IELTS Academic Reading Practice Test 179

Chess has fascinated computer scientists for longer than almost any other game. It offers no hidden information and no element of luck: both players can see every piece on the board, and nothing is decided by dice. In principle, therefore, it looks like a problem of pure calculation. The number of possible games, however, is immense: Claude Shannon, an American mathematician, estimated in 1950 that some 10^120 could be played, far more than the number of atoms in the observable universe. No machine could examine all of them, so chess became a testing ground for methods of searching intelligently.

Long before the computer, however, the most famous chess-playing machine in Europe had already made its name, and it contained no electronics at all. The Turk was unveiled in 1770 by Wolfgang von Kempelen, an official at the Habsburg court, and consisted of a cabinet behind which sat a life-sized figure dressed in Turkish robes. It toured Europe for decades, defeating many strong players. Its secret, eventually established beyond doubt, was that a human player was concealed inside the cabinet, moving the pieces from a cramped compartment. The machine was taken to the United States and destroyed by fire in Philadelphia in 1854.

The first machine that genuinely played chess, without a hidden operator, was built by the Spanish engineer Leonardo Torres y Quevedo and completed in 1912. El Ajedrecista played a king and rook against a lone king, moving the pieces with electromagnets beneath a metal board and sensing replies through electrical contacts. It could not play a full game: its skill was limited to this single endgame. It signalled when it had delivered check and refused moves that broke the rules, and Torres y Quevedo demonstrated it at the University of Paris in 1914.

The theoretical foundations were laid by Shannon and by the British mathematician Alan Turing. Shannon's paper of 1950, 'Programming a Computer for Playing Chess', described how a program might assess a position by adding up the values of the pieces, and how it could look ahead by considering replies and counter-replies — the procedure known as minimax. Turing wrote a program with David Champernowne, but no computer was available to him, so he carried out its instructions himself, move by move. Turing's program lost the game it played against a colleague in 1952.

Early machines were weak. In 1951 Dietrich Prinz, working at the University of Manchester, wrote a program for the Ferranti Mark I that could solve mating problems requiring checkmate in two moves; it could not play a game. In 1956 a program on the MANIAC I computer at Los Alamos played on a simplified six-by-six board without bishops and defeated a young woman who had learned the rules a week earlier. Two years later, Alex Bernstein and his colleagues at IBM produced a program that played a complete game.

Progress over the next three decades came from faster hardware and better search rather than from any single insight. The most important technique was alpha-beta pruning, which allows a program to ignore branches of the search tree that cannot affect the outcome; a well-ordered search can therefore reach much deeper in the same time. In 1968 the British computer scientist David Levy accepted a bet that no computer would beat him in a match within ten years. He won it in 1978.

In 1988 a machine called Deep Thought, built by a team at Carnegie Mellon University that included Feng-hsiung Hsu, defeated the Danish grandmaster Bent Larsen. The following year the world champion Garry Kasparov beat Deep Thought in a two-game match, but IBM noticed the team and hired it to build a machine capable of beating Kasparov himself. The result, Deep Blue, was a parallel computer with special-purpose chess chips that could evaluate some 200 million positions a second. A match in New York in May 1997 was won by the machine, 3½–2½. IBM retired Deep Blue shortly afterwards.

Deep Blue's victory did not mean that chess had been solved by brute force. It did, however, end the era in which the best human players could expect to win. The next advance came from a different direction: learning. In December 2017 researchers at DeepMind, a London company owned by Google, reported that a program called AlphaZero had been given nothing except the rules of the game and had then taught itself by playing against itself. Its neural network was trained by a method known as reinforcement learning, in which the program is rewarded for results rather than instructed how to achieve them.

AlphaZero's strength lay in the way it chose which positions to examine. Like earlier programs, it searched ahead, but a neural network guided that search, estimating which moves were worth considering and how favourable each position was. In a hundred-game match against Stockfish, then the strongest conventional engine, AlphaZero won 28 games, drew 72 and lost none. Stockfish examined some 70 million positions a second, while AlphaZero looked at roughly 80,000, a fraction as many, yet chose better ones.

