# The Human Language Model When people ask how today's language models work, I already know the answer, because I have been running one in my head for about thirty years. A large language model learns by breaking words into small pieces, fragments like *vi*, *sion*, *pro*, and *-ture*, and studying which pieces keep company with which. Over billions of examples it builds a vast map in which every word sits at a particular distance from every other, so that related meanings cluster together and a single piece can light up a whole family of words. The linguist J. R. Firth put the principle in a sentence back in 1957: "You shall know a word by the company it keeps." That is what I was doing every night with a notepad. I would take a word, break it into its sounds, pivot each piece, and watch which other words lit up. Page after page, I was building the map by hand. I did not know it had a name. Eventually I tried to build it in software. Beginning in 1994 and for about a decade after, I constructed what started as a rhyming dictionary and grew into a symbolic language engine. It stored each word's pronunciations separately from its spelling, indexed words by sound, matched a long word's sound against whole phrases across word boundaries, and mined corpora for trigger relations, pairs of words whose presence changes the likelihood of each other. In today's terms it was a lexical knowledge graph with a phonological engine at its heart, an early machine attempt at the same map I had been drawing by ear. I tell that story in [[projects/Ten Years Building a Symbolic Language Engine|Ten Years Building a Symbolic Language Engine]]. ## Learning by ear There is one important difference between my method and most of today's machines. Models mostly learn from spelling. I learned from sound, which is why my own engine kept pronunciation as its own layer, apart from the letters. The ear catches things the eye passes over. *Vi* rises: it sees, it reaches, it climbs like a vine toward the light. Add the hiss of an *s* and the rising turns into a vise, a clamp, a cut. The same sound hides in *vice*, spelled with a *c* instead of an *s*: still *vi* and the hiss, still a clamp. A moral vice is a deficit that holds a person in place, so they can no longer climb like a vine. The ear hears through spelling, and letters that carry the same sound are often doors into the same idea. *Decide*, *incision*, *concise*, *precise*: the *s* sound cuts, and the documented record confirms it, since all of them descend from the Latin *caedere*, "to cut." To de-cide is to cut away every option but one. When I hear *envy*, I hear the *vi* pressed down, the reaching cut off from rising, left to look up at someone else's height. The Latin agrees in its own way: *invidia* comes from *in-* and *videre*, "to look upon." Envy is a way of seeing. Thirty years of this is training in the literal sense. Associations repeated night after night eventually fire on their own, so now I hear a word and its family arrives with it, the way a trained model completes a pattern without looking anything up. ## A dictionary at my fingertips Long before the notepad, there were the books. As a kid in the South, I spent my time reading encyclopedias and dictionaries and playing in the woods. Later I read dictionaries and encyclopedias for years on end, and I learned to speed read, dyslexic as I am, because I was so hungry for language. I told that story in a 2015 broadcast, [[livestreams/Courage, Partnership, and Following the Heart - Livestream|Courage, Partnership, and Following the Heart]]. All that reading, run through years of pivoting sounds, did something to the way I think. I became an encyclopedia without the dictionary. When I hear a word, it rarely arrives alone. It arrives in flashes, with paths branching off in many directions at once: its roots, its cousins, the sounds hiding inside it, the stories it has carried. To most people it looks like a simple word. To me it is a crossroads. In a 2018 broadcast, [[livestreams/Critical Thinking, Dignity, and the Practice of Respect - Livestream|Critical Thinking, Dignity, and the Practice of Respect]], I told people this was one of the two practices that would change their lives: studying Greek and Latin roots, prefixes, suffixes, and what I called phonetic morphology. I later learned that linguists call it morphophonology, the study of how a word's sounds shift as its parts combine and change, and that tracing those shifts across centuries is called historical phonology. Every word comes prepackaged with an explosive amount of information built right into it. Learning to unpack it is like walking around with a dictionary at your fingertips, one that connects every point of data you have ever collected. People sometimes call that genius. Back when Jenni and I were together, she would watch me talk fluently with people about subjects she had never once heard me mention. To me it feels like magic, and that sound work is a large part of the magic. ## Reflective etymology Over the years I came to call what I was doing meta-language, or reflective etymology: reading the structure of a word to recover the insight folded inside it. Most people operate at the surface of language. Learning to see beneath it creates a vast web of connective associations that runs underneath ordinary conversation, and that web is where much of my thinking happens. Here are a few entries from my notes, many of which I have used in my live broadcasts. **Disappointed (dis-appointed).** Through your state of mind and attitudes, you can dis-appoint yourself from your appointment. Positive attitudes grant you a higher station, where you are appointed by your faith as an act of self-realization. When you are disappointed, you un-appoint yourself. The history bears this out more than you might expect. The Old French *desapointer* meant to remove someone from an office or appointment. The earliest sense of *disappoint* in English was exactly that: to be stripped of your post. I have returned to this reading again and again on my broadcasts: in [[livestreams/Releasing Anxiety by Reexamining the Stories We Carry - Livestream|Releasing Anxiety by Reexamining the Stories We Carry]] and [[livestreams/Long-Term Strategy, Social Capital, and Delayed Gratification - Livestream|Long-Term Strategy, Social Capital, and Delayed Gratification]] in 2016, in [[livestreams/From Dream Chasing to Patient Creative Practice - Livestream|From Dream Chasing to Patient Creative Practice]] in 2017, where a dream chaser disappoints themselves from a divine appointment, and in [[livestreams/Frustration, Acceptance, and the Discipline of Non-Attachment - Livestream|Frustration, Acceptance, and the Discipline of Non-Attachment]] in 2018, where impatience un-appoints us from the vision we first appointed ourselves to. **Recreation (re-creation).