“…a must-read for serious students of economic policy.” - Glenn Hubbard, Dean, Columbia Business School, and former Chairman of the Council of Economic Advisers
“Unintended Consequences is full of substance, it is one of the must-read books of the year, and once I finish it I will be giving it a second read through right away.” - Tyler Cowen, Professor, George Mason University
“Unintended Consequences is far smarter and more thought-provoking than most economics written for the general public” - Greg Mankiw, Harvard University, Former Chairman of the Council of Economic Advisors
“Unintended Consequences provides a provocative interpretation of the causes of the global financial crisis and the policies needed to return to rapid growth. Whether you agree or not, this analysis is well worth reading.” - Nouriel Roubini, New York University; Chairman, Roubini Global Economics
“There are an amazing number of good ideas and interesting points made in Unintended Consequences. The thinking underlying it, and the obvious depth of understanding of the author, are very impressive.” - Steven Levitt, coauthor of Freakonomics; 2004 John Bates Clark Medal
“…reminds us that inequality sends a signal of what society lacks most, in America’s case, entrepreneurship and risk taking.” - Lawrence Lindsey, CEO, The Lindsey Group, former Director of the National Economic Council
“…a comprehensive explanation of the modern economy.” - Julian Robertson, Founder, Tiger Management
“A full-throated defense of economic dynamism.” - The Wall Street Journal
“…a must-read for serious students of economic policy.” - Glenn Hubbard, Dean, Columbia Business School, and former Chairman of the Council of Economic Advisers
“Unintended Consequences represents the most cogent and persuasive analysis of the Financial Crisis to date.” - Andrei Shleifer, 1999 John Bates Clark Medal Winner
“Unintended Consequences should be read by anyone who takes for granted the superiority of progressive taxation and has not thought carefully about the trade-offs involved.” - The New Republic
“…reminds us that inequality sends a signal of what society lacks most, in America’s case, entrepreneurship and risk taking.” - Lawrence Lindsey, CEO, The Lindsey Group, former Director of the National Economic Council
AI Summary.The U.S. college wage premium fell from 0.626 to 0.575 after four decades of expansion, implying the first sustained negative growth in relative demand for college-educated labor. The authors estimate that weaker wage growth in occupations highly exposed to artificial intelligence accounts for roughly 28% of that decline, though the patterns are not causal.
José Azar, Mireia Gine and Javier Sanz-EspínSocial Science Research Network
Date Posted:
Is Database:
Database
The college wage premium flattened in the mid-2010s and has fallen ~8% since 2022. The authors argue that this compression reflects a broad decline in the returns to formal schooling, rather than a decline in the upper tail.
Is AI exposure driving the decline in the college wage premium?
Core argument: The U.S. college wage premium fell from 0.626 in 2022 to 0.575 in 2026 after expanding for four decades, with Current Population Survey data implying the first sustained negative relative demand growth for college labor.
After expanding for four decades, the U.S. college wage premium [dropped] sharply from 0.626 in 2022 to 0.575 in 2026. Current Population Survey data through 2026 implies an unprecedented drop in relative demand for college labor—the first sustained negative relative demand growth. Post-2022 wage growth slowed disproportionately in high-exposure occupations, which employ a disproportionate share of college graduates. By 2026, going from zero occupational AI exposure to full exposure had a negative effect on wages of−0.086. Combined with the college–non-college exposure gap, this mechanism accounts for roughly 28% of the total drop in the college wage premium from 2022 to 2026. While non-causal, these patterns indicate that task displacement in AI-exposed white-collar occupations plays a quantitatively meaningful role in the recent compression of the aggregate skill premium.
Takeaways by Macro Roundup® AI
The U.S. college wage premium fell from 0.626 in 2022 to 0.575 in 2026 after expanding for four decades, with Current Population Survey data implying the first sustained negative relative demand growth for college labor.
By 2026, moving from zero occupational AI exposure to full exposure had a -0.086 effect on wages, and combined with the college–non-college exposure gap this mechanism accounts for roughly 28% of the 2022–2026 drop in the college wage premium.
Post-2022 wage growth slowed disproportionately in high-exposure occupations that employ a disproportionate share of college graduates, patterns that, while non-causal, indicate task displacement in AI-exposed white-collar occupations plays a quantitatively meaningful role in compressing the skill premium.
