A developer tested replacing traditional code with AI by building a contact manager where every HTTP request goes directly to an LLM with database, web response, and memory tools. The AI dynamically creates schemas, generates UIs, and evolves based on feedback without traditional routes or controllers. While functionally successful, it's extremely slow (30-60 seconds per request), expensive ($0.05 per request), and lacks UI consistency. The experiment suggests that when AI inference speeds improve significantly, we might question the need for generating code at all rather than just improving code generation.