Notes on research, systems, and things I learn while building.
Some posts are hidden. Click Unlock to reveal them.
Rejecting the two dominant "genius narratives" — head-start and leverage — and unpacking three anti-utilitarian cognitive algorithms Feynman left behind: playfulness, fighting knowledge vanity, and the twelve-problems trick that grows real taste.拒绝"抢跑"与"杠杆"两种天才叙事，拆解费曼留下的三条反功利认知算法——从"游戏心态"到"对抗知识虚荣"，再到"12 个问题"养出 Taste。
Read more ->Starting from predictive coding theory, explaining the neural mechanisms of musical pleasure, and modeling conversation as directional choices in vector space — revealing the shared information-theoretic foundation of both.从预测编码理论出发，解释音乐愉悦感的神经机制，并将对话建模为向量空间中的方向选择——揭示二者共享的信息论基础。
Read more ->A quick-start guide for the new FastVideo training framework (fastvideo/train), covering code checkout, how to run training, example commands, and links to the related PR and RFC.FastVideo 新训练框架（fastvideo/train）的快速上手指南，包括代码拉取、训练运行示例及相关 PR/RFC 链接。
Read more ->FFN + SRAM fused kernel: improves perplexity/downstream accuracy while cutting peak memory by 3–5x and reaching up to 1.08x inference speedup.FFN + SRAM 在线融合 kernel，在提升困惑度/下游指标的同时把峰值显存降 3–5x，并带来最高 1.08x 推理加速。
Read more ->A quick cheat-sheet of four widely used distillation recipes for diffusion/flow-style video generation models, with minimal PyTorch-like pseudocode.核心直觉： 利用教师模型将纯噪声和真实数据一一对应，学生模型通过学习两点之间的直线速度 来拉直生成轨迹。
Read more ->