For decades, the digital world has treated words as the ultimate carriers of meaning. Large Language Models (LLMs) like ChatGPT and Gemini ingest trillions of them—an avalanche of text tokens—just to produce coherent sentences. Yet, beneath the awe of their linguistic power lies a problem that is increasingly hard to ignore: inefficiency.
Each word, punctuation mark, and even space becomes a discrete data point to process. The result is immense computational load, ballooning energy consumption, and training pipelines that strain both hardware and the planet.
But what if the bottleneck isn’t in the size of the models, but in the form of the data?
Rethinking the Token
A “token” is the atomic unit of meaning for an LLM—the smallest chunk it understands. In English, these are usually word fragments: “run”, “ing”, “ly”. To understand a simple sentence, the model may process dozens of tokens; to capture an idea, millions more. The system, in effect, reads the world one syllable at a time.
Yet humans don’t think in words alone. We think in relations, images, and associations. A sketch, a symbol, a gesture—each can encapsulate what might take a paragraph to explain.
That old adage, a picture is worth a thousand words, may now be more than a cliché; it might be a blueprint for a new kind of efficiency in machine understanding.
Pictographs as Meaning Units
In a forthcoming paper, I propose a reimagined token system: one built not on words, but pictographs—small, structured icons that carry dense meaning.
If I wanted to express "Spelling is illogical", I might use a single icon to encode "spelling", “therefore” to represent logic, ! to represent “not”, and ↔ "meaning in normal conversation”. Together, they form short, composable sentences that are both human-legible and machine-readable, something like:
abc ∴ ! ↔
abc -> spelling (text)
! -> not (computer code)
∴ -> logical (maths)
↔ -> normal conversation
This is more than shorthand; it’s semantic compression. Each pictograph bundles what used to be several tokens into one meaning-rich unit — an echo of how Chinese characters combine context, domain knowledge and specific meaning, or how mathematical notation condenses entire logical statements into symbols.
Think how Maths takes a 'word' problem and encapsulates it into symbols.
"A notebook costs $4 and a pen costs $2. If a student buys x notebooks and y pens, how much money does the student spend in total?"
TC=4x+2y
This is a 100-fold compression. This is massive considering the exponential nature of the size data training sets required to facilitate learning the substance of the statement.
Efficiency Through Density
An image-based token carries multiple layers at once:
- Meaning (the object or action it represents)
- Relation (its logical or causal ties to others)
- Context (domain, time, or emotional tone)
Instead of parsing thousands of isolated word-tokens, a model could operate on a far smaller, denser vocabulary—closer to how humans intuit patterns visually and relationally.
Learning from Chinese Script
The system borrows inspiration from Chinese writing, where characters combine radicals (semantic classifiers) and contextual information into highly compact signs.
Our icons do something similar: a core symbol for logic (∴) joined with a domain radical (↔ for conversation, ⚖ for morality, ∑ for mathematics). This preserves nuance while dramatically cutting redundancy.
Why This Matters
The promise of such a system is profound. If successful, it could:
- Reduce model size and energy consumption by minimizing redundant textual data.
- Increase interpretability, since each icon has a transparent structure and defined meaning.
- Bridge modalities, allowing seamless movement between image, text, and symbolic reasoning.
LLMs might no longer need to be lumbering giants powered by endless text. They could become smaller, faster, and more human—thinking in pictures, patterns, and compressed ideas rather than brute-force strings of words.
A Glimpse Ahead
This research — to be released in November — outlines the notation, grammar, and early parser design for such an icon-based semantic system. It sketches a path from verbal sprawl to visual precision, showing how we might finally bring the intelligence revolution back into proportion.
Because in the end, perhaps the future of language models isn’t in teaching machines to read more, but in helping them see better.