Ordinary Engineers, Not Heroic Inventors

Tim O'Reilly · rss · 2026-07-07 source ↗ Reader 🖥️

practical · mixed — Builds a theoretical historical explanation (Ding's diffusion theory vs. leading-sector theory, evidenced by the Japan/US case and the electrification precedent) in order to argue practically: companies and countries should stop chasing frontier-model capability and instead invest in organizational skill infrastructure and incentive design. The historical/theoretical apparatus is entirely in service of a prescriptive conclusion about corporate and national AI strategy.

Structure

Key terms

• Leading Sector (LS) theory Skill Infrastructure (Ding)
  Ding's name for the conventional account of technological revolutions: new technologies create fast-growing industries (steel, railroads, automobiles, semiconductors), and whichever country/company dominates invention in that industry captures the monopoly profits and economic linkages — 'if you win the leading sector, you win the era.'
• Diffusion theory
  Ding's alternative account: national/corporate economic leadership comes from diffusing a general-purpose technology more quickly and broadly across the whole economy than rivals, not from inventing it. The win goes to whoever embeds the technology into the widest range of ordinary productive work.
• General-purpose technology (GPT)
  A foundational technology (steam engine, electricity, the computer, and by extension AI) that doesn't just create profit in one industry but spreads across the whole economy over decades.
• Skill infrastructure Skill Infrastructure (Ding)
  Ding's term for the education and training systems that widen the pool of people who can actually work with a GPT — the institutions that build engineering skill at scale, standardize good practice, and tie research to industry. O'Reilly extends it to a firm-level analogue: the internal apparatus for spreading skill and compounding what employees learn.
• Ordinary engineers, not heroic inventors
  Ding's (quoted) formulation: when widespread GPT adoption is the priority, the people and institutions that matter are the broad base of practitioners who diffuse the technology into everyday work, not the famous inventors/researchers celebrated by the leading-sector narrative.
• AI-native (vs. AI-assisted) AI-Native Company
  Guido's distinction (quoted): AI-assisted means giving people tools without changing the system around them; AI-native means rebuilding the operating system of the organization — standardized toolchain, explicit rules, a capability ladder, adoption sprints, reusable artifacts, and safe autonomy — so diffusion compounds.
• Sovereign AI
  Redefined here as 'AI designed for diffusion, not just raw increases in capability' — a predictable consequence of a GPT diffusing into different societies, firms and institutions that adapt it to their own needs, rather than a matter of national competition for frontier capability.
• Mechanism design
  Described as 'reverse game theory': starting from the outcome you want and designing the rules of the game (incentives) to produce it — applied here to the problem of getting employees to share AI-learned workflows.

