🧠Artificial Intelligence (AI)

Artificial intelligence is software that performs tasks which normally require human judgment: recognizing a face, understanding a sentence, ranking search results, flagging a fraudulent charge. That is the whole definition. No consciousness, no magic, no robot uprising: a program that handles a judgment-shaped task counts as AI.

Two ways to build it

There have only ever been two basic approaches. Rules: a human writes the logic by hand, if the email contains this phrase, flag it. Rule systems are predictable and auditable, but brittle: nobody can write rules for every way a cat can appear in a photo. Learning: instead of writing rules, you show the program thousands of examples and let it adjust internal numbers until its guesses become accurate. Nearly everything called AI today, from photo search to ChatGPT, is the second kind: machine learning.

You already use it every day

Your spam filter learned from millions of labeled emails. Your phone keyboard predicts the next word with a small language model. Maps picks routes with learned traffic patterns. Streaming services rank what you might watch next. Credit card networks score every swipe for fraud in milliseconds. None of these systems understands anything in the human sense; each one turns your input into numbers, runs learned arithmetic on them, and outputs a decision that is usually right.

Where the name comes from

The term artificial intelligence was coined for a 1956 summer workshop at Dartmouth College, organized by John McCarthy with Marvin Minsky, Claude Shannon, and Nathaniel Rochester. The ambition in the proposal, that every aspect of learning can in principle be described precisely enough for a machine to simulate it, is still the field's north star, and still unfinished. The honest one-line status of the field today: machines are superhuman at many narrow judgment tasks and human-level at none of the broad ones.

1943. Warren McCulloch and Walter Pitts publish the first mathematical neuron. Textbooks call it a weighted sum; it was actually a binary logic gate with an integer threshold and no ability to learn at all. The founding idea was that neural events could be treated with propositional logic.

Naming, hype, and the first winters

1950. Alan Turing asks whether machines can think and proposes the imitation game as a substitute for the question. 1956. The Dartmouth workshop names the field. 1958. Frank Rosenblatt publishes the perceptron, a probabilistic theory of learning in a hypothetical nervous system (the error-correction rule and convergence proof came in 1962 work by Rosenblatt, Block, and Novikoff). Enormous press hype follows, collides with the era's limits, and by the mid-1970s funding collapses: the first AI winter. A second follows in the late 1980s when expert systems, the rules-based approach, hit their ceiling.

The learning era

1986. Rumelhart, Hinton, and Williams popularize backpropagation for training multi-layer networks; the underlying reverse-mode differentiation had been published by Linnainmaa in 1970, and the 1986 paper's real headline was that hidden layers learn useful internal representations. 1991-1994. Hochreiter, then Bengio, Simard, and Frasconi, diagnose why deep networks were failing: vanishing gradients. 2012. AlexNet, a GPU-trained deep network, crushes the ImageNet competition and ends every remaining argument about whether deep learning works. 2015. Networks pass the human benchmark on ImageNet classification.

The scale era

2017. The transformer architecture (Attention Is All You Need) makes it practical to train vastly larger sequence models. 2018-2020. Pretraining on internet-scale text produces GPT-2, then GPT-3, and next-token prediction quietly becomes the most economically important algorithm in the world. 2022. ChatGPT puts a post-trained LLM in everyone's browser. Since then: multimodal models, reasoning training, and agents. The ingredient under all of it is still the 1958-shaped unit: a weighted sum, a bias, a nonlinearity, adjusted by gradient descent.