Consider a friar wandering 13th-century Europe. A Bible amongst his few threadbare possessions, its pages made from exceptionally thin parchment and its text written in letters small enough to fit the entire work into a single portable volume.

The tumultuous centuries of Reformation that followed the creation of Pocket Bibles brought further changes to relationships between individuals, information and institutions. Four developments in this historic process are particularly pertinent to today's AI revolution. 

• Standardisation gave biblical texts a more consistent structure, much as common architectures and formats make AI models easier to move between systems. 

• Miniaturisation made complete Bibles portable, while compression and efficiency can bring useful AI onto personal computers even as the largest models continue to grow. 

• Translation of texts into peoples native tongues widened the population capable of reading scripture. Similarly AI opens the opportunity for expert knowledge to be widely accessible. 

• Production of texts radically expanded once the printing press reduced the cost of reproduction. Today, open-weight models can be copied, adapted and operated beyond the infrastructure of their original developers.

None of these developments caused the Reformation alone, nor did they occur as stages in a planned process. Together, they gradually changed who could possess, understand and reproduce information, reducing central control and catalysing new institutions. AI is passing through analogous processes at a greatly accelerated rate. The consequences may depend as much on decentralisation as on frontier capability.

Standardisation

In the 13th century, the Paris Bible standardised the structure of religious text into chapter divisions that remain familiar today. This made it easier for people in different places to navigate and refer to the same passages. 

AI has its own forms of standardisation. Common model formats allow models to move between researchers, developers and hardware. A model developed by one organisation can easily and independently run elsewhere without specialised infrastructure. 

Standards influence how decentralised a technology can become, even when they appear less significant than improvements in capability. A shared structure makes information or technology easier to move beyond its original institution.

Miniaturisation

Pocket Bibles changed a different constraint; reducing the physical size at which a Bible could be possessed. Bookmakers used extremely thin parchment and tiny writing to produce volumes suited to travelling clergy. Compression did not increase the knowledge in the book, but it personalised the particulars under which that knowledge could be possessed and ultimately used.

AI is scaling on several axes. Frontier models increasingly require vast compute, placing their development and operation beyond the resources of individuals. Meanwhile quantisation, distillation and improvements in hardware are making useful models accessible on consumer computers.

A laptop does not need to reproduce everything available in a data centre for this form of decentralisation to matter. A Pocket Bible did not need to contain more material than a monastic library for its portability to be useful, and a local model does not need to match the most capable frontier model for individual ownership to have value. 

Centralised systems still have substantial advantages, just as medieval universities and libraries could provide resources far beyond those carried by an individual reader. Cloud models can use much greater computing power. Open-weight models offer greater control over availability, content and opinion. 

Decentralisation does not require one arrangement to replace the other, and may instead depend upon useful AI being able to exist at several scales.

Translation

Despite addressing portability, Pocket Bibles left other barriers intact in Pre-Reformation Europe. Texts remained in Latin and were only legible to a literate minority. In the 14th century, followers of the theologian John Wycliffe produced the first complete Bible in English. While requiring substantial effort to produce by hand, the translation carried an important idea: people who did not know Latin could now encounter scripture in a language they understood.

Much of the information available to us today has an equivalent barrier created by specialist knowledge. A scientific paper can be freely available while assuming years of specialist education. Programming documentation can contain the answer to a problem while remaining inaccessible to someone unfamiliar with its concepts. AI can translate some of this material into vernacular appropriate to the user, reducing the expertise required to engage with information that was already available.

Wycliffe’s ideas did not remain in England. They travelled to Bohemia and influenced the reformer Jan Hus, who was burnt as a heretic in 1415. Once information moves across community lines it can acquire meanings and purposes its originators did not intend. Decentralised AI introduces the same loss of censorial control when models can be possessed, adapted and used outside the organisations that created them.

Wider access to knowledge leaves questions of interpretation unresolved. Translating a Bible into English or German inevitably involves decisions about how its words should be rendered. An LLM explaining a scientific paper or legal judgment can introduce interpretations of its own. Removing one intermediary can create another, so the value of decentralisation lies in widening participation rather than guaranteeing that everyone who participates will reach a particular conclusion.

Production

Printing changed popular participation during the fifteenth century by making books faster and cheaper to produce. A century later, Martin Luther published his German New Testament in 1522, followed by William Tyndale's English New Testament in 1526, signalling the start of the upheaval we now call the Reformation. Neither man invented the idea of translating the Bible into ordinary languages, but printing allowed knowledge to circulate in previously unimaginable quantities. 

Open-weight AI introduces a comparable change in production and distribution. A cloud service can serve millions of people while remaining a single centrally governed system; open model weights can produce independently operated copies across many computers. Those copies remain usable irrespective of the original developer. 

Printing became part of a longer feedback loop between access to information and the ability to use it. The Protestant emphasis on reading scripture and the growing supply of printed material in native languages increased the value of learning to read, which in turn created greater demand for printed text.

Ideally, AI will create a similar feedback loop between access to expertise and the acquisition of expertise. Someone who can use a model to understand programming faces a lower barrier to learning it; someone who can work through unfamiliar scientific literature faces a lower barrier to understanding. 

Some of these people will then produce software, research and explanations of their own. Whether this becomes comparable to the expansion of literacy remains unknown, but it illustrates how easier access to knowledge can eventually affect the capabilities of the people using it.

During the Reformation, improved production circulated scholarship alongside polemic and propaganda. Open models can likewise support beneficial and harmful applications using the same underlying capability. Decentralisation increases participation and reduces any single institution's ability to determine what everyone else can do, leaving societies to manage the consequences rather than preordaining a path forward.

An Uncertain Reformation

Standardisation, miniaturisation, translation and production interacted to spark widespread changes in religious, social and intellectual life. The Reformation emerged organically over generations of cultural evolution rather than following ipso facto from a single development.

These same interactions may matter more for AI than any single technical advance. Most importantly, consequences will dance to the fiddle of culture not engineering.

The changes that took place during and around the Reformation were uncertain at the time. The shape of history is much clearer to us now than it was to anyone living through it. 

We have no comparable distance from the development of AI. 

It’s difficult to predict how cloud and local systems may eventually complement each other, or which will win out for every day use. Open-weight models may become central to the technology or prove to be just one stage in a distribution model that changes again. 

Centralisation is an important variable within this uncertain system because it shapes who determines the course of development. Techno-feudalism is used to describe a system in which a population depends on a few monolithic owners of digital infrastructure. Intelligence can become inexpensive and widely accessible while the infrastructure and the technology’s future remains controlled by a universal church.

Open weights and local inference offer a decentralised possibility, allowing some models to leave their creators' infrastructure and continue operating elsewhere. This does not create complete independence: personal computers still depend upon semiconductors, software ecosystems and electricity. But it does increase the congregation capable of influencing the technology’s future and its consequences. 

The history surrounding the Reformation and the decentralisation of religious knowledge provides a framework for thinking about AI without providing a prediction. The democratisation of knowledge in the Reformation did not make participation equal or remove existing authorities, but created circumstances for a cultural evolution whose form emerged over generations. 

Reforming elements in AI may democratise machine intelligence, distributing some control beyond the institutions capable of developing the largest systems. Preserving an ecology of technical arrangements matters despite the consequences remaining uncertain if this process is to be a broad social evolution.