yv (@yvbbrjdr) header
yv (@yvbbrjdr) profile picture

exists as 451; opinions are my own; Creator of @LANDropApp, @AthenaAGI, LMRouter; MTS @MicrosoftAI pretraining; Ex-@NVIDIA RISC-V sec; EN @realJerryZhou

AS8075
Joined July 2013
Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)Photo by yv (@yvbbrjdr)
淡推一个星期,暂时再见
1
7
867
I wanted to be professional, so I kept quiet all day. Now, the opinions are my own. 1. Jensen’s significance to AI is different. Without his long-term bet on GPU computing and the CUDA ecosystem, AI might have been delayed by many years. Forget LLMs or AGI. Even ResNet might have come much later. I don’t believe any of this was historically inevitable. It was one man’s bet, going all in with NVIDIA. People in the industry understand how much vision and conviction that took. Never forget where we came from. His place in history as a godfather of AI will always be irreplaceable to me. 2. Jensen’s significance to NVIDIA is different, too. He took NVIDIA to the top of the world by market capitalization and kept it there for a long time. CEOs of this caliber are rare, even in the entire history of business. 3. I’ve worked in AI and used NVIDIA GPUs since my very first job. As a family member of an NVIDIA employee who witnessed the company’s extraordinary rise in valuation, I’ve also personally benefited from what he built. I will never kick him when he’s down, and I don’t feel I’m in any position to judge his actions. You can’t benefit from someone and then turn around, forget what they’ve done for you, or even stab them in the back. 4. SGLang Summit was incredibly, incredibly lively, yet also incredibly, incredibly bittersweet. Its opening and that news were only a few hours apart, but those few hours made a huge difference in what we could ultimately present. Some things we can no longer share will have to remain privately treasured. I’m deeply grateful I had the chance to see them. They were a tremendous source of encouragement for us. We will keep moving forward. But today, beneath all the excitement, the underlying feeling is sadness.
8
44
2
484
29,049
Today is the day of reckoning.
11
1,090
I mean, I'm just sad that the world is what it is right now.
When the world suddenly turns against you, the press, the street, the politicians, even your own employees. Hold up, the board; hold up, @JensenHuang!
1
7
878
When the world suddenly turns against you, the press, the street, the politicians, even your own employees. Hold up, the board; hold up, @JensenHuang!
2
1
20
1,909
老黄会涅槃重生
All I’m going to say is that I hope board of $NVDA stands by Jensen, despite media and political pressures. The same way $TSLA and SpaceX stood up for Elon Musk around similar allegations. The AI trade won’t end off this but for Nvidia’s best interest, he should remain CEO. Otherwise I see Nvidia’s valuation crashing due to extreme uncertainty, with a lot of capital rotating to $AMD and other competitors.
3
16
3,736
最近看到好多speculate董事会会把他踢出去的FUD猜想,我只能说你们不懂NVIDIA。
3
13
3,711
从小到大,我们都被规训要遵守规则、遵循普世价值,殊不知这是最大的骗局之一。
1
1
13
1,225
看来早有预感😂
作为前员工,总觉得老黄越来越狂妄了
1
1
1
34
8,088
请还我一个正常的TL谢谢
1
11
2,286
无论你多么聪明,都存在针对你的骗局
所谓骗局,就是给你下套,让你以为你不得不做某件事,否则后果很严重。聪明的人可以看穿一些低端骗局,但是同样有针对聪明人的骗局。 从这个角度看,人类社会就是一个针对所有人的巨大骗局。
9
1,302
yv retweeted
Dario Amodei testified under oath that he didn't know his home address
PACER access obtained ✅
107
190
134
6,028
1,192,874
By napkin math, pretraining this model takes roughly 8.57e23 FLOPs. Given 68 hours of pretraining time and 100% goodput, that works out to 3.5 EFLOPS. A single GB300 GPU gives 2.5 PFLOPS BF16 roofline, so the MFU is (140000/N)% (BF16 denominator), with N being the number of GB300 GPUs. Now is the speculation part. Pretraining usually takes 2^n GPUs, so if they used 8192 GPUs the MFU is 17.1%. Other numbers are either too high or too low.
wow impressive details! curious about the mfu or trained on how many blackwells (so people can do math on their own 👀, but I guess it is sensitive to be released) high efficiency smell anyways. is this the first seriously published native fp4 trained model?
3
33
4,993
北冥有鱼,其名为___。
The becoming begins.
50
17
2
461
22,082
涂改logo也就算了,涂还没涂干净🤪
Congrats to @vllm_project & @sgl_project on launching Day 0 @Kimi_Moonshot support across AMD, NVIDIA, Rubin, and TPU! Interesting to see that SGLang also has Day 0 support for Trainium 2/3!
1
1
18
4,364
Kimi K3.1 recalls instructions it was never given on Terminal Bench. They match the hidden tests. K3.1 while solving: "Instructions likely distilled pitfalls from previous agents." 🤯 When we edit a task to contradict them, K3.1 refuses the new instructions and fails. Other models follow the edit. Is it terminal(ly) benchmaxxed? 🧵
41
90
43
1,061
177,934
如果你觉得设计芯片就和写软件代码一样,你就会像这个人一样大错特错
Chip design is almost the perfect fit for AI: 1. It’s easily verifiable. You can simulate it, check functional correctness, estimate power etc before any fabbing. 2. Enormous search space. AI can find optimizations humans would never think of. 3. Self compounding(recursive). Better chips would make it easier to make even better chips and so on. 4. Huge economic incentive. Tiny improvements are worth billions so every company would push it hard. This will be the first industry in the physical space to see massive AI-powered disruption.
1
7
1,480
大部分人所说的RSI,基本都没有R和S只有I
1
7
1,452
yv retweeted
第一次看到Anthropic的创意的时候,感觉是非常惊艳; 第二次看到MoonShot的模仿时,感觉也可圈可点; 而你这个属于是模型和宣传片都拉完了😄
14
4
2
310
19,305
机器学习本质上就是数字信号处理。训练中的各种功能就像效果器,以特定的方式变换信号,可能会引入噪声,也可能不会。Bug 则只会引入噪声,但这些噪声可能小到不足以产生显著影响,因此很难被排查出来。
Machine Learning is actually Digital Signal Processing. Training features are like effects units that transform the signal in specific ways, and may or may not add noise. Bugs only add noise, but they may be too small to be significant, making them hard to hunt down.
6
772