WTF?: What's the Future and Why It's Up to Us

by Tim O'Reilly

“No, communication is terrible!” he said. The reason can be explained by the old joke: “One person sits and drinks. Two people clink and drink. The more people you add, the higher the ratio of clinking to drinking.” What you want is a situation where people “clink” only with the people doing shared work, not with everyone they touch. This is simple math. Communication gets worse as team size grows.

Kim explained to me that “writing the press release first is a mechanism to make customer obsession concrete.”

A modern technology engineering organization (or an entire organization like Amazon or Spotify) seeks to have high alignment and high autonomy. Everyone knows what the goal is, but they are empowered to find their own way to do it.

In a presentation I heard General McChrystal give at the New York Times New Work Summit in the summer of 2016, he said, “I tell people, ‘Don’t follow my orders. Follow the orders I would have given you if I were there and knew what you know.’” That is, understand our shared objective, and use your best judgment about how to achieve it.

From the point of view of the company offering an online service, software has gone from being a thing to a process, and ultimately, a series of business workflows.

At that 2003 all-day Amazon all-hands meeting, Jeff Bezos gave the opening talk. It was called “It’s Still Day 1.” He described the history of electricity, with vivid historical photographs of the nests of wires coming down from light sockets in the ceiling to power new kinds of electric devices. The standardized power plug hadn’t yet been invented. He showed how factories still drove the machines on their assembly lines with huge centralized motors, with belts and pulleys carrying the power, just as they had in the age of steam, not yet having realized that they could bring electricity directly to small motors where the work was actually happening. Often, when new technology is first deployed, it amplifies the worst features of the old way of doing business. Only gradually do individuals and organizations realize, through a cascading network of innovations, how to put new technology properly to work.

But Google got it. When an independent developer named Paul Rademacher deciphered the Google Maps data format, he realized that he could build new custom maps by combining data from multiple sources. He built a site called housingmaps.com that showed apartment listings from Craigslist on a Google map—and Google saw the opportunity. Instead of shutting down Paul’s hack, they celebrated it. They hired him, and opened up an API to make mashups easier. This was a transformative breakthrough that led to Google’s dominance in online mapping. As more and more developers built applications for Google Maps, the platform got more powerful and drew more users. It became the classic thick marketplace, where users came for the apps, and apps came for the users.

The App Store is so central to our experience of the smartphone today that it’s easy to forget that the first iPhone didn’t have an app store.

It’s easy to forget just how generative government interventions can be. Larry Page and Sergey Brin’s research project at Stanford, which led to Google, was funded by the National Science Foundation’s Digital Library program. Were the NSF an investor rather than a grant maker for the public good, that investment alone would have repaid more than the entire NSF budget for the years the grant was made. In fact, the market value of Google is greater than the entire amount of taxpayer dollars spent on the NSF since it was first founded in 1952.

As I ran through the park, I couldn’t help but think of the park as a metaphor for all that government does for its citizens. Our roads, our trains, our water and sewers, our universal access to electricity, heat, and telecommunications. Our schools. Our protection from fire and flood, from crime and from foreign enemies. Our rule of law. I know that many of these services cost more than they should, and accomplish less than they could. Some are tragically at odds with core American values—I grieve for police violence against people of color, unnecessary foreign wars, a rule of law that too often seems to favor the rich and powerful over the rights of all. Yet I also think of all the ways that government is the platform upon which our economy and our society are built, with many analogies to the way that iOS and the App Store are the platform for the Apple smartphone economy.

As we’ve discussed, creating a thick marketplace is the first requirement of any platform. This is not a given. A thick marketplace requires both producers (in Apple’s case, app developers) and consumers. In the smartphone space, Apple and Google were able to build thick marketplaces, but Microsoft, for all its past success, was unable to do so. Not enough people bought their phones, which were late to market, and so app developers were unwilling to build new apps for Windows Mobile, which confirmed customers in their decision to avoid the phone. What is the equivalent for government? For “thick marketplace,” read “flourishing economy.”

But there’s an important lesson from the wealth of nations: If the population doesn’t have enough money to buy goods and services on offer either from its own sellers or from those trading from other countries, the country remains poor.