Programs are now part of the way chess is played and taught. Grandmasters prepare openings with engines, and the strongest chess played anywhere is produced by machines. In freestyle tournaments, where competitors may consult programs, teams of amateurs working with several engines have beaten both strong grandmasters and the engines playing alone — a reminder that the most effective arrangement is not always the one that replaces people entirely.

The phrase 'the Internet of Things' is usually credited to Kevin Ashton, who used it in 1999 while working on a project to attach radio-frequency identification tags to products in a supermarket supply chain. What he described then was a fairly narrow idea: objects that could identify themselves to computers without a human typing anything. The term has since expanded to cover almost any device with a sensor, a processor and a network connection, from a domestic heating controller to a soil probe in a field. Estimates of how many such devices exist vary widely, but they are commonly given in the tens of billions.

A typical node is a small computer with three jobs. A sensor converts a physical quantity — temperature, vibration, the concentration of a gas — into an electrical signal, which a microcontroller turns into numbers. A radio then transmits those numbers, often by a short hop to a nearby receiver. An actuator does the opposite of a sensor, turning a signal into a physical effect: a valve opens, a motor turns. Because most nodes are battery-powered or rely on energy scavenged from their surroundings, they spend nearly all their time asleep, waking briefly to measure and transmit. The energy cost of sending a few bytes by radio can exceed that of thousands of calculations, so designers compress data before it leaves the device.

The connections between nodes and the internet take several forms. Short-range technologies such as Bluetooth Low Energy, Zigbee and Wi-Fi work over metres and suit devices in a home or factory. Wide-area networks are needed for sensors in fields, pipelines or mines, and here two approaches compete. LoRaWAN, an open standard, encodes data in chirps — pulses that sweep across a range of frequencies — and in the countryside a single gateway can pick up signals from sensors tens of kilometres away, though the data rate is only a few hundred bits a second. Narrowband IoT, by contrast, uses licensed mobile spectrum and is offered by telephone companies, which means it inherits their billing systems and their coverage maps.

Raw data rarely travel straight from a sensor to a data centre. A gateway collects readings from many nodes, checks them and forwards them in batches, often using a protocol such as MQTT, which needs little bandwidth and copes with unreliable connections by publishing messages to a broker rather than addressing each recipient directly. Increasingly, some analysis happens at the edge of the network — on the gateway or even on the node itself — because a factory that must stop a machine within milliseconds cannot wait while a message travels to another country and back. In industry, a digital twin, a model kept up to date by sensor readings, allows engineers to try out changes in a simulation before making them on the real equipment.

The applications that have grown fastest are unglamorous. Electricity and gas meters report consumption without a visit from a reader, and water utilities use acoustic sensors to listen for the leaks that waste much of the water in ageing pipe networks. Factories fit vibration and temperature sensors to pumps and motors so that maintenance can be scheduled before a failure stops production. Farmers place probes in the soil to decide when to irrigate, and wearables send a patient's heart rhythm to a clinic. Cities adjust traffic signals according to the queue at each junction.

Security is the weakest point. Many cheap devices ship with default passwords that are never changed, and their makers rarely issue software updates, so a flaw discovered years after a product was sold may never be fixed. Because such devices are numerous, always connected and poorly defended, they make ideal recruits for a botnet. The Mirai malware, which appeared in 2016, took control of hundreds of thousands of cameras and home routers and used them to flood a company that manages part of the internet's domain system; on 21 October 2016 many popular websites became unreachable. A microcontroller with a few kilobytes of memory can barely run the encryption that a laptop takes for granted, and a device with no screen and no keyboard is difficult to update even when a patch exists.

Standards have been another obstacle. Products from different manufacturers have often spoken incompatible languages, so a household that buys a lamp, a thermostat and a doorbell from three companies may need three applications and three accounts. In 2022 the Connectivity Standards Alliance, an industry group, released the first version of Matter, a common application layer for smart-home products, and several large companies have promised to support it. Privacy raises a different question, since a meter that records when a room is occupied reveals more about its occupants than the supplier needs to know.

