IELTS Academic Reading Practice Test 108

For the first half of the twentieth century, electronics ran on glass. A vacuum tube, or valve, was a sealed bulb with the air removed, containing a wire filament that glowed when heated and released electrons into the empty space inside. By controlling the flow of those electrons with a small voltage, a valve could amplify a weak signal, which made radio, telephony and the first electronic computers possible. But valves had obvious drawbacks. They were fragile, gave off a great deal of heat, needed time to warm up and consumed large amounts of power. The computer known as ENIAC, built in Pennsylvania in the 1940s, contained roughly 18,000 of them.

A better device, it was widely believed, would have to be made from a solid rather than from a vacuum. In 1945 the American Telephone and Telegraph Company set up a research group at its Bell Telephone Laboratories in New Jersey to pursue the idea. It was led by William Shockley, a brilliant and ambitious physicist, and included John Bardeen, a quiet theorist, and Walter Brattain, an outstanding experimenter. Their early efforts were frustrated by the behaviour of the surface of semiconductors, where electrons seemed to become trapped.

Bardeen and Brattain eventually worked out that the trapped electrons formed a barrier which blocked the effect they were trying to exploit. Their solution was to press two very fine gold contacts, separated by a tiny gap, against the surface of a crystal of germanium. On 23 December 1947 they demonstrated the resulting device to their colleagues. When a small current was fed into one contact, a much larger current flowed out of the other, so the arrangement could amplify power. Shockley saw at once that this was important.

Shockley therefore spent the following weeks developing a design of his own. Instead of pressing contacts onto the surface, he proposed a sandwich of three layers of germanium, each treated so that its electrical properties differed from those of its neighbours. This junction transistor proved far more robust than the point-contact device and much better suited to manufacture, and within a few years it had replaced it. The name given to the whole family of devices, 'transistor', was coined by a Bell engineer, John Pierce, as a contraction of 'transfer resistor'.

Bell Labs did not keep the invention to itself. Its parent company, AT&T, was under scrutiny from antitrust authorities, and licensing the technology was both a legal safeguard and a way of sharing the cost of research. In 1952 the laboratory held a symposium for firms willing to pay a fee, and representatives of dozens of companies attended. Transistors soon appeared in hearing aids, and in 1954 the first transistor radio, the Regency TR-1, went on sale in the United States.

The material also changed. Germanium worked, but it was sensitive to heat and its behaviour was difficult to control. Silicon, more abundant and more stable, was harder to purify but promised better results. In 1954 researchers at Texas Instruments and at Bell Labs independently produced working silicon transistors, and the industry gradually shifted to the new material, which remains the basis of almost every chip manufactured today. Silicon also had a natural oxide that could be grown on its surface, a property that later proved decisive.

The next problem was interconnection. A circuit containing thousands of separate transistors, each wired to the next, was expensive to assemble and liable to fail at any joint. In 1958 Jack Kilby, working at Texas Instruments, built a circuit in which all the components were formed within a single piece of germanium. A year later Robert Noyce, at the Fairchild Semiconductor Corporation, devised a more practical version using silicon, with connections laid down as thin strips of metal and protected by a layer of oxide. The integrated circuit, or chip, had been invented.

Two Bell Labs researchers, Dawon Kahng and Martin Atalla, had meanwhile designed a quite different kind of transistor. In their field-effect device, a voltage applied to an electrode separated from the semiconductor by a thin insulating layer of oxide controlled the current flowing beneath it. Because it needed very little power and could be made extremely small, this metal-oxide-semiconductor transistor, or MOSFET, became the dominant form. By the end of the century it was being produced by the billion, and it remains the type found in almost every processor and memory chip in use.

In 1965 Gordon Moore, then at Fairchild, noticed that the number of components that could be packed onto a chip had been doubling roughly every year, and predicted that the trend would continue. He later revised the period to about two years. Moore's law, as the observation became known, was never a law of physics, but for half a century it described the industry's ambitions so accurately that firms organised their research around it. The consequences were extraordinary: a device that once filled a room now sits in a pocket.

