Comment by Razengan
1 month ago
Wait.. If "distillation" produces better results than the "original" models, why don't the US model owners do it to themselves?
Have ChatGPT 5.6 talk to itself to produce 5.7 or whatever?
And since they already know which requests came from China or looked like distillation, can't they just replay those same prompts?
They already do and everyone has been using that concept for years now, it's known as RLAIF.