Research is now directed at removing the battery and the cable. Energy harvesting draws power from light, vibration or the difference in temperature between two points, though the amounts are tiny and intermittent. Ambient backscatter takes a different approach: instead of generating a radio signal, a tag reflects the waves already in the air and modulates them, so that it can transmit while using almost no power. TinyML, the running of small machine-learning models on microcontrollers, allows a device to decide that its data are uninteresting and to stay silent. Federated learning offers a way to train models on data spread across thousands of devices without collecting the data in one place, which reduces both the volume of traffic and the risk to privacy.

Estonia is the only country where internet voting is a routine part of national elections rather than an experiment. Voters there have been able to cast a ballot online in local elections since 2005, and in 2007 the country held the first parliamentary election anywhere in the world in which internet votes counted; in 2019 more than two-fifths of all votes were cast online. Voting requires an identity card and personal codes. Supporters present the country as proof that the technical problems have been solved. I do not think the evidence supports quite so confident a conclusion, but neither do I accept that remote voting is a fantasy.

The arguments in favour are substantial. A ballot cast from home costs far less to administer than one cast in a polling station, which needs premises, staff and paper. People who find it hard to travel — those with limited mobility, those in hospital, those caring for young children — can vote without difficulty, and so can citizens living abroad, a group that postal voting serves badly. Whether the system raises turnout is disputed: studies of the Estonian case have found modest effects at best, and turnout in national elections there has stayed close to the levels recorded before it was introduced.

The case against begins with the device. A vote cast from a kitchen table passes through an operating system, a browser and a network that no election official controls or can inspect. Malicious software on the voter's own machine could record a choice and change it, or display one candidate on the screen while transmitting another. In a polling station, by contrast, a person marks a paper, folds it and watches it go into a box. The infrastructure is exposed too. A denial-of-service flood that leaves the service unreachable for a few hours may not change a result, but it corrodes confidence, and the servers, the software and the machines used to prepare the ballot all become targets.

Cryptographers have answers to the first of these problems, though they are demanding. End-to-end verifiable systems produce a receipt that allows each voter to check that a particular encrypted ballot was included in the final count, while techniques such as mix-nets allow officials to add up votes without ever decrypting an individual one. Observers can audit the result without learning how anyone voted. They solve the problem of the dishonest server. They do not solve the problem of the dishonest voter's laptop, because no design can control software running on a machine the election cannot see.

Coercion is the second difficulty. Voting at home moves the act into a place where an employer, a landlord or a relative may be watching, and a ballot bought with money or a threat is hard to detect when it is cast on a sofa. Paper voting is not immune, but the polling station offers some protection through privacy. Estonia's response is elegant: a voter may cast an electronic vote more than once and only the last is counted, so that someone who has been coerced can later vote again in safety, and a paper ballot cast at a polling station overrides anything sent online. The protection is real but incomplete, since a voter who never returns to the terminal remains exposed.

Other countries have weighed the balance and stepped back. Norway ran trials in 2011 and 2013 and then stopped. The German constitutional court ruled in 2009 that voting machines must be understandable to citizens without specialist knowledge. The Netherlands gave up paperless voting machines in 2007 after campaigners showed how easily a machine's output could be manipulated, and Ireland never used the machines it had bought. None of these decisions was reached because the technology failed to work. Each was made because the public could not be given a convincing account of why the technology should be trusted.

The digital divide supplies a further objection, though not a decisive one. If online voting ever replaced the paper ballot, it would exclude those without skills, equipment or a reliable connection, and those groups are already the least likely to vote. As an additional channel, however, it excludes nobody. The cost of running two systems, though, is not trivial, and there is a subtler danger: as online voting becomes normal, the pressure to close polling stations and save money grows.

My own view is that internet voting has been justified only where it is offered as a supplement, and only where three conditions are met. The system must be end-to-end verifiable, so that the result can be checked from outside. Voters must be able to replace an electronic vote, and to override it at a polling station, so that coercion can be undone. The software must be published and scrutinised, because a secret system cannot be audited by anyone. Countries that cannot meet these conditions should keep counting paper ballots in public, by hand.

None of this is an argument for refusing the technology for ever. Estonia's system has worked, and the convenience it offers is genuine. But the argument that a service can be built is not the same as the argument that it should be universal, or that its adoption should be rushed. Democracy is an unusual kind of system: the loser must accept the result, and that acceptance depends on a shared confidence in the count. Anything that weakens that confidence costs more than it saves.