Dian

用产品和工程视角聊 AI、芯片和国际科技博弈｜少喊口号，多看事实

[Dian @DominicWoung]

@wk_haunted 这其实是越南国旗

[Dian @DominicWoung]

@woshiok @SuperBrahmin

[Dian @DominicWoung]

@blyen46 Check out Muse, your personal AI agent. Redeem my code in Settings within 48 hours of joining and we'll both get 1 billion Muse tokens.

Code: CLQEZN
https://t.co/LNbx1XcfvK

[Dian @DominicWoung]

@FinanceYF5 Check out Muse, your personal AI agent. Redeem my code in Settings within 48 hours of joining and we'll both get 1 billion Muse tokens.

Code: CLQEZN
https://t.co/LNbx1XcfvK

[Dian @DominicWoung]

快来看看你的个人 AI 智能体 Muse 。在加入后的 48 小时内通过“设置”兑现我的邀请码，我们就能分别获得 10 亿个 Muse 词元。

邀请码：CLQEZN
https://t.co/LNbx1XcfvK

[Dian @DominicWoung]

@BlockInsight214 邀请码：CLQEZN

[Dian @DominicWoung]

@yedabinbinah 邀请码：CLQEZN

[Dian @DominicWoung]

@SwordXyj 邀请码：CLQEZN

[Dian @DominicWoung]

@jinglian @alexandr_wang 邀请码：CLQEZN

[Dian @DominicWoung]

@zhandoudage @TerribleMaps 气笑了

[Dian @DominicWoung]

@LiuweijiaVip @0xlangeai @li9292 @nopinduoduo CLQEZN

[Dian @DominicWoung]

Check out Muse, your personal AI agent. Redeem my code in Settings within 48 hours of joining and we'll both get 1 billion Muse tokens.

Code: CLQEZN
https://t.co/LNbx1XcfvK

[Dian @DominicWoung]

@AdrianMasa0015 @Flintlockxoxo @Chinaembmanila son of monkeys，Put your head in the shit, bastard.

[Dian @DominicWoung]

@TaoHong25462487 @leo114119 factory

[Dian @DominicWoung]

@averse_side @Flintlockxoxo @Chinaembmanila Sure enough, it's a monkey-human hybrid

[Dian @DominicWoung]

@leo114119 185

[Dian @DominicWoung]

@Flintlockxoxo @averse_side @Chinaembmanila You repeat it 10,000 times, and the Filipinos can't get close to Huangyan Island. What are you happy about, idiot?

[ChineseEmbassyManila @Chinaembmanila]

Remarks by Chinese Embassy Spokesperson Ji Lingpeng on AFP Spokesperson’s Malicious Accusations Against China

Q: On September 29, the Armed Forces of the Philippines (AFP) spokesperson for “the West Philippine Sea” Roy Vincent Trinidad said in a virtual interview that China could be behind the circulation of a fake letter allegedly signed by Defense Secretary Gilberto Teodoro Jr. Does the Chinese Embassy have any comment on that?

A: We firmly reject and strongly condemn the groundless smears and accusations against China made by Mr. Roy Vincent Trinidad .

As spokesperson of the AFP, Mr. Roy Vincent Trinidad should have spoken based on facts, evidence, and responsibility. AFP reportedly filed a case with the NBI for investigation on the allegedly faked letter yesterday. How come the AFP spokesperson has possibly drawn some conclusion even before the NBI?

By his logic, can we also claim that every smear and malicious accusation against China in the Philippines points to the AFP because it stands to benefit?

The people are not to be fooled. This shabby tactic of throwing mud at others for one’s own political interests will not work. We urge certain Philippine individuals and forces to stop stirring up anti-China sentiment, sowing hatred, and deliberately fueling division and confrontation. Otherwise, they may ask for humiliations and troubles for their own.

[Zhang Heqing @zhang_heqing]

人民英雄 永垂不朽
“三年以来，在人民解放战争和人民革命中牺牲的人民英雄们永垂不朽！三十年以来，在人民解放战争和人民革命中牺牲的人民英雄们永垂不朽！由此上溯到一千八百四十年，从那时起，为了反对内外敌人，争取民族独立和人民自由幸福，在历次斗争中牺牲的人民英雄们永垂不朽！”

[Max For AI @MaxForAI]

🚨DeepSeek又开源啦！

刚刚，@DeepSeek 正式开源了一整套面向华为昇腾的基础设施组件。

简单说，就是把此前在 NVIDIA 平台上跑通的那套核心基础设施，开始系统性地搬到昇腾上：

TileLang、DeepGEMM、DeepEP、TileKernels、FlashMLA、DeepSelect，基本一一对应。

这里面我最关注的其实是 TileLang。

它本质上是一个用来写高性能 AI 算子的高级语言。以前想把 GPU/NPU 性能榨干，开发者往往需要直接写 CUDA、Ascend C 这种很底层的代码，复杂，而且换一套芯片又得重新适配。

TileLang 相当于往上抽了一层：你用更简单、接近 Python 的方式写 GEMM、Attention、量化这些算子，它再负责把代码编译到 NVIDIA、AMD、昇腾等不同硬件上。

有点像 Triton，但 TileLang 从一开始就在往多硬件后端走。

更重要的是，它不是只追求“好写”，而是希望好写的同时还能把硬件性能吃满。

DeepSeek V4 系列训练里的大量算子，现在已经是用 TileLang 实现的。

这次昇腾版本直接把底层 Ascend C 指令封装起来，并且 DeepSeek 训练里用到的每一个 TileLang 算子，目前在昇腾上都有对应实现。

另外几个开源的仓库也基本把LLM训练最核心的底层能力补齐了：

DeepGEMM 做矩阵计算；
DeepEP 做 MoE 大规模跨卡通信；
TileKernels 放各种常用训练/推理算子；
FlashMLA 做稀疏注意力；
DeepSelect 做高性能数据筛选。

我翻了一下刚开源的仓库，DeepGEMM Ascend 在官方给出的部分 Dense GEMM 测试里，最高已经做到硬件理论上限的 99.8%。

更有意思的是，这次 DeepSeek 还明确提到了和华为一起做 Ascend 950 的 128 卡超节点，对计算和通信做了深度优化。

国产算力过去一个很现实的问题就是：硬件出来了，但真正好用的软件栈、算子库、通信库和开发工具，往往还得重新磨很久。

现在 DeepSeek 相当于把自己已经验证过的一套东西直接开出来。

希望后面这套栈真的能在更多团队、更多国产芯片上复用。