Andrej Karpathy's Software 3.0 framework has evolved into Software 3.1 with AI Functions from Strands Labs. This new approach moves AI-generated code from development-time text generation to runtime execution, where LLMs produce native Python objects that run inside applications and are verified automatically through post-conditions.

Unlike traditional Software 3.0 tools like Cursor or ChatGPT that generate code strings for human review, AI Functions execute LLM-generated code at call time. The framework uses the `@ai_function` decorator to intercept function calls, generate implementations via prompt, execute them in-process, and return typed Python objects like DataFrames or Pydantic models. Post-conditions act as automated validators that trigger retries with error feedback, eliminating the need for human code review.

This runtime execution model fundamentally changes the AI-software relationship. Instead of generating text that humans must integrate and test separately, AI Functions create live objects your application uses directly. The verification happens continuously with every function call rather than as a one-time pre-deployment gate. This represents a genuine architectural shift in how AI integrates with production software, moving from human-in-the-loop development assistance to autonomous runtime code generation with built-in quality controls.