WORDS
PREDICT THE NEXT TOKENThe training objective looked much smaller than the behavior that emerged.FIELD GUIDE · LLMS, ANIMALS & THE LANGUAGE OF THOUGHT
IS LANGUAGE THE SECRET OF INTELLIGENCE?
We trained machines on words and got reasoning, coding and abstraction back. That is astonishing evidence about language. It is not proof that thought itself is made of words.
UPDATED 2026-09-22
THE PROBLEM
WE FED IT WORDS. IT LEARNED MORE THAN WORDS.
That is evidence about language. It is not proof that thought is made of language.
A large language model learns statistical structure by predicting pieces of language. Scale that process far enough and the model becomes useful at tasks nobody explicitly programmed one by one: translation, question answering, coding, analogy, arithmetic and more. The philosophical temptation is obvious: maybe language is the hidden machinery of intelligence. The catch is that human brains can reason when language is badly damaged, animals solve problems without human language, and language itself is packed with knowledge produced by minds before the model ever saw the text.
REASONING
NOW DO ARITHMETICLanguage models acquired abilities nobody hand-coded one by one.APHASIA
LANGUAGE GONE. LOGIC STAYS.Human reasoning can survive profound language impairment.ANIMALS
NO ENGLISH REQUIREDCausal inference, planning, curiosity, and tool use exist without human language.THE FIGHT
IS LANGUAGE THE ENGINE, THE INTERFACE, OR THE ARCHIVE?
If language is the core machinery of general intelligence, then the success of LLMs is evidence that AI found the right substrate surprisingly early. If language is mostly an interface over deeper reasoning systems, then LLMs may be learning from a compressed record of human thought without revealing how biological intelligence itself works. The answer changes how we interpret model capabilities, animal minds, education, human uniqueness, and the path toward more general AI.
THE LANGUAGE-FIRST CASE
LANGUAGE MAY BE A COGNITIVE SUPERPOWER
Language gives minds a compact system for naming abstractions, combining concepts, receiving instructions, rehearsing possibilities, and importing discoveries made by other people. LLMs are striking because broad abilities emerge from learning that linguistic structure at scale. Experiments also show that natural-language instructions can scaffold flexible generalization in neural models, and cognitive theories argue that language can shape or enhance human thought even when thought is not reducible to language.
Language may be less like a dictionary and more like a programmable control panel for abilities underneath it.
THE THOUGHT-FIRST CASE
LANGUAGE MAY BE THE RECORD, NOT THE ENGINE
Human neuroscience separates core language processing from several systems used for knowledge and reasoning. People with severe aphasia can retain logical abilities, and many nonhuman animals show forms of causal inference, planning, information seeking, and tool use without human language. On this view, LLMs succeed partly because language is full of traces left by intelligent minds rather than because intelligence itself is made of language.
A library contains civilization. That does not mean civilization is made of paper.
THE AMPLIFIER CASE
LANGUAGE COULD BE THE MULTIPLIER
The cleanest synthesis is that substantial cognition can exist without natural language, while language radically expands what minds can share, recombine, preserve, and learn from one another. Cultural-transmission experiments show that languages themselves become structured as they pass through generations. Language may therefore amplify intelligence at the individual level and multiply it at the cultural level without being the original source of thought.
Fire existed before the library. The library lets one person's fire-starting trick survive for centuries.
THE WEIRD SHIT
THE HUMAN BRAIN KEEPS RUINING THE CLEAN STORY
TEXT PREDICTION PRODUCED SURPRISINGLY BROAD BEHAVIOR
The GPT-3 paper showed that scaling an autoregressive language model trained on large text corpora substantially improved few-shot performance across translation, question answering, cloze tasks, word manipulation, arithmetic, and other tasks without task-specific gradient updates at evaluation time.
That is strong evidence that linguistic training data contains reusable structure far beyond memorizing phrases. It does not identify the mechanism of human intelligence.LANGUAGE AND REASONING ARE NOT ONE BRAIN NETWORK
A 2024 Nature perspective reviewing neuroscience and related evidence argues that the core language network is distinct from systems used for many forms of thought, and that complex thought does not require natural language.
