Generative AI at the Crossroads: Light Bulb, Dynamo, or Microscope?

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Generative AI at the Crossroads: Light Bulb, Dynamo, or Microscope? Martin Neil Baily, David M. Byrne, Aidan T. Kane, Paul E. Soto 2025-053 Please cite this paper as: Baily, Martin Neil, David M. Byrne, Aidan T. Kane, and Paul E. Soto (2025). “Generative AI at the Crossroads: Light Bulb, Dynamo, or Microscope?,” Finance and Economics Discussion Series 2025-053. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2025.053. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Generative AI at the Crossroads: ∗ Light Bulb, Dynamo, or Microscope? Martin Neil Baily David M. Byrne Aidan T. Kane Paul E. Soto June 27, 2025 Abstract With the advent of generative AI (genAI), the potential scope of artificial intelligence has increased dramatically, but the future effect of genAI on productivity remains uncertain. The effect of the technology on the innovation process is a crucial open question. Some inventions, such as the light bulb, temporarily raise productivity growth as adoption spreads, but the effect fades when the market is saturated; that is, the level of output per hour is permanently higher but the growth rate is not. In contrast, two types of technologies stand out as having longer-lived effects on productivity growth. First, there are technologies known as general-purpose technologies (GPTs). GPTs (1) are widely adopted, (2) spur abundant knock-on innovations (new goods and services, process efficiencies, and business reorganization), and (3) show continual improvement, refreshing this innovation cycle; the electric dynamo is an example. Second, there are inventions of methods of invention (IMIs). IMIs increase the efficiency of the research and development process via improvements to observation, analysis, communication, or organization; the compound microscope is an example. We show that GenAI has the characteristics of both a GPT and an IMI—an encouraging sign that genAI will raise the level of productivity. Even so, genAI’s contribution to productivity growth will depend on the speed with which that level is attained and, historically, the process for integrating revolutionary technologies into the economy is a protracted one. ∗Authors are listed in alphabetical order, not in order of relative contribution. Baily and Kane are at the Brookings Institution (mbaily@brookings.edu and akane@brookings.edu). Byrne and Soto are at the Federal Reserve Board of Governors (david.m.byrne@frb.gov and paul.e.soto@frb.gov). The views expressed here are not represented to be the views of the staff or trustees of The Brookings Institution nor of the Federal Reserve. The authors are grateful to Michael Chui, Leland Crane, Avi Goldfarb, Bob Gordon, Shane Greenstein, Anton Korinek, James Manyika, Sid Srinivasan, Scott Stern, and Bill Whyman for helpful conversations. 1

1 Introduction In early 2023, OpenAI grabbed the world’s attention with ChatGPT, the first of several recently released “generative AI” (genAI) programs that use a computer model of human discourse to respond to natural-language questions. The scope of AI has expanded dramatically with the advent of genAI, including to tasks previously seen as quintessentially human, such as competition-level mathematics (fig. 1 on the following page). Indeed, more and more challenging benchmark tests have been needed to assess progress as genAI has matched human performance on one task after another.1 In an encouraging sign, field test evidence of productivity improvements from genAI in practical applications has also emerged, including for writing, computer programming, and responding to call center inquiries (table 1 on page 4).2 Although some companies do credit genAI with improvement to their bottom line, it remains to be seen whether widely-used cost-effective business applications will follow from these successful field tests.3 Web searches for AI and downloads of the ChatGPT app have soared, sparking intense competition for leadership in genAI (section 1 on page 5) andthecomputationalintensityoftraininggenAImodelsandprocessinguser requests has led to a massive increase in data center construction and spending on AI-related semiconductor chips (fig. 3 on page 6). While optimists see potential for genAI to spur an IT-fueled productivity boom compara- 1For more on benchmarks, see Maslej et al. (2024). The external validity of such benchmarks—that is, how much they tell us about performance on practical tasks seeminglyrelatedtothetests—isamatterofsomedebate(Liaoetal.,2021). Newbenchmarks, such as the ARC-AGI and the Graduate-Level Google-Proof Q&A (GPQA) benchmarks, have been introduced to better evaluate advanced capabilities. Recent models, especially “reasoning” models such as o3 from OpenAI and others discussed in section 3.3.1 on page 21, have performed strongly on these more demanding tasks. Haupt and Brynjolfsson (2025) argue that benchmarks that measure how well AI and humans can jointly perform tasks are needed to shed light on practical AI use. 2Brynjolfssonetal.(2023)findthatcallcenteroperatorsbecame14%moreproductive when they used the technology; Peng, Kalliamvakou, Cihon and Demirer (2023) find that accesstoGitHubCopilotenabledprogrammerstocompletetasks56%faster;andNoyand Zhang (2023) find that writers who used ChatGPT worked more quickly and produced higher-quality outputs. 3McKinsey (2025) reports that a large share of their survey respondents—primarily large corporations—credit AI with cost reductions in some business functions, but 80% “aren’t seeing a tangible impact on enterprise-level [earnings before interest and taxes] from their use of genAI.” 2

Figure 1: AI Benchmark Performance Note: The“humanbaseline”conceptusedvariesbytask. Formorechallengingtasks,thebaselinetends toreflectexpert-levelperformance. Source: Reproduced with permission from the 2024 AI Index Report, Stanford Institute for HumancenteredArtificialIntelligence. ble to the late 1990s and early 2000s, more downbeat observers see claims about its capabilities as overstated and highlight potential headwinds, such as regulations to guard against unintended harms and the colossal energy requirements for training and running genAI systems.4 It is too early to adopt either view with confidence. Whether practical genAI applications will be consequential enough to raise aggregate productivity growth remains to be seen. GenAI may be no more important than previous innovations in information technology (IT) already reflected in the historical trend, including its predecessors in the field of AI.5 With only “green shoots” of quantitative evidence in hand that genAI will raise productivity, we frame the prospective effect of genAI in qualitative terms: We ask what class of innovation it may be. Some labor-saving innovations, such as the light bulb, temporarily raise productivity growth as 4For an optimistic view, see Kurzweil (2024). For a more cautious perspective see Narayanan and Kapoor (2024). 5Moreover, the present era of modest productivity growth despite mature machine learning, cloud computing, and smartphones should temper expectations for another IT boom. 3

Table 1: Selected GenAI Productivity Field Studies Study Task Results Noy and Zhang (2023) Writing ChatGPT speeds, improves writing. Writers shift from drafting to idea generation and editing. Brynjolfsson, Li, Customer Moreissuesresolvedusingconver- Ramsey (2023) service sational assistant. Dell’Aqua et al. (2023) Various GPT-4 increases task completion, speed, and quality. Peng et al. (2023) Coding Using GitHub Copilot, programmers complete tasks faster Cui et al. (2024) Coding GitHub Copilot raises task completion. adoption spreads, but the effect fades when the market is saturated; that is, the level of output per hour is permanently higher but the growth rate is not. Other widely used technologies, such as the electric dynamo, spur knock-on innovations—new products, process improvements, and business reorganization—and continually refresh this adoption cycle through ongoing improvementinthecoretechnology(David,1990). Theboosttoproductivity growth from these general-purpose technologies (GPTs) may last longer. Yet other inventions, such as the compound microscope, increase the efficiency of the research and development process; these “inventions of methods of invention” (IMIs) yield a sustained increase in productivity growth by lowering the cost of research and development. We first define “generative AI,” then consider the evidence that genAI is a GPT, reviewing indicators for the scope of diffusion, the extent of knockon innovations, and signs of ongoing progress in the core technology. To assess its status as an IMI, we discuss evidence that it increases the efficiency of observation, analysis, communication, and organization in research and review several indicators (patents, earnings calls, and query topics). For both questions, we reference case studies for the financial, health care, and information sectors and for the electricity generation industry (Baily and Kane, 2025a,b; Kane and Baily, 2025a,b). We conclude there is substantial evidence that genAI is both a GPT and an IMI, an encouraging sign for higher productivity in the future. 4

Figure 2: Indicators of Interest in GenAI (a) GenAI Mobile App Downloads (b) Web Searches for AI Note:AppsareChatGPT,Claude,DeepSeek,andPerplex- Note: IncludesrelatedtermsinGoogle’s“AI”topic. ity. IncludesAndroidandiOS.Androiddownloadinforma- Source: GoogleTrends. tionnotavailableforChina. Doesnotaccountforaccess viaapplicationprograminterface. Source: appfigures. Thereisasubstantialliteratureonthequestionofwhethermachinelearning, which preceded genAI, may be a GPT (Cockburn, Henderson and Stern, 2018; Trajtenberg, 2018; Bresnahan, 2019; Goldfarb, Taska and Teodoridis, 2023; Bresnahan, 2024) and Cockburn, Henderson and Stern (2018) discuss the possibility that machine learning is an IMI.6 There is little work focused specifically on genAI. Eloundou, Manning, Mishkin and Rock (2024), a prominent exception, considers the prospects for genAI to be a GPT; relative to that work, we draw on a broader set of indicators and consider evidence for whether genAI is both a GPT and an IMI. Our focus on characteristics of the technology itself and its integration into business processes complements the large literature on the labor market impact of genAI.7 Our qualitative assessment of genAI also serves as a primer to inform discussion of AI in the context other literatures, such as creative destruction and endogenous growth (Akcigit and Van Reenen, 2023). 6Our focus is primarily on the U.S. economy; Filippucci et al. (2024) take a global perspective. 7SeeAgrawal,GansandGoldfarb(2023);AcemogluandRestrepo(2020);Brynjolfsson, Li and Raymond (2023); Eloundou, Manning, Mishkin and Rock (2024), among others. 5

Figure 3: Indicators of U.S. AI-Related Investment (a) Data Center Construction (b) Application-Specific Chips Note: Nominalvalueofconstructionputinplace. Note: Application-specific chips include GPUs, Source: CensusBureau. TPUs,andASICsforotherapplications. Source: SemiconductorIndustryAssociation. 2 What is Generative AI? “Artificial intelligence” (AI) is an umbrella term encompassing a variety of algorithms deployed on computers to mimic human thought, communication, and choices, such as machine learning, computer vision, and generative models. (See Appendix A for a discussion of several influential definitions of AI.) AI systems achieve these objectives by constructing and calibrating mathematical models of complex patterns found in training data. The most widely known implementations of genAI use a computer model of the human discourse found on the internet to respond to natural-language prompts (questions or directives), though genAI systems take other forms as well as we discuss below. We narrow our focus to genAI for several reasons. First, the broad and varied use of “AI” makes a coherent discussion of its effects on productivity difficult. Second, the productivity impact of genAI is largely in the future, in contrast to AI types already in use for which the effects on productivity are, in principle, an empirical question. Third, the human-like behavior of generative models has made concerns about the disruptive effects of AI particularly salient. Although systems which respond to prompts with natural language text were a part of the AI field from its inception, early models were not grounded 6

in a model of human language. (For a short history of the field, see Appendix B.) For example, rudimentary chatbots, such as ELIZA the psychotherapist, text autocompletion, and “expert systems,” such as MYCIN for infection diagnosis, appeared in the 1960s and 1970s. These systems had matured by the 2010s, when sophisticated voice-driven chatbots like Alexa and Siri had been introduced, news outlets were auto-generating routine stories, and IBM’s Watson, famous for beating humans at Jeopardy in 2011, was repurposed to provide advice on a host of topics. These were symbolic, rules-based systems, albeit with an element of randomness in their output. Richer generative models appeared after the development of large language models (LLMs). LLMs, which represent the meanings of words and their relationships by locations in a high-dimensional space, emerged in the 2010s. Word2Vec, most notably, encoded words as vectors of numerical values, and while the values do not correspond to specific, interpretable characteristics, they capture semantic relationships in an abstract space (Mikolov et al., 2013b). For example, while humans would represent a dress by its color, size, and hem length, say, the characteristics chosen by Word2Vec would not have a readily apparent interpretation (Bajari and Chernozhukov, 2018). Aided by advances in computational power and big data processing techniques, genAI developers pushed the field forward and achieved a breakthrough with the Transformer architecture in 2017 (Vaswani et al., 2017). (The technical features of the Transformer are discussed in a box below.) This model allows for a richer representation of word meaning in its encoding by accounting for context. Many GenAI models use LLMs to encode the tokens (words, phrases, or parts of words) in the input in this fashion. After embedding the input text into locations within a high-dimensional space of abstract characteristics, they draw on the information embedded nearby to guide the prediction of the next token, weaving the output into a relevant natural-language response. Becausethesemodelsarecreatedasneuralnetworks, whichareextremely flexible, theydifferfrompreviousmodels, likeIBM’sWatson, inthattheyare not symbolic. That is, they do not have a predetermined logical structure. The effect of this added flexibility and the richness enabled by their massive scale is that the range of possible generated content is more open-ended than earlier systems. For example, earlier symbolic attempts were capable of producing formulaic news stories about a firm’s quarterly earnings or the outcome of a baseball game, but ChatGPT can provide a convincing nuanced response to prompts such as “write a short story in the style of Edgar Allan 7

Poe.” Importantly, genAI can support a wider array of applications than natural language tools for learning and creativity. Broadly stated, genAI systems produce contextually appropriate artifacts using an open-ended stochastic process that draws from patterns in a dataset. The artifacts may be a variety of things other than text, including computer code, images, music, chemical structures, game environments, mathematical proofs, or dance moves, to name a few. And, while the chat window user interface makes genAI accessible, it is not found in all applications. For example, in generative adversarial networks (GANs), neural networks interact with each other, not humans, to generate output. Landmark AI Models: The Transformer Thetransformerarchitecture,introducedbyVaswanietal.(2017),was a game changer in AI, particularly as the engine behind genAI models. Its key innovation, the “attention mechanism,” steers models to focus selectively on relevant parts of the prompt, enabling more efficient and accurate processing of language. This breakthrough has powered major advancements in natural language understanding, translation, and generation, forming the backbone of today’s most advanced genAI systems. Transformers process input data through a series of layers (steps), each consisting of an attention mechanism followed by a multilayer perceptron (MLP, defined below), proceeding as follows. First, a representation of the prompt (input text) suitable for analysis by the model is created. Specifically, the prompt is broken into tokens(smaller pieces whichmaybe phrases, words, orparts ofwords). The tokens are converted into embeddings (numerical vector representations) which encode the semantic and syntactic meaning of each token. Loosely speaking, for each token, the closest of the other tokens, as measured by the distance between their embeddings, are the ones most important to understanding its meaning. Second, the attention mechanism processes the matrix of token embeddings using three large matrices called the “query,” the “key,” and the “value.” For each token in the input, the query is compared to the keys of all tokens to compute attention scores, which are used to form 8

The Transformer (continued) a weighted average of values. This step allows each token’s representation to incorporate information from other tokens in the prompt based on their contextual relevance. Third, the data passes through an MLP, a type of neural network. While the attention mechanism focuses on pairwise interactions between tokens, the MLP applies nonlinear functions (in contrast to the linear attention mechanism) in refining the token representations. This sequence—of computing the attention mechanism followed by the MLP—is repeated multiple times depending on how many layers are in the model (for example, the Llama-3 model has 32 layers), enabling the model to capture increasingly abstract features of the input text. The performance gains from scaling of this system through increasing the size of these matrices—along with larger training datasets and improvements in hardware and processing algorithms—underpins the rising ability to handle complex language tasks. 3 Is GenAI a General Purpose Technology? While the evidence for genAI-driven productivity in specific tasks is intriguing (table 1 on page 4), to look ahead to its future productivity impact, we would like to know (1) if genAI will be widely adopted, (2) the extent of related innovations, and (3) whether genAI will continue to improve. That is, we would like to know if genAI is a general-purpose technology (GPT).8 Through widespread adoption, downstream complementary innovation, and sustained innovation in the core technology, GPTs have long-lasting effects on productivity. Examples of GPTs are shown in table 2 on the following page. As described by Lipsey, Carlaw and Bekar (2005), “big GPT shocks change almost everything in a society and revitalize the growth process by creating an agenda for the creation of new products, new processes, and 8Note that the “GPT” acronym in the names of OpenAI genAI models stands for “generativepre-trainedtransformer.” Wewilluse“GPT”tomean“general-purposetechnology” exclusively in this paper except when referring to OpenAI models. 9