Getting a robust marketplace off the ground and keeping it often requires strong government intervention. In his book Bad Samaritans, Korean economist Ha-Joon Chang describes how South Korea used central planning and targeted investment in specific industries to build a highly successful economy.

Government and tech platforms must each provide core services that the “apps” or other services rely on. Despite the prevailing belief that the United States economy is largely a “free market,” none of it works without fundamental infrastructure. In the 1930s, the Tennessee Valley Authority and the Rural Electrification Administration built dams and power distribution systems and established the idea that access to electricity was a fundamental right of every citizen. Telecommunications followed the same pattern, with a commitment to universal service enforced by the Federal Communications Commission. And of course, our national highway system, created in the 1950s, enabled interstate commerce and accelerated the growth of our economy. These are among the fundamental platform services of our country,

Every Silicon Valley firm builds two intertwined systems: the application that serves users, and a hidden set of applications that they use to understand what is happening so that they can continuously improve their service.

It had been set up in 2011 under the leadership of Mike Bracken, the former head of digital for the Guardian. Mike had soon attracted top talent from Britain’s technology and digital media circles, and the GDS had been described by one prominent VC as “the best startup in Europe we can’t invest in.”

To succeed, platforms can’t just offer apps, or services. They have to effectively set and adjust the rules that govern the behavior of the platform participants.

“Some of you, not all of you, are working right now on another app for people to share pictures of food or a social network for dogs. I am here to tell you that your country has a better use for your talents.” He listed a set of urgent problems the government needs help with, and concluded, “All of these are design and information-processing problems, and all of these are matters of life or death to millions of citizens and all of them are things you can fix if you choose to.”

If government’s slow, change-resistant technology procurement processes mean that it is five or six years behind the private sector, the three or four exponential generations of Moore’s Law that have passed will make its capabilities ten times worse. And in classic “news from the future” style, that’s exactly what we see. Amazon can deliver packages within hours of your order; Google can tell you in near-real time that there’s an accident up ahead and to take a different route. Yet the VA takes eighteen months just to determine whether discharged soldiers are eligible for benefits.

While this paper was focused on language translation, it summed up the approach that has been essential to the success of Google’s core search service. Its insight, that “simple models and a lot of data trump more elaborate models based on less data,” has been fundamental to progress in field after field, and is at the heart of many Silicon Valley companies. It is even more central to the latest breakthroughs in artificial intelligence.

As Father John Culkin so aptly summarized the ideas of Marshall McLuhan, “We shape our tools, and thereafter our tools shape us.”

Using machine learning, the developer starts out with a hypothesis, just like before, but instead of producing a handcrafted algorithm to process the data, she collects a set of training data reflecting that hypothesis, then feeds the data into a program that outputs a model—a mathematical representation of features to be looked for in the data. This cycle is repeated again and again, with the program making minute adjustments to the model, gradually modifying the hypothesis using a technique such as gradient descent until it more perfectly matches the data. In short, the refined model is learned from the data. That model can then be turned loose on real-world data similar to that in the training data set.

Deep learning uses layers of recognizers. Before you can recognize a dog, you have to be able to recognize shapes. Before you can recognize shapes, you have to be able to recognize edges, so that you can distinguish a shape from its background. These successive stages of recognition each produce a compressed mathematical representation that is passed up to the next layer. Getting the compression right is key. If you try to compress too much, you can’t represent the richness of what is going on, and you get errors. If you try to compress too little, the network will memorize the training examples perfectly, but will not generalize well to novel inputs.

Machine learning has become a bigger part of Google Search. In 2016, Google announced RankBrain, a machine learning model that helps to identify pages that are about the subject of a user’s query but that might not actually contain the words in the query. This can be especially helpful for queries that have never been seen before. According to Google, RankBrain’s opinion has become the third most important among the more than two hundred factors that it uses to rank pages.

Possibly the most important questions in AI are not the design of new algorithms, but how to make sure that the data sets with which we train them are not inherently biased. Cathy O’Neil’s book Weapons of Math Destruction is essential reading on this topic.

