)	Robot arm control (infinite positions)
Agents	Single-agent	Multi-agent	Sudoku solver	Multiplayer games, multi-agent AI systems

Why environment type matters for design: A fully observable, deterministic, episodic, discrete environment (like chess) can in principle be solved with a lookup table. A partially observable, stochastic, sequential, continuous, multi-agent environment (like autonomous driving) requires probabilistic reasoning, memory, planning, and robustness to uncertainty — all simultaneously. The gap in engineering complexity is enormous.

Types of Agents In-depth

Agent Sophistication Spectrum — From Reflex to Learning

🔦

Simple Reflex Agent

IF-THEN rules that map current percepts directly to actions. No memory. No history. Example: thermostat (IF temp < setpoint THEN heat on). Fast and auditable but completely brittle — fails whenever current percept doesn't capture all relevant state.

🗺️

Model-Based Reflex Agent

Maintains an internal state that tracks the world — knows what it can't currently perceive. Can handle partial observability. Example: a robot vacuum that tracks which areas have been cleaned even when not currently there.

🎯

Goal-Based Agent

Has an explicit goal and searches for action sequences to achieve it. Uses search algorithms (A*, BFS) to plan. More flexible than reflex agents — multiple paths to the same goal. Requires a model of how actions change the world.

📊

Utility-Based Agent

Uses a utility function — a graded preference ordering over states, not just goal/no-goal. Chooses actions maximising expected utility. Handles stochastic outcomes naturally. Foundation of decision theory. Modern RL agents and LLM agents with reward models are utility-based.

🎓

Learning Agent

Any agent type augmented with a learning component that modifies behaviour based on experience. Composed of: learning element (improves performance), performance element (selects actions), critic (evaluates against a standard), problem generator (suggests exploratory actions). All modern AI systems are learning agents.

State-Space Problem Formulation Core

Element	Definition	Example: 8-Puzzle
Initial state	Starting configuration	Random tile arrangement
Actions	Set of possible moves from each state	Slide tile left, right, up, or down
Transition model	Result of taking each action in each state	New tile arrangement after sliding
Goal test	Determines if current state is goal	Tiles in order 1-2-3-4-5-6-7-8-blank
Path cost	Numeric cost of a path	Number of moves taken

Why search-based AI was replaced: The 8-puzzle has ~181,000 reachable states. The 15-puzzle has ~10¹². Chess: ~10⁴⁶. Go: ~10¹⁷⁰. Real-world AI problems are combinatorially intractable for exhaustive search — which is precisely why machine learning (pattern recognition from data) displaced search as the dominant paradigm for perception, language, and unstructured reasoning tasks.

1.6

Chapter 1.6 · Schools of Thought

Key Paradigms & Schools of Thought

AI is not one field — it is a collection of competing intellectual traditions with different assumptions about what intelligence is and how to build it. Understanding the camps explains why researchers from different traditions argue past each other, and why hybrid approaches are gaining traction.

Connectionism vs. Symbolicism In-depth

Dimension	Symbolicism	Connectionism	Current Status
Core claim	Intelligence = symbol manipulation under logical rules	Intelligence emerges from densely connected simple units	Connectionism dominant (2012–present)
Key figures	McCarthy, Minsky, Newell, Simon	Rosenblatt, Rumelhart, Hinton, LeCun, Bengio	Hinton, LeCun, Bengio won 2018 Turing Award
Strengths	Interpretable, logically consistent, handles explicit structured knowledge	Learns from data, generalises to new inputs, handles noise	Both needed; neither sufficient alone
Weaknesses	Brittle, doesn't scale, knowledge acquisition bottleneck	Black box, data-hungry, fails on distribution shift	Interpretability and robustness remain unsolved
Modern form	Knowledge graphs, formal verification, constraint solvers	Large language models, diffusion models, transformers	Hybrid (neuro-symbolic) is active research frontier

Neuro-Symbolic AI — The Hybrid Frontier
Systems combining neural pattern recognition with symbolic reasoning. AlphaGeometry (2024): uses a neural model to generate proof steps and a symbolic geometry solver to verify them — solves IMO-level geometry problems. Program synthesis: neural nets generate code candidates; symbolic testing verifies correctness. This is the current research direction for systematic generalisation beyond interpolation.

