The Picking-Winners Problem Is Solved. The Reading Problem Can't Be Programmed.

Yuen Yuen Ang shows how states discover winners under uncertainty. Ken Opalo maps two structural barriers to industrial policy. Taken together, they point to another factor: professional formation.

The World Bank’s rehabilitation of industrial policy as a legitimate development strategy has generated some of the most productive writing on the subject in years. Two responses are worth reading together. Yuen Yuen Ang’s directed improvisation framework reframes what the state can actually do under genuine uncertainty: not pick winners in advance, but create conditions in which winners prove themselves. Ken Opalo’s caution against faddism names what the Bank’s advocacy papers over: the coalition industrial policy requires doesn’t form by mandate, and the Bank’s own staff were not trained to execute what they have now been tasked with promoting. Both are, I think, exactly right. This post builds on them: not to find gaps, but to follow the thread their arguments jointly pull, toward another factor that the standard toolkit does not address.

---

Israel’s Yozma programme is often told as a story about money. In 1993, the government put up $100 million, seeded ten venture funds, and stood back while a private VC industry took root. The detail that matters is smaller and stranger. Each of those ten funds was required to partner with a foreign venture capital firm.¹ The state did not back the start-ups. It backed funds and ensured each fund included someone who already knew how to tell a good bet from a bad one.

That requirement is the whole game. Yuen Yuen Ang’s recent piece draws a foundational distinction between risk (negative outcomes whose probabilities can be calculated) and genuine uncertainty, in which the range of possible outcomes is not yet known. Under genuine uncertainty, she argues, the state cannot pick winners in advance, because nobody can. What it can do is act like a theatrical director: set the stage, issue (constructively) ambiguous signals about what kinds of activity are welcome, watch the pilots, then endorse what works and draw red lines against the worst. Discovery follows proof of concept; it does not precede it.

Her term for this is directed improvisation. It settles the picking problem. The reading problem is a different question: someone still has to be able to interpret the signal.

In Yozma, the state didn’t build that capacity; it rented it from investors who arrived with their judgment intact.

Ang’s e-commerce example runs the same move by another route. In 2000, Beijing issued a directive about “internet information services” (narrowly defined as webpage design and online information provision) with no specific plan for what would follow. They hadn’t foreseen e-commerce. Chinese merchants built cross-border trading platforms regardless, and as goods moved between domestic sellers and foreign buyers, transaction volumes and customs records accumulated automatically.

By 2015, the State Council approved a Zhejiang proposal to formalise and scale what had already proven itself, directing local authorities to “support bold innovation” and “adjust according to the times.” The cross-border data did the assessing; no official had to interpret the signal from scratch. The Zhejiang government’s endorsement followed the proof. It did not manufacture it.

What both cases share is a validation source that sits outside the domestic political apparatus and resists fabrication: international markets in one case, cross-border data trails in the other. Joe Studwell’s insight about export discipline is this: connecting firms to external selection pressure makes deception impossible in a way domestic monitoring never can, because the assessor has no stake in the answer. A domestic ministry can be lobbied, captured, or simply told what it wants to hear. An export market cannot.

So, what happens when that external anchor isn’t there?

The assessor problem travels

The Venture Investment Fund was New Zealand’s version of Yozma: launched in 2002, $100 million, the same fund-of-funds structure, the same buy-back option. No foreign VC partner was required.² By 2018, the return was -3.29 percent.³ None of the original five funds exercised the buy-back. Trade Me, Xero, Vend, and Rocket Lab all emerged during those years. Not one received VIF backing.⁴

Ang’s own field research in Nigeria documents the pattern: the country has produced the world’s second-largest film industry by volume, yet banks won’t lend to entrepreneurs in the sector without physical collateral, and technocrats overlook what doesn’t look ‘strategic.’ Proof of concept is not the problem. What’s missing is the thing Yozma imported and that Chinese e-commerce generated for free: a source of validation external to the domestic ecosystem and hard to fake. There are no foreign co-investors calibrating the sector, and no natural cross-border data flows providing hard evidence of what works. The signals that document Nollywood’s success stay inside the domestic economy, legible to the people already in it and inaccessible to the banks and ministries that would have to act on them.

