On the sun-baked plains of the American Midwest in 1892, a revolution was loudly sputtering to life: the tractor, an engine which signaled the end of the era of animal power and the beginning of the age of machine power. This machine was not just a piece of equipment; the tractor was a manifestation of an exponential shift in energy density, from animal metabolism to coal burning, empowered by discoveries in thermodynamics. But diffusion of the tractor, screeching across the horizon, took much longer than expected.

At the beginning of the 20th century, 40% of American workers were farmers, comprising 15% of the economy. The tractor created a significant reduction in marginal unit cost of food, freeing approximately three acres of cropland per horse previously needed for feed. Yet by 1920, only 4% of American farms had a tractor. This is the ‘diffusion deficit’: the difference between the availability of an innovation and its diffusion throughout its potential markets. But once the tractor diffused, however, the effects were enormous: agricultural mechanisation would eventually raise American GDP by 8%. Time and time again similar patterns have shown up in the spread of technologies through the economy, as we sit on the precipice of yet another industrial shift we may yet learn from the past. 

Diffusion Deficits

Technologists and scientists often equate technological innovation with immediate social and economic change. But just because a technology exists does not mean it will necessarily have a strong effect on society in the short run––even if the technology is profitable and sensible to adopt. The tractor was one of the most impactful innovations in recent centuries, but it took over 70 years for the potential of this technology to be realised at scale. 

If we understand an “innovation” to be an idea, practice, or technology that is perceived as new, then “diffusion” is the process by which an innovation spreads throughout a population or social system over time. Everett Rogers first popularized the notion of “technological diffusion" in 1962, when he observed that adoption of a new innovation follows a predictable pattern across populations. Rogers proposed a bell curve with five distinct adopter categories: innovators (2.5%), early adopters (13.5%), early majority (34%), late majority (34%), and laggards (16%). This distribution of adopters remained remarkably consistent, regardless of the particular innovation being studied, from agricultural practices to technological advancements.

For a given technology, Rogers identified five crucial factors that influence adoption rates: relative advantage (the perceived benefits of using a new technology over existing alternatives), compatibility (alignment of the technology with existing social values and needs), complexity (the ease of understanding and using the technology), trialability (the ability of new users to experiment with the technology), and observability (the visibility of positive results of others using the technology). Rogers’ theory also emphasizes the importance of communication channels and social systems in the diffusion process, noting that innovations rarely spread through pure technical merit alone, but rather through social networks and interpersonal influence. The importance of word-of-mouth recommendations for diffusion explains why many superior technologies fail while inferior ones sometimes achieve widespread adoption—for diffusion, the social mechanisms of technological recommendation often matter more than the intrinsic qualities of the innovation itself.

A classic example of how Rogers' diffusion theory played out in a real-world scenario is the VHS vs. Betamax format war of the late 1970s and early 1980s. Despite Betamax offering superior video quality, VHS ultimately dominated the market due to VHS's superior compatibility with user needs—specifically, VHS allowed users to record TV for two hours, compared to Betamax's one hour. Home video recording had existed for some time, but what spurred its adoption was this seemingly minor advantage, which triggered powerful network effects. As more consumers purchased VHS recorders, more video rental stores stocked VHS tapes––creating a positive feedback loop where the availability of VHS content reinforced its superior market position. The Betamax failure shows that even with better technology, compatibility and network effects can lead to an inferior product dominating the market.

General-purpose technologies (GPTs) take even longer to be widely adopted by society, but their effects are transformational once they are. Exactly what is classed as a “general purpose” technology has been the source of some debate amongst historical economists: a “general purpose” technology is a one which initially has much scope for improvement, has many uses, eventually comes to be widely used, and also has many social spillover effects. Writing, agriculture, steam power, electricity, railroads, and information technology are all paradigmatic examples of general purpose technologies. General purpose technologies (or GPTs) have even more specific diffusion challenges than regular technologies, even if they follow the same adoption curve as regular technologies.Crucially, GPTs require specialized skills and training to be useful in proportion to their potential. But the diffusion of those skills will necessarily lag behind the invention itself.

In many cases, productivity initially drops when a GPT is introduced because workers and organizations face a learning curve​. For example, firms needed operators and engineers to implement long-range telegraphs and telephones––skills that were scarce in the 1870s. After the invention of the telephone,  it took time to develop educational programs to upskill workers and for ordinary people to gain experience with the new systems. But once this infrastructure was in place, the telephone took off. Inertia also plays a role in slowing diffusion of GPTs––people comfortable with older solutions may be slow to trust or learn a radically new tool.

Cutting edge invention often speeds ahead of what organizations and society can absorb––creating a gap between what could be done with the available technology, and what is done in practice. [something like] as a result, we can’t learn a lot about whether or not a new technology will be impactful just by looking at productivity charts soon after its invention: the full effect of a technology may only be felt years, decades, or even a full century after the fact.

Industrial reorganisation and its consequences

Behind the diffusion deceit lie the problems inherent in implementing a new technology, chief among them industrial reorganisation. Diffusion slows down (particularly for GPTs) because firms don’t just need to adopt a new technology to benefit from the innovation: they often need to redesign themselves from the ground up to take full advantage of the new technology. The context in which a new technology will be deployed often requires as much care, effort, and innovation as the science itself.  And process innovation takes longer than scientific invention: so the effects of the most important technologies often also take the longest to show up in markets.

