Inference by Turing Post

When Will We Train Once and Learn Forever? Insights from Dev Rishi, CEO and co-founder ⁨@Predibase ​

May 30, 2025
In this engaging discussion, Devvret Rishi, CEO and co-founder of Predibase, dives into the future of AI modeling. He explains the revolutionary concept of continuous learning and reinforcement fine-tuning (RFT), which could surpass traditional methods. Dev shares insights on the challenges of inference in production and the significance of specialized models over generalist ones. He addresses the gaps in open-source model evaluation and offers a glimpse into the smarter, more agentic AI workflows on the horizon.
Ask episode
AI Snips
Chapters
Books
Transcript
Episode notes
INSIGHT

Continuous Model Improvement Today

  • Train once and learn forever is already happening via continuous model improvement pipelines.
  • Static models are giving way to models that improve during production with user feedback and customization.
ADVICE

Leverage Reinforcement Fine-Tuning

  • Use reinforcement fine-tuning (RFT) to customize models with small data via reward functions.
  • Set up continuous feedback loops with RFT to improve models online rather than one-off tuning.
ANECDOTE

Healthcare Uses RFT Feedback Loop

  • Early healthcare customers use LLMs as judges and clinician feedback for continuous tuning.
  • They label few conversations and feed this into RFT loops for ongoing model improvement.
Get the Snipd Podcast app to discover more snips from this episode
Get the app