The Perez Proposition

The support agents could use the assistant's suggested replies, change them or ignore them. The assistant followed the conversation so far, offering possible responses and links to the software company's internal documentation. Its advice went to the agent alone. There was room between producing an answer and sending it: a person could take part of what was offered, supply something else, or decide that none of it would help.

In the study's November 2023 account, access to the assistant raised productivity, measured as issues resolved per hour, by about 14 percent. The gains were uneven, and the study did not calculate a complete return on the investment. It nevertheless found an improvement in work performed, with no deterioration in average resolution rates or surveyed customer satisfaction.¹ The service could retain its purpose and its customers while the means of performing it changed. An employer could then reconsider how much of the surrounding procedure to keep.

Carlota Perez's larger history concerns the spread of these choices into a practical understanding of how production should be organized. A key factor, in her early formulation, combines a visibly falling relative cost with a prospect of ample supply, pervasive uses and interwoven technical and organizational changes. Its importance grows as people cease treating it as an exceptional purchase and begin designing around its availability. A techno-economic paradigm is this wider development, not the name awarded to an impressive machine.²

The price of finishing

Computing has a long history of becoming cheaper, but the history changes with the thing being priced. William Nordhaus compared the cost of performing bundles of computational tasks, reaching back through mechanical and manual calculation and forward to machines available in 2006. The spectacular improvement in his reconstruction was measured through changing performance tests, not by counting transistors and assuming that their multiplication had delivered equivalent useful work. He also identified important complements that the comparison did not adequately include. Software and communications helped determine what the faster machine could accomplish.³

A separate saving occurs when less computation is needed to obtain a given technical result. Anson Ho and colleagues estimated that, across language-model evaluations on WikiText and Penn Treebank from 2012 to 2023, algorithmic progress roughly halved the training computation required for a fixed performance level every eight months. The uncertainty around that estimate is substantial; it measures a historical rate on particular tests, not a promise about every future task. More economical training methods and faster chips can reinforce one another without becoming the same achievement.⁴

Neither history directly prices the customer's resolved problem. That requires following the output into use. A cheap draft may be expensive to check; a more expensive model may need less correction. If the supposed saving is obtained by handing difficult work to someone omitted from the account, the operation has changed its bookkeeper more successfully than its cost. A new way of performing the work may remove a handoff, make a previously troublesome check straightforward, or allow a person to deal with more cases without lowering the quality of the result. The support study gives evidence of a gain at that intermediate level: labour productivity within a specified service. Its value does not vanish because the researchers have not also measured every cost borne by the employer.

To insist that the surrounding workflow remain unchanged would exclude much of the achievement under investigation. Consider a proposed document service that extracts information, produces a draft and directs uncertain passages to a reviewer. This is a possible design, not an account of a particular deployment. Comparing it with an earlier procedure requires accounting for the review, the missed errors and the work needed to keep the service operating. It does not require preserving the old sequence of reading and transcription. The purpose and the quality of the result make the comparison meaningful; the altered means of reaching it may be the discovery.⁵

If checking and repair consume the whole apparent saving, the proposed arrangement has failed its economic claim. That finding can coexist with substantial progress on the underlying model. If, instead, the arrangement provides a better service at an acceptable additional cost, a test concerned only with making the old service cheaper would miss its value. A price decline becomes economically consequential through what people can obtain with it, including things they had previously gone without.

A supply worth designing around

The promise of abundance concerns a future on which someone is prepared to act. It can influence a design before the promised supply is fully built, which is why it can also mislead. A company that reorganizes work around an available service makes a different commitment from one that buys occasional assistance and can do without it. The more thoroughly the new capability enters the operation, the more consequential its terms of provision become.

A recipient can acquire a trained model without reproducing its development. Frontier training need not be the unit of the recipient's commitment. Obtaining routine outputs, financing a new frontier model and finding power for a particular installation are different undertakings. A shortage affecting one can raise the cost of another, but the connection has to be followed. A local queue is not a verdict on the availability of a service across an economy.

Coal required mines, drainage, transport and users who had learned what to do with the delivered fuel. Computation does not become historically exceptional because it, too, requires several things at once. The question is whether those things can be supplied on terms that sustain the activities being built around them. For some uses, a change of supplier or a less demanding model may preserve the service; for others, specialized capacity or operating knowledge may make substitution difficult. Where dependable provision repeatedly fails, businesses cannot sustain the arrangements they have built around it. Counting bottlenecks cannot answer that question without examining their consequences.⁶

Dependable provision can also make control more consequential. A capability can be widely used while access to it remains concentrated. Its price may permit an organization to undertake work that was previously uneconomic even as the supplier acquires influence over how that work continues. Abundance for the purchaser and independence from the supplier are different possibilities. The distinction matters precisely when the service is useful enough to become ordinary.

