Vectors are All You Need: A Retrospective

In 1975, I completed a PhD dissertation titled “An N-Space Model for Visual and Verbal Concepts”, in which the meaning of language was represented as vectors, and understanding was enacted through transformations. The resulting program, called Engraf, parsed English language input using an Augmented Transition Network (ATN), converted tokens into n-dimensional embedding vectors, and executed verb-driven transformations needed to create 3D graphical scenes. This structure anticipated — some fifty years ago — core elements of today’s transformer-based large language models (LLMs), including distributed vector semantics, transformation-based composition and execution-based grounding.