Bardeen, Brattain and Shockley shared the Nobel Prize in Physics in 1956. Bardeen later won a second, for work on superconductivity, and remains the only person to have received the physics prize twice. Shockley moved to California and founded a company whose early employees went on to establish Fairchild and, eventually, Intel, in the region that became known as Silicon Valley. The name, of course, refers to the element at the heart of the device.

An ordinary computer stores information in bits, each of which is either a zero or a one. A quantum computer uses qubits, which obey the rules of quantum mechanics and can therefore occupy a combination of both values at once. This property is called superposition. Two qubits can hold four values at once, three can hold eight, and the number of possibilities doubles with every qubit added, so that a machine of only three hundred qubits would describe more states than there are atoms in the visible universe. That extraordinary capacity is the source of the excitement, and also of the misunderstanding, because those states cannot simply be read out one after another.

Qubits can also be entangled, meaning that their states become linked in a way that has no counterpart in ordinary physics: measuring one instantly constrains what can be found in the other, however far apart they are. Entanglement alone does not produce a calculation. The work is done by interference. A quantum algorithm is arranged so that the paths leading to wrong answers cancel each other out, like waves out of phase, while the paths leading to the right answer reinforce. At the end the state is measured, and the superposition collapses into a single result, which is why a quantum computation usually has to be repeated many times to give a usable answer.

The idea that one quantum system might be used to simulate another was argued by the physicist Richard Feynman in the early 1980s. Nature, he pointed out, is not classical and cannot be modelled efficiently by classical means. The field became urgent in 1994, when Peter Shor, then at Bell Laboratories, described an algorithm by which a quantum computer could factor very large numbers. Much of the world's commerce is protected by encryption whose security depends on the difficulty of factoring such numbers, so a working machine would make some widely used methods of protecting data obsolete.

Most machines built so far use superconducting circuits. A qubit is a small loop of metal in which an electric current oscillates, and which can hold two energy levels at once. The loop must be kept at a temperature only a few thousandths of a kelvin above absolute zero, colder than the space between the stars, and it sits at the bottom of a dilution refrigerator that fills the room. Operations are performed by sending pulses of microwaves down cables, and a two-qubit operation takes a few tens of nanoseconds.

Competing technologies have different strengths. Trapped ions, held in place by electromagnetic fields and manipulated with lasers, are slower but far more reliable and can be entangled over greater distances. Neutral atoms, also held by light or magnetic fields, are easier to arrange in large regular arrays, and photonic systems can work at room temperature but are difficult to store. Each approach has its own engineers, its own funding and its own enthusiasts. No one yet knows which of them, if any, will dominate.

The central difficulty is that a quantum state is extraordinarily fragile. Physicists call its loss decoherence, and it can be triggered by something as slight as a stray vibration or a fluctuation of temperature, or even by the very pulses used to control the qubits. A useful calculation may require millions of operations, so even a low error rate will eventually produce nonsense. Classical computers cope by storing each bit three times and taking a majority vote, but a quantum state cannot be copied in this way. Correction therefore requires many physical qubits to create a single reliable, or logical, one, and the best-developed scheme needs thousands.

Where does this leave the field? Today's largest machines use several hundred to around a thousand qubits and are described by their designers as noisy and intermediate in scale. In 2019 Google reported that a processor of fifty-three qubits had completed in a few minutes a calculation that would take a classical supercomputer thousands of years, and IBM immediately argued that the comparison was unfair and could be improved upon. Such milestones are genuinely impressive, but they are demonstrations of arithmetic rather than of usefulness. Errors dominate long before a problem of commercial interest can be tackled.

Where hope persists is in problems that are themselves quantum. Predicting the behaviour of a catalyst or a new drug molecule requires accounting for the interactions of many electrons, and classical computers handle this badly; a machine built from qubits is, in a sense, doing the same thing as the system it is studying. Chemists and materials scientists are therefore the most likely early beneficiaries, with optimisation, machine learning and the simulation of financial systems further off. A quantum computer would not speed up everything; for most everyday tasks a laptop is and will remain better.

It is also unlikely that these machines will arrive as replacements. The most plausible picture is a hybrid arrangement, in which a conventional computer prepares the problem, runs the parts it does well and interprets the results, while a quantum processor performs a small number of specialised steps. Governments and technology companies are investing heavily in the hope of exactly that. Whether the reward is the transformation of chemistry, a modest acceleration of a few niche problems, or a useful lesson about the limits of computing, no one can yet say.

