IELTS Academic Reading Practice Test 152

The industrial robot is a child of the factory rather than of science fiction. The word itself comes from the Czech writer Karel Čapek's 1920 play R.U.R., but Čapek's robots were manufactured people of flesh and blood, not machinery. What appeared on the factory floor four decades later was more modest: a mechanical arm that could be taught a sequence of movements and then repeat them. It did so tirelessly, and to within a fraction of a millimetre. Its ancestors were the automatic machines of the early twentieth century, which could stamp or turn a single component in enormous numbers but could do nothing else.

The crucial step was to make a machine's movements programmable. During the Second World War, engineers had developed servo-mechanisms that used feedback to hold a gun platform steady, and after it the same principles were applied to machine tools. In 1952, with funding from the United States Air Force, the Massachusetts Institute of Technology demonstrated a milling machine controlled by a punched paper tape. Numerical control, as it became known, allowed a machine to cut complicated shapes, but every new part still demanded a new programme. The machine also had no way of knowing whether it had picked up the right piece of metal.

The man who saw the limitations of this approach was George Devol, an American inventor with a long record of patents in electronics and packaging. In 1954 he applied for a patent for a 'Programmed Article Transfer' device: an arm that could be moved through a series of positions by hand, store each position in a magnetic memory, and afterwards repeat the sequence automatically. The patent was granted in 1961, by which time Devol had met Joseph Engelberger, an engineer who immediately recognised the commercial possibilities. Together they founded a company called Unimation.

The first machine they built, the Unimate, was driven by hydraulic pressure and weighed about two tonnes. It was not a clever device. It had no eyes, no sense of touch and no computer, and its memory was a magnetic drum on which a sequence of arm positions could be stored and played back. To record those positions, a worker took hold of the arm and guided it through the movements required. In 1961 a Unimate was installed in a General Motors plant in Trenton, New Jersey, where it took finished parts from a die-casting machine. The job was hot, heavy and dull.

Progress thereafter was slow. The machines cost as much as a house, broke down often and leaked hydraulic fluid onto the floor. Because they moved blindly from one recorded position to the next, they had to be kept inside a fence; anyone who happened to be in the way would be struck. By the late 1960s, General Motors was using Unimates to spot-weld car bodies, and a few other manufacturers had begun to experiment. Even so, the total number of industrial robots in the world remained in the hundreds.

Two developments transformed the field in the 1970s. The first was the replacement of hydraulics by electric motors, which made machines cleaner, quieter and far more accurate. In 1973 the German firm KUKA built an electrically driven arm with six axes, allowing it to reach almost any point in the space around it. A year later the Swedish company ASEA introduced an all-electric robot that became a commercial success. The second was the microprocessor. Instead of a fixed sequence of positions, a robot could now carry a program written in a language, edit it and store many programs at once.

Japan adopted the technology with enthusiasm. A Japanese firm imported its first American robot in 1967, and in 1968 Kawasaki Heavy Industries signed a licensing agreement with Unimation. Faced with a labour shortage and a rapidly expanding car industry, Japanese manufacturers installed robots in far greater numbers than their competitors. By the 1980s the country was building most of the world's industrial robots and operating the largest share of them. Robot density — the number of machines per thousand manufacturing workers — became a measure of industrial maturity, a statistic in which South Korea and Singapore later led the world.

Safety remained a problem. In 1979 a worker at a Ford casting plant in Michigan was killed by a robot arm, an accident often described as the first of its kind. The standard response was physical separation: robots worked inside cages, and those cages were fitted with interlock switches, so that the machine stopped if a gate was opened. Modern standards, set out in documents such as ISO 10218, still assume that a conventional industrial robot is dangerous and must be kept away from people.

That assumption has since been challenged. In the 1990s engineers began to design collaborative robots, or cobots, fitted with force sensors that stop the arm when it meets an unexpected obstacle. They were light enough to be mounted on a bench without a fence. A Danish manufacturer began selling such an arm in the late 2000s, and it was soon adopted for tasks too small or too varied for a traditional robot. Today machine vision and machine learning allow a robot to pick up an object it has never seen, though the human hand remains far more versatile.

