The Translation Tax

For most of computing history, the human being bore the full cost of translation. The command-line interface accepted a finite set of commands, each with a precise syntax; to communicate with the machine, the user had to compress a rich intention—multidimensional, partially tacit, context-laden—into the narrow vocabulary the channel recognized. This compression was lossy, in Claude Shannon’s precise sense: below the entropy rate of the original intention, so information was necessarily destroyed. The cognitive overhead of performing this compression while simultaneously thinking about the problem to be solved is the translation tax: the bandwidth consumed by encoding rather than by thought. Each translation—vision into specification, specification into architecture, architecture into tickets, tickets into code—imposed its own tax, and the taxes compounded across the pipeline. The natural language interface abolished the tax: when the machine “learned to meet you on yours,” the compression from intention to channel format nearly disappeared, and the cognitive bandwidth that had been consumed by