Effective AI Agents Need Data Flywheels, Not The Next Biggest LLM

Learn how to build effective AI agents through data flywheels rather than relying on the largest available language models in this 17-minute conference talk. Discover NVIDIA's approach to creating self-improving AI systems that continuously learn from real-world data and agent interactions to evaluate, retrain, and optimize smaller, faster models that can match large LLM performance while significantly reducing costs and computational requirements. Explore the implementation of data flywheels using NVIDIA's NeMo microservices to achieve lower total cost of ownership and faster inference speeds. Gain insights into a practical framework for building data flywheel systems for your own AI agent applications, focusing on model distillation techniques and generative AI optimization strategies.