Capital Designs Its Successor

The difficulty was keeping the diamond at the proper angle. In the older glass-cutting tool described by Charles Babbage, the stone sat in a conical ferrule, and an apprentice had to learn both where its edge would cut and how to maintain that inclination as the hand moved. Even seven years could leave the work indifferently done. An improved mounting put the diamond in a small piece of squared brass. Someone who knew the work filed one side until, held against a ruler, it guided a clean cut; a swivel between the mounting and its handle allowed the beginner's hand to wander a little without taking the stone with it. Much of the difficulty formerly assigned to the apprentice had been given a mechanical arrangement.¹

Babbage was pleased by the time and glass this could save. There is no reason to begrudge him the pleasure. A long apprenticeship is not made desirable by the quantity of material spoiled during it, and a person who can now cut glass has acquired a useful freedom. Yet the skilled hand has not quite vanished from the account. It prepared the instrument with which another hand can work. What had to be supplied repeatedly in the act of cutting has been supplied in advance, to the mounting, and the purchaser need not acquire the same dexterity to obtain its benefit.

That small difference contains more than one possible history. A glazier might accomplish more with the improved tool, a beginner enter the trade more easily, or an employer find less reason to pay for long experience at this particular operation. Babbage's description does not decide among them. It does show why an invention that preserves the usefulness of knowledge need not preserve the circumstances under which its possessor was paid. The glass still requires the right angle. It no longer requires everyone who cuts it to have learned how to hold that angle unaided.

Knowledge has been taking up residence in equipment for a very long time. The more arresting question now is what happens when equipment participates in devising further equipment—when the thing purchased helps its purchaser investigate what should be made next. Babbage was interested in how much intelligence an undertaking required, how it could be distributed, and what each portion would cost.

An easy rate

The French government had commissioned de Prony to prepare logarithmic and trigonometric tables for a decimal division of the circle. A table could put a long calculation within reach of someone who needed its result but could not have derived it. Producing the table was another matter. Babbage's account turns on the discovery that mental labour could be divided much as Adam Smith had divided the manufacture of pins. A small number of eminent mathematicians selected expressions suitable for extensive numerical work. Another group converted the formulae into the instructions and starting values needed by the calculators, distributed the work and checked the returns. The largest group performed addition and subtraction. Babbage gives their numbers as five or six, seven or eight, and between sixty and eighty respectively. These are the proportions in his description of the enterprise, not a census of intelligence. The group doing most of the calculation was required to understand the least of its mathematical design.²

An elegant formula is not necessarily the cheapest formula to put through a roomful of people. The mathematicians were selecting among ways of expressing the same function, looking for one whose numerical consequences could be obtained through simple operations carried out by many hands. The intellectual achievement included discovering how little intellectual preparation each of those hands would need. What counted as a good method therefore depended partly on the organization that would use it. Mathematics was not merely handed to administration after the important thinking had been finished; the anticipated administration entered the choice of mathematics.

Nor did simplicity at the bottom eliminate the work between the levels. Numbers had to be prepared, instructions made usable, calculations received and checked. Babbage remarks that the second group could verify results without repeating or examining the whole of the lowest group's work. The prior account he quotes also describes two workshops calculating independently, each providing a check on the other. The duplication was part of the economy. A cheap result that could not be trusted would have required a different reckoning.

Babbage presents the people adding and subtracting as essential to the arrangement and their services as cheap to obtain. He moves between the two judgments without apparent difficulty. After observing that the greatest exertions fell to the least knowledgeable group, he writes: “Such labour can always be purchased at an easy rate.” He is looking at the enterprise from the purchaser's side, where the vast quantity of necessary work does not carry the weight of the scarce knowledge required to organize it. Usefulness to an undertaking and the price at which a participant can be replaced are different properties.

