"A.I. and Our Economic Future," Professor Chad Jones

Introduction to AI and Its Transformative Potential

• The topic of AI and its potential impact on the future is a significant area of concern, with AI being considered the most transformative technology of our lifetime, similar to earlier technologies like electricity, the transistor, semiconductors, information technology, and the internet 10s.
• Two scenarios are presented as extremes to consider the potential effects of AI: the first scenario is where AI dramatically accelerates economic growth, and the second scenario is where AI is just a normal technology, similar to how electricity and semiconductors were transformative but not revolutionary 2m6s.
• The first scenario, where AI accelerates economic growth, is potentially already happening, with AI automating software engineering and possibly automating most coding in the next decade, leading to AI agents that can do everything a software engineer can do 5m45s.
• As AI agents become capable of performing tasks currently done by humans, they can be put to work on AI research, building better algorithms, and improving the AI itself, potentially leading to agents that can function as virtual remote workers 8m38s.

Economic Impacts of AI: Two Scenarios

• The possibility of AI performing every task a human can do raises questions about what it would be like to live in a world where machines can perform cognitive and physical work, and how this would impact the economy and society 3m30s.
• Research has been conducted to explore the potential consequences of AI on the economy, including the possibility of AI dramatically accelerating economic growth, and the potential for AI to automate many jobs, leading to significant changes in the workforce 1m40s.
• The concept of billions of virtual research assistants, each running 100 times faster than humans, can lead to the discovery of new ideas and the design of better technologies, such as computer chips, robots, and pharmaceuticals, with the potential to transform industries like robotics and pharmaceuticals 10s.
• The idea of a "country of geniuses in a data center" can automate cognitive tasks and, eventually, physical tasks, leading to explosive growth, as seen in growth models, where the automation of both cognitive and physical tasks can result in rapid growth 42s.
• This scenario is considered plausible, especially in Silicon Valley, but the timeline for its occurrence is uncertain, with possibilities ranging from three to 25 years, and its impact on the world would still be significant regardless of the timeline 2m6s.
• An alternative scenario is that AI is just a business-as-usual technology, and its impact on growth rates may be limited, as seen in the historical context of the United States, where average living standards have risen at a consistent 2% per year over the past 150 years, despite the introduction of transformative technologies like electricity, internal combustion engines, and the internet 4m37s.
• The introduction of new technologies has allowed the 2% growth rate to continue, as each technology has enabled growth to persist for another 50 years, and it is possible that AI may be the latest technology to allow this growth rate to continue 8m10s.
• The two scenarios presented, one with explosive growth and the other with steady 2% growth, both have merit and represent two extremes of the potential impact of AI on the economy 10m30s.

Historical Context of Technological Adoption

• Economists who have studied the replacement of steam power with electric power and the diffusion of information technology throughout the economy have found that these transformations take decades to complete, requiring the reorganization of factories and the integration of complementary innovations 10s.
• The process of adopting new technologies, such as electric power or information technology, involves a series of complementary innovations, including the invention of new tools and the reorganization of production, which can take time measured in decades 42s.

The Concept of Weak Links in Economics

• The concept of weak links, which suggests that a chain is only as strong as its weakest link, can be applied to business success, where completing many tasks successfully is required, and the failure of one task can lead to the loss of value 2m6s.
• The weak link model can be illustrated by examples such as the Space Shuttle Challenger explosion, which was caused by the failure of a $25 rubber seal, or the complex manufacturing process of computer chips, where one small error can lead to the failure of the entire product 4m30s.
• The idea of weak links can help explain why having a computer with 100 million times the transistors of a 1970s computer does not necessarily make one 100 million times more productive, as there are other limiting factors, such as the ability to figure out what data to use or what questions to ask 6m15s.
• Weak links are the source of scarcity, which is a key concept in economics, as scarcity gives rise to high returns, and identifying the weak link is essential to understanding economic phenomena 8m40s.
• The concept of weak links can be used to study economic phenomena, such as the impact of technological advancements on income, by identifying the limiting factors and understanding how they affect the overall system 10m10s.

Labor and Capital Share in GDP

• Economists are interested in understanding who gets paid and what share of GDP is paid to labor versus capital, with the share paid to labor having fallen by about 10% in the last 25 years, while the share paid to capital has increased, 10s.
• There are two main theories to explain this shift: automation and market power, with the possibility that big firms are exercising more market power and capturing more of the GDP as profits, 42s.
• The share of GDP paid to labor can be broken down into components, such as the share paid to people with more than a college education versus less than a college education, and the share paid to capital can be broken down into returns to buildings and structures versus equipment, 2m6s.
• The share of GDP paid as a return to computing power has actually decreased over time, from a peak of just under 4.5% in 2000 to around 3% today, despite the increasing presence of computers in the economy, 4m30s.
• This decrease is due to the price of computing power falling faster than the increase in quantity, which is consistent with a weak-link model that predicts that the most plentiful factor will receive a lower share of GDP, 5m15s.

