Erik Brynjolfsson may well be the best possible person with whom to discuss the potential of current AI systems as he is the rare combination of a world-leading economist focused on understanding the evolution of productivity, and a lifelong practitioner and proponent of machine learning and AI systems. As a result, he is uniquely able to translate the emerging technical capabilities into possible economic futures, not only in terms of the value to individual businesses but also to the economy as a whole.
Brynjolfsson occupies a unique position as both a Stanford University professor and the head of the digital economy lab at the Institute for Human Centered AI, which allows him the freedom to pursue analyses that are not compromised by a particular corporate or financial agenda, but are still grounded in economic reality and, at the same time, also account for the human part of the equation. Indeed, one of the primary conclusions of our conversation is that augmentation of human tasks is where the real economic gains are to be found, rather than in replacing human activity by automation, and consequently that, as he puts it, “We need to treat humans as an end and not just a means to an end.” But what is particularly striking is that this seemingly facile and human-validating imperative is anything but that: it is based on hard economic facts and rigorous analyses that are gradually revealed over the course of our conversation.
Brynjolfsson was born in Denmark to an Icelandic father and a French mother, before the family moved to U.S. as part of the “brain drain” migration, with his father being recruited into the Apollo program in the 1960s. His Icelandic heritage is evident not only in the naming of his children according to Icelandic tradition (Eriksson) and stature (Icelandic men rank in the top 5 globally in height), but in his appreciation for mythology, which he invokes when describing the current moment in technological time, “All through history, people have written myths about robots, about golems, about other creatures, where somebody created an entity that looked like a human, talked like a human. It was always mythology, but now our generation, you and I, are living through this amazing time where people are actually creating these entities, mythological technologies. And so, they’re not just mythological anymore.” He is, of course, referring to the preternatural ability of LLMs to imitate human language and to interact in the most human of ways, but also implicitly in a way that can be deceptive.
His Massachusetts origin story continued after high school as he studied applied mathematics at Harvard University, before eventually undertaking a Ph.D at MIT. His interest in both the theory and the practice of AI systems was apparent early on, “after college, I taught a course on building expert systems…and cofounded a company and we were doing interesting things, but I felt like I wanted to get in a little deeper and really understand the material.” His goal was to do a joint Ph.D in artificial intelligence and economics, but no such program or faculty sponsor existed at that time, so he chose to study at MIT due to the excellence of both the business school and computer science departments at that august institution.
Reflecting back on those early days, he feels a familiar sense of excitement, “in the late 80s, [expert systems] were all the rage—AI was going to be people like me handwriting the rules of experts and codifying it. It worked okay, but it really didn’t scale. So, I’m really excited about the time in history we are in today where neural nets are allowing machines to learn from data, instead of us humans having to kind of hand code them.” Similar to Gary Marcus, he argues he had a measure of prescience about this evolution, albeit not about the rate at which it occurred, “I wouldn’t say it surprised me. To be frank, from when I was pretty young, I thought this was coming. I didn’t know exactly how fast or how slow it would come. Parts of it were slower than I expected, but the past few years have gone way faster than I expected.”
As with all the interviews in this series, our conversation is anchored by aligning on a definition of human intelligence to serve as a common foundation for the discussion of the artificial form. Brynjolfsson agrees with the canonical definition that has emerged from the series, that intelligence is comprised of a few essential components:
- The ability to process inputs from multiple sources (and senses).
- The ability to derive an essential understanding.
- The ability to produce original output(s).
- The ability to do this with little or no training.
But he extends this definition by reference to his training as an economist—that something of value must be produced, so intelligent behavior cannot be “just creating meaningless theorems or meaningless content, but stuff that actually is useful. I think that’s also an important part of intelligence.”
He is not a fan of nebulous definitions of different forms of AI, as his background demands something more concrete, “people talk about AGI, artificial general intelligence, even super intelligence, and they have a lot of different technical definitions. As an economist, what I’m most interested in is how it’s going to change the economy, how it’s going to transform it.” At a recent conference he co-founded on what he likes to call “Transformative AI” (TAI), a primary definition of TAI was an “AI technology that is sufficiently impactful in the economy that it creates a transition comparable to what we saw between the Agricultural Age and the Industrial Revolution.” When pushed on what this means in practice, he clarifies that this means a couple of orders of magnitude in productivity growth, but he sees this happening over a vastly shorter period of time than the 100 years or so it took for the full effect of the industrial revolution to take hold—more like a few decades this time around. He also references Dario Amodei’s vision for “powerful AI” to describe the end state: “Imagine you had a country of geniuses in a data center. You know, hundreds of thousands or millions of Einsteins working in parallel, maybe working 10x or 100x faster than a human brain can, and with expertise in not just physics, but chemistry, economics, marketing, journalism…in all the cognitive areas.”
