In the summer of 2018, wildfires in the forests of British Columbia coupled with winds from the north were driving hazy smoke across the normally crystal-blue skies of Washington State. More than one-third of the nation was experiencing moderate to exceptional drought, yet crops like wheat, corn, and soybeans were beginning to pour into markets at abundant rates, driving down prices for farmers, including the farmer I was driving to visit. Meanwhile, in Washington, DC, a pitiful scrum of white nationalists and a much larger group of antifascists had spent the weekend exchanging heated insults in the streets of the nation’s capital. The farm bill—the federal government’s primary instrument for agricultural policy—languished over the summer break in a reconciliation committee while a trade war had broken out, prompting the administration to propose giving farmers $12 billion in federal funding as a stopgap measure while international negotiations proceeded.
It was against this backdrop that I was visiting a small farm about fifteen miles east of Microsoft’s sprawling campus in Redmond. A group of developers in the company’s AI and Research group had invited me to see how they are using low-cost sensors and digital mapping to gather agricultural data like soil temperature and moisture to improve crop outcomes. A smallholder farmer just outside of Carnation, a community named for the evaporated milk company, had allowed the researchers access to his land in exchange for valuable data.
Seattle’s rush hour traffic was spilling onto the rural roads just outside the farm. We parked our car beside a big red barn and then had to sprint across the road to avoid being hit by commuters racing home. Not all farmers are at the end of a dirt road in a flyover state. In his thought-provoking book Hinterland: America’s New Landscape of Class and Conflict, Phil A. Neel divides his observations between a far and a near hinterland—deep rural and just barely. While Richard Florida’s work on the creative class explores why economic development is concentrated in cities and metropolitan areas, Neel explains what’s happening just beyond. “But beyond the city, where there is little question of inclusion, it becomes clear that these populations are also unified by something else: the commonality that comes from being increasingly surplus to the economy, though also paradoxically integral to it.” Rural regions, he concludes pessimistically, are becoming wastelands for global production.
Not if Dr. Ranveer Chandra, a Microsoft researcher, has anything to say about it. Ranveer is a cheerful product of the famed Indian Institutes for Technology and Cornell University’s computer science department. He’s walking us through endless rows of produce ripening under today’s filtered Pacific Northwest sunshine. Ranveer grew up in the Indian state of Bihar, which borders Nepal. He worked a small ten-acre farm with his grandparents, who grew sugarcane, wheat, rice, and mangoes. His passion is to create technologies that both help rural people like his grandparents and the friends he left behind, and that help to increase food production by up to 70 percent to feed the growing population of the world. And while he believes in the power of tech, he believes even more in the human element that is needed to grow food, to cultivate the earth.
“We want to supplement the farmer’s knowledge,” he tells us while unpacking a small white drone from a case. “Rather than intuition alone, we can use data. Rather than water and pesticide everywhere, we can be more precise so that it’s only used where it’s needed.”
With that he flips open a Surface book and pulls out his mobile phone. He’ll need these to monitor the images and data as they flow in. The drone lifts off about one hundred feet in the air. From his notebook we can see the rows of plants from above, and follow its path as it zigs and zags across the entire acreage, collecting the soil’s pH levels, moisture, and temperature. With this data it can build a heat map to help the farmer pinpoint problem areas. The data that’s collected, as much as 200 megabytes of images per acre, is transported to the cloud’s edge back at the farmer’s house, where another computer stores and analyzes the information in real time.
There has been a lot of focus in recent years on cloud computing, rightly so, but edge computing is also fundamental to the future of AI in precision farming, medicine, and other applications.
With edge computing, data is run and analyzed on a nearby device, close to where it’s being generated, rather than flowing to a faraway data center. This way data can be analyzed in real time, rather than being hostage to a rural area’s low-latency, slow connectivity to the cloud. With an estimated 25.1 billion devices expected to be connected to the Internet by 2021, edge computing will empower and transform the Internet of things (IoT), like this farm, for years to come.
