Fifty years ago, productivity growth in advanced economies began to slow down. Productivity growth — the component of GDP growth that is not due to increases in labor and capital — is the primary driver of rising incomes. When it slows, so does economic growth as a whole. This makes it an urgent trend to understand. Unfortunately, the most popular explanation for why it’s happening might be wrong. The most widely endorsed reason productivity growth has faltered is that we are running out of good ideas. As this narrative has it, the many scientific and technology advances responsible for driving economic growth in the past were low-hanging fruit. Now the tree is more barren. Novel advances, we should expect, are harder to come by, and historical growth may thus be difficult to sustain. In the extreme, this may lead to the end of progress altogether. This story began in 2020, with the publication of “Are Ideas Getting Harder to Find?,” by economists Nicholas Bloom and colleagues. Bloom et al. looked across many sectors, from agriculture to medicine to computing. In each field, productivity measures have grown at the same rate as before. This sounds like good news, except that the number of researchers in each of these fields has exploded. In other words, each researcher produces much less than they used to — something you might expect if ideas really are getting harder to find. The progress studies movement and the metascience community have risen, in part, in response to this challenge. Both seek ways to rethink how we do research: by making our research institutions more efficient or by increasing science funding. But there's a growing body of evidence that suggests ideas are not, in fact, getting harder to find. Instead, the problem appears to be that markets have become less effective at translating breakthrough technologies into productivity gains. Researchers appear to be continuing to generate valuable innovations at historical rates. It’s just that these innovations face greater barriers to commercialization, and innovative firms thus fail to gain market share. All this suggests that the constraint on growth isn’t in our universities or labs or R&D departments, but in our markets.
Anagh Banerjee
Why ideas matter
Historically, the task of growth theory has been to rationalize this graph:
Through two world wars, the Great Depression, the global financial crisis, and the Cold War, US real GDP per capita has grown steadily at 2% per year. This consistency is remarkable, so much so that it has motivated economists to search for a near-immutable source of economic growth: something fundamental that drives growth across long sweeps of time. Sure, growth can be affected in the short run by policies and events of the day — tariffs, wars, demographic transitions, educational booms — but it would be an incredible coincidence if the combined impact of every economic policy and external event in history happened to net out as a constant growth rate! So what could this immutable mechanism be? It took economists decades to understand how sustained exponential growth is even possible. Exponential growth requires increasing returns to scale — doubling all the inputs into production must more than double the scale of economic output. Why? Each year output has to be reinvested as capital to produce next year’s output. But there’s diminishing returns to capital alone, so each year a country would need to reinvest a larger fraction of its annual output to get the same growth rate. Eventually, it would need to reinvest more than 100% of its output, which is impossible. It is only through increasing returns to scale that we can counteract the diminishing returns to capital, allowing us to maintain exponential growth. So it’s easy! We just posit increasing returns to scale in production, and we now have an explanation for why growth is exponential — decades of research not necessary. The problem is that increasing returns to scale violates our basic intuitions. Imagine you own a factory that produces 100 cars a day. You create an exact duplicate of this factory in another city, with the same employees, same equipment. How many cars would you expect from both factories combined? The intuitive assumption is 200. Increasing returns demands that we can somehow get more than double the cars from creating the second factory — which is hard to justify. The economist Paul Romer won a Nobel Prize for resolving this puzzle: Increasing returns to scale comes from ideas. In his view, it is a mistake to think that the only inputs to the car factory are the workers and machines. There are also blueprints for the cars being produced, instructions for how the machines should be laid out, and the concept of an assembly line process to organize the workers. Romer’s work elegantly solved the problem posed by our factory duplication thought experiment. In setting up the second factory, we needed only to duplicate the workers and the machines — we didn’t need to duplicate the designs or the idea of an assembly line. Once created, ideas can be used in perpetuity, which is how we can double the factory’s output while doubling only its physical inputs. This key property of ideas — that they can be used by everyone at the same time — has become the fundamental explanation for exponential growth. This is why it matters so much if ideas truly are getting harder to find: If idea production is slowing down, it threatens the foundation that allows growth to continue.
