Summary of Key Points
This article explores the millennia-old debate about the relationship between language and thought, presenting a groundbreaking conclusion based on the latest research in brain science from MIT and advancements in the field of AI: **Language is a tool for communication, not a necessary condition for thinking (especially logical reasoning)$. Studies on aphasic patients (who cannot speak but can solve logical problems) and brain imaging techniques have shown that the core processes of thinking occur outside of language areas. Evidence from AI research—such as the perspectives of Yang Likun, Claude’s concept of the “J-space”, and the modular structure of large models—also supports this conclusion. This suggests the future direction for AI: it should first develop a “world model” similar to that of infants, then learn reasoning skills, and only later acquire language abilities, rather than relying solely on language training.
A Millennium-Old Philosophical Debate Finally Resolved by Science
From Plato’s view that “thinking is a silent dialogue of the soul” to Wittgenstein’s early belief that “language is a prison for the world”, there has always been confusion about whether one can think without language. MIT’s research has dispelled this myth through experiments:
- The “miracles” of aphasic patients: Two patients who could barely speak or understand language were able to solve complex logical problems and explain patterns using gestures, demonstrating that complex reasoning is possible without language.
- The truth revealed by brain scans: When healthy individuals engage in reasoning, the language areas of the brain are virtually inactive, while the “multiscale network” (responsible for solving complex problems and working memory) becomes active. The neural mechanisms for deductive (using existing rules to draw conclusions) and inductive (identifying new patterns) reasoning differ; in the former case, even the multiscale network is hardly used.
Conclusion: Thinking, especially reasoning, is an independent “background process” that occurs independently of language, with language merely serving as a medium for conveying the results.
AI Expert Yang Likun Supports This Conclusion
Yang Likun, a Turing Award winner, has criticized large language models (LLMs) for their lack of physical intuition and causal understanding. MIT’s research supports his argument: If the human brain does not use language areas for reasoning, how can AI trained solely with language learn true intelligence? It’s like trying to play the piano by only looking at sheet music—never mastering the skill. Yang believes that the core of human intelligence lies in “predicting real-world events”, a process that does not rely on language. MIT’s findings confirm that non-verbal reasoning is key to intelligence, and AI should focus on developing this ability rather than memorizing language patterns.
AI Models Also Have Their Own “Silent Thinking Spaces”
Anthropic discovered a mysterious “J-space” within the Claude large model: it processes reasoning silently without producing text. This space did not emerge through training but is an inherent part of the model’s functionality, similar to the background processes in the human brain. This further validates MIT’s conclusion that language is merely an output interface, not the core computing mechanism. Even pure language models must use non-verbal processes for reasoning.
Large Models and Human Brains Share Similar Structural Patterns
Another MIT study found that large models have a modular organization similar to the human brain: different neural clusters are dedicated to specific tasks (language, logic, physics, social reasoning). Experiments showed that disabling the language module does not affect physical reasoning abilities (although speech becomes impaired), while disabling the physics module results in fluent sentences but incorrect conclusions (e.g., suggesting that a ball will float in the air). This indicates that intelligence, whether based on carbon (the human brain) or silicon (AI), follows similar organizational principles—with clear divisions of labor between language and reasoning functions.
The Root Cause of AI’s “Delusions” and the Way Forward
Why do large models sometimes produce illogical outputs? Because they learn from compressed representations of human knowledge (e.g., “fire is hot” as a generalization), without real-world experience to verify these concepts. They can generate grammatically correct sentences that do not reflect reality. To improve AI, we should follow the developmental sequence of infants:
1. Underlying world model: Understand fundamental physical principles and causal relationships.
2. Intermediate reasoning engine: Learn to identify patterns and draw logical conclusions.
3. Top-level language interface: Finally, acquire language skills to express thought processes.
Current large models follow the reverse approach, focusing first on language before attempting to develop intelligence from it, which is not effective.
Conclusion
More than two thousand years ago, Plato believed that thinking was a silent dialogue. Today, brain scans reveal that true thinking involves silent processes beneath the surface of language. Language is a remarkable invention of intelligence, but it is not intelligence itself. This conclusion not only resolves philosophical debates but also points the way for AI: instead of focusing on language, we should emulate the development of infants and create systems that can understand the world.