"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis The Dawn of Dynamic AI: RFT Comes Online, w/ Predibase CEO Dev Rishi, from Inference by Turing Post
250 snips
Jul 16, 2025 
Dev Rishi, CEO and co-founder of Predibase, dives into the revolutionary shift from static to continuously learning AI systems. He explains how reinforcement learning can adapt via ongoing user feedback, showcasing its potential in healthcare and finance. Rishi also discusses the challenges of implementing these dynamic models, like reward hacking and maintaining quality. The conversation highlights the possibilities of 'practical specialized intelligence' as a more stable alternative to traditional AGI, and how it can reshape various economic niches.
AI Snips
Chapters
Books
Transcript
Episode notes
Build Feedback Data Pipelines
- Collect prompts and responses automatically from production to build feedback datasets.
- Use small amounts of user feedback to fine-tune models continuously with techniques like Direct Preference Optimization (DPO).
Agentic Workflows Are Fragile
- Agentic AI workflows involve multi-turn interactions and tool calls to fulfill tasks.
- Current agent implementations are brittle; low call accuracy compounds to poor user experience.
LLMs Shifted Company Focus
- The rise of large pre-trained LLMs changed Predibase's mission from democratizing all deep learning to focusing on fine-tuning LLMs.
- The user base shifted from NLP experts to general AI engineers entering the field after ChatGPT.
