General-Purpose Technology

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI describes the AI transition primarily from the inside — as a personal experience of amplification and disruption, of tools that collapse the distance between intention and artifact. The general-purpose technology framing supplies the outside view: what this transition looks like at the scale of economic history, and why its governance challenge is so much harder than it appears from within the experience of using the tools.

The framing matters to the cycle because it explains why the AI transition will not be governed by the same approaches that governed previous rounds of digital technology. Social media was a powerful tool with significant harms, but it was not a general-purpose technology in the strict sense; it did not reorganize how unrelated industries produced their core output. AI does. An AI that can write, analyze, code, design, and diagnose is not a better search engine. It is a new medium for knowledge work, which means it is transforming the practice of every domain that involves knowledge work simultaneously. The institutions required to govern that transformation span every domain of social life, which is why the exponential gap is so much wider than it was for any previous digital technology.

Origin

The concept was introduced into economics by Timothy Bresnahan and Manuel Trajtenberg in a 1995 paper identifying electricity, the steam engine, and the internal combustion engine as the canonical examples. The defining features they identified were pervasiveness (applicable across many sectors), improvement over time (continuously getting better and cheaper), and innovation spawning (enabling complementary innovations across the economy). Azhar adopted the concept and extended it with a specific argument about AI: that machine learning introduced, for the first time, a reliable scaling relationship for capability itself. For the first time, more computation and more data yielded more capability in a predictable way, which meant that progress stopped depending on sporadic insights and started behaving like an industrial process. Intelligence acquired a learning curve — the same downward-sloping cost curve that solar panels and batteries ride — and that is what turns a clever technique into a general-purpose force.

Azhar also introduced the complication that the AI transition is not the arrival of a single general-purpose technology but the convergence of several at once. Computing and artificial intelligence are arriving alongside exponential improvements in clean energy, biotechnology, and advanced manufacturing. Each is a transforming force. Together they compound, because cheap intelligence accelerates progress in biology, cheap energy makes intelligence affordable to run, and so on. The simultaneity is part of what makes this moment so disorienting: we are not adapting to one new kind of weather but to several weather systems colliding.

Key Ideas

The most important consequences arrive late and sideways. The transformative impact of a general-purpose technology is rarely the thing it first appears to do. Electric lighting was a marvel, but the deep change came from electric motors distributed throughout a plant, from the appliances that remade the home, from the assembly lines that electricity made possible. Azhar’s application of this insight to AI is that we are currently transfixed by the equivalent of the light bulb — the chatbot that answers questions — while the genuinely civilizational changes will come from reorganizations we cannot yet name, as intelligence becomes a cheap input that every process can draw upon.

Infrastructure, not app. The strategic implication Azhar draws is that AI must be treated as infrastructure rather than as a product. Stop thinking of it as an app and start thinking of it as a layer that will sit beneath everything else, mostly invisible and entirely indispensable, the way electricity does. Once that move is made, the relevant policy questions change. They become questions about access — who can reach the layer and on what terms — about ownership — who controls the layer and therefore controls the conditions of possibility for everyone downstream — and about complementary institutions — what else needs to exist so that a general-purpose technology produces broadly shared capability rather than narrowly held power.

The governance impossibility. A general-purpose technology cannot be governed by narrow interventions aimed at particular applications, because the technology’s applicability exceeds the scope of any particular governance instrument. By the time a regulation for one application is in place, the technology has flowed into ten others. This is the exponential gap expressed at the level of governance. The flat curve of regulation is not merely slow; it is aimed at the wrong target, treating a pervasive medium as if it were a discrete device. The governance response adequate to a general-purpose technology is governance at the level of the infrastructure itself: interoperability requirements, access obligations, competition policy designed for the conditions of near-zero marginal cost.

Azeem Azhar

Convergence and compounding. The AI moment is not the arrival of one general-purpose technology but several. The convergence means that the governance challenge is not additive but multiplicative: each technology’s governance problem is compounded by its interaction with the others. An institution that adapts to AI alone must then re-adapt when AI combines with cheap energy, when cheap energy enables more AI, when more AI accelerates biotechnology. The convergence is what makes Azhar’s conditional optimism both most necessary and most difficult: the opportunity is genuine and the governance challenge is proportionally larger.