Take a look around your house, office, or even the next store you visit, and you’ll start to notice that Internet-connected devices are bringing us closer than ever before to a world of ubiquitous computing and ambient intelligence. As these Internet-of-things devices become increasingly commonplace, people will start to expect computing to be more integrated into their lives, to anticipate, understand, and seamlessly meet their needs. They will expect software to respond to spoken natural language, gestures, body language, and emotion, and for it to understand the physical world and the rich context surrounding each user as they navigate their personal life, their work, and the world around them.
This trend has more promise than just bringing additional convenience, productivity, and connections to our everyday lives. Smart sensors and devices are breathing new life into industrial equipment from factories to farms, helping us navigate and plan for more sustainable urban cities and bringing the power of the cloud to some of the world’s most remote destinations. With the power of AI enabling these devices to intelligently respond to the world they are sensing, we will see new breakthroughs in critical areas that benefit humanity like health care, conservation, sustainability, accessibility, disaster recovery, and more.
We call this next wave of computing the intelligent edge and intelligent cloud. And these new technologies may be just as transformative for problems rooted in the physical world as AI has been for the consumer Internet. The raw material for building things with modern AI is data. The data that you have dictate what the AI is able to use. If those data are clicks on search results, ads, likes, and shares, then you can build AI systems that can optimize the quality of your search results, that can show you better ads, and that can learn what information to show you in a news feed based on what appear to be your preferences. If those data are soil moisture, pH, temperature, humidity, weather forecasts, and images of crops, then you can build AI systems that can help farmers make better decisions about what to plant and when, how to irrigate more effectively, how to deploy fertilizers and supplements, and when to harvest. The net effects of making better decisions about these things are higher crop yields with less environmental impact.
This extension of the power of modern AI into the physical world is truly exciting. I would argue that AI will ultimately have a far greater positive impact on our lives in the era of the intelligent edge and cloud than it has had through the aperture of the consumer Internet alone. Smart sensors that can collect data and act on physical events as they happen will enable us to build models that have knowledge about, and the ability to interact with, an incredible diversity of things in the physical world. And those models will be able to assist humans in making better, quicker, higher-quality decisions, or used in systems that take autonomous action to supplement humans when environmental conditions and scale merit.
In the world of conservation, the intelligent edge and cloud have special promise. Conservation efforts are often constrained by the cost and scale of the effort (e.g., observing animals over millions of acres of their habitat). Even if you could afford it, you couldn’t hire enough humans to do all the animal and habitat conservation work that exists in the world today. Disney’s Animal Kingdom is already leveraging the intelligent edge to study the purple martin bird. They worked with Microsoft to develop hundreds of tiny “smart houses” in Disney’s Animal Kingdom to learn more about the species and help inspire a new generation of conservationists in the parks. The scientists now have unprecedented insight into the nesting behavior of the purple martins.
At a grander scale, there are efforts underway to deploy sensors—cameras, microphones, motion detectors, and more—at massive scale for several conservation tasks. Smart Parks, a nonprofit wildlife conservation organization that is trying to prevent poaching of endangered or threatened species, has deployed a network of sensing technologies in Rwanda, Tanzania, and India to monitor and alert for potential poaching activity over thousands of square kilometers of habitat. Another conservation organization, Rainforest Connection, is transforming recycled smartphones into smart sensors to detect illegal deforestation in rain forests so that authorities can be alerted and act before these precious ecosystems can be harmed. In the case of both conservation efforts, deploying the sensors and getting the data is the first step in a process where AI will be able to deliver higher-precision detection of harmful activity and better response automation to handle the truly vast scope of the conservation task. Our human resources are too scarce to solve these problems without the help of AI.
Another noteworthy area where the intelligent edge and cloud paradigm can have a huge potential positive impact is climate change and conservation of our natural resources. How do we transition to renewable sources of energy as quickly as possible while immediately mitigating the deleterious effects of consuming carbon-based fuels (oil, gas, coal, wood)? One of the challenges in navigating this transition is economic: it will be extremely expensive, costing tens or hundreds of trillions of dollars, to fully complete; and using abstinence as a strategy for managing the cost of the transition is fraught with political peril, particularly in parts of the world with rapidly expanding economies where higher energy consumption goes part and parcel with a higher standard of living. The costs of not making the transition are catastrophically high, which means that this is not a problem that will just go away. Nor will it get better over time.
