News for hackers. Stuff that matters.
Re:Chinese military researchers tap US AI models to train de... by verdverm on Friday July 31, 2026 @03:13PM (#49127397)
1. Everyone has been doing distillation for a long time

2. You ideally want outputs from multiple models, not a single one

3. Distillation (or a model trace) is insufficient on its own (a) you need a sufficiently strong base (b) crafting RL rewards is an art

4. You are conflating DeepSeek with Moonshot (K3)

Re:Chinese military researchers tap US AI models to train de... by relug on Tuesday August 04, 2026 @03:55PM (#49174161)
no the newwest deepseek flash model is better than opus5 (or i cant tell difference anymore)