AI Introduction

What "artificial intelligence" means#

Artificial intelligence (AI) is the broad field of building systems that perform tasks which normally require human intelligence — understanding language, recognizing images, making decisions under uncertainty, planning, playing games. That's a deliberately wide definition, because AI as a field spans everything from a chess-playing program with hand-coded rules from the 1960s to a modern model that writes code. The unifying thread isn't a specific technique — it's the goal: get a machine to do something that looks intelligent.

Because the definition is about the goal rather than the method, AI is best understood as a set of nested subfields, each narrower and more specific than the last.

The nested scope: AI ⊃ ML ⊃ deep learning ⊃ LLMs#

It helps to picture this as concentric circles, from broadest to narrowest:

┌───────────────────────────────────────────────────────┐
│ Artificial Intelligence                                │
│  (any technique that produces intelligent behavior)     │
│  ┌─────────────────────────────────────────────────┐   │
│  │ Machine Learning                                  │   │
│  │  (systems that learn patterns from data,          │   │
│  │   rather than following hand-coded rules)         │   │
│  │  ┌───────────────────────────────────────────┐   │   │
│  │  │ Deep Learning                               │   │   │
│  │  │  (ML using multi-layer neural networks)     │   │   │
│  │  │  ┌─────────────────────────────────────┐   │   │   │
│  │  │  │ Large Language Models (LLMs)          │   │   │   │
│  │  │  │  (deep learning applied to text,      │   │   │   │
│  │  │  │   trained on next-token prediction)   │   │   │   │
│  │  │  └─────────────────────────────────────┘   │   │   │
│  │  └───────────────────────────────────────────┘   │   │
│  └─────────────────────────────────────────────────┘   │
└───────────────────────────────────────────────────────┘

• AI is the field. A rule-based tax-filing wizard built from a thousand if/else statements is AI — it produces useful, seemingly intelligent behavior, even though nothing in it "learns" anything.
• Machine learning (ML) is a subset of AI: systems that improve at a task by learning statistical patterns from data instead of being explicitly programmed with rules for every case. A spam filter trained on thousands of labeled emails is ML.
• Deep learning is a subset of ML: specifically, ML done with multi-layer artificial neural networks. It's what made computer vision, speech recognition, and language modeling take a huge leap forward starting around 2012, because it can automatically learn useful features from raw data (pixels, audio waveforms, text) instead of requiring a human to hand-engineer them first.
• Large language models (LLMs) are a specific, recent application of deep learning: neural networks (almost always the transformer architecture) trained on enormous amounts of text to predict the next piece of text, which turns out to produce a system that can converse, summarize, translate, and reason about language.

Every LLM is deep learning; not all deep learning is language modeling (image classifiers and speech-to-text systems are deep learning too, but not LLMs). Every deep learning system is machine learning; not all machine learning uses neural networks (a decision tree or linear regression model is ML but not deep learning). Every machine learning system is AI; not all AI learns from data (a hand-coded rules engine is AI but not ML).

A brief history#

Symbolic AI (1950s–1980s). Early AI assumed intelligence could be captured as logic: explicit rules and symbol manipulation. Expert systems encoded human knowledge as if-then rules (for example, medical diagnosis systems with thousands of hand-written rules). This worked for narrow, well-understood domains but didn't scale — nobody can hand-write enough rules to cover the messiness of natural language or vision. This limitation, and the funding drought that followed, is often called the "AI winter."

Statistical machine learning (1990s–2000s). Instead of hand-coding rules, researchers built systems that learned patterns from labeled examples using statistics and probability — spam filters using naive Bayes, fraud detection with decision trees, handwriting recognition with support vector machines. These methods needed far less manual rule-writing, but they typically required a human to hand-engineer the input features (for example, "does this email contain the word 'free'?") before the algorithm could learn from them.

Deep learning (2010s). Neural networks had existed since the 1950s, but three things converged around 2012 to make them dramatically more effective: much larger labeled datasets, much cheaper parallel compute (GPUs), and algorithmic refinements. Deep networks could now learn useful features directly from raw pixels or audio, without a human hand-engineering them first. This produced step-change improvements in image recognition, speech recognition, and machine translation.

Transformers and LLMs (2017–present). A 2017 architecture called the transformer (built around a mechanism called self-attention) turned out to scale exceptionally well with more data and more compute, and to be far more parallelizable to train than the recurrent networks that came before it. Applied to text at massive scale, it produced large language models — a single general-purpose model that, unlike earlier task-specific systems, can perform translation, summarization, question-answering, and code generation without being separately built for each task.

Common mistakes#

• Using "AI" and "machine learning" interchangeably in technical contexts — a rules engine is AI but not ML, and the distinction matters when scoping a project (do you actually need labeled data and training, or would explicit rules be simpler and more maintainable?).
• Assuming "deep learning" and "AI" are synonyms — many production systems (fraud rules, pricing logic, basic recommendation heuristics) are AI without touching a neural network at all.
• Treating LLMs as the endpoint of AI rather than one (currently very prominent) branch of it — computer vision, robotics, and classical ML are separate, still-active branches that don't run through a language model at all.