Cheating Is All You Need

Published on the Sourcegraph blog in March 2023, a few months after ChatGPT shipped. The framing is a veteran sounding an alarm: an industry-shaking thing is happening in software engineering and most engineers are sleeping through it. The argument has two halves. First, a demolition of the skeptics: engineers who say "you can't trust LLM-generated code" are repeating the exact mistake Steve and his Amazon colleagues made when they waved off an early AWS demo — and besides, the objection is incoherent, because software engineering exists as a discipline precisely because you can never trust any code, yours or anyone else's. He proves the upside by example: ChatGPT one-shots a working Emacs-Lisp function from a sloppy English prompt (the demo on the cover). Second, the technical payload, and the part that aged best. The context window is tiny — an index card next to a textbook — so the only differentiator among the jillion coding assistants is a data moat: a fast, queryable sidecar that can fetch the right code to stuff onto that cheat sheet. AI coding is a search problem.

The pivot point — the essay where Steve, after forty years of programming and a famous skeptical streak, went on the record all-in. "Context is everything, it's a search problem" prefigured the entire RAG and context-engineering wave; "data moat" passed straight into the industry's vocabulary. Everything Steve has written and built since — the junior-developer essays, Gas Town, Vibe Coding — runs downstream of this one.