Why AI Needs Its Linnaeus to Become a Revolution

Summary

Samir Varma uses the history of marginal economics and plant taxonomy to argue that a correct idea becomes a revolution only after others can use it cheaply. His four-part Tetrad asks whether an idea has measurable substance, a structure large enough to hold it, a process that non-experts can run, and an economic selection mechanism that gives people reasons to adopt it. Jevons had the insight and much of the mathematical structure by 1862, but marginalism spread only after later economists built diagrams, optimization problems, equilibrium models, and estimation methods that turned it into routine work. Linnaeus similarly defeated the more biologically complete system of John Ray because counting and standardized naming were easier to teach, replicate, and share. Applied to AI, the essay says capability is substantial but measured economic effects remain limited: a Danish study found no change in earnings or hours across 25,000 workers through December 2024, while a Stanford study found a descriptive 19% employment gap for workers aged 22 to 25 in exposed occupations, concentrated in hiring and not established as causal. An MIT NANDA sample reported 95% of organizations receiving no return from generative AI, though the author stresses that its interviews and denominator make the figure preliminary. By contrast, a customer-support deployment involving 5,172 agents increased issues resolved per hour by 15% by codifying expert behavior for less experienced workers. Varma argues that AI is being forced into an automation frame when the strongest measured gains often come from complementing workers. Adoption should spread fastest where outputs are cheap to check, mistakes are contained, and someone can redesign the workflow. Coding is furthest ahead because tests, version control, and local authority reduce those costs. The essay concludes that the decisive question is whether organizations perform the “filing” needed to make AI teachable and repeatable, while interpretability attempts to build instruments for what predictive systems leave unexplained.