Propositions

• explicitJapan led the world in semiconductors, consumer electronics, and computer hardware in the 1980s yet did not overtake the United States in the subsequent information-revolution economy.conf: high
• explicitLeading Sector (LS) theory holds that whoever dominates invention in a new leading-sector industry captures its monopoly profits and economic linkages, and thereby 'wins the era.'conf: high
• explicitDiffusion theory holds that national and corporate economic leadership comes from diffusing a general-purpose technology more quickly and broadly across ordinary productive work than rivals, not from inventing it.conf: high
  extendsHoward Yu's Post — American AI models went from 70% of usage to 30% in 12 months · p8 — Diffusion theory provides theoretical grounding for the empirically observed pattern that cheap-and-distributed beats best-and-gated.
• explicitThe US kept its lead over Japan in the information-systems revolution by diffusing computing 'up the stack' via software into every sector of the economy, rather than by out-competing in electronic hardware.conf: medium
• explicitToday's AI industry and the national strategy forming around it operate on the (implicit) assumption of Leading Sector theory: whoever has the biggest and best model wins, and everyone else is an also-ran.conf: high
• explicitestablishedGPT diffusion depends on 'skill infrastructure' — education and training systems that widen the pool of people who can work with the technology.conf: high
  supports ←How Organizations Use AI: Evidence From ChatGPT · p4 — Firm-level empirical evidence (SG&A/R&D/software stocks predicting adoption) for Ding's 'skill infrastructure' claim that GPT diffusion depends on accumulated organizational capability, not raw access.
  supports ←The End of Programming as We Know It · p3 — Same author returning to the same Bessen/diffusion-theory mechanism in a later essay; this article is the earlier, fuller statement of the 'skill infrastructure diffuses slowly' argument that the later piece applies specifically to GPT/AI diffusion.
• explicit"It is ordinary engineers, not heroic inventors, who matter" when the priority is widespread GPT adoption rather than invention.conf: high
• explicitDiffusion is an organizational problem, not a technical one, which is why it takes decades to pay off.conf: high
  supportsKunstig intelligens gir neppe en ny, industriell revolusjon · p5 — Confirms the 'decades before broad economic growth' pattern and supplies the organizational (not technical) explanation for why the lag happens.
  supportsSimon Wardley's Post — Industrialisation of ML: nine notes on what to expect · p2 — Same claim in different vocabulary: diffusion / co-evolution of practice with an industrializing technology takes years, not quarters.
• explicitestablishedEarly electrification produced almost no productivity gain because factories bolted new power sources onto old shaft-and-belt layouts; the productivity gains appeared only decades later, once a new generation of engineers redesigned plants around what electricity actually made possible.conf: high
  extends17) Jonas Slørdahl Skjærpe om svinghjuleffekter, liminalfaser og kunsten å ta heis · p10 — Supplies the concrete historical mechanism (Paul David's electrification account) behind the claim that AI needs decades of organizational restructuring, like the tractor or the car.
  supports ←The loudest warning about AI and jobs yet · p8 — The historical electrification-lag finding offers exactly the mechanism Newton says economists find puzzling: labor-market substitution can precede measurable aggregate productivity gains by years, because the gains appear only once work is reorganized around the new technology.
  supports ←AI as an Enterprise Operating System · p13 — Same author's electrification-productivity-lag parallel (early electrification produced no gain until factories were redesigned) is restated here via computing/Solow, applied to AI.
• explicitcorroboratedThe organizational know-how to reorganize work around a new general-purpose technology accumulates gradually, person by person and plant by plant.conf: high
  supports17) Jonas Slørdahl Skjærpe om svinghjuleffekter, liminalfaser og kunsten å ta heis · p12 — Both claim the organizational learning process cannot be skipped or shortcut on the way to real productivity gains.
  supports ←The End of Programming as We Know It · p4 — Both make the same point in different words: the organizational know-how/workforce-wide skill needed to realize a new technology's benefit accumulates gradually and must be broad, not just held by a few early adopters.
• explicitThe rapid pace of frontier-model improvement makes organizational skill infrastructure more valuable, not less, because it is the only asset that survives each model generation.conf: high
• explicitdevelopedEnterprise AI strategies built around procurement (choosing a vendor/model) or a moonshot lab/demo treat AI as a sector to be won, which misdirects effort away from where the durable value actually lies.conf: high
  extendsWe've Stopped Arguing About Whether AI Changes Software · p13 — Diffusion theory explains why a compounding adoption gap becomes a permanent structural moat rather than a closable lag.
  extends ←How Leaders Can Use AI to Solve Real Business Problems · p2 — O'Reilly's critique of procurement/moonshot-driven AI strategy gets a concrete mechanism here: labs' debt-driven need to sell urgency is why 'choosing a vendor' framings misdirect effort.
• explicitdevelopedBuilding an 'AI-native' operating system (versus merely giving people AI-assisted tools) requires standardizing the toolchain, writing down clear usage rules, building a capability/maturity ladder, running adoption sprints, packaging learning into reusable artifacts, and making autonomy safe via sandboxing.conf: high
  extends ←AI as an Enterprise Operating System · p2 — O'Reilly's own prior piece states the abstract checklist for an 'AI-native operating system' (standardized toolchain, usage rules, capability/maturity ladder, adoption sprints, reusable artifacts, sandboxing); this piece is the concrete Trail of Bits case study instantiating that exact list.
• explicitcorroboratedEffective enterprise AI transformation needs leadership that gets its own hands dirty, a 'lab' that turns individual discoveries into shared tools, and 'the crowd' whose daily work generates most applied discoveries.conf: high
  supports ←Model the Transformation You Expect Employees to Deliver · p8 — Both claims converge on the same mechanism: oreilly-ordinary-engineers holds that effective enterprise AI transformation needs leadership that 'gets its own hands dirty'; this article's DBS case (Gupta visiting branches incognito, entering the DeepRacer competition himself) is a concrete, non-AI-specific instance of the same hands-on-leadership mechanism driving transformation success.
• explicitInside a firm (unlike a nation), diffusion collides with incentives: an employee who shares a workflow that automates part of her own job may be acting against her own career interest under status/indispensability-based incentive structures.conf: high
• explicitcorroboratedThe binding constraint on enterprise AI adoption is not raw employee skill but whether an organization's incentives make sharing what you've learned raise your status rather than lower it.conf: high
  supportsDORA Community Discussion - Operationalizing DORA within a large organization · p12 — Skill infrastructure and incentive design are the 'strong foundation' without which AI amplifies chaos instead of performance.
  extends ←AI as an Enterprise Operating System · p14 — Ordinary-engineers argues the binding constraint on enterprise AI diffusion is whether incentives reward sharing what you've learned; this piece adds that at the industry level, firms that have solved this are themselves not sharing, extending the incentive problem up a level.
  supports ←This Week in AI: Chips, Checks, and Changing Jobs · p8 — The SAP-vs-IKEA contrast is a concrete, current illustration of the production-viability/ordinary-engineers thesis that the binding constraint on enterprise AI adoption is whether an organization's incentives make sharing what's been learned raise or lower an employee's status, not raw employee skill or which talent-acquisition path (external hiring vs. internal retraining) is chosen.
• explicit'Sovereign AI' is best understood as AI designed for diffusion — adapted by different societies, firms, and institutions to their own needs — rather than as national competition for frontier capability.conf: high
• explicitThe electrification analogy has a limit worth heeding: the unit-drive revolution decentralized the application of power (many small motors) while generation remained centralized (the grid) — a pattern now straining under AI data centers' demand.conf: high
• implicitA patient, adaptive diffusion strategy beats a fast, flashy invention-race strategy over the long run of a technological transition — for nations and for companies.conf: medium