But Brynjolfsson makes a further critical observation that is highly relevant to current AI systems, that one must distinguish between the different domains of knowledge upon which an intelligent system must operate, in particular between codified and embodied knowledge, in order to measure progress towards full human-equivalent intelligence.
Not All Knowledge Is the Same
The distinction between codified and embodied knowledge is crucial for understanding how knowledge is acquired and utilized in various contexts. Codified knowledge, often referred to as “know-how,” can be thought of as the knowledge that can be expressed in words; it can be articulated and applied in a systematic manner. In contrast, embodied knowledge, or “know-what,” is often tacit and intuitive, allowing individuals to perform tasks without conscious thought. Codified knowledge is often associated with formal education and systematic learning, while embodied knowledge is more intuitive and often learned through sensorimotor-based personal experience. It is intuitively obvious that current AI systems are significantly more adept in the codified knowledge realm, based on their almost exclusive training using formal and written knowledge, and the relative dearth of comparable descriptions of the physical, embodied realm. Brynjolfsson sees this as “Moravec’s paradox on steroids,” referring to the famous observation by Hans Moravec that it is “comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.”
He observes that “We’ve been very fortunate that there’s a lot of written data about a lot of our cognitive processes, and that has allowed us to train LLMs that do a lot of cognitive tasks, but there’s not the same kind of detailed data about physical processes. That’s something that we humans understand, like picking up a glass of water or breaking a pencil or buttoning a shirt, they require a lot of detailed data about forces and directions and so forth. That hasn’t been codified anywhere. It’s not on the internet, and that means that we don’t have the training data for these physical systems.”
But Brynjolfsson does not think we will have to wait for embodied “world models” before seeing significant economic gains, “No question, we’re going to have a huge productivity improvement just from the cognitive part. Now, many people have this mantra, or this goal, of solving all of the problems, and we won’t be able to solve those until we have the world models. But from an economist’s perspective, there are trillions, tens of trillions, maybe hundreds of trillions, of dollars’ worth of value that can be created from, quote, ‘just the cognitive side’,” he conjectures.
The Other GPT
The obvious question is why such productivity gains are not yet apparent. Indeed, multiple recent studies have reported that 95 percent of AI projects have yet to reach real commercial deployment or that only 39 percent of businesses see any EBIT impact currently due to AI. But this is where Brynjolfsson’s unique combination of expertise comes to the fore, to provide a compelling narrative, supported by rigorous analyses, on what is going on, and moreover, that this situation is even to be expected early in the technological lifecycle. He argues that early productivity deficits are actually most prevalent for so-called General Purpose Technologies or GPTs, a set of technologies that continuously improve over time, impacting multiple different areas and skills across the economic spectrum, with many unexpected “spillover” effects, of which AI is widely thought to be the latest example. For clarity, this “GPT” acronym is unrelated to the use of the same acronym for Generative Pre-trained Transformer—the foundational technology of LLMs such as chatGPT: a perverse coincidence of acronyms if ever there were one.
Brynjolfsson clearly sees the current set of AI technologies as not only the latest GPT in a very exclusive club, but he also argues that it is potentially the most important: “AI is not the first GPT, but it may be the last one, because it is our ultimate invention that allows us to do other things. Earlier GPTs were the steam engine, electricity, and the internal combustion engine. These are technologies that are pervasive, they affect almost all parts of the economy, they improve over time, and most importantly is that last one that they spawn complementary innovations—they trigger innovations in other technologies and in business processes, even in human skills. I don’t think there’s any technology that fits better than AI. It is, in many ways, the ultimate GPT.” However, the critical point is that the timescale of the impact of such technologies is invariably much longer than one would expect. This is a result of what Rodney Brooks characterized as “exponential thinking” about such technologies that are perceived as magical, and the resultant applicability of Amara’s Law that states that we tend to overestimate the short-term impact and underestimate the long-term impact of technological change.