Beyond the expense account

An invoice is easy to rename. A recurring payment for inference can appear in a budget while the work, authority and dependencies of the firm remain much as they were. Conversely, an organization can change substantially without inventing a new accounting category. A service that makes existing staff faster is one achievement; a service that allows work to be divided differently, performed by different combinations of people and machinery, or offered to customers previously beyond reach opens further choices. These changes have to be found in the operation, not read out of the label on the expense.

The support assistant suggests how close the two achievements can lie. Advice could arrive during an encounter, alongside the managers' weekly feedback sessions described in the study. The documented productivity gain is evidence about that use; it does not establish a subsequent reorganization of training or supervision. The firm nevertheless has a new means of supplying help at the point where it is needed. What it can dispense with, and what it may newly attempt, depend on how it uses that means.

Saving time in one operation can change what other operations are worth attempting. If routine preparation becomes inexpensive, an organization may decide to examine more alternatives rather than dismiss the people who prepared the old ones. It may instead offer a cheaper service, reduce staffing, enlarge review, or find that the apparent economy disappears outside a narrow class of cases. A proposal to broaden service fails if the new cases cannot be handled usefully; a proposal to reduce total cost fails if the work merely reappears elsewhere. The availability of another use does not rescue the proposal that failed.

A rearrangement begins to acquire Perez's wider significance when it becomes available beyond the circumstances that first sustained it. Other organizations must be able to learn from, adapt and improve the method. The common element is unlikely to be an identical chart of departments. It may be a changed judgment about what is worth checking, how much preparation a decision can afford, or which kinds of service can support an undertaking. As complementary skills, equipment and institutions develop around those judgments, an individual experiment can become a direction in which many experiments are made.⁷

One successful workplace cannot supply that history. A catalogue of domains in which models have produced plausible answers cannot supply it either. The broader case requires independently observed uses that remain useful after the work of making them usable is included. Repeated failure outside unusually supported settings would narrow the proposition. Durable success within several industries would strengthen it, even if their departments retained familiar names and their transition refused to meet a five-year deadline.

What a new age would add

Historical computing became vastly less costly under defined performance measures; language-model training became more economical on the tests examined; and a deployed assistant improved labour productivity in an actual service. These findings concern different objects, but together they give a concrete basis for investigating redesign. They do not establish its economy-wide diffusion, and they do not require that diffusion to have occurred before the gains count.

A new techno-economic paradigm would add a stronger claim: that a distinct set of technical and organizational principles is coming to guide production broadly, supported by investment and institutional change. The alternative is that learned systems extend the existing information revolution into work it had not previously absorbed. Much could change under that description. The historical boundary between revolutions need not coincide with the boundary between a person's secure position and its disappearance, or between an undertaking that could not be begun and one that now can.

Finance helps decide which of these attempts can proceed. A crash can interrupt a useful reorganization; a continued flow of credit can keep an unsuccessful one alive. Neither event by itself tells us what the changed operation could accomplish. The terms on which it can be financed may determine whether anyone gets to find out.⁸

The case for computational capital therefore rests on a change in the means available for productive work, not on winning a historical classification. Where an organization can obtain useful cognitive work through an owned capability, it can reconsider what it must assemble through employment, what it can purchase, and what it can undertake at all. Whether those choices yield a saving, a better product or a failed investment depends on the arrangement actually made. Among the work a firm can seek to obtain in this way is the work of conceiving its next investment. It need not wait for a new age before asking the machinery to help decide what to build.