A large language model is, at heart, a statistical machine for predicting what comes next. Text is broken into tokens — words, parts of words or punctuation marks — and the model's task is to estimate, at every position, the probability of each token in its vocabulary. A reply is generated one token at a time: the model selects a token, adds it to the text and repeats the process with the enlarged sequence. It does not always select the most probable one; a controlled amount of randomness is introduced so that answers vary. Nothing in this process requires the model to consult a store of facts.

The architecture that made this practical was presented in 2017 by a team of researchers at Google in a paper entitled 'Attention is All You Need'. Earlier systems read a sentence strictly in order, one word after another, and struggled to connect words that were far apart. The transformer, as the new design was called, processes all the tokens at once. Its central mechanism allows every token to weigh its relationship with every other token, so that the model can relate a pronoun to its noun, or a verb to its subject, regardless of the distance between them. The design was quickly taken up by research groups and companies alike.

Before any of this can happen, words must be turned into numbers. Each token is mapped to a vector — a list of several thousand values — and these vectors are adjusted during training so that tokens used in similar contexts end up close together in this mathematical space. Directions in the space come to stand for relationships: the step from one word to another may correspond to a change of tense, of number or of gender. The model is not told any of these rules. They are a by-product of its efforts to predict text. The geometry of this space is a consequence of use, not of any dictionary definition.

Training is a process of repeated guesswork. The model is shown a passage with its final token concealed and must assign probabilities to every possible continuation; its settings are then adjusted, by a procedure called gradient descent, so that the correct token becomes slightly more likely. Repeated across billions of examples, this gradually shapes the network. The numbers that are adjusted in this way are called the model's parameters, and their number has grown enormously: GPT-3, released by OpenAI in 2020, contained 175 billion of them.

For a time, progress depended simply on making models bigger. In 2020 researchers described scaling laws, showing that the error made by a model falls predictably, as a power law, when the number of parameters, the amount of text and the computing power are all increased. Two years later a study by DeepMind suggested that many models had been trained on too little text for their size, and that a smaller model given more data could match a much larger one. Claims that entirely new abilities appear suddenly at a certain scale remain disputed.

A model trained only to continue text may respond to a question with further questions. To turn it into an assistant, developers fine-tune it on examples of the behaviour they want, and then apply reinforcement learning from human feedback. In this technique, people rank competing answers, a second model learns to predict those preferences, and the assistant is rewarded for producing answers that the second model scores highly. The second model is known as a reward model. The method is powerful, but it teaches a model to imitate the preferences of its raters rather than to be accurate.

The most conspicuous weakness of these systems is that they state falsehoods with complete confidence. Because a model is optimised to produce plausible sequences of tokens, not to establish the truth, it will invent a date, a quotation or a scientific reference when the pattern demands one, and it has no reliable sense of what it does not know. Nor is its knowledge current: the parameters are fixed when training ends, so anything that happened afterwards is invisible to it unless the information is supplied in the prompt. Contamination is a further difficulty, since test questions may already have appeared in the training text.

The text used for training is largely scraped from the web, and its sheer quantity matters as much as its quality. Filters remove duplicated and offensive material, but the remaining corpus still reflects the prejudices and gaps of its sources, and a model trained on it will reproduce them. The use of such material has also provoked disputes over copyright, as authors and publishers argue that their work was taken without permission. Training consumes prodigious quantities of electricity, and the computers that serve the finished model consume more.

Research now concentrates on making models cheaper, more reliable and easier to check. One promising approach is retrieval-augmented generation, in which the model is handed relevant documents to consult before it answers, so that its replies rest on texts that can be inspected. Others connect models to calculators, databases or software so that they can act rather than merely talk. A further priority is evaluation: measuring what a model can actually do is surprisingly difficult, and the benchmarks used to compare systems quickly become part of the training data.

