API Distillation
A technique that trains a student model on input-output pairs obtained through extensive calls to a teacher model's API. It is imitation at the behavioral level and does not involve direct copying of model weights.
Timeline (last 90 days)
Some labs developing closed-source large models have publicly stated that they have observed third parties using API outputs to distill competing models, which they regard as effectively copying their models' capabilities.
Copying weights usually clearly violates the license agreement, whereas API distillation operates in the interpretive space of the terms of service.
All Facts (2)
Copying weights usually clearly violates the license agreement, whereas API distillation operates in the interpretive space of the terms of service.
50%UnverifiedSome labs developing closed-source large models have publicly stated that they have observed third parties using API outputs to distill competing models, which they regard as effectively copying their models' capabilities.
50%