IELTS Academic Reading Practice Test 151

For most of the nineteenth century, manufactured goods were made by skilled workers who built each item in a single place, fetching parts from stores or shaping them by hand. A carriage, a clock or a sewing machine took shape at one bench, and no two products were exactly alike. The idea that the product might instead travel past a series of workers, each of whom performed one small task, appeared first in less celebrated settings. The shipyards of Venice assembled galleys along a channel, with materials delivered in sequence. The overhead trolleys of Chicago's slaughterhouses, where carcasses moved past butchers who each removed a particular cut, are often cited as an inspiration for the men who later reorganised car production.

In 1901 the American manufacturer Ransom Olds introduced a system that carried this principle into the motor industry. His Curved Dash runabout was a simple vehicle, and it was built in a factory where the cars, mounted on wheeled trolleys, were pulled along a line of stationary workers who each added a component. Output rose quickly, and the little car became the first mass-produced automobile in the United States. Even so, the method was only a partial one. The cars moved, but the pace was set by people, and heavy parts were still lifted by hand.

Henry Ford went further. The Model T, introduced in 1908, was designed for simplicity, but the demand for it outstripped anything the company could supply from existing methods. In April 1913, at the Highland Park plant outside Detroit, engineers experimented with moving lines. The first component to be assembled this way was not the car itself but the flywheel magneto, a small generator that provided the ignition current. Instead of one skilled man building a complete unit, the work was divided so that each worker performed a single operation as the component passed. The time taken to make a magneto fell from about twenty minutes to thirteen.

Later that year the same principle was applied to the chassis, the frame and running gear on which the body is mounted. At first the chassis was pulled along the floor by a rope, with the workers staying where they were while the vehicle advanced past them; later a chain-driven conveyor was installed. Parts arrived beside each worker down inclined chutes, so that nobody had to leave a position to fetch anything. Assembling a chassis had taken about twelve and a half hours; with the moving line, the time fell to ninety-three minutes.

The effect on prices was striking. As volume grew, the price of a Model T fell from 850 dollars in 1908 to less than 300 dollars in the mid-1920s, while the number produced each year rose into the millions. Other manufacturers had to follow. Ford's share of the American market was enormous, and a company that could not match the new costs was soon in difficulty. Nor did the technique stay within the car industry. Vacuum cleaners, radios, sewing machines and much else came to be built on moving lines, first in the United States and then in Europe, where manufacturers sent engineers to study the Highland Park plant at first hand.

For the people on the line, the experience was less celebrated. A man might tighten two bolts on hundreds of engines in a day, and the monotony, together with the pace, made the work unpopular. Turnover at Highland Park was extraordinary: in 1913 the company hired more than 50,000 people to fill roughly 14,000 positions. In January 1914 Ford responded by announcing a minimum wage of five dollars a day, nearly double the going rate, and cutting the working day from nine hours to eight. The extra money was not automatic. To qualify for the profit share, workers had to satisfy company investigators, who visited their homes and enquired into their habits, and the improvement in productivity that followed was considerable.

When war came, the method was applied to weapons. Ford built a vast plant at Willow Run, Michigan, to manufacture B-24 bombers on moving lines, and although early output disappointed, the factory eventually delivered aircraft in large numbers. Other countries copied the approach. By the middle of the twentieth century the moving line was the organising principle of factories on both sides of the Atlantic, and the measure by which industrial capacity was counted.

Not everyone was satisfied with it. At Toyota in Japan, Taiichi Ohno and his colleagues concluded that a line designed only to produce as much as possible at any cost produced waste as well as cars. Their alternative, later called lean production, gave workers the power to halt the line when a fault appeared, and required that parts arrive only as they were needed, a system known as just-in-time. The result was both fewer defects and less inventory, and from the 1970s Western manufacturers began to study it closely.

