- The goal of LLM routing has evolved and become more sophisticated
- Earliest idea was simply: here’s my prompt, what is the cheapest LLM which can handle it?
- Complexity added when you consider the human cost of failures, incorrect answers, suboptimal answers, and hallucinations
- Complexity added when tasks are agentic instead of one-shot
- Complexity added in multi-turn or with other contextual info beyond the current prompt