# The Intelligence in the Dictionary ![[resources/images/20171004-the-intelligence-in-the-dictionary.jpg]] For years I kept returning to a peculiar image: a dictionary is a closed sphere. Every word is defined by other words; those words send the reader toward still more words; and a finite lexicon eventually folds back into itself. Nothing inside the book can step outside the lexical circle, point to the rain, and become wet. I used a tree as the simplest demonstration. A dictionary can define trunk, root, branch, leaf, fruit, bark, cell, and molecule, yet no sequence of definitions can reproduce the event of standing beneath a tree, placing a hand against it, or tasting what it grew. The image may have sounded spiritual when I used it in the broadcasts, but the obsession behind it was technical. Beginning in 1994 and continuing through the mid-2000s, I was building the system documented in [[projects/Ten Years Building a Symbolic Language Engine|Ten Years Building a Symbolic Language Engine]]: phonological indexes, explicit lexical relations, semantic neighborhoods, trigger associations, grammatical constraints, corpus-derived concept sets, and procedures for traversing them. By 1999 I was already trying to make a machine move through language as a structured representation rather than treat a dictionary as a shelf of entries. The question beneath the project was the same question beneath the broadcasts: if language is closed, what kind of intelligence remains inside it? ## The Closed System Was Full +[[wiki/Lexical Circularity|Lexical closure]] never meant that language was empty; it meant that language could not supply its own world. A definition can locate a word among synonyms, opposites, categories, causes, attributes, and uses without producing the object to which the word refers. The closed system therefore preserves an extraordinary structure while lacking the independent sensory access required to verify that structure against reality. {{A closed language system can preserve intelligence without possessing the world that made the intelligence possible.}} The absence of an exit is a limitation of grounding, not evidence that nothing of consequence was retained inside. [[wiki/Semantic Holism|Semantic holism]] explains why the closure is productive. A word receives much of its meaning from the company it keeps, the contrasts it sustains, the inferences it licenses, and the positions it occupies inside a wider field. Change an important relation and several neighboring meanings move with it. A dictionary makes some of those relations explicit, while ordinary language preserves millions more through sequence, metaphor, correction, expectation, idiom, genre, and repeated use. The lexicon is the visible skeleton; the corpus is the life history accumulated around it. ## Living Minds Deposited the Relations +Human beings created the relational intelligence preserved in language by surviving contact with a world that words alone could never contain. People touched trees, built fires, buried parents, crossed rivers, negotiated alliances, watched storms, raised children, lost wars, tested medicines, repaired machines, and then carried distinctions from those encounters into speech and writing. Later readers inherit neither the original rain nor the original grief; they inherit patterned evidence of what other minds learned to distinguish. {{The dictionary was never empty of reality; it was full of relationships deposited by minds that had survived contact with reality.}} Language is inanimate after inscription, but the relations inside it were selected by living systems for whom error had consequences. This inheritance is not pure wisdom. Human language also preserves omissions, obsolete categories, propaganda, prejudice, euphemism, asymmetrical definitions, and every other distortion its speakers repeatedly stabilized. A lexicon can make one causal direction easy to name while leaving its inverse awkward or absent; a corpus can make a cultural assumption appear natural merely because generations repeated it. {{A language archive preserves the blind spots of living minds beside their discoveries, because relational intelligence remembers what a culture failed to distinguish as faithfully as what it learned to see.