The Pattern of Invention: A Historical View of Transformative Technologies

Transformative technologies tend to arrive the same way. They externalise a capability — move it out of the individual and into a tool that others can use, copy, or own. Then a small group controls it. Then, sometimes, it spreads. AI has arrived at the first stage, with the fastest adoption on record and the same early concentration every previous case showed.

The sequence is recognisable across thousands of years. It is also not automatic, and the exceptions are the most useful part of the story.

Two Kinds of Externalisation

The earliest examples did not just make a task easier; they changed what humans are.

Fire, controlled at least 300,000 years ago, externalised energy — warmth, protection and cooked food without direct physical effort, extending active hours into the night and human range into colder climates. Language, roughly 50,000–100,000 years ago, externalised thought: ideas could survive the person who had them, travel, and accumulate across generations. Agriculture, around 12,000 years ago, externalised food production, so labour invested once could be harvested repeatedly.

A second wave changed what societies could do without changing human biology. The wheel (c. 3500 BCE) externalised transport. Writing and paper externalised memory — knowledge no longer had to live in a head or on clay. The abacus externalised calculation, letting empires manage taxation and trade at a scale mental arithmetic could not reach. Gunpowder externalised force, and the castle and the armoured knight lost the monopoly they had held for a millennium.

Both waves follow the same first step, and both begin concentrated: priesthoods, imperial workshops, military establishments, wealthy merchants. The timeframes then vary enormously — centuries for paper to travel from Han China to Europe, decades for gunpowder to reshape European warfare.

The Modern Compression

The pace accelerates, but not the shape.

The printing press (c. 1440) externalised knowledge reproduction. It needed capital, so early presses belonged to wealthy merchants and religious institutions — then within fifty years, more than two hundred European cities had one. The steam engine (1712) externalised physical labour, displacing artisans and weavers hard enough to produce the Luddite protests of 1811–1816, before new professions and higher living standards emerged over the following decades. Electricity was first a luxury of wealthy urban districts, with proprietary early networks designed to lock customers in; open, interoperable standards and regulated public grids turned it into a universal utility in about fifty years. Computing and the internet began inside governments, corporations and universities — ARPANET was a closed military-academic network — and reached global consumer adoption in roughly fifteen years once open protocols (TCP/IP, HTTP) made permissionless connection possible.

"Broad adoption" here means the point at which a capability is used by a large share of the people it could serve, at a price they can pay. Measured that way, one trend is unambiguous: the speed of adoption has compressed dramatically across history. What took centuries for paper and fifty years for electricity now takes years.

The second observation matters for how AI is understood: its transformation is deep but different. Like the abacus, it externalises a cognitive capability — calculation then, synthesis now. Like fire or agriculture, it does not change human biology. It changes workflow and capability, which is precisely why the engineering around it decides the outcome.

The Pattern, and the Cases That Break It

Across the record, four steps recur:

1. A transformative invention externalises a core human capability.
2. It is initially controlled by a small group — those with the resources, knowledge or authority to deploy it.
3. It expands, through trade, competition, standardisation or regulation.
4. It becomes widely accessible, often at a scale its early adopters could not have imagined.

Steps 1 and 2 are close to universal. Step 3 and 4 are not, and pretending otherwise would make the pattern useless. Three cases show why:

• Nuclear power was a transformative energy technology with state backing and enormous capital — and it stayed concentrated and expensive for decades, because cost, regulation and public risk perception never lined up to make distribution work.
• Supersonic passenger flight (Concorde) was technically spectacular and never distributed at all: the economics and the noise rules never permitted a second life.
• Nuclear weapons were deliberately kept from spreading, which is the clearest demonstration that concentration can be a choice maintained by policy rather than a phase that ends on its own.

So the useful reading of history is not "every technology eventually spreads". It is: technologies spread when the economics, the standards, or the rules push them that way — and stay concentrated when they do not. Distribution is a decision made by people, repeatedly, not a force of nature.

Where AI Sits

AI externalises cognitive synthesis: recognising patterns, generating text, analysing data, drafting solutions. It runs on top of every previous invention — the electricity grid, the knowledge the press and paper produced, the computation that grew from the abacus, the connectivity of the internet.

As of 2024–2026:

• the most capable models are developed and controlled by a small number of large companies;
• access is mediated through proprietary platforms and paid services;
• the infrastructure underneath — data centres, specialised chips, training data — is concentrated;
• hundreds of millions of people use AI daily, most through interfaces they do not control.

The speed of adoption is unprecedented. The pattern of initial concentration is not. And because step 3 is not automatic, the question worth asking is not whether AI will follow the historical sequence. It is which of the three cases above AI most resembles — nuclear power, Concorde, or Linux.

That outcome will not be decided by the technology. It will be decided by architecture, which is where Part 2 takes the argument. And the engineering discipline that makes any of it trustworthy is the same discipline this site keeps returning to: requirements as engineering, architecture before code, and gates that can actually fail.

The Orchestrator's Takeaway

Do not read history as a promise. Concentration is the default; distribution is the exception that has to be engineered, paid for, and defended. When you adopt a transformative technology, ask which of the pattern's later steps is actually happening — and who is doing the work that makes it happen.

Next in the Series

Part 2 — Why Architecture Dictates Autonomy draws the conclusion: whether AI completes the pattern is an architectural choice, and the Linux precedent is the case worth learning from — including where it does not apply.

For the full method, see The DevOps Engineer's Guide to Effective AI Usage.