Arguments

Narratives

• Japan's 1980s dominance in semiconductors, consumer electronics, and hardware, followed by its failure to overtake the US in the information revolution.
  → The essay's evidentiary backbone: supports diffusion theory over Leading Sector theory as the explanation for how national economic power actually shifts.
• Paul David's 'Dynamo and the Computer' account of early factory electrification — motors bolted onto old shaft-and-belt layouts producing no productivity gain, until a new generation of engineers redesigned plants around what electricity made possible.
  → Supports the claim that diffusion is organizational rather than technical, and grounds the essay's later speculation about centralized vs. decentralized AI compute (grid vs. unit-drive).
• James Bessen's 'Learning by Doing' and Arthur Herman's 'Freedom's Forge' (Bill Knudsen's shop-floor-to-CEO rise and the WWII industrialization of US auto manufacturing into defense production).
  → Supports the claim that organizational know-how accumulates bottom-up, person by person, team by team — not from a central plan or a single heroic figure.
• Dan Guido's account of making Trail of Bits 'AI native,' including its six-step playbook (standardize toolchain, write down rules, build a capability ladder, run adoption sprints, package learning, make autonomy safe) and its specific tool-use rules (e.g., Cursor barred from client code).
  → Illustrates concretely what firm-level 'skill infrastructure' looks like in practice, and supports the claim that AI-native requires system redesign, not just tool provision.
• DJ Patil's account of the unified, clean data infrastructure ('the tidy house') that let Devoted Health move quickly once AI arrived.
  → Supports the claim that boring infrastructure work (data organization) is itself part of skill infrastructure and a precondition for fast AI diffusion.
• Historical railroad-gauge incompatibility as a cautionary analogy for AI standards.
  → Supports the argument for interoperable protocols as a middle layer between one universal AI system and fragmented incompatible ones.
• The Unix/Bell Labs research-community history of shared software development despite AT&T's proprietary licensing.
  → Supports the claim that open source's real driver of collaborative diffusion is modular, protocol-centric architecture, not licensing per se.