As we collect more data about physical phenomena, using AI to optimize power production, transmission, and consumption may allow us to fundamentally alter the economics and speed of the clean energy transition. Already Schneider Electric is using the intelligent edge in oil fields to monitor and configure pump settings and operations remotely, sending personnel on-site only when necessary for repair or maintenance when, for example, intelligent pump monitoring indicates that something will go wrong. This contributes to overall worker safety and improved resource management. But there is also the potential to make our electric power generators more efficient, to more efficiently generate and transmit power to points of consumption when needed, and to make power consumers themselves more efficient. The holy grail here, in my opinion, is using the intelligent edge and cloud, and sophisticated AI, to globally coordinate the production and consumption of power to achieve higher efficiency. We do this now with crude mechanisms like higher prices when demand is high, and in the worst case, rolling brownouts when the grid is oversubscribed. We perhaps already possess the technological capability to make generation and consumption of energy smart, with the networks and cloud infrastructure to do this global coordination.
You may have noticed a pattern emerging from the agriculture, conservation, and energy examples. We as a country, in most cases as a species, face some challenging problems that may at first seem to be zero-sum. In other words, the problem is so much bigger than the resources available to solve it that we are not only forced to solve a small subset of the problem, but we sometimes are confronted with extremely contentious public debates about how to allocate our finite resources to solve even a constrained version of the problem. One of the themes of this book is that AI and advanced automation can be a tool to turn these constrained, zero-sum problems into non-zero-sum ones. In other words, AI can be used to create a new type of abundance that can then be used to break the zero-sum gridlock, whether that’s how to feed a growing population, how to conserve our precious natural resources, or how to solve climate change. AI isn’t a miracle and can’t completely solve these gigantic problems by itself, but it can be an extremely effective tool in helping us make progress that might otherwise be impossible.
We need to give all organizations and developers the tools to build these kinds of increasingly ambitious solutions that span the intelligent edge and intelligent cloud. Moreover, these tools must give developers strong security foundations and help them to place security at the very core of their solutions. Devices on the edge handle some of our most sensitive business and personal data in our homes, workplaces, and sometimes physically remote places.
As we talk about an increasingly connected world with smart sensors and edge computing, we can’t avoid thinking about and investing in security. To protect our data and privacy, security needs to be baked in from the silicon to the cloud. This has been one of the central design principles of Microsoft’s intelligent edge products and services. Many cloud providers are working hard on this. Azure Sphere is our intelligent edge solution to power and protect connected microcontroller unit (MCU)–powered devices. There are nine billion of these MCU-powered devices shipping every year, which power everything from household stoves and refrigerators to industrial equipment. With more processing power than traditional MCUs and a holistic security approach, Azure Sphere will, we believe, make our increasingly connected world safer. In addition, Azure IoT Edge enables you to run cloud intelligence directly on IoT devices and includes security—from device provisioning and management to hardware and cloud services—that runs on top of the devices. Azure Stack, just one of our many tools to power hybrid scenarios, offers customers the flexibility to securely deploy in the cloud, on-premises or at the intelligent edge.
Every cloud provider should possess all these building blocks, and every developer and solutions architect should have a choice of tools in their arsenal for designing better security in a sensor-rich world of edge computing.
There are several innovations Ranveer and his team are working on to advance precision agriculture in their experimental work on the farm in Carnation. The first is solving the lack of Internet connectivity in rural America. From a primitive shed in the midst of the field, a solar-powered television antenna exploits what’s known as TV white space—unused portions of UHF and VHF radio waves that can be used to spread the Internet and Wi-Fi at very low costs. This arrangement has transformed the surrounding field into an agrarian Internet-of-things network. The second innovation is the aerial mapping to feed the data model. Using machine learning techniques like visual similarity and spatial smoothness, the aerial imagery makes studying a field like studying a gene or a cell. It provides the farmer with the pulse of the farm. Finally, the team is working to bring down costs and increase ease of use for farmers, particularly for air, plant, and soil sensors.
