Ordinary Engineers, Not Heroic Inventors

In the 1980s, Japan led the world in semiconductors, consumer electronics, and computer hardware, the industries everyone assumed would decide the next phase of economic power. Japan won them and still did not overtake the United States in the information revolution that followed. Jeff Ding, a political scientist at George Washington University, opens his book Technology and the Rise of Great Powers with the history of the first and second industrial revolutions and the third, the information revolution. The explanation he gives for who wins and who loses applies to companies as well as it does to nations, and very much to the current trajectory of AI.

Ding contrasts two theories of how technological revolutions reshape economic power. The conventional one he calls the leading sector model, or LS theory. It goes like this: New technologies create fast-growing new industries like steel and railroads and automobiles and semiconductors, and the country that dominates invention in those sectors captures the monopoly profits and the upstream and downstream economic linkages that come with them. As the story goes, if you win the leading sector, you win the era. Britain won in the first industrial revolution through its mastery of steam power, and then was surpassed by the US in the second through its leadership in electrification, the internal combustion engine, and mass manufacturing.

This is pretty clearly the working hypothesis of today’s AI industry and the national strategy that is forming around that industry. The company and the country with the biggest and best models wins. Everyone else is an also-ran.

Ding offers another explanation, which he calls diffusion theory. He points out that general-purpose technologies, foundational ones like the steam engine, electricity, and the computer, don’t just create massive profits and productivity gains in a single industry but instead spread across the whole economy. National economic leadership comes not from inventing the new sector but from diffusing the general-purpose technology more quickly and more broadly than your rivals. This happens over decades. The win goes to whoever most successfully embeds the technology into a wide range of ordinary productive work. This is how the US kept its lead over Japan rather than being surpassed by it.

This is obviously aligned with the thinking of Arvind Narayanan and Sayash Kapoor in “AI as Normal Technology,” which Ding cites in his book.

A big part of what enables diffusion is what Ding calls skill infrastructure, the education and training systems that widen the pool of people who can actually work with the technology. When the priority is widespread adoption rather than invention, he argues, the institutions that matter are the ones that build engineering skill at scale, standardize good practice, and tie research to industry. He writes:

GPT diffusion theory highlights the importance of GPT [General Purpose Technology] skill infrastructure. Education and training systems that widen the pool of engineering skills and knowledge linked to a GPT. When widespread adoption of GPTs is the priority, it is ordinary engineers, not heroic inventors, who matter.

Music to my ears, as it should be to yours: “It is ordinary engineers, not heroic inventors, who matter.”

That is not how the current AI narrative goes. Everyone is fixated on the labs, the frontier models, and the most famous researchers. And that fixation shapes enterprise strategy. Inside many companies AI strategy is a procurement decision: Which model and which vendor and which flagship tool should we choose? Or it’s a moonshot to stand up a lab and build an impressive demo and hire your own famous developer. Both approaches treat AI as a sector to be won. Ding’s argument is that the breakthrough sector itself is not where the long-term value for national power lives. And I believe that the same applies to corporate success. The value is in how widely and how well the technology gets embedded into the work of the people you already employ. The company that puts AI to work in finance and support and legal and sales and operations, across every unglamorous process, as well as in product and engineering, outperforms its competitors and drives its industry forward.

Diffusion is organizational, not technical

The reason diffusion takes a long time is that it is an organizational problem and not a technical one. In his oft-cited 1990 paper “The Dynamo and the Computer,” Paul David answered a quip from Robert Solow that you could “see computers everywhere except in the productivity statistics” by looking at the history of electrification, and more specifically, electric motors. When factories first electrified, they bolted a giant electric motor where the steam engine used to be and kept driving the same shafts and belts through the same Rube Goldberg system. Productivity barely moved.

MACHINE SHOP NORTH/NORTHEAST INCLUDING OVERHEAD LINE SHAFTING. MOSTLY BELT DRIVEN WITH ONE ROPE DRIVEN LATHE IN MIDDLE GROUND. POWER COMES FROM KNIGHT TURBINE ON FAR WALL. This image is available from the United States Library of Congress’s Prints and Photographs division under the digital ID hhh.ca2269. Public Domain.

The gains came decades later, when a new generation of entrepreneurs, factory architects, and electrical engineers redesigned the plant around what electricity actually made possible, with many small motors each driving its own machine and the factory floor laid out for the flow of work.

David’s account has since become a paradigmatic example of how technology transformation actually works. This historical analogy suggests that the future might not be ever bigger and smarter centralized AI models but a decentralized network of AI rightsized for thousands or millions of specialized tasks. Yes, there will still be big centralized AI dynamos somewhere, but most of the action will be with smaller (perhaps open source) models distributed throughout the economy.