Related Articles:
Looking for the Ladder— The article argues that the hiring downturn in occupations most exposed to generative AI reflects interest-rate sensitivity rather than AI itself, since roughly 38% of workers in the most AI-exposed quintile work in information, finance and insurance, and professional and technical services, versus under 2% in the least-exposed quintile.
How Students and Recent Grads are Responding to the Rise of AI— Undergraduate enrollment in degrees tied to the most AI-exposed jobs has risen 8%, indicating students are moving toward rather than away from fields exposed to artificial intelligence. A decline in computer science degrees is more than offset by growth in other exposed fields such as engineering.
AI and Young-adult Jobs: The Real Mystery— Employment rates have declined for young adults both with and without college degrees, with non-degree holders faring marginally worse, and employment growth for young workers of all education levels has lagged that of other workers. Artificial intelligence can replace some lower-skilled jobs, but those jobs are too few to drive youth labor market weakness.
AI Summary.An analysis of Epoch's Capabilities Index data, controlling for developer and model family, finds no statistically significant breakpoint in AI progress; a segmented model fits no better than a straight-line trend (p = 0.615), with the estimated slope change ranging from -8.4 to +23.4 points per year.
Paul KedroskyApplied Complexity
Date Posted:
Is Database:
Database
Kedrosky argues AI capabilities continue to improve, but “the full composite data shows flattening relative gains, not acceleration…rolling relative model gains have fallen from their 2024 peak, while model dispersion has narrowed sharply.”
Has AI capability progress accelerated beyond its linear trend?
Core argument: Using all observations from Epoch's Capabilities Index and controlling for developer and model family, a piecewise model shows no statistically significant breakpoint in AI capability trends (p = 0.615) and does not improve on a purely linear trend.
Using all Epoch’s Capabilities Index observations, and controlling for developer and model family, there is no statistically significant breakpoint. A piecewise model—which splits the series into intervals and applies a sub-function to each segment—does not improve on a purely linear trend: p = 0.615, The estimated change in slope has a confidence interval of -8.4 to +23.4 points per year. In short, the maths shows there is no model acceleration, contrary to claims, and as expected. The result comes from selecting frontier observations only, choosing a breakpoint, ignoring variance collapse, and fitting separate lines on either side.
Takeaways by Macro Roundup® AI
Using all observations from Epoch's Capabilities Index and controlling for developer and model family, a piecewise model shows no statistically significant breakpoint in AI capability trends (p = 0.615) and does not improve on a purely linear trend.
The estimated change in slope of the AI capabilities trend carries a confidence interval of -8.4 to +23.4 points per year.
Apparent AI acceleration results from selecting frontier observations only, choosing a breakpoint, ignoring variance collapse, and fitting separate lines on either side of that breakpoint.
Related Articles:
Why .400 Hitters Disappeared — and What It Means for AI— Artificial intelligence models' capability gains are shrinking in percentage terms, falling from about 15% relative to the baseline early on to roughly 10% for the same absolute increment, narrowing the gap between frontier and good-enough models. The author argues this points to a commoditizing market where price, not minor model differences, becomes the main differentiator, pressuring frontier firms' margins.
Chart of the Day: Small Models are Closing the Gap to Frontier AI— Achieving 60% on MMLU, a test of academic reasoning, once required a 540bn-parameter model but was reached by a 3.8bn-parameter model two years later, a 142x reduction in parameters. Models small enough to run on phones or sensors shift inference from data centers to devices, weakening the case for paying cloud providers per token.
Anthropic’s Best AI Model Struggles To Attract Users As Cheaper Tools Thrive— Spending on Anthropic's largest and priciest AI model has plateaued at about 11% of total outlay on the company's tools, according to payments data from 70,000 companies, breaking a pattern of corporate users defaulting to the most powerful models. Analysts and investors cite its high price and older models' adequacy.
AI Summary.An index blending artificial intelligence token prices and usage is ambiguous when it softens, since token prices have fallen more than 90% since 2023 while total spending has roughly doubled over the past year. Allianz Research measures a nearly 46% growth gap between artificial intelligence investment and sales, wider than the 32% divergence during the 2001 telecom bust.
Jan-Patrick Barnert and Michael MsikaBloomberg
Date Posted:
Is Database:
Database
The Silicon Data LLM Token Expenditure Index, which tracks AI token prices, is down ~20% from its recent peak in May. The metric is a rough “proxy for marginal willingness to pay.”
