
Fine-tuning on company codebases

Building Windsurf with Varun Mohan
Topic 6 of 16
Published May 7, 2025 · Clip 20:42–23:57 (3:15)
The host asks whether fine-tuning models on company codebases was worth it. Varun describes the elaborate infrastructure they built, including self-hosted fine-tuning that runs backpropagation on preemptible GPU layers between inference requests. Their finding: fine-tuning was only a modest bump compared to great personalization and retrieval. He frames the lesson as climbing the easier hill first and avoiding technical debt from premature complexity.