← Reference · Nestor G Pestelos Jr
Reference Document
Artificial Intelligence
A citable reference on Artificial Intelligence (AI): the rational agent paradigm, historical taxonomy, statistical foundation models, and autonomy bounds.
See Also & Related References
- 📖 Reference: Machine Learning — Empirical risk minimization, learning paradigms, and loss optimization.
- 📖 Reference: Deep Neural Networks — Multi-layer composite functions, non-linear activations, and backpropagation.
- 📖 Reference: Large Language Models (LLMs) — Transformer decoders, autoregressive tokens, and production gateways.
- 🌳 LLM System Design Knowledge Tree — Prerequisite curriculum map and mastery frontier.
Jump to Section
1. Definition: The Rational Agent Paradigm
Artificial Intelligence (AI) is the computational study and engineering of systems that perceive their environment and execute actions that maximize the expected probability of successfully achieving defined goals (Russell & Norvig, 2020).
Modern technical definitions reject superficial "human-likeness" (the Turing test framing) in favor of rationality: an agent is rational if it selects an action \(a \in A\) that maximizes expected utility given its percept history \(P\) and internal knowledge base:
$$a^* = \arg\max_{a \in A} \mathbb{E}\big[U(s') \mid P, a\big]$$
2. The Hierarchy of Artificial Intelligence
Artificial Intelligence is a broad umbrella discipline containing progressively specialized nested subfields:
| Tier | Definition & Scope | Primary Mechanism |
|---|---|---|
| Artificial Intelligence (AI) | The overarching field of building goal-directed agents that perceive, reason, plan, and act. | Search algorithms, formal logic, knowledge graphs, heuristic optimization, machine learning. |
| Machine Learning (ML) | A subfield of AI focused on algorithms that learn statistical parameters from data without explicit manual rules. | Empirical risk minimization, loss optimization, decision trees, gradient descent. |
| Deep Learning (DL) | A subfield of ML utilizing multi-layered neural networks (\(L \ge 2\)) for automatic representation learning. | Backpropagation, non-linear activation functions (ReLU, GELU), GPU tensor operations. |
| Generative Foundation Models (LLMs) | Massive deep networks trained self-supervised on internet-scale corpora for cross-domain generation and reasoning. | Self-attention transformers, autoregressive next-token prediction, RLHF alignment. |
3. Historical Paradigm Shifts
- Symbolic AI / GOFAI (1950s–1980s): Explicit rule-based systems, first-order logic inference engines, and heuristic tree searches (e.g. A* search, Expert Systems). Fragile to ambiguity and combinatorial explosion.
- Statistical Machine Learning (1990s–2000s): Shift to probabilistic formulations, Support Vector Machines (SVMs), Random Forests, and Bayesian networks with manual feature engineering.
- Deep Learning Era (2012–2020): Breakthrough of large-scale backpropagation on GPUs (AlexNet, ResNet, Transformer). Replaced manual feature engineering with end-to-end representation learning.
- Foundation Models & Agentic AI (2020s–Present): Emergence of billion-parameter autoregressive models capable of general in-context reasoning, tool orchestration, and autonomous execution loops.
4. The PEAS Agent Environment Framework
In formal AI design (Russell & Norvig), every agent system is specified by its PEAS descriptors:
- Performance Measure (P): The mathematical utility function \(U\) defining success (e.g., latency, accuracy, profit, safety constraints).
- Environment (E): The operational domain (observable vs. partially observable, deterministic vs. stochastic, static vs. dynamic, discrete vs. continuous).
- Actuators (A): Mechanisms through which the agent acts (e.g., API calls, shell execution, database writes, physical motors).
- Sensors (S): Mechanisms through which the agent receives percepts (e.g., HTTP payloads, file streams, cameras, microphones).
5. Major Subfields of AI
| Subfield | Core Problem | Key Techniques |
|---|---|---|
| Natural Language Processing (NLP) | Understanding, translating, and generating human language. | Autoregressive Transformers, tokenizers, dense vector embeddings. |
| Computer Vision (CV) | Extracting structured geometric and semantic meaning from pixels. | Convolutional networks (CNNs), Vision Transformers (ViT), diffusion models. |
| Planning & Multi-Agent Systems | Coordinating multi-step action sequences across competing/cooperating agents. | Hierarchical Task Networks (HTN), Monte Carlo Tree Search (MCTS), Game Theory. |
| Robotics & Control | Embodied physical interaction in continuous dynamic environments. | Model Predictive Control (MPC), Reinforcement Learning, Sim-to-Real transfer. |