Collective Intelligence Community Podcast OpenPipe Co-Founder on Reinforcement Learning - David Corbitt | Ep. 22
Jul 29, 2025
David Corbitt, co-founder and Chief Product Officer of OpenPipe, discusses innovative ways to enhance small AI models. He shares insights on fine-tuning and reinforcement learning that empower these models to outperform larger alternatives. The conversation reveals how OpenPipe reduces costs and latency, making AI more accessible. Listeners learn about the significance of realistic training scenarios and the role of human feedback in developing reliable AI. Corbitt also emphasizes the enduring value of specialized models in a landscape dominated by large, closed systems.
AI Snips
Chapters
Transcript
Episode notes
Small Models Beating Big Models
- OpenPipe focuses on making small open models outperform large closed models for specific tasks.
- This improves quality, cost, latency, and independence from central providers.
Fine-Tune When Tasks Diverge
- Fine-tune models when your task differs from big labs' training data to increase reliability.
- Switch to smaller fine-tuned models to reduce inference cost and latency.
GPU Time Drives Costs
- Training and inference costs are dominated by GPU time (H100/H200).
- LoRA-style parameter-efficient tuning greatly reduced training costs.