Every few months a new demonstration arrives. A program translates a speech into another language in the speaker's own voice; a telephone app reads a menu or a road sign; a company announces that it has automated the translation of its entire catalogue. Each time, commentators draw the same conclusion: the days of the professional translator are numbered. Anyone who has worked in the field for twenty years will have heard this prediction at least once a decade. It is not that the prediction is wholly wrong. It is that it answers the wrong question. Whether a machine can translate matters less than what kind of translation it can do, for whom, and at what cost to meaning.

The idea is not new. In 1954, at Georgetown University in Washington, a team working with the computer company IBM turned a few dozen Russian sentences into English, and newspapers reported that translation would soon be solved. Governments invested heavily in rule-based systems, which attempted to encode grammar and vocabulary as explicit instructions, and in 1966 an American committee concluded that the results did not justify the expense, and funding collapsed. Statistical methods, which learned from millions of translated documents rather than from rules written by linguists, revived the field from the late 1980s, and in 2016 neural networks produced a sharp improvement in fluency.

Contemporary systems are trained on enormous collections of parallel texts: novels, legal codes, parliamentary records, subtitles and websites. A neural network learns to predict, word by word, what a human translator would probably have written in that context, and the most successful architectures use an internal mechanism called attention to weigh how strongly each part of the source sentence should influence each part of the output. The result is often remarkably smooth. What the system lacks, however, is any model of the world the words describe, or of the person who will read them.

None of this makes the technology useless; quite the opposite. For a traveller asking for directions, a researcher skimming an article in an unfamiliar language or a small charity trying to read a foreign supplier's documents, instant translation is a genuine gift, and its cost is close to nothing. For internal correspondence, rough drafts and texts whose only purpose is to convey information, machine output is often good enough. It is fast, it is cheap, and it never tires.

Where it falters is where language carries more than information. Consider the ordinary English pronoun 'you'. French, German and many other languages require a choice between a familiar and a formal form, and that choice depends on the relationship between the speaker and the person addressed, which the sentence itself rarely states. Idioms, jokes, puns, proverbs and religious references are worse still, since their meaning is embedded in a culture rather than in a dictionary. A machine will usually produce something plausible in such cases, and plausibility is precisely what makes the errors hard to spot.

The economics then become awkward. Because machine output looks finished, clients increasingly ask for it to be checked rather than translated, and expect to pay accordingly. In practice, correcting a text that is subtly wrong can take longer than translating it from scratch, because the editor must first work out what the original meant and then decide how far the draft has departed from it. Agencies that sell this service as an equivalent product are, for many kinds of text, offering a false economy: the saving appears on the invoice and the loss appears somewhere else.

I should be honest about the counter-argument. Sceptics have underestimated these systems repeatedly, and each improvement has removed a task that translators once regarded as safe from automation. Fluency has risen; errors have become rarer; the technology will go on improving. I would not bet against a machine on any single, narrowly defined task. But translation is rarely a single, narrowly defined task.

Two things remain stubbornly human. The first is responsibility. A contract, a medical leaflet or a court statement requires someone who can be held to account for the words, and no client can hold a neural network to account. The second is interpreting, where meaning is carried by tone, hesitation, gesture and the timing of a pause as much as by vocabulary, and where the interpreter must judge, in real time, what the speaker intended rather than what the speaker said.

The nature of the profession will change, as it has changed before. Translators are becoming editors, terminologists and advisers on style and culture, and the training they need is different from what their predecessors received. Literary translation, which depends on voice, rhythm and the deliberate ambiguity of poetry, continues to attract practitioners and readers. What worries me is not that machines will take over, but that fluency will be mistaken for accuracy, and that the people best able to spot the difference will be the first to be dispensed with.

The question in my title, then, is badly framed. Machines will not replace translators; they will replace certain tasks, and they will make some kinds of translation so cheap that they will happen where none happened before. That is a gain. But wherever accuracy matters, or voice matters, or someone has to answer for the result, a human being will still be required. The danger is that we shall forget this until a contract is misread, a prescription misunderstood or a novel spoiled.