Riley Argues Language Isn't Thought, Straining the LLM-AGI Bet

Every AGI pitch on stage right now leans on one implicit bet: scale up language and thinking will follow. In an essay for The Verge, Benjamin Riley, founder of Cognitive Resonance, argues that bet is miscast at the level of neuroscience. His line is blunt: "human thinking is largely independent of human language," and modeling one ever more sophisticatedly will not conjure the other.

The case rests on evidence he says has been piling up for years. Functional MRI work shows different brain regions activating for different cognitive activities, meaning the neurons doing a math problem are not the neurons handling a sentence. People who lose language after brain injury, he notes, still solve math problems, follow nonverbal instructions, and read emotion. He points to a Nature commentary summarizing decades of that research. If language were the substrate of thought, none of that would hold.

The industry read-across is where this gets uncomfortable. Frontier roadmaps assume that scale plus data eventually cross into general intelligence. Riley's counter is that the architecture is doing something narrower: "LLMs are simply tools that emulate the communicative function of language, not the separate and distinct cognitive process of thinking and reasoning, no matter how many data centers we build." Take that seriously and the capex story shifts from a bet on emergent cognition to a bet on very good text generation.

He softens none of it in the close. Even a strong model, on his account, is "forever trapped in the vocabulary we've encoded in our data and trained it upon," a dead-metaphor machine useful for remixing what we already know but not for producing new thought.

The honest caveat is that this is one operator's synthesis, aimed squarely at what Riley calls "AI-exuberant CEOs," not a peer-reviewed refutation, and plenty of researchers believe language and reasoning are more entangled than he allows. What the reporting doesn't give you is a falsifiable test the industry could run to check whether the next scaling jump is buying thought or fluent stenography. The useful move for anyone allocating budget is to stop treating LLM benchmarks as evidence of reasoning and start asking, per use case, whether emulated communication is actually all you need. For a surprising amount of enterprise work, it is, and pricing accordingly is the calmer bet than waiting on AGI.