The grey-to-endorsement sequence (the process Ang calls adaptive policy communication: issuing ambiguous signals first, then endorsing pilots that prove themselves) works when evidence arrives in a form institutions can read. For anything whose success is denominated in domestic reputation rather than foreign exchange, that condition has to be built; for cross-border transactions, it is essentially given.

Export revenue is hard to manipulate because the verification runs through a system with no interest in the answer. Film streaming data inside a domestic ecosystem closes the feedback loop within the same network that generates the signal. Creative value judgments are contested by nature, which means the disagreement itself can be used to block endorsement indefinitely. More rigorous monitoring does not change this.

In practice, the problem is architectural, not technical: what matters is whether the assessor has a stake in the answer, and no refinement of measurement resolves that question. For industries whose success is entirely measured by domestic outcomes — film reputation, food system quality, service performance — no anchor with these properties can be constructed; the signal always closes within the same network that generates it.

The grey-to-endorsement sequence works when the signal validates itself. Solving the picking problem is what brings the reading problem into view: who reads signals that no external party will read for you.

These two problems compound. Where external anchors are absent, assessment depends on officials who can process domestic evidence without being captured by the network that generates it. Where such officials are absent, external anchors are the only reliable input. Strip both away, and the assessment mechanism does not degrade; it ceases to function.

The directed improvisation model solves the reading problem at the firm level: it creates conditions in which firms generate legible evidence without the state having to assess it directly. Professional formation is the same problem one level down: whether the officials receiving that evidence have the internal judgment to process it faithfully, without an external verdict to shelter behind. They are not separate design problems; they are the same problem at different levels.

The judgment forms slowly

Reading signals that no external party will read for you requires officials who can assess domestic evidence under political pressure, hold a view against incumbents who benefit from ambiguity, and enforce what they find against connected interests, without an external verdict to shelter behind. That capacity is professional formation: judgment accumulated within a working cohort through real, consequential decisions. It is what neither technical assistance nor structural design can install.

Capacity is the constraint beneath the one Ken Opalo names. His critique of the World Bank’s rehabilitation of industrial policy starts with faddism: the institution has cycled through policy fashions for decades without accountability, and its new enthusiasm reflects geopolitical pressure from wealthy nations doing their own industrial programmes, not genuine intellectual conversion. Studwell’s line is apt: successful developing-country governments are the ones able to “decide what they need from the catwalk and politely decline what they do not.”⁵ This takes taste. Taste cannot be transmitted in a technical assistance package.

Opalo flags two structural barriers. Industrial policy requires a durable coalition spanning bureaucrats, businesses, and politicians: an alignment that takes years to form and cannot be projectised or contracted out. And the Bank’s own staff were trained to distrust industrial policy; you cannot suddenly task them with executing it. Both problems are deep rather than incidental. A third is professional formation, distinct from coalition in kind: where a coalition aligns interests across sectors, formation builds shared professional expectations within a practitioner cohort, and no political settlement produces the latter.

None of the cases usually cited as models of successful industrial policy is a template; the political conditions that produced each were specific to moments that have passed. What they provide is a timeline: how long it actually took to build institutions capable of running a coherent industrial strategy.

Korea’s Economic Planning Board was created in 1961 and spent the next two decades accumulating the authority and the people to run a coordinated industrial strategy.⁶ Taiwan’s Industrial Technology Research Institute, founded in 1973, spun TSMC and UMC off only after years of building a corps capable of judging which technologies were worth pursuing. Singapore’s Economic Development Board began in 1961 and learned its trade deals one by one.⁷ Ireland’s Industrial Development Authority, reoriented toward targeted FDI attraction from 1958, had built sector-specific networks and judgment to consistently draw electronics and pharmaceutical investment by the mid-1970s.⁸

What these bodies shared was not a mandate (those can be written in a month) but a body of professional judgment accumulated over hundreds of real, consequential decisions, in environments where feedback was genuine, and the stakes were high. The question that follows is not whether the government has the institutional form but whether it has the formation depth: how long the relevant professional community has been doing work whose outcomes were at stake.