If language were the universal medium of human thought, this separation would be difficult to explain.SEVERE LANGUAGE DAMAGE CAN LEAVE FORMAL LOGIC INTACT
A 2026 PNAS study combined fMRI with tests of people who had extensive damage to language areas and severe aphasia. The authors reported intact inductive and deductive logical reasoning and found that the language network was not required for those tasks.
At least some sophisticated human reasoning can proceed without usable natural-language representations.ANIMALS DO COGNITIVE WORK WITHOUT HUMAN LANGUAGE
Comparative cognition research documents causal and inferential abilities in nonhuman animals, while recent reviews distinguish the broader cognitive capacities humans share with other species from the unusually abstract, decontextualized forms that become prominent in humans.
Language may help explain the scale and flexibility of human cognition without being a prerequisite for every form of intelligence.LANGUAGE CAN SCAFFOLD NEW TASKS
A 2024 Nature Neuroscience study trained neural models to perform families of psychophysical tasks and found that natural-language instructions supported compositional generalization to previously unseen tasks.
Language can act like a reusable control signal that helps combine existing skills in new ways.LANGUAGE IS A CULTURAL TRANSMISSION MACHINE
Iterated-learning experiments have shown that artificial languages passed between human learners become more structured and easier to transmit across generations.
Language does not merely express one person's intelligence. It lets cognition accumulate outside any one skull.THE PEOPLE WITH A STAKE
WHY THIS OLD PHILOSOPHY FIGHT GOT EXPENSIVE
AI LABS
A language-first interpretation makes continued scaling, better data, longer context, tools, and instruction following look like a direct road toward broader intelligence. A thought-first interpretation suggests important missing machinery may still sit outside language modeling.
COGNITIVE SCIENTISTS
LLMs create a new experimental object for old questions about symbols, reasoning, concepts, and communication. But similarity in behavior does not automatically imply the same internal mechanism.
ANIMAL COGNITION
If intelligence requires human-like language, sophisticated animal problem solving becomes hard to explain. Comparative research documents cognitive abilities in species without human language while asking what language uniquely adds in humans.
EVERYONE LEARNING WITH AI
If language is a powerful scaffold rather than merely a reporting channel, conversation itself can become a way to assemble unfamiliar skills and concepts. That makes the interface between human and machine part of the cognitive system.
THE UNANSWERED QUESTION
WHAT WOULD ACTUALLY PROVE LANGUAGE IS THE KEY?
- A convincing demonstration that broad human reasoning consistently collapses when natural-language processing is unavailable, rather than surviving in other cognitive systems.
- AI systems trained on rich nonlinguistic sensory and action data outperforming language-centered systems across general reasoning would weaken the claim that language is the privileged route to intelligence.
- Conversely, language-only systems repeatedly acquiring broad new cognitive abilities that comparable nonlinguistic systems cannot learn would strengthen the language-first case.
- Mechanistic evidence showing that successful LLM reasoning depends on internal structures that closely match the representations used in human reasoning, rather than merely producing similar answers.
- Comparative evidence identifying which cognitive jumps appear only when a species gains a sufficiently expressive communication system and cumulative culture.
TAKE THIS TO DINNER: LLMs show that words carry far more cognitive structure than we realized. They do not prove that intelligence itself is linguistic.
RECEIPTS
LLMs made the old language-versus-thought argument newly empirical. The receipts run from GPT-3 to aphasia, animal cognition, cultural transmission, and neural models of instruction.
- Language Models are Few-Shot LearnersOpenAI / arXiv
- Language is primarily a tool for communication rather than thoughtNature
- Evidence from formal logical reasoning reveals that the language of thought is not natural languagePNAS
- The development of human causal learning and reasoningNature Reviews Psychology
- Natural language instructions induce compositional generalization in networks of neuronsNature Neuroscience
- Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human languagePNAS
- The Centrality of Language in Human CognitionLanguage Learning