A Substack essay published March 13, 2026 is circulating among programmers for a single, blunt phrase: AI-generated code is "statistically related to the correct useful one." Written by Robert Herron under his newsletter "he's not joking," the piece — titled simply "computers" — skewers the LLM coding boom with sardonic precision rather than technical argument.
Herron's method signals his thesis before he states it. He refers to Claude, Grok, Gemini, and their peers by scrambled and invented names — "avacado," "agent claw" — to underscore what he sees as their interchangeability in popular discourse. It is a rhetorical move that registers as dismissal without quite committing to one.
The piece's sharpest work is its map of the three-way split dividing the programming community. Herron identifies those who enjoy the productivity boost of AI autocomplete, those who <a href="/news/2026-03-14-beej-hall-ai-code-authorship">celebrate software creation opening up to non-programmers</a>, and a third camp of skeptics who question whether probabilistic output can meet the <a href="/news/2026-03-14-ai-is-great-at-writing-code-terrible-at-making-engineering-decisions">deterministic precision production software demands</a>. That phrase — statistically related to the correct useful one — puts a name on exactly what the skeptics have been circling.
What distinguishes the essay from routine anti-hype commentary is its historical framing. Herron traces the current moment back through decades of accumulated programmer knowledge — including the Stack Overflow and Reddit content scraped to train the models now under debate — positioning LLM coding tools as artifacts built on top of that collective body of work rather than a clean break from it. The argument is less about what the models can do than about where they actually came from.
The piece has landed as a cultural marker more than a technical brief. In early 2026, as agentic coding tools have attracted serious enterprise investment and increasingly bold capability claims, Herron's essay gives practitioner skeptics a quotable frame. Whether "statistically related to the correct useful one" becomes a lasting shorthand for the limits of LLM-generated code in production will depend on whether those skeptics are proven right.