The TV show The Americans, a cold-war drama set in Washington, DC, in the 1980s, features a robot that delivers mail and classified documents around the FBI office. That an autonomous vehicle existed in the 1980s might seem surprising. Marketed as the Mailmobile, it had first appeared a decade earlier.1
To guide the Mailmobile, a technician would lay out a chemical trail that gave off ultraviolet light from the mail room along the carpeted floors to various offices. The robot used a sensor to slowly follow the trail (at less than one mile per hour) until the chemical markings signaled it to stop. The Mailmobile cost between $10,000 and $12,000 (about $50,000 in today’s dollars), and for an extra fee, the company could attach a sensor to detect obstacles in its path. Otherwise, it just beeped a lot to warn people it was coming. In an office where a human took two hours to deliver the mail, the Mailmobile completed the job in twenty minutes, not stopping for office banter.
The mail robot required careful planning. Even some simple but perhaps costly office reallocations might have been necessary to accommodate the robot’s operation. It could deal with only small variations in its environment.
Even today, many automated rail systems worldwide have extensive installation requirements. For example, the Copenhagen metro uses no drivers, but it works because trains operate in a carefully preplanned setting; only a limited number of sensors inform the robot about its environment.
These limitations are a common feature of most machines and equipment. They are designed to operate in rigid environments. Compared with most equipment on factory floors, the mail robot was notable because many offices could install it relatively easily. But, for the most part, robots need a tightly controlled and standardized environment in which to operate because the equipment does not tolerate uncertainty.
More “Ifs”
All machines—both hard and soft—are essentially programmed using the classic if-then logic. The “if” part specifies a scenario, environmental condition, or piece of information. The “then” part tells the machine what to do for each of the “ifs” (and “if nots” and “elses”): “If the chemical trail is no longer detected, then stop.” The mail robot had no ability to see its surroundings and could only operate in an environment that artificially reduced the “ifs” it could deal with.
If it could distinguish between more situations—more “ifs”—and even if it didn’t change what it did, essentially stop or go at any point, it could have been used in many more places. A modern-day Roomba—the automated vacuum cleaning robot from iRobot—is able to do this and roam freely around rooms with sensors to prevent it from falling down stairs or getting stuck in corners, along with a memory to ensure it covers the floor in a timely fashion.
If a robot operates outside, it needs to move more slowly to avoid slipping when the ground is wet. Two possible situations (or states) arise—dry and wet. If the robot’s motion is also influenced by whether it is light or dark, whether a human is moving in the vicinity or not, whether rush items are in that batch of mail, if it is okay to run over squirrels but not cats, and a variety of other factors, and if the rules are sensitive to interactions (it is okay to run over squirrels if it is dark, but not if it is light), then the number of situations—the number of “ifs”—grows radically.
Better prediction identifies more “ifs.” With more “ifs,” a mail robot can react to more situations. A prediction machine enables the robot to identify that wet dark environments with a human running twenty feet behind and a cat up ahead might require slowing down, but wet dark environments with a human standing twenty feet behind and a squirrel ahead might not. The prediction machine enables the robot to move around without a preplanned trail or track. Our new Mailmobile can operate in more environments without much additional cost.
Delivery robots abound. Warehouses have autonomous delivery systems that can predict their environment and adjust accordingly. Fleets of Kiva robots transport products inside Amazon’s vast fulfillment centers. Startups are experimenting with delivery robots that take packages (or pizza) onto sidewalks and streets from businesses to homes and back again.
Robots can now do this because they can now use data from sophisticated sensors to predict their environment and then receive instructions for how to handle it. We don’t often conceptualize this as prediction, but fundamentally it is. And as it keeps getting cheaper, the robots will get better and better.