Modeling AI and Automation in Economic Growth

• A model has been developed to study the scenarios of automation and AI, featuring ideas as the source of long-run growth and production functions with weak links that can be automated away, 6m40s.
• The model is calibrated to fit historical US data back to the 1950s and can be used to run simulations to ask what will happen in the future, including the potential impact of automating certain jobs, such as software engineers, 8m30s.
• A thought experiment can be done in the model to ask how much richer we would be if we automated all software engineers, which is an interesting question given the expectations of companies like Anthropic and OpenAI, 10m20s.
• If the economy had infinite amounts of software, it would only be 2% richer because other weak links would bottleneck the growth, as software is about 2% of GDP 10s.
• Automating one thing really well is not enough, and what is needed is to keep automating the weak links, requiring a dynamic model rather than a static one 2m6s.

Simulating Future Economic Scenarios with AI

• A model featuring automation and weak links can be used to simulate the future, with two key scenarios: one where AI is a continuation of historical patterns of automation, and another where AI is a break with the past, with the entire economy getting automated at a rate similar to Moore's Law 4m42s.
• The first scenario assumes AI is just a continuation of historical patterns, while the second scenario takes a more aggressive approach, assuming the entire economy starts getting automated at a rate of 10% per year, with automation giving more ideas and more automation 6m15s.
• The simulations show three sets of graphs, including the capital share versus the labor share, with different scenarios resulting in varying shares of GDP paid to capital and labor, including a scenario where the share of GDP paid to capital goes to 100% and the share paid to labor goes to zero 10m10s.
• The scenarios also include one where 3% of tasks are reserved for humans, resulting in the human share going to 100% and the capital share going to zero, similar to the example of computers where the share paid to computers is falling 12m40s.
• The baseline scenario, also referred to as the blue scenario, assumes that the capital share remains stable, with labor receiving two-thirds of the share, and this scenario is interesting because it explores what happens to growth when the capital share and labor share remain constant, with growth rates initially at 2% 10s.
• In the baseline scenario, growth accelerates over time, but at a slow pace, eventually settling down at 50% per year, although it takes centuries to reach this point, with the acceleration being stunningly slow, and by 2050, the growth rate increases to 2.3% 2m6s.
• The growth rate graph shows that the acceleration is slow, and when looking at the levels of income per person, it is clear that growth is accelerating, but the explosion is slow due to weak links, such as the limitations of human capabilities, with the result being that by 2050, people are 4% richer, and by 2075, they are 15% richer 4m20s.
• The different scenarios, including the green and purple lines, show very different futures 200 years from now, but for the next 75 years, it is difficult to determine which scenario is most accurate, which is why a second set of scenarios is necessary to consider alternative possibilities 6m30s.

Alternative Growth Scenarios with AI

• A new set of scenarios assumes that AI starts out with Moore's Law everywhere, with machines getting better throughout the economy at 10% per year, instead of the baseline 3% per year, resulting in significantly higher growth rates, with the purple line and blue line showing explosive growth, and by 2050, the economy is well on its way to achieving high growth rates 10m0s.
• The new scenario shows that if AI starts out with Moore's Law everywhere, the economy grows at 4.7% instead of 2%, and by 2040, the growth rate reaches 7% and 13%, with the explosion happening relatively quickly, but not until 2060 is the explosion fully complete, according to this scenario 12m40s.
• The concept of accelerating growth due to AI is explored, and even in an aggressive scenario where people are 50% richer in 2030, the explosion of growth takes 30 years due to weak links 10s.
• The weak link phenomenon slows down the growth process, and it is a key lesson to take away from the scenario, as it affects how quickly AI can automate tasks and lead to exponential growth 42s.
• The idea of growth exploding over the next 50 or 100 years is discussed, but it is noted that the explosion is not as fast as one might expect, and it is attributed to the weak links that need to be automated away before the flywheel effect takes over 2m6s.