Table 2: Examples of General Purpose Technologies Technology Initial Impact Domestication of plants 9000–8000 BCE Writing 3400–3200 BCE Iron 1200 BCE Waterwheel Early medieval period Three-masted sailing ship 15th century Printing 15th century Factory system Mid 18th century Steam engine Late 18th century Railway Mid 19th century Internal combustion engine Late 19th century Electricity Early 20th century Motor vehicle Early 20th century Mass production, continuous process factory Early 20th century Lean production Late 20th century Computer Late 20th century Internet Late 20th century Source: Adapted from Lipsey, Carlaw and Bekar (2005). new organizational forms.” For example, new products that followed the development of the (electronic) computer include office productivity software, ATM machines, and the routing equipment that directs traffic around the internet. New processes that followed the development of reliable electricity include the production of hydrogen by electrolysis, salvaging scrap steel using electric arc furnaces, and the fabrication process for semiconductor chips. And, a new organizational form that followed the introduction of the threemasted sailing ship was the joint-stock company, used to finance the voyages of large-scale trading firms. We briefly describe the three criteria for a technology to be a GPT, then examine in detail the evidence that genAI meets each one. Diffusion Themorewidespreadtheapplicationofatechnology,thegreater the potential impact on aggregate productivity (Hulten, 1978). That said, however widely adopted, the direct productivity effect of a single invention is bounded. Productivity growth will be higher during the adoption transition, 10

but will return to its underlying trend when diffusion is complete (Solow, 1956). To illustrate using the light bulb, once suitable filaments were developed and the inexpensive incandescent light bulb was available, the technologywasgraduallyadoptedinworkplaces, raisingproductivitythroughbetter visibility and lower risk of accident (Abdou, 1997). Lighting is necessary for nearly all human labor, so the potential effect on the level of productivity from the light bulb was noteworthy, but once the light bulb market was saturated, it delivered no further (direct) level effect and the increment to productivity growth present during the transition disappeared.9 Knock-on Innovation Technologies that spur further innovation can deliver a longer-lived impetus to productivity growth. The greater persistence of elevated growth is the result of a series of overlapping classical “light bulb” growth effects.10 The electric dynamo is an example. The dynamo uses electromagnetism to convert mechanical energy produced by a prime mover—a steam engine, say—to electromagnetic energy, which is then conveyed by wires and converted back to mechanical energy by a motor used to drive machinery in another location. Existing systems conveyed mechanical energy directly to machinery through a set of belts. The dynamo/wiring/motor system is more energy efficient than the belt system except in very simple arrangements, so the replacement of existing factory systems yielded productivity gains. In addition, the dynamo enabled a more flexible organization of production (David, 1990). The less centralized factory designs adopted by firmsinresponsearethemselvesaproductivity-enhancinginnovationspurred by the dynamo. Ongoing Core Innovation When a technology continues to improve over time, the new productivity level—fixed in the Solow (1956) model—becomes a moving target. Ongoing innovation translates into greater technical performance at a lower cost, a form of productivity gain. Moreover, the price typically follows the production cost downward spuring greater adoption. (Productivity gains are, in a sense, embedded in the capital (Kaldor, 1957).) Solid state electronics is an example. Relentless increases in the number of 9The light bulb was not solely a terminal innovation, of course. The surge in demand for electricity represented by light bulbs led to centralized power stations (David, 1990). 10Complementary innovations may increase productivity by raising the effectiveness of theGPTaswell,suchasraisingtheoperatingrateofcomputersbytheinventionofcloud computing. 11

transistors on each semiconductor chip has driven the price of computing lower, making it cost-effective both to embed electronics in a greater variety of devices (like inexpensive toys) and to enhance devices with more and more electronic capability (like smartphones). 3.1 Diffusion Although comprehensive measures of the diffusion of genAI, specifically, are limited, recent trends in the diffusion of AI, generally, may indicate the underlying influence of genAI. Surveys show AI adoption rising, particularly in large corporations where AI use is concentrated. Even so, a large majority of firms still don’t see an application for AI in their business. Analyses of the text of job descriptions suggest that AI can be used for a broad range of workplace tasks, indicating that the potential for diffusion among firms is high.11 At the same time, the share of job postings mentioning AI skills is modest, indicating that firms are taking a cautious approach to hiring workers to focus on AI use. Meanwhile, from the worker’s perspective, AI adoption seems widespread; surveys of individuals document that a large share of workers are already AI users. Adoption surveys Distilling a single message from the available surveys ofAIuseisdifficultatfirstglance. TheCensusBureau’sBusinessTrendsand Outlook Survey (BTOS) finds roughly 9% of firms use AI, while McKinsey reports that 72% of firms do so (fig. 4 on page 14). On closer inspection, these surveys are consistent with one another and reveal important nuances in the state of AI adoption with respect to firm size and business functions. Combining the results of these surveys points to far higher AI adoption for large firms than small ones.12 The BTOS is a representative sample of 200,000 U.S. firms, only a handful of which are large corporations (Bonney 11See Acemoglu et al. (2020), Brynjolfsson et al. (2018), Felten et al. (2019), Webb (2019), Eloundou et al. (2024). 12There may well be significant heterogeneity within small firms on this question. Because new firms, which are typically small, may begin life as digitally native firms, they may adopt AI more easily. The pace of business applications with a high-propensity of turningintobusinesseswithpayroll,reportedbytheCensusBureau,hasmovedupsignificantlyinthewakeofthepandemic. Notethatinformationonnewfirmswillappearwith a delay in the BTOS, for which the sample is drawn from the Census Business Register (Bayard et al., 2018). 12

et al., 2024). The McKinsey survey, in contrast, is a convenience sample with heavy representation from large corporations (McKinsey, 2024).13 The U.S. firm size distribution is highly skewed and large corporations have thousands of employees (Kondo et al., 2023). The threshold for the largest firm-size group in the BTOS is 250 employees, for which BTOS reports 14.9% adoption, andtheBTOSreportsthatforsmallestfirm-sizegroup, firmswithfewer than 5 employees, 9.7% were using AI in June, 2025. These surveys also suggest that AI use may be less prevalent in core business functions than in support functions. Differences in the definition of adoption between the surveys point to this conclusion. The BTOS survey asks, “In the last two weeks, did this business use Artificial Intelligence (AI) in producing goods or services? (Examples of AI: machine learning, natural language processing, virtual agents, voice recognition, etc.).” The McKinsey questionisratherlessrestrictive,askingrespondentsifthey“useAIinatleast one business function” That is, the BTOS asks about use in core business functions (“producing goods or services”), while McKinsey includes non-core function like “marketing and sales,” which is the function with greatest AI use in their survey. Evidence about adoption for genAI specifically is more limited. Only the McKinsey survey asks separately about genAI, finding that use among their (primarily large firm) respondents surged from roughly one-third in 2023 to two-thirds in 2024. Bick, Blandin and Deming (2024) survey workers (a representative sample of 18-64 year-olds in the United States) and find 40% of respondents used genAI. As posited in Crane, Green and Soto (2025), high-level managers may underestimate the extent to which their employees are using AI tools, or AI users may be highly concentrated in large firms (roughly half of the U.S. workforce) or AI-intensive industries. Consistent with the idea that AI users may be concentrated in AI-intensive industries, Sergeyuk et al. (2025) survey programmers and find 86% of them use AI. A recent trend in industry composition is suggestive as well. Decker and Haltiwanger (2024) identify a step up in the share of business establishments—reported by the Bureau of Labor Statistics in the Quarterly Census of Employment and Wages—in industries with a high share of employees in science, technology, engineering, and math 13Although McKinsey states that “participants representing the full range of regions, industries, company sizes, functional specialties, and tenures,” responses are not weighted by the relative prevalence of their characteristics in the population. We thank Michael Chui for confirming that this is a fair characterization of the survey. 13

fields. We speculate this employee composition may be paired with higher AI use. Figure 4: AI Use Over Time (a) Census BTOS (United States) (b) McKinsey (Global) Note: For BTOS, respondents were asked about AI use in producing goods or services during the past twoweeksandanticipatedinthenextsixmonths. ForMcKinsey,respondentswereaskedifthey“useAI inatleastonebusinessfunction” Source: CensusBureau,BusinessTrendsandOutlookSurvey;McKinsey,“TheStateofAIinEarly2024.” Job postings Job posting data from Lightcast (formerly Burning Glass) categorized using terms associated with AI by Acemoglu, Autor, Hazell and Restrepo (2020), extended with terms not prevalent at the time of their analysis (such as “genAI”), show the share of job postings currently related to AI to be roughly 4 percent, using a broad definition of AI including a cluster of related skills (fig. 5 on the following page). Importantly, that share moved up (from no more than 2 percent) around 2017, well before practical genAI applications became prevalent, consistent with earlier forms of AI driving the increase. AI-related job postings have moved up only modestly since then, suggesting that explicit use of genAI-related skills are not crucial to many jobs. In some sectors, most notably the information sector (shown in the graph), the share of job postings referencing AI is substantially higher. Case study evidence x The information sector has adopted genAI rapidly. Among U.S. programmers, 92%wereusingAItools(includinggenAI)asofJune, 2023(Shani 14

Figure 5: AI-Related Job Postings Note: Jobs classified using AI-related terms found in job descriptions, as described in Acemoglu et al. (2020). ListofAI-relatedtermsupdatedbyLightcast. Source: Lightcast. and GitHub Staff, 2023). Other occupations within the information sector use genAI in their work as well. Of graphic designers and illustrators, 69% used genAI in their work in 2023, employing tools like generative fill and large-scale text-to-image models such as DALL-E.14 In the U.s. healthcare system, adoption has been slow for IT generally and AI (Poon et al., 2006; Goldfarb et al., 2020). However, genAI may be an exception: A majority of physicians already used or were planning to use genAI in 2025 for generating chart summaries, creating discharge instructions, and an array of other tasks (AMA Augmented Intelligence Research, 2025). Radiology is an enlightening example of how genAI has changed AI useinhealthcare. While30%ofradiologistsalreadyusedAIasof2020(Allen et al., 2021), the emergence of genAI spread AI to all stages of the radiology workflow, including the health system (e.g. imaging need prediction, claims processing), clinicians (e.g. prior authorization, communication), technologists (e.g. patient tracking, report generation), and radiologists (e.g. image quality assessment, diagnosis) (Burnside et al., 2025). Notwithstanding these inroads, AI use faces substantial hurdles in healthcare. For example, Agarwal et al. (2023) document that combining human analysis with AI diagnostics can yield disappointing results—radiologists 14See“CreativeprosseegenerativeAIaspartoftheirfuture,”onblog.adobe.com,March 21, 2023. 15

have difficulty determining how much confidence to assign the AI output. And, although Sahni et al. (2023) conclude that machine learning has the potential to assist with an array of tasks including diagnosis, treatment choice, and managing records, they assess that for hospitals, AI adoption often cannot be justified on financial factors alone. GenAI is used for many tasks in finance as well. Companies can use genAI to lower the cost of creating client-specific portfolios (Joshi, 2025), improveexistingautomatedsystemsthatrespondtoclientrequests(McKinsey & Company, 2025), assist with regulatory compliance (Agarwal et al., 2024) and with loan underwriting (Wang, 2023). In electricity generation and distribution, firms are experimenting with genAI, with approximately 33% piloting genAI in their customer service operations.15 Some companies use genAI in their load forecasting processes, while others use the technology to simulate equipment degradation and inform predictive maintenance, areas where machine learning was already in use (Gao et al., 2024). 3.2 Knock-on Innovation In the course of adopting new technologies, firms retool, retrain, and reorganize to better exploit their productivity potential.16 This process involves knock-on innovation in products, production processes, and operations. In the case of genAI, knock-on product innovation includes the diverse array of user interface software for generative models and the emergence of more capable robots. Process innovation includes genAI-enhanced approaches to product design and production line operation. Organizational innovations include centralized data governance and optimization of supply chains and data centers. Bresnahan (2024) observed that early machine learning (pre-generative AI) adoption was concentrated in places where complementary innovation was less necessary, such as in firms that were highly digitized from their founding (digital natives). In such firms, adoption of AI was more straightforward, involving substitution of AI for existing IT capital or deploying AI to undertake tasks not previously part of operations. The legendary email 15See “A third of utilities have begun to pilot generative AI for customer service, other uses: report,” on www.utilitydrive.com, July 12, 2023. 16Bresnahan and Greenstein (1996) discuss this process in detail, which they refer to as “co-invention.” 16

from Jeff Bezos instructing internal teams at Amazon to exclusively use APIs to deliver data and functionality would have landed rather differently at a non-native company, for example.17 In 2018, nearly 10 years after internet giants had begun using machine learning at scale (e.g. Amazon’s random forest demand forecasting (2009) and Google’s Panda search algorithm (2011)), AI began to gain traction at other firms (fig. 4 on page 14).18 Digital natives will surely lead the charge for genAI as well. For other firms, the pace and success of knock-on innovation will be a key determinant of the scale and timing of productivity effects from genAI.19 3.2.1 Products User interfaces (UIs) provide a channel through which requests and responses can pass between the model and the user, whether a human user or another system, such as a vehicle or a robot. In the early days of genAI, users accessed genAI through their own Python programs or through websites such as the OpenAI Playground. A major shift occurred in November 2022, when OpenAI released ChatGPT, a conversation-style “chatbot” that made genAI interactions significantly more accessible to a broader audience. Since then, several new interfaces have emerged. In 2023, OpenAI introduced Custom GPTs, enabling users to create specialized LLMs for specific domains, such as LegalGPT for legal matters. In 2024, OpenAI announced integration of their ChatGPT model to Apple’s Siri voice assistant and Google launched NotebookLM , which made it easy to upload documents and transform them into interactive discussions. In addition, there are “copilots” that integrate AI into existing user workstreams, notably GitHub Copilot (computer programming) and Microsoft 365 Copilot (office productivity). 17Even at Amazon, continuous digital transformation is challenging, as Yegge (2011) relates in his account of the aftermath of the Bezos email. 18Bresnahan(2019)describestheuseofmachinelearningatAmazon,Facebook,Google, andNetflix,particularlyformatching(e.g. consumerstoproducts)andforuserinterfaces and notes “there is little application of [artificial intelligence technologies] outside the Internet Giants as of spring 2018.” Bughin and Van Zeebroeck (2018) notes that “only a fraction of companies—about 10 percent—have tried to diffuse AI across the enterprise, ...An additional quarter of companies have tested AI to a limited extent.” 19Theemergenceofcloudservicesmayhavecatalyzedmachinelearningadoptionoutside of digital natives and AI as a service, such as Azure OpenAI Service may play a similar role for genAI. 17

System interfaces allowhardwareandsoftwaresystemstoaccessthecore AI system. For example, Nvidia’s Isaac Software Development Kit (SDK) facilitates the integration of AI into robotics.20 Access to AI through SDK helps the robot with environmental integration problems, such as simultaneously tracking its location and mapping its environment (SLAM). Multimodal models, which take in inputs of different kinds (text, images, sensor readings) and can output instructions to the robot, such as the rotation and torque for a joint (Reed et al., 2022; Brohan et al., 2023) System interfaces also make possible the design of agentic AI systems, which collect information during operation, interpret context, make decisions and act in pursuit of goals autonomously (Park et al., 2023).21 Examples include AutoGen, from Microsoft, and CrewAI. For example, an airline reservation app using agentic AI would be capable of handling the complex steps of searching for available flights, proactively aligning the user’s travel options with their preferences, making the booking, adjusting to disruptions to the reservation process (e.g. sporadic price changes), and including special accommodations along the way (e.g. upgrading seats if available within the user’s preferences). 3.2.2 Production processes Product design is a use of genAI that will be self-evident to users of tools such as DALL-E to create whimsical images; powerful genAI tools are also capable of designing products of all kinds that meet technical and aesthetic specifications. Perhaps less obviously, the design process itself can be transformed through knock-on innovation. Saadi and Yang (2023) interviewed designers and observed: Rather than thinking about how to create several one-off designs, designers may consider how to create a system for design that would allow the design tool to generate a large number of valid 20Robot capabilities have advanced in tandem with AI progress. Theseus, a symbolic AI robot mouse that navigated mazes was developed in the 1950s (Wallace, 1952). Industrial robots were trained using machine learning beginning in the 1990s (Arinez et al., 2020; Soori et al., 2023). Even more sophisticated robots, which can learn from their environments, integrating sensor data, are available today, including civilian and military autonomous vehicles (Knight, 2016; Pierson and Gashler, 2017). 21For more detail on agents, see Wiesinger et al. (2024). For an array of agentic AI use cases, see Hall (2024). 18

outputs. This can involve setting the appropriate specifications, manufacturing methods, and product architecture early in the process to input into computational tools. Production line operation Serradilla et al. (2022) provides an overview of the use of deep learning, including genAI such as generative adverserial networks, to optimize line configuration, throughput, efficiency, and carbon footprint. Predictive maintenance using synthetic data and scenario simulation is another application for genAI in industry. Sai et al. (2024) provides examples of production optimization outside of manufacturing as well. 3.2.3 Organization Amongtheorganizationalinnovationsaretheimportanceofcross-functional teams with access to data that spans the enterprise, breaking down barriers between business units, optimization of supply chains, and reallocation of employees to de-emphasize repetitive writing tasks (Iansiti and Lakhani, 2020). 3.2.4 Case study evidence Task-specific user interface programs are already in use in the health care industry. For example, one hospital system used genAI to field a flood of questions during the COVID-19 pandemic (Wittbold et al., 2020). AI is also used to transcribe conversations and dictations to create clinical notes (Handa and Sorensen, 2023). Hornback et al. (2025) document how Fast Healthcare Interoperability Resources (FHIR), a protocol for the secure exchange of sensitive health information, has evolved to support emerging AI technologies. These efforts include the integration of BERT, a transformer model, for encoding unstructured text data (Peterson et al., 2020) and the development of a generative model, FHIR-GPT (Li et al., 2024). In finance, JP Morgan Chase adopted AI for contract review using software called COiN (Contract Intelligence) and reported saving 360,000 hours of costly lawyers and loan officers (Weiss, 2017). In the information sector, Peng et al. (2023) document that the use of GitHub Copilot markedly increased the productivity of programmers, particularly less experienced ones. And, data center networks have evolved to better support genAI and genAI has contributed to network optimization beyond the contributions of machine learn- 19

ing (Liu et al., 2024b). GenAI has proved useful in the energy sector as well. Choi et al. (2024) describe a custom interface to genAI designed to assist control room operators in balancing supply and demand on electrical grids. 3.3 Ongoing Core Innovation Since the introduction of the Transformer, genAI models have steadily pushed out the frontier of capability. At first, advances were largely driven by increasing the scale of the model, computational power used (“compute”, in the lingo of AI), and the size of the training dataset (fig. 6). AI scientists and engineers often focus on the “scaling laws” that describe the benchmark performance effects of increasing each of these inputs (Kaplan et al., 2020). More recently, tactics to use compute more efficiently and to refine models for specific applications have been a focus as well. Figure 6: Number of Parameters and Training Dataset Size Note: Only models with reported parameter size and training dataset size are included in the dataset. Forinstance,GPT-4isexcludedasitsparametersizeisnotknown. Source: Epoch(2024)withmajorprocessingbyOurWorldinData. Importantly, economic performance (productivity) only rises when more can be accomplished while holding input costs fixed. In other words, we are looking for shifts in scaling law functions, not movement along the curves. 20

