Chapter 16 The invention of symbol
9 minute read
For nearly four billion years, biological evolution remained trapped inside a single channel of memory. Every adaptation had to pass through the vertical replication of DNA from parent to offspring. A wolf or a hominid could spend decades learning the contours of a valley or the craft of flaking stone. Yet when that animal died, every trace of its hard-won neural mapping perished with its brain. Only variations that altered survival and reproduction could trickle through the slow genetic sieve across thousands of generations.
Between 300,000 and 100,000 years ago in East and Southern Africa, a second inheritance system caught fire among early Homo sapiens. Cumulative cultural evolution began to outpace the gene. Its engine was symbolic language, an open-ended protocol that encodes structured thoughts into acoustic pressure waves between 20 and 20,000 hertz. Through speech, one mind could reconstruct a complex mental graph directly inside another mind across a campfire. Knowledge no longer died with the body that discovered it. Each generation could stand on the insights of its ancestors, turning human communities into collective brains that spanned centuries.
16.1Recursive grammar and the architecture of language
What separates human speech from every other form of animal communication on Earth? Nature is full of vivid signals. Vervet monkeys utter distinct alarm calls for leopards, eagles, and pythons, sending their troop racing up trees or scanning the grass. Honeybees perform figure-eight waggle dances in the dark of the hive to report the angle and distance of distant wildflowers. Songbirds weave intricate melodies to defend territory and attract mates. Yet every one of these animal codes remains a closed repertory of fixed signals. A monkey cannot join the call for leopard with the call for eagle to describe a novel predator. Nor can it tell a story about yesterday's storm or tomorrow's hunt.
In 1960, the linguist Charles Hockett mapped the architectural features that set human language apart, culminating in duality of patterning and discrete infinity. Human speech operates on two distinct combinatorial levels at once. At the first level, phonology, the vocal tract carves continuous breath into 15 to 80 meaningless sound units called phonemes. This vocal precision became possible after the human larynx descended in the throat. At the same time, cortical motor circuits linked to the FOXP2 gene evolved fine control over lips, tongue, and breath. Taken in isolation, individual phonemes carry no meaning at all. Yet by permuting that tiny alphabet of sounds, a community builds tens of thousands of meaningful words and roots called morphemes.
At the second level, syntax, speakers combine those finite words through recursive rules into a countably infinite repertoire of sentences. As the logician Gottlob Frege observed, the meaning of a complex sentence flows systematically from its parts and their grammatical arrangement. Because the rules act on abstract structures rather than fixed strings, speakers routinely utter and understand completely novel sentences.
In 1956, Noam Chomsky unified linguistics and computer science by showing how formal rewrite grammars generate infinite languages from finite means. Every formal grammar has four distinct components. It combines a finite set of abstract grammatical categories, a disjoint finite vocabulary of terminal words, a set of rewrite rules, and a designated start symbol among the categories. Chomsky arranged all formal grammars into a four-tiered hierarchy of rising computational power. Each tier corresponds to an abstract machine with a specific kind of memory.
At the base of the hierarchy sit Type 3 regular grammars, whose right-linear rewrite rules extend a phrase by one word at a time from left to right. A regular grammar corresponds to a finite-state automaton, a machine with a fixed number of internal states and no external memory. One step higher sit Type 2 context-free grammars, which expand a single grammatical category into a phrase regardless of the surrounding words. Recognizing a general context-free language requires a nondeterministic pushdown automaton equipped with an expandable stack memory to track nested phrases. Above these lie Type 1 context-sensitive grammars, whose rules depend on the surrounding grammatical neighborhood and never shrink the length of a string. These require a linear-bounded automaton whose memory scales proportionally to the length of the sentence. At the summit sit Type 0 unrestricted grammars, which possess the full computing power of a universal Turing machine.
In his 1957 monograph Syntactic Structures, Chomsky proved that human language cannot be generated by any Type 3 finite-state machine. The reason lies in a structural hallmark of human thought called center-embedding. A speaker can nest one relative clause inside another to arbitrary depth. Consider the classic English sentence, "The rat that the cat that the dog chased killed ate the malt." Here three noun phrases stack up in succession before their three matching verbs arrive in reverse order, following the nested pattern of opening subjects paired with closing verbs.
By the pumping lemma for regular languages, no finite-state machine with a fixed number of states can match that nested symmetry once the embedding depth exceeds the number of states. To link each subject to its verb across an intervening clause, the listener must hold unfinished phrases on a mental stack. Even more strikingly, as the linguists Riny Huybregts and Stuart Shieber demonstrated, Dutch and Swiss German subordinate clauses interleave nouns and verbs in parallel cross-serial order rather than mirror order. By the pumping lemma for context-free languages, those parallel cross-serial dependencies break Type 2 context-free grammars as well. Human syntax therefore occupies the mildly context-sensitive region strictly between Type 2 and Type 1 grammars.
Decades later, in his Minimalist Program, Chomsky distilled the core engine of human syntax down to a single recursive operation called Merge. Given any two syntactic objects, Merge combines them into an unordered binary set:
That new set can then serve as an input to the next stage of combination. Notice the deep bridge back to the foundations of mathematics in Book I. Human syntax is the set-theoretic axiom of pairing enacted in living neural circuits, recursively assembling hierarchical trees of concepts inside the brain.
