How Do You Speak Pidgin to a Probability Distribution? (worksonmymachine.ai)

The piece argues that frameworks aren’t obsolete just because LLMs can write code; they’re the shared vocabulary that lets humans and models communicate efficiently. Using a pidgin/creole analogy, the author explains how popular frameworks (Rails, React) become “native” to models that saw them during training, while bespoke or internal frameworks are unseen “pidgins” that require repeated teaching. That mismatch explains why senior devs (native speakers) get reliable outputs from AI assistants, while beginners or teams with custom stacks get brittle, unsafe results. The upshot: frameworks matter for interoperability, maintainability, and for making model-generated code comprehensible months later. To address this, the author introduces VSM (Viable System Model), an agent framework (v0.2.0) for building CLI-based, tool-using agents backed by LLMs that are designed to be teachable. Key features: a CLI generator to scaffold projects, MCP tool support to plug into existing tool servers, and “meta-tools” that let agents introspect code (methods, arguments, purpose) and explain their own operations. Future releases aim to let agents generate and run their own tools inside the framework. Technically, VSM reframes frameworks as active translators — self-documenting, interactive layers that help probability distributions learn domain-specific idioms and reduce friction (and inequality) in AI-assisted development.