JundeWu2025-01-27 21:15:06我之前说OpenAI>Deepseek，所以追赶OpenAI的临门一脚到底差在哪里？先说结论，我认为是超高质量的RLHF (Reinforcement Learning from Human Feedback)，也就是人类反馈 Deepseek-R1这次的训练，仅利用了rule-based outcome reward，也就是数学题答案对错/测代码能不能跑通，训练出来了超强的逻辑，在math/c#OpenAI#DeepSeek#RLHF#Reinforcement Learning from Human Feedback#rule-based outcome reward#逻辑能力#math/c
