I consider Definite Optimism as Human Capital to be my most creative piece. Unfortunately, it’s oblique and meandering. So I thought to write a followup to lay out its premises more directly and to offer a restatement of its ideas.
The goal of both pieces is to broaden the terms in which we discuss “technology.” Technology should be understood in three distinct forms: as processes embedded into tools (like pots, pans, and stoves); explicit instructions (like recipes); and as process knowledge, or what we can also refer to as tacit knowledge, know-how, and technical experience. Process knowledge is the kind of knowledge that’s hard to write down as an instruction. You can give someone a well-equipped kitchen and an extraordinarily detailed recipe, but unless he already has some cooking experience, we shouldn’t expect him to prepare a great dish.
I submit that we have two big biases when we talk about technology. First, we think about it too much in terms of tools and recipes, when really we should think about it more in terms of process knowledge and technical experience. Second, most of us focus too much on the digital world and not enough on the industrial world. Our obsession with the digital world has pushed our expectation of the technological future in the direction of cyberpunk dystopia; I hope instead that we can look forward to a joyful vision of the technological future, driven by advances in industry.
Process knowledge is represented by an experienced workforce. I’ve been studying the semiconductor industry, and that has helped to clarify my thoughts on technological innovation more broadly. It’s easy to identify all three forms of technology in the production of semiconductors: tools, instructions, and process knowledge. The three firms most responsible for executing Moore’s Law—TSMC, Intel, and Samsung—make use of each. All three companies invest north of $10 billion a year to push forward that technological frontier.
The tools and IP held by these firms are easy to observe. I think that the process knowledge they possess is even more important. The process knowledge can also be referred to as technical and industrial expertise; in the case of semiconductors, that includes knowledge of how to store wafers, how to enter a clean room, how much electric current should be used at different stages of the fab process, and countless other things. This kind of knowledge is won by experience. Anyone with detailed instructions but no experience actually fabricating chips is likely to make a mess.
I believe that technology ultimately progresses because of people and the deepening of the process knowledge they possess. I see the creation of new tools and IP as certifications that we’ve accumulated process knowledge. Instead of seeing tools and IP as the ultimate ends of technological progress, I’d like to view them as milestones in the training of better scientists, engineers, and technicians.
The accumulated process knowledge plus capital allows the semiconductor companies to continue to produce ever-more sophisticated chips. The doubling of transistor density every 24 months wouldn’t be possible if these firms didn’t already possess deep pools of process knowledge. It’s not just about the tools, which any sufficiently-capitalized firm can buy; or the blueprints, which are hard to follow without experience of what went into codifying them. The US has many decades of experience in designing and fabricating semiconductors, and it has developed the talent ecosystem that succeeds in pushing Moore’s Law forward. This cluster of talent allows the US to maintain its lead on a critically-important technology.
The US industrial base has been in decline. But sustained innovation in semiconductors is an exception in US manufacturing. The country used to nurture vibrant communities of engineering practice (a term I like from Brad DeLong), which is another way to talk about the accumulated process knowledge in many segments of industry. But not all communities of engineering practice have been in good shape.
The real output of the US manufacturing sector is at a lower level than before the 2008 recession; that means that there has not been real growth in US manufacturing for an entire decade. (In fact, this measure may be too rosy—the ITIF has put forward an argument that manufacturing output measures are skewed by excessive quality adjustments in computer speeds. Take away computers, which fewer and fewer people are buying these days, and US real output in manufacturing would be meaningfully lower.) Manufacturing employment peaked in 1979 at nearly 20 million workers; it fell to 17 million in 2000, 14 million in 2008, and stands at 12 million today. The US population has grown by 40% since 1979, while the number of manufacturing workers has nearly halved.
When firms and factories go away, the accumulated process knowledge disappears too. Industrial experience, scaling expertise, and all the things that come with learning-by-doing would decay. I visited Germany earlier this year to talk to people in industry. One point Germans kept bringing up was that the US has de-industrialized itself and scattered its production networks. While Germany responded to globalization by moving up the value chain, the US manufacturing base mostly responded by abandoning production.
Brad Setser has shown that the US stands out amongst rich countries for its low level of manufactured goods exports. Call me simple-minded, but I believe that the world’s most developed country ought to be responsible for exporting goods around the world. Instead, the US runs both a trade deficit and a current account deficit. The US trade deficit is high not only because it imports a lot of goods; it’s high also because it doesn’t export very much. In order for other countries to import more from the US, first it should have better goods to sell.
Knowledge should circulate throughout the supply chain, flowing both up and down the stack. Successful industries tend to cluster into tight-knit production networks. The easiest way to appreciate the marvel of clusters is to look at Silicon Valley, where capital, academia, a large pool of eager labor, and companies both large and small sit next to each other. To any of us who have spent some time in Silicon Valley, it’s obvious that this concentration of economic linkages is part of the magic that makes the system work.
There are many other examples of industrial clusters. Taiwan’s semiconductor industry was founded and is still centered around a small industrial park south of Taipei. Silicon Valley is so-named because it was the center of semiconductor production (and it has enough toxic Superfund sites nearby to prove that heritage). It’s not just chips: autos, electronics, biotech, aviation, and machine parts all tend to be geographically clustered.
