Interpolative Software

Open original

The author argues that much of the excitement around AI coding comes from the fact that most software is interpolative: code close enough to what models have already seen that a short prompt gets you the result. Novel software, by contrast, still needs heavy human guidance, and the practical job is to turn novel problems into interpolative ones.

  • Interpolative software is code within a model's training distribution, measured by how few bits of information a prompt needs to specify it (issue trackers, web forms, CI systems, and similar).
  • Opus 5.5 is described as especially strong at generating interpolative software, and the author says agents rebuilt Anki and Linear from a few prompts with no-op drop-in results for the features he cared about.
  • Novel work, like a distributed system with unusual requirements at his job, still needs diligent human shepherding, and the author recommends pushing as much of a problem as possible into the interpolative domain.

The core idea is that software sits on a spectrum from "interpolative" to "novel," and that the split determines how well AI coding agents work. Interpolative software is code that falls squarely inside a model's training distribution: todo apps, clones of existing sites, React front-ends, issue trackers, web forms, CI systems, workflow engines, anything with a canonical open-source reference design. The author's test is practical: how many bits of information do you need to put in a prompt to get what you want? If a few sentences do it, the software is interpolative. Novel software is the reverse, built from components the models have rarely or never seen, or combined in ways agents can't intuit without heavy steering. The author stresses this is a description, not a value judgment, and that much of a modern engineer's job is turning a novel problem into an interpolative one by supplying the right bits so agents can handle the bulk of implementation. The evidence is concrete. The author's team is rebuilding a large distributed system with unusual requirements, and after most of a year of heavy agent and token use, it still needs diligent human shepherding to perform as designed. By contrast, he had agents rebuild large parts of Anki and a hackable Linear-style issue tracker from one-shot prompts, and both were usable drop-in replacements for the features he cared about. He credits Claude Opus 5.5 with unusual strength at generating high-quality interpolative code, and says GPT's current models trail slightly here. He also notes that what counts as interpolative keeps growing as RL environments diversify, and he points to METR's task time horizon metric as a rough proxy, though he treats it as imperfect. The piece also makes a historical argument: even before AI, much software work was reassembling known patterns in slightly new configurations, which he sees as the real reason the open-source movement of 2010 to 2020 took off. Standardizing on shared components was the rational move when the patterns were hard to implement well. The post is short and mostly a reference note for a framing he has been using for months, and it draws a parallel to Gwern's "Adding Bits Beats AI Slop" essay.

Ask about this article