Jeremy Howard went on to cofound Enlitic, a company that is using machine learning to review diagnostic radiology images, as well as scanning many other kinds of clinical data to determine the likelihood and urgency of a problem that should be looked at more closely by a human doctor. Given that more than 300 million radiology images are taken each year in the United States alone, you can guess at the power of machine learning to bring down the cost and improve the quality of healthcare.

What do all these forms of regulation have in common? 1. A clear understanding of the desired outcome. 2. Real-time measurement to determine if that outcome is being achieved. 3. Algorithms (i.e., a set of rules) that make continuous adjustments to achieve the outcome. 4. Periodic, deeper analysis of whether the algorithms themselves are correct and performing as expected.

There are those who say that government should just stay out of regulating many areas, and let “the market” sort things out. But bad actors take advantage of a vacuum in the absence of proactive management. Just as companies like Google, Facebook, Apple, Amazon, and Microsoft build regulatory mechanisms to manage their platforms, government exists as a platform to ensure the success of our society, and that platform needs to be well regulated.

The best regulations encourage the regulated party to take on the problem themselves. This is not “self-regulation” in the sense that government simply trusts the market to do the right thing. Instead, it is a matter of creating the right incentives. For example, the Fair Credit Billing Act of 1974 made consumers responsible for only $50 of any fraudulent credit card charges, making it in the industry’s own self-interest to police fraud aggressively.

When data is provided in reusable digital formats, the private sector can aid in ferreting out problems as well as building new services that provide consumer and citizen value.

It is said that “that government is best which governs least.” Unfortunately, evidence shows that this isn’t true. Without the rule of law, capricious power sets the rules, usually to the benefit of a powerful few. What people really mean by “governs least” is that the rules are aligned with their interests. In an economy tuned to the interests of the few, the rules are often unfair to the rest. An economy tuned to the interests of the majority may seem unfair to some, but John Rawls’s “veil of ignorance”—the idea that the best rules for a political or economic order are those that would be chosen by people who had no prior knowledge of their place in that order—is a convincing argument that that government is best that governs for most.

In the language of Eric Ries’s popular Lean Startup methodology, the first version is referred to as “minimum viable product (MVP),” defined as “that version of a new product which allows a team to collect the maximum amount of validated learning about customers with the least effort.”

Think your hotel housekeeper works for Hyatt or Westin? Chances are good they work for Hospitality Staffing Solutions. Think those Amazon warehouse workers who pack your holiday gifts work for Amazon? Think again. It’s likely Integrity Staffing Solutions.

But in today’s world, this has given way to a kind of continuous partial employment for most low-wage workers at large companies, where workplace scheduling software from vendors like ADP, Oracle, Kronos, Reflexis, and SAP lets retailers and fast-food companies build larger-than-needed on-demand labor pools to meet peak demand, and then parcel out the work in short shifts and in such a way that no one gets full-time hours.

More powerfully, it is the notion that workers are just a cost to be eliminated rather than an asset to be developed. Ultimately, the segregation of workers into privileged and unprivileged classes, and the moral and financial calculus that drives that segregation, has to stop. Over time, we will realize that this is an existential imperative for our economy, not just a moral imperative.

“Zeynep Ton has proven what great leaders know instinctively—an engaged, well-paid workforce that is treated with dignity and respect creates outsized returns for investors. She demonstrates that the race to the bottom in retail employment doesn’t have to be the only game being played.” Economists have long recognized this phenomenon. They call wages higher than the lowest that the market would otherwise offer “efficiency wages.” That is, they represent the wage premium that an employer pays for reduced turnover, higher employee quality, lower training costs, and many other significant benefits.

The algorithm is the new shift boss. What regulators and politicians should be paying attention to is the fitness function driving the algorithm, and whether the resulting business rules increase or decrease the opportunities offered to workers, or whether they are simply designed to increase corporate profits.

But Cadwalladr ignored the scale at which Google operates, and the way that scale fundamentally changes the necessary solution. Google, Facebook, Twitter, and their like need to be understood as a new thing, which doesn’t fit neatly into the old map. That new thing operates by different rules—not by whim or an unwillingness to incur the costs of curation, but by necessity.

The result of any Google search is the result of prodigious efforts to retrieve and rank every page on the web—30 trillion of them, from 250 billion unique web domain names, according to former Google VP of search Amit Singhal—and to serve them up in response to more than 5 billion searches a day.