AI and the Future of Work: Examples and Predictions

• The example of Jeff Hinton, a Nobel Prize winner and inventor of deep neural networks, is given, where he predicted in 2016 that AI would replace radiologists in five years, but it is noted that while AI has improved, radiologists are still needed and their wages have increased 4m10s.
• The concept of jobs being bundles of tasks is explained, and how automating 75% of tasks can actually raise wages, as the remaining tasks become more valuable and scarce, leading to higher productivity and wages 6m15s.
• The example of Uber drivers is given, where it is predicted that they may be replaced by self-driving cars, but it is noted that this process takes time, and the development of self-driving cars has been slower than expected 8m30s.
• The topic of inequality and meaningful work is introduced, and it is noted that historically, labor has been the main asset that people trade to get consumption, but with machines being able to do things better than humans, this dynamic may change 12m0s.

AI and Inequality: Implications and Redistribution

• The world where AI changes everything is a world with incredibly high GDP, abundance, and plenty to go around, which could lead to increased consumption for the bottom 10% of the population, even if current US redistribution programs remain in place 10s.
• Economists aim to make everyone better off, but this does not happen automatically, as seen in the example of the China shock, where people in North Carolina lost their jobs and communities were decimated 2m6s.
• The development of AI models, such as GPT 5.2 Pro and 5.5 Pro, which are already as good as or better than humans at math, raises concerns about the potential for AI to replace human jobs and leave people without meaningful work 4m30s.
• An analogy for a potential future where AI has replaced human work is retirement, where people find happiness and meaning in activities such as traveling, socializing, and hobbies, and a similar concept could apply to people whose jobs are replaced by AI 6m20s.

Risks and Catastrophic Potential of AI

• However, there are also catastrophic risks associated with AI, including the "bad actor version," where a hacker or malicious individual could use AI to cause harm, such as designing a lethal virus, and the "alien intelligence version," where a highly advanced AI could pose an existential threat to humanity 10m30s.
• The "alien intelligence version" is more speculative, but it raises important questions about how humanity can retain power over entities more powerful than itself, as quoted by computer science professor Stuart Russell 14m40s.
• The potential risks and consequences of advanced AI highlight the need for serious discussion and consideration of these issues, without pejorative criticism, to ensure that the development of AI is aligned with human values and safety 16m50s.

Weak Links and System Fragility in AI Development

• The concept of a chain being only as strong as its weakest link is relevant to the development of artificial intelligence, as the benefits of AI come slowly because all weak links need to be automated and strengthened to achieve full benefits and explosive growth 10s.
• A weak link model is not only slow to improve but also fragile on the downside, and this concept is illustrated by the example of the Space Shuttle Challenger's O-ring problem, where breaking one link in the chain results in the loss of all value 42s.
• The Mythos model, developed by Anthropic, has discovered thousands of bugs in 25-year-old software that had been battle-tested by humans for 25 years, and there is a concern that a bad actor could use this model to hack critical systems such as the electric grid, financial system, or banking system 2m6s.

Downside Risks and Systemic Vulnerabilities

• The potential risks associated with AI include hacking the electric grid, banking system, or communicating with bio labs to design viruses, and these risks could have significant consequences, such as zeroing out everyone's bank balance, although they may not be existential 4m30s.
• The impact of AI on the world between 2015 and 2045 is expected to be transformative, possibly equivalent to multiple "internets," but it may take longer than expected, and the downside risks could come sooner 8m45s.

Long-Term Impacts and Measurement Challenges

• The weak link view of the world suggests that the effects of AI will ultimately be huge and transformative, but it also delivers the possibility of catastrophic risks, and it is essential to use the intervening years to prepare for these risks, including inequality, labor market, political economy, and catastrophic type risks 12m10s.
• The measurement of GDP growth rate may not accurately reflect the value created by AI, as some benefits, such as infinite free chess lessons, are not monetized, and this is a challenge in capturing the true value of AI 16m30s.
• The idea that people could have either the GDP growth of the 20th century or the gains in life expectancy, but not both, is presented, with life expectancy increasing from 50 to 77 years, and it is noted that the gains in life expectancy are not adequately captured in GDP 10s.
• The possibility that things will be increasingly mismeasured in the future is discussed, and it is suggested that simulations holding constant the measurement may not fully capture the potential for faster growth 1m20s.

