The Shape of Intelligence

Before anyone built a thinking machine, people argued about what thinking was. In 1936 Alan Turing described a machine that could carry out any procedure written as a finite list of steps, which gave "computation" a precise meaning. In 1943 Warren McCulloch and Walter Pitts wrote the nerve cell as a switch: add up the inputs, fire if the total crosses a threshold. In 1950 Turing set the argument aside with a test: if a machine can hold a conversation you cannot tell from a person's, call it intelligent. In 1955 four researchers proposed a summer study at Dartmouth on the conjecture that every feature of intelligence could be described precisely enough for a machine to simulate it. They named the field before they had a method. That order of events matters. The definition of intelligence was never settled first; it was rewritten each time a new method arrived.

The first working programs reasoned with symbols. In 1956 Allen Newell, Herbert Simon and Cliff Shaw's Logic Theorist proved theorems from Whitehead and Russell's Principia Mathematica by searching through possible steps. Rules seemed to be the way in: write down what an expert knows as conditions and actions, and let the machine apply them. Joseph Weizenbaum's ELIZA (1966) showed how far a short list of pattern rules could go, and how quickly people read understanding into it. A rival programme grew alongside. Frank Rosenblatt's perceptron (1958) learned its own weights from examples, and the Mark I perceptron machine set them with motors turning potentiometers. In 1969 Marvin Minsky and Seymour Papert showed the limits of a single layer, and neural approaches lost most of their funding for a decade. Rules were legible and brittle: every case had to be written down, and the world kept producing cases nobody had written.

Expectations ran ahead of results twice. In 1973 Sir James Lighthill's report for the British Science Research Council found that the field had not delivered on its promises, and funding in the United Kingdom fell away. In the 1980s expert systems revived the money: Digital Equipment's XCON configured computer orders with thousands of rules, and Japan's Fifth Generation project (1982) set out to build machines that reasoned. Companies were founded to sell hardware for the Lisp language. By 1987 that market had collapsed, and the phrase "AI winter" had been coined at a 1984 conference by people who saw it coming. Investment and progress did not move together. While budgets fell, small groups kept working: backpropagation for training multilayer networks was published in 1986, Yann LeCun's convolutional network read handwritten digits by 1989, and Gerald Tesauro's TD-Gammon learned backgammon from self-play in 1992. The quiet work was done in the cold.

Learning from examples means giving up on writing the rule. You choose a model with adjustable parameters, show it labelled examples, and move the parameters to reduce the error. A perceptron draws a straight boundary; a network with a hidden layer can bend it. Three things decide what gets learned: the examples, the objective the model is judged by, and its capacity to represent a boundary at all. Statistics supplied the method through the 1990s: support vector machines (1995), probabilistic models, and in 1998 the MNIST digit set that became the field's shared examination. Data grew with the web. ImageNet (2009) grew to more than fourteen million labelled images, labelled by people paid per task through Amazon's Mechanical Turk. A model that fits its examples has not been proved right; it has been fitted. What happens on the next example, from a slightly different world, is the question the examples cannot answer.

In 2012 a network called AlexNet, trained on two consumer graphics cards for about a week, cut the error rate in the ImageNet competition by a margin nobody had seen. Nothing in it was new on its own. Convolutional networks were twenty years old, the data was three years old, and the graphics cards had been built for games. What changed was that the ingredients met at the right sizes. Within four years the same recipe read speech, translated text and, in AlphaGo (2016), beat one of the strongest Go players in the world. Each result rested on things that are easy to leave out of the story: the people who labelled the images, the electricity and cooling for the machines, the benchmarks that decided what counted as progress, and the companies that could afford to run the experiment again. Results emerged from several ingredients interacting, and the ingredients had owners.

In 2017 eight researchers at Google described the transformer, a network that reads a sentence all at once and lets every word attend to every other. Attention weights say which words matter for which. Training is simple to state: predict the next token, again and again, on more text than a person could read in a thousand lifetimes. What the model learns is a representation shaped by context: the same word means different things after different words, and the model's numbers move accordingly. By 2018 BERT set records across language tasks; in 2019 GPT-2 wrote paragraphs; in 2020 GPT-3, with 175 billion parameters, did tasks it had never been trained for when shown a few examples in its prompt. Prediction turned out to produce behaviour. Whether the model understands anything is a question people still argue about. What it does is measurable: a distribution over what comes next.

On 30 November 2022 OpenAI released ChatGPT, a conversational interface on a model tuned with human feedback, and by one estimate a hundred million people used it within two months. The interface changed who was in the conversation. Fluent text is not the same as true text: the same model that writes a working program will invent a court case that never happened, in the same confident sentences. A third property arrived with tools. Models that call a calculator, search the web or run code can act, and an action can succeed or fail separately from whether the words were right. Regulation followed the interface: the European Union's AI Act entered into force on 1 August 2024. By 2025 "agents" meant models running multi-step tasks with software of their own, and the argument moved from what the models could say to what they should be allowed to do.

The instruments in this exhibition were made by people who chose their goals. Rules were chosen for legibility, examples for coverage, prediction for scale. Each choice left something unresolved: cases nobody wrote down, examples nobody collected, outputs nobody checked. The choices ahead are of the same kind, and the technology does not settle them. Who decides what a model is for, and who answers for it when it acts? Which measures of success count: accuracy on a benchmark, harm avoided, energy used, people paid fairly for the data they produced? Should capability be released as fast as it is built? Researchers disagree in public about how much risk the current path carries, and the disagreements are worth reading in their own words. This is not a prediction of where the story goes. It is a fork with the paths unlit, and the lamps are held by people.