与Linux&Git创始人Linus Torvalds 合影留念✌️

发一条竟然获得了43.1K粉丝

优秀是可以培养的

？？怎么首页开始给我推一堆外国人学中文的帖子？

Congratulations~

很佩服. @elonmusk ，说开源就开源，而且是持续开源

看来真是本人？

A post takes me back to my early days in the Apache world, when my mentor’s very first piece of advice was, “All important discussions happen on the mailing list.” At the time I privately thought it felt a bit dated — it was only later that I came to appreciate the design wisdom behind it. Over the years I’ve watched quite a few project teams move their discussions into Slack, Discord, or enterprise messaging platforms. It feels lively at first, but once key people leave or team structures shift, large amounts of context simply evaporate. By contrast, the projects that stuck with mailing lists allow you to go back five, even eight years later and still trace exactly how a decision moved from disagreement to rough consensus. That kind of depth in time is something instant messaging can never provide. What moves me isn’t really the mailing list itself, but the point the post brings out — that it was “intentionally shaped to reduce structural barriers to participation.” As someone who often can only open the laptop and reply to threads after getting the kids to sleep, I know all too well that without a thoroughly asynchronous, open mechanism that isn’t tied to any single organisation’s account system, people like us who have to squeeze contributions into the cracks of life would have been excluded from decision-making long ago. This isn’t just a technical choice; it reflects a deep understanding of equitable participation. The changes AI is bringing feel significant too. Using tools to semantically search and summarise years of Apache mailing list archives delivers a level of efficiency I could hardly have imagined in earlier days. But the precondition for all of that is precisely the publicly archived, meticulously preserved record that was never locked into any one company’s platform. When the data is open, the discussion can truly be open. I’m really grateful that in an era where everything prizes “fast,” the ASF continues to hold onto this infrastructure principle — one that may look clumsy on the surface but is, in reality, deeply far-sighted.

666

明晚20:00啊，不得不看了

不得不说追觅俞浩还是挺有意思的一个人 200多个BU都想独立上市，他这么说了，有人真在这么做，比如京东

Cool～

哥们会说中文了

即将在法国巴黎举办的 GOSIM 2025，@kaiyuanshe 有一些免费票，欢迎申请哈～ 限量免费赠票 ● 共20张免费票，先到先得 ● 赠票注册截止时间：5月1日 免费注册链接：eventbrite.com/e/19840138408…

有在法国的可以去看看～

到底是哪些傻X自媒体鼓吹DeepSeek人才流失严重的？

DeepSeek V4 技术报告3.1节白纸黑字写了，V4 从训练到推理的全链路，已经完整适配了华为昇腾 NPU。英伟达的 CUDA 生态不再是唯一选项。报告里甚至详细介绍了他们自研的细粒度专家并行方案“MegaMoE”，能在 NVIDIA GPU 和华为昇腾NPU上都跑出 1.5 到 1.96 倍的推理加速。横跨两大硬件生态做底层优化的模型团队，respect～

这是好事儿啊

DeepSeek V4的架构改动，用“伤筋动骨”来形容都不为过。他们重新设计了长文本的处理方式，搞了一套叫CSA（压缩稀疏注意力） 和HCA（重度压缩注意力） 的混合注意力架构。CSA先把连续4个token的KV缓存压缩成一个，然后通过一个叫“索引器”的组件，让每个query只关注top-k个压缩后的KV块。HCA更狠，压缩率直接拉到128。报告里直接给了一个让人头皮发麻的数据：百万token长文本场景下，V4 Pro的推理计算量只有V3.2的27%，KV缓存占用只有V3.2的10%。