Accordingly, we focus below on (a) how innovations in model architecture (the algorithms used for distilling information from data) raise genAI model capabilities without raising training costs, (b) how hardware innovations lower the cost of computation, and (c) how richer datasets (with more information per token) can be brought to bear on training. 3.3.1 Model Development Since the introduction of the Transformer, model development has progressed at a blistering pace, including improvements attributable to ramping upmodelscale, trainingdatasetsize, andtheamountofcomputeemployed— the scaling laws discussed above—and the introduction of novel model concepts and techniques to increase the efficiency of model training. An important catalyst in this process has been a rise in open-source models, which has accelerated experimentation across text-generation models, and since the second half of 2024, multi-modal models (fig. 7). Figure 7: Open-Source AI Models Source: HuggingFace. 21

ThetrainingofgenAImodels(optimalcalibrationofitsparameters)takes place in two stages: pre-training and fine-tuning. Pre-training produces a broadly-applicable “foundation model”; fine-tuning refines the foundation modelforaspecificapplication. Thetrainedmodelisthenusedininference— responding to user requests. Efforts early in the wave of genAI improvement thatfollowedtheTransformerfocusedonpre-training, buttheescalatingcost of making progress in that stage has led researchers to explore improvements in fine-tuning and inference, as discussed below.22 Inadditiontotrainingandinferenceinnovations,advancesinperformance have come from novel model concepts. Mamba, introduced in 2023, achieved subquadratic-time sequence modelling by avoiding the pairwise comparison among tokens used in the attention mechanism, meaning that as input texts lengthen, the computational burden increases at a slower pace than the Transformer (Gu and Dao, 2023).23 Small-scale models, with lower computational requirements, have been a focus for some applications as well, such as personal devices and lower resource settings, making LLMs more accessible to the average user. Microsoft’s Phi and models from Mistral AI, in particular, have shown relatively strong performance given their size (Jiang et al., 2023; Abdin et al., 2024). Pre-training Between2018and2022, akeypre-trainingtacticintheeffort to improve genAI performance was to increase model size. Size is measured in the number of estimated parameters, most notably the weights that determine how much influence each neuron has on each of the others in a neural network. For example, GPT models initially had 117 million parameters in 2018 (GPT-1), then 1.5 billion in 2019 (GPT-2), and a staggering 175 billion in 2020 (GPT-3). Unfortunately, costs typically rise quadratically when parameters are added: each word (token) in the input sequence has to be compared to all of the others in the attention mechanism, as discussed in the box on the Transformer. Remarkably, the cost of training a model of a given size was halved approximately every eight months through 2024 because of improvements in algorithmic efficiency (Ho et al., 2024). But, by 2022, the direction of innovation had begun to shift; scaling laws indicated diminishing returns to 22See“AIscalinglawsareshowingdiminishingreturnsforcingAIlabstochangecourse,” on techcrunch.com, November 20, 2024. 23In particular, Mamba estimates the parameters of a latent state space structure. 22

model size for foundation models, and focus turned to optimizing training efficiency. Hoffmann et al. (2022), for example, note that raising the number of parameters faster than the amount of data used leads to “undertrained” (imprecisely estimated) parameters that are less useful for inference and recommend scaling the model size no faster than the dataset size. Fine-tuning As the returns to model size scaling have diminished, researchershavebegunfocusingmoreattentiononfine-tuningfoundationmodels for specific tasks. That is, developers have used domain-specific training data to increase the model’s expertise beyond the capabilities of foundation models for narrowly defined questions. Several techniques have been developed to improve this type of foundation model adaptation. Transfer learning involves taking a foundation model that has been fine-tuned for a specific task and adapting it (finetuning it again) for a related task, a process which typically involves only a modest amount of modification. Instruction tuning provides the model with guidance for specific scenarios that enhance its performance for targeted use cases. For example, the Alpaca model is a fine-tuned version of Meta’s Llama model, refined by adding a set of instructions and desired outputs to the training process (Taori et al., 2023).24 Reinforcement learning has been extensively applied in AI for fine-tuning, allowing the model to refine its behavior based on a reward function (Mnih et al., 2013). A particularly influential approach along these lines was reinforcement learning from human feedback (RLHF), a technique that aligns the model’s output with human preferences by learning explicitly from human reactions to the model’s responses to their queries (Christiano et al., 2017). For example, the model may provide the user with several responses to a query and ask the user to rank them, then use the ranking to improve performance in future queries. This technique gained widespread attention in the context of genAI with the release of InstructGPT in 2022 (Ouyang et al., 2022). Inference Inference costs (in terms of electricity, time, compute, and carbon emissions) have risen with the popularity of genAI, leading to a focus 24ThesupplementaltrainingsetshowedLlamathedesiredresponsetoparticularqueries, such as “Instruction: Brainstorm a list of possible New Year’s resolutions. Output: lose weight, exercise more, eat healthier.” 23

on techniques to make this step more efficient.25 One of the most important innovations in this area has been the Mixture of Experts (MoE) approach, an architecture that activates only a subset of model parameters in response to queries (Jacobs, Jordan, Nowlan and Hinton, 1991). A router determines which “expert” (smaller model) to activate based on the input. This adjustment allows the model to employ only a subset of the billions of parameters in the original foundation model, drastically reducing computational costs (Shazeer et al., 2017). Pruning is another inference refinement; here extraneous parameters are removed outright from the model (Cetin et al., 2024). Developers also incorporate distillation, a compression-like technique that uses knowledge from large complex models to inform smaller, less costly models (Hinton, Vinyals and Dean, 2015). Quantization reduces the level of accuracy in order to reduce costs in terms of computation and memory requirements (e.g. moving from 32-bit to 8-bit floating point precision). For example, in 2023 Microsoft developed BitNet, an LLM with competitive performance that transforms floating point parameters in the model to a terniary digit (0, 1, or -1) (Wang et al., 2023). Token caching involves temporarily storing information anticipated to be needed in future inference steps.26 For example, the prompts sent by a user to a model for inference may tend to be similar, implying that by caching processed tokens (words, phrases), anticipated computation costs can be reduced (Pope et al., 2023).27 The recently introduced DeepSeek R1 model leverages several of the techniques described above to deliver a substantial performance improvement compared to existing models (See the box, “Landmark AI Models: DeepSeek AI.”). 25After setting a flat fee for access to ChatGPT, OpenAI CEO Sam Altman was surprisedbythefloodofuserrequests. See“SamAltmansayshe’slosingmoneyonOpenAI’s $200-per-month subscriptions: ‘People use it much more than we expected’.” Economic modelshavelongpredictedthisoutcomewhenthepriceofthemarginalunitissettozero. See, for example, the discussion of overuse of common pasture in The Wealth of Nations (Smith, 1776). 26The fact that caching, a technique first employed in computing in the 1960s (on IBM System/360 mainframes), was first introduced to genAI inference in 2023 suggests there maybeotherbasicnumericalmethodsthathaveyettobeleveraged. Ifso,thisbodeswell for future efficiency improvements. 27Sakana AI recently introduced an approach using neural networks called NAMMs to optimallydecidewhethertokeepordiscardtokens,savingupto75%ofcachememory. See “New LLM optimization technique slashes memory costs up to 75%,” at venturebeat.com, Ben Dickson, December 13, 2024. 24

In some cases, recent efforts have acted to extended inference time to enhanceperformance. Extendinginferencetimeraisescosts, butonthemargin, resources may be better used for responding to queries than for additional training. OpenAI’s recent o1 model exemplifies the benefits of this approach via its superior performance across various domains, particularly reasoningheavy tasks.28 And, constraining the model to provide a response grounded in logic can lead to better results as well; chain-of-thought (CoT) reasoning guides the LLM to articulate a series of steps in its inference (Wei, Wang, Schuurmans, Bosma, Xia, Chi, Le, Zhou et al., 2022). Landmark AI Models: DeepSeek R1 A salient example of recent model innovation occurred in January, 2025 when DeepSeek unveiled R1 (short for “reasoning”) (DeepSeek- AI, 2025). This model set a new bar for cost-effective, high-quality multi-modal models. DeepSeek stunned the AI research community by reporting the costs for training this 671 billion parameter model to be $6 million.a Although observers noted this figure does not account for the substantial reliance on R&D by other AI developers, the news led toamaterialchangeinperceptionsoftheroleofChineseAIcompanies. News of DeepSeek R1 coincided with a one-day market valuation drop of nearly $600 million for NVIDIA as DeepSeek had achieved leading edge performance with far less of NVIDIA GPU-provided compute than previously believed necessary. And, the DeepSeek R1 release may have spurred a major competitor to accelerate their product offerings: OpenAI soon updated o3-mini, its extended inference model, lowered its price, and offered a free-trial version. Deepseek R1 blends several concepts that had been known in the genAI field for some time to improve performance and slash inference costs including mixture of experts, chain-of-thought reasoning, reinforcement learning, distillation, and quantization. Deepseek R1 endeavored to optimally combine all of these techniques to reduce costs. Some of these techniques were already commonly used in frontier models but others, such as their novel approach to reinforcement learning, called Group Relative Policy Optimization (GRPO), were not. 28See“OpenAIscientistNoamBrownstunsTEDAIConference: ‘20secondsofthinking worth 100,000x more data’,” at venturebeat.com, Michael Nun˜ez, October 23, 2024. 25

DeepSeek R1 (continued) Relative to other frontier model developers, DeepSeek shifted attention from supervised learning during the fine-tuning stage to reinforcement learning. aSee the technical report for DeepSeek V3 (DeepSeek, 2024): “...DeepSeek-V3 costs only 2.788M GPU hours for its full training. Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576million. Notethattheaforementionedcostsincludeonlytheofficialtraining of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data.” Agents Anotherdirectionforprogresscurrentlyreceivingintenseattention— distinct from the pre-traning/refinement/inference optimization approach— isthecreationofAIagents. Agentic AIsystemsdevelopstrategiestopursue broad goals and recalibrate in response to their environment, in contrast to tool-based AI, which has a stable structure and calibration and is equipped only to respond to carefully crafted requests. While agents of different kinds are commonplace in the home and at work, both in physical form, such as self-driving cars, and in virtual form, such as web browsers, agentic AI is distinguished by its autonomy in pursuit of more abstractly specified goals. One definition comes from Boston Consulting Group:29 Put simply, AI agents are artificial intelligence that use tools to accomplish goals. AI agents have the ability to remember across tasks and changing states; they can use one or more AI models to complete tasks; and they can decide when to access internal or external systems on a user’s behalf. This enables AI agents to make decisions and take actions autonomously with minimal human oversight. AI agents extend capabilities of AI to wider use in business, particularly to people with little technical knowledge or skill. For example, in the case study of healthcare, we noted that LLMs are being used to help with the paperwork burden faced by healthcare professionals, including making appointments, following up on treatment protocols and submitting insurance 29See “AI Agents,” at bcg.com. 26

claims. Specialized AI agents can be programmed to provide this rather specific type of assistance, operating in combination with LLMs. However, they require programming and testing and cannot simply apply an off-the-shelf AI program. Correctly specifying the objective function for the AI agent is difficult. It may inadvertently be guided to maximize its programmed reward while subverting the real objective of the user.30 Moreover, in extreme cases bad actors may hijack agents in business or military uses, leading to disastrous outcomes. 3.3.2 Hardware GenAI model training and inference has massive computational requirements, making ongoing innovation in electronic hardware (and related hardware, such as cooling systems), essential to continued technical advance. Progress in this area has been rapid in recent years. GenAI processing relies heavily on graphics processing units (GPUs). Like the image processing tasks GPUs were first designed for, training and inference for deep neural networks requires a large number of identical, independent computations which can be run in parallel on a GPU. (In contrast, microprocessing units (MPUs) perform computations sequentially, in the main.) GenAI processing also relies to a lesser extent on field-programmable gate arrays (FPGAs), which are more flexible than GPUs, particularly for inference. And, application-specific integrated circuits (ASICs) are used for specific steps in estimation as well.31 For example, tensor processing units (TPUs) are customized for matrix multiplication, heavily used in neural networks. They consisting of thousands of multipliers and adders connected to each other to form a large physical matrix. Storing the matrix parameters in on-chip registers drastically reduces the need to access off-chip memory, dramatically increasing computational efficiency. Successive GPUs released by NVIDIA have delivered leaps in AI performance achieved by improvements in the power consumption and computational power of the processing cores—known as Compute Unified Device Architecture (CUDA)—adding and refining TPUs on the GPU chip, and increasing and refining cache (on-chip) memory. The history of NVIDIA GPU 30See “Faulty reward functions in the wild,” at openai.com, December 21, 2016. 31ThelogicalcircuitsonFPGAscanbereconfiguredthroughprogrammingandtailored to specific algorithms. The circuitry in ASICs is customized for specific tasks at the time of fabrication and cannot be changed. Technically, GPUs are ASICs as well. 27

Table 3: Price of Compute, Selected Nvidia GPUs Model Year Price TFLOPS Price/TFLOP Transistors GeForce 8800 GT 2007 $349 0.3 $1,163 0.8B GeForce RTX 4060 2024 $299 15.1 $20 18.9B Change (ann. rate) NA –1% 23% –24% 19% Source: TechPowerUp. generations illustrates this progression.32 CUDA, first introduced in 2007, explicitly supported non-graphics computing; prior to CUDA, programmers were required to reframe computations in terms of graphics operations.33 As AI became a prominent use case for NVIDIA GPUs, new architectures were increasingly optimized for deep learning—beginning with Pascal (2016)— and for genAI—beginning with Hopper (2022). On-chip TPUs were first introduced with the Volta microarchitecture in 2017 and successive GPU generations—Turing (2018), Ampere (2020), Hopper (2022), and Blackwell (2024)—each included improved TPUs. While the engineering performance of leading edge GPUs has rocketed upwards in recent years, prices have increased dramatically as well. Fortunately for productivity, holding performance constant, the price of GPUs has moved down: In 2007, a $349 GPU provided 0.3 teraflops (TFLOPS) of compute and in 2024, a $299 GPU delivered 15.1 TFLOPS, implying an average annual rate of price decline of 24% that persisted for 17 years (table 3). As shown in fig. 8 on the next page, this example is representative of a broader trend: the cost efficiency of computation has improved dramatically. Moreover, performance measured in TFLOPS likely understates the advance in AI hardware performance over this period as some design changes operate to reduce the number of TFLOPS needed for a given level of AI training or inference. For example, GPUs have taken on board additional logic blocks to accelerate matrix operations (Tensor cores) and GPUs have been partitioned to provide more flexibile use of TFLOPS (Lino, 2024). Continuation of this trend of declining computation cost is not guaranteed. What seems like inexorable progress from a distance is in fact the 32This discussion was substantially improved by a discussion with Claude AI and research on Wikipedia. 33With CUDA and associated tools, programmers can use standard programming techniques in C/C++. 28

Figure 8: Price Indices of GPU Improvements (a) Price per TFLOP Note: PriceperTFLOPS(trillionfloating-pointoperationspersecond). Thebluelinerepresentsthebest fitlineforNVIDIAGPUs,andtheorangelinerepresentsthebestfitlineforAMDGPUs. Source: TechPowerUp. (b) Price per vRAM Note: Price per vRAM in GB (video random access memory). The blue line represents the best fit line forNVIDIAGPUs,andtheorangelinerepresentsthebestfitlineforAMDGPUs. Source: TechPowerUp. 29

result of a long sequence of difficult engineering feats when seen up close.34 Historically, a key contributor to falling costs in the semiconductor industry has been steady miniaturization at the leading edge of the chip industry. Through 2003, the linear dimensions of the features (e.g. transistors) on leading edge chips were reduced by 30 percent with each generation (node), yielding a (0.7∗0.7 =) 50 percent reduction in the space they occupied on a chip every two years (fig. 9b on the following page). From that year forward, the dimensions of these features were reduced far more slowly, though the industry, using the term “effective node,” contended that performance gains with each generation matched the historical trend. The apparent slowdown in cost improvements of TFLOPS and vRAM in fig. 8 on the previous page may well reflect that development. Since then, chip innovation has relied less heavily on miniaturization. For example, increasing “die size” (the surface area of the chip) has allowed the number of transistors per chip to continue to climb (fig. 9a on the following page).35 The proximate cause of the miniaturization slowdown was the end of a regularity known as “Dennard scaling” whereby power usage was largely unchanged even as more electronic activity was squeezed into the same surface area. Because heat generation is increasing in power usage, cooling cost began to rise with each node. Consequently, development efforts in the computing sector shifted to a balance of computing speed and power consumption. Energy efficiency merits greater attention as concerns have arisen that power demand for genAI use may outpace growth in power-generation capacity, throttling genAI-led productivity advances. Two recent developments temper that concern. The energy efficiency of the most efficient industrial supercomputers, which include the data centers of major IT service providers, has roughly doubled since 2022, when genAI use began to climb, and U.S. electricity generation is forecast to expand substantially in coming years (fig. 10 on the following page).36 34To get a sense of the staggering array of engineering problems involved in moving between chip generations, flip through any edition of the International Technology Roadmap for Semiconductors on the worldwide web. 35Since “Moore’s Law” is a prediction that the number of transistors per die (chip) will double every two years, it arguably still holds true. Of course, because Moore (1965, 1975) does not contain a testable hypothesis, debates about whether Moore’s Law holds are largely semantic. As is evident from fig. 9a on the next page, there is no single value for transistors per chip at any point in time. 36U.S. datacenter power consumption was 176 terawatt hours in 2023 and forecasts for datacenter power use in 2030 range from 200 to 600 terawatt hours. The U.S. Energy 30

