Cheating Is All You Need

Written at a dead sprint. Steve admits up front that earlier attempts at this post ran 25-plus pages and died because the field kept outrunning them, so this version got banged out in one sitting, and it reads that way in the best sense: money volcanoes, a "meh-nado" of skeptical coworkers, a raw LLM likened to a Harvard CS grad four hours into a fading mushroom trip, calculator-watch arithmetic on what a code-drafting machine is worth. Under the whooping sits a patient, jargon-free anatomy lesson — how a coding assistant actually talks to a model, and why the cheating image in the title is doing real work — aimed at the two-thirds of programmers who, per Sourcegraph's own internal poll, were still at "meh."

Fair warning, which Steve also gives himself in the opening: the back stretch is a product plug and a recruiting ad for Cody, his employer's then-newborn assistant, and should be read as one. Time has been uneven with the rest. The index-card context windows he measures in kilobytes ballooned within two years, and the flattest-wrong line is "Nobody can differentiate on the LLM; they're all about the same" — the model wars settled that decisively the other way. Three years on it reads as a bulletin from the brief window when serious engineers were still asking whether ChatGPT could write working code at all, delivered by someone jumping up and down and pointing at the lava.