More “Thens”
George Stigler, a Nobel Prize–winning economist, reportedly remarked: “People who have never missed a flight have spent too long in airports.”2 While a peculiar logic is in operation here, the counterargument is strong: you can get work done or relax just as easily at the airport as elsewhere, and it might give you some peace of mind to get there early to avoid the hassles of missing a flight. Thus was born the airport lounge. Airlines invented it to provide passengers (or at least wealthy or frequent-flying ones) a convenient and quiet space to wait for their flights. The lounge exists because you are likely to arrive early for your flight. Someone who is perennially late would only use a lounge in transit or when a flight is delayed or to weep when they miss their flight to Bali. The lounge is there to provide some wiggle room, a bit of a buffer for when your arrival time is less than precise (which is likely to be quite often).
Suppose you have a flight at 10 a.m. Airline guidelines say you should arrive sixty minutes beforehand. You could arrive at 9 a.m. and make your flight. Given that, what time should you leave for the airport?
You usually can get to the airport in thirty minutes, which might allow you to leave at 8:30 a.m., but that does not account for traffic disruption. When flying back to Toronto from a New York meeting about this very book, we three experienced such bad traffic to LaGuardia Airport that we ended up walking the last mile along the freeway. That could easily add another thirty minutes (more, if you are risk averse). Now you are back to 8 a.m., which is when you leave every time you don’t know what traffic is going to be like. As a result, you usually end up spending thirty minutes or more in the lounge.
Apps such as Waze provide very accurate travel times from your current location to the airport. Such apps monitor both real time and historic traffic patterns to both forecast and update the quickest routes. Pair that with Google Assistant, and you can account for any delays that might appear for your flights with other apps that monitor historical delays or the location of a connecting aircraft. Together, these apps mean that you can reliably trust the prediction, which opens up new options such as “unless there is a traffic problem, leave later and go directly to the gate” or “if there is flight delay, leave later.”
Better prediction, by reducing or eliminating a key source of uncertainty, eliminates your need to have a place to wait at the airport. More critically, better prediction enables new actions. Rather than having a hard-wired rule to leave two hours before your flight, you can have a contingent rule that takes information and then tells you when to leave. Those contingent rules are if-then statements and enable more “thens” (leave early, on time, or later), depending on more reliable predictions. So, in addition to producing more “ifs,” prediction expands opportunities by increasing the number of feasible “thens.”
Mail robots and airport lounges have something in common: they are both imperfect solutions to uncertainty, and they both will be undermined by better prediction.
More “Ifs” and “Thens”
Better prediction allows you to predict more things more often, reducing uncertainty. Each new prediction also has an indirect effect: it makes choices feasible that you would not have considered before. And you don’t have to explicitly code the “ifs” and “thens.” You can train the prediction machine with examples. Voilà! Problems that were not previously understood as prediction problems may now be tackled as such. We were compromising without recognizing it.
Such compromises are a key aspect of how humans make decisions. Economics Nobel Prize–winner Herbert Simon called this “satisficing.” While classical economics models superintelligent beings making perfectly rational decisions, Simon recognized and emphasized in his work that humans cannot cope with complexity. Instead, they satisfice, doing the best they can to meet their objectives. Thinking is difficult, so people take shortcuts.
Simon was a polymath. In addition to a Nobel, he also won the Turing Award, often called the Nobel of computing, for “contributions to artificial intelligence.” His economics and computing contributions were related. Echoing his thoughts on humans, his 1976 Turing Award lecture emphasized that computers “have limited processing resources; in a finite number of steps over a finite interval of time, they can execute only a finite number of processes.” He recognized that computers—like humans—satisfice.3
The mail robots and airport lounge are examples of satisficing in the absence of good prediction. Such examples are everywhere. It will take practice and time to imagine the possibilities enabled by better prediction. It is not intuitive for most people to think of airport lounges as a solution to poor prediction and that they will be less valuable in an era of powerful prediction machines. We are so used to satisficing that we don’t even think of some decisions as involving a prediction.
In the translation example earlier in the book, specialists saw automatic language translation not as a prediction problem but as a linguistic one. The traditional linguistic approach used a dictionary to translate word by word, coupled with some grammatical rules. This was satisficing; it led to poor results because of too many ifs. Translation became a prediction problem when researchers recognized that translation could happen sentence by sentence or even paragraph by paragraph.