Today the machinery is different. Welding, painting and much of the assembly of a modern car are carried out by robots that never tire and need no wage, but the underlying logic — a product moving through a sequence of fixed, repeated operations — is unchanged. The moving line did not merely speed up production; it reorganised the way work itself was designed and paid for. The warehouses that supply online shopping, where robots carry shelves to a human packer, follow a similar principle.

Ask a dozen people how they found the film they watched last night and most will mention a recommendation. Streaming services, online shops, music platforms and social networks all decide what to put in front of each user by predicting, from past behaviour, which items that person is likely to choose. The scale is difficult to imagine: a large video service may hold hundreds of thousands of titles, while a shopping site may list hundreds of millions of products. No human editor could make those choices, and no single ranking can suit everyone, so the selection is made by software trained on records of what millions of users have done.

The earliest practical methods relied on what is called collaborative filtering. The principle is simple: people who agreed in the past tend to agree in the future. If two users have both enjoyed a large number of the same films, the system predicts that one will like a film the other has already seen. In the early 2000s the online retailer Amazon described an item-to-item version of this idea, in which each product is matched to a list of similar products rather than to similar customers. Interest in the field increased sharply in 2006, when Netflix offered a prize of one million dollars to anyone who could improve its own predictions by ten per cent; the winning team, announced in 2009, combined several different methods.

A second family of methods uses the characteristics of the items themselves. A song can be described by its tempo, key and instrumentation, a news article by the words it contains and the topics it mentions, a product by its price, size and category. A profile of the user's past choices is then compared with these descriptions, so that the system can recommend items that resemble what the user has already enjoyed. Content-based filtering works well when a new item has few users, because it does not depend on other people's behaviour, and it can explain its choices in terms that a person understands. Its weakness is that it tends to be narrow, always offering more of the same.

Neither approach, however, captures the subtlety of real taste. Modern systems therefore learn a compact numerical description of every user and every item, known as a vector or embedding. Each film might be represented by a few hundred numbers, and each user by a similar list. The numbers themselves have no obvious meaning, but the model learns them by adjusting the lists until its predictions about known behaviour are as accurate as possible. The distance between a user's vector and an item's vector then becomes a measure of likely interest. The technique, known as matrix factorisation, became standard in commercial systems.

Even a good model cannot compare every user with every item; the arithmetic would be impossible. Large systems therefore work in two stages. The first, called retrieval, uses fast, simple comparisons to reduce a catalogue of millions to a few hundred plausible candidates. The second uses a much heavier model, often a deep neural network with many layers, to rank those candidates by the probability that the user will watch, click or buy. Because this second stage sees only a few hundred items, it can take account of the context, such as the time of day and what the person has done in the current session.

The model, however, can only optimise whatever it is told to optimise. If the measure of success is the number of clicks, the system will learn to attract clicks, which is not the same as showing people what they value. Watch time, shares and purchases are each different targets, and each produces a different service. Firms decide between them by running controlled experiments in which groups of users are shown slightly different versions, and the effects are compared. The practice is powerful and largely invisible, and the choice of target is a commercial decision rather than a technical one.

Several problems follow from learning only from the past. A new item has no history, so the system has no reason to show it, a difficulty known as the cold start problem; the same is true of a new user. Systems also tend to favour items that already attract attention, because recommending something popular is rarely a visible failure, and the result is a bias towards a small number of titles. Critics have argued further that personalisation may enclose each user in a filter bubble, a term coined by the writer Eli Pariser in 2011 to describe a world in which people encounter mainly views they already hold. The evidence for such an effect is contested, and its strength may depend on the type of content and the design of the service.

Researchers are now trying to build systems that are more diverse, more transparent and less self-reinforcing. One line of work deliberately includes a degree of randomness, sacrificing a little short-term accuracy in order to learn about items the system would otherwise ignore. Others measure the long-term effect of recommendations rather than the immediate click, or attempt to show why an item was chosen. Regulation has begun to take an interest as well. European law now requires the largest platforms to offer users at least one option for recommendations that are not based on profiling their behaviour, and to explain the main parameters used by their systems.

