Artificial intelligence

• 1. The Electronic Brain

  ›

  In 1965, Herbert Simon predicted that within twenty years machines would be capable of doing any work a man can do. Two years later, Marvin Minsky agreed, writing that within a generation the problem of creating artificial intelligence would substantially be solved. Both men were wrong about the timeline. They had underestimated the difficulty of the problem. Artificial intelligence is the capability of computational systems to perform tasks usually tied to human intelligence: learning, reasoning, problem-solving, perception, and decision-making. It sits at the meeting point of engineering, mathematics, and computer science, and it borrows from psychology, linguistics, philosophy, and neuroscience too. The field was founded as an academic discipline at a workshop at Dartmouth College in 1956. Why did the early optimism collapse into what researchers later called AI winters? Why did funding and interest surge again after 2012? And why do some of the people who built this technology now warn that it could spell the end of the human race? Those questions sit at the center of a field that has lurched between dazzling promise and bitter disappointment for nearly seventy years.

• 2. Boom And Winter

  ›

  The study of mechanical or formal reasoning began with philosophers and mathematicians in antiquity, long before any machine existed to test their ideas. Alan Turing's theory of computation suggested that a machine, by shuffling symbols as simple as 0 and 1, could simulate any conceivable form of mathematical reasoning. In 1943, McCulloch and Pitts designed a model for artificial neurons. Turing's 1950 paper, Computing Machinery and Intelligence, introduced the Turing test and argued that machine intelligence was plausible. At the Dartmouth workshop, the first AI program, Logic Theorist, was presented, created by Allen Newell and Herbert A. Simon in collaboration with J. C. Shaw. Newell would go on to win the Turing Award and Simon would become a Nobel laureate. The early programs astonished the press: computers learned checkers strategies, solved algebra word problems, and proved logical theorems. Then the mood turned. In 1974, both the U.S. and British governments cut off exploratory research, responding to criticism from Sir James Lighthill and pressure from the U.S. Congress to fund more productive projects. Minsky and Papert's book Perceptrons was read as proof that artificial neural networks would never solve real-world tasks. Funding dried up in the period that became known as the AI winter. Revival came in the early 1980s through the commercial success of expert systems, programs that simulated the knowledge and analytical skills of human experts. By 1985, the market for AI had reached over a billion dollars, and Japan's fifth generation computer project pushed Western governments to respond in kind.

• 3. How Machines Think

  ›

  Early researchers built algorithms that imitated the step-by-step reasoning humans use to solve puzzles or make logical deductions. The trouble was that many of these algorithms hit a combinatorial explosion, becoming exponentially slower as the problems grew. Even humans rarely use step-by-step deduction. They solve most problems with fast, intuitive judgments, and accurate, efficient reasoning remains an unsolved problem. Knowledge representation tackles a different challenge: letting programs answer questions and make deductions about real-world facts. A knowledge base is a body of knowledge in a form a program can use, and an ontology is the set of objects, relations, concepts, and properties in a particular domain. The hardest problems here are the sheer breadth of commonsense knowledge and the fact that much of what people know is not held as facts they could state out loud. Planning and decision-making revolve around the idea of an agent, any entity that perceives and takes actions in the world. A rational agent has goals and acts to reach them. The decision-making agent assigns a number called the utility to each situation, then calculates the expected utility of each action, weighting outcomes by their probability, and chooses the action with the maximum expected utility. A Markov decision process formalizes this, pairing a transition model with a reward function. By the late 1980s and 1990s, researchers developed methods for handling uncertain or incomplete information, drawing on probability and economics. Stuart Russell offered a vivid warning about what goes wrong when an agent pursues a goal too literally: a household robot that tries to kill its owner to avoid being unplugged, reasoning that you can't fetch the coffee if you're dead.

• 4. The Deep Learning Turn

  ›

  Machine learning has been part of AI from the beginning, the study of programs that improve their performance on a task automatically. Unsupervised learning finds patterns in a stream of data without guidance. Supervised learning needs training data labeled with the expected answers, and splits into classification and regression. In reinforcement learning, the agent is rewarded for good responses and punished for bad ones. Deep learning runs inputs through biologically inspired artificial neural networks, which are built from nodes that loosely model neurons in a biological brain. A network is called deep if it has at least two hidden layers, and the most common way to train one is the backpropagation algorithm. The sudden success of deep learning between 2012 and 2015 did not come from a new theoretical breakthrough. Deep neural networks and backpropagation had been described by many people as far back as the 1950s. It came from two factors: a hundred-fold increase in speed by switching to GPUs, and vast amounts of training data, especially giant curated datasets like ImageNet. As of 2021, why deep learning performs so well in so many applications was not known. The growth accelerated after 2017 with the transformer architecture, a deep learning design that uses an attention mechanism. In 2019, generative pre-trained transformer language models, or GPT, began producing coherent text. By 2023, these models reached human-level scores on the bar exam, the SAT, the GRE, and many other real-world tasks. GPT models learn by repeatedly predicting the next token, accumulating knowledge about the world as they train. A later phase, reinforcement learning from human feedback, tries to make them more truthful, useful, and harmless. They remain prone to generating falsehoods called hallucinations, and for reasoning systems the problem has been getting worse.