I ask Ranveer if these AI tools are going to augment or replace farmers. He smiles and says he gets that question a lot, including from a congressional committee chairman who had just visited the week before.
“We are purposefully not automating. The goal is to augment.”
He notes that all of this requires humans. There is setup and assembly required. Human analysis and judgment are required. He said county extension agents are quite enthusiastic because they will be able to offer even better advice and guidance to farmers. And Ranveer’s group is working with Future Farmers of America and North Dakota State University to train students.
Not far from North Dakota, in the great agricultural state of Nebraska, former Microsoft executive Jeff Raikes now runs his family’s large farm and ranch in a community located between Omaha and Lincoln. I never worked with Jeff, but he is the legendary creator of the Microsoft Office productivity suite and has long used computers to help run the family farm. Not surprisingly, the Raikes farm continues to pioneer the nexus of tech and agriculture. He notes that in the 1940s it took three hundred people to run the family farm. Today it takes far fewer. Wi-Fi–enabled feedlots, for example, provide cattle just the right amount of food and water.
But it’s not as simple as that. Farming has become more technical. He points out that he learned to drive a tractor when he was seven years old, but today would be unqualified to run any of the tractors on his farm.
“You have to be technically literate to work on this farm,” he said.
Raikes has a theory that a smallholder farmer today is one that is below the “minimum viable economic unit.” By his estimation, in the last century that MVE was about 40 acres. In the 1970s it was about 400 acres, and today it’s 1,000 to 1,500 acres—depending on the soil and weather conditions of where your farm is located. In addition to great crop productivity and lower environmental impact, AI technologies for agriculture are already becoming accessible to small farmers, and those technologies could significantly change the MVE, giving small farms located practically anywhere the ability to prosper and to create jobs in their local communities. By making small farms easier to start, cheaper to operate, more productive, and more competitive in the marketplace, we will get more small farms in more communities along with the jobs that accompany them.
Granted, this means that the future of farming may look a bit different. Future farmers will need some minimum level of technical skill. But I don’t see this as some great obstacle. Training for these technical skills should come from school, in perhaps a twenty-first-century version of the “Ag” classes I took in middle and high school. Farming has never been easy, and good farmers have always been great technicians with a mastery of soil, seed, and the technical equipment that has been making farms more productive at a regular pace for hundreds of years. AI will be just another tool in the farmer’s arsenal. Moreover, if we all do our job right building an AI platform that’s useful for agriculture—and conservation, energy production, medicine, etc.—it may very well be one of the easiest tools to master. We’ll talk more later about how cloud platforms have a variety of economic incentives to make AI more accessible, how open source is a democratizing influence for AI tools, and how new breakthroughs like machine teaching can further lower the barriers to entry for AI.
In addition to the immediate economic stimulus that AI and advanced automation could potentially bring to rural economies by making small businesses easier to start and more competitive, both the AI platforms and the AI businesses that they power are without doubt going to bring with them brand-new jobs that don’t exist today, and that even for the futurists among us might be difficult to imagine. According to LinkedIn’s 2017 US Emerging Jobs Report, 65 percent of children entering primary school today will ultimately hold jobs that don’t yet exist. When I joined Google in 2003, data scientist wasn’t yet a role. Today data science is one of the fastest-growing professions in the world, and we have a shortage of over 150,000 of them to fill open roles, according to LinkedIn’s August 2018 Workforce Report.15 It is difficult to predict what new jobs AI may soon create, but it’s not difficult to believe that the jobs will come based on the history of technology revolutions both recent and further in the past.
Two new types of AI jobs are already beginning to become important. There is a growing number of independent software vendors (ISVs) who specialize in helping customers build their products and run their businesses with AI. Those ISVs are popping up wherever there are businesses needing help on their AI journey, and those customers are increasingly outside of Silicon Valley. And then there are machine teachers.