On a quiet stretch of motorway in the early hours, a heavy lorry holds its lane, keeps its distance from the vehicle in front and slows for a curve. There is nobody in the cab. Runs of this kind, first demonstrated on public roads a few years ago, are no longer startling to engineers. What is striking is how quickly the argument about them has settled into two positions: that driverless lorries will put millions out of work, or that they are a distant fantasy that need not trouble anyone yet. Both, I think, are wrong, and the reason lies in the peculiar nature of lorry driving as an occupation.

The technology is furthest advanced where the driving is easiest. Motorways are engineered environments: lanes are marked, junctions are graded, and pedestrians and cyclists are absent. Platooning, in which several lorries travel close together under electronic control so that the followers save fuel by slipping into the leader's slipstream, has been tested on public roads in Europe. A project funded by the European Union demonstrated a convoy on a Spanish motorway in 2012. In the United States, a small company ran a lorry with nobody aboard along a stretch of public toll road in Florida in 2019. Neither demonstration was a commercial service, but together they showed that the essential functions could be performed.

What is easy on a motorway is hard almost everywhere else. A lorry's journey begins at a loading bay, where pallets must be checked against paperwork, secured with straps and, increasingly, loaded by the driver. It ends in a town centre, at a delivery entrance with a customer who is not ready to receive the goods. Weather, roadworks and unexpected obstructions demand judgement. The motorway portion, which attracts the attention, is perhaps the simplest part of the job; the parts that are hardest to automate are those that involve other people.

The case for automation rests mainly on safety. Heavy vehicles are involved in a disproportionate share of serious injuries, and investigations in several countries have attributed the great majority of crashes to human error: distraction, fatigue and misjudgement. To this must be added better enforcement of the rules, since a lorry that cannot exceed the speed limit, or drive longer than the law allows, would remove a persistent problem. These arguments are strong, and they are the reason why the case cannot be dismissed as the enthusiasm of engineers.

There is an economic argument too. Labour is a large part of the cost of moving goods by road, and a lorry that drives itself could run through the night, when roads are clear and demand for delivery is growing. Platooning reduces fuel consumption. The savings would be considerable, and in an industry with narrow margins they would be pursued.

Yet the number of people involved is not small. In the United States, analyses of census data have repeatedly found that lorry driving is the most common occupation in a large number of states, and the industry employs several million people. These are jobs that do not require a university degree, that pay better than much other work open to people with few qualifications, and that are often in regions where alternative employment is scarce. A technology that removed them would concentrate the loss on particular places and particular people.

The optimists reply that this is the wrong way to think about automation. New technology changes tasks rather than abolishing them wholesale; the driver of the future will supervise, load and unload, and take over by remote control when a vehicle meets a problem it cannot handle. On this view the job will become cleaner and safer, and the numbers employed will fall only slowly. Some economists point to the cash machine, which coincided for a time with an increase in the number of bank clerks.

What neither side emphasises is the period in between. The most likely future is not a choice between lorries with drivers and lorries without, but a long stretch in which part of a route is automated and the rest is not. Once the motorway section is performed by a machine, the driver who remains to do the urban deliveries is no longer doing a job that requires long training; the case for a skilled wage weakens, and the work may be reorganised into short shifts at a depot. The employment may survive while the good job disappears. The optimists are not wrong about the past; they are wrong to assume that a job which survives is the same job.

If that analysis is right, the argument should move from the number of jobs to their quality. Regulators who are asked to approve trials should require operators to state clearly who is responsible when something goes wrong. They should also insist that a person supervising several vehicles from a control room is treated as a driver, with limits on hours and proper training. Public money would be better spent on retraining, and on support for freight depots in towns where driving is a major employer, than on subsidies for the technology.

None of this means that automated lorries should be refused. The safety gains are real, and the countries that resist the technology will not be safer, merely less productive. It means that the debate has been asking the wrong question. The interesting question is what kind of work the people who remain in the industry will do, and on what terms.