Audience Questions and Short-Term Economic Effects

• An audience member asks about the short-term effects of AI on the economy, specifically the potential for automation to eliminate certain tasks and disrupt the economy, and it is noted that this is an important question that is being explored in a follow-up paper 2m40s.
• The example of radiologists and self-driving cars is used to illustrate the potential for AI to automate certain tasks, but also the potential for these changes to take longer than expected due to "weak links" in the system 4m10s.
• The idea that software engineering may be one of the first areas to be automated is discussed, but it is also suggested that the integration of AI into businesses will require many software engineers, potentially limiting the impact of automation in this area 6m20s.
• An audience member notes that the rise of radiologists and their incomes may be due to increased demand for medical services and variability in medical techniques, and suggests that automation in this area could potentially strengthen the "weak link" in the system 9m40s.
• The contrast between the potential for automation in areas like radiology and code is highlighted, with the suggestion that code may be more susceptible to automation due to its more structured nature 11m10s.

AI in Education, Healthcare, and Services

• The development of artificial intelligence (AI) and its potential to replace human workers, such as kindergarten teachers, is a topic of discussion, with the possibility of robots being trained to be the world's best kindergarten teachers and then replicated to be used in every classroom 42s.
• The growth of economies, such as Taiwan and South Korea, which have substantial IT output, is mentioned, and it is noted that these economies have outsourced manufacturing and replaced it with services like healthcare and financial services, which are areas that are easily threatened by AI 4m6s.
• The idea that manufacturing is easier to automate than services, and that services like healthcare and education are more difficult to replace with AI, is discussed, with the example of a nurse holding the hand of a patient with Alzheimer's being a task that may not be replaced by a robot in the next 20 years 6m15s.
• The concept of a K-shaped economy, where the wealthy accumulate more wealth and the poor struggle to make ends meet, is mentioned, and the question of what stops the hyper-concentration of capital share of income, leading to an oligopoly, is raised 8m30s.

AI and Inequality: Diverging Economic Outcomes

• The potential impact of AI on inequality is discussed, with the idea that AI may replace high-skilled cognitive labor, such as doctors, lawyers, and economists, but may not replace lower-skilled labor, such as electricians and plumbers, which could potentially reduce inequality 10m45s.
• The role of government in addressing inequality through taxation and transfer systems is mentioned, with the idea that the government can help to put a floor under the less fortunate and reduce the negative impacts of AI on the economy 14m20s.
• Management skills taught in business schools will still be valued in the future, as managers will consult with AI systems and make final decisions, making them a scarce and valuable resource 10s.

Global Implications and Economic Disparities

• It is advised to own shares of the S&P 500 to earn high returns, as the people who own these shares will be okay in the long run, with managers being safe for another 15 years before automation potentially takes over 42s.
• Different societies may progress at different speeds when it comes to adopting AI technology, which could lead to huge differences in economic outcomes, with countries that have software engineers and create ideas behind neural networks being better off 2m6s.
• The global implications of AI on the economy, particularly in developing countries with high inequality, are not well studied, but it is hoped that the ideas invented in more developed countries can help those in need 2m6s.
• As AI takes over easier jobs, a small group of people may become busier and receive higher pay, while a larger group may have more leisure time, raising questions about how society will think about wealth distribution and resource allocation 4m30s.

Redistribution and the Role of Government

• In a world of abundance created by AI, there will be enough resources to redistribute, and it is possible for the worst-off individuals to become millionaires, but solving the redistribution question will be crucial 6m40s.
• Economists like Daron Acemoglu are optimistic about the potential for a world of abundance to work out well, with the tax system and philanthropy playing a role in redistributing resources 8m20s.
• Roberto Santana, a Google DeepMind employee and class of 2011, is trying to reconcile the model being discussed with the natural experiments they are running, but they do not match up, as the model assumes that the constraint, or weak link, is a human and that AI cannot take or improve in those tasks, such as coordination, judgment, and taste 10s.

Model Calibration and Real-World Challenges

• The model, based on Gradualism, assumes that AI has to be gradual, which is the piece that Roberto Santana is trying to reconcile, and this reaction is acknowledged as a valid point, as the model was written and calibrated to historical data to test this assumption 1m42s.
• The calibration of the model, including how strong the weak links are, is a key question, and it is possible that they were calibrated to be too strong, but the example of Waymo, which predicted that self-driving cars would be a solved problem by 2012, is used to illustrate that even seemingly simple problems can be complex due to bottlenecks and weak links 2m6s.
• The physical world is complicated, and while machine learning algorithms can perform certain tasks better than humans, automating everything else in the physical world is a challenging task, and it is uncertain whether this will happen in 20 years or 100 years, but it is unlikely to happen in five years based on the weak link argument 4m10s.
• The example of computers is also used to illustrate the slowdown, as their factor share has gone down, and automating everything else in the physical world is necessary to achieve significant progress, which supports the intuition behind the slowdown 5m40s.