archive – Ethan Ferdosi
summary
Claims that multi-LLM setups outperform single models for evolving requirements and show three implementable patterns: static routing (UI components tied to one model), LLM-assisted classifier (route via a model that classifies prompts), and semantic embedding + vector DB (embed prompts and similarity-search against reference vectors). Notes practical tradeoffs: extra model calls raise cost and latency, caching can cut repeated inference, vector stores add complexity, and hybrid approaches balance accuracy, cost, and domain routing.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:15Why use multiple LMs
- 3:01Multi-model routing use cases
- 5:26Multi-LM routing challenges
- 6:29Static routing strategy
- 8:17Dynamic routing strategies
- 10:00LLM-assisted routing deep dive
- 11:46Semantic routing with embeddings
- 14:13Hybrid routing benefits
- 14:55Implementation considerations
- 16:19Summary
- 16:41Q&A
related talks
30mLLMs: Beyond the Hype - A Practical Journey to Scale
32mLLM Future Is Multi-Model | Erik Davtyan #JSConfAM25
30mWhere there's a will, there's a way: The mad genius of LLMs as a classifiers
24mLLM Inference Optimization From Simple To Sophisticated
30mLLM-Based Recommendation Systems
30mLLM Revolution: Necessity or Excess? A Study of the Most Efficient Solutions
47mGive Your LLMs a Left Brain
46mArchitecting Efficient Agents with Semantic Routing and Mixture-of-Tools