From Code to Knowledge

Software culture still treats code as the place where capability lives: a new feature means new code, and new code means a development cycle with requirements, use cases, sprints, pipelines and releases. This essay argues that the history of artificial intelligence, read critically, points the other way. Each lasting advance in AI moved knowledge out of code and into a separate, changeable medium. Expert-system shells pulled rules out of programs. The semantic web pulled meaning out of applications. Large language models are use-case-agnostic mathematics, so the code is general and the behaviour comes from data. The idea that connects these advances is knowledge liquidity: how cheaply, quickly and safely what a system knows can be changed without touching its code. If liquidity is the goal, the need for new code should shrink over time. Code should settle into a stable library of universal components and algorithms, and new applications should be defined mostly as knowledge in a well-chosen representation. That in turn sets new requirements for programming languages and data formats. The essay also argues that the field's current formats (XML, JSON, YAML) were adopted by imitation, not by design. That is a small example of the unconscious evolution that a field now building intelligence can no longer afford. A fixed-arity representation derived from natural sentence structure, the seven-slot