Questions addressed

• ✅ Why did Japan, despite leading semiconductors, consumer electronics, and hardware in the 1980s, not overtake the US in the information revolution that followed?
  Diffusion theory: the US won not by competing in the leading sector (hardware) but by diffusing computing more broadly and quickly via software across the whole economy.
• ✅ What actually drives national (and corporate) economic leadership through a technological revolution — invention or diffusion?
  Diffusion, mediated by skill infrastructure, not invention or capability leadership in a leading sector.
• ✅ Why does technology diffusion take decades rather than happening as soon as the technology becomes available?
  Because it is an organizational reorganization problem (redesigning work around what the technology makes possible), not a technical availability problem — illustrated via electrification.
• ✅ What should companies do differently in their AI strategy given diffusion theory?
  Build firm-level skill infrastructure / an 'AI-native operating system' (Guido's six steps, Mollick's leadership/lab/crowd model) instead of treating AI strategy as procurement or a moonshot lab.
• ⭕ How does a firm solve the incentive problem where sharing an AI-learned workflow can threaten the sharer's own job security?
  Named as a mechanism-design problem and gestured at ('reward sharing so it raises status rather than lowers it'), but no concrete compensation, promotion, or governance mechanism is specified.
• ⭕ Should AI compute and models be centralized (frontier labs, big models via API) or decentralized (many small, rightsized, possibly open-source models)?
  Raised via the electrification grid/unit-drive analogy and left as an open design choice ('we should design for a future of multiple AIs'), not resolved.
• ⭕ What does 'sovereign AI' actually mean once diffusion, not frontier capability, is understood as the real game — and does this reframing change how AI policy debates are currently conducted?
  Redefined conceptually (AI designed for diffusion, adapted to different societies) but the essay does not address whether or how this reframing would change the actual national AI policy and arms-race discourse it criticizes.

Gaps

• incompleteThe essay names the firm-level incentive problem (sharing a workflow that automates your own job can be self-defeating) and invokes 'mechanism design' as the fix, but never specifies a concrete mechanism — no compensation model, promotion criterion, or governance structure is proposed beyond 'reward sharing so it raises status.'
  Read next: Mechanism-design literature on incentive design inside firms (e.g., Milgrom & Roberts); the corpus's own psychological-costs-of-AI-adoption and status-theater material for how real organizations have (or haven't) solved this in practice.
  closed byAI as an Enterprise Operating System — That piece names the firm-level incentive problem ('sharing a workflow that automates your own job can be self-defeating') and invokes mechanism design as the fix but specifies no concrete mechanism. This piece supplies one: Trail of Bits ties AI capability directly to its existing performance-review ladder (redefining the top rung as 'invents new ways to work,' not 'uses AI most'), gives skills-repository contributors explicit credit for encoding their expertise, and tracks a cross-team metric (what fraction of your data/workflows you made reusable for others) — a compensation-adjacent, promotion-linked mechanism where none was previously specified.
• uninformedDing's Leading-Sector-theory-vs-diffusion-theory dichotomy is presented as though it were the field's main framework, without engaging the substantial pre-existing literature on general-purpose technology diffusion (Bresnahan & Trajtenberg's original GPT economics) or classical diffusion-of-innovation theory (Rogers), leaving unclear how novel Ding's contribution actually is relative to this scholarship.
  Read next: Bresnahan & Trajtenberg, 'General Purpose Technologies: 'Engines of Growth'?'; Everett Rogers, Diffusion of Innovations.
• illogicalA single historical case comparison (Japan vs. the US, 1980s-1990s) is used to license a general theory that is then applied to firms and to the current AI moment with little qualification — but this one comparison cannot by itself rule out Japan's well-documented 1990s asset-price collapse, keiretsu financing structure, and demographic decline as competing explanations for its stagnation relative to diffusion theory.
  Read next: Literature on Japan's 'lost decades' (asset bubble collapse, structural/demographic explanations) to test whether diffusion theory survives as an explanation once the standard macroeconomic account is on the table.
• incompleteThe argument for decentralized, rightsized AI models over centralized frontier models is made entirely by historical analogy (electrification's unit-drive vs. grid) but never engages current model-training economics — the capital intensity and returns-to-scale that presently favor centralization — which could make the electrification precedent misleading for AI specifically.
  Read next: The corpus's compute-infrastructure and ai-unit-economics material on why frontier-model economics currently favor centralization, to test whether the electrification analogy actually holds for AI.
• incompleteThe 'ordinary engineers, not heroic inventors' thesis assumes a steady supply of trainable ordinary engineers, but doesn't address the corpus's recurring concern that the traditional on-ramp for junior engineers (writing code) is itself being displaced by agents — if that on-ramp collapses, it's unclear where the next generation of 'ordinary engineers' who do the diffusing work comes from.
  Read next: The corpus's ai-labor-disruption and systemic-apprenticeship material (e.g., extract-stopped-arguing-whether-ai-changes-software#p8 on the disappearing entry-level rung) for the supply-side half of this diffusion story.
  closed byWhy Mentoring Matters More in the AI Era — Directly answers the supply-side worry about who trains the next generation once the organic on-ramp (routine task repetition) collapses: intentionally engineered mentoring, not organic exposure, becomes the substitute on-ramp.