Across every case with a detailed institutional record, this formation, the process by which practitioners come to hold standards not as rules to be followed but as something whose violation would be a form of self-betrayal, took between fifteen and twenty-five years. The fastest documented instances ran to fifteen years; none I can identify ran shorter.

No documented case has produced the formation depth required to enforce graduated exit in fewer than fifteen years; most agencies currently tasked with industrial policy execution are substantially younger than this. The mechanism is not skill transfer, which can be accelerated. It happens within a working cohort, through direct observation of how colleagues perform under pressure, across enough consequential decisions for mutual predictability to take hold.

When it doesn’t

What professional formation produces, when it produces anything durable, is not a rulebook but mutual predictability: shared expectations about how each member will act under pressure. It is that predictability, built through real decisions rather than instruction, that makes the standard self-sustaining: violation is costly not because it breaks a rule but because it breaks what the community has learned to expect of you. In institutions where the community’s judgment carries career weight, that cost is material rather than merely social. This is not transmitted across generations; it is accumulated within them, through direct experience that neither cultural inheritance nor training can replicate.

That distinction is invisible under normal operating conditions. The official who holds standards because they are monitored, and the official who holds them because a violation would be a form of self-betrayal, behave identically when the monitoring is functioning. Remove the external check (a loss of political protection, a direct assault on the institution, a withdrawal of donor attention), and they diverge immediately. Standard institutional assessments, staff qualifications, process compliance, and audit results cannot detect this difference because they are conducted under the conditions in which it doesn’t exist. The gap becomes visible only when the external check fails, by which point it is too late to close it.

This test cannot be run in advance, which is why there is no shortcut for running it over time. Better monitoring does not narrow the gap between these two types of officials; it only keeps the gap invisible. What produces an official for whom the external check is redundant is a question of professional formation, not of monitoring design.

The EPB’s analytical corps was not the product of a capacity-building programme; it was the product of two decades of consequential work under conditions in which getting it wrong carried career costs. Korea created those conditions by design. What the corps developed through them was something no design can impose: a standard that no longer needed external conditions to hold it in place.

When Shinjin Motors failed to compete with Hyundai’s Pony in the mid-1970s, EPB officials transferred its credit allocation to Daewoo.⁹ The enforcement was not a matter of following a rule that could not be avoided; it was a matter of meeting what the corps had come to expect of itself. Within a professional community where that expectation carried career weight, the cost of not enforcing was real before it reached any formal hierarchy.

Ajaokuta is what the absence of that community looks like in operation. The plant received billions in capital and a statutory mandate; no corps existed whose accumulated judgment would have made non-performance costly before it reached formal hierarchy. There was no professional expectation to meet, so there was no cost to not meeting it.

What Opalo calls the competency gap, Bank staff trained to distrust industrial policy cannot suddenly execute it, is a professional formation problem, not a training problem. Formation depth is a property of the cohort, not the individual, which is why it cannot be resolved by replacing or retraining the people.

Which stones will hold?

Formation depth determines which instruments a given system can actually enforce — the constraint that comes before instrument choice, not after it. Ang’s directed improvisation model, and the learning-by-doing literature more broadly, converge on demand-led support as the right process. The harder question is whether the system deploying that support can enforce the discipline the instrument presupposes.

If judgment is the binding constraint, the instruments have to be matched to it. In practice, the same tool does opposite things in different hands. Export discipline forced Korean firms to become competitive because the state could credibly pull support from those that didn’t perform, and because the officials enforcing that discipline faced their own incentive structure, one that rewarded enforcement over accommodation. Korea redesigned both fitness functions at once. Attempts to replicate export discipline have typically redesigned the firm-level fitness function without simultaneously redesigning the official-level one.