Figure 9: Progress on Moore’s Law: Two Perspectives (a) Transistor Count (b) Transistor Size Figure 10: Indicators of Power Demand and Supply (a) Data Center Efficiency (b) U.S. Electricity Generation Note: Data center efficiency indicator is gigaflops per watt of supercomputers labeled “industry” or “vendor”. Source: Top500.orgforefficiency. U.S.EnergyInformationAdministrationforelectricitygeneration. 31

3.3.3 Datasets GenAI models “learn” by adjusting parameters to best represent the content of large amounts of text (and other media), allowing them to estimate the probability that a given word or phrase should appear next in the sequence it generates in response to a prompt. Loosely speaking, the larger the corpus of text available to the model, the better it can estimate these probabilities. Figure 6onpage20illustratestheincreaseovertimein the size of the datasets used to train the models. The change in the coloration of the circles, from purple to blue to green to yellow, indicates the rapid increase in the size of the training datasets. Some observers believe that we may be approaching the limit of what we can learn from public text data.37 A crucial nuance to this aspect of model improvement is that access to more information, not more text in itself, is needed to continue to improve genAI models.38 To illustrate with an extreme example, doubling the size of the training text by exactly duplicating the corpus will yield no improvement to the model. Good quality data enables the model to learn the underlying language structure efficiently, which requires thedatatospanthefullextentofthelanguagewithminimalredundancyand noise. Consequently, there are two looming challenges with respect to training data. First, more obscure or sensitive topics may have little coverage in public-facing content. Second, diminishing marginal returns to training will set in as developers move from information-rich content, such as Wikipedia and scientific articles, to more inane text, like social media posts. One approach to mitigating the content constraint is transfer learning, where a model pre-trained with public data is improved by further training using proprietary data.39 This not only increases the size of the dataset, but may increase the scope as well. Developers are also wrestling with the question of “domain generalization,” where a model is used to generate content on topics outside the scope of the training dataset (Zhou et al., 2022). As the information content of the marginal text from the internet falls, Information Administration’s long-term forecast is for an capacity to increase by 532 terawatt hours between 2023 and 2030. 37Villalobos et al. (2022) predict that public high-quality text data may become scarce as early as 2026. 38In terms of information theory, models improve from entropy, the expected amount one will learn from the data generation process producing the text (Shannon, 1948). 39Cockburnetal.(2018)notethatthisraisestheissueofmarketstructureasapotential constraint on progress in AI. 32

other techniques for generating data for training become more attractive. In one approach, small localized modifications of the training data can be introduced. For example, the performance of an image recognition model may be improved by supplementing the training set of labeled images with their mirror images.40 Such variations in input data constrain the model parameter search process in a useful way. If an image is a dog, say, the model should recognize that its mirror image is a dog as well, a desirable property known as “regularity.” Another approach to augmentation is the use of “synthetic data” created via generative models to emulate the patterns and characteristics of real data (Liu et al., 2024a). For example, an LLM designed to tackle mathematical questions might be trained on a dataset of questions generated by another LLM, using bootstrapping techniques to create similar questions from a human-produced training dataset (Yan et al., 2025). This approach is attractive for medical imagery as well, where creating training data, such as CT scans, is both resource-intensive and constrained by privacy concerns (Guo et al., 2025).41 Last, datasetscanbeaugmentedbyharvestinginformationcollectedwith sensors, particularly in physical environments such as industrial robots and autonomous vehicles (Feng et al., 2019). This approach offers the prospect of a broader domain of use for genAI models and further diffusion of the technology. 3.4 The Case that GenAI is a GPT To summarize, although it is early to tell how widespread the use of genAI will be, the case that generative AI is a general-purpose technology is compelling, supported by the impressive record of knock-on innovation and ongoing core innovation. Of the three GPT criteria, widespread adoption is the most difficult to 40This approach was taken by the developers of AlexNet, a model which revolutionized the field (Krizhevsky et al., 2012). 41Theusefulnessofsyntheticdataisamatterofsomedebate. Someobservershaveraised concerns that training with synthetic data (and AI-generated text increasingly present on the internet) will yield low-quality or even nonsensical results, a phenomenon known as “model collapse” (Alemohammad et al., 2023; Shumailov et al., 2023). Others have argued that model collapse only occurs when the original training text is replaced by model-generated text (Gerstgrasser et al., 2024). 33

arguethatgenAIhasmet. AlthoughsomefieldstudieshaveprovidedencouragingresultsthatgenAImayraiseproductivity, outsideoflargecorporations, few firms have adopted the technology. The share of jobs requiring AI skills is low and has moved up only modestly, suggesting that firms are taking a cautious approach. The ultimate test of whether genAI is a GPT will be the profitability of genAI use at scale in a business environment and such stories are hard to come by at present. That said, use among individuals is high, perhaps unbeknownst to their employers, and with genAI increasingly folded into office productivity software (such as Microsoft 365), its use may become so unremarkable that firms and workers may not be aware it is in use. The case that genAI meets the knock-on innovation criterion is somewhat stronger. Key areas of product innovation include user interface software and interfacewithrobotics,wherethegenAImodelenablesfarmoresophisticated applications. Productionprocessinnovationincludesdigitaltwinstoimprove production line efficiency. Organizational innovations include restructuring of the product design business function. What share of organizations can justify the digital transformation needed for genAI, such as centralized data governance,particularlyfornon-digital-nativecompanies,remainstobeseen. Coretechnologyinnovationisthecriteriaforwhichthecaseistheclearest. GenAI performance has moved up at a blistering pace since the introduction of the Transformer thanks to increasingly large datasets and application of more computing power. More importantly, performance has risen while holding inputs constant (data, compute, and model size), driven by algorithmic improvements. If this trend continues, the direct cost of using genAI will fall, spurring greater adoption. Artificial General Intelligence In1960, HerbertA.Simon, winnerofthe1975Turing Award and1978 Nobel Memorial Prize in Economic Sciences wrote, “within the very near future—much less than twenty-five years [before 1985]—we shall have the technical capability of substituting machines for any and all human functions in organizations” (Simon, 1960). This is the concept now known as “artificial general intelligence” (AGI).a Demonstrating the feasibility of AGI is a long-run objective of many AI researchers but is not a particularly interesting exercise for economists. Technical feasibility is a necessary condition, but far from 34

AGI (continued) a sufficient one, for a technology to raise productivity. It is technically feasible to turn lead into gold, for example, but only at prohibitive cost (Matson, 2014). And, conjecture about future technology, however well informed, is not sufficient for a useful productivity forecast, which must answer the question of when, not just if, the technology will appear and, a fortoriori, when it will be practical to use. Moreover,achievinghuman-levelperformanceonalltaskswouldlikelyentail devoting resources to improving performance on tasks AI is ill-suited for at the expense of tasks where improvement would be more readily achieved. Dell’Acqua et al. (2023) explore this issue and provide extensive evidence this is a first-order impediment to AGI. Some present-day IT leaders believe AGI is imminent: In 2024, Elon Musk predicted the arrival of AGI within two years, Sam Altman predicted its arrival by 2025, and Dario Amodei expected AGI by 2026. Others are skeptical: Yann LeCun speculated it may never be achieved.b Historically, technology forecasts have been highly unreliable. Berkeley (1949), for example, foresaw a machine that would read handwritten text; noteworthy practical use of handwriting recognition systems U.S. Post Office came fifty years later. Fortunately, AGI is not a precondition for genAI to be a GPT, nor do we need a forecast for the timing of the arrival of AGI to forecast the effects of genAI on productivity. aRemarkably, Cˇapek (1920) had already envisioned a world with robots that performed all human tasks, including ones involving physical manipulation of the environment. At that time, although punchcard tabulators programmable with plugboards existed, most “computers” were human (Grier, 2007). bSee “Tesla’s Musk predicts AI will be smarter than the smartest human next year,” Reuters, April 8, 2024; “Here’s how far we are from AGI, according to the people developing it,” Business Insider, November 9, 2024; “Meta AI Head: ChatGPT Will Never Reach Human Intelligence,” PYMNTS, May 22, 2024. 35

4 Is GenAI an Invention of a Method of Invention? In classical “light bulb” growth models (commonly known as “Solow- Swann” models), the source of total factor productivity (TFP)—the part of productivity growth not attributable to capital accumulation—is unspecified. About the importance of TFP that emerged when this model was brought to the data, Abramovitz (1956) famously observed, “since we know little about the causes of productivity increase, the indicated importance of this element [TFP] may be taken as some sort of measure of our ignorance about the causes of economic growth in the United States and some sort of indication of where we need to concentrate our attention.” “Endogenous growth” models have since appeared, including some that add a research sector to produce new technologies in response to incentives (Romer, 1994; Aghion and Howitt, 1992; Akcigit, 2023).42 Like other sectors, efficiency in the research sector can be increased by the use of appropriate capital, such as an invention of a method of invention (IMI).43 Griliches (1957) noted that the hybridization process developed for creating new varieties of corn played this role. Other examples of IMIs are shown in table 4 on the following page, grouped into observational, analytical, communication, and organizational tools. We consider below whether genAI falls into these categories and how it can contribute to research productivity beyond what is contributed by machine learning. We then review a number of broad indicators of the role of AI in research, including patent filings, the share of AI use accounted for by users with research jobs and tasks, and new evidence on the prevalence of AI references in company conference calls. Prior to the appearance of genAI, AI had already diffused across a wide range of scientific disciplines (Carobene et al., 2024). And, it had already been shown to improve the efficiency of research. Cockburn, Henderson and Stern (2018) note that pre-generative AI assists with the “labor-intensive 42Ofcourse,inactualityTFPisnot simplytheoutputofaresearchsector. TFPresults when firms choose, in response to research results, to make complementary investment in intangibles (Brynjolfsson et al., 2021) in the context of government policy (Baily et al.,forthcoming)andisimportantlyaffectedbybusinessdynamism(Decker,Haltiwanger, Jarmin and Miranda, 2017), labor market efficiency (Davis and Haltiwanger, 2014), and market structure (Goettler and Gordon, 2011). 43The term originates from Whitehead (1925), according to Mowery and Rosenberg (1999). 36

Table 4: Examples of Inventions of Methods of Invention Observational tools Telescope 1608 CE Compound microscope 1620 CE Pendulum clock 1656 CE DNA sequencer 1973 CE Analytical tools Mainframe (IBM S/360) 1964 CE Personal computer (IBM PC) 1981 CE Machine learning 1998 CE Communication tools Printing press (Gutenberg) 1439 CE Internet protocol (TCP/IP) 1975 CE Organizational innovations Scientific societies (Accademia dei Lincei) 1603 CE Corporate labs (GE) 1900 CE Government labs (U.S. NRL) 1923 CE Big science (Oak Ridge) 1961 CE Source: Authors’ judgment. search with high marginal cost of search” involved in many types of R&D. Put differently, AI improves prediction, a point emphasized by Agrawal et al. (2018), including predicting how materials might behave. Examples of phenomenal success are well known. Scientists have made major advances toward practical nuclear fusion using reinforcement learning techniques to adjust the magnetic system that contains the plasma in a fusion reactor (Degrave et al., 2022; Seo et al., 2024). Richardson et al. (2020) used the “knowledge graph”—which encodes relationships among scientific publications using machine learning—created by BenevolentAI to produce a novel treatment for COVID-19. Machine learning has also been used extensively for predicting the properties of novel metal alloys, economizing on physical experimentation and computer simulations (Hart et al., 2021). Gazzani and Natoli (2024) find evidence that pre-generative AI innovation has increased industrial productivity and lowered prices. Our focus is on the question of whether genAI enables additional efficiencies in R&D beyond these and other improvements provided by machine learning. In particular, we ask whether 37

genAI enhances measurement, analysis, communication, and organization of invention. GenAI as an observational tool Observational tools, such as microscopes, telescopes, and cameras produce imperfect images due to defects in their components and variation in the environment. GenAI provides a tool to impute imperfect portions of images as well as missing observations in datasets of all kinds in a fashion more consistent with the apparent properties of the underlying phenomena. For example, generative techniques for image enhancement, which rely on an implicit model of the manifold of the data generating process—closer to the actual physics, say, of a remote galaxy—perform better than techniques, such as splines, relying solely on smoothness assumptions (i.e. that nature does not make leaps) (Liu et al., 2018; Lugmayr et al., 2022). GenAI as an Analytical Tool Like the compound microscope for physical phenomena, Christian (2020) notes that LLMs serve as a kind of microscope to look at social phenomena. Caliskan et al. (2017), for example, find that “text corpora contain recoverable and accurate imprints of our historic biases.” This new visibility may promote and support analysis of social science questions not previously tractable. There has been an explosion of sentiment analysis and other forms of NLP in recent years fueled by this capability of genAI.44 While the identification of underlying sentiment (encoding) is strictly speaking a function of the LLM, conveying the discovered sentiment to the user is necessarily a generative process. Korinek (2023) documents a variety of potential roles for genAI in the economic research process; that genAI may play a similar role in many other fields is a reasonable conjecture.45 44Sentiment analysis is possible with earlier forms of AI but the capabilities of genAI models are vastly greater (Gentzkow et al., 2019; Dell, 2025). 45EarlyversionsofthispapercitedToner-Rodgers(2024),whichpurportedtoshowthat genAI substantially accelerates the discovery process in materials science. The veracity of thatworkhassincebeenquestionedbytheauthor’sinstitutionandprominentresearchers in the field. Credible empirical evidence on the effect of genAI on scientific research efficiency would be a timely contribution to resolving the uncertainty around the effects of genAI on productivity. 38

Table 5: The Stages of a Research Project Conceptual reviewing the literature, formulating the broad problem, identifying specific goal Planning determining research design and procedures, identifying resource needs, procuring funding Empirical collecting data, preparing data for analysis Analytical identifying data features, testing hypotheses, interpreting results Dissemination communicating to audience (colleagues, industry, policymakers, public, students) in written, visual, and oral form GenAI-Supported Organizational Innovation Institutional organization plays a central role in the effectiveness of R&D (Mowery and Rosenberg, 1999), as do informal associations into professional networks (Wang and Barab´asi, 2021) and geographic clusters (Porter and Stern, 2001). Consequently, the method of invention for any given research program properly includes the institutions involved. Emerging applications of AI “digital twins” offer the prospect of R&D with a reduced institutional footprint in many areas of study. Among these are drug discovery (Bordukova et al., 2024), industrial research (Tao et al., 2024), and materials science (Kalidindi et al., 2022). For example, generalized adverserial networks (GANs) may provide an alternative to animal testing for toxicology (Chen et al., 2022). GenAI as a Communication Tool Although empirical and analytical stages of research projects focus on measurement and calculation, many aspects of the research process involve manipulating language. GenAI may be employed in the writing tasks involved in the conceptual, planning, and dissemination stages of research projects, such as drafting literature reviews, grant applications, and seminar slides (table 5). Whether, on net, genAI improves the efficiency of such tasks once the review and editing of the documents drafted by genAI is accounted for is an open question. If so, genAI mayplayasimilarroletotheprintingpressandwordprocessingasacatalyst to the invention process. 39

Research agents AI agents (discussed in section 3.2.1 on page 17) have emerged that endeavor to automate the core of research entirely, generating research questions, designing and conducting experiments, and reporting results. Thus, research agents may play the role of an observational, analytical, and communication IMI all at once. Examples include Google’s AI co-scientist and Sakana’s The AI Scientist (Gottweis et al., 2025; Lu et al., 2024).46 Opinions of the significance of research agents vary widely. Importantly, the design, conduct, and communication of experiments is only a portion of the activities of a scientist (table 6 on the next page). Even so, Lu et al. (2024) report that The AI Scientist can generate publishable research for as little as $15 per journal article, a striking finding. On the other hand, Beel et al. (2025) evaluate Sakana’s agent and conclude, Our evaluation of the AI Scientist reveals critical shortcomings. The system’s literature reviews produced poor novelty assessments,oftenmisclassifyingestablishedconcepts(e.g.,micro-batching for stochastic gradient descent) as novel. It also struggles with experiment execution: 42% of experiments failed due to coding errors, while others produced flawed or misleading results. Code modifications were minimal, averaging 8% more characters per iteration, suggesting limited adaptability. Generated manuscripts were poorly substantiated, with a median of five citations, most outdated (only five of 34 from 2020 or later). Structural errors were frequent, including missing figures, repeated sections, and placeholder text like “Conclusions Here”. Some papers contained hallucinated numerical results. Despite these flaws, the AI Scientistrepresentsaleapforwardinresearchautomation. Itgenerates full research manuscripts with minimal human input, challenging expectations of AI-driven science. Moreover, genAI research agents may have a subtle but important limitation: uncovering the fundamental features of phenomena. Li et al. (2022) argue that genAI does have that capability. They trained a generative model to play the board game Othello without providing the rules of the game then demonstrated that the model can play appropriately in a setting not found 46Stoughton (2023) documents co-scientist innovation at the National Science Foundation. 40

