AI labs train models through thousands of experiments with different weights, hyperparameters, and data slices. Current source control systems weren't designed for this workflow.
Create a Git-like system that efficiently stores model checkpoints, tracks lineage of outputs, and allows branching/merging of model versions. Focus on handling large binary files and metadata.
AI labs would pay for better experiment tracking to avoid wasted compute costs and lost research directions.
Start with a CLI that snapshots model states with minimal storage overhead, then add visualization of model lineage.
Risk: Existing tools like DVC might extend to cover this need first.