Across the range of countries where the same instrument has been tried, the direction of effect is not random: it is predictable from a small set of observable properties of the system the instrument enters. Where the state can and does enforce the rule behind the instrument, the tool selects for performance. Where conditionality is absent or capturable, the same instrument entrenches the existing distribution of power. Credit allocation, national champions, and local-content rules: each amplifies whatever the system was already selecting for, rather than replacing it.

In systems where Opalo’s coalition-formation barriers are structural rather than incidental, this dynamic does not merely produce neutral failure. Indonesia’s 1996 national car programme offers a clearly documented instance: subsidies and import duty exemptions routed through a politically connected producer deepened the cronyist network rather than building manufacturing capacity, and the government never enforced the performance benchmarks that would otherwise have been required.¹⁰

Each programme that routes new resource flows through existing distributional networks increases the number of actors with a stake in those networks, deepens the expectations that hold the arrangement in place, and makes the next reform attempt harder. The instruments don’t just fail to build industrial capacity; they consolidate the equilibrium that blocks it.

The officials administering them are responding rationally to the selection environment they actually face: one that rewards accommodation of connected actors, not enforcement against them. The instrument fails because it was designed for an official whose incentive structure does not exist in that system.

The documented pattern is sharper than the standard conditionality-failure account: the instrument does not convert the system’s direction; it amplifies it. In systems selected for performance, the result has been stronger performance; in systems selected for distribution, stronger distribution. The consistent finding is that instrument choice matters less than what the system was already selecting for.

At sufficient formation depth, these tools are viable; where formation depth is insufficient, they will be captured. Get the match wrong, and it scarcely matters which industry was picked, because the instrument will be turned against its own purpose. The stones that hold are the ones the system was already selecting for; the rest will give way the moment weight is placed on them.

The window and the timeline

Professional formation is the variable that determines which stones hold, and also the slowest one. The political window to launch an industrial policy is measured in the lifespan of a reform coalition, a few electoral cycles at most. The formation timeline runs to twenty-five years at minimum: that is how long it takes to build the corps that can discipline connected firms, enforce graduated exit, and distinguish a genuine developmental bet from a rent. The window and the timeline operate at different speeds, and no policy decision closes the gap.

Bolivia under Morales illustrates the mismatch on one axis: thirteen years in office, resource revenues near their peak, explicit industrialisation plans; yet, flagship projects failed, and Bolivia exited the period as structurally hydrocarbon-dependent as it entered it.

In systems where different components change at different speeds, the slowest component constrains the system regardless of what the faster variables do. Political orientation can shift within months because it is a coordination problem: it moves when enough key actors converge on a new expectation, and signals can drive that change quickly. Formal institutional architecture can be redesigned in a few years because it is a design problem: the right decisions, made by the right people, with enough resources, can produce new rules and structures within a political cycle.

Professional formation is different in kind, not just degree: a community problem. Standards are held deeply only by practitioners who have seen each other perform under pressure in real decisions over years, and who have, through direct experience, built both the judgment and the mutual knowledge of who can be trusted to exercise it. No signal produces this, and no design decision installs it. It forms, or it doesn’t, at a pace the system cannot accelerate.

Professional formation and political orientation operate at different intrinsic speeds, and the faster one cannot drag the slower one along.

There is a structural reason this mismatch is stable, one that gives Opalo’s faddism critique its structural depth. The coalition that invests in formation does so for successors, not for itself. Formation investment is a public good for future governments, which means the party that pays and the party that benefits are different actors, separated by electoral time. That is not a coordination failure; it is an incentive problem, and it persists regardless of how much any individual leader cares about long-run development. The governing coalition that would have to invest is judged on outcomes within a political cycle; the return on formation investment arrives a generation later. No amount of political will changes the structure of that misalignment.

The missing factor

Ang’s model solves the picking problem: directed improvisation under genuine uncertainty is the right process, and the state cannot identify winners before the proof is in. What the debate around it has not asked is whether the institutions receiving the proof can read it. Opalo identifies two structural barriers to that reading capacity. Read together, their arguments bring a prior variable into view: not the absence of a framework, but the absence of a professional corps formed through enough consequential work to hold standards when external verification is removed. Professional formation is the prior determinant of instrument outcomes: whether an instrument enforces or extracts depends on whether the agency running it has, through real decisions over a generation, built the judgment to uphold standards without an external verdict to shelter behind.

