I can't wait for all the x250 sample distills of Mythos and GPT-5.6
A Reddit user is questioning the effectiveness of current model distillation techniques, particularly those using a small number of samples like 250. They recall a positive instance with Qwen R1 8B but haven't found other distilled models to be superior to their base versions since then. The user expresses skepticism about whether new models like Mythos or GPT-5.6 will yield significant improvements through such limited distillation, lamenting the perceived decline in quality for these methods. AI
IMPACT Raises questions about the practical utility and quality improvements offered by current AI model distillation methods.