Why does a softening AI token index send ambiguous signals?
Core argument: Token prices have fallen more than 90% since 2023 while total token spend has roughly doubled over the past year, supporting a benign read in which cheaper tokens expand the market and an index pause reflects digestion.
A softer index doesn’t mean AI is getting cheaper. The gauge blends prices and usage, meaning a dip can imply very different scenarios: either list prices are falling, or demand is shifting toward cheaper models. It could also point to a genuine softening in what buyers are prepared to shoulder. Each of these possibilities carries different implications. Let’s start with a benign read: While token prices have collapsed more than 90% since 2023, total spend has roughly doubled since last year. Cheaper tokens have expanded the market. This means that an index pause is simply digestion, while demand is real and capex is money well spent. The bull case for Nvidia rests here. Now for the interpretation that’s keeping people up at night: Bears warn that sustained weakness in the index could end the trade that saw nearly the entire AI cohort rally hard this cycle. It’s token spending that justifies the next capex order, and the bill is already looking stretched. Allianz Research said there’s nearly a 46% growth gap between AI investment and sales. That’s worse than the 32% divergence measured during the 2001 telecom bust.
Takeaways by Macro Roundup® AI
Token prices have fallen more than 90% since 2023 while total token spend has roughly doubled over the past year, supporting a benign read in which cheaper tokens expand the market and an index pause reflects digestion.
Allianz Research put the growth gap between AI investment and sales at nearly 46%, wider than the 32% divergence measured during the 2001 telecom bust.
The index blends prices and usage, so a dip can imply falling list prices, demand shifting toward cheaper models, or a genuine softening in what buyers are prepared to shoulder.
Related Articles:
AI: The ROI Runway Could Be Long Outside the Tech Sector— AI company valuations rest on the assumption that profit margins across non-tech firms will eventually rise, but outside a handful of sectors adoption is slow, creating a divergence between front-loaded valuations and slower cash flows. If the per-use costs of AI models converge toward zero, revenue may be insufficient for all major cloud providers.
Semiquincententacles— Leading artificial-intelligence labs are moving to usage-based token pricing on a cost-plus basis that includes total ownership costs, pushing token prices higher and giving some enterprises reason to migrate to cheaper open models. Frontier models remain necessary for cybersecurity, distillation training, agents, and scientific discovery, where maximum performance is required.
Chart of the Day: Small Models are Closing the Gap to Frontier AI— Achieving 60% on MMLU, a test of academic reasoning, once required a 540bn-parameter model but was reached by a 3.8bn-parameter model two years later, a 142x reduction in parameters. Models small enough to run on phones or sensors shift inference from data centers to devices, weakening the case for paying cloud providers per token.
AI Summary.A surge in US business formation is being driven by solo entrepreneurs, with growth in AI tool usage and in solopreneur activity in high-AI-adoption sectors suggesting artificial intelligence accounts for a meaningful portion of the increase. The authors argue AI is lowering barriers to starting and growing a business by expanding what one-person firms can do.
Ernie Tedeschi, Marisa Rama and Chris CruickshankStripe Economics
Date Posted:
Is Database:
Database
New US business applications have surged since late 2024, driven mainly by “likely nonemployer” filings representing “solopreneurs” rather than businesses expected to hire. Data suggest AI may be enabling this by filling capability gaps without hiring.
Is AI driving the surge in one-person business formation?
Core argument: The recent surge in US business formation is not reflected in high-propensity applications, but the evidence points toward a structural increase in genuine small business activity over 2025 and 2026, driven by solopreneurs.
While the recent surge in US business formation is not reflected in high-propensity applications, the evidence points toward a structural increase in genuine small business activity over 2025 and 2026, driven by solopreneurs. Preliminary evidence of growth in AI tool usage and solopreneur growth in high-AI-adoption sectors suggests that advances in AI are responsible for a meaningful portion of this growth. We believe AI is lowering barriers to business formation and growth by expanding the capabilities of solopreneurs, further improving tools and platforms that cater to new businesses, and creating a new set of opportunities for entrepreneurs to pursue. Whether the magnitude of this effect is as large as the most optimistic readings of the data suggest remains to be seen, but we believe we might be in the early innings of a fundamental acceleration in business formation—which could have ripple effects throughout the economy.