Proximity makes it easier to generate process knowledge. But what happens when we tear apart these production networks by separating design and manufacturing? Sometimes it’s no big deal, sometimes it works out great. But I believe that in most cases, dislocation makes it more difficult to maintain process knowledge.
Both the design process and production process generate useful information, and dislocation makes it difficult for that information to circulate. I think we tend to discount how much knowledge we can gain in the course of production, as well as how it should feed back into the design process. Maybe it’s easier to appreciate that with an example from computing. Arjun Narayan tells me that good software design requires a deep understanding of chips, and vice versa. The best developers are those who understand how processes interact both up and down the stack.
What happens when we stop the flow of knowledge up the stack? I think that the weakness of the US industrial robotics sector is instructive. The US has little position in making high-end precision manufacturing equipment. When it comes to factory automation systems, machine tools, robot arms, and other types of production machinery, the most advanced suppliers are in Japan, Germany, and Switzerland. I think the reason that the US has little position can be tied directly to the departure of firms from so many segments of manufacturing. How do engineers work on the design of automation systems if they don’t have exposure to industrial processes?
A quote from the article: “A report to President Barack Obama on advanced manufacturing, prepared by his council of science advisers in 2012, concluded that the ‘hard truth’ was that the U.S. lagged other rich nations on manufacturing innovation.”
I don’t see enough Americans being troubled by the idea that America isn’t making advanced industrial robots. It might be fine to think that robots will be doing all the manufacturing work in the future; but someone has to build these robots, and own the IP of advanced robotmaking, and for the most part, that someone is not the US. It can’t be an accident that the countries with the healthiest communities of engineering practice are also in the lead in designing tools for the sector. They’re able to embed knowledge into new tools, because they continue to generate process knowledge.
Let’s try to preserve process knowledge. The decline of industrial work makes it harder to accumulate process knowledge. If a state has lost most of its jobs for electrical engineers, civil engineers, or nuclear engineers, then fewer young people will enter into these fields. Technological development slows down, and it turns into a self-reinforcing cycle of decline. I think we should try to hold on to process knowledge.
Japan’s Ise Grand Shrine is an extraordinary example in that genre. Every 20 years, caretakers completely tear down the shrine and build it anew. The wooden shrine has been rebuilt again and again for 1,200 years. Locals want to make sure that they don’t ever forget the production knowledge that goes into constructing the shrine. There’s a very clear sense that the older generation wants to teach the building techniques to the younger generation: “I will leave these duties to you next time.”
Regularly tearing down and rebuilding a wooden temple might not sound like a great use of time. But I’m not sure if local priorities are entirely screwed up here. These people understand that it’s too difficult to write down every instruction necessary for building even a single wooden structure; imagine how much more difficult it is to create instructions for a machinery part, or a chip. Every so often we discover ancient tools of which we have no idea how to use. These shrine caretakers have decided that preservation of production knowledge is important, and I find that admirable.
Building a vast industrial base and practicing learning-by-doing used to be the American way. Brad DeLong again: “When the technologies of the second industrial revolution arrived, the United States with its cotton and wide market, and its rich natural resources, and its communities of engineering excellence, was able to leap ahead—and in fact greatly surpass Britain in manufacturing productivity pretty much everywhere. So that the 20th century became an American century, rather than a second British century, in large part because of the bets Hamilton had induced the United States to make on not simply following comparative advantage.”
The future should be more than services. Isn’t manufacturing always the low value-added stuff, and that the future should be driven by services instead? I’m not so sure. I’m skeptical that we can pin all our hopes on the services sector because it tends to have two big problems: a lot of it is winner-take-all, and much of the rest is zero-sum.
One of the issues with services jobs in the US is that most of the gains are captured by very few workers. Two services sectors are enormously productive: tech and finance. But other sectors, like retail, hospitality, and food services don’t generate productivity growth quite as quickly. Therefore, while aggregate production can rise, consumption doesn’t tend to keep up. That’s because overall production gains in services are asymmetrically generated by hedge fund managers and machine learning engineers, not the hotel workers and retail staff, and these highly productive workers can only consume so much.
Another issue with a lot of service work is that much of it is zero-sum, a point made very well by Adair Turner. Too many service jobs are meant to cancel out the efforts of other service jobs, for example in litigation, where a plaintiff’s lawyer creates a job for the defendant’s lawyer. And often the zero-sumness is asymmetric: a dozen hackers make a theft, and companies everywhere subsequently need to spend collective billions on staff or contractors to protect themselves; a few criminal plague a state, and the government subsequently needs to hire hundreds of officers to make people feel safe; a few people commit accounting fraud, and the ensuing uproar forces companies and banks to ramp up the size of their compliance departments by tens of thousands in the aggregate.
There’s an entertaining line in the Brad Setser piece I linked to earlier. He tells us that one of the reasons that the US has such a high surplus in the services trade is that Americans have a low propensity to travel abroad. I don’t view that as a great way to earn a trade surplus.
My favorite genre of the Bloomberg column has become Noah Smith dunking on the United Kingdom. Services make up about 80% of the British economy, and that has brought along a host of problems. These include low levels of productivity growth over the last two decades, extraordinary vulnerability to the financial crisis, and low levels of R&D spending by its biggest companies. Matt Klein has put forward a fun claim: “Take out Grea
