DataFramed #241 Getting Generative AI Into Production with Lin Qiao, CEO and Co-Founder of Fireworks AI
Sep 5, 2024
Lin Qiao, CEO and Co-Founder of Fireworks AI, is a former engineering leader at Meta and LinkedIn, specializing in generative AI. She discusses the transition of generative AI from experimentation to mainstream use, emphasizing the importance of fine-tuning models for specific business needs. Lin addresses the challenges of AI integration, including cost efficiency and managing hallucinations. She also explores the rising competitiveness of smaller models and the critical role of quality evaluation datasets in deploying effective AI solutions.
AI Snips
Chapters
Transcript
Episode notes
Generative AI Production Challenges
- Generative AI applications face new production challenges compared to pre-generative AI apps.
- These include maintaining low latency with large models, managing costs, and controlling hallucinations.
Fine-Tuning for Better Alignment
- Fine-tune foundational models to align with specific business problems and vocabularies.
- This improves performance compared to using off-the-shelf models.
Fine-Tuning Process
- Create an evaluation dataset to set target goals.
- Identify failure cases and use them for supervised fine-tuning or preference-based tuning.