A small aircraft descends over a suburban street, hovers, lowers a cardboard box on a wire and climbs away. Films of such scenes have been used for a decade to suggest that the way goods reach our homes is about to be transformed. The machines certainly fly: companies in several countries now make regular deliveries by drone, and regulators have begun to license them. What is less clear is whether the practice deserves to spread. My own view is that drone delivery is a genuine advance in a few narrowly defined situations and a poor substitute for a van in most others.

The devices themselves are of two broad kinds. Multirotor drones, which look like large quadcopters, take off vertically, hover and land on a small pad; they typically carry two or three kilograms over perhaps twenty kilometres. Fixed-wing models fly like small aircraft, launched from a catapult and recovered by a hook, and can cover far greater distances. Zipline, an American company, began using aircraft of the second kind in Rwanda in 2016 to carry blood for transfusions, and similar services followed in Ghana and elsewhere. Wing, owned by Google's parent company, started delivering to households in Australia in 2019, and Amazon began a limited service in the United States in 2022.

The strongest case for the technology is medical. Blood products spoil, hospitals cannot stock every type, and a patient's need is urgent. In countries where the roads are poor and the rainy season long, a journey that takes hours by motorbike may take twenty minutes by air. The same argument applies to vaccines, insulin, antivenom for snakebite and emergency equipment. Here the drone is not competing with an efficient van but replacing an unreliable and sometimes impossible journey, and the value of the cargo is high enough to justify the cost. Few people would object to a drone carrying blood.

Remote communities and disaster zones present a similar case. When an earthquake has broken bridges and blocked roads, or when a village lies beyond the end of any road, an aircraft that needs no runway is a lifeline. Aid agencies have used small drones to survey damage and to carry samples, and there is no serious dispute about the value of such work. These uses have something in common: the destination is hard to reach, the payload is light and the alternative is bad or nonexistent.

The case for delivering pizzas, books and groceries is much weaker. The usual argument is that drones remove vans from congested streets, and so cut both traffic and emissions. This is plausible only if each flight replaces a whole van journey. In reality a delivery van carries dozens of parcels on a route that is planned to be efficient, whereas a drone must fly out and back for a single small package. Because a small aircraft carries one parcel while a van carries dozens, the energy used for each parcel is often greater, not smaller. Where deliveries are short and dense, cargo bicycles are the better answer.

Noise is the drawback that residents notice first. A drone is small, but its propellers turn fast, and the resulting whine carries. A single delivery is a brief nuisance; the objection is to the hundredth delivery of the day, arriving over the same gardens. A drone descending to a suburban pad at seven in the morning will be heard by everyone within a hundred metres, including people who never ordered anything. Studies of environmental noise have long shown that sounds imposed by others are the most irritating, even at low volume.

Safety raises harder questions. A drone must not fall on a person, collide with another aircraft or be ingested by an engine. Its batteries can fail, its propellers can ice up and its radio link can be lost. For this reason regulators in most countries have required the pilot to keep the aircraft in sight, which confines flights to a few hundred metres unless special permission is granted. Rules allowing flight beyond the pilot's view are being drafted, and a system for managing low-altitude airspace is slowly taking shape, but neither is yet complete.

There are quieter concerns. A delivery drone carries a camera, and it may fly over gardens, windows and children playing. Even if the images are discarded, the sense of being watched from above is unwelcome, and the law of trespass offers limited protection against an object passing overhead. The benefits, meanwhile, are unevenly distributed: the company collects the revenue, the paying customer receives a parcel slightly sooner, and the noise and the risk fall on neighbours who were not consulted.

None of this argues for banning the technology. Used for medical supplies, for isolated communities and in emergencies, drone delivery does work that nothing else can do at the same speed. The mistake is to treat it as a general replacement for the delivery van. Rules should therefore distinguish between these cases: permissions should be generous where lives or isolation are at stake, and strict where the only benefit is convenience. Above all, residents should be able to see where flights will pass and to object before a service begins. Treated as a specialised tool rather than a revolution, drone delivery earns its place.