🏳️⚧️ System Hacker.
Santa Cruz, CA
Joined July 2012
- Tweets17.6K
- Following7.1K
- Followers13.7K
- Likes88K
vickieGPT retweeted
BREAKING: Google just proved AI data centers in space actually work
Project Suncatcher's first satellite, built with @planet and launched on a SpaceX rocket with 4 TPUs on board, verified the physics: Google sent a prompt from Earth, a TPU in orbit ran Gemma inference, and the full answer was beamed back down.
SpaceX is launching its StarMind satellites with NVIDIA GPUs on Starship in late 2027.
Kardashev Type II civilization INCOMING
vickieGPT retweeted
I'm giving a talk at #SC26 about the architecture and design of a 900 PB file system we deployed. It shares a session with one talk on liquid coolant chemistry and another on liquid coolant velocity though. Guess I'll be the odd one out.
sc26.conference-program.com/…
#HPC
vickieGPT retweeted
现在看Intel:
Xeon 6 最多由 3 个 Compute Die + 2 个 I/O Die（共 5 颗芯片）的2.5D EMIB.
XEON7 delay:
为了适应AI/HPC的超高带宽(取消了原计划的 8 通道入门级设计，改为全线聚焦 16 通道内存（支持 MRDIMM 2）的高端 HPC 与 AI 虚拟化市场), 将compute die上的内存控制器移到Fabric die上去: 这属于一个大架构的大改. 然而Fabric die, AMD很多年前就独立出来了.
然后FAB还在亏损, 然后高价位买了设备扩厂. 这是在TSMC/Samsung一再提高封装价格的情况下. 如果某天TSMC/Samsung 降价, INTC 更难受了.
回头看AMD:
1. chiplet是主动选择的创新.
2. FABLESS 是主动选择的商业模式.
没有创新, 没有自己有力的商业模式, 就很难有力量.
vickieGPT retweeted
heard in Silicon Valley:
Security researchers will soon make more than AI researchers.
vickieGPT retweeted
Chip design is all about talent. All the best talent has already left Qualcomm.
Geekerwan findings are striking. Qualcomm latest Snapdragon moved to a 2nm process and pushed CPU clocks to 5GHz, yet apparently failed to deliver consistent improvements in performance per watt. Meanwhile both Apple and MediaTek delivered more convincing efficiency gains. $QCOM
vickieGPT retweeted
事实已经证明了如下几个观点：
1）未来，企业AI服务和Token支出会大幅下降，因为提供不了实际价值。参考Now财报
2）Capex严重影响公司运营，参考MSFT，GOOG，AMZN
3）存储增速见顶，市场对此已经失去兴趣，参考SNDK
我们不难得出如下结论：
1）Anthropic IPO后股价至少腰斩
2）民主党胜选后，投资熊市将开启并贯穿至2027
2027年将是大科技砍Capex，并且美联储继续加息的一年
What a year it would be?
只能定投指数了
所有在这一次牛市中，把运气当实力的人
都会死的很惨
不过我不怕他们看到我的帖子，得到什么启发
因为他们大概已经把我拉黑了
而且就算看到了，也不会信
谢谢大家
😅😅😅
vickieGPT retweeted
DeepSeek and Moonshot distilled Astra. Let them
In @TheEconomist this week, we argue that distillation is good for consumers and good for America.
The biggest losers are China's AI labs. They are sacrificing long-term innovation for short-term gains
vickieGPT retweeted
SemiAnalysis has already leaked something similar
But this graph doesn't include BSPD routing benefits so it undersells density of nodes like 14A
It also does not account for transistor perf, which impacts IP density
This is why Intel CTO @magicsilicon disagrees with this graph
vickieGPT retweeted
💥炸了！米哈游创始人刘伟公开批评梁文峰：看不上只会卷低端，放言自己要搞核聚变造飞船！
米哈游创始人刘伟（大伟哥）在校招演讲中直接把DeepSeek当反面教材：认为对方只盯着中低难度token、拼性价比，最终只会消耗这些“低端”任务。
他抛出米哈游路线：持续往“塔尖”走，真正模型能力应该用来做核聚变、造宇宙飞船、治愈癌症，而不是跟应届生抢饭碗。
网友直接炸了：自家大模型连水花都没见着，就先看不上已经把API价格打到“水电煤”级别、跑通商业闭环的DeepSeek？之前还放话三年砸1000亿冲第一梯队，结果现在用还没成型的项目去踩一个已经落地的对手。
有人调侃：踏实卖大米养活十几亿人的，被跳出来说“我要造飞船”……
大模型的塔尖，是靠海量显卡、数据和无数试错堆出来的，不是一句“往塔尖走”就能直接飞上去的。
#DeepSeek #米哈游 #刘伟 #梁文峰 #AI大模型 #AI普惠
🚨米哈游这波操作太离谱了！自研大模型已备案，中文名就叫“无用”！
备案号：Shanghai-WuYong-202604100142
备案时间：2026年5月15日
主体：上海米哈游秘法科技有限公司
同时在多个类别狂铺商标：
HoYoUseless / HoYoWuyong / 米哈游无用模型
覆盖1类化学原料、2类颜料、3类日化、5类医药、6类金属、10类医疗器械、45类法律服务……
申请日期全是2026-09-20。
一个叫“无用”的模型，却在各种实体与服务类里全线占坑。
结合刘伟此前说的“3年最多1000亿，做不成就算放个大烟花”，这名字选得有点太有幽默感了。
#米哈游 #无用大模型 #AI大模型 #HoYoUseless #崩坏星穹铁道 #原神 #绝区零 #米哈游AI #技术宅拯救世界
vickieGPT retweeted
Little preview: our custom built plugin is +10% faster than TTNN
Prefill and decode load hardware differently, and the split varies with each new agent and model.
Blackhole® runs both phases on the same chip, letting you move capacity as workloads change. Our engineers break down how: bit.ly/4zp75OY
vickieGPT retweeted
I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model.
Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching.
We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards.
As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!
vickieGPT retweeted
The great irony here is that Cerebras has previously criticized Nvidia Blackwell packaging for CTE mismatch between the dies, interposer, and organic substrate while claiming that their monolithic WSE would solve such thermal expansion issues through the absence of an interposer and its "proprietary packaging."
Yet now Cerebras (according to this research) has to run background, non-functional workloads to keep its wafers warm and avoid damaging thermal cycles.
Even with a uniform CTE, a wafer this large is still susceptible to mechanical stress from spatial temperature gradients (imagine one corner of the chip running significantly hotter, and thus expanding relative to the rest).
Even funnier given that their S-1 claims "thermal expansion-tolerant packaging" without mentioning any requirement to artificially keep wafers warm.
sec.gov/Archives/edgar/data/…
vickieGPT retweeted
there will be more demand for computation after continued learning becomes mainstream
Today on the blog, learn about our new instance of a “sleep” phase in continual learning, which enables LLMs to continuously integrate new facts and enhance reasoning while mitigating catastrophic forgetting or static knowledge cutoffs.
More →goo.gle/4iICkyD
ALT Overview of our sleep framework: Input streams from the active phase are stored in short-term memory before entering the sleep cycle, which bifurcates into Consolidation (transferring fragile short-term tokens into expanded long-term parametric memory using low-rank structures and MoE) and Dreaming (autonomous self-improvement via synthetic routing with random experts and RL refinement).
