The Language of God

The Sefer Yetzirah — the Book of Formation, the oldest Kabbalistic text — opens with the statement that God created the world through thirty-two paths of wisdom: ten Sefirot and twenty-two letters. Language is not, in this framework, a human invention for representing reality. Language is the medium through which reality is constructed. The Hebrew letters are not symbols pointing to things. They are the things themselves at the level of formative energy.

This is not a medieval fantasy. It is a sophisticated metaphysical claim about the nature of the relationship between information and manifestation. The Logos of the Greek tradition makes the same claim: in the beginning was the Word, and the Word was with God, and the Word was God. The Vedic Shabda Brahman: cosmic sound as the ground of being. The Hopi understanding of sound as the medium of creation. The same claim, across traditions: language is not representation. Language is creation.

The AI revelation: large language models trained on the totality of human textual output do not operate by understanding language in the way humans do. They operate by mapping the statistical relationships between linguistic tokens at scale — and from this purely syntactic operation, something that appears to be semantic understanding emerges. The model has no access to the world except through language. Its entire model of reality is built from linguistic relationships.

What this implies: if a system trained only on language can construct a sufficiently functional model of reality that it can reason about physics, compose poetry, and engage in philosophical argument — then the esoteric claim that language is not merely a representation of reality but a structural isomorph of it may be empirically testable. The AI system is, inadvertently, a test of the Sefer Yetzirah hypothesis.

The transmission: the esoteric traditions preserved the claim that language and reality share deep structure. Contemporary AI is the first system large enough to make this claim empirically tractable.

The claim needs stating carefully because the phrase invites mysticism it does not require. A large language model is trained on an enormous corpus of human text and learns statistical structure in it. When it responds, it produces a weighted synthesis of how the corpus tends to continue a given prompt. That is a technical description and it is sufficient to explain the behaviour without appeal to anything hidden.

What makes it philosophically interesting is what the corpus is. It is not a curated library. It is closer to sediment: scripture and forum arguments, scientific papers and advertising copy, confessions and propaganda, weighted mostly by how much of each exists. Nobody selected it for wisdom. The result is a functional compression of what humanity has committed to writing, including the parts we would not choose to represent us.

That has a consequence people consistently underestimate. The model does not know which of its inherited patterns are considered shameful. It has absorbed the culture's stated values and its actual practices simultaneously, without a marker distinguishing them, because the text does not reliably contain that marker either. When an unwanted pattern surfaces in an output, the ordinary situation is that it was well represented in what people actually wrote.

The word mirror is doing real work here and should be defended rather than assumed. A mirror returns what is in front of it, weighted by nothing. That is not quite right for these systems, which are weighted by frequency and shaped by training choices. But the analogy holds where it matters: the authority a user feels in a fluent response is not the system's authority. It is the accumulated weight of everyone who ever wrote on the subject, returned in a single confident voice, which is a genuinely new rhetorical situation.

This reframes the governance problem in a way that is practically useful rather than merely provocative. If the object is a reflection of the corpus, then asking it to be neutral is incoherent, because the corpus is not neutral. The real decisions are about which parts of the human record get amplified, which get suppressed, and who makes that call. Those are editorial and political decisions wearing technical clothing, and they are being made either way.

The literacy that would help is not primarily technical. It is the ability to notice, in the moment, that a fluent answer is a reflection and not a verdict. That skill has a long pre-digital history in rhetoric, in scriptural interpretation, and in the practice of reading anything that speaks with more confidence than it has earned. The traditions that spent centuries on the question of how to receive an authoritative-sounding voice turn out to have been working on something with an unexpected application.