Translation with prediction machines involves predicting the likely equivalent sentence in the other language. Statistics enable the computer to choose the best translation by predicting the ifs—which sentence a professional translator is most likely to use based on translation matching in the data. It relies on, remarkably, no linguistic rules. A pioneer of this field, Frederick Jelinek remarked, “Every time I fire a linguist, the performance of the speech recognizer goes up.”4 Clearly, this is a scary development for linguists and translators. All sorts of other tasks—including image recognition, shopping, and conversation—are being identified as complex prediction problems that are amenable to the application of machine learning.
By enabling more complex decisions, better prediction can lower risk. For instance, one of the practical applications of recent AI is in radiology. Much of what radiologists currently do involves taking images and then identifying issues of concern. They predict abnormalities in images.
AIs are increasingly able to perform that prediction function at human levels of accuracy or better, which can assist radiologists and other medical specialists in making decisions that have an impact on patients. The critical performance metric is the accuracy of the diagnosis: whether the machine predicts a disease when the patient is ill and predicts no disease when the patient is healthy.
But we must consider what such decisions involve. Suppose doctors suspect a lump and must decide how to determine if it is cancerous. One option is medical imaging. Another option is something more invasive, like a biopsy. A biopsy has the advantage of being highly likely to provide an accurate diagnosis. The problem, of course, is that a biopsy is invasive; thus, both doctors and patients prefer to avoid it if the likelihood is low that the issue is serious. One job of a radiologist is to provide a reason not to conduct an invasive procedure. The ideal is to perform a procedure only to confirm a serious diagnosis. The biopsy offers insurance against the risk of not treating a deadly disease, but it comes at a cost. The decision to undertake the biopsy depends on how costly and invasive the biopsy itself is and how bad it would be to overlook the disease. Doctors use these factors to decide whether the biopsy is worth the physical and monetary costs of the invasive procedure.
With a reliable diagnosis from an image, patients can forgo the invasive biopsy. They can take an action that, absent the prediction, would be too risky. They no longer have to compromise. Advances in AI mean less need for satisficing and more “ifs” and more “thens.” More complexity with less risk. This transforms decision-making by expanding options.
KEY POINTS
- Enhanced prediction enables decision-makers, whether human or machine, to handle more “ifs” and more “thens.” That leads to better outcomes. For example, in the case of navigation, illustrated in this chapter with the mail robot, prediction machines liberate autonomous vehicles from their previous limitation of operating only in controlled environments. These settings are characterized by their limited number of “ifs” (or states). Prediction machines allow autonomous vehicles to operate in uncontrolled environments, like on a city street, because rather than having to code all the potential “ifs” in advance, the machine can instead learn to predict what a human controller would do in any particular situation. Similarly, the example of airport lounges illustrates how enhanced prediction facilitates more “thens” (e.g., “then leave at time X or Y or Z,” depending on the prediction of how long it will take to get to the airport at a particular time on a particular day), rather than always leaving early “just in case” and then spending extra time waiting in the airport lounge.
- In the absence of good prediction, we do a lot of “satisficing,” making decisions that are “good enough” given the information available. Always leaving early for the airport and often waiting once you arrive because you’re early is an example of satisficing. That solution is not optimal, but it’s good enough given the information available. The mail robot and the airport lounge are both inventions designed in response to satisficing. Prediction machines will reduce the need to satisfice and thus reduce the returns to investing in solutions like mail robot systems and airport lounges.
- We are so used to satisficing in our businesses and in our social lives that it will take practice to imagine the vast array of transformations possible as a result of prediction machines that can handle more “ifs” and “thens” and, thus, more complex decisions in more complex environments. It’s not intuitive for most people to think of airport lounges as a solution to poor prediction and that they will be less valuable in an era of powerful prediction machines. Another example is the use of biopsies, which largely exist in response to weaknesses in prediction from medical images. As the confidence in prediction machines goes up, the impact from medical imaging AIs may be much greater on the jobs associated with conducting biopsies because, like airport lounges, this costly and invasive procedure was invented in response to poor prediction. Airport lounges and biopsies are both risk management solutions. Prediction machines will provide new and better methods for managing risk.