The irony. Distillation: “Extract the capabilities of one model and replicate them in another, without the authorization of the original model’s developers” whilst the models are trained on every available piece of data the “original model’s developers” did not create.