Segment Anything
ReadA promptable segmentation model trained on a billion masks that transfers without fine-tuning. Shows what a foundation model looks like outside language.
My notes
Picked this as the one vision paper in the LLM stretch. The data engine section is the real lesson — the model helped label the data that trained the next model.
Data Engine
- Assisted manual: annotators click, SAM proposes masks.
- Semi-automatic: SAM fills in confident masks, humans do the rest.
- Fully automatic: grid of point prompts → 1.1B masks on 11M images.
Same flywheel shape I see in LLM products: ship a model, use it to generate/label data, train a better model. Worth designing for from day one.