Shooting the J
A recent study by Anthropic attempted to impute productivity gains by estimating the value of time saved by tasks undertaken with their LLM, Claude, resulting in an estimated additional productivity growth of 1.8 percent per annum over the next decade, effectively doubling the average growth rate since 2019 (1.8 percent), or since 1947 (2.1 percent on average). However, such claims must be regarded with a healthy dose of skepticism, as too many assumptions are made regarding how task time savings are utilized in practice and whether individual time savings manifest directly as real business productivity growth or are lost in the morass of operating inefficiencies.
Importantly, Brynjolfsson has been able to put this AI revolution on a more rigorous economic footing, with the observation that such profound technological transitions are accompanied by what he calls “productivity J-curves,” which he explains as follows: “We’re mismeasuring some of the outputs, and we’re also mis-measuring some of the inputs. On the input side, what we’re mismeasuring is the value of intangibles comprised of new business processes, new skills, new ways of organizing work. Every time you have a powerful GPT, it also means there are these complementary investments. Often, those investments are significantly larger than the initial investment in the GPT, maybe 10 times larger, according to some of my research. Since we don’t measure those well, we get this unfortunate situation, where companies invest in new business processes and new skills that takes a lot of time and effort, managerial attention, even payments to consultants and others, but it doesn’t show up as more output. It just shows up as spending. We’re not measuring the intangible value [that is being accrued]. And to some extent, [businesses] will be critiqued for that spending because on paper, it looks like their productivity has actually gone down. But if these are wise investments, these investments will tend to pay off. Then you start harvesting the intangibles as you have a new system that can produce twice as much output without more physical capital, without more labor, so your productivity is much higher.”
To summarize this phenomenon: multi-factor productivity (a common measure of productivity) is defined as the output volume of a business or economy divided by the amount of capital, labor and materials consumed in producing that output. So, if there are increased capital and labor expenditures in the early phase of the technological transition of a business (or the economy as a whole), without tangible output increase, the productivity growth will be computed to be negative—the downward “loop” of the J. Conversely, when enhanced output is observed in the longer term, the classical calculation will not associate these gains with the prior increase in spending that led to the accumulated efficiency, and the productivity growth will be calculated to be anomalously high, giving rise to abrupt upward slant of the J.
It is challenging to directly observe these J-curves in productivity data, as there are multiple overlapping curves associated with the different component elements, but Brynjolfsson and his collaborators have developed a methodology based on an economic quantity known as Tobin’s q factor, that compares the market capitalization of a company to its book value and associates any discrepancy with the intangible output value of a company at any point in time, which is correctly accounted for by the “wisdom of the crowd.” Their analyses reveal that there is significant intangible value in hardware and software-based investments up to 2017, that go some way to explaining some part of productivity growth anomalies. Moreover, they argue that, since 2017, it is reasonable to conclude that investments in AI systems may now be the primary contributor to the productivity shortfall.
This intuitive J-curve model also explains the famous observation by the economist Robert Solow, that “one can see the computer age everywhere but in the productivity statistics,” as there is an inherent lag in the measured productivity growth due to the mis-accounting for the output intangibles in the early going. Indeed, Brynjolfsson notes that this effect is typical for GPTs, noting that “with general purpose technologies like electricity it took 20 to 30 years to work its way through the J curve and start having the upswing.” He anticipates that the comparable J-curve for AI will, however, be significantly compressed, “The best evidence suggests that everything’s happening much more quickly. This time, with AI, we’ve seen returns sometimes just in a matter of a few months. For the economy, more broadly, I do think that within five years, we’ll be seeing much bigger gains.” In fact, based on his recent analysis of employment and wage data (more on this later), he argues that “we may already be beginning to turn the corner on significant productivity gains…I wouldn’t be surprised if the next year we saw above-trend productivity growth starting to happen, and by the end of five years, it would be way above that.”
One of the most compelling findings he and his collaborators have made about these J-curves is that cost leadership strategies worsen the J curve initial declines, whereas market expansion that targets new markets and new customers will attenuate these declines: a profoundly important doctrine that should inform all corporate business and national economic policies and strategies.