Table 6: Scientific Activity beyond Research Projects Conceptual designing a research program (a connected set of research projects); packaging program to influence appropriate audiences (e.g. writing textbooks) Leadership playing executive and advisory roles in local and profession-wide academic and government specialist communities Mentoring supervising research, advising and instructing students Support fostering buy-in to research program from institutional leadership Networking recruiting collaborators and maintaining relationships Commercialization translating research results into practical applications in the training data, concluding that genAI has created an “emergent world model.” Other research has challenged this conclusion. jylin04 et al. (2024) argue that the model is employing a “bag of heuristics,” rather than a set of game rules.47 This question is a crucial one in determining the capabilities of genAI to contribute to science. Without a model of the underlying structure of the physical or social phenomenon under study, one cannot articulate its fundamental laws. This limitation may arise naturally from the training process; humans learn the fundamentals of science from textbooks, but these laws may not be the rhetorical foundation for the verbal exchanges on the topic found in the training corpus. 47Ausefulentrypointtothisongoingdebateis“LLMsandWorldModels,” byMelanie Mitchell, February 13, 2025, found at the AI: A Guide for Thinking Humans Substack blog. 41

4.1 Indicators of GenAI Research and of GenAI Use in Research We discuss below a set of indicators of genAI research (patents) and of genAI use in research (conference call transcripts and genAI queries).48 Patents AI-related patents issued by the United States Patent and Trademark Office (USPTO) increased markedly following the advent of genAI, suggesting a related surge in genAI research (fig. 11 on the following page).49 The USPTO index of AI-related patents began climbing in 2018, shortly after the publication of the seminal paper by Vaswani et al. (2017) which introduced the Transformer architecture, quickly reaching a level 50 percent higher, which it has sustained since 2019. We also observe that increases in patent activity for AI modalities particularly related to genAI—natural language processing (NLP), vision, speech, and knowledge processing—have risen even further. This suggests that the recent surge in patenting activity is not merely a reflection of advancements in machine learning. GenAI Prompts Handa et al. (2025) provide a rich set of information on actual genAI use in their Anthropic Economic Index (AEI), a useful complement to the detailed work on the potential impact of genAI based on analysis of job descriptions from Eloundou et al. (2024). The AEI assigns millions of conversations from Claude (Anthropic’s premier genAI system) to roughly 3,500 tasks defined by the U.S. Department of Labor’s O*NET Dataset.50 An equal fraction of each task’s percentage share of all prompts is then apportioned to each occupation which includes that task in O*NET. 48For evidence of the potential for genAI use in research based on job descriptions, see Eloundou et al. (2024) who note that “scientists and researchers” and “technologists” are the job groups most highly exposed to LLMs, and that “this suggests that when LLMs improve they have potential to cause downstream improvements in R&D productivity for workers in sectors deploying them.” 49Pairolero et al. (2025) use BERT-based embeddings to refine a previous iteration of their patent classification methodology that used Word2Vec. We use their most conservative threshold of 93% probability to identify AI-related patents. We refer the reader to the USPTO website hosting their data for more information: https://www.uspto.gov/ippolicy/economic-research/research-datasets/artificial-intelligence-patent-dataset. 50Patterns of Anthropic use may not be representative of the patterns for all genAI programs, of course. See section 7.1 of Tamkin et al. (2024) for a discussion of related work. 42

Figure 11: AI Mentions in Scientific Patents Source: Artificial Intelligence Patent Dataset (2023), U.S. Patent Office. Table 7 on the next page shows the the estimated share of prompts accounted for by occupational groups, their employment share, and the ratio of the two. (If prompts were equally distributed across all workers, these ratios would each be equal to 1.) “Computer & mathematical occupations”, which includes the computer programmers for whom genAI use is reputed to be especially intense, have the highest ratio of prevalence of genAI use to occupational prevalence, 10.9. Use intensity is nearly as high among scientists, who account for 7.1 times as many prompts as would be found if prompts were equally distributed across workers.51 Other occupational groups with high relative prevalence of genAI use include “arts, design, sports, entertainment & media”; “architecture & engineering”; and “educational instruction & library”. The remaining 87.6% of employment is accounted for by occupations which AEI found had a share of Claude prompts roughly equal to or lower than their share of employment, highlighting the very concentrated nature of genAI adoption in the economy at present. Table 8 on page 45 shows the prevalence of selected O*NET tasks related to scientific discovery among Claude AI prompts. These tasks collectively account for only 0.9% of all prompts, revealing that the share of prompts accounted for by scientists (6.4%) includes far more than scientific discov- 51Naturally, they may be using genAI for computer programming tasks as well. 43

Table 7: Occupations with High GenAI Usage Job Type Prompt Share Empl. Share Ratio Computer & mathematical 37.2 3.4 10.9 Arts, design, sports, entertainment, 10.9 1.4 7.8 & media Life, physical, & social science 6.4 0.9 7.1 Architecture & engineering 4.5 1.7 2.6 Educational instruction & library 9.3 5.8 1.6 Memo: Other occupations 31.8 87.6 0.4 Note: Percent share of prompts submitted to Claude AI and linked to tasks by Anthropic. Task weights are apportioned equally to all occupations which include that task in O*NET. Source: Anthropic Economic Index. ery. These discovery tasks are most commonly related to the creation of mathematical or statistical models of technical phenomena, such as business, scientific, and engineering, either to foster understanding of the phenomena or to predict how modeled systems would perform. Such tasks account for 86.5% of the scientific tasks Claude AI is asked to help with. Other tasks include the advancement of mathematical science (9.0% of scientific prompts) and the design of research projects (4.5%). Figure 12 on page 46 illustrates significant automation and augmentation of tasks among our groupings of research occupations: programmers exhibit the highest automation rate, with nearly half of the requests handled by genAI being automation tasks. Social science researchers show slightly lower automationrates,witheconomistsshowingnearly23%oftheirpromptsbeing automation focused. Notably, for hard science researchers (e.g., physicists, biochemists),theshareoftheirgenAIuseforautomationisnearly15%higher than their natural science counterparts. This difference likely reflects AI’s strength in data-intensive and simulation-based research such as those found in hard sciences like physics and materials science. Conference Call Mentions GenAI’s integration into the invention process is also revealed through firm communication. We analyze quarterly earnings calls, which are routine events where firm executives discuss com- 44

Table 8: O*NET Scientific Task Prevalence in Claude AI Prompts Task Share (pct.) Modelling & Prediction 86.5 conduct logical analyses of business, scientific, engi- 46.1 neering, and other technical problems, formulating mathematical models of problems for solution by computers. design or develop software systems, using scientific 16.9 analysis and mathematical models to predict and measure outcome and consequences of design. complete models and simulations, using manual or 15.7 automated tools, to analyze or predict system performance under different operating conditions. develop mathematical or statistical models of phe- 4.5 nomena to be used for analysis or for computational simulation. designcomputersimulationstomodelphysicaldata 2.2 so that it can be better understood. developsoftwareapplicationsorprogrammingtouse 1.1 for statistical modeling and graphic analysis. Other Tasks 13.5 develop new principles and new relationships be- 9.0 tween existing mathematical principles to advance mathematical science. design research projects that apply valid scien- 4.5 tific techniques and use information obtained from baselines or historical data to structure uncompromised and efficient analyses. Note: Share of all Anthropic prompts accounted for by these tasks is 0.9%. Task-level prompts labeled by Anthropic. Source: Anthropic Economic Index. 45

Figure 12: GenAI Automation vs. Augmentation in Researcher Roles Note: Authors’calculations. Source: AnthropicEconomicIndex. pany performance, future projects, and key developments with investors and analysts. Figure 13 on the following page plots the count of the number of firms referencing AI in the context of research as indicated by the firm mentioning an AI-specific term (“machine learning,” “deep learning,” “artificial intelligence,” “genAI, or “generative AI”) within a research-related context (within 10 words of “inventi-”, “research-”, or “discover”). A sudden rise appears in 2023, with approximately 60 public companies per quarter mentioningsuchusage. ThisincreaseinintegrationofAIwithR&Disillustrative of the role it has begun to play in innovation in a corporate context. Two examples are provided below. 46

Figure 13: Mentions of AI Usage for Research in Conference Calls Note: Authors’estimates. DatathroughQ42024. Source: S&PCapitaldatabasemergedwithCompustat. In 2024, John Wiley & Sons, a publishing company, announced expansions in compound databases leveraging advanced AI and its capability to accelerate scientific discoveries: Wiley has just released two new database collections using advanced AI techniques to significantly expand the number of compounds available for analysis from food-related compounds to industrial compounds. The end goal here to help scientists reach better conclusions faster . Similarly, Cadence Design Systems, an electronic systems design firm highlighted the potential of AI to automate fields such as biology and life sciences. And then the third phase of AI adoption is AI applied to areas that were not automated in the past, okay? So I think that may take longer, maybe 5 years plus, but that has to be driven to digital biology and life sciences. I mean there’s a huge application of AI. 4.2 Is GenAI an IMI? OurassessmentisthatthereisastrongcasethatgenAIisanIMI.Indeed, it is a multifaceted IMI, having characteristics of an observational IMI (e.g. 47

image enhancement), an analytical IMI (e.g. sentiment analysis), an organizational IMI (e.g. digital twins), and a communication IMI (e.g. document drafting). Whether the excitement over genAI research agents will prove to be merited or not, genAI’s ability to augment these four dimensions of invention suggests that the idea generation process is becoming more productive. It also appears that it is taking root within the research community. The signal from the US PTO’s AI database is that AI patents surged when the use of genAI became practical. GenAI prompt analysis from Anthropic points to relatively intense use of AI by scientists as well as in adjacent fields like computing and engineering. And, corporate earnings calls increasingly mention genAI while discussing their research efforts. 5 Conclusion The release of ChatGPT in late 2022 was a stark inflection point in public interestingenAIandpredictionsofafirst-orderimpactonproductivityinthe future soon followed.52 Yet, as exciting as progress in genAI is from a science and engineering standpoint, its economic effects are highly uncertain. For firmstojustifythereorganizationandothercomplementarycapitalneededto exploitgenAI,thereturnfromthetechnology,lessthetotalcostofownership, must be high enough. Field studies do point to efficiency gains in selected business functions and many firms have experimented with the technology, but only a small share of them attest to material improvements to their bottom lines from the technology thus far. To complement the limited empirical evidence, we ask what the characteristics of genAI suggest its future impact on productivity may be. GenAI has features typical of both a general-purpose technology—headed toward being widely used, stimulating related innovation, and displaying ongoing improvement in (economic) performance—and an invention of a method of invention—raising the efficiency of R&D through improvements to observation, analysis, communication, or organization. Because both GPTs and IMIs promote productivity growth for extended periods, it is reasonable to expect genAI will have a noteworthy impact on 52Goldman Sachs analysts forecast that genAI will eventually increase in U.S. labor productivity by 15%. (See “Global Economics Comment: AI Productivity and Labor Market Impacts Are Still Small (For Now),” Joseph Briggs and Sarah Dong, March 14, 2025. 48

productivity. Importantly, genAI’s potential for productivity does not depend on the elusive goal of reaching artificial general intelligence (AGI). It canqualifyasaGPTandIMIwellbeforethearrivalofAGI.Themainhurdle is diffusion. Complementary innovations like interfaces, robotics, and agents, for example, are emerging, and technological progress is ongoing. Yet, outside of the tech sector, firm-level adoption in production processes is still modest. As an IMI, the case is stronger: genAI usage is gaining traction within the scientific community via workflows and patents. Even so, we offer several cautionary observations. First, we expect that genAI will boost productivity growth relative to the counterfactual economy without it. If the growth effect of machine learning (and other IT innovations) is waning, the impact of genAI will have to match the impact of machine learning on the likes of Amazon and Facebook for the economy to maintain the underlying trend.53 Second, the GPT effect on productivity is inherently slow as it involves complementary investment. For example, the effect on the productivity level of solid-state computing was large, but it played out over decades, damping the effect on productivity growth. The tech boom was a long time coming: Massive advances in computational technology, including the invention of the solid state transistor, had accumulated by the end of the 1940s and a steady decline in computing costs had begun (Nordhaus, 2007).54 Third, investment to deploy new technologies is fraught with risk. If genAI is a widely adopted “killer app” that defines a new era of IT, the computing capacity needed to deliver genAI to millions of simultaneous users will be massive. Anticipation of this outcome helps explain the wave of investment in data centers and electricity generation. But, building to meet anticipated demand can lead to disastrous consequences, as illustrated by the history of railroad expansion and the associated boom-bust cycles in the 19th 53Bresnahan (2024) notes that the spread of earlier AI technologies to other companies slowed once the digitally native companies had jumped in. 54Predictions of an IT-infused future of abundance soon followed, but noteworthy productivitygainsonlyappearedintandemwithtime-consumingcomplementaryinvestment, such as business re-organization. For example, Berkeley (1949), based on observation of the handful of computers in existence, eagerly anticipated automatic address books, libraries, translators, typists, and stenographers, as well as business process optimization, psychologicaltestingandtraining,weatherforecasting,andevenweathercontrol. Others, such as Martin Jr. (1960), cautioned that productivity gains would be hard-won: “The data-processing system of an organization is of almost un-imaginable complexity. The introduction of a computer usually involves widespread changes in this complex system.” 49

Century.55 For IT systems, the capacity forecasting problem is compounded by the progress of technology, which drives down the hardware investment required to deliver a given level of service.56 A critical further concern is the availability of electrical power to accommodate the demands of data centers supporting widespread genAI use (Pilz, Mahmood and Heim, 2025). And, of course, R&D is inherently risky because the returns are erratic: The chain of choices made between insight from research and greater output per hour is long. In short, our modal forecast is for a noteworthy contribution of genAI to the level of labor productivity, but the range of plausible outcomes is wide, with respect to both the magnitude of the total contribution and the contour of that impact over time (hence, the productivity growth rate). 55On the British experience in the 1840s, see Campbell and Turner (2012) and Odlyzko (2012). 56This forecasting challenge for fiber optic telecommunications systems, combined with duplicative effort by competing networks, was a major factor behind the economic downturn in 2001 (Couper et al., 2003; Doms, 2004). On the capital overhang in the telecommunications network in the early 2000s, see Hecht (2016): “The post-bubble network was vastly overbuilt and underused. In late 2002, consulting firm TeleGeography estimated only 10 percent of the long-haul fibers installed in Europe and North America carried any signals, and that only 10 percent of the wavelengths in those fibers were lit. ...Soon, creditors were trying to unload dark-fiber networks for pennies on the dollar.” 50

References Abdin, Marah, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl et al., “Phi-3 technical report: A highly capable language model locally on your phone,” arXiv preprint arXiv:2404.14219, 2024. Abdou, Ossama A, “Effects of Luminous Environment on Worker Productivity in Building Spaces,” Journal of Architectural Engineering, 1997, 3 (3), 124–132. Abramovitz, Moses, “Resource and output trends in the United States since 1870,” American Economic Review, 1956, pp. 1–23. Acemoglu, Daron and Pascual Restrepo, “The Wrong Kind of AI? ArtificialIntelligenceandtheFutureofLabourDemand,” Cambridge Journal of Regions, Economy and Society, 2020, 13 (1), 25–35. , David Autor, Jonathon Hazell, and Pascual Restrepo, “AI and jobs: Evidence from online vacancies,” Technical Report, National Bureau of Economic Research 2020. Agarwal, Nikhil, Alex Moehring, Pranav Rajpurkar, and Tobias Salz, “Combining Human Expertise with Artificial Intelligence: Experimental Evidence from Radiology,” NBER Working Papers, 2023. Agarwal, Rahul, Andreas Kremer, Ida Kristensen, and Angela Luget, “How Generative AI can help Banks Manage Risk and Compliance,” Technical Report, McKinsey & Company 2024. Aghion, Philippe and Peter Howitt, “A Model of Growth through Creative Destruction,” Econometrica, 1992, 60 (2). Agrawal, Ajay, Joshua Gans, and Avi Goldfarb, Prediction Machines: The Simple Economics of Artificial Intelligence Boston: Harvard Business Review Press 2018. , , and , “The Turing Transformation: Artificial Intelligence, Intelligence Augmentation, and Skill Premiums,” Brookings Policy Brief, 2023. 51

Akcigit, Ufuk, “Creative Destruction and Economic Growth,” The Economics of Creative Destruction, 2023. and John Van Reenen, The Economics of Creative Destruction, Harvard University Press Cambridge, 2023. Alemohammad, Sina, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G Baraniuk, “Self-Consuming Generative Models go Mad,” arXiv preprint arXiv:2307.01850, 2023. Allen, Bibb, Sheela Agarwal, Laura Coombs, Christoph Wald, and Keith Dreyer, “2020 ACR Data Science Institute artificial intelligence survey,” Journal of the American College of Radiology, 2021, 18 (8), 1153– 1159. AMA Augmented Intelligence Research, “PhysicianSentimentsaround the Use of AI in Heath Care: Motivations, Opportunities, Risks, and Use Cases,” American Medical Association, 2025. Arinez, Jorge F, Qing Chang, Robert X Gao, Chengying Xu, and Jianjing Zhang, “ArtificialIntelligenceinAdvancedManufacturing: Current Status and Future Outlook,” Journal of Manufacturing Science and Engineering, 2020, 142 (11), 110804. Baily, Martin and Aidan Kane, “AI in the Finance Sector Case Study,” Technical Report, Brookings Institution 2025. and , “AI in the Healthcare Sector Case Study,” Technical Report, Brookings Institution 2025. , David Byrne, Aidan Kane, and Paul Soto, “Productivity Policy in the United States,” forthcoming. Forthcoming. Bajari, Patrick L. and Victor Chernozhukov, “Quality-Adjusted Price Indices Powered by ML and AI with an Application to Apparel,” 2018. Bar-Hillel, Yehoshua, “A Demonstration of the Nonfeasibility of Fully Automatic High Quality Translation,” Advances in Computers, 1960, 1, 158–163. 52