A notorious instance of the difficulty in implementing a new technology within existing systems is the replacement of steam power with electric motors in manufacturing. Electric motors promised increased manufacturing productivity, but their application in factories initially yielded few efficiency gains. Late-19th-century factories used central steam engines to drive machines via belts and line shafts. When electricity came, managers first just swapped the steam engine for an electric motor, but left the centralised-power layout intact​. As a result, the electric motor did not help, because factory owners simply “overlaid one technical system upon a preexisting stratum.”  

Change did not come until thirty years later, when a new generation of factory designers built “electric native” plants with distributed “unit drives”––small motors on each machine–– and flexible layouts​.  No longer tethered to the central drive shaft, individual machines could now operate independently, each calibrated to its optimal speed and schedule rather than the uniform rhythm imposed by belt-driven systems. This reorganization enabled a fundamental transformation in manufacturing, what we might call the true "electrification dividend."The innovation (electric motor) had been available for decades. But the implementation (optimal use of that motor) was where true progress lay. 

Other GPTs show the same pattern: an innovation cannot be utilised until complementary innovations arise in the processes that are needed to yield its capabilities. In the 1970s and 80s, economists puzzled over the “productivity paradox” of IT––computers had been invented, and seemed self-evidently useful, but companies hadn’t yet figured out how to deploy them effectively, so productivity growth remained modest. Between 1973 and 1995, U.S. non-farm labour-productivity growth fell to roughly 1.4 % per year, compared with 2.8 % in the post-war boom. Micro-panel work showed that only the most IT-intensive plants enjoyed early gains, masking progress in aggregates, as IT forward firms gained but most buyers did not. It took complementary innovations like user-friendly software, networking, and new management practices in the 1990s for IT’s impact to be clearly felt. The benefits of computation and digitalisation in offices did not fully materialize until information flows were re-engineered to leverage digital capabilities. Suddenly, industries like retail trade, wholesale, and logistics – epitomised by Walmart’s data-driven supply-chain redesign – saw productivity levels leap 40-60 % relative to 1987 trends.

Old laws, or lack of standards, can also slow down implementation; for example, early in the automobile era, some countries had regulations requiring cars to travel at walking pace with a person ahead (like Britain’s 19th-century “Red Flag” laws) – rules designed for safety that severely hindered the new technology’s use. Similarly, the telephone’s spread was hindered until standardized exchanges and phone numbering systems were established replacing systems where every region had a different switchboard and routing logic. Interoperability enabled the “killer app” of the telephone: the long-distance call. 

The product problem or wrapper wrangling

Given the importance of diffusion it is fascinating to observe how it is considered regarding the current crop of transformational technology. In the highly financialised space of AI research, for instance, so many resources are directed towards the abstract goal of ‘AGI’ rather than any specific product.

Perhaps our obsession with “AGI” has distracted the AI industry—investors, entrepreneurs, and the frontier labs—from actually making an impact on firms and consumers. You could ask, “Why bother putting all this work into making today’s LLMs do something when next year or the year after we’ll have systems that can do everything?” This is why White House Advisor Dean Ball suggests Manus, a product then built upon Anthropic’s Claude foundation model, has been pejoratively described as a ‘wrapper’ (software that scaffolds a frontier lab’s foundation model).  We tend to overlook technologies like Manus because they defy our conventional understanding of "AI progress.” What Manus’s critics misunderstood is that a wrapper is, in essence, progress in diffusion, rather than in hard technology.  

Would we claim that the train was a “wrapper” on the steam engine? Is a newspaper “a wrapper” on the printing press? Was the bicycle a “wrapper” of the wheel? By the same logic, one would argue AI is just a wrapper on the semiconductor. Every novel technology nests within it the theoretical and physical innovations of the past. GPT’s cannot exist in a vacuum, as pure potential: they need to be specialised to meet the real needs of people and firms. This specialisation has always been a crucial component of human progress: utility can only be extracted from innovation when a GPT finds a form factor that allows it to fulfill a real need. Whether it be celestial navigation, steam power or, indeed, the Generative Pretrained Transformer. 

With the world’s focus squarely on the technical innovation taking place at scaling labs, progress in AI diffusion strategies have less veneer. But the more transformative a technology, the more exhaustive and comprehensive its diffusion strategy must be. Building a good wrapper requires a profound understanding of diffusion, an incredibly valuable skill. We ought to celebrate Charles P. Day, James Bray Griffith, and Oscar Perrigo, and the other engineers who first understood that electrification meant rethinking both power delivery and layout, as much as we do Tesla and Edison.

Meanwhile, AGI promises to be our last innovation, and it may well be – but it will still require complementary innovations to effect its change on the real world like previous GPTs. An AI-native organisation, one that fully harnesses AI, has not been invented – save for in the copy of YC applicants' slide decks. Part of what’s missing is that so much of our financial, legal, and organisational frameworks are not interoperable with AI, these institutions will have to change. The burden will fall upon humans to make a world which can work with AI. Only then will we see a new kind of autonomous firm evolve and grow, a natural product of the economy and its integration with machine systems.