Not So Fast! There’s More Mismeasurement…
Beyond the industrial J-curve dynamics that result from the internal accumulation of intangible (mismeasured) benefits inside companies and economies, Brynjolfsson identifies a second type of mismeasurement that relates to the inability to quantify the true value of digital goods, compared to the physical goods-based world for which the classic Gross Domestic Product (GDP) calculations were designed. He argues that digital goods are often provided at zero direct cost to the user as they are subsidized by other means, “As the economy becomes more digital, we’re all getting more and more free digital goods; we’re getting better quality. The average American spends about eight and a half hours a day looking at screens of various sizes as they consume that content, most of it they don’t pay for. The thing about GDP is, while it measures all the things that are bought and sold in the economy…if something has zero price, it has zero weight in GDP, so we’re missing a lot of the value.”
He and his team have quantified this second measurement gap by creating a modified GDP standard they have named GDP-B where “the B stands for benefits, measuring not what you pay for something, but the benefits you get. Official GDP is very good at measuring tons of steel cars, grain etc., but historically, we just didn’t get free goods.” He gives a couple of examples to illustrate the point, “with Wikipedia, maybe I’d be willing to pay $10 or $50 a month, even if I pay zero. That’s the benefit I’m getting.” And in their study of the value users attributed to different digital applications such as Snapchat, WhatsApp and Facebook from 2003 to 2017, as well as to digital hardware such as smartphone cameras which are also subject to undervaluation in economic accounting, it was found that there was significant missing growth from GDP calculations (on the order of a percentage point or two) due to these effects. Moreover, Brynjolfsson argues that the rise of AI services will exacerbate this effect: “For instance, take the case of chat bots like ChatGPT, that’s worth about $100 billion in terms of consumer surplus, in terms of the value that people are getting, even if most people pay zero for it because they’re using the free service. And just to be clear, even when they pay $20 for it, they are almost surely getting more than $20 worth of value from it because they’re willing to pay more. So, you can go through all the goods in the economy and see how much consumer surplus there is versus how much they’re paying. And what we find is that there’s a big mismatch.” The net effect is that, since national productivity is calculated by GDP divided by labor and capital expended, this economic metric will almost certainly be significantly underestimated in the AI era, compounding the J-curve dynamics described above.
What About Me? The Impact on Jobs and Wages
But Brynjolfson’s story isn’t quite complete; the most urgent debate currently is on the human impact of AI on employment and the incomes we can expect in the near- and longer-term future. In a paper entitled “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”—continuing the recent animal-themed AI analogies of Emily Bender (stochastic parrots), Yann LeCun (cats) and Phil Tetlock (foxes and hedgehogs)—Brynjolfsson and collaborators use a wealth of payroll data from ADP to identify the following early portents of the coming critical shifts:
- There are significant declines in early-career employment in occupations that are most exposed to AI, such as software development and customer service. Importantly, no such declines are seen for more experienced workers in same roles.
- Overall employment continues to grow, but young worker employment has been stagnant since 2022, with non-AI-exposed young workers experiencing comparable job growth to older workers (6-9 percent) that is offset by a 6 percent decline in employment in AI-exposed jobs for 22 to 25-year-olds.
- In general, the employment declines are seen in jobs where work can be automated not just augmented.
- These changes correlate with the rise of LLMs in 2022, rather than factors such as remote working or computer-related occupations or out-sourcing prevalent areas.
- These employment trends were independent of college education-level associated with the role, although non-college-level roles saw diminished employment impact in AI-impacted roles up to the age of 40.
- There was no discernable impact on wages in any of the analyses.
Brynjolfsson proffers an intuitive explanation for these age-related effects, namely that younger workers rely more on codified knowledge, whereas older workers rely on tacit or embodied knowledge they have accumulated as on-the-job expertise and heuristics (“tips and tricks”). Consequently, since AI is better able to automate roles that rely on codified knowledge, as discussed above, there is a larger (and negative) impact on younger workers in AI-impacted roles. The reverse is true for older workers in any role.