Probabilistic & Bayesian AI Core

🎲

Bayesian AI

Models uncertainty explicitly using probability distributions. Bayes' theorem updates beliefs when new evidence arrives: P(H|E) ∝ P(E|H) × P(H). Principled framework for reasoning under uncertainty — but computing exact posteriors is often intractable, requiring approximations (MCMC, variational inference).

Applications: spam filtering, medical diagnosis, sensor fusion in robotics, weather forecasting

🔗

Probabilistic Graphical Models

Bayesian networks: directed acyclic graphs where nodes are random variables and edges are conditional dependencies. Efficient inference in structured domains. Hidden Markov Models (HMMs): sequences of hidden states with observable outputs. Dominated speech recognition from 1970s–2010s until deep learning.

Still widely used in robotics, bioinformatics, and probabilistic programming (Stan, Pyro)

Evolutionary Computation Reference

Inspired by biological evolution. Genetic algorithms maintain a population of candidate solutions, select for fitness, recombine (crossover) and mutate. Used for optimisation where the landscape is rugged and gradient-based methods fail. Active in neural architecture search (NAS), hyperparameter optimisation, and hardware co-design.

Embodied & Situated AI Reference

Rodney Brooks (MIT) argued in the 1980s that traditional AI was fundamentally misguided — intelligence doesn't require symbolic internal representations. His subsumption architecture layered simple reactive behaviours to produce surprisingly complex robot behaviour. "Intelligence without representation." Now re-emerging: physical AI labs (Figure, Boston Dynamics, 1X) combine embodied robotics with LLM reasoning.

Cognitive Architectures Reference

ACT-R (Anderson): models human cognition with procedural memory, declarative memory, and attention modules — makes testable predictions verified against reaction time data. SOAR (Newell, Laird, Rosenbloom): unified theory of cognition with problem spaces, operators, and impasse-driven learning. Both influenced the architecture of modern AI agents in Domain 8. OpenCog: open-source AGI architecture combining probabilistic logic networks with deep learning.

1.7

Chapter 1.7 · Current Landscape

The AI Landscape Today & Tomorrow

This chapter connects Domain 1's history and theory to the present, and gives you a map of where each subsequent domain fits. AI in 2026 is defined by foundation models as the new default paradigm — but with hard limitations that motivate everything that follows in this curriculum.

The Current AI Landscape Core

🏛️

Foundation Models — The New Paradigm

Large models pre-trained on internet-scale data, adaptable to almost any task. The shift: from task-specific models trained from scratch → general-purpose models fine-tuned for tasks. One model (GPT-4o, Claude 3.7, Gemini 1.5) outperforms hundreds of specialised predecessors.

Open weights: LLaMA 3, Mistral, Qwen, DeepSeek-V3, Phi-3
Closed API: GPT-4o, Claude 3.7, Gemini 1.5 Pro, Command R+
Specialised: Codex/StarCoder (code), Whisper (audio), SAM (segmentation)

⚖️

Open vs. Closed Source

Closed: Better safety filtering, proprietary advantage, pay-per-use, rate limits

Open weights: Full control, privacy, fine-tuning, zero inference cost, safety risk if misused

Compute scaling trends: training cost for frontier models grows ~4× per year. GPT-4: ~$100M. Next-generation frontier: estimated $1B+.

AI Capability Taxonomy Core

Capability	What It Means	State of the Art (2026)	Domain Coverage
Perception	Interpreting sensory data (vision, audio, text)	Superhuman on ImageNet; near-human speech; strong OCR	Domain 6 (CV), Domain 5 (NLP)
Reasoning	Drawing valid inferences from information	Good at formal logic; poor at common-sense causality	Domain 5 (LLMs), Domain 1 (logic)
Planning	Generating action sequences toward goals	Improving with o1/o3 reasoning models; still brittle at scale	Domain 8 (Agents), Domain 7 (RL)
Generation	Creating novel text, images, audio, code, video	Human-competitive in text/image; superhuman in code assist	Domain 5 (LLMs), Domain 6 (CV)
Action	Taking actions in physical or digital environments	Early stage; computer-use agents emerging; physical robots	Domain 8 (Agents), Domain 7 (RL)