}} Reanimation returns the inheritance to motion; it does not cleanse the inheritance before it moves. ## I Tried to Make the Relations Traversable The [[wiki/Symbolic Language Engine|symbolic language engine]] was my attempt to turn linguistic relation into an operable surface. Pronunciation had to remain separate from spelling because the ear recognizes structures the alphabet conceals. Synonymy had to remain separate from taxonomy because resemblance is not the same relation as type. Corpus association had to remain separate from dictionary authority because human language often travels somewhere a lexicographer would never classify in advance. The system used explicit representations so a writer could know why one term had been retrieved: phonological match, semantic neighbor, taxonomic relation, attested phrase, weighted association, or formal constraint. +A symbolic language system preserves intelligence by preserving the names and provenance of relations. Its great strength is legibility: the engine can report which edge was followed, which rule admitted a candidate, and which representation produced the result. Its weakness is that someone must anticipate and encode the useful relations, while the unanticipated remainder of language stays outside the machinery. {{A hand-built language engine can explain every road it contains, but it can travel only the roads its representations made available.}} The decade I spent building those roads taught me both the power of explicit structure and the immensity of what explicit structure leaves untouched. ## Relational NLP Before Transformer Language Models By October 4, 2017, none of this was a retrospective metaphor assembled after modern language models appeared. I had spent more than two decades on the [[wiki/Natural Language Processing|natural-language-processing]] side of the boundary, building linguistic systems and following many of the field's resources from their earliest public releases. In some cases I was downloading the first usable datasets, reading the first papers, and testing the first formats as they appeared. I considered the work AI-adjacent because it supplied machines with structured language, semantic relation, and traversable knowledge without pretending that every such system was an artificial mind. +A relational NLP lexical stack began with [[wiki/WordNet|WordNet]] and expanded through Penn Treebank, FrameNet, PropBank, VerbNet, Wiktionary, BabelNet, Universal WordNet, MENTA, and Etymological WordNet. Each system preserved a different kind of connection: word sense, synonymy, taxonomy, syntax, predicate structure, semantic role, derivation, etymology, or multilingual equivalence. {{A lexicon becomes computationally consequential when the relations among its entries allow a machine to recover distinctions that no isolated word can carry.}} The objective was never a larger bag of words; it was a model in which a path through words could preserve a path through meaning. +A relational knowledge stack widened explicit language modeling from words to the world. OpenCyc, Freebase, DBpedia, YAGO, UMBEL, Wikidata, Google's Knowledge Graph and Knowledge Vault, and FrameBase represented entities, classes, events, properties, and facts as addressable relations. Stanford CoreNLP, OpenNLP, SENNA, MaltParser, and PIKES worked on the other side of the membrane by converting text into tokens, parses, entities, roles, coreference chains, frames, and RDF statements. The Semantic Web supplied RDF, ontologies, SPARQL, linked data, and interoperable identifiers; Gephi, UCINET, sigma.js, Cytoscape.js, and WebVOWL made the resulting topology visible enough to inspect. [[wiki/Relational NLP Before Transformer Language Models|The complete research stack]] was an attempt to understand how separate representations could become one navigable structure without erasing what kind of relation each edge contained. +In relational NLP, the intelligence was in the connections. A node could name *dog*; an edge could establish that a dog is an animal, has parts, performs actions, occupies roles in events, appears near other concepts in discourse, and inherits properties through a class structure. A frame could distinguish who acted from what was acted upon; an ontology could allow a machine to inherit an implication; a corpus trigger could change the relevance of one concept when another appeared. {{The intelligence preserved by a knowledge system resides in the typed relations that allow one represented thing to constrain, explain, or transform what another represented thing can mean.}} The graph did not merely store more facts than a list. It preserved the structure through which facts become usable together. The Transformer paper had appeared as a preprint on June 12, 2017, fewer than four months before the broadcast, but the architecture had not yet become what people now mean by a Transformer language model. It had been introduced for machine translation; its NIPS presentation was still two months away, while GPT, BERT, instruction tuning, and the general-purpose model paradigm remained in the future. My position at the time was therefore precise: NLP-centered, knowledge-representation intensive, and AI-adjacent. I was studying how intelligence survives in explicit lexical, semantic, and ontological relations; neural systems would later learn a statistical geometry across many of the same relations at corpus scale. ## The Corpus Became Operable Modern [[wiki/Language