Augmentation Over Automation
Bryjolfsson’s observations are consistent with the prior finding that technologies that automate inexpert tasks may decrease employment levels, as the remainder of the role is more “expert” and therefore is applicable to fewer people; conversely, technology that replaces expert tasks increases employment by making the role accessible to more workers. Notably, the same study revealed that wages should be impacted—going up in the former case (due to fewer qualified workers creating enhanced market demand) and down in the latter case (due to the supply of more qualified workers). Importantly, these wage effects are not yet visible in Brynjolfsson’s study, but he points out that there is precedent for a rise in employment and real wages—the internet or IT revolution led to a rise in both after an initial adjustment period.
Perhaps the most critical point of the study is the fact that augmentation once again is the preferred state. In fact, Brynjolfsson observes that throughout history “augmentation is the norm. I think a lot of people have this intuition that you need to replace work in order for productivity to grow. Nothing could be further from the truth; right now, the value of an hour of human labor is maybe 50 times more than it was two or three hundred years ago, because humans have all these tools that can allow us to do new things.” He also points out that most of the value that the economy has created comes from new goods and services not automation of the production of existing goods, citing the much-used statistic that 60 percent of people are now employed in occupations that did not exist in 1940.
He also argues that the set of tasks that can be performed by humans augmented by machines is undeniably greater than can be performed by either humans or machines working in isolation; his prior work on the suitability of machine learning found many occupations in which machines could contribute some tasks, but zero occupations out of 950 in which machine learning could do 100 percent of the necessary tasks. So, augmentation is the rule, and complete automation is the exception to that rule. This finding is consistent with that observed in Phil Tetlock’s experiments with LLMs interacting with people in prediction tasks; the human-machine interaction was found to be beneficial for all participants independent of the skill of the LLM. Similarly, a recent study of the ability of LLMs to perform economically valuable tasks, found there was still a measurable gap between human and machine capabilities, but when humans and LLMs worked in concert, task completion could be expedited and the cost reduced by more than 50 percent.
Mind the Trap
This would seem to be very positive news that should serve to ground the rather nihilistic or baseless discussions of human work in an AI-enabled future. But, there is a trap that must be avoided, one that Brynjolfsson calls the “Turing Trap,” after the well-known test of intelligence contemplated by Alan Turing: “The Turing test is the idea that the ultimate measure of AI is to make a machine that perfectly imitates a human so that you can’t tell which is the human and which is the machine…And I remember when I first read that, I thought, ‘Oh, that’s amazing. That’s a clear test; if a machine can do that, it’s a test of intelligence. Now, I don’t think it’s a great test and, worse, I think it’s a terrible goal, because if you make a machine that perfectly imitates a human, what does that do to the value of human labor? It drives it down, because [the machine] becomes a perfect substitute. And the more you have a substitute, the more we could have a fall in human labor. And the lower wages [will be] and that will lead to less economic power and less political power for most people.”
He goes on to explain that there may be some rebalancing of power to offset this, as some early studies have shown that less skilled workers may benefit disproportionately from the use of AI. This can be understood by AI capturing some of the expertise from the best workers and “recycling” it to the less skilled workers; conversely, the very best workers experience essentially zero gains as they were getting back knowledge they already possessed in their domain. This is, in effect, a version of the phenomenon which David Eagleman calls the “intelligence echo illusion”—that (expert) human knowledge is being discovered, summarized and echoed back to other humans by LLMs, making this knowledge more widely accessible. Similarly, recent studies of the role of LLMs in assisting coding tasks have shown negative impact on expert coders (due to the greater time spent correcting errors in LLM-generated code), in contrast to the widespread reports of productivity gains for inexperienced coders.
However, Brynjolfsson remains concerned that “AI will not end up being such a leveling force,” even if there are such benefits for lower-skilled workers, due to a manifest inequity in to whom the financial benefits accrue, “as AI gets more and more powerful, I think more and more of the returns are going to be captured by people who own and control the AI, who own and control capital, and overall, the value of labor could fall. Zoe Hitzig and I have a paper that [outlines how this] could lead to a world where almost all the decision-making power, almost all the economic value, and ultimately all the political value, is concentrated in a small number of people. And that could be…. very disempowering for the rest of us.”
