Cody Is Cheating

The second installment of Steve's Cheating With Cody series, written six weeks after Cheating Is All You Need and during the most chaotic stretch of the early LLM era — the week Geoffrey Hinton left Google, the leaked "We have no moat, and neither does OpenAI" memo, and LLaMA's weights escaping to Discord. Steve's thesis distills to a formula: AI App = Data Moat + LLM. As the models themselves commoditize into "little pocket Tamagotchi," the defensible part of any AI product is the proprietary data feeding the context window — and Cody's, he argues, is the trio of Sourcegraph backends most people didn't know it had: Embeddings + SCIP Graph + Code Search. The middle of the piece is a point-by-point teardown of Copilot's "moat" (good search, plastic-imitation code graph, a mess of embeddings), and the close makes the deeper claim that you can embed the code graph itself to lift every phase of an LLM workflow.

What makes it land, and what makes it Steve, is the comedy welded to the argument — most memorably a multi-page Socratic dialogue with an exasperated LLM that ends with the conclusion that "embeddings are basically LLM poo," 768 mystical numbers you squeeze out of a 175-billion-parameter beast and just have to believe mean something. It's the missing middle of the trilogy (Cheating Is All You Need → Cody Is Cheating → All You Need Is Cody), and the spot where Steve announces he's coming out of "blogging retirement" because the field is now moving too fast to cover any less often than weekly.