Vanishing Gradients Episode 28: Beyond Supervised Learning: The Rise of In-Context Learning with LLMs
Jun 9, 2024
Alan Nichol, Co-founder and CTO of Rasa, shares insights on using LLMs in chatbots, the evolution of conversational AI, and the challenges of supervised learning. He emphasizes the importance of balancing traditional techniques with new advancements. The podcast also includes a live demo of Rasa's CALM system, showcasing the separation of business logic from language models for reliable conversational flow execution.
AI Snips
Chapters
Transcript
Episode notes
Separate Logic From Language
- Separating business logic from language understanding avoids unmaintainable nests of statements.
- Control flow should be deterministic while language models act as translators into that logic.
ChatGPT Was An Industry Shock
- ChatGPT's launch felt like a "nuclear bomb" in the industry and reversed cancelled projects.
- Large companies reconsidered chatbot investments when public expectations suddenly rose.
Redesign From First Principles
- Re-evaluate primitives when adding LLMs and design new building blocks from first principles.
- Use LLMs where they add qualitative new abilities, not as a drop-in replacement for your whole stack.