Machine teaching is exactly what it sounds like. If you would like to train a model to look for aphids, for instance—insects that can harm crops and forests—knowledgeable people need to examine thousands of images and label how many aphids are on a plant, and at what life cycle. There are tens of thousands of teachers employed right now labeling data and in other ways producing the specific type of knowledge that machine learning systems need to digest in order to build accurate models for the tasks in which they are to be employed. If you think about all the uses to which AI is going to be put, and all the little tasks that we are going to want AI to do, we are going to need lots of machine teachers to help train the AIs. Those are real jobs for the future, and they are ones where geographic concentration is neither necessary nor particularly beneficial. Later in the book we’ll talk about machine teaching more broadly, which is much more powerful than just data labeling and has the very real potential to create lots of geographically diverse teaching jobs for the millions and millions of AI models that our future economy will be built upon.
This question of what jobs will look like in an age of AI and advanced automation, particularly those jobs that could theoretically be completely automated away, is a question preoccupying a lot of economists. Writing for the journal Science, Erik Brynjolfsson and Tom Mitchell examine the workforce implications of machine learning and conclude that for human tasks to be automated, those tasks would need to meet all of these eight criteria:
That may sound like some farm chores, but it doesn’t sound like the work of a farmer to me.
PricewaterhouseCoopers (PWC), the global accounting firm, published a report in July 2018 that looked at this question in the United Kingdom. They estimate that the countervailing displacement and income effects on employment are likely to broadly balance each other out over the next twenty years, with the share of existing jobs displaced by AI likely to be approximately equal to the additional jobs that are created. PWC identified several policy areas where action could help to maximize the benefits of AI, including mitigation strategies such as job retraining for displaced workers.
Walking back to our cars, I tell Ranveer we need to figure out how to get his team’s work to scale. I love that they’ve started with small farms, because the lessons learned can apply to millions of smallholder farmers around the world who rely on farming for survival. By some estimates, half the people in the world work in agriculture, and many of those are self-employed as farmers. The work Ranveer is doing can help them increase yields and keep costs down with the goal of moving out of poverty.
Curious about the news I was reading from Washington, DC, I spoke by phone that same week with Burton Eller, the national legislative director for the National Grange, an organization that describes itself as working to “bring America’s farmers, ranchers and other rural residents the resources they require to stay current and competitive in today’s ever-changing global and local economies.” In addition to helping to shape farm policy since 1975, Burton is also a farmer in a region of Virginia not too far from Gladys, where I grew up. He runs 150 acres of beef cattle on the Virginia and Tennessee border.
From his office overlooking Lafayette Square and the White House beyond, Burton has seen administrations come and go. He explains that except in rare instances, the farm bill is written by Congress, not the president. An exciting time of bipartisanship was the 1970s when Senator George McGovern of South Dakota, a Democrat, and Senator Robert Dole of Kansas, a Republican, worked together to say the farm bill is needed both to feed people and to help create strong farms. They worked toward a more aligned policy, trying to balance the interests of food and farm. An estimated 80 percent of the bill is focused on food programs for the needy such as the Supplemental Nutrition Assistance Program, or SNAP for short. The remaining 20 percent deals with farm supports. One such program is the Rural Microentrepreneur Assistance Program, which provides low-interest loans and grants to support business creation in small towns and rural areas. But these programs are not backed by mandatory funding. Over the summer of 2018, the Aspen Institute, which has long hosted a rural-issues coalition, noted that even though the Rural Development component (Title VI of the Food, Conservation, and Energy Act of 2008) accounts for less than 1 percent of the farm bill’s proposed outlays, “we know firsthand that rural America relies on this funding.”
Over the past several decades farm policy has become increasingly market-based. Burton remembers when farm bills were more about paying farmers not to produce in order to manage the nation’s supply, demand, and ultimately prices.
“Now it’s market-based,” he said. “It’s up to the farmer and the good Lord if it fails. Now there is disaster insurance. We’ve changed dramatically.”
But little prepared him for the seismic change in tone surrounding his issues since 2016. He said in the days and weeks following the election, members called to ask what he was seeing and hearing. Looking out his window at Lafayette Square, he joked that people seemed to be gathering for a spelling bee.
“They are learning how to spell r-u-r-a-l,” he laughed. “I’ve never seen such sensitivity to rural.”
