AI Basics

What is AI?

Artificial intelligence is software that learns from examples rather than following hand-written rules. A traditional spam filter might be programmed with explicit rules — flag any email containing the word "free." An AI-powered spam filter learns what spam looks like by studying millions of examples, then keeps improving as it sees more. That shift, from writing rules by hand to learning patterns from data, is the core idea behind modern AI.

Most of what people call "AI" today is actually a stack of related ideas. They nest inside each other:

Machine Learning

Instead of a programmer writing rules, a machine learning model is shown thousands — or millions — of examples and figures out the patterns on its own. Show it 50,000 labeled photos of cats and dogs, and it will learn to tell them apart without anyone writing a rule about ear shape or whiskers. The same approach works for detecting fraud, predicting equipment failure, and recommending what to watch next.

There are two main flavors: supervised learning (learning from labeled examples — "this is a cat, this is not") and unsupervised learning (finding patterns in unlabeled data on its own — used for clustering and anomaly detection).

Deep Learning & Neural Networks

Deep learning stacks many layers of pattern-matching on top of each other. The first layer might learn edges, the second textures, the third shapes, the fourth faces. Each layer builds on the one below. "Deep" just means the network has many layers — and it turns out depth is what unlocks the really hard problems: understanding images, transcribing speech, and reading language.

A neural network is the architecture that makes this possible. It's a mesh of simple mathematical units — neurons — organized into layers. Input enters from the left, output exits on the right, and signals pass through every layer in between. During training, the strengths of the connections between neurons are adjusted millions of times until the network reliably produces correct outputs.

Signals (pulses) propagate from the input layer → through a hidden layer → to output nodes. Each connection weight is adjusted during training.

Natural Language Processing

Natural language processing (NLP) is the branch of AI focused on understanding and generating human language. Early NLP relied on hand-crafted grammars and dictionaries — systems that worked in narrow domains but broke down on informal, messy, real-world text. Modern NLP uses deep learning to treat language as sequences of tokens and learn statistical relationships between them at massive scale. The result is AI that can summarize documents, answer questions, translate between languages, and hold a conversation.

Large Language Models

A large language model (LLM) is a deep learning model trained specifically on text — trained to predict the next word given everything that came before. That's a deceptively simple objective. Scaled up with enough data and compute, it produces systems capable of reasoning, writing, coding, and summarizing across virtually any topic. GPT-4, Claude, and Gemini are all LLMs.

The underlying architecture — the transformer, introduced by Google in 2017 — is what made this possible. Transformers can process long sequences of text in parallel and learn which parts of a sequence are relevant to each other, a mechanism called attention. For a closer look at how this all works under the hood, see Decoding AI Conversations. And if you want to use these models more effectively day-to-day, Prompt Engineering 101 covers the practical techniques that make the biggest difference.

Computer Vision

Computer vision enables machines to interpret images and video. It's one of deep learning's great success stories — the 2012 ImageNet breakthrough (where a deep neural network dramatically outperformed all prior approaches to image classification) is widely credited with igniting the current AI era. Today, computer vision powers medical imaging, autonomous vehicles, manufacturing quality control, and the face unlock on your phone.

---