The Meaning of Meaning: What Does It Mean to Understand?

When you type a single word into a search engine—say, "python"—you expect it to understand what you mean. Are you looking for the slithering reptile, the powerful programming language, or perhaps the comedy troupe? For decades, the answer has been a frustrating guess. The search engine doesn't truly "understand" your intent; it performs a mechanical tally of letters, counting how often "python" appears on a webpage and matching it against other words like "snake," "code," or "Monty." This mismatch between human intention and machine processing is not a minor bug; it is the fundamental chasm of the information age. We communicate in a world rich with meaning, nuance, and context, while our machines have historically operated in a barren landscape of symbols, devoid of semantics. This book is about how we are finally building a bridge across that chasm, a bridge made not of steel and concrete but of mathematics and data. The central thesis is both radical and simple: meaning itself can be transformed into a precise, computational language—a language of numbers and geometry. This first chapter establishes the problem in its full historical and philosophical depth. We will trace the lineage of symbol-processing AI, define the elusive concepts of "semantics" and "context" with rigor, and confront the profound gap between syntactic matching and true understanding. By the end, you will see why the old approach was doomed to fail and how a new paradigm—one that represents words, ideas, and concepts as points in a high-dimensional space—holds the promise of a machine that finally gets it.

The Age of Keywords: A Legacy of Symbol Manipulation

To understand the revolution we are living through, we must first appreciate the architecture of the old world. For the first half-century of computing, the dominant paradigm was symbolic artificial intelligence (often called "GOFAI," or Good Old-Fashioned AI). Its core assumption, inherited from the logicians and mathematicians who founded computer science, was that intelligence is the manipulation of abstract symbols according to formal rules. In this view, knowledge is a vast graph of interconnected facts, and reasoning is a process of traversing that graph using logical inference. A classic example is the expert system: a program that diagnoses medical conditions by applying a set of "if-then" rules crafted by human doctors. The system does not "know" what a fever is in any human sense; it merely recognizes the symbol `<FEVER>` in its working memory and fires the rule `IF <FEVER> AND <RASH> THEN <MEASLES>`. The power and the limitation are the same: it is a master of syntax, but it possesses zero semantics. It can process the symbol "bank" without ever conceiving of a riverbank or a financial institution.

This symbolic approach dominated early information retrieval. The first search engines, like Archie (1990) and later AltaVista (1995), were essentially automated librarians operating on a Boolean logic model. You gave them keywords—`cat AND video`—and they returned documents containing those exact strings. The engine's "understanding" was limited to set theory: a document either contained the term or it did not. There was no gradation, no association, no sense that "kitten" might be related to "cat" even if the word never appeared. The user was expected to be a precise librarian, anticipating the exact vocabulary the document's author used. This created a brutal skill gap: effective searching required learning arcane query syntax (`+`, `-`, `"exact phrase"`, `OR`) and guessing the author's lexical choices.

The commercial explosion of the web with Google (1998) did not initially change this fundamental model. Google's famous PageRank algorithm was a monumental advance in *ranking*—determining which of the millions of documents containing your keywords was most *important*—but it did not change the *matching* phase. It still relied on keyword frequency and placement. The query "Java" would return pages about coffee, the island, and the programming language, ranked by a combination of keyword prominence and link popularity. The engine could not disambiguate because it had no model of what "Java" *meant*. It had no concept of a *semantic field* or a *topic cluster*. It was a brilliant, scalable version of the same old symbol-counting.

This keyword paradigm had several fatal flaws from a human-centric perspective:

1. The Vocabulary Mismatch Problem: The user and the document author operate with different lexicons. A search for "heart attack" will miss a medical article titled "myocardial infarction." The system has no mechanism to bridge this lexical gap because it treats words as arbitrary strings. 2. The Polysemy Problem: A single word has multiple, unrelated meanings (*polysemy*). "Crane" can be a bird, a construction machine, or a ballet position. A keyword system returns a mashup of all senses, forcing the user to sift through noise. 3. The Synonymy Problem: Different words express the same concept (*synonymy*). "Big," "large," "enormous," "vast" are treated as entirely separate symbols. A document using only "enormous" will be missed by a query for "big." 4. The Context Ignorance Problem: Meaning is derived from surrounding words. "I went to the *bank* to deposit money" versus "I sat on the river *bank*." A keyword system sees the same symbol in both and cannot differentiate.

These are not mere inconveniences; they are symptoms of a deeper disease: the system has no representation of *meaning*. It operates on the surface form of language, the ink on the page, without access to the conceptual world that the ink refers to. This is the symbol-grounding problem: how do symbols (like the word "dog") get connected to the real-world concept (the furry, barking animal)? In a pure symbol system, they don't. They are just symbols pointing to other symbols, in an endless, meaningless loop. A dictionary defines "dog" as "a domesticated canine," but the system has no sensory experience of a canine, no concept of domestication, no understanding of loyalty or barking. It is a game of textual chess with no board.

Defining the Elusive: What Is Semantics, Anyway?

Before we can solve the problem, we must name it with precision. What do we mean by semantics? In linguistics and philosophy, semantics is the study of meaning. But "meaning" is a notoriously slippery concept. For our purposes, we can define it operationally: semantics is the relationship between a linguistic sign (a word, phrase, sentence) and the concept or referent it stands for in the world, as well as its relationship to other signs. This definition has two critical components: *reference* (the connection to the external world or mental concept) and *sense* (the network of relationships to other words).

Consider the word "bachelor." Its *reference* is the set of all unmarried adult males in the world. Its *sense* is defined by its relationships: it is synonymous with "unmarried man," antonymous with "husband," hypernymous (a broader category) with "man" or "person," and hyponymous (a narrower category) with "confirmed bachelor." A purely symbolic system can manipulate the string "bachelor" and even build a thesaurus of related strings, but it cannot *ground* the reference. It cannot *know* what it means to be an adult male, or what marriage is. It has the sense relations (if encoded) but lacks the experiential referent.

This is where context becomes paramount. The meaning of a sign is not fixed; it is *situated*. The sentence "The software has a bug" is straightforward. But "The entomologist found a bug" uses the same word with a completely different referent. The context—the surrounding words "entomologist" and "found"—activates a different sense of the word. Human understanding is a constant, dynamic process of context activation. We don't look up words in a mental dictionary in isolation; we use the entire