Large Language Model (LLM) AI systems actualize the relational meaning at the heart of century-old Philosophical Structuralism.

Structuralism was influential from 1910 through 1970 and focused on meaning and culture as products of larger language or social systems called structures.

Ferdinand de Saussure was a foundational structuralist thinker specializing in language and semiotics with his study of signs, which can be written or spoken words, pictures, symbols.

According to Saussure, signs have two parts. There's the signifier, which is the actual marks on a page or spoken sounds, and there is the signified, or the concept being referenced.

A signifier, like the word "dog" you see on the screen or piece of paper, has no essential connection to the concept of a dog, the signified, in your mind. There's nothing dog-like about "dog."

Instead, signifiers acquire meaning from the structure of language itself. "Dog" refers to, or means, the concept of dog because of its relationship to and difference from every other word in the whole language structure. "Dog" means dog because "dog" is not "cat" or "cow" or "person" but is sort of like "puppy" or "canine" and so on.

LLMs, like ChatGPT, instantiate this century-old structuralist insight.

LLMs break down paragraphs of a prompt into individual tokens, which can be whole words, word fragments or even individual characters.

Each token is represented by a long list of statistical values, called a vector. For example, a token "dog" could have a vector like [0.3, -1.2, …] with thousands of individual values.

Vector values are learned from a model's training data and encode each token's relational context within that data. For example, the tokens that often follow or are followed by the token, or tokens that are often used in similar or different contexts. Vectors reflect a token's similarity to and difference from all other tokens within an LLM's vocabulary.

Structuralism asserts signifiers, like written words, acquire meaning from their relations within the entire structure of a language. LLM tokens, represented by vectors encoding relational patterns, also have a meaning that emerges from the model's internal structure!

Just as Saussure thought no essential connection exists between signifiers (like written words) and signified (concepts), LLM tokens also have no connection to the underlying concepts.

When LLMs analyze or output text, there's no conceptual link to the tokens. The system is stringing tokens together based on the structural conventions and differences encoded within vectors.