Watch: AI Functions Deep Dive:                                  Andrej Karpathy has a... Tagged with ai, llm, programming, softwaredevelopment.

Dev.to is a community platform for developers, offering resources, discussions, and networking opportunities to empower individuals in their coding journey. Developers can learn new programming languages, frameworks, and tools, engage in discussions on best practices, and share their experiences with fellow developers to enhance their skills and grow their professional network.

Andrej Karpathy's Software 1.0/2.0/3.0 framework is extended with a proposed 'Software 3.1' concept: AI Functions. Unlike Software 3.0 where LLMs generate text/code for humans to integrate, AI Functions (via the `@ai_function` decorator from Strands Labs) generate and execute code at runtime inside your Python process, returning native objects like DataFrames or Pydantic models. The key innovation is post-conditions — Python assertions or even other AI functions — that validate output on every call and feed failures back to the LLM for automatic retries. The post covers structured output with Pydantic, local code execution with explicit import restrictions, multi-agent composition, async parallel workflows, and using test suites as post-conditions. Security tradeoffs of running LLM-generated code in-process are acknowledged, with sandboxing recommended for production use.

Software 3.1? - AI Functions

<p>This sounds like a great idea unless somebody uses it to replace things that should actually be done with code.</p>
<p>Just imagine ifs and loops being non deterministic 😂</p>
