Welcome to AI book reviews, a series of posts that explore the latest literature on artificial intelligence.
Despite its promising advances, artificial intelligence has yet to cause a transformational change in many industries. And in many cases, the problem is not necessarily with the technology but with the way we perceive it.
Power and Prediction, a new book by Ajay Agrawal, Joshua Gans, and Avi Goldfarb, explores the fundamental challenges standing in the way of AI adoption in different industries. A sequel to their acclaimed Prediction Machines, the new book discusses what needs to change before organizations can benefit from the full potential of advances in artificial intelligence.
From point solutions and applications to AI systems, Agrawal, Gans, and Goldfarb study the success and failure of AI in different fields. They also provide important insights from past technological revolutions and show how rethinking and redesigning our systems from scratch can help create true value based on powerful machine learning and deep learning algorithms.
Point solutions vs AI systems
Today’s AI systems are prediction machines, which means they can predict what happens in the future based on past data. This is basically what every mathematical model does. But thanks to the availability of very large amounts of data and compute as well as advances in deep learning algorithms, we’ve been able to create models that can make predictions on complex information such as images, text, and multi-dimensional data.
In Power and Prediction, the authors break down the value of AI into three categories: point solutions, application solutions, and systems solutions.
Until now, most of what we’ve seen are point solutions and application solutions. These are AI systems that replace tasks that previously required prediction. For example, in financial services, one of the tasks is to predict which transactions are fraudulent. A machine learning model trained on the right data can take over this task. Point solutions are the low-hanging fruit of AI because adopting them requires minimal investments and changes to the underlying system.
Another example of a point solution is analyzing radiology scans. There are now several deep learning models that can detect various diseases from x-ray and MRI scans at a level that matches experienced radiologists.
They are automating one of the many tasks that a radiologist performs without making any changes to the underlying patient care system.
AI systems can provide much greater value by automating new tasks and problems that are unaddressed in current applications and systems. However, AI systems require a blank-slate approach, in which you redesign the entire processes, workflows, and applications that not only solve existing problems, but also new ones. To make them work, AI systems often require new organizational structures and alignment of goals and incentives. This makes AI systems harder and riskier, but also more rewarding.
“System solutions are typically harder to implement than point solutions or application solutions because the AI-enhanced decision impacts other decisions in the system,” the authors of Power and Prediction write. “Whereas point solutions and application solutions often reinforce existing systems, system solutions, by definition, upend existing solutions and therefore often result in disruption. However, in many cases, system solutions are likely to generate the greatest overall return to investments in AI.”
The Between Times of AI
In Power and Prediction, the authors suggest that we are in the “Between Times” of AI, “after witnessing the power of this technology and before its widespread adoption.” This is why point and application solutions are currently the more attractive and sought-after use cases of AI.
There is a historic precedent for this. For example, in the late nineteenth century, when electricity was beginning to be industrialized, its first applications were point solutions. For factories, this meant lowering the cost of energy by replacing their steam engines with electromotors. Changing the source of power did not require redesigning factories.
However, the real value proposition of electricity was decoupling the machine from the power source. This enabled new factory designs that were impossible with steam power, and they were more productive and less costly. But it took decades for this adoption to take place because it required fundamental changes, breaking habits, and upfront investments that incumbents were not willing to make. The entrepreneurs who did take advantage of the opportunity managed to take the lead and capture a huge part of markets that later replaced the old ones.
You can see these shifts in many other industries, such as the rise of online shopping, the advent of personal computers, and the transition from print to digital media.
AI is an infrastructural technology, whose impact tech leaders have compared to that of electricity. Therefore, it requires a new mindset and daring exploration.
The authors of Power and Prediction write, “AI-driven industry transformation takes time. It’s not obvious how to do it at first. Many will likely experiment and fail because they misunderstand demand, or they can’t get the unit economics to work. Eventually, someone will succeed and establish a pathway to profitability. Others will try to imitate. The industry leader will attempt to create moats to protect its advantage. Sometimes it will succeed. Regardless, the industry will transform, and as always, there will be winners and losers.”
Breaking the rules
Per Power and Prediction, “When you don’t have something, you don’t just give up. You compensate for it. If you don’t have the information you need to make an informed choice, you insulate yourself from the consequences of having to do things blindly. Thus, when AI prediction comes along, it shouldn’t be a surprise that the opportunities for its use are not immediately obvious. Would-be decision-makers have built up a scaffolding based on not having that information.”
AI opportunities are difficult to spot because they are usually hidden behind rigid rules and procedures that are working well and have been established for a long time. These rules compensate for our lack of information. They enable us to make decisions without being able to predict the exact outcome. And they help build systems that, while not optimal, work reliably and for many cases.
The key to finding these opportunities is to first, understand the power of prediction machines, and second, find places where prediction can replace hard-set rules. A very interesting example that the authors explore in the book is the use of AI in education.
Thanks to machine learning algorithms and historical data, we can predict how students will perform, where they will excel, and where they will struggle. This gives us the opportunity to provide more personalized content to each student.
But these predictive models can’t be of much help in the current education system, which is founded on age-based curricula with a single teacher per class. This system was established because we didn’t have ways to exactly measure the individual learning capabilities of students through their education trajectory.
To be able to take full advantage of machine learning, we need to rethink the education system in a new way. This new system will replace the age-based curriculum with personalized discussions, group projects, and teacher support and can result in a much bigger impact on overall education and personal growth and development.
“The age-based curriculum rule is the glue that holds together much of the modern education system, and so an AI that personalizes learning content can only provide limited benefit in that system,” the authors of Power and Prediction write. “The primary challenge for unleashing the potential of a personalized education AI is not building the prediction model, but unsticking education from the age-based curriculum rule that currently glues the system together.”
The successful adoption of AI requires what the authors of Power and Prediction call the “system mindset,” which stands in contrast to the “task mindset.” The task mindset focuses on cost savings. The system mindset focuses on value creation. The task mindset focuses on automating individual tasks. The system mindset recognizes the need to reconstitute systems that generate value based on machine prediction and human decisions.
We have already seen this happen in some industries and large tech companies such as Amazon and Google, which have shaped profitable systems that recommend personalized content based on AI predictions.
Perhaps one of the important elements of the system mindset is the power shift that takes place with the adoption of AI. As systems change, so do the people who have the power to make decisions.
“While AI cannot hand a decision to a machine, it can change which human is making the decision. Machines don’t have power, but when deployed, they can change who does,” the authors of Power and Prediction write. “When machines change who makes decisions, the underlying system must change. The engineers who build the machines need to understand the consequences of the judgment they embed into their products. The people who used to decide in the moment may no longer be needed.”
One hypothetical example that the authors explore in the book is heart attack risk. Currently, such risk assessments are made through tests at hospitals, and the decisions are made by specialist doctors who administer the tests.
Say we are able to build AI systems that can predict heart attack risk based on data collected by wearables such as a smartwatch. Then, it might be possible to move those predictions out of the triage space in a hospital’s emergency department to a patient’s home. In this case, many patients will never need to go to the hospital, after being diagnosed with something that a pharmacist or a primary care physician can help treat at home.
Regardless of where we stand on the scientific and philosophical debates surrounding AI, what we can all agree on is that prediction machines have much to offer and we are only scratching the surface. Being able to leverage their full potential starts by going back to the drawing board and rethinking how we would design our systems if we had the power to predict.