We should be wary of making Babbage's three divisions into the three rungs of a ladder. Nothing in this account shows the person performing subtraction being prepared to select the analytical method. Elsewhere in the same chapter he cites parliamentary testimony about the subdivision of watchmaking: a boy apprenticed to one branch might complete his term without being able to work in another. Learning a trade could mean learning a fragment of it. The older workshop was quite capable of economizing on a worker's understanding.³

But Babbage's prospect did not end with the purchase of cheap calculation. Once a calculating engine could replace the third group, he expected the analysts to turn their attention toward simplifying its application. The machine would alter the work above it. He also observed that the first group's intellectual effort, having established the method once, would be smaller on a subsequent occasion. Even before the engine arrived, the initial act of organization had produced something that could be used again. The seventeen folio volumes in his account were one product; a method of producing tables was another.

There is an old temptation to locate human security in the work immediately above the work being mechanized. Someone must understand the whole, devise the next method, repair the instrument or decide what it ought to do. Often this has been a sound description of where additional human work appeared. It is less satisfactory as a guarantee. Babbage himself was already following the effects of machinery upward into the arrangement of intellectual labour, while reserving that arrangement for people. Our question concerns the reservation. An owned system can become useful in some of the work by which the next system is conceived, tested and improved. The intelligence accumulated in a productive asset then enters further acts of production through more than the repetition of a settled instruction.

Marx saw the intellectual powers embodied in machinery becoming a power of its proprietor over the people who worked with it; he also distinguished a capitalist's command of science from the capitalist's personal understanding of it. One can grant that insight without accepting that it exhausts every relation between a tool and its user. The person with Babbage's improved cutter may be more capable than before. A distributional argument that has no place for that gain will misunderstand the attraction of the technology it criticizes.⁴

The blurred designs

In chip design, the relation between equipment and the devising of equipment has become unusually close. AlphaChip places large circuit components, called macros, within chip layouts. It learns a placement policy through repeated trials and feedback, and applies it to further design problems. Conventional software performs other operations in the flow. The 2021 paper describes further optimization and detailed routing after placement, with a third party continuing the production flow.⁵ The firm chooses the programme and commits resources to it; the design system participates in preparing a choice the firm can act upon.⁶

The method entered a field already crowded with automation, in which experienced designers worked with powerful tools. Google reports using AlphaChip in successive generations of its own machine-learning processors. Its comparative advantage over other placement methods remains disputed.⁷

What makes that use consequential here is the relationship among the stages. Computation contributes to a design that, after further engineering and manufacture, becomes equipment for more computation. The contribution is partial. Partial contributions are how most productive undertakings work. A new component does not have to displace every other contributor to change the undertaking's possibilities, any more than a better mounting had to make the diamond or supply the glass.

During training, AlphaChip produces placements and evaluates them; its learned policy is useful insofar as it carries something from those trials into a new problem. Human knowledge enters through the design of the problem, the software, the criteria and the surrounding engineering, rather than solely as examples of choices to imitate. The importance of prior training on relevant blocks makes earlier design problems part of the preparation for later ones. An institution's experience can become material for an instrument it will use again.⁸

There is a revealing reticence in the original paper: some images of the resulting layouts are deliberately blurred because the designs are proprietary. The publication shows a technical achievement while withholding part of the thing achieved. Later releases of software and a pretrained model make the method more accessible, but a release is not the same object as the production organization that generated and used it. The organization has other designs, tools, obligations and decisions to bring to the next attempt. A release can help an outsider begin without supplying everything the releasing institution has accumulated. How much of that accumulation the outsider needs will depend on what they want to attempt.⁹

Who is asked back

An owner who obtains a useful means of searching for further designs acquires more than a past solution to reproduce. The instrument can change the terms on which expertise must be obtained again.

Consider an engineer whose judgment must be sought before an investment can proceed. Besides payment for the completed task, there is a reason for the undertaking to return. Repeated consultation can bring knowledge of customers, acquaintance with collaborators and the opportunity to propose a different project. The engineer may come to understand a difficulty that the original assignment only partly addressed. Something worth attempting later can emerge through work undertaken for another purpose, with people who did not set out to become its collaborators.