Facebook is similarly huge. In 2013, the social network disclosed that nearly 5 billion pieces of content were posted every day.

The idea that Google or Facebook can solve the problem simply by hiring teams of human editors or fact checkers, or use outside media organizations to combat fake news, hate speech, or other objectionable results, removing or demoting them one at a time, indicates that people have little idea of the scale or nature of the problem.

“A lie will have gone halfway around the world before the truth has had time to tie on its shoes.”

Air Force Colonel John Boyd, “the father of the F-16,” introduced the term OODA loop (“Observe-Orient-Decide-Act”) to describe why agility is more important in combat than pure firepower. Both fighters are trying to understand the situation, decide what to do, and then act. If you can think more quickly, you can “get inside the OODA loop of your enemy” and disrupt his decision making. “The key is to obscure your intentions and make them unpredictable to your opponent while you simultaneously clarify his intentions,” wrote Boyd’s colleague Harry Hillaker in his eulogy to Boyd. “That is, operate at a faster tempo to generate rapidly changing conditions that inhibit your opponent from adapting or reacting to those changes and that suppress or destroy his awareness. Thus, a hodgepodge of confusion and disorder occur to cause him to over- or under-react to conditions or activities that appear to be uncertain, ambiguous, or incomprehensible.”

The distance between human enthusiasm and the fundamentals can also be measured for news, using many signals that can be verified algorithmically by a computer, often more quickly and thoroughly than they can be verified by humans.

“Cracks were the centerpiece of the investigation. They could not be eliminated. They were everywhere, permeating the structure, too small to be seen. The structure could not be made perfect, it was inherently flawed, and the goal of engineering design was not to certify the airframe free of cracks but to make it tolerate them.” So too, the essence of algorithm design is not to eliminate all error, but to make results robust in the face of error. The fundamental question to ask is not whether Facebook should be curating the newsfeed, but how.

During earlier research on economic differences between the twenty regional governments of Italy, Putnam had noticed that there was a close correlation between civic engagement and prosperity. “These communities did not become civic simply because they were rich. The historical record strongly suggests precisely the opposite: They have become rich because they were civic.” Social capital is as important as financial capital in the wealth of nations.

Ray Dalio, the founder and executive chairman of Bridgewater Associates, uses a similar approach to creating what he calls an “idea meritocracy” at his company, the largest hedge fund in the world. As members of the firm debate investments or ideas, they rate the assertions of the other participants, assembling them into a matrix that highlights agreement and disagreement. Everyone is urged to be “radically transparent” with their opinions, and the newest associate is welcome to tell Ray himself that he is wrong. Bridgewater takes the further step of applying an algorithm to the matrix, which takes into account factors such as past performance, expertise on the particular topic, and other ways of weighting individual opinions. The goal is to combine the best of human insight and the ability of computer algorithms to sum up and clarify the points of agreement and disagreement.

As Henry Farrell wrote to me in another email: “Processes of intellectual discovery are all about arguments between different (and sometimes stylized) positions. To use a machine learning analogy stolen from my collaborator, Cosma Shalizi—all of us put together are at best an ensemble of weak learners, each of which only grasps a few of the terms in a very long and complicated vector that we’re trying to model. It plausibly helps if we start from very different positions (each weak learner sees a different set of terms) as long as each of these positions reflect some aspect of the truth and then, and only then, try to converge on a shared model of the problem.”

The race to the bottom has in part been a result of the primary shift of news industry revenue from subscription to advertising and from a secure base of local readers to chasing readers via social media. Subscription-based publications have an incentive to serve their readers; advertising-based publications have an incentive to serve their advertisers.

He had come to realize that however successful, Medium hadn’t gone far enough in breaking with the past. He concluded that the broken system is ad-driven Internet media itself. “It simply doesn’t serve people. In fact, it’s not designed to,” he wrote. “The vast majority of articles, videos, and other ‘content’ we all consume on a daily basis is paid for—directly or indirectly—by corporations who are funding it in order to advance their goals. And it is measured, amplified, and rewarded based on its ability to do that. Period. As a result, we get . . . well, what we get. And it’s getting worse.”