Benchmarks & Measuring Progress In-depth

Goodhart's Law applied to AI benchmarks: "When a measure becomes a target, it ceases to be a good measure." AI benchmarks are routinely gamed — models are fine-tuned on test distributions, inflating scores without genuine capability improvement. Benchmark saturation happens faster than the underlying capability warrants. Always ask: what does this benchmark actually measure, and what can't it measure?

Benchmark	Domain	Current Status	Key Concern
ImageNet (ILSVRC)	Computer Vision	Models exceed human ~97%+	Real-world robustness far lower; adversarial examples break all models
GLUE / SuperGLUE	NLP	Saturated — models exceed human baseline	Saturated within 2 years of creation; replaced by harder benchmarks
MMLU	Knowledge (57 subjects)	GPT-4 ~87%; Claude 90%+	Multiple-choice is gameable; doesn't test applied or procedural knowledge
BIG-Bench Hard	Reasoning	Frontier models pass most tasks	Being saturated; successor benchmarks in development
HumanEval / SWE-bench	Code	~90% HumanEval; ~50% SWE-bench	SWE-bench is more realistic but still a limited test suite of GitHub issues
ARC-AGI	Abstract Reasoning	~75–80% best models (2025)	Designed to resist pattern matching; remains the hardest general reasoning eval

Open Problems in AI In-depth

🧩

Common-Sense Reasoning

Humans know ice cream melts in sun, dropped objects fall, elephants don't fit in cars — without being told. LLMs absorbed much of this from text but fail unpredictably on novel physical scenarios. The Winogrande benchmark reveals systematic errors humans find trivial. Commonsense is the gap between pattern matching and genuine world understanding.

🔀

Causal Inference

Current AI is fundamentally correlational. Correlation ≠ causation. "Countries with more hospitals have more disease" — a correlational model would conclude hospitals cause disease. Judea Pearl's causal hierarchy (association → intervention → counterfactual) identifies exactly what's missing. No current neural system reliably reasons causally from observational data.

🗓️

Long-Horizon Planning

Current agents reliably handle ~10–50 step tasks. Real-world projects span hundreds of interdependent steps over days or weeks. Errors compound: one wrong action early invalidates downstream planning. Agents also suffer context window limitations (memory) and lack persistent state across sessions.

🎯

Sample Efficiency & Robustness

Humans learn from a handful of examples; LLMs require billions. Sample efficiency — learning more from less data — is largely unsolved. Robustness — consistent performance under distribution shift — is equally unsolved. Both are critical for safety-critical deployment (medicine, autonomous vehicles, infrastructure).

📋 Domain 01 — Key Takeaways

AI = systems that perceive, reason, learn, and act — not a single technology; a collection of competing approaches
All current AI is ANI (Narrow). AGI timelines are genuinely debated; ASI is theoretical. Precision about which level you mean matters enormously.
AI history is a cycle of overpromise → collapse → breakthrough. The current wave is real — but understanding the winters prevents misreading today's hype.
AI winters were caused by combinatorial explosion + knowledge acquisition bottleneck + overpromising timelines — not bad science
Symbolic AI failed to scale but wasn't wrong — it lives on in knowledge graphs, formal methods, and neuro-symbolic hybrids
The Turing Test, Chinese Room, and symbol grounding problem are actively debated in the context of LLMs — they motivate alignment, interpretability, and AI rights discourse
The PEAS framework (Performance, Environment, Actuators, Sensors) formally describes any agent — including modern LLM-based agents in Domain 8
Connectionism dominates today; Bayesian, evolutionary, and embodied approaches remain active and complementary
Benchmarks are routinely gamed — benchmark saturation does not mean a capability is solved. Always interrogate what the benchmark actually measures.
Hard open problems: common-sense reasoning, causal inference, long-horizon planning, sample efficiency, robustness

← Foundation Overview Domain 02: Mathematics →