The critical point is that there is a tendency towards a greater concentration of technological and economic power that in turn leads to a greater concentration of political power that risks trapping a powerless majority into an unhappy equilibrium: the Turing Trap. Brynjolfsson explains that the trap is effectively set when technologies automate human labor, as this tends to reduce the marginal value of workers’ contributions, and more of the gains go to the owners, entrepreneurs, inventors and architects of the new systems. In contrast, when technologies augment human capabilities, more of the gains go to human workers. Moreover, he explains that the trap has a seductive appeal: “The risks of the Turing Trap are amplified because three key groups of people—technologists, businesspeople and policy-makers—each find it alluring: Technologists have sought to replicate human intelligence for decades to address the recurring challenge of what computers could not do; businesses seek to reduce the largest line-item on their financial statements—human labor costs, in part because many investors prefer “scalable” business models that can grow without hiring more people; and last, governments increasingly set tax and regulatory policies that favor capital investment and facile, expedited growth over labor investment and societal good.”
But Change Is Good
So, where does Brynjolfsson come out on the pros and the cons of AI—the risks of falling into the trap versus the potential for significant productivity gains, combined with employment and even wage gains? In short, he is confident that, by understanding the essential factors and the historical precedents that should inform our view of likely futures, we can successfully navigate these tricky waters as we have many times in technological history, “there’s no question that technology has always been destroying jobs and it’s always been creating jobs. There’s this churn, and although it’s a common instinct, the worst thing you can do is try to freeze in place the existing jobs. You have to lean into the dynamism and into the new jobs, even as the old ones get replaced. Now, the exact set of jobs that appear and disappear changes with each technology. Sometimes it’s the less skilled workers that are hurt and the more skilled workers that benefit. And that happened during much of the late 20th century. Economists call it skill-biased technical change.” But, in this case, it seems clear from his work that the skills in question that will be impacted are the more codified knowledge-based ones, rather than roles that rely on embodied or physical expertise.
He is also quick to point out that the biggest gains will not be from enhancing individual tasks, but rather from reimaging entire workflows, processes and organizations, using Amazon as the example, “back in 1995 Jeff Bezos was looking at how bookstores could use technology to really automate. Imagine if he’d gone into a bookstore and said, there’s a cashier—we could automate that cashier. We’re going to put a robot where the human is…[that] wouldn’t have been very transformative. Maybe you would have taken some labor costs out. But unfortunately, that is the way most people are using most technologies; they look at what they’re already doing and try to automate one little bit, replace one worker, cut costs. The bigger transformation is looking at the whole process and thinking, let’s just reinvent this. We don’t need a physical bookstore. We can have things delivered. We can have people buy things on a browser, in a totally different world.” This observation is backed by recent studies that show that end-to-end workflow re-imagining results in a 60 percent premium, relative to the impact of each element being AI-enabled in isolation.
Brynjolfsson takes the definitive position that “the solution is not to slow down technology but rather eliminate or reverse incentives for automation over augmentation…as automating labor unlocks less value that augmenting it to create something new.”
And although the move towards TAI may result in an unprecedented rate of change, he calibrates this by reference to the internet age, noting that “in 2025 only about 18 percent of retail is online now. So, it does take a long time for this to diffuse and spread. Nobody would say, I don’t think that internet is important, or that e-commerce doesn’t work. It’s just that it takes a while for it to really be worked through. I think AI is going to be a lot faster, but we still should be realistic about the timelines.”
Parting Thoughts
Our conversation concludes with a tour de table of how he sees the future across a number of critical dimensions—the protections we need to put in place; how to measure progress towards augmentation and; the burning question of the moment: whether we are currently in an AI-driven financial bubble.
On the missing protections: Brynjolfsson says, “we are unquestionably going into unprecedented time. We’ve never created an entity that’s more intelligent than us. So, I think we want to tread pretty carefully about this, and we should take it seriously. We are not spending nearly enough energy, time, or government attention on safety research. How do we make these machines safe? We should probably be increasing the efforts on safety research 10-fold, [so] we can have machines that are powerful to help cure cancer and solve these other problems, which are also controllable. Our understanding of economics, of our institutions, of our organizations, is lagging way behind. We’re not spending billions or hundreds of billions of dollars on that…that gap between what the technology can do and our economic understanding is where most of the challenges and also most of the opportunities lie for the coming decade.”
He points out that for every dollar spent on technology, companies may need to spend nine dollars on intangible human capital, which is similar to the 10:20:70 rule proposed by Boston Consulting Group, in which 10 percent of the spend is on the AI algorithmic development, 20 percent on the supporting AI infrastructure and 70 percent on the retraining, reimagining and reorganizing of people and processes.
