How is Deepseek actually doing this? Are they just feeding claude's answers into their own models as their own model as training data to improve reasoning? How exactly one train it's model on output of other? what's enginnering inovlved here?

I'd love breakdown of how thsi is executed at scale.

Backstory:

Anthropic recently accused Deepseek,Minimax,Moonshot of using lots of fake accounts to generate exchanges with claude, using the outputs to train the model and called it "distillation attack".