An employer need have promised none of this. It is part of what makes the relation valuable without being a separately purchased service. If a design system makes some of the consultation unnecessary, a saving in the cost of the assignment can accompany a change in who encounters the next problem. A profession can grow and its members remain well paid while particular people lose occasions to acquire the knowledge and associations through which they formerly entered further work. Counting the occupation would not tell us what had happened to that opportunity.

This is also why describing a tool as an assistant settles so little about its consequences over time. A team can become more capable with its help while the employer acquires further means of obtaining proposals without returning to the same people. Those who understand the system's failures may become more valuable as its use expands; others may find that what they contributed to making it useful supplies little reason to consult them again. The knowledge can remain effective while the relation through which it was supplied changes.¹⁰

It would be a poor defense of human work to assume that every loss of such a position is a public injury. Making a useful competence easier to acquire can be an advance for people who did not possess it. A reduced return to the incumbent's expertise does not cancel their gain. Education itself can widen opportunity while reducing the premium enjoyed by those who were educated first. The difficulty is to distinguish what has become more widely available from what has merely become unnecessary to the purchaser.¹¹

A cheaper means of investigating an idea can matter most to someone who previously could not afford to investigate it at all. There is a difference between speeding an approved programme and enabling a programme to exist. A small undertaking might never have assembled a design team through hiring; access to a capable tool can make its participants' knowledge of a problem, customer or use worth acting on. They can begin to ask questions whose complementary work was formerly beyond their means. The expertise they lack has become obtainable without requiring them to persuade a large establishment to take up their purpose.

Their independence need not extend to manufacturing every machine they use. Productive independence has rarely meant owning an entire supply chain. What matters is whether access to the necessary complements leaves an undertaking room to develop. A rented tool can help establish a business, and continued dependence on its supplier can become a point at which that business is disciplined. A low price at the beginning tells us little about the cost of changing suppliers after work has accumulated in the arrangement. Access, reliability and the ability to carry the work elsewhere therefore belong inside the account of what cheap intellectual labour makes possible.

The same questions apply to the engineer who leaves. Knowledge acquired while serving one undertaking may become useful in another, but knowing what to attempt and having the means to attempt it are different achievements. Finance, customers, collaborators and contractual obligations remain consequential. The invitation to the next discussion is one way of acquiring a future position; being able to proceed without that invitation is another. A tool that weakens the necessity of the first may help provide the second.

Nor can new productive activity be counted as a new human position merely because it is new. In Acemoglu and Restrepo's analysis, tasks in which labour has a comparative advantage reinstate labour in production; an activity performed entirely by a machine is not that reinstatement. Whether cheaper production and new demand generate enough work elsewhere is a further question. Here the difficulty is how people become able to enter the work that follows.¹²

Learning the rest of the trade

The most efficient way to complete an assignment is not always the way that prepares someone to undertake a more difficult one. Where paid work also provides instruction, the result delivered to the customer accounts for only part of what is happening. A person can learn why a specification took its present form, where a method fails, or which apparently minor variation changes the problem. The assignment ends; some of that understanding remains available for work nobody had yet thought to commission.

An organization can replace an expensive operation without deciding that this education was worthless. The education may never have appeared as a separate expense or purpose. It occurred in doing the work, and changing how the work is done changes the occasions on which it can occur. The loss, where there is one, is easy to overlook when the new result arrives more quickly and is at least as good.