Brynjolfsson also sees important lessons that can be learned from the free trade initiatives of the last decades in that the benefits that arose from focusing on core capabilities and sourcing the remainder from elsewhere (that specialized in other domains) did not guarantee that every person in every country came out ahead. In reality, the economic benefits were not redistributed, resulting in the recent populist backlash and the rise of vehicles such as tariffs as a mechanism to attempt to redress this imbalance, albeit a distinctly faulty one.
So, in short, Brynjolfsson argues that, this time around, incentives and policies need to be put in place to support the requisite investment in people and prevent manifest societal disruption by the pervasive deployment of AI.
Centaur metrics are more sensible: Brynjolfsson argues that “one of the biggest weaknesses of systems right now, is that we can’t really interrogate them…the AI researchers have often set up the wrong benchmarks…too many of these benchmarks are black box metrics. Most prominent evaluations of machine learning systems consider the systems in isolation from humans, leading to easily saturated benchmarks, hard-to-formalize human-centered desiderata and a bias of technological development toward human replacement instead of human augmentation. So, Andy Haupt and I have created a set of “centaur benchmarks”—you know, half human, half machine, and we are hoping more technologists will focus on having the machine and human work together.” Brynjolfsson believes that only by using these types of benchmarks will we maximize value by measuring progress towards human labor augmentation rather than simple automation.
A bubble or not a bubble, that is the question: Brynjolfsson conjectures that “we could easily be in a bubble. We’ve been in bubbles before. [But] I can tell you that the fundamentals are pretty strong, that the technology really is working, and there’s a path to even more powerful technology. What I’m less sure of is how well those companies will be able to capture that value. One of the things I think a lot of investors underestimate is that especially when you have big productivity gains, it doesn’t necessarily lead to big profits. You can have a company or an industry become twice or 10 times more productive, and paradoxically, that can lead to falling profits because it leads to over-capacity and falling prices, and ends up creating a lot of value for consumers, but not necessarily for stockholders. It’s a reasonable hypothesis that an awful lot of infrastructure will have been built out that has tremendous intangible value that can’t be realized because essentially, there’s a glut relative to the usage. Hopefully demand will grow equally or even more, and then the investors will get their payoff. But there’s no guarantee that will happen. That said I’m probably closer to the bullish end of the spectrum in terms of seeing on the ground how these technologies can solve more and more problems. And I think there’s a lot of pent-up demand, in medicine, in finance, in retailing and manufacturing, in every industry, to have more intelligence to solve more problems.” In essence, he agrees with the prior observation in this series that AI systems are solving for the digital information overload we created in the computer age, by helping humans overcome their inability to deal with this information glut.
He also believes that investments in these cognitive tools may help us create the world models required to describe embodied, physical systems, by “recursive self-improvement that allows us to understand the world better than we could just with our own brains,” essentially arguing that there will be a bootstrapping effect that allows continuous value realization from each generation of AI infrastructure deployment that will mitigate the risk of a major “bubble bursting” revaluation.
Despite the cautionary tone of these latter remarks, there is little doubt that Brynjolfsson is indeed “bullish” about the future, which he sees as one in which we will use new human-augmentation benchmarks to successfully navigate the path to an unprecedented AI-augmented productivity regime that is defined by growth in both wages and employment, without excessive centralization of wealth and power, allowing humankind as a whole to benefit from the ultimate GPT: GPTs and their evolutionary offspring.
It is tempting to bet against this idealistic vision, as the alignment required of all the relevant factors and players would itself be unprecedented in a competitive free market environment. But, as Brynjolfsson likes to point out, he bet Robert Gordon that productivity growth would return to prior levels (1.8 percent per year, on average) between 2020 and 2030, driven by AI, a bet that that Brynjolfsson reports that Gordon concedes that he currently looks like winning. There is an Icelandic expression, “þetta reddast,” meaning “everything will work out” that is deeply embedded in that culture, and it is perhaps no surprise that this same attitude seems to inform Brynjolfsson’s outlook. To use another popular Icelandic expression, if indeed this comes to pass as he foresees, that would truly be “the raisin at the end of the hot dog”.