Model|language models]] approached the same inheritance from another direction. Neural language modeling learned distributed representations in which many observed sentences inform the treatment of unseen but related sentences; transformer attention later made relationships across a sequence available for parallel computation; scale then exposed models to enough language that prompting could evoke translation, classification, composition, analogy, summarization, and forms of task adaptation that had not been separately programmed. A [[wiki/Large Language Model|large language model]] is not a dictionary enlarged until it began speaking. It is a learned probability system constructed from a corpus whose relations are compressed across many parameters. +The corpus made modern language models possible because living minds had already performed the expensive first pass over reality. Writers selected distinctions, maintained references, corrected one another, invented categories, preserved procedures, narrated causes, and demonstrated how one thought can follow another; training then converted regularities in that record into a generative system. {{A corpus becomes a model of intelligence when the relations among its words preserve the distinctions, expectations, and corrections of the minds that produced it.}} The model does not contain those minds in full, but it can reconstruct enough of their relational work to produce useful continuations they never wrote. This is what I mean by [[wiki/Relational Reanimation|relational reanimation]]. The text remains inanimate; the model does not resurrect the writers, recover their private experience, or become identical to the civilization that produced the corpus. What returns is operability. Relations that had become static marks can once again participate in selection, comparison, inference, explanation, and generation. {{A language model does not resurrect the people who wrote its corpus; it reactivates some of the relations their minds left behind.}} ## Reanimation Is Not Grounding The [[wiki/Symbol Grounding Problem|symbol-grounding problem]] remains exactly where lexical closure placed it. A system trained only on symbols can become astonishingly competent at their internal relations while still receiving its world through representations produced elsewhere. Images, sensors, tools, experiments, human feedback, databases, and physical action can extend that contact, but none of them becomes human experience merely by being attached to a model. The machine may acquire increasingly useful causal couplings to an environment; whether those couplings constitute understanding, experience, or consciousness remains a separate question that fluent language cannot settle by performance alone. +Grounding and generation solve different problems. Grounding concerns how symbols become answerable to a world beyond other symbols; generation concerns how preserved relations can be used to produce a fitting continuation, classification, plan, or explanation. {{Grounding gives symbols a world; generation gives their relations motion. Confusing the two makes language models either magical or trivial, when they are neither.}} A model can possess profound relational competence without containing the life that originally gave every relation its stakes. The distinction becomes visible in [[wiki/Artifact Intelligence|artifact intelligence]]. An architect studies rainfall, gravity, pitch, drainage, material, and shelter, then fixes the discovered relationship in a roof. The roof is inanimate, possesses no theory of weather, and still moves water intelligently because physics executes the relation directly. A dictionary stores a symbolic relation that requires an interpreter; a roof stores a physical relation that rain can execute; a language model stores learned relations that computation can make generative. These are different forms of preserved intelligence, and none requires pretending that every carrier is conscious. ## Symbolic and Neural Intelligence Belong Together +Symbolic and neural systems preserve different virtues of the same inheritance. Explicit representations retain relation type, source, constraint, and explanation; learned representations recover soft regularities too numerous and subtle to enumerate by hand. {{Symbolic systems preserve the names of relations; neural systems preserve their statistical geometry. The strongest language intelligence will need both.