But there is no reason to give unaided difficulty a monopoly on instruction. Babbage's watchmaker could finish an apprenticeship knowing one narrow branch of the trade. An instrument that brings more of the operation within reach can supply material for a wider understanding. A person can learn through the comparison of alternatives that would have taken too long to produce unaided, or through an explanation available at the moment a difficulty arises. Computational assistance can participate in such an education; its presence does not establish dependence.¹³

Much turns on what the organization permits the learner to do with that help. It can make time to follow a result through its reasons, try a different approach and discover why an apparently plausible answer fails. It can also use the same capability to spare experienced people the trouble of explanation, leaving newcomers a faster means of finishing and fewer occasions to understand the decisions within the work. Successful delivery can conceal that difference for a considerable time. It becomes consequential when the person encounters a difficulty for which following the familiar procedure is inadequate.

Understanding is not simply the ability to perform without equipment. The engineer who needs instruments does not thereby lack engineering knowledge. The question is whether the person can recognize a difficulty, pursue it, judge an answer and find another course when the present arrangement fails. Some of that competence can be developed with computational help. Time spent understanding why an alternative fails may contribute little to the present delivery and much to the person who will encounter the next difficulty. The two purposes give different accounts of what the time has cost.

A business has substantial reasons to provide those occasions when it expects to need the resulting expertise. Its interest in competent people is real. But a business can reasonably decline a project that someone trained there has come to regard as worth pursuing. The proposal may be ill judged; it may also demand a change of purpose the owners have no reason to undertake. Their refusal does not decide its value for everyone else. The person who has learned enough to disagree still needs some practical way of putting that judgment to work.

The public interest extends beyond the competence an employer expects to need. Existing undertakings have purposes and obligations; even an excellent business has reasons to leave promising uses of what its people know unexplored. Someone who sees a different use should have a practical chance to pursue it. That does not oblige the employer to finance the attempt. It does mean that the employer's refusal should not acquire the force of a verdict on whether the attempt is worth making. The distinction matters only where there are means of proceeding elsewhere.

The next machine and the next person able to take responsibility for a productive undertaking do not arrive through the same process. Equipment can be improved without renewing the circumstances in which someone learned enough, earned enough or met the right collaborators to propose its replacement. An owner can acquire an increasingly capable means of preparing future investments and dispense with some of the consultation through which people formerly helped prepare them. The machines may be better. The people may remain employed. What diminishes in such an arrangement is their part in originating the uses to which productive capacity will be put.

I would not preserve the difficulty of cutting glass to protect the dignity of a long apprenticeship. The gain is worth having. The political difficulty begins where an owner can obtain what it needs without a contributor's judgment, while the contributor still cannot obtain the means to pursue a different purpose. An invitation to offer suggestions would do little to change this relation. The person needs an opportunity to undertake something the present owner would not have chosen.

Capital has always helped reproduce capital. It now has further means of participating in the choices by which its successors are designed. Those means can equip a new entrant as well as an established firm; their wider availability can put a refusal in its proper place, as one owner's decision about one owner's resources. Where access remains restricted, the loss of an occasion to be consulted can carry much further. An undertaking can become less dependent on its people's judgment while leaving those people dependent on its purposes.

Source notes

Footnotes

1. Charles Babbage, On the Economy of Machinery and Manufactures, third edition, enlarged (London: Charles Knight, 1832), §8, pp. 10–11. The facsimile identifies the edition; Project Gutenberg 4238 supplies searchable text. The reported apprenticeship and the tool's operation are Babbage's observations, not a measured employment outcome. ↩

2. Babbage, §§242–246, pp. 191–196. The numerical organization and checking arrangements are his account. His footnote identifies the earlier Note sur la publication… des grandes tables logarithmiques et trigonométriques de M. de Prony (Didot, 1 December 1820), p. 7. The quoted sentence is on p. 195. The chapter infers the distinction between necessary work and an individual's economic position; it does not reconstruct de Prony's payroll. Babbage's calculating engine appears here as his expectation. ↩

3. Babbage, §§250–251, pp. 201–202, on the purchase of differentiated skill and the watchmaking testimony. His 102 branches are a reported subdivision, not a claim that every watchmaking apprenticeship had this form. ↩

