Is Language the Boundary of Human Intelligence?

Video by Grok Imagine using ChatGPT generated AI image as input

"Language is the boundary of human intelligence" has been a prevailing view following the rise of Large Language Model Artificial Intelligence (LLMs).

Artificial Intelligence has long promised more than it delivered. Since the concept of Artificial Intelligence was first coined in the 1950s, it has gone through numerous development paradigms supported by different mathematical theories. Yet the results were often underwhelming. In the absence of sufficient returns on investment, AI endured two "AI winters," at times reduced to a marginal academic discipline.

The Key Breakthrough in Recent AI: Hardware Compute + Digitization + Massive Language Text

It was not until the 2010s—when Deep Learning techniques were combined with massive corpora of human language as training data that the birth of Transformer-based Large Language Models brought about a genuine technological breakthrough.

For this reason, textual data that carries human knowledge through language has widely been regarded as the key to transformative change. Of course, the digital technologies that enable large-scale storage and retrieval of text, along with the hardware that provides computational power, are the critical infrastructure that allows language to function as a carrier of intelligence.

Current breakthroughs in AI seem to demonstrate that the most valuable asset of human civilization is not merely information itself, but the logical structures and modes of thinking organized through language. As long as we can continue to digitize such knowledge efficiently and provide sufficient computational resources, the upper bound of artificial intelligence may keep rising.

Language as the Compression Algorithm of Human Experience and Knowledge

Taking a step back, why has language rather than the various mathematical theories repeatedly tested by scientists been more effective in propelling AI forward?

The likely reason is that language does not merely record and disseminate knowledge. It enables knowledge to be transmitted across generations, across geographical boundaries, and even across the divide between humans and machines.

In other words, language can:

• Record and compress human experience, indirectly carrying the "world model" of a linguistic community;

• Enable intergenerational and cross-regional transmission of knowledge;

• Make knowledge transferable between humans and machines.

Put differently, Large Language Models achieve their feats not because they understand in a human sense, but because they have absorbed the most efficient representational system humanity has ever devised. Language is not a mere heap of symbols, it embodies thousands of years of accumulated knowledge and experience, transmitted across time and space, and most crucially made transferable between humans and machines.

By feeding vast amounts of text into Deep Learning models, we are effectively delivering a compressed "world model" to machines. Language abstracts complex reality, allowing knowledge to detach from the human body and achieve a form of "digital resurrection" on silicon chips.

Non-Linguistic Learning, Experience, and Expression

However, the view that "language is the boundary" has not won universal support among AI scholars.

2018 Turing Award laureate Yann LeCun argued that intelligence cannot be fully encapsulated by language alone. Otherwise, Large Language Models would already possess the self-learning and environmental adaptability that cats, dogs, and other non-linguistic animals display.

In their August 2022 paper, "AI and the Limits of Language", Yann LeCun and New York University scholar Jacob Browning further contended that language carries only a small fraction of human knowledge. Most human knowledge and nearly all animal knowledge is non-linguistic. Therefore, Large Language Models trained solely on textual data will never attain the full spectrum of human intelligence.

At a time when the media was overwhelmingly praising Large Language Models such as ChatGPT, Yann LeCun even reiterated on social media that before we reach "God-like AI," we must first achieve "dog-like AI".

Spatial Intelligence: Visualizing the Physical World

Fei-Fei Li, a pioneer often dubbed the ‘Godmother of AI’, notes that Spatial Intelligence—which must grapple with the laws of physics—remains beyond the grasp of language models alone. The same, she suggests, is true for the even loftier goal of Artificial General Intelligence (AGI).

Recently, DeepSeek-OCR, launched by DeepSeek, highlighted this trend. Through an OCR-based visualization mechanism, the model renders text into images, performs contextual compression and logical verification visually, and then converts the result back into compressed text. This move from "1D text" to "2D visual compression" acknowledges that language has gaps that only non-verbal mediums can bridge.

The core innovation of DeepSeek-OCR lies in what might be termed "visual-as-compression" (Contexts Optical Compression):

• It can compress a 1,000-word document into approximately 100 visual tokens, achieving roughly a tenfold compression while maintaining about 97% accuracy;

• By rendering long text into images and extracting visual features, it alleviates computational and memory pressures faced by large language models when handling ultra-long contexts, opening up the possibility of "infinite context".

DeepSeek's Contexts Optical Compression approach suggests that the path to "infinite context" lies in looking at the world, not just reading about it. In other words, the current progress in AI suggests that intelligence's frontier extends well beyond prose.

From LLM to LMM: The Multimodal Evolution

The public still tends to conflate AI with the "Large Language Model" (LLM). But the frontier has already moved toward the "Large Multimodal Model" (LMM). These systems do not just read; they see, hear, and synthesise information across different modes.

The evolution of LLMs into multimodal systems is, in itself, a concession: it proves that the boundaries of intelligence are far wider than the lexicon.

The Perils of a Linguistic Bias

To treat language as the sole boundary of intelligence is to risk a subtle but consequential bias. It privileges eloquence over experience and textual mastery over embodied exploration. It is the epistemological equivalent of praising "reading ten thousand books" while neglecting the value of "travelling ten thousand miles".

Art offers a reminder. Painting, sculpture, music and dance emerge where language falters. They probe emotional and physical realities that resist tidy articulation. Words can gesture towards such experiences; they rarely exhaust them.

An AI ecosystem trained exclusively on text risks institutionalising a kind of bookish evolutionism: a belief that intelligence consists primarily in verbal dexterity. Yet much of what enables humans to survive and innovate—motor coordination, spatial reasoning, tacit skill, emotional attunement develops through embodied interaction with the world.

Reading matters. But so does doing.

Language as Portal, Not Perimeter

None of this diminishes the achievement of Large Language Models. Their success vindicates decades of research and demonstrates that language is a formidable substrate for modelling aspects of thought.

The danger lies elsewhere: in mistaking fluency for comprehension, verbosity for wisdom. When the ability to generate plausible sentences is equated with understanding, both humans and machines may be overestimated. Meanwhile, the opacity and unpredictability of reality are underestimated.

Language is not the boundary of human intelligence. It is an entrance—a doorway through which we access and organise complexity.

True intelligence, whether biological or artificial, demands that we step beyond that doorway: into perception, action, embodiment and the unruly texture of the physical world.

Language opened the current chapter of AI. It is unlikely to close the book.

Note:

This is the English translation of my article published on Malaysian Chinese News Portal Oriental Daily. The translation was by ChatGPT and Google Gemini.
This article was developed using Google NotebookLM, Gemini, ChatGPT, and DeepSeek to analyse materials related to the 2026 International Chinese Debating Championship semi-final topic: "Language is/is not the boundary of humanity".