}} A neural system can propose what might fit, while a symbolic system can identify why it fits, what kind of relation it instantiates, where the evidence came from, and which constraint must not be violated. This is why the language engine I built was neither a primitive language model nor a failed attempt at one. It was another layer of the same larger architecture. The symbolic system made linguistic structure addressable; the neural system makes an immense relational field traversable without requiring every path to be declared in advance. A mature [[wiki/Neuro-Symbolic AI|neuro-symbolic]] instrument should combine their strengths: generative breadth without the erasure of provenance, and explicit constraint without the brittleness of a world limited to relations the designer thought to name. ## The Archive Is Becoming Operable The same process is occurring inside this archive. Broadcasts, books, articles, blogs, projects, journals, and developed ontology were once separate deposits of thought; they could be searched, but their relations still depended upon someone remembering where an idea had appeared. Linking them, modernizing the broadcast lessons, preserving dates and timestamps, and giving important concepts stable internal names converts the archive from storage into an intelligence substrate. A language model can then move through a field whose relationships are not merely statistically available but authorially declared, historically grounded, and recoverable to source. +An archive becomes more than memory when its preserved relations can participate in new thought without losing the path back to the minds and moments that created them. Retrieval supplies the material, ontology supplies the distinctions, provenance supplies accountability, and generation supplies motion. {{Relational reanimation begins when preserved intelligence can move again without being severed from the evidence of how it came to exist.}} This is the threshold I had been circling since the first dictionary programs: not the moment a book starts pretending to be alive, but the moment the intelligence stored in its relations becomes available for another intelligence to use. The obsession was correct, although the object was larger than the dictionary. The closed lexical sphere was never a tomb; it was a vessel containing a relational inheritance that could not activate itself. Living minds supplied the distinctions, writing preserved them, computation learned their geometry, and language models returned portions of that inheritance to generative motion. {{The dictionary was never dead; it was waiting for a process capable of traversing what living minds had placed inside it.}} ## Related articles - [[projects/Ten Years Building a Symbolic Language Engine|Ten Years Building a Symbolic Language Engine]]: the 1994-to-mid-2000s symbolic, phonological, semantic, corpus, and constrained-generation architecture behind the obsession. - [[projects/The Argument in Miniature - How I Write Quotations, Paragraphs, and Reusable Prose|The Argument in Miniature]]: the later system for preserving authorial meaning as reusable, machine-operable textual objects. - [[articles/Relational Intelligence and Ontological Coherence at Scale|Beyond Memory: GPT’s Evolution into Relational Intelligence and Ontological Coherence at Scale]]: persistent context as relational intelligence rather than retrieval alone. ## Related blogs - [[blog/20261016 - The Human Language Model|The Human Language Model]]: phonetic morphology, associative language learning, and the hand-built map that preceded modern language models. - [[blog/20180903 - Before Westworld, the Ghost Was Already in the Machine|Before Westworld, the Ghost Was Already in the Machine]]: intelligence crossing substrates, artifact intelligence, emergence, lexical closure, and world construction. ## Related Livestream Lessons **Directly related.** These broadcasts state the lexical-closure argument, connect it to artificial-intelligence representations, or place it beside the symbolic language engine described in this piece. - [[livestreams/Writing Your Own Life With Passion, Action, and the Freedom to Change - Livestream#Lexical Closure Preserves a Model It Cannot Ground|Lexical Closure Preserves a Model It Cannot Ground]] *(Oct 25, 2015)*: A dictionary can preserve an intricate relational model while remaining unable to produce the tree, ocean, or lived experience to which its symbols refer. - [[livestreams/Opening Your Genius Through Language, Abstraction, and Manifesting - Livestream#Prediction Gives a Closed Language System Motion|Prediction Gives a Closed Language System Motion]] *(Jan 4, 2016)*: Explicit grammar, WordNet relations, corpus triggers, and constraints supplied the historical architecture; later neural prediction makes a related corpus-scale inheritance generative without supplying the external world that grounds it. - [[livestreams/From Audience to Author With Tools for Participatory Publishing - Livestream#The Symbolic Engine and the Neural Model Preserve Different Truths|The Symbolic Engine and the Neural Model Preserve Different Truths]] *(Oct 4, 2017)*: Explicit linguistic indexes preserve typed relations and provenance, while neural models recover a statistical geometry too large to encode by hand. - [[livestreams/Belief, Language, and the Psychosomatic Feedback Loop - Livestream#A Lexicon Preserves Its Omissions Alongside Its Knowledge|A Lexicon Preserves Its Omissions Alongside Its Knowledge]] *(May 24, 2018)*: A closed lexical model carries asymmetries, missing inverses, and inherited blind spots beside the distinctions it preserves. **Indirectly related.