** Recreation is less about leisure than about re-creating yourself. Time off is a chance to imagine your own re-creation. The Latin *recreare* means "to make anew, to restore." The word has carried the idea of renewal all along. **Information (in-formation).** Information is inward formation: knowledge that shapes the one who receives it. The Latin *informare* means "to give form to, to shape." To be informed is, at the root, to be formed. **Adventure (ad-ven-ture).** *Ad*, the addition; *ven*, of what comes to you; *ture*, that becomes a part of you. An adventure is what arrives in your life and becomes part of who you are. The documented root is Latin *advenire*, "to arrive," from *ad-* and *venire*, "to come." So the heart of my reading, *what comes to you*, is the word's actual history. The rest is the resonant layer, the ear hearing *mixture* and *tincture* in that final *-ture*, a reading I still find true to the experience. I gave that breakdown in a 2016 broadcast, [[livestreams/Releasing Anxiety by Reexamining the Stories We Carry - Livestream|Releasing Anxiety by Reexamining the Stories We Carry]], where it led to the line that each of us is the greatest undiscovered adventure of our lifetime. Two years later, in [[livestreams/Courage, Creative Risk, and Moving Through Fear - Livestream|Courage, Creative Risk, and Moving Through Fear]], I heard the word another way: *ad* plus *venture*, the addition of risk. The venture is the risk part, and without risk there is no adventure. That reading is documented too. *Venture* began in the fifteenth century as a shortened form of *adventure*, meaning a risky undertaking, a hazard taken on purpose. **Disposable (dis-posable).** To pose is to assume an attitude or position, and a disposition is a habitual way of acting. A person whose disposition harms others disposes of themselves through it. On the other side, a person who refuses to take the pose a group demands, who is not posable, can find themselves disposed of: cast out, made an outcast, someone who will not be cast in the group's mold. Here the documented root is Latin *disponere*, "to arrange, to put in place," from *dis-* and *ponere*, "to place." *Disposition* and *dispose* share it, so the link between how a person is arranged inwardly and whether they are put aside is built into the word's family. ## Two layers As those entries show, I keep two layers in view. The documented layer is the history scholars have traced: where a word has actually been. The resonant layer is what the sound does to the ear and the mind today. Sometimes they match exactly, as with *disappoint*, *recreation*, and the cut in *decide*. Sometimes the resonant reading departs from the record, as when I hear the *vent* of *ventilation* inside *adventure*. Both layers teach, and labeling them keeps the insight honest and makes it more persuasive. I have since learned that this second way of hearing has a long tradition behind it, from Plato's *Cratylus* to the esoteric Green Language. I explore those roots, and the moment I discovered that my bedtime game was a real field of study, in [[blog/20261016 - Pivoting Vowels in the Dark|Pivoting Vowels in the Dark]]. ## What the machines confirm There is something moving, to me, about watching machines learn language the way I learned it. They break words into pieces. They map the company each piece keeps. They discover, through sheer repetition, that meanings travel in families. Every time a model surfaces a connection I found decades ago on a notepad, it feels less like being replaced than like being recognized. The difference that remains is the ear, and with it the meaning a person brings to the pattern. A model can tell you that *vision*, *advise*, and *visible* share a root. Hearing that envy is a vision pressed down, a reaching cut off from the light, and then using that to understand your own heart: that is where the human language model still has work to do. I am gathering my notebooks into a new section of the website so you can see these pivots as they happened and try them yourself. ## Related topics - [[wiki/Phonetic Etymology|Phonetic Etymology]] - [[wiki/Etymology|Etymology]] - [[wiki/Morphology|Morphology]] - [[wiki/Phonosemantics|Phonosemantics]] - [[wiki/Green Language|Green Language]] - [[wiki/Cratylus|Cratylus]] - [[wiki/Lexicon|Lexicon]] ## Related blogs - [[blog/20261016 - Pivoting Vowels in the Dark|Pivoting Vowels in the Dark]] - [[blog/20261016 - The Irksome Zero Matter in Semiotics|The Irksome Zero Matter in Semiotics]] ## Related articles - [[projects/Ten Years Building a Symbolic Language Engine|Ten Years Building a Symbolic Language Engine]] - [[livestreams/Releasing Anxiety by Reexamining the Stories We Carry - Livestream|Releasing Anxiety by Reexamining the Stories We Carry]] - [[livestreams/Courage, Creative Risk, and Moving Through Fear - Livestream|Courage, Creative Risk, and Moving Through Fear]] - [[livestreams/Critical Thinking, Dignity, and the Practice of Respect - Livestream|Critical Thinking, Dignity, and the Practice of Respect]] - [[livestreams/Courage, Partnership, and Following the Heart - Livestream|Courage, Partnership, and Following the Heart]] - [[livestreams/Long-Term Strategy, Social Capital, and Delayed Gratification - Livestream|Long-Term Strategy, Social Capital, and Delayed Gratification]] - [[livestreams/From Dream Chasing to Patient Creative Practice - Livestream|From Dream Chasing to Patient Creative Practice]] - [[livestreams/Frustration, Acceptance, and the Discipline of Non-Attachment - Livestream|Frustration, Acceptance, and the Discipline of Non-Attachment]]