The real risk in the World Bank’s rehabilitation of industrial policy is not bad technical advice. It is that the new legitimacy licenses governments to launch programmes they lack the professional depth to execute, and then calls the launch itself capacity-building. Standard capacity-building (training programmes, technical advisors, project-cycle skills transfer) has not replaced professional formation in any case on record: it operates at the wrong level — producing portable credentials that travel with individuals rather than the professional identity a community builds through consequential decisions over years — and no extension of programme duration changes this.

The World Bank’s own evaluators documented the pattern in 2008: across a decade of public-sector reform support, technical programmes (financial management, tax administration) regularly improved across borrowing countries, while civil-service performance, the dimension requiring genuine behavioural change, improved in fewer than half.¹¹ Programme duration is not the variable: no externally delivered input produces the community-formation mechanism, however long the delivery runs.

The toolkit external partners currently deploy — technical assistance, project-cycle skills transfer, performance management frameworks — has no instrument calibrated to formation depth; this is why the diagnostic question is also an institutional design challenge. The question neither the report nor most of the commentary asks is not which sectors to target or which instruments to use, but which countries are making the formation investment now: not in instruments, but in the slow work of forming the people who could read the signals and enforce the discipline a generation from today. That is the thing whose answer is already in the record — the age and performance history of the implementing agency — and the more consequential.

It is the question external partners are structurally unable to answer, because the honest answer forecloses the programme cycle as the unit of response. Everything else is downstream of who, in twenty years, will be able to tell a winner from a rent.

---

Notes

1. OECD (2025). Benchmarking government support for venture capital: Israel. https://www.oecd.org/en/publications/benchmarking-government-support-for-venture-capital_82cd3fe1-en/israel_b5c8cc2e-en.html

2. Lerner, J., Moore, D., & Shepherd, S. (2005). A study of New Zealand’s venture capital market and implications for public policy. LECG Limited for the Ministry of Research, Science and Technology. https://thehub.sia.govt.nz/assets/documents/MoRST%20Venture%20Capital%20Study%202005.pdf: New Zealand Government. (2006). Budget 2006 questions and answers: Venture Investment Fund. Beehive. https://www.beehive.govt.nz/sites/default/files/Venture%20Investment%20Fund%20backgrounder.pdf

3. The Treasury (New Zealand). (2018, December 7; released 2019, August). Cabinet Paper DEV-18-SUB-0316: Deepening New Zealand’s early stage capital markets. https://www.treasury.govt.nz/publications/cabinet-paper/dev-18-sub-0316-deepening-new-zealands-early-stage-capital-markets

4. Simpson, R. (n.d.). Rocket fuel. https://rowansimpson.com/essays/venture-capital/

5. Studwell, J. (2026). How Africa Works. Grove Atlantic.

6. Korea Development Institute. (2013). Operation of the Economic Planning Board in the era of high economic growth in Korea. KDI School. https://www.kdi.re.kr/eng/research/reportView?pub_no=13671

7. Economic Development Board. NLB Infopedia, National Library Board Singapore. https://eresources.nlb.gov.sg/infopedia/articles/SIP_2018-01-08_135544.html

8. Barry, F., & Ó Fathartaigh, M. (2015). The Industrial Development Authority, 1949–58: establishment, evolution and expansion of influence. Irish Historical Studies, 39(155), 460–478.

9. Amsden, A. H. (1989). Asia’s next giant: South Korea and late industrialization. Oxford University Press.

10. WTO Dispute Settlement DS54. Indonesia — Certain Measures Affecting the Automobile Industry. https://www.wto.org/english/tratop_e/dispu_e/cases_e/ds54_e.htm

11. Independent Evaluation Group. (2008). Public sector reform: What works and why? An IEG evaluation of World Bank support. Washington, DC: World Bank. ISBN 978-0-8213-7589-1.