Yet how did such an elegant formal engine arise in biological history? Modern cognitive and evolutionary linguists, including Michael Tomasello, Adele Goldberg, Morten Christiansen, and Simon Kirby, uncovered the cultural mechanism. Recursive grammar did not require a sudden genetic miracle. Instead, language evolved by adapting itself to the human brain across thousands of generations of cultural transmission. In iterated learning experiments, each generation of learners must infer an entire language from a small sample of childhood utterances. Unstructured holistic calls cannot survive that narrow memory bottleneck. Only compositional grammars have low enough algorithmic description length to be learned reliably from sparse examples. Under the relentless pressure of cultural transmission, chaotic signals spontaneously crystallize into regular, hierarchical syntax.
16.2Information entropy and the invention of script
Once spoken language emerged, how efficiently did it pack information into sound? In 1948 and 1951, Claude Shannon turned his mathematical theory of communication onto printed and spoken English. He modeled a language as a stationary stochastic source that emits a stream of characters from a finite alphabet. Shannon defined the entropy rate of the language as the average information carried by each new character given all preceding text:
Consider printed English with its 26-letter alphabet. Equal letter probabilities would give a maximum capacity of 4.70 bits per letter, or 4.75 bits per symbol when including a blank space as a 27th character. Real language is heavily skewed. Among the 26 letters of English text, the letter e occurs 12.7% of the time whereas the letter z appears only 0.07% of the time. At the level of whole words, vocabulary obeys George Kingsley Zipf's power law, where the frequency of the nth most common word is inversely proportional to its rank. Accounting for single-letter frequencies alone drops the 26-letter entropy rate from 4.70 down to 4.14 bits per letter, or 4.03 bits per symbol with spaces included.
Shannon then measured long-range grammatical and semantic predictability across whole sentences by having human readers guess upcoming letters in a text. He found that the true asymptotic entropy rate of English falls to between 1.0 and 1.3 bits per letter. Comparing this actual information rate to the 26-letter maximum capacity of 4.70 bits per letter reveals that human language is roughly 75% redundant, with .
Why does every natural language waste three-quarters of its theoretical symbol capacity? Shannon's noisy-channel coding theorem supplies the answer. Spoken conversation rarely takes place in a silent laboratory; it unfolds across windy plains, crackly campfires, and crowded rooms. That 75% structural redundancy acts as a built-in error-correcting code. It allows a listener's brain to reconstruct missing syllables and misheard consonants without losing the speaker's message.
Even more remarkably, comparative studies led by François Pellegrino and Christophe Coupé across 17 languages uncovered a universal trade-off between syllable complexity and speaking speed. Languages such as Mandarin and English pack high information density into complex syllables at roughly 7 bits per syllable. Speakers articulate those dense syllables at a measured pace of about 5.5 syllables per second. Conversely, languages such as Spanish and Japanese use simpler consonant-vowel syllables that carry only about 5 bits per syllable. Speakers utter those lighter syllables rapidly at nearly 7.8 syllables per second. When you multiply information per syllable by syllables per second, tongues across the globe converge on roughly 39.15 bits per second. Regardless of grammar or vocabulary, every human language pours meaning through the same biophysical bottleneck in the auditory and prefrontal cortex.
Despite its power, spoken language suffered from one fatal physical flaw. Acoustic pressure waves dissipate into thermal motion within milliseconds of leaving the lips. For tens of thousands of years, human culture could preserve only what living elders could hold in biological memory before they died.
Then, in four independent crucibles of early agriculture, humanity learned to freeze ephemeral speech into solid matter. Between 3400 and 3200 BCE in Mesopotamian Uruk, temple clerks pressed clay accounting tokens for grain and sheep into wet clay tablets. That bookkeeping gave birth to Sumerian cuneiform, while Egyptian scribes at Abydos along the Nile developed hieroglyphs. Around 1200 BCE in late Shang dynasty China at Anyang, diviners carved oracle bone script into ox shoulder blades and turtle shells. Around 600 BCE in Oaxaca, Zapotec scribes and later Maya scribes around 300 BCE chiselled glyphs into Mesoamerican limestone monuments. Writing had entered the world.
Across the centuries, writing systems evolved through three structural stages of algorithmic compression. The earliest scripts were logographic and morphosyllabic systems that assigned a distinct visual glyph to every word or morpheme. Because a working vocabulary spans thousands of concepts, logographic scripts required 1,000 to 10,000 intricate characters. Memorizing that vast inventory demanded a decade of specialized training, confining literacy to a small caste of palace and temple scribes.
To escape that visual explosion, ancient scribes exploited the rebus principle. They borrowed the picture of a concrete object to represent any abstract word or syllable that shared the same spoken sound. That insight produced syllabaries such as Mycenaean Linear B, Japanese kana, and Sequoyah's Cherokee script. By mapping symbols onto whole spoken syllables, syllabaries compressed the sign inventory down to between 50 and 100 characters.
The final compression occurred between 1800 and 800 BCE along the eastern shore of the Mediterranean. Semitic-speaking miners in the Sinai adapted Egyptian hieroglyphs into the Proto-Sinaitic script, recording only consonant sounds. By 1050 BCE, Phoenician merchants standardized that system into a 22-letter consonantal abjad, and around 800 BCE Greek speakers repurposed surplus consonant signs as dedicated letters for vowels. By splitting the syllable all the way down to the atomic phoneme, the alphabet compressed written speech into just 22 to 30 characters. A script that once required 10 years of priestly apprenticeship could now be mastered by a child in a few weeks. By slashing the cognitive cost of literacy, the alphabet broke the scribal monopoly on external memory and ignited philosophy and science—all sustained by the thermodynamic surplus of the agrarian state.