A Word That Means Too Much
Say "AI" and everyone pictures something different: a chatbot that writes your emails, a car that drives itself, a sci-fi machine plotting against humanity, a phone that recognizes your face. These can't all be the same thing — yet we reach for one word. So before any math, we need a sharper question: what is artificial intelligence actually trying to do, and how is it different from the ordinary software already running everywhere around you?
As a field, AI has a simple, ambitious aim: build machines that do things which, when a human does them, we call intelligent — recognizing a face, understanding a sentence, planning a route, learning from a mistake. The field was named at a 1956 summer meeting (the Dartmouth Workshop), where a handful of researchers proposed that "every aspect of learning or intelligence can in principle be so precisely described that a machine can be made to simulate it." That bold bet is the whole adventure. Whether it is fully true is still open — but it has carried us a remarkably long way.
How AI Differs From Ordinary Software
Ordinary software follows rules a person wrote out in advance. "If the cart total is over $50, subtract $5" — a human decided that rule, typed it in, and the computer obeys it exactly, forever. This is wonderful when the rules are knowable. But how would you write the rules for "is this photo a cat?" Try, and you drown: cats come in a thousand colors, poses, lightings, and breeds, half-hidden behind furniture. Nobody can spell out every case in advance.
The deepest current in modern AI flips this around. Instead of writing the rules, you show the machine thousands of examples — photos already labeled "cat" or "not cat" — and let it find the pattern on its own. That is machine learning: the program is not told the rule, it learns the rule from data. The shift sounds small but it is profound — we go from programming the answer to programming the way to discover the answer.
Three Nested Circles: AI, ML, Deep Learning
The three words people mix up most are best pictured as circles inside circles. Artificial intelligence is the big outer circle: any technique that makes machines act smart, rules and learning alike. Inside it sits machine learning: the subset that learns from data instead of being hand-programmed. And inside that sits deep learning: machine learning done with neural networks that stack many layers, loosely inspired by how brains wire neurons together.
So all deep learning is machine learning, and all machine learning is AI — but not the other way around. A chess engine that searches moves by brute force is AI but not machine learning. A simple spam filter that learns from labeled emails is machine learning but not deep learning. And the chatbots everyone talks about — built on a kind of large language model — are deep learning, the innermost circle. Keeping the circles straight saves you from a lot of confused headlines.
Artificial Intelligence (make machines act smart)
└─ Machine Learning (learn the rules from data)
└─ Deep Learning (learn with many-layered neural networks)
examples:
chess by brute-force search ... AI only
spam filter from examples ..... AI + ML
image / language models ....... AI + ML + Deep LearningThis nesting also tells a history. Early AI leaned on symbolic rules and search; the learning-from-data view, called connectionism, was sidelined for years, even through funding droughts known as the AI winters when promises outran results. Only when fast computers and oceans of data arrived did deep learning surge to the center. The lesson many took from this — that general methods which scale with computation tend to win out over hand-built cleverness — even has a name: the bitter lesson.
Narrow vs General: One Trick or Many?
Here is the single most important distinction for keeping your expectations honest. Almost every AI you will ever meet is narrow AI: brilliant at one specific job and helpless outside it. A program that crushes grandmasters at chess cannot tie a shoelace, hold a conversation, or even play checkers. A medical model that spots tumors in scans has no idea what a tumor is, what a patient is, or that anything exists beyond its narrow lane. Narrow does not mean weak — these systems are often wildly superhuman at their one task. They just can't transfer that skill to anything new on their own.
The contrast is artificial general intelligence (AGI): a hypothetical machine that could learn and reason across any domain the way a person can, picking up a new skill it was never built for. It does not exist. Today's chatbots feel general because they can chat about almost anything, but underneath they are doing one narrow thing extremely well — predicting plausible next words — not understanding the world the way you do. The flexible feel is real and useful; mistaking it for genuine general understanding is the most common modern overclaim.
What Does "Intelligence" Even Mean Here?
Philosophers have argued about "intelligence" forever, so AI mostly sidesteps the question and uses an operational test: never mind what's going on inside — can the machine produce the behavior we'd call intelligent? In 1950 Alan Turing proposed exactly this. His Turing test asks not "can a machine think?" but "can it converse so well that a human can't tell it from a person?" Judge by behavior, not by some inner spark we can't measure.
A useful working lens many practitioners adopt is the agent view: an intelligent system is anything that perceives its situation and acts to achieve a goal. By that yardstick a thermostat is a (very dim) agent and a self-driving car is a sophisticated one. "Intelligence" becomes a dial — how well an agent reaches its goals across how many situations — not a yes/no badge. This is deliberately humble, and it's a big reason the field makes progress instead of getting stuck on definitions.
Where You Already Meet AI — and Its Limits
AI is no longer exotic; it quietly threads through your day. The spam filter sorting your inbox learned from millions of labeled emails — that's supervised learning, learning from examples with the right answers attached. The route your map app picks, the photos your phone groups by face, the songs a service queues next, the autocomplete finishing your sentence — each is a narrow model trained for one job. The most visible recent face is the chatbot, powered by a foundation model trained on a huge sweep of text and then adapted to many tasks.
The landmark wins fit the same pattern. Deep Blue beat the world chess champion in 1997 mostly by calculating millions of positions — closer to brute force than learning. Two decades later AlphaGo mastered Go, a game too vast to brute-force, by learning from games and from playing itself. Both are stunning — and both are narrow: neither can do anything but its one game. That gap between dazzling specialist skill and zero general competence is the signature of all AI today.
So end with an honest ledger. Today's AI is genuinely transformative at recognizing patterns, generating fluent text and images, and optimizing within well-defined goals. It is also brittle: it can fail in childish ways just outside its training, it has no common sense or grasp of cause and effect, and language models routinely produce hallucinations — confident, fluent statements that are simply false. The rest of this ladder builds the real machinery beneath these systems, so you can tell the genuine power from the hype — and there is plenty of both.