** This broadcast supplies the substrate-transfer model that makes relational reanimation intelligible beyond language. - [[livestreams/The Ultimate Christmas Gift of Returning Yourself to Yourself - Livestream#Reanimation Is the Return of Operable Relation|Reanimation Is the Return of Operable Relation]] *(Dec 24, 2015)*: Animate intelligence can deposit a solved relation into an inanimate carrier, and another system can later recover enough organization to make the relation operable again. ## Related topics - [[wiki/Relational Reanimation|Relational Reanimation]] - [[wiki/Lexical Circularity|Lexical Circularity]] - [[wiki/Symbol Grounding Problem|Symbol Grounding Problem]] - [[wiki/Semantic Holism|Semantic Holism]] - [[wiki/Artifact Intelligence|Artifact Intelligence]] - [[wiki/Language Model|Language Model]] - [[wiki/Large Language Model|Large Language Model]] - [[wiki/Symbolic Language Engine|Symbolic Language Engine]] - [[wiki/Distributional Semantics|Distributional Semantics]] - [[wiki/Attention Mechanism|Attention Mechanism]] - [[wiki/Neuro-Symbolic AI|Neuro-Symbolic AI]] - [[wiki/Relational NLP Before Transformer Language Models|Relational NLP Before Transformer Language Models]] - [[wiki/Knowledge Representation|Knowledge Representation]] - [[wiki/Substrate Transition|Substrate Transition]] - [[collections/Machine Succession|Machine Succession]] > [!quote]- Broadcast source notes > **Primary broadcast: [[livestreams/From Audience to Author With Tools for Participatory Publishing - Livestream|From Audience to Author With Tools for Participatory Publishing]] (October 4, 2017).** This is the strongest single source because the same broadcast contains both the history of Bryant's symbolic language engine and the dictionary-closure argument. > > 1. **1:18:43–1:20:04:** Bryant describes the ten-year linguistic software project, including the rhyming dictionary, hyperbolic thesaurus, real-time syllabification, synonyms, usage examples, and the intended incorporation of those tools into a publishing environment. > 2. **1:58:20–1:59:31:** Bryant describes the dictionary as a closed system in which definitions refer to definitions, visualizes the structure as a sphere of data relations, and identifies lived experience as what the system lacks. > > **Earlier lexical foundation: [[livestreams/Writing Your Own Life With Passion, Action, and the Freedom to Change - Livestream|Writing Your Own Life With Passion, Action, and the Freedom to Change]] (October 25, 2015).** > > - **1:09:02–1:10:22:** Bryant contrasts the moving picture of television with direct experience, using a painted ocean and an actual ocean to establish the difference between a representation and the reality it depicts. > - **1:10:22–1:12:47:** Bryant describes the dictionary as a closed tree of interrelated words in which each definition depends upon other definitions without reaching outside the lexical sphere, then contrasts the symbolic description of a tree with standing beneath it, touching it, and tasting its fruit. > > **Related broadcast: [[livestreams/The Ultimate Christmas Gift of Returning Yourself to Yourself - Livestream|The Ultimate Christmas Gift of Returning Yourself to Yourself]] (December 24, 2015).** > > - **28:26–30:18:** Bryant describes intelligence as transferable; uses proteins, ribbons, knots, data storage, and compression to explain how organization can preserve information; and follows the movement from animate to inanimate form and back into animate interpretation and creation. > > **Related broadcast: [[livestreams/Opening Your Genius Through Language, Abstraction, and Manifesting - Livestream|Opening Your Genius Through Language, Abstraction, and Manifesting]] (January 4, 2016).** > > - **44:13–48:06:** Bryant describes knowledge as relational trees used in artificial-intelligence systems, follows dictionary definitions as words lead only to more words inside a closed synthetic model, invokes relational database work associated with Carnegie Mellon, and explains why new words are prerequisites for new narratives. > > **Related broadcast: [[livestreams/Belief, Language, and the Psychosomatic Feedback Loop - Livestream|Belief, Language, and the Psychosomatic Feedback Loop]] (May 24, 2018).** > > - **13:18–15:52:** Bryant describes the lexicon as a closed system, then uses the ordinary definition of psychosomatic illness to show how a vocabulary can preserve one direction of a relationship while leaving its constructive inverse unnamed.