Jonathan Rodden of Stanford, quoted in the Economist, reports that nearly half the variance in the county-level vote shares in the presidential election of 2016 could be explained solely by their number of voters per square kilometer.
Federal initiatives are increasingly shaded with a dose of “here’s what this can do for rural people.” One example is Federal Communications Commission (FCC) policy, which is prioritizing the spread of broadband Internet into rural communities. It helps, he said, that the FCC commissioner, Ajit Pai, grew up in the rural southwest corner of Kansas. A memorandum of understanding was signed among various federal departments that manage land to bring down barriers that might inhibit the widening of rural broadband. Even health initiatives now carry the noble purpose of telemedicine for rural people.
“The federal government is not going to spend money on taking 5G [fifth-generation wireless systems] to every corner of the country. But it is paying a lot of attention to rural areas.”
I wrote earlier that in my hometown, churches played a big role in helping families get through rough times. The tragedy of hunger is often too big, though, for one church or one nonprofit. I was encouraged to see churches band together to end hunger through an advocacy group in Washington, DC, called Bread for the World. As the 2018 farm bill was debated in Congress, Bread issued a revealing report, “The Jobs Challenge.” The report faced head-on a dispute between Republicans and Democrats over how much of a work requirement should be included in hunger programs. In it, ministers explored the “dignity of work,” and make the case for policies that will help create more job opportunities. One that caught my eye was this: “In rural areas, invest in high-speed Internet to overcome barriers to jobs, education and social services.”
Bestselling author John Grisham is known to the world for his The Pelican Brief, The Firm, and other legal thrillers, many set in the South. But his relative Vaughn Grisham perhaps surpasses his renown in the world of rural economic development. Vaughn Grisham’s Tupelo: The Evolution of a Community is something of a classic of its own kind. According to its foreword, “The Tupelo Story does not offer a one-size-fits-all blueprint for community development. . . . And yet, for many of his listeners, whether they live in Maine or Montana, Grisham’s narrative does serve as a sort of catalyst, inspiring them to imagine how they might best build upon the one asset that all small communities have in common, regardless of their external circumstances—their people.”16 One of the guiding principles of the Tupelo model is that local people must address local problems. Each person should be treated as a resource. So the community development process begins with the development of people. The goal of community development is to help people help themselves. Community development must be done both locally and regionally if the full benefits are to be achieved. Never turn the community development process over to any agency that does not involve the people of the community. Expenditures for community development are an investment—not a subsidy—and will return gains to the investors. So people with money have both the responsibility and an interest in investing in the development of their own community.
The intersection of AI, agriculture, and rural development is a good case study for the broader set of ideas and issues that will come up repeatedly as we think about AI’s benefits and costs on other segments of the workforce. I don’t purport to have all the answers, but there are some clear things we must do if we want to realize a democratized version of AI and advanced automation whose benefits will be equitably distributed, and where rural development is a first-order objective, not an afterthought.
First and foremost among these is really ensuring that we build AI as a platform to serve both the needs of big technology companies and the very long list of AI applications that folks will want and need to build to serve their businesses, large and small, local and global. As I’ve mentioned, companies with cloud platforms have a very clear economic incentive to make AI more accessible: storing data for AI, training AI models, and using AI in applications results in the consumption of cloud services. I’ve also mentioned that many AI tools are being built as open source software, which means that communities of developers can easily form around the tools and participate in their creation. Having lots of people developing these tools makes it more likely that they will serve a broader set of purposes than if individual companies develop them. Perhaps more important, the open part of open source means that the tools are available for anyone to use who can clear what, for professional developers, are modest hurdles.
All of us should be encouraging more of this, and making sure that we don’t inadvertently break anything that is already working to democratize access to AI.
Second, we need to ensure that AI development marketplaces are emerging in a way that serves a diverse set of businesses and entrepreneurs. There is a lot to unpack here, and we’ll cover this in more detail in chapter 8. But in short, we need the AI platform companies to make it easy for ISVs serving geographically diverse businesses to get started with AI infrastructure and development. We need the AI platform companies to help with the education and onboarding of these businesses, and even to focus on solutions and industry- or market-specific tools to help these businesses get started on their AI journey. We need community colleges, universities, and incubators to put programs in place to help individuals learn and practice the skills that these local businesses will need. And we need communities that are enthusiastically supportive of this type of development and transformation.
