Returning to the Source

In 1994, for a master’s thesis, I argued that meaning doesn’t travel. It is not a payload loaded into words at one end and unpacked at the other — it is an effect of statistical and differential structure in language, structure no single speaker authors. I built the case by running Claude Shannon’s model of communication through Saussure and Derrida. It was an argument you could make with citations and no evidence. Three decades later the evidence arrived, at scale, as transformers.

The thesis was called Returning to the “Source” — a reconsideration of Shannon and Weaver’s mathematical model of communication. The textbook read of Shannon treats his model as plumbing: a sender encodes a message, a channel carries it, a receiver decodes it, and “meaning” is whatever the sender stuffed in. I thought that read threw away the most interesting thing Shannon had found — that language has measurable statistical structure, and that this structure, not the sender’s intention, is what makes communication possible at all. The transformer era has since made that argument concrete in a way I couldn’t have proven then. Three legs of it, in particular.

The statistics were the meaning

Shannon’s quieter result — the one in Prediction and Entropy of Printed English — is that you can predict the next letter of English from the letters before it, because language is shot through with redundancy. He built his approximations from Markov chains, the same device Andrey Markov used when he sat down and counted the vowels and consonants in Eugene Onegin. Meaning, on this view, is not a thing sitting behind the symbols. It is a pattern in how the symbols fall.

A large language model is that experiment run at planetary scale. Next-token prediction is Shannon’s printed-English game with a trillion parameters and a continent of text behind it. The thesis bet that statistical structure was doing the real work; the transformers are what happens when you take that bet seriously enough to spend a billion dollars on it.

Meaning is not a thing behind the symbols. It is a pattern in how the symbols fall.

Meaning is difference, deferred

Shannon alone gives you redundancy, not meaning. To get from one to the other the thesis turned to Saussure, who held that a sign means nothing in itself — it means only by its differences from every other sign. Derrida pushed that further with différance: meaning is not only differential but deferred, never fully present, always pointing onward to other signs.

This sounds like philosophy until you look at how a transformer represents a word. An embedding is meaning as position in a differential space — a word is defined by what it sits near and far from, which is exactly Saussure’s “differences without positive terms.” Attention is deferral made mechanical — a token’s meaning is conditioned on its context, resolved relationally across the sequence rather than retrieved from a fixed dictionary. The machinery that makes language models work is the machinery the thesis described, built in silicon.

A voice with no one behind it

The thesis closed on a smaller, stranger claim, and I’ll keep it light because it travels furthest: if meaning is an effect of structure, then so, in part, is the speaker. The subject who seems to be talking is also produced by the differential field, not standing outside it issuing messages into it. You don’t have to accept the full Lacanian apparatus to notice that a language model demonstrates the deflationary version cleanly — it produces a fluent, consistent voice with no self behind it, a speaking position generated by the statistics rather than a person who owns them. The persona is an effect of the pattern. That is a claim about language, not only about machines.

Two branches, reconverged

There is a historical irony I didn’t fully appreciate at the time. Information theory and French post-structuralism are usually treated as opposite intellectual worlds — engineering on one side, the humanities on the other. But they share a root: both descend from the structural linguistics of Saussure and Jakobson, and Jakobson personally carried those ideas between Shannon’s world and Paris. The two branches split after the war and spent fifty years not speaking to each other. Large language models are where they reconverge — a statistical, differential, relational account of meaning that an engineer and a literary theorist would both recognize. The thesis was an early attempt to hold both branches in one hand. The transformers vindicated the grip.

Where the lens began

I tell this story on a UX-research site for a reason. The commitment that runs through everything I’ve done since — that meaning is situated and relational, not a payload; that you study the system of relations, not the isolated unit — did not start in a usability lab. It started here, in 1994, before the PhD, as an argument about Shannon. The method I’ve carried through thirty years of fieldwork (The Practice, and its six traditions) is this thesis turned into a way of watching people. The same relational picture of meaning scales up into the claim that the intelligence is in the coordination, not in any single unit. I simply didn’t have the machines yet to show that the underlying picture of language was right.

Now we do. The transformers didn’t discover a new theory of meaning. They confirmed an old one.

Essay — foundations. A 1994 master’s thesis read under the thirty-year lens. ← All essays