What "AI" Means, and How We Got Here

"AI" is a label, and a slippery one. The spam filter in your email is AI. So is the thermostat that learns when you get home, the camera that reads license plates, the app that turns your voice memo into text, and the chatbot that will write a eulogy for your goldfish in the style of a sea shanty. These machines work in very different ways and fail in very different ways. If you follow AI as a public issue, or just use it, the first skill is telling them apart.

Here is a working definition that holds up: artificial intelligence is the name for computer systems that do jobs we used to think needed human judgment. Recognizing a face. Ranking which job applicant to call first. Translating Spanish. Writing a paragraph. The definition moves over time. Nobody calls a calculator AI anymore, though in 1950 a machine doing long division looked like magic.

Two ways to make a computer act smart

Almost everything in AI history falls into one of two camps.

The first camp writes rules. A person who knows the subject sits down and spells out the logic: if the patient has a fever and the culture shows this bacterium, then suggest that drug. The computer follows the rules exactly and fast. It never learns anything new unless a person adds a rule.

The second camp shows examples. Instead of writing the rule for spotting spam, you hand the computer 100,000 emails already marked spam or not spam and let it work out which patterns predict which label. That is machine learning. Nobody types the rule. The rule is found in the data, and it is stored as a pile of numbers that no human wrote.

A short history, in eras

1956: the name

The field got its name at a 1956 summer workshop at Dartmouth College organized by John McCarthy. His proposal bet that every feature of intelligence could be described precisely enough for a machine to copy it. That bet turned out to be harder than a summer.

1960s to 1980s: rules and expert systems

Early AI was mostly logic and rules. The peak was the expert system. MYCIN, built at Stanford in the early 1970s, used about 600 hand-written rules to identify bacteria behind serious infections and suggest antibiotics, and it could explain why it asked each question. It performed well in tests and was never used in routine care, partly because nobody could settle who was responsible when a computer gave medical advice. In the 1980s companies spent heavily on expert systems for loan approval, equipment repair and the like.

Rules have a ceiling. The world has too many exceptions, and every new exception needs a person to write a new rule. When the promises outran the results, money dried up. Those dry spells are called AI winters, and there were two big ones, in the 1970s and the late 1980s.

1986 to 2000s: learning from data

In 1986 David Rumelhart, Geoffrey Hinton and Ronald Williams published a paper in Nature on backpropagation, a practical way to train neural networks with several layers. Computers were too slow and data too scarce for it to shine yet. Meanwhile simpler machine learning methods quietly took over real jobs: spam filters, credit card fraud alerts, the Netflix and Amazon recommendation engines.

The famous 1997 win of IBM's Deep Blue over chess champion Garry Kasparov was mostly brute force. IBM says it evaluated 200 million positions per second, scored with rules tuned by chess experts. Impressive engineering. Not much learning.

2012: deep learning arrives

The moment researchers point to is the 2012 ImageNet contest, where programs sorted photos into 1,000 categories. A team from the University of Toronto (Alex Krizhevsky, Ilya Sutskever and Hinton) entered a neural network with 60 million adjustable numbers, trained on gaming graphics cards. Its top-5 error was about 15.3 percent. Second place was about 26.2 percent. In a contest usually won by a point or two, that gap ended the argument. Deep learning, meaning neural networks with many layers, became the default for images, then speech, then language.

In 2016 DeepMind's AlphaGo, which learned from human games and then from playing itself, beat top Go players, a game long thought too large for computers. In 2024 John Hopfield and Geoffrey Hinton shared the Nobel Prize in Physics for foundational work on machine learning with artificial neural networks.

2017: the transformer

In June 2017 eight researchers, most at Google, posted a paper called "Attention Is All You Need." It described the transformer, a network that looks at a whole passage at once and works out which words matter to which other words. Older language networks read one word at a time, like a person running a finger along a line. The transformer could train in parallel, which meant it could use far more data and far more chips. The paper's biggest translation model trained in 3.5 days on eight GPUs and beat the previous best scores. Every major chatbot today is a descendant. The T in GPT stands for transformer.

2022: the ChatGPT moment

OpenAI released ChatGPT on November 30, 2022. Under the hood it was a GPT-3.5 model fine-tuned with human feedback, built on research published earlier that year. What was new was the package: a free chat window anyone could type into. Analysts at UBS estimated it reached 100 million monthly users in about two months, which would make it the fastest growing consumer app on record at the time. That number is an outside estimate from analysts. Nobody disputes the stampede.