Third, we need to make it much, much easier to train models. We’ll talk about machine teaching and transfer learning later, but in short, we need new techniques and breakthroughs to reduce the effort required to teach an AI system how to do a new job. Right now, this training task is expensive and requires huge amounts of labeled training data, which has made state-of-the-art AI a thing that’s mostly only accessible to large companies with both lots of expertise and resources. There’s a lot to be hopeful about here, with both AI platform providers and a whole host of data labeling start-ups racing to solve this problem. And there have been multiple technological breakthroughs in these areas as I’ve written this book.
There’s certainly opportunity to do more: for governments to provide incentives for the creation of machine teaching companies, and for them to fund research on machine teaching, transfer learning, and other techniques that make AI training less costly. In February 2019, the president of the United States signed the “Executive Order on Maintaining American Leadership in Artificial Intelligence,”17 which in Section 5(a) directs all agency heads to identify sources of data within their agencies that might be useful for AI, and to work to make that data available for nonfederal use. This, and other government investments and incentives to make more data publicly available for training AI models, is a very good thing.
There is also a very real opportunity for the emergence of open source and commercial marketplaces for pretrained models as transfer learning becomes more effective. Transfer learning, as we briefly touched upon, allows someone to use a model trained for one task to more easily build models to accomplish different tasks. Right now, pretrained models are available for very general types of tasks, like recognizing the objects in an image. For the most part, these pretrained models are too general to be useful in the types of highly specific tasks that businesses are going to want to accomplish, which is why the current best practice for most applications is training models from scratch. With transfer learning, this could change, making it possible for a huge number of pretrained models to be used in a huge number of tasks that the original model trainer could never have anticipated. Such pretrained model marketplaces could dramatically lower the cost that a small, rural business might have to incur to build an AI model to solve a problem.
Fourth, we should pick some very specific problems at the intersection of AI and rural, and bring together stakeholders from AI platform providers, rural businesses and communities, and government to try to solve those problems. The problems could be ones in agriculture, or health care, or transportation, or energy. Basically, whatever matters to the community that will demonstrate how AI can be used to tackle the problem and others like it, and how AI empowering rural businesses can create jobs and local stimulus. Once we have demonstrations that these projects are good business and good policy, then more will happen organically. This would be another great purpose to which a grand AI initiative, on the scale of the Apollo program, could be put.
Fifth, and finally, is that we really do need to invest in rural infrastructure. The intelligent edge and cloud can’t enable any AI transformations if our rural communities don’t have access to high-speed networks. Microsoft has been developing and deploying high-speed wireless infrastructure for rural communities with its Airband initiative.18 Airband uses reclaimed television radio frequencies to try to solve the last-mile networking problem in rural communities, where low population density makes wired infrastructure very costly to deploy. Whether it’s Airband, fiber, 5G, or something else entirely, it is absolutely critical to get these rural communities connected. In addition to being table stakes for the intelligent edge and cloud, and a foundation for rural businesses to benefit from AI, high-speed Internet is also a necessity for the digital workforce of the future to want to live in an area.
The other critical piece of infrastructure that we must invest in is high-quality technical education and training that is geographically and economically accessible to residents of rural communities. It is crazy to assume that we are going to be able to train and re-skill all of the workers needed to make the rural businesses of the future hum if the cost of that training is tens of thousands of dollars or more, or if you have to uproot your life and move to a new community while you go through that training. Some folks, like me and my wife, will do this. Others cannot or will not. It’s not their fault that they want to live in the communities where they were born and raised, near their family and friends and the local institutions that they cherish. That is, in fact, a great thing. And it is on the collective us to find better ways to allow them to live the life they want to live, and to get the training they need, to have economic security for themselves and their families.
These are ideas I can see working back home in central Virginia, and ideas that I hope will guide rural policy in Washington, DC.