Opus 5.5 improves interpolative AI coding

The author argues that modern AI coding tools excel at generating interpolative software, which lies squarely within a model's training distribution and can be specified with brief prompts. In contrast, novel software—code requiring combinations of components absent from training data—remains difficult for agents and demands significant human oversight. The author suggests that much of a software engineer's role now involves reframing problems so they can be solved through interpolation. Open article →