Takeaways by Macro Roundup® AI
The recent surge in US business formation is not reflected in high-propensity applications, but the evidence points toward a structural increase in genuine small business activity over 2025 and 2026, driven by solopreneurs.
Preliminary evidence of growth in AI tool usage and solopreneur growth in high-AI-adoption sectors suggests that advances in AI are responsible for a meaningful portion of this growth.
AI appears to be lowering barriers to business formation and growth by expanding solopreneur capabilities, improving tools and platforms serving new businesses, and creating opportunities for entrepreneurs, with possible ripple effects throughout the economy.
Related Articles:
Solopreneurs, Solow, and the SaaSpocalypse— For three decades after commercial electricity delivery began in 1882, inflation-adjusted output per worker grew just 0.5% annually, accelerating to more than double that pace only after factories were redesigned around electric motors rather than retrofitted from steam. The author argues a near-term absence of productivity acceleration from AI is not evidence none is coming.
AI-Native Firms∗— The authors find AI-native U.S. startups are 25% smaller than non-AI startups in the same industry and cohort, with 13% higher engineer shares, roughly 15% lower shares of entry-level workers and managers, flatter hierarchies, and comparable valuations. They argue embedding AI into products lets firms scale knowledge work without large teams.
Which Entrepreneurs Boost Productivity?— The authors find that a minority of entrepreneurs are "transformative" — generating disproportionate productivity gains, scoring high on IQ tests, being well-educated, and hiring research and development workers. They argue that economies with more such entrepreneurs and R&D workers sustain higher long-run productivity growth.
AI Summary.Achieving 60% on MMLU, a test of academic reasoning, once required a 540bn-parameter model but was reached by a 3.8bn-parameter model two years later, a 142x reduction in parameters. Models small enough to run on phones or sensors shift inference from data centers to devices, weakening the case for paying cloud providers per token.
Paul KedroskyApplied Complexity
Date Posted:
Is Database:
Database
Small models have increasingly closed the performance gap with frontier models. To Kedrosky, this implies “edge compute becomes real,” which in turn, at the margin, means reduced relative demand for data center-based versus local compute.
Are small AI models closing the gap with frontier models?
Core argument: Reaching 60% on MMLU, a test of academic reasoning, required a 540bn-parameter model in 2022 but only a 3.8bn-parameter model by 2024, a 142-fold parameter reduction in two years.
A quiet story in AI isn't that large models keep getting better, it's that small ones are catching up quickly. In 2022, reaching 60% on MMLU, a test of academic reasoning, required a 540-billion-parameter model. By 2024, a 3.8-billion-parameter model hit the same threshold — a 142-fold parameter reduction in two years. The direction is clear, and the implications are wide-ranging. Edge compute becomes real. Models that fit on a phone or a sensor don't need a data center. Inference shifts from hyperscaler to device, which is a different supply chain and a different competitive moat than anyone has built for. The data center API business model has a half-life. If a fine-tuned 7B model running locally can handle 80% of enterprise use cases, the case for paying per-token to a cloud provider shrinks fast. The most exposed providers are those whose moats are model quality rather than distribution, i.e., the capex-heavy frontier models. Compute becomes less scarce.
Takeaways by Macro Roundup® AI
Reaching 60% on MMLU, a test of academic reasoning, required a 540bn-parameter model in 2022 but only a 3.8bn-parameter model by 2024, a 142-fold parameter reduction in two years.
Models small enough to run on a phone or sensor shift inference from hyperscalers to devices, which involves a different supply chain and competitive moat than providers have built for.
The case for paying per-token to a cloud provider shrinks if a fine-tuned 7B model running locally can handle 80% of enterprise use cases, leaving capex-heavy frontier providers whose moats rest on model quality rather than distribution most exposed.
Related Articles:
Commoditization, Orchestration, and the New AI Stack— The author argues that as incremental gains from larger AI training runs shrink, the return on additional computing power falls, weakening the case for the training spending that dominates data center demand. Survivors of model commoditization will be those owning the orchestration layer or the applications where outcomes are priced.
The Geopolitics of AI: Decoding the New Global Operating System— The performance gap between the top American and Chinese AI models narrowed from 9.3% to 1.7%, with Chinese firms relying on cheaper open-source models—DeepSeek prices usage 20-40x below OpenAI's ChatGPT. This price advantage could win adoption in developing countries and build dependency on Chinese AI ecosystems.