The simple argument for declining idea productivity
To test whether our research efforts are getting less bang for their buck today than they did decades ago, Bloom et al. use a key feature that descends from Romer’s original model — that productivity growth can be represented by the following equation: Productivity growth = Research productivity * Number of researchers Whether total factor productivity, or agricultural yields, or chip density, growth of any productivity measure should be a direct function of the number of researchers working in that sector and their research productivity. This means that if we observe productivity growth in any given sector, and we know the number of researchers working in that sector, we can infer research productivity. Thus, the authors compile data on productivity growth and researcher counts across a number of different sectors to estimate whether research productivity has been falling, which according to the authors, means ideas must be getting harder to find. Perhaps their most compelling evidence comes from Moore's Law, the famous observation that the number of transistors on a computer chip doubles roughly every two years. This doubling represents a constant 35% annual growth rate in chip density that has held for 50 years. On its face, Moore’s Law seems like a refutation of any diminishment in technological progress. Yet maintaining the exponential growth in chip density has required exponential increases in effort. Bloom et al. compiled R&D spending data from dozens of semiconductor firms over time and found that the effective number of researchers working to advance Moore's Law increased by a factor of 18 between 1971 and 2014. Meanwhile, the growth rate of chip density has stayed constant. Put differently, it takes 18 times as many researchers today to achieve the same rate of improvement in chip density as it did in the early 1970s. This implies that research productivity in semiconductors has fallen at an average rate of 7% per year. Look at agricultural productivity and you see a similar pattern. The authors measure crop yield growth across major US crops. Between 1969 and 2009, yield growth for these crops averaged a steady 1.5% per year, but the research effort directed toward improving yields has grown by between sixfold and 24-fold, depending on the crop. Zoom all the way out, and the pattern still holds. Across the economy as a whole, R&D efforts have increased by a factor of 20 since the 1930s, yet productivity growth has become slower.
These results are unambiguous. Research effort has gone up, yet productivity growth is not budging. This seems like clear evidence that something about productivity growth is getting harder. But whether the problem is a lack of new ideas is much less obvious.
Measuring idea productivity directly
The idea-based growth model is successful as a simple description of how exponential growth could occur. The problem is we’ve taken it too literally. Bloom et al. assume that idea production is the only factor behind productivity growth. For example, their agricultural case study uses crop yield growth as the sole variable for new ideas. This allows them to sidestep difficulties in defining ideas and measuring their impact, but it also rules out the possibility that factors other than ideas are the real reason yields are stagnating. Imagine that agricultural R&D spending was highly effective, and that in the past few decades it led to a stream of new seed varieties that were each higher-yield than the last. What if those seeds were not actually being purchased by farmers — maybe because farmers were unaware that they existed or because adopting a new seed is risky? We would observe crop yields stagnating despite R&D spending effectively creating more productive crops. Bloom et al.’s measure of "ideas" combines actual research innovations with other necessary conditions for research innovations to translate into higher output. After being invented, technologies have to be successfully commercialized, marketed, and adopted at scale before they can have large effects on economic output. What if we’re still just as good at producing ideas, but we’ve become much worse at capitalizing on them? This is exactly the argument made by Teresa Fort and colleagues in a paper from April of this year: “Growth Is Getting Harder to Find, Not Ideas.” Fort et al. use the census of firms linked to US patent filings to capture economy-wide invention, rather than focusing on sector-specific case studies. Most importantly, they measure idea production more directly, by estimating the relationship between R&D spending and new patents rather than inferring idea production from firm growth. Since patents represent technologies that are novel enough to be given intellectual property protection, and also economically valuable enough to be worth patenting, they serve as a more direct measure of “ideas.” Fort et al. find that, across firms, research expenditure today continues to be associated with a proportional increase in patents similar to the 1980s. They use a variety of measures to get at this, but the most transparent one is to measure the ratio between patents and R&D expenditures for each firm. Doing this, they find that the average firm’s patent-to-R&D ratio has actually increased by 50% since 1977 — contrary to a story in which R&D effort is becoming less effective. While there is enough variability in this ratio that Fort et al. can’t be confident that it has actually increased, we can certainly say that it hasn’t fallen in the way that Bloom et al. would predict. The obvious question is whether these patents might represent less generative and useful ideas, something like more incremental advances than patents of the past. Maybe the low-hanging fruit really is gone, and new patents are capturing less useful ideas. Fort et al. address this issue by focusing on breakthrough patents, a measure of technological innovation defined by Kelly et al., and showing that their results still hold. For a technology to count as a breakthrough, it must be generative — technologies that come after must build on it. This idea is the basis for Kelly et al.’s measurement of a patent’s significance. They score a patent as breakthrough if its text is different from patents that came before it but similar to the text of patents that came after it. Patents that scored in the top 5% on this measure included the elevator, the typewriter, the telephone, and frozen foods — giving us some assurance that this measure really selects high-quality technologies. Fort et al. show that their results are not simply coming from more incremental patents over time. Not only has the number of patents filed per R&D dollar increased, but the number of breakthrough patents per R&D dollar has also increased. Firms produce three times more breakthrough patents per R&D dollar than they did in 1977. This analysis suggests that Bloom et al. jumped the gun by attributing the slowdown in productivity growth to declining research productivity. If you infer research productivity only from output growth, it’s hard to find. But if you look at new idea production through the lens of patent data, we appear to be as generative as ever. So there must be some other failure in translating new technologies into productivity growth. What could that be?
... continue reading