Using local LLMs vs always going to the hosted models

A discussion on the trade-offs between running open-weight LLMs locally versus using hosted frontier models. The key insight is a hybrid approach: use local models (like Gemma) for trivial or planning tasks, and reserve expensive hosted models (Claude Sonnet/Opus, Gemini) for complex, large-scale code generation. The conversation touches on specialized hardware (Apple Silicon, Windows AI PCs) enabling efficient local inference, and the vision of AI coding agents that intelligently route tasks between local and remote models based on complexity and cost.