Misanthropic: On Mythos, Bad Human Behaviors and Systems Vulnerabilities— Anthropic reports that its Mythos model has detected thousands of high-severity cyber vulnerabilities, including ones found by chaining multiple obscure software weaknesses, and that these hacking skills are emergent rather than a designed capability. Mythos completed a corporate network attack simulation estimated to take an expert over 10 hours, which no other frontier model had solved.
AI Summary.Euro-area income per person falls behind the US when measured in constant 2021 prices but gains on the US when measured using each year's prices, suggesting Europe is holding its own in applying technology. The author argues Europe's real risk is being cut off from strategically important technologies.
Paul KrugmanPaul Krugman Wonks Out
Date Posted:
Is Database:
Database
In 2021 prices, per capita GDP in the Euro area is falling relative to that of the US, but the opposite is true if one takes into account the benefits to Europe from falling tech prices. “The big benefits of IT come from applying it, [not] creating it.”
Does Europe's income gap with the US depend on price measurement?
Core argument: Euro-area GDP per capita relative to the US declines over the past 25 years when measured in constant 2021 prices, as the World Bank does, but rises when measured at each year's prices (PPP).
Let’s start by looking at GDP per capita in Europe (actually the euro area) as a percentage of GDP per capita in the US. If we do this using constant prices — the World Bank uses 2021 prices — we get the line in Chart 1 labeled “2021 prices.” This line shows Europe falling behind over the past 25 years. If, however, we simply use prices in each given year, we get the line labeled “PPP,” which shows Europe gaining on the US. We get a similar picture if we look at GDP per worker-hour, where the black line is calculated using 2021 prices and the blue line is calculated using PPP. IT progress is passed on to all consumers via lower prices. The big benefits of IT come from applying it, rather than creating it. And as I’ve tried to show, the data show Europe holding its own in the relative value of the goods it produces, indicating that European economies are doing fine when it comes to applying technological advances. What should worry Europe, instead, are the geopolitical implications of US/Chinese leadership in advanced technology. The risk of being cut off from strategically important technologies, once minimal, is now very real. And that risk, rather than misleading numbers about trends in real GDP per worker hour, is what should concern European policymakers.
Takeaways by Macro Roundup® AI
Euro-area GDP per capita relative to the US declines over the past 25 years when measured in constant 2021 prices, as the World Bank does, but rises when measured at each year's prices (PPP).
The same divergence appears in GDP per worker-hour, where the constant-2021-price series and the PPP series point in opposite directions for Europe relative to the US.
The author argues that Europe is holding its own in the relative value of the goods it produces and in applying technological advances, and that the real concern is the risk of being cut off from strategically important technologies amid US and Chinese leadership.
Related Articles:
Europe v America: Who’s Really Winning?— Real GDP data show the US economy growing substantially faster than the European Union's, yet on a purchasing-power basis adjusting for price levels, the EU economy remains only slightly smaller than the US, as it was earlier. The author argues US dominance in information technology, where productivity growth is rapid and passed on as lower prices, reconciles the two measures.
Europe’s Tech Lag: Does It Matter?— If technology producers face competition, faster measured productivity growth in a tech-specialized country is passed to consumers in both countries through lower prices rather than higher relative living standards. Real GDP per worker over 1998-2024 diverged more between California and the rest of the US than between the US and Europe, largely reflecting industry composition.
From IT to AI: What Explains US Productivity Outperformance?— Larger information-technology quality adjustments and hours-worked measurement distortions explain about 30% of the average annual US-Euro area total factor productivity growth gap since 1995, leaving an adjusted gap near 0.25pp. The authors estimate greater US intangible investment spillovers contribute about 0.1pp, more efficient resource allocation about 0.1pp, and better management and larger firms about 0.05pp.
AI Summary.Using an AI model to estimate how much earlier each of 190 major inventions could plausibly have appeared, the author finds most gaps are small: 64% had an "earliest plausible" date within 50 years of the actual date, and 90% had an "earliest straightforward" date within 50 years.
Brian PotterConstruction Physics
Date Posted:
Is Database:
Database
An analysis using Claude finds that for more than half of 190 inventions, the delay between the earliest date at which the invention became possible and the invention’s realization was 10 years or less. Since 1900, the fraction rose from ~50 to 75%.
How much earlier could major inventions plausibly have appeared?
