Software is now eating itself (devpg.substack.com)

Generative AI is collapsing the fixed costs of software development and creating what the author calls “software eating itself”: instead of big, long-lived products, we’re entering an era of cheap, purpose-built micro-apps created from natural-language prompts. Tools like Cursor, Claude (Sonnet), and GitHub Copilot can turn descriptions into working code, shrinking what used to take teams into single-person, single-morning efforts. The marginal cost of creating software is approaching zero, so engineering trade-offs flip—maintenance and scalability matter less than speed, utility, and disposability. For the AI/ML community this is both an opportunity and a technical inflection point. Models and toolchains will be judged on prompt fidelity, composability, runtime orchestration, and safe API/tool use rather than only raw accuracy. Expect demand for lightweight deployment, fast inference, retrieval-augmented generation, provenance/versioning, automated testing, and security/privacy controls for ephemeral apps. Product thinking shifts toward ambient, context-aware logic embedded in workflows, while infrastructure economics move from monoliths to many transient services. That changes research and engineering priorities: prompt engineering, model grounding, tool integration, and orchestration frameworks become core primitives for building the next wave of hyper-personalized, disposable software.