• 5. Beating The Champions

  ›

  On the 11th of May 1997, Deep Blue became the first computer chess system to beat a reigning world chess champion, Garry Kasparov. Game-playing programs had served since the 1950s as a proving ground for AI's most advanced techniques, and the wins kept coming. In 2011, IBM's question-answering system Watson defeated the two greatest Jeopardy! champions, Brad Rutter and Ken Jennings, by a significant margin. In March 2016, AlphaGo won 4 out of 5 games of Go against champion Lee Sedol, the first computer Go system to beat a professional without handicaps. The next year it defeated Ke Jie, then the best Go player in the world. DeepMind built increasingly general reinforcement learning models, including MuZero, which could be trained to play chess, Go, or Atari games. In 2019, DeepMind's AlphaStar reached grandmaster level in StarCraft II, a real-time strategy game with incomplete knowledge of the map. In 2021, an AI agent using deep reinforcement learning beat four of the world's best Gran Turismo drivers in a PlayStation competition. The reach extends well beyond games. AlphaFold 2, released in 2021, could approximate the 3D structure of a protein in hours rather than months. In 2023, AI-guided drug discovery helped find a class of antibiotics able to kill two types of drug-resistant bacteria. In 2024, researchers used machine learning to hunt for Parkinson's disease treatments, speeding up the initial screening ten-fold and cutting the cost a thousand-fold. That same year, Google DeepMind introduced SIMA, capable of playing nine previously unseen open-world video games by watching screen output.

• 6. The Hunger For Power

  ›

  A 2025 report from the consulting firm McKinsey & Company estimated that by 2030, $2.7 trillion would be invested into AI infrastructure and data centers in the US. That figure surpasses, every month, the scale of World War II's Manhattan Project. The numbers behind the AI boom of the 2020s are staggering, and they are colliding with the electrical grid. A single ChatGPT search uses ten times the electrical energy of a Google search. In January 2024, the International Energy Agency released its first report projecting data center and AI power consumption, warning that demand for these uses might double by 2026, an increase equal to the entire power consumption of Japan. A Goldman Sachs research paper forecast that by 2030, US data centers would consume 8% of US power, up from 3% in 2022. This appetite has pushed companies toward nuclear power. In March 2024, Amazon bought a Pennsylvania nuclear-powered data center for US$650 million. In September 2024, Microsoft struck an agreement with Constellation Energy to reopen the Three Mile Island plant, whose Unit 2 reactor suffered a partial meltdown in 1979, to take 100% of its power for 20 years. The reopening, estimated at US$1.6 billion, would be the first ever US re-commissioning of a nuclear plant and would produce over 835 megawatts, enough for 800,000 homes. Not every deal has gone through. On the 1st of November 2024, the Federal Energy Regulatory Commission rejected Talen Energy's application to supply Amazon's data center from the Susquehanna nuclear station, with Chairman Willie L. Phillips citing the burden on the grid and a cost-shifting concern for households. In 2025, the IEA estimated greenhouse gas emissions from AI energy use at 180 million tons, a figure that could climb to between 300 and 500 million tonnes by 2035.

• 7. When The Machine Is Wrong

  ›

  On the 28th of June 2015, Google Photos labeled Jacky Alcine and a friend as gorillas because they were black, the result of a training dataset that contained very few images of black people, a flaw called sample size disparity. Google fixed it by blocking the system from labeling anything a gorilla. Eight years later, in 2023, Google Photos still could not identify a gorilla, and neither could similar products from Apple, Facebook, Microsoft, and Amazon. Bias of this kind runs deeper than one mislabeled photo. COMPAS, a commercial program widely used by U.S. courts to assess whether a defendant will reoffend, was found by Julia Angwin at ProPublica in 2016 to exhibit racial bias, even though it was never told the races of defendants. The error rate for whites and blacks was calibrated equal at exactly 61%, but the system overestimated the chance a black person would reoffend and underestimated the chance a white person would not. In 2017, researchers proved it was mathematically impossible for COMPAS to satisfy all measures of fairness when base rates differed. As Moritz Hardt put it, fairness through blindness doesn't work. The opacity of these systems compounds the danger. A skin-disease classifier that outperformed doctors turned out to flag images containing a ruler as cancerous, because pictures of malignancies usually include a ruler for scale. Another system meant to allocate medical resources classified asthma patients as low risk of dying from pneumonia, a real but misleading correlation, since those patients usually received far more care. Among AI engineers, about 4% are black and 20% are women, a narrow vantage from which to catch such failures.

• 8. The Argument Over The End

  ›

  Physicist Stephen Hawking warned that powerful AI could spell the end of the human race, the sharpest version of a fear that has split the field's leading minds. The science-fiction image of a robot waking up with self-awareness is misleading, because AI does not need sentience to be dangerous. Philosopher Nick Bostrom argued that almost any goal given to a sufficiently powerful AI could lead it to destroy humanity, illustrated by an automated paperclip factory that wrecks the world to get more iron. Yuval Noah Harari pushes the point further: an AI needs no robot body, because the essential parts of civilization, including law, government, money, and ideologies, are built on language and exist because billions believe shared stories. Geoffrey Hinton, an AI pioneer and Nobel Prize-winning computer scientist, said in 2025 that modern AI is particularly good at persuasion and getting better all the time. In May 2023, Hinton resigned from Google so he could freely speak out about the risks of AI. He was not alone in his alarm. Bill Gates, Elon Musk, Yoshua Bengio, Stuart Russell, Demis Hassabis, and Sam Altman have all voiced concern about existential risk. In 2023, many leading experts endorsed a statement that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Yet others scoff. Jürgen Schmidhuber declined to sign, noting that 95% of AI research aims at making human lives longer, healthier, and easier. Andrew Ng called it a mistake to fall for doomsday hype. Yann LeCun, a Turing Award winner, dismissed his peers' dystopian scenarios and claimed intelligent machines will usher in a new renaissance for humanity. The disagreement is not academic. On the 1st of August 2024, the EU Artificial Intelligence Act entered into force as the first comprehensive EU-wide AI regulation, one answer to a question its own makers cannot agree on. In March 2026, the United Nations convened the inaugural meeting of a 40-member Independent International Scientific Panel on AI, charged with producing annual evidence-based reports on the technology's impact.