Source notes

Footnotes

1. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, Generative AI at Work, NBER Working Paper 31161, November 2023 version, §§2.2–4.2, pp. 9–15, Tables 2–3. Paper. The figures and locators refer to this working-paper version. The main analysis uses staggered adoption and difference-in-differences, following a small initial randomized pilot; bulk adoption occurred November 2020–February 2021. The preferred levels specification in Table 2, column 3, reports 13.8%, rounded in the narrative. The 5,179-agent dataset includes quality measures for only a subset; effects differ by skill and tenure, with some adverse quality results among the most skilled. The mean findings do not establish uniform improvement, a full investment return or a reorganization of the whole firm. The later discussion of what an employer could investigate is this chapter's inference. ↩

2. Carlota Perez, “Microelectronics, Long Waves and World Structural Change”, World Development 13 (1985), 441–463; inspected author copy, internal pp. 6–8, §2(a). The four conditions are paraphrased; they concern a key factor within an interacting paradigm. See also Technological Revolutions and Financial Capital (Edward Elgar, 2002), introduction, pp. xvii–xix, and chapter 3, pp. 22–26, 31–35. These author-hosted excerpts retain the book's pagination. The manuscript's comparisons of purposes, outcomes and changed requirements are operational adaptations, not tests specified by Perez. The possibility of AI developing within the information-and-communications paradigm is the present interpretation. ↩

3. William D. Nordhaus, “Two Centuries of Productivity Growth in Computing,” Journal of Economic History 67 (2007), 128–159, especially pp. 134–136, 142–147 and Tables 5–8; published article. The inspected reproduction is the published 2007 text, with observations through 2006. Its changing performance bundles and capital/labour costs are not a uniform hardware FLOP-price series. Rates in Tables 7–8 are logarithmic. The account of omitted complements is on p. 136. No single permanent rate of decline or current price is inferred. ↩

4. Anson Ho and colleagues, “Algorithmic Progress in Language Models”, arXiv:2403.05812v1 (9 March 2024), §§2–3.1, 4.2 and Appendix E. The preferred model estimates an effective-compute doubling time of 8.4 months, 95% confidence interval 4.5–14.3 months; the body rounds the equivalent saving in training computation. Data quality and efficiency in using data are not separately identified. Evaluation differences, sparse early data and model specification constrain the estimate. Hardware progress and deployment costs are separate. Jaime Sevilla and colleagues, “Compute Trends Across Three Eras of Machine Learning”, arXiv:2202.05924v2 (9 March 2022), §3, Tables 2–4, investigates a different quantity: computation used in milestone training runs, 1952–2022. Increasing training budgets can coexist with greater efficiency at a fixed result. Neither paper establishes a future rate of improvement. ↩

5. The document service is explicitly hypothetical. The comparison concerns similar purposes and outcome quality while counting changed labour, verification, integration, operation and error consequences. Domain-specific questions remain useful: for support, resolutions, escalation and satisfaction; for code, accepted work, review, defects and maintenance; for documents, completeness, accuracy and correction; for research, reproducible outcomes; for physical operations, interventions and site variation; for consequential decisions, the applicable performance and institutional requirements. This is a research checklist, not a list of demonstrated profitable deployments. A gain in labour productivity is not by itself a fall in total operating cost, and neither alone measures the return on investment. ↩

6. The historical fuel account is in Chapter 4, Ghost Acreage; the distinction between a trained artifact and its production requirements is in Chapter 10, Time Crystallized in Weights. These selected chapters remain unchanged. Appendix C, §§C.2 and C.4, keeps its infrastructure and price observations at their stated dates and scopes. Its records do not by themselves establish either pervasive availability or its failure. The proposed test concerns dependable service for the productive arrangement in question, including practical substitutes; it is not an inference from data-center electricity shares. ↩

7. Perez, Technological Revolutions and Financial Capital, chapter 6, pp. 60–63, distinguishes uneven sectoral and geographic diffusion from synchronized aggregate cycles. The chapter adopts that distinction, not her categorical dismissal on pp. 61–62 of long-term constant-money comparisons. Chapters 2–3 of this volume retain their positive account of economic measurement. The conditions proposed here for testing organizational change belong to this manuscript; they carry no arbitrary rolling deadline. ↩

8. Hyman P. Minsky, “The Financial Instability Hypothesis”, Working Paper 74 (May 1992), pp. 2–4, 6–8, inspected in the institute's reproduction. Minsky follows promised payments, uncertain receipts and the tendency of successful conditions to encourage more fragile financing; he supplies no timetable for a new productive order. This directly inspected account replaces the earlier unlocated citation to the 1986 book. Perez's 2002 introduction, p. xviii, itself distinguishes her technology-and-finance inquiry from Minsky's; chapter 7, pp. 71–73, distinguishes financial and production functions. The separation between continuing operations and existing financial claims is developed in this volume's selected Chapter 23. No crash timetable, compulsory financing reversal or survival of every original firm is made a condition of useful reorganization. ↩