Bayard, Kimberly, Emin Dinlersoz, Timothy Dunne, John Haltiwanger, Javier Miranda, and John Stevens, “Early-Stage Business Formation: An Analysis of Applications for Employer Identification Numbers,” Technical Report, National Bureau of Economic Research 2018. Beel, Joeran, Min-Yen Kan, and Moritz Baumgart, “An Evaluation of Sakana’s AI Scientist for Autonomous Research: Wishful Thinking or an Emerging Reality Towards’ Artificial General Research Intelligence’(AGRI)?,” arXiv preprint arXiv:2502.14297, 2025. Berkeley, Edmund Callis, Giant Brains or Machines that Think, John Wiley, 1949. Bick, Alexander, Adam Blandin, and David J. Deming, “The Rapid Adoption of Generative AI,” NBER Working Paper Series, 2024. Bonney, Kathryn, Cory Breaux, Cathy Buffington, Emin Dinlersoz, Lucia S Foster, Nathan Goldschlag, John C Haltiwanger, Zachary Kroff, and Keith Savage, “Tracking Firm Use of AI in Real Time: A Snapshot from the Business Trends and Outlook Survey,” Technical Report, National Bureau of Economic Research 2024. Bordukova, Maria, Nikita Makarov, Raul Rodriguez-Esteban, Fabian Schmich, and Michael P Menden, “Generative Artificial Intelligence Empowers Digital Twins in Drug Discovery and Clinical Trials,” Expert Opinion on Drug Discovery, 2024, 19 (1), 33–42. Bresnahan, Timothy, “Artificial Intelligence Technologies and Aggregate Growth Prospects,” Prospects for Economic Growth in the United States, 2019, pp. 132–172. , “What Innovation Paths for AI to become a GPT?,” Journal of Economics & Management Strategy, 2024, 33 (2), 305–316. and Shane Greenstein, “Technical Progress and Co-Invention in Computing and in the Uses of Computers,” Brookings Papers on Economic Activity. Microeconomics, 1996, 1996, 1–83. Brohan, Anthony, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny 53

Driess, Avinava Dubey, Chelsea Finn et al.,“Rt-2: Vision-Language- Action Models Transfer Web Knowledge to Robotic Control,” arXiv preprint arXiv:2307.15818, 2023. Brown, Tom, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell et al., “Language Models are Few-Shot Learners,” Advances in Neural Information Processing Systems (NeurIPS), 2020, 33, 1877–1901. Brynjolfsson, Erik, “The Turing Trap: The Promise & Peril of Human- Like Artificial Intelligence,” Daedalus, 2022, 151 (2), 272–287. , Daniel Rock, and Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies,” American Economic Journal: Macroeconomics, 2021, 13 (1), 333–72. , Danielle Li, and Lindsey R Raymond, “Generative AI at work,” Technical Report, National Bureau of Economic Research 2023. , Tom Mitchell, and Daniel Rock, “What can Machines Learn and What Does it Mean for Occupations and the Economy?,” in “AEA Papers and Proceedings,” Vol. 108 American Economic Association 2014 Broadway, Suite 305, Nashville, TN 37203 2018, pp. 43–47. Buchanan, Bruce G and Reid G Smith, “Fundamentals of Expert Systems,” Annual Review of Computer Science, 1988, 3 (1), 23–58. Bughin, Jacques and Nicolas Van Zeebroeck, “Artificial Intelligence: Why a Digital Base is Critical,” The McKinsey Quarterly, 2018. Burnside, Elizabeth S, Thomas M Grist, Michael R Lasarev, John W Garrett, and Elizabeth A Morris, “Artificial Intelligence in Radiology: A Leadership Survey,” Journal of the American College of Radiology, 2025. Caliskan, Aylin, Joanna J Bryson, and Arvind Narayanan, “Semantics Derived Automatically from Language Corpora contain Human-like Biases,” Science, 2017, 356 (6334), 183–186. 54

Campbell, Gareth and John D Turner, “Dispelling the Myth of the Naive Investor During the British Railway Mania, 1845–1846,” Business History Review, 2012, 86 (1), 3–41. Carobene, Anna, Andrea Padoan, Federico Cabitza, Giuseppe Banfi, and Mario Plebani, “Rising Adoption of Artificial Intelligence in Scientific Publishing: Evaluating the Role, Risks, and Ethical Implications in Paper Drafting and Review Process,” Clinical Chemistry and Laboratory Medicine (CCLM), 2024, 62 (5), 835–843. Cetin, Edoardo, Qi Sun, Tianyu Zhao, and Yujin Tang, “An Evolved Universal Transformer Memory,” arXiv preprint arXiv:2410.13166, 2024. Chen, Xi, Ruth Roberts, Weida Tong, and Zhichao Liu, “Tox-GAN: AnArtificialIntelligenceApproachAlternativetoAnimalStudies—ACase Study with Toxicogenomics,” Toxicological Sciences, 2022, 186 (2), 242– 259. Choi, Seong Lok, Rishabh Jain, Patrick Emami, Karin Wadsack, Fei Ding, Hongfei Sun, Kenny Gruchalla, Junho Hong, Hongming Zhang, Ziangqi Zhu, and Benjamin Kroposki, “eGridGPT: Trustworthy AI in the Control Room,” Technical Report, National Renewable Energy Laboratory 2024. Christian, Brian, The Alignment Problem: Machine Learning and Human Values, W. W. Norton & Company, 2020. Christiano, Paul F, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei, “Deep Reinforcement Learning from Human Preferences,” Advances in Neural Information Processing Systems, 2017, 30. Cockburn, Iain M, Rebecca Henderson, and Scott Stern, “The Impact of Artificial Intelligence on Innovation: An Exploratory Analysis,” in “The economics of artificial intelligence: An agenda,” University of Chicago Press, 2018, pp. 115–146. Couper, Elise, John P Hejkal, and Alexander L Wolman, “Boom and Bust in Telecommunications,” FRB Richmond Economic Quarterly, 2003, 89 (4), 1–24. 55

Crane, Leland, Michael Green, and Paul Soto, “Measuring AI Uptake in the Workplace,” FEDS Notes, 2025. David, Paul A, “The dynamo and the computer: an historical perspective on the modern productivity paradox,” The American Economic Review, 1990, 80 (2), 355–361. Davis, Steven J and John Haltiwanger, “Labor market fluidity and economic performance,” Technical Report, National Bureau of Economic Research 2014. Decker, Ryan A, John Haltiwanger, Ron S Jarmin, and Javier Miranda, “Declining dynamism, allocative efficiency, and the productivity slowdown,” American Economic Review, 2017, 107 (5), 322–26. Decker, Ryan and John Haltiwanger, “High Tech Business Entry in the Pandemic Era,” FEDS Notes, 2024. DeepSeek, “DeepSeek-V3 Technical Report,” arXiv preprint arXiv:2412.19437, 2024. DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,” arXiv preprint arXiv:2501.12948, 2025. Degrave, Jonas, Federico Felici, Jonas Buchli, Michael Neunert, Brendan Tracey, Francesco Carpanese, Timo Ewalds, Roland Hafner, Abbas Abdolmaleki, Diego de Las Casas et al., “Magnetic Control of Tokamak Plasmas through Deep Reinforcement Learning,” Nature, 2022, 602 (7897), 414–419. Dell, Melissa, “Deep Learning for Economists,” Journal of Economic Literature, 2025, 63 (1), 5–58. Dell’Acqua, Fabrizio, Edward McFowland III, Ethan R Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, Fran¸cois Candelon, and Karim R Lakhani, “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” Harvard Business School Technology & Operations Mgt. Unit Working Paper, 2023, 24 (013). 56

Devlin, Jacob, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018. Doms, Mark, “The Boom and the Bust in Information Technology Investment,” Economic Review-Federal Reserve Bank of San Francisco, 2004, pp. 19–34. Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, “GPTs are GPTs: Labor Market Impact Potential of LLMs,” Science, 2024, 384 (6702), 1306–1308. Felten, Edward W, Manav Raj, and Robert Seamans, “The occupational impact of artificial intelligence: Labor, skills, and polarization,” NYU Stern School of Business, 2019. Feng, Di, Xiao Wei, Lars Rosenbaum, Atsuto Maki, and Klaus Dietmayer, “Deep Active Learning for Efficient Training of a Lidar 3d Object Detector,” in “2019 IEEE Intelligent Vehicles Symposium (IV)” IEEE 2019, pp. 667–674. Ferrucci, David A, “Introduction to “this is watson”,” IBM Journal of Research and Development, 2012, 56 (3.4), 1–1. Filippucci, Francesco, Peter Gal, Cecilia Susanna Jona Lasinio, Alvaro Leandro, Giuseppe Nicoletti et al., “The Impact of Artificial Intelligence on Productivity, Distribution and Growth,” Organisation of Economic Cooperation and Development, 2024. Gao, Mingyang, Suyang Zhou, Wei Gu, Zhi Wu, Haiquan Liu, and Aihua Zhou, “A General Framework for Load Forecasting based on Pretrained Large Language Model,” arXiv preprint arXiv:2406.11336, 2024. Gazzani, Andrea Giovanni and Filippo Natoli, “The Macroeconomic Effects of AI-based Innovation,” SSRN, 2024. Gentzkow, Matthew, Bryan Kelly, and Matt Taddy, “Text as Data,” Journal of Economic Literature, 2019, 57 (3), 535–574. Gerstgrasser, Matthias, Rylan Schaeffer, Apratim Dey, Rafael Rafailov, Henry Sleight, John Hughes, Tomasz Korbak, Rajashree Agrawal, Dhruv Pai, Andrey Gromov et al., “Is Model 57

Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data,” arXiv preprint arXiv:2404.01413, 2024. Giczy, Alexander V, Nicholas A Pairolero, and Andrew A Toole, “Identifying Artificial Intelligence (AI) Invention: A Novel AI Patent Dataset,” The Journal of Technology Transfer, 2022, 47 (2), 476–505. Gill, T Grandon, “Early Expert Systems: Where are They Now?,” MIS quarterly, 1995, pp. 51–81. Goettler, Ronald L and Brett R Gordon, “Does AMD spur Intel to innovate more?,” Journal of Political Economy, 2011, 119 (6), 1141–1200. Goldfarb, Avi, Bledi Taska, and Florenta Teodoridis, “Artificial intelligence in health care? evidence from online job postings,” in “AEA Papers and Proceedings,” Vol. 110 2020, pp. 400–404. , , and , “Could machine learning be a general purpose technology? a comparison of emerging technologies using data from online job postings,” Research Policy, 2023, 52 (1), 104653. Gottweis, Juraj, Wei-Hung Weng, Alexander Daryin, Tao Tu, Anil Palepu, Petar Sirkovic, Artiom Myaskovsky, Felix Weissenberger, Keran Rong, Ryutaro Tanno et al., “Towards an AI Co-scientist,” arXiv preprint arXiv:2502.18864, 2025. Grier, David Alan, When Computers were Human, Princeton University Press, 2007. Griliches, Zvi, “Hybrid Corn: An Exploration in the Economics of Technological Change,” Econometrica, 1957. Gu, Albert and Tri Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752, 2023. Guo, Pengfei, Can Zhao, Dong Yang, Ziyue Xu, Vishwesh Nath, Yucheng Tang, Benjamin Simon, Mason Belue, Stephanie Harmon, Baris Turkbey et al., “Maisi: Medical AI for Synthetic Imaging,” in “2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)” IEEE 2025, pp. 4430–4441. 58

Hall, Brian, “321 Real-world Gen AI Use cases from the World’s Leading Organizations,” Technical Report, Google 2024. Hancock, John T and Taghi M Khoshgoftaar, “Survey on categorical data for neural networks,” Journal of big data, 2020, 7 (1), 28. Handa, Kunal, Alex Tamkin, Miles McCain, Saffron Huang, Esin Durmus, Sarah Heck, Jared Mueller, Jerry Hong, Stuart Ritchie, Tim Belonax et al., “Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations,” arXiv preprint arXiv:2503.04761, 2025. Handa, Sarthak and Jared Sorensen, “Automatically Create Clinical Documentation with Generative AI,” Technical Report, Amazon Web Services 2023. Hardesty, Larry, “The History of Amazon’s Recommendation Algorithm,” Amazon Science, 2019, 22. Hart, Gus LW, Tim Mueller, Cormac Toher, and Stefano Curtarolo, “Machine Learning for Alloys,” Nature Reviews Materials, 2021, 6 (8), 730–755. Haupt, Andreas and Erik Brynjolfsson,“AIShouldNotBeanImitation Game: Centaur Evaluations,” Technical Report 2025. Available at https: //www.andyhaupt.com/assets/papers/Centaur_Evaluations.pdf. Hecht, Jeff, “Boom, Bubble, Bust: The Fiber Optic Mania,” Optics & Photonics News, 2016. Hennessy, John L and David A Patterson, Computer Architecture: A Quantitative Approach, Elsevier, 2011. Hinton, Geoffrey E and Ruslan R Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,” Science, 2006, 313 (5786), 504–507. Hinton, Geoffrey, Oriol Vinyals, and Jeff Dean, “Distilling the Knowledge in a Neural Network,” 2015. 59

Ho, Anson, Tamay Besiroglu, Ege Erdil, David Owen, Robi Rahman, Zifan Carl Guo, David Atkinson, Neil Thompson, and Jaime Sevilla, “Algorithmic Progress in Language Models,” arXiv preprint arXiv:2403.05812, 2024. Hoffmann, Jordan, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark et al., “Training Compute-Optimal Large Language Models,” arXiv preprint arXiv:2203.15556, 2022. Hornback, Andrew, Marteau Benoit, Shaun QY Tan, Kyungbeom Kim, Oankar Patil, Joshua Traynelis, Yuanda Zhu, Felipe Giuste, and May D Wang, “FHIR in Focus: Enabling Biomedical Data Harmonization for Intelligent Healthcare Systems,” Authorea Preprints, 2025. Hulten, Charles R, “Growth accounting with intermediate inputs,” The Review of Economic Studies, 1978, 45 (3), 511–518. Iansiti, Marco and Karim R Lakhani, Competing in the age of AI: Strategy and leadership when algorithms and networks run the world, Harvard Business Press, 2020. Jacobs, Robert A, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton, “Adaptive Mixtures of Local Experts,” Neural Computation, 1991, 3 (1), 79–87. James, Conrad D, James B Aimone, Nadine E Miner, Craig M Vineyard, Fredrick H Rothganger, Kristofor D Carlson, Samuel A Mulder, Timothy J Draelos, Aleksandra Faust, Matthew J Marinella et al., “A Historical Survey of Algorithms and Hardware Architectures for Neural-Inspired and Neuromorphic Computing Applications,” Biologically Inspired Cognitive Architectures, 2017, 19, 49–64. Jiang, Albert Q, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier et al., “Mistral 7B,” arXiv preprint arXiv:2310.06825, 2023. 60

Joshi, Satyadhar, “Generative AI in Investment and Portfolio Management: Comprehensive Review of Current Applications and Future Directions,” Technical Report, preprints.org 2025. Jr., E. Wainright Martin, “Practical Problems of Introducing a Computer,” Business Horizons, 1960, 3 (3), 4–86. jylin04, JackS, Adam Karvonen, and Can, “OthelloGPT Learned a Bag of Heuristics,” LESSWRONG, 2024. Kaldor, Nicholas, “A Model of Economic Growth,” The economic journal, 1957, 67 (268), 591–624. Kalidindi, Surya R, Michael Buzzy, Brad L Boyce, and Remi Dingreville, “Digital Twins for Materials,” Frontiers in Materials, 2022, 9, 818535. Kane, Aidan and Martin Baily, “AI in the Electricity Sector Case Study,” Technical Report, Brookings Institution 2025. and , “AI in the Information Sector Case Study,” Technical Report, Brookings Institution 2025. Kaplan, Jared, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei, “Scaling Laws for Neural Language Models,” arXiv preprint arXiv:2001.08361, 2020. Keohane, Joe, “What News-Writing Bots Mean for the Future of Journalism,” Wired, 2017, 16, 2017. Knight, Will, “Japanese Robotics Giant Gives Its Arms Some Brains,” MIT Technology Review, 2016. Kondo, Illenin O, Logan T Lewis, and Andrea Stella, “Heavy Tailed but not Zipf: Firm and Establishment Size in the United States,” Journal of Applied Econometrics, 2023, 38 (5), 767–785. Korinek, Anton, “Generative AI for Economic Research: Use Cases and Implications for Economists,” Journal of Economic Literature, 2023, 61 (4), 1281–1317. 61

Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E Hinton, “Imagenet classificationwithdeepconvolutionalneuralnetworks,” Advances in neural information processing systems, 2012, 25. Kurzweil, Ray, The Singularity Is Nearer: When We Merge with AI, Random House, 2024. LeCun, Yann, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel, “Backpropagation Applied to Handwritten Zip Code Recognition,” Neural Computation, 1989, 1 (4), 541–551. Li, Kenneth, Aspen K Hopkins, David Bau, Fernanda Vi´egas, Hanspeter Pfister, and Martin Wattenberg, “Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task,” arXiv preprint arXiv:2210.13382, 2022. Li, Yikuan, Hanyin Wang, Halid Z Yerebakan, Yoshihisa Shinagawa, and Yuan Luo, “FHIR-GPT Enhances Health Interoperability with Large Language Models,” medRxiv prepring, 2024, 1 (8), AIcs2300301. Liao, Thomas, Rohan Taori, Inioluwa Deborah Raji, and Ludwig Schmidt, “Are We Learning Yet? A Meta Review of Evaluation Failures across Machine Learning,” in “Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)” 2021. Lino, Giro, “Nvidia GPU Evolution: From GeForce to AI Powerhouse,” girolino.com, 2024. Lipsey, Richard G, Kenneth I Carlaw, and Clifford T Bekar, Economic transformations: general purpose technologies and long-term economic growth, OUP Oxford, 2005. Liu, Guilin, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro,“ImageInpaintingforIrregularHoles using Partial Convolutions,” in “Proceedings of the European conference on computer vision (ECCV)” 2018, pp. 85–100. 62

Liu, Ruibo, Jerry Wei, Fangyu Liu, Chenglei Si, Yanzhe Zhang, Jinmeng Rao, Steven Zheng, Daiyi Peng, Diyi Yang, Denny Zhou et al., “Best practices and lessons learned on synthetic data for language models,” arXiv preprint arXiv:2404.07503, 2024. Liu, Yinqiu, Hongyang Du, Dusit Niyato, Jiawen Kang, Zehui Xiong, Yonggang Wen, and Dong In Kim, “Generative AI in Data Center Networking: Fundamentals, Perspectives, and Case Study,” arXiv preprint, 2024. Lu, Chris, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha, “The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery,” arXiv preprint arXiv:2408.06292, 2024. Lugmayr, Andreas, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool, “Repaint: Inpainting using DenoisingDiffusionProbabilisticModels,” in“ProceedingsoftheIEEE/CVF conference on computer vision and pattern recognition” 2022, pp. 11461– 11471. Markov, Andre˘ı Andreevich, “An example of statistical investigation of the text Eugene Onegin concerning the connection of samples in chains,” Science in Context, 2006, 19 (4), 591–600. Maslej, Nestor, Loredana Fattorini, Raymond Perrault, Vanessa Parli, Anka Reuel, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Juan Carlos Niebles, Yoav Shoham, Russell Wald, and Jack Clark, “Artificial Intelligence Index Report 2024,” 2024. Matson, John, “Fact or Fiction? Lead can be Turned into Gold,” Scientific American, 2014. McKinsey & Company McKinsey & Company, “Banking on Innovation: How ING uses Generative AI to Put People First,” Technical Report 2025. McCulloch, Warren S and Walter Pitts, “A Logical Calculus of the Ideas Immanent in Nervous Activity,” The Bulletin of Mathematical Biophysics, 1943, 5, 115–133. 63

McKinsey, “The state of AI in early 2024: Gen AI adoption spikes and starts to generate value,” McKinsey & Company, 2024. , “The state of AI: How Organizations are Rewiring to Capture Value,” McKinsey & Company, 2025. Mikolov, Tomas, Ilya Sutskever, and Quoc Le, “Learning the meaning behind words,” Google Open Source Blog, 2013, 14. , Kai Chen, Greg Corrado, and Jeffrey Dean, “Efficient Estimation ofWordRepresentationsinVectorSpace,” arXiv preprint arXiv:1301.3781, 2013. Minsky, Marvin, “A Neural-Analogue Calculator Based Upon a Probability Model of Reinforcement,” Technical Report, Harvard University Psychological Laboratories 1952. and Seymour A. Papert, “Perceptrons,” Cambridge, MA: MIT Press, 1969, 6 (318-362), 7. Mnih, Volodymyr, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller, “Playing Atari with Deep Reinforcement Learning,” 2013. Moor, James, “The Dartmouth College Artificial Intelligence Conference: The Next Fifty Years,” AI Magazine, 2006, 27 (4), 87–87. Moore, Gordon E., “Cramming more components onto integrated circuits,” Electronics Magazine, 1965, 4, 114–117. Moore, Gordon E,“Progressindigitalintegratedelectronics,”in“Electron devices meeting,” Vol. 21 Washington, DC 1975, pp. 11–13. Mowery, David C and Nathan Rosenberg, Paths of Innovation: Technological Change in 20th-Century America, Cambridge University Press, 1999. Narayanan, Arvind and Sayash Kapoor, AI Snake Oil: What Artificial Intelligence Can Do, What it Can’t, and How to Tell the Difference, Princeton University Press, 2024. 64

Newell, Allen, John Calman Shaw, and Herbert A Simon, “Elements of a Theory of Human Problem Solving.,” Psychological Review, 1958, 65 (3), 151. Nilsson, Nils J, The quest for artificial intelligence, Cambridge University Press, 2009. Nordhaus, William D, “Two centuries of productivity growth in computing,” The Journal of Economic History, 2007, 67 (1), 128–159. Noy, Shakked and Whitney Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Available at SSRN 4375283, 2023. Odlyzko, Andrew, “The Railway Mania—Fraud, Disappointed Expectations and the Modern Economy,” Journal of the Railway and Canal Historical Society, 2012, 215 (2), 1–16. Olson, Parmy, Supremacy: AI, ChatGPT, and the Race that Will Change the World, St. Martin’s Press, 2024. OpenAI, “Introducing ChatGPT,” 2022. , “GPT-4 Technical Report,” arXiv preprint arXiv:2303.08774, 2023. Ouyang, Long, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray et al., “Training Language Models to Follow Instructions with Human Feedback,” Advances in Neural Information Processing Systems, 2022, 35, 27730–27744. Pairolero, Nicholas A, Alexander V Giczy, Gerard Torres, Tisa Islam Erana, Mark A Finlayson, and Andrew A Toole, “The Artificial Intelligence Patent Dataset (AIPD) 2023 Update,” The Journal of Technology Transfer, 2025, pp. 1–24. Park, Joon Sung, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein, “Generative Agents: Interactive Simulacra of Human Behavior,” in “Proceedings of the 36th AnnualAcmSymposiumonUserInterfaceSoftwareandTechnology”2023, pp. 1–22. 65

Peng, Sida, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer, “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot,” arXiv preprint arXiv:2302.06590, 2023. Peterson, Kevin J, Guoqian Jiang, and Hongfang Liu, “A Corpus- Driven Standardization Framework for Encoding Clinical Problems with HL7 FHIR,” Journal of Biomedical Informatics, 2020, 110, 103541. Pierson, Harry A and Michael S Gashler, “Deep Learning in Robotics: a Review of Recent Research,” Advanced Robotics, 2017, 31 (16), 821–835. Pilz, Konstantin F., Yusuf Mahmood, and Lennart Heim, “AI’s Power Requirements Under Exponential Growth,” RAND Research Report, 2025. Poon, Eric G, Ashish K Jha, Melissa Christino, Melissa M Honour, Rushika Fernandopulle, Blackford Middleton, Joseph Newhouse, Lucian Leape, David W Bates, David Blumenthal et al.,“Assessing the Level of Healthcare Information Technology Adoption in the United States: ASnapshot,” BMC medical informatics and decision making, 2006, 6, 1–9. Pope, Reiner, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean, “Efficiently Scaling Transformer Inference,” Proceedings of Machine Learning and Systems, 2023, 5, 606–624. Porter, Michael E and Scott Stern, “Innovation: location matters,” MIT Sloan Management Review, 2001. Radford, Alec, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever,“LanguageModelsareUnsupervisedMultitask Learners,” OpenAI Blog, 2019. Reed, Scott, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg et al., “A Generalist Agent,” arXiv preprint arXiv:2205.06175, 2022. Richardson, Peter, Ivan Griffin, Catherine Tucker, Dan Smith, Olly Oechsle, Anne Phelan, Michael Rawling, Edward Savory, 66

and Justin Stebbing,“BaricitinibasPotentialTreatmentfor2019-nCoV Acute Respiratory Disease,” The Lancet, 2020, 395 (10223), e30–e31. Romer, Paul M, “Theorigins of endogenousgrowth,” Journal of Economic perspectives, 1994, 8 (1), 3–22. Rosenblatt, Frank, “The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain.,” Psychological review, 1958, 65 (6), 386. Russell, Stuart and Peter Norvig, Artificial Intelligence: A Modern Approach, Prentice Hall, 2009. Saadi, Jana I and Maria C Yang, “Generative Design: Reframing the RoleoftheDesignerinEarly-StageDesignProcess,”Journal of Mechanical Design, 2023, 145 (4), 041411. Sahami, Mehran, Susan Dumais, David Heckerman, and Eric Horvitz, “A Bayesian approach to filtering junk e-mail,” in “Learning for Text Categorization: Papers from the 1998 workshop,” Vol. 62 Citeseer 1998, pp. 98–105. Sahni, Nikhil, George Stein, Rodney Zemmel, and David M Cutler, “The Potential Impact of Artificial Intelligence on Healthcare Spending,” TechnicalReport, NationalBureauofEconomicResearchCambridge, MA, USA: 2023. Sai, Siva, Revant Sai, and Vinay Chamola, “GenerativeAIforIndustry 5.0: AnalyzingtheimpactofChatGPT,DALLE,andothermodels,” IEEE Open Journal of the Communications Society, 2024. Searle, John R, “Minds, Brains, and Programs,” Behavioral and Brain Sciences, 1980, 3 (3), 417–424. Selfridge, Oliver G., “Pandemonium: A Paradigm for Learning,” Technical Report, Mechanisation of Thought Processes: Proceedings of a Sumposium Held at the National Physical Laboratory 1958. Seo, Jaemin, SangKyeun Kim, Azarakhsh Jalalvand, Rory Conlin, Andrew Rothstein, Joseph Abbate, Keith Erickson, Josiah Wai, Ricardo Shousha, and Egemen Kolemen, “Avoiding Fusion Plasma 67

Tearing Instability with Deep Reinforcement Learning,” Nature, 2024, 626 (8000), 746–751. Sergeyuk, Agnia, Yaroslav Golubev, Timofey Bryksin, and Iftekhar Ahmed, “Using AI-based Coding Assistants in practice: State of affairs, perceptions, and ways forward,” Information and Software Technology, 2025, 178, 107610. Serradilla, Oscar, Ekhi Zugasti, Jon Rodriguez, and Urko Zurutuza, “DeepLearningModelsforPredictiveMaintenance: ASurvey, Comparison, Challenges and Prospects,” Applied Intelligence, 2022, 52 (10), 10934–10964. Shani, Inbal and GitHub Staff, “Survey Reveals AI’s Impact on the Developer Experience,” GitHub Blog, 2023. Shannon, Claude Elwood, “A Mathematical Theory of Communication,” The Bell System Technical Journal, 1948, 27 (3), 379–423. Shazeer, Noam, AzaliaMirhoseini, KrzysztofMaziarz, AndyDavis, Quoc V Le, Geoffrey E Hinton, and Jeff Dean, “Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer,” in “Proceedings of the 5th International Conference on Learning Representations (ICLR)” 2017. Shortliffe, Edward H, “Mycin: A Knowledge-Based Computer Program AppliedtoInfectiousDiseases,” in“ProceedingsoftheAnnualSymposium on Computer Application in Medical Care” American Medical Informatics Association 1977, p. 66. Shumailov, Ilia, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson, “The Curse of Recursion: Training on Generated Data Makes Models Forget,” arXiv preprint arXiv:2305.17493, 2023. Simon, Herbert A.,“TheCorporation: WillitbeManagedbyMachines?,” in M Ashen and G. Bach, eds., Management and the Corporations, ABC- CLIO, 1960. Smith, Adam, An Inquiry into the Nature and Causes of the Wealth of Nations, 1st ed., London: W. Strahan and T. Cadell, 1776. 68

Solow, Robert M., “A Contribution to the Theory of Economic Growth,” Quarterly Journal of Economics, 1956, 70 (1), 65–94. Soori, Mohsen, Behrooz Arezoo, and Roza Dastres, “Artificial Intelligence, Machine Learning and Deep Learning in Advanced Robotics, a Review,” Cognitive Robotics, 2023, 3, 54–70. Srihari, Sargur N and Edward J Kuebert,“IntegrationofHand-Written Address Interpretation Technology into the United States Postal Service Remote Computer Reader System,” in “Proceedings of the Fourth International Conference on Document Analysis and Recognition,” Vol. 2 IEEE 1997, pp. 892–896. Stoughton, Jason, “Meet ‘Coscientist,’ Your AI Lab Partner,” Science Matters, 2023. Tamkin, Alex, Miles McCain, Kunal Handa, Esin Durmus, Liane Lovitt, Ankur Rathi, Saffron Huang, Alfred Mountfield, Jerry Hong, Stuart Ritchie et al., “Clio: Privacy-Preserving Insights into Real-World AI Use,” arXiv preprint arXiv:2412.13678, 2024. Tao, Fei, He Zhang, and Chenyuan Zhang, “Advancements and Challenges of Digital Twins in Industry,” Nature Computational Science, 2024, 4 (3), 169–177. Taori, Rohan, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto, “Stanford Alpaca: An Instruction-following Llama Model,” Stanford Center for Research on Foundation Models, 2023. Toner-Rodgers, Aidan, “Artificial Intelligence, Scientific Discovery, and Product Innovation,” arXiv preprint arXiv:2412.17866, 2024. Trajtenberg, Manuel, “AI as the Next GPT: A Political-Economy Perspective,” Technical Report, National Bureau of Economic Research 2018. Turing, Alan, “Computing Machinery and Intelligence,” Mind, 1950, 59 (236), 433–60. 69

Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, L(cid:32) ukasz Kaiser, and Illia Polosukhin, “Attention is All You Need,” Advances in Neural Information Processing Systems, 2017, 30. ˇ ΩCapek Cˇapek, Karel, R.U.R. (Rossum’s Universal Robots), Penguin Clasics, 1920. Villalobos, Pablo, Jaime Sevilla, Lennart Heim, Tamay Besiroglu, Marius Hobbhahn, and Anson Ho, “Will we run out of data? an analysis of the limits of scaling datasets in machine learning,” arXiv preprint arXiv:2211.04325, 2022. Wallace, Richard A, “The Maze Solving Computer,” in “Proceedings of the 1952 ACM National Meeting (Pittsburgh)” 1952, pp. 119–125. Wang, Dashun and Albert-L´aszlo´ Barab´asi, The Science of Science, Cambridge University Press, 2021. Wang, Hongyu, Shuming Ma, Li Dong, Shaohan Huang, Huaijie Wang, Lingxiao Ma, Fan Yang, Ruiping Wang, Yi Wu, and Furu Wei, “Bitnet: Scaling1-bittransformersforlargelanguagemodels,” arXiv preprint arXiv:2310.11453, 2023. Wang, Yizhu, “Banks, Credit Unions Testing AI Models for Underwriting in Credit Cycle,” Technical Report, S&P Global 2023. Webb, Michael, “The impact of artificial intelligence on the labor market,” Available at SSRN 3482150, 2019. Wei, Jason, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” Advances in Neural Information Processing Systems, 2022, 35, 24824–24837. Weiss, Debra Cassens, “JP Morgan Chase Uses Tech to Save 360,000 Hours of Annual Work by Lawyers and Loan Officers,” ABA Journal, 2017. Weizenbaum, Joseph, “ELIZA—A Computer Program for the Study of Natural Language Communication between Man and Machine,” Communications of the ACM, 1966, 9 (1), 36–45. 70

Whitehead, Alfred North, Science and the Modern World: Lowell Lectures, 1925, New American Library, 1925. and Bertrand Russell, Principia Mathematica, Vol. 2, Cambridge University Press, 1927. Wiesinger, Julia, Patrick Marlow, and Vladimir Vuskovic, “Agents,” Technical Report, Google 2024. Wittbold, Kelly A., Carroll Colleen, Marco Iansiti, Haipeng Mark Zhang, and Adam B. Landman, “How Hospitals Are Using AI to Battle Covid-19,” Harvard Business Review, 2020. Wooldridge, Michael, A Brief History of Artificial Intelligence: What it is, Where we are, and Where we are Going, Flatiron Books, 2021. Yan, Yuchen, Jin Jiang, Yang Liu, Yixin Cao, Xin Xu, Mengdi Zhang, Xunliang Cai, and Jian Shao, “Sˆ3cmath: Spontaneous Step- Level Self-Correction Makes Large Language Models better Mathematical Reasoners,” in “Proceedings of the AAAI Conference on Artificial Intelligence,” Vol. 39 2025, pp. 25588–25596. Yegge, Steve,“Stevey’sGooglePlatformsRant,”TechnicalReport,GitHub Gist 2011. Zhou, Kaiyang, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy, “Domain Generalization: A Survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 45 (4), 4396–4415. 71