But there’s more to the story than right-sizing the technology so that it can fit into specialized tasks. The know-how to reorganize work around it had to be built up one person and one plant at a time. This gradual, bottom-up growth of knowledge about how to apply a new technology is also the point of one of my favorite books about the first industrial revolution, James Bessen’s Learning by Doing. It’s also one of the key messages from Arthur Herman’s Freedom’s Forge, which tells the story of the rapid military industrialization of the US in response to the challenges of World War II. (This story may be newly relevant today as AI and drones transform modern warfare.) Herman called out Bill Knudsen’s bottom-up knowledge of the industry as a critical element in his success transforming the auto industry into a defense powerhouse. (Knudsen was the CEO of General Motors, but he had risen up the ranks from the shop floor.)

That is also the whole story of enterprise AI right now. The latest and greatest model is widely available. What takes time to develop is the organizational know-how to redesign work around it. Most of that know-how does not live in the labs that trained the model. It lives in ordinary practitioners, and it accumulates the way David and Bessen and Ding have described, person by person and team by team, as people work out what the technology is good for in the specific context of their own industry and their own jobs.

What skill infrastructure looks like inside a company

Ding’s national version of GPT skill infrastructure is engineering education, standardized best practice, and strong links between universities and industry. My firm-level version of his vision is the internal apparatus for spreading skill and compounding what people learn. The problem with most enterprise AI transformation programs is that they treat AI as a subject to be taught rather than a capability to be built. Training is part of it, but only part. The harder part is the set of mechanisms that apply AI to the actual problems of the business, then capture each new discovery and turn it into something the whole organization can use, so that learning compounds instead of hiding away in a thousand private workflows.

In “The End of Programming as We Know It,” I made the case that AI expands who can build rather than replacing the people who build today. This means that a company’s best source of applied R&D is the everyday experimentation of the people it already has. The job is to make that experimentation visible, shareable, and rewarded. It is also the framework we are building into O’Reilly’s enterprise AI transformation programs.

We base our ideas about effective AI transformation in part on ideas we’ve taken from Wharton business school professor and author Ethan Mollick and from Dan Guido, the CEO of AI security firm Trail of Bits.

Join Dan Guido and Tim online at the Live with Tim O’Reilly event taking place on July 9. You can register here.

Mollick suggests solving the enterprise transformation problem takes three things: leadership that not only sets the conditions and incentives but gives a good example by getting their own hands dirty with AI; a lab that turns individual discoveries into tools everyone can use; and the crowd, meaning everyone else, whose daily work is where most applied discoveries actually happen. This is a great way to think about applied corporate AI adoption.

Guido adds a number of other elements to AI transformation strategy as we conceive it at O’Reilly. As he put it in his essay “How We Made Trail of Bits AI Native (So Far)”: “AI works. Most companies are using it wrong. They give people tools without changing the system. That’s the gap between AI-assisted and AI-native. One is a tool, the other is an operating system.” To build that “operating system,” he suggests that a company must:

Standardize its toolchain. This step seems boring and perhaps even unnecessarily restrictive but according to Guido, without a shared standard across an enterprise, you get zero organizational leverage. While experimentation is encouraged and different departments may have different tools, it’s important to constrain the possibilities so that you don’t get a sprawling set of incompatible workflows. That does not mean that the toolchain becomes fixed, just that organizational discipline is important. New capabilities and tools appear at a furious pace. A key corporate capability thus becomes how to evaluate and select tools at enterprise scale as well as how to govern the toolchain over time as the ecosystem evolves.

Write down the rules. When large language models were new, enterprise AI handbooks were full of warnings: Watch out for hallucinations. Watch out for putting in PII or proprietary company data. Beware of copyright infringement. Check and compensate for bias. And so on and on and on. As Mollick noted, such handbooks often discouraged adoption. Guido simply argues for clarity: what tools are approved, especially for sensitive data. For example, among their rules at Trail of Bits:  “Cursor can’t be used on client code (except blockchain engagements; use Claude Code or Continue.dev instead). Meeting recorders are disallowed for client meetings conducted under legal privilege.” He notes, “The handbook doesn’t just list what’s approved. It explains the risk model behind each decision, so people understand why….Once you have policy, you can safely push harder on adoption.”

Build a capability ladder. Every company needs an “AI maturity matrix” to help employees understand where they are in their AI journey and measure their progress. This is not an exhaustive list of tools and techniques to master. The spine of the Trail of Bits maturity matrix is not specific technical skills but the pathway from resistance or lack of engagement (stage 0) to comfort with using a job-relevant set of AI tools (stage 1), to proactively seeking out

[truncated for AI cost control]