4. Karl Marx, Capital, volume I (1867), chapter 15, §4 and note 23; English translation by Samuel Moore and Edward Aveling, edited by Friedrich Engels (1887). The appropriation of accumulated knowledge and separation of conception from execution are serious antecedents of this chapter's argument. No general theory of distribution is claimed as new here. ↩

5. Azalia Mirhoseini and colleagues, “A graph placement methodology for fast chip design”, Nature 594 (2021), 207–212, especially Fig. 1 and Methods, “Use in a production setting.” The production flow included conventional optimization and subsequent work by a third party. The 2022 correction updates code availability. ↩

6. Chapter 14 supplies the proposed measure of computational participation in successor production and the unresolved growth question; Chapter 19 develops principal and contractual authority; Chapter 23 examines commitments and adjustment; Chapter 27 develops claims on the surplus. The present argument entails neither compulsory collateral nor a necessary settlement asset. Removing their extended discussion here changes neither distinction. ↩

7. Chung-Kuan Cheng and colleagues, “Assessment of Reinforcement Learning for Macro Placement”, ISPD 2023; “An Updated Assessment of Reinforcement Learning for Macro Placement”, author version, §§I, V–VIII. The updated study tests released pretrained weights and stronger baselines. Nature's article history records closure of its investigation on 26 September 2024. Deployment is reported in Anna Goldie and Azalia Mirhoseini's September 2024 developer account. These records establish neither effects on designers' earnings nor universal superiority over competing tools. ↩

8. Anna Goldie and colleagues, 2024 addendum, Nature 634, E10–E11, “Pre-training,” “Initial placement” and “Clarifications to the paper.” Relevant prior blocks, resources and evaluation conditions qualify the transfer claim. The intellectual accumulation discussed in the text includes human-designed problems and machine-generated trials; it is not a claim that the model simply copied human layouts. ↩

9. Mirhoseini and colleagues, Extended Data Fig. 5, identifies the blurred designs as proprietary. The 2024 addendum describes the release of a pretrained checkpoint. The interpretation of the public picture and the wider organization is this chapter's. Neither the blurred images nor differences between a release and its producing organization establish an insurmountable advantage or a monopoly rent. ↩

10. Erik Brynjolfsson, “The Turing Trap”, Dædalus 151(2) (2022), 272–287; inspected author version, pp. 2–5. The relation between substitution, augmentation and bargaining is an antecedent, not a thesis refuted by this chapter. The discussion of consultation, collaborators and subsequent undertakings develops an economic possibility; the AlphaChip records do not establish these effects on its designers. ↩

11. Claudia Goldin and Lawrence Katz, The Race between Education and Technology: The Evolution of U.S. Educational Wage Differentials, 1890 to 2005, NBER Working Paper 12984 (2007), pp. 26–29; the circulated draft identifies itself as chapter 8 of the book published in 2008. This is the version inspected here. Education's widening of opportunity is not equivalent to preserving each incumbent's scarcity premium. ↩

12. Daron Acemoglu and Pascual Restrepo, Artificial Intelligence, Automation and Work, NBER Working Paper 24196 (2018), pp. 1–3, 9–11; David Autor, “Why Are There Still So Many Jobs?”, Journal of Economic Perspectives 29(3) (2015), 3–30, especially pp. 5–8 and 26–28. Task displacement, total employment, absolute labour income and labour's income share are different outcomes. Chapter 20 develops the aggregate adjustment question. The present chapter supplies no aggregate forecast. ↩

13. Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work, NBER Working Paper 31161, November 2023 version, §5.2, pp. 20–21. The authors report suggestive evidence that workers exposed to recommendations handled conversations faster during later outages. The outage evidence is noisy and nonrandom. It concerns conversation duration, not independent mastery, promotion or a long-term career effect. The body retains learning through assistance as a possibility; no result about chip-design careers is inferred. ↩