A Definitions of AI We illustrate the varied use of the term “artificial intelligence” by discussing four influential definitions. Alan Turing devised a broad, conceptual definition—the“Turingtest”—todetermineifasystemwasindistinguishable from humans in 1950. The Dartmouth Project, a seminal meeting for the AI field in 1956, provided a more demanding analytical definition with concrete criteria for what AI must do. In 2020, the National Artificial Intelligence Initiative placed a definition into U.S. law with a rather different set of criteria. And, since 2020, the U.S. Patent and Trademark Office has used an algorithmicapproach: alargelanguagemodel(LLM)supplementedbyreinforcement learning with human feedback (RLHF) that can classify any technology as AI or not based on similarity to descriptions of eight related areas of science and engineering. Importantly, while the sets of technologies meeting these four definitions have substantial overlap, they are far from identical. Thus, it is crucial to stipulate the scope of analysis when discussing “AI” and its economic effects. A.1 Alan Turing Turing (1950) offered this definition of “thinking machines”: A machine can be said to think if it can imitate human responses wellenoughthatahumaninterlocutorcannotreliablydistinguish between the machine and a human. This, of course, is the origin of the “Turing test” which is often referenced to gauge effective artificial intelligence. We judge that for most observers, passing the Turing test would be a sufficient condition for a system to be AI, but it isn’t a necessary condition in current usage. After all, few people would be fooled into thinking the machine learning-based recommendation engines used on social media sites are human. Moreover, mimicking and replacinghumansisnotthesolegoaloftheAIfield. Indeedtheadversesocial consequences of that approach is a concern of some scholars who recommend a focus on using AI to complement human activity instead. Brynjolfsson (2022) observes: We can work on challenges that are easy for machines and hard for humans, rather than hard for machines and easy for humans. 72

The first option offers the opportunity of growing and sharing the economic pie by augmenting the workforce with tools and platforms. The second option risks dividing the economic pie among an ever-smaller number of people by creating automation that displaces ever-more types of workers. Another shortcoming of this definition has been raised by philosophers who have disputed the use of the Turing test to assess whether a machine can think. Searle (1980) offers the counterexample (the “Chinese room argument”) of an individual, ignorant of Chinese, passing the Turing Test by using an instruction manual to connect questions posed in Chinese characters to appropriate responses without understanding their meaning. A.2 The Dartmouth Project The term “artificial intelligence” can be traced to the summer of 1956, when Dartmouth College professor John McCarthy convened the seminal “Summer Research Project on Artificial Intelligence” (Nilsson, 2009). The project proposal stated its premise and objectives and contained an implicit definition of AI (in italics):57 The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. TheDartmouthAIdefinitionisfarbroaderthantheTuringtest. Systems that pass the Turing test would surely be included in the scope of the Dartmouth definition by virtue of using language, provided a system need not satisfy all the criteria at once. The project also envisioned systems forming abstractions, anticipating the flexible models found in neural network systems. That is, the training of these models is agnostic about the analytical structure, rather than calibrating a predetermined functional form. And, the 57He chose the term “Artificial Intelligence” to distinguish the field from “automata theory”—a branch of computer science focused on rule-based mathematical models of computation—and “cybernetics”—a field focused on control systems, feedback, and communication in machines and living things. 73

idea that AI systems will improve themselves hints at the concept of artificial general intelligence. Last, the “solve kinds of problems now reserved for humans” criterion is likely the vernacular definition many casual observers would provide for AI if pressed to do so, though whether present-day observers would reserve the same set of problems for humans as observers in 1956 is unclear. A.3 National Artificial Intelligence Initiative Naturally, as the topic of AI has become a focus of public policy in recent years, the U.S. legal system has required a definition. One is provided in the National Artificial Intelligence Initiative (NAII) (15 USC 9401(3)): The term “artificial intelligence” means a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments. Artificial intelligence systems use machine and human-based inputs to (A) perceive real and virtual environments; (B) abstract such perceptions into models through analysis in an automated manner; and (C) use model inference to formulate options for information or action. Like the Dartmouth definition, this law provides concrete criteria for a system to be AI, but the sets of criteria are rather different. Unlike the Dartmouth definition, the law requires that systems collect observations from the environment. Like the Dartmouth definition, though, the NAII definition requires that AI systems form abstractions and one can loosely compare the “formulate options for information and action” criterion to the “solve problems” criterion in the Dartmouth definition. However, there is no mention of language in the government definition, an important part of the Dartmouth and Turing definitions, nor any mention of self improvement. A.4 The U.S. Patent and Trademark Office The U.S. Patent and Trademark Office (USPTO), identifies patents that “contain” AI for the Artificial Intelligence Patent Dataset. This exercise re- 74

quires a definition that provides an unequivocal declaration for each patent, unlike the other definitions discussed above. To that end, the USPTO identified eight “AI component technologies” and trained a large language model to identify patents referencing those technologies using an iterative supervised learning process where subject experts identified examples of AI and non-AI for each technology and reviewed the AI algorithm determinations for accuracy (Giczy et al., 2022): 1. Knowledge processing: The field of knowledge processing contains methods to represent facts about the world and to derive new facts (or knowledge) from a knowledge base. For example, expert systems generally contain a knowledge base and an inference method to obtain new facts from that knowledge base. 2. Speech: Speech recognition includes methods to understand a sequence of words given an acoustic signal. For example, the noisy channel model is a statistical approach used to identify the most likely sequence of words given verbal input using Bayes’ rule (Russell and Norvig, 2009). 3. AI hardware: The field of AI hardware includes physical hardware designed to implement artificial intelligence software. For example, Google designed the Tensor Processing Unit (TPU) to run neural network algorithms more efficiently. AI hardware may include logic circuitry, memory, video, processors, and solid-state technologies. It may also include embedded software that implements other AI component technologies, such as machine learning algorithms. 4. Evolutionary computation: Evolutionary computation contains a set of computational methods utilizing aspects of nature and, specifically, evolution (Russell and Norvig, 2009). For example, genetic algorithms include methods for selecting algorithm variants through the selection of optimal random mutations by maximizing fitness. 5. Natural language processing: Naturallanguageprocessingcontains methods for understanding and using data encoded in human natural language. For example, language models represent probability distributions of language expressions (Russell and Norvig, 2009). 75

6. Machine learning: The field of machine learning contains a broad class of computational learning models. For example, supervised learning classification models are algorithms that learn to classify observations based on pre-labeled training data. Machine learning includes, among other techniques, neural networks, fuzzy logic, adaptive systems, probabilistic networks, regression, and intelligent searching. 7. Computer vision: The field of computer vision contains methods to extract and understand information from visual input, including images and videos. For example, edge detection identifies the boundaries and borders contained in an image. Additional areas of computer vision include object recognition, manipulation (e.g., transformation, enhancement, or restoration), color processing, and conversion. 8. Planning/control: The field of planning and control contains methods to identify and execute plans to achieve specified goals. Key aspects of planning include representing actions and states of the world, reasoning about the effects of actions, and efficiently searching over potential plans. Modern control theory includes methods to maximize objectives over time (Russell and Norvig, 2009). For example, stochastic optimal control considers dynamic optimization in uncertain environments. Additionally, planning and control includes data systems for administration/management (e.g., managing an organization and its employees, including inventory, workflow, forecasting, and time management), adaptive control systems, and models or simulators of systems. Among the concrete challenges faced by this classification effort are what non-software components of AI systems are included. Tensor Processing Units (TPUs), for example, are chips designed to accelerate AI inference, while graphics processing units (GPUs) were originally designed for graphics rendering but have since been re-engineered to accelerate AI model training. As challenging as this question is, the inclusion of hardware is essential for a definition that is conceptually stable over time with respect to the function of the technology; the division of tasks between hardware and software varies over time within functionally identical IT systems (Hennessy and Patterson, 2011). Further distinction by type of AI would be a welcome refinement. Cockburn et al. (2018) found distinguishing among robotic, symbolic, and neural 76

Table 9: Distinguishing Features of AI Models symbolic vs. connectionist deterministic vs. stochastic discriminative vs. generative network AI technologies as useful for their work on innovation. In light of recent developments in the theory and application of AI, identifying patents as related to generative AI would be useful for research as well. B A Short History of AI The Dartmouth Summer Research Project, in 1956, is often used as a rough marker of the beginning of the AI field, though the scientists in attendance did not share a theory of what the field entailed (Moor, 2006).58 And, foundational work took place before the Dartmouth project. For example, McCulloch and Pitts (1943) had studied the use of artificial neurons, Shannon (1948) had identified Markov Chains as a potential basis for generating new content, and Turing (1950) had introduced a test for machine intelligence (now known as the Turing test) whereby a human attempts to determine if a hidden interlocutor is a computer or another human.59 We sketch subsequent developments in AI theory and application below. As will be apparent, substantial progress on AI preceded the explosion of attention to AI that followed the introduction of ChatGPT in 2022. B.1 Early AI Research Following the Dartmouth project, AI research developed models distinguished along several dimensions (table 9). • Symbolic AI encoded a system of explicit rules in computer programs. For example, Newell et al. (1958) designed a model called “Logic Theorist” that successfully proved 38 theorems from Whitehead and Russell’s Principia Mathematica (Whitehead and Russell, 1927). 58Formoreontheconference, seeNilsson(2009), Wooldridge(2021), andOlson(2024). 59Indeed, Andrey Markov identified language as a use for his mathematical structures as early as 1906 (Markov, 2006). 77

Connectionist AI, in contrast, allowed complex rules to emerge organically, sacrificing interpretability for flexibility (James et al., 2017). The Perceptron model in Rosenblatt (1958), foundational for this approach, was a single neuron, used to combine signals from multiple input channels to classify images. Later models, such as MADALINE (Multiple Adaptive Linear Neuron), combined multiple layers of neurons together. These networks combine input signals into an output signal with the weight given to each input signal evolving during the training process. • Though early models were typically deterministic AI systems, where a given input would consistently produce the same output, stochastic AI models emerged as well where the path taken was not predetermined. For example, SNARC (Stochastic Neural Analogue Reinforcement Calculator), simulated a rat navigating a maze by random experimentation with different paths (Minsky, 1952). • Most early efforts focused on classification—discriminative AI—such as using the Perceptron to label pictures. But, the ambition to create a generative AI system that would respond to questions with an appropriate free-form text response was already present in this era. In 1966, theELIZAchatbotprovidedarudimentarysimulationofaconversation with a psychotherapist. Unlike present-day AI chatbots, ELIZA used a deterministic, symbolic logic approach, relying on pattern matching and word substitution (Weizenbaum, 1966).60 In addition to these theoretical characteristics, applied AI systems are distinguished by the type of training used in their development. Some use reinforcement learning, interacting with the environment to refine the model. Others use predictive learning, where the system is trained in advance of use. Predictive learning primarily took the form of supervised learning in this period, such as when the Perceptron was trained with labeled pictures. However, Selfridge (1958) was a major advance in algorithms forpatternrecognition, whichwasfoundationalforunsupervised learning, where the system develops classification categories without guidance. Early efforts to apply theoretical models were limited by advances in computinghardware. GreaterAIsystemcapabilitytypicallyrequiresalarger 60Strictly speaking, some AI models, such as the “expert systems” described below, are neither generative or discriminative, so our classification scheme is not exhaustive. 78

model,wheresizeismeasuredinthenumberofparameters(fig.6onpage20). Earlymodels,likeTheseus—aroboticmaze-solvingmouse—andthePerceptron— the rudimentary neuron mentioned above—had tens or hundreds of parameters. Recent models, like DALL-E, Llama, and GPT-3 have hundreds of billions of parameters. Moreover, more complicated models typically require larger training datasets (fig. 6 on page 20). Computational requirements for model training and application rise with the size of the model and the size of the training dataset. B.2 Emergence of Practical AI EarlypracticalapplicationsofAIwerefoundinsolvingclassificationproblems in high-volume communication systems. LeCun et al. (1989) developed theLeNetmodeladoptedbytheU.S.PostalServicetoreadhand-writtenZIP codes. The post office was soon reading entire handwritten addresses using AI as well (Srihari and Kuebert, 1997). Another early practical application of AI was the identification of spam email (Sahami et al., 1998). Notwithstanding these advances in the 1980s and 1990s, interest in the connectionist approach, and neural networks in particular, had fallen off with the downbeat assessment of Minsky and Papert (1969). Interest returned with the insights provided by Hinton and Salakhutdinov (2006), who introducedadvancesintrainingmethods(greedylayer-wisepre-training)and efficient use of large datasets (dimensionality reduction). A key innovation that followed soonafter was the convolutional neural network (CNN), an advance in connectionist AI that focused on rapid development of a coarse representation of the image which revealed some features, like edges, but not others. Krizhevsky et al. (2012) demonstrated the power of CNN by leaping ahead of other competitors in the ImageNet Large Scale Visual Recognition Challenge, a benchmark for computer vision, with their AlexNet system. Referring to the characteristics described above, these early image systems were connectionist, deterministic, discriminative, and trained by predictive learning. AlexNet also demonstrated the value of data augmentation by adding mirror images of training pictures to the training set. In the news industry in this period, symbolic models such as Cyborg at the AP and Heliograf at the Washington Post were used to write articles. Unlike present-day generative models, these systems relied on structured data—tagged as sports scores or stock prices, for example—not on models of the language as a whole, making them a kind of proto-generative AI. In 79

Table 10: History of Amazon Demand Forecasting Techniques Year Forecasting Technique 2007 Time-Series Models 2009 Random Forest 2011 Seasonality Models 2013 Sparse Quantile Random Forest 2015 Feed-Forward Networks 2017 Multi-Horizon Quantile Recurrent Forecaster 2020 MQ Transformer Source: Reproduced from Hardesty (2019) 2013, articles on major company financial announcements and sports events were generated with these systems and by 2017, these models were used for expanded coverage of sparsely populated news markets and small companies, and were even used to generate rudimentary news videos (Keohane, 2017). Symbolic AI was put to practical use in this period as well. Expert systems leveraged a large trove of domain-specific information and a set of rules (encoded in an “inference engine”) provided by specialists to provide guidance, such as medical diagnoses (Buchanan and Smith, 1988). Examples include MYCIN, used to diagnose infectious diseases, and IBM’s Watson, deployed in medical and other applications (Shortliffe, 1977; Ferrucci, 2012). Expert systems fell out of favor over time due to their cost of development, limited reliability, and narrow field of application (Gill, 1995). Most importantly, as emphasized by Agrawal, Gans and Goldfarb (2018), AI practicioners soon realized that discriminative AI could be recast as “prediction in the statistical sense of using existing data to fill in missing information” and these models were soon used in a diverse set of prediction problems. (Indeed, these systems are often now referred to as “predictive AI.”) Amazon, for example, first used AI to forecast demand for its products in 2009, then continuously updated its forecasting approach to adopt emerging AI techniques (table 10). Machine learning, another term for this approach to AI, is credited by Amazon and other major IT companies with increasing profitability during this period (Bresnahan, 2019). 80

B.3 Generative AI PublicinterestinAIsurgedinlate2022withtheappearanceofChatGPT, a user interface for a viable genAI system—one that can respond to natural language prompts with human-like (coherent, nuanced, context-specific) responses in the form of text, images, videos, and sounds. The event was the culmination of a roughly 10-year period of advances in large language models (LLMs). LLMs are a mathematical representation of the linguistic relationships among the “tokens” (words, groups of words, and portions of words) found in a “corpus” (set of texts or other media). A key advantage of LLMs is their ability to reduce unstructured text to flexible structural representations that do not rely on a small set of variables specified in advance. Rudimentary early attempts at language models encoded words as long vectors of zeros with a single element—the index assigned to the specific word—marked with a “1” (Hancock and Khoshgoftaar, 2020). In 2013, Google introduced a richer method known as the Word2Vec model (Mikolov et al., 2013a). Words are represented with dense vectors known as embeddings such that the distance between two embeddings reflects the semantic similarity between the represented words. Word2Vec revolutionized many natural language processing (NLP) tasks, such as classification and translation. A limitation of Word2Vec encodings is that a word’s representation is the same regardless of the context. For example, the model will represent the word “bank” with the same embedding whether the input text is “I withdrew money from my bank account” or “I went fishing down at the river bank.” Since Word2Vec assigns fixed embeddings, it cannot distinguish whether “bank” refers to a financial institution or a riverbank, impeding its understanding of the text.61 A major breakthrough in addressing this shortcoming came with the introduction of the Transformer model, an architecture that creates contextaware word representations by efficiently processing the entire input text at once (Vaswani et al., 2017). This architecture is defined by two princi- 61Computer scientists have wrestled with this word sense disambiguation problem since the 1950s. Bar-Hillel (1960) in discussing the prospects for fully automatic high-quality translation, offered this assessment: “What such a suggestion amounts to, if taken seriously, is the requirement that a translation machine should not only be supplied with a dictionary but also with a universal encyclopedia. This is surely utterly chimerical and hardly deserves any further discussion.” It appears we now have such a universal encyclopedia in hand. 81

pal characteristics: the attention mechanism and positional encodings. Theattentionmechanismenablesthemodeltoassign, foreachtoken, varying degrees of relevance to different parts of the input, allowing it to understand context in longer sequences of text. Positional encodings ensure that word order is meaningfully integrated into the model’s processing of inputs. (See the box, “Landmark AI Models: The Transformer,” for more detail.) This advancement changed how machines encoded text, shifting the focus within NLP toward a deeper understanding of language.62 With the advent of the Transformer, genAI flourished. Most notably, text generation improved at a blistering pace in this period, though NVIDIA’s GAN (2017) and DALL-E (2022) were striking advances in image generation; audio generation was dominated by WaveNet (2016), a different architecture, for a time, but eventually the Transformer approach was used for speech generation as well. Especially prominent were a series of models producedbyOpenAI:GPT-2(Radfordetal.,2019), GPT-3(Brownetal.,2020), ChatGPT (OpenAI, 2022), GPT-4 (OpenAI, 2023), and others. Successive models were increasingly human-like in their knowledge and creativity, eventually passing Turing tests due to their coherent and contextually relevant output. 62Particularly important was the introduction of the BERT model the following year(Devlin,2018). Thefinal‘T’inBERTstandsfor‘Transformer’(bidirectionalencoder representations from transformers) 82