Closes gaps elsewhere

• Why Enterprise AI Has a Leadership Problem · g2 — Ding's 'skill infrastructure' argument — that GPT diffusion depends on education/training systems widening the pool of capable practitioners, not on capital deployed to the technology itself — makes the case for why an infra-heavy spending split is backwards, directly answering the gap's call for infrastructure-first vs. people-first sequencing literature.
• Kunstig intelligens gir neppe en ny, industriell revolusjon · g3 — Paul David's 'Dynamo and the Computer' account (cited here in full) supplies one of the concrete diffusion-lag mechanisms the gap says is missing from the productivity-paradox debate: electrification's payoff was delayed because factories kept old plant layouts, not because the technology itself was weak.
• Please stop the AI Confidence Theater · g4 — The LS-theory/diffusion-theory distinction directly names the tension the gap flags: demonstrated capability ('AI really can do crazy stuff') and diffused, reliable value are different axes, and Japan's hardware dominance without economic overtake is a historical case where capability leadership failed to translate into economic leadership.
• We've Stopped Arguing About Whether AI Changes Software · g5 — Diffusion theory (vs. leading-sector theory), evidenced by the Japan/US case, gives the mechanism for why a compounding adoption gap becomes a structural moat rather than a closable lag — it's decades of skill-infrastructure accumulation that laggards cannot fast-forward, not flywheel logic alone.

Related documents

• Howard Yu's Post — American AI models went from 70% of usage to 30% in 12 months
• Simon Wardley's Post — Industrialisation of ML: nine notes on what to expect
• Kunstig intelligens gir neppe en ny, industriell revolusjon
• Nils Brede Moe's Post
  Code Was Never the Bottleneck Learning Culture & Slack Technology Diffusion & Transformation Curves
• 9) Nils Brede Moe om AI-paradokset, autonomi, læringskultur og fremtidens team
  Code Was Never the Bottleneck Enterprise AI Adoption & Readiness Learning Culture & Slack
• How Organizations Use AI: Evidence From ChatGPT
  Enterprise AI Adoption & Readiness Skill Infrastructure (Ding) Technology Diffusion & Transformation Curves
• AI as an Enterprise Operating System
  AI-Native Company Enterprise AI Adoption & Readiness Technology Diffusion & Transformation Curves
• The End of Programming as We Know It
  Code Was Never the Bottleneck Skill Infrastructure (Ding) Technology Diffusion & Transformation Curves
• Measuring AI impact, assessing readiness, and new data trends
  Code Was Never the Bottleneck Enterprise AI Adoption & Readiness
• Finding Comfort in the Uncertainty
  Code Was Never the Bottleneck Enterprise AI Adoption & Readiness