Core argument: Of 166 major inventions for which Claude Opus 4.7 estimated a date, 107 (64%) had an "earliest plausible" invention date within 50 years of the actual date, and 150 (90%) had an "earliest straightforward" date within that window.
I used a list of 190 major inventions. For each invention, I asked Claude Opus 4.7 how much earlier it could have been invented. The graph shows how much earlier, on average, each invention could have appeared for both the “earliest plausible” and “earliest straightforward” date ranges. We can clearly see a few trends on this graph. One is that for most inventions, the gap between when they could have been invented and when they were actually invented is not particularly large. Of the 166 inventions Claude estimated a date for, 107 of them (64%) had an “earliest plausible” date 50 years or less from the actual date, and 150 of them (90%) had an “earliest straightforward” date 50 years or less from the actual date. For more than half the inventions, the average earliest straightforward date of invention was 10 years or less from the actual date.
Takeaways by Macro Roundup® AI
Of 166 major inventions for which Claude Opus 4.7 estimated a date, 107 (64%) had an "earliest plausible" invention date within 50 years of the actual date, and 150 (90%) had an "earliest straightforward" date within that window.
For more than half of the inventions assessed, the average "earliest straightforward" date of invention was 10 years or less from the actual date of invention.
Asking Claude Opus 4.7 how much earlier each of 190 major inventions could have been invented suggests that for most inventions the gap between the earliest possible and actual date is not particularly large.
Related Articles:
Attention (And Money) Is All You Need: Why Universities Are Struggling to Keep AI Talent— The annual compensation gap between top industry and academic artificial intelligence researchers widened from roughly $294,000 to $1.5m, coinciding with accelerating academic out-migration toward large incumbent technology firms. Researchers who move to industry cut paper production by about 63% while increasing patent filings nearly sixfold, consistent with a shift from open science toward proprietary innovation.
Never Enough: Dynamic Status Incentives in Organizations— Data on over 5,000 German fighter pilots show performance rose markedly as pilots approached the eligibility threshold for each medal in a tiered awards ladder and slumped immediately after receiving it, with the effect repeating at each successive rung. The authors estimate the ladder added 3,000–5,000 aerial victories, 5–10% of the total.
Are Ideas Getting Harder to Find?— The authors estimate U.S. research productivity has fallen by a factor of 41 since the 1930s, averaging declines above 5% per year across sectors. Sustaining constant growth in output per person requires doubling research effort every 13 years, challenging growth models that assume constant research productivity.
AI Summary.Sectors with the highest artificial intelligence adoption rates have also seen the strongest growth in new business applications, which one analysis interprets as evidence that AI is lowering the barriers to starting a company.
Torsten SløkApollo
Date Posted:
Is Database:
Database
Since 2022, sectors with the highest AI adoption rates have also seen the strongest growth in new business applications, suggesting to Torsten Sløk that “AI is lowering the barriers to starting a company.”
Is AI lowering the barriers to starting a company?
Core argument: Sectors with the highest AI adoption rates have recorded the strongest growth in new business applications since 2022.
Sectors with the highest AI adoption rates have also seen the strongest growth in new business applications since 2022, showing that AI is lowering the barriers to starting a company.
Takeaways by Macro Roundup® AI
Sectors with the highest AI adoption rates have recorded the strongest growth in new business applications since 2022.
Torsten Sløk reads the link between AI adoption and new business applications as a sign that AI is lowering the barriers to starting a company.
Related Articles:
Generative AI and Entrepreneurship— The authors find that startups with a larger share of tasks doable by generative AI cut employment beginning immediately after ChatGPT's release, with technically proficient founders restructuring teams toward higher-skill roles. Venture investors made smaller but more numerous deals, and new firm entry offset the decline in total startup employment.
Attention (And Money) Is All You Need: Why Universities Are Struggling to Keep AI Talent— The annual compensation gap between top industry and academic artificial intelligence researchers widened from roughly $294,000 to $1.5m, coinciding with accelerating academic out-migration toward large incumbent technology firms. Researchers who move to industry cut paper production by about 63% while increasing patent filings nearly sixfold, consistent with a shift from open science toward proprietary innovation.
Which Entrepreneurs Boost Productivity?— The authors find that a minority of entrepreneurs are "transformative" — generating disproportionate productivity gains, scoring high on IQ tests, being well-educated, and hiring research and development workers. They argue that economies with more such entrepreneurs and R&D workers sustain higher long-run productivity growth.
