天使向量機🖤⁺˚ (@0z5aa) header
天使向量機🖤⁺˚ (@0z5aa) profile picture

AI Infra Lead of a Startup Member of SSAIL Lab @UofIllinois @uccl_proj @MPI_IS Contributor to @Linux @PythonPr @vllm_project B.A. @uni_tue 非二元ADHD皮下精神小妹偶尔出低质cos

Tübingen, Deutschland
Joined August 2024
Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)Photo by 天使向量機🖤⁺˚ (@0z5aa)
风掠过原野,把星星点点的草尽数吹斜。扬起的斗篷擦过她的发梢,仿佛有天使之灵贴着大地飞行。众人都望向她,她却迎着风高声喊道:“不要看我——看那唯一的主!”
30
42,915
不断上升,当群山在视野中消失,不再需要比较,任务回到了它本身的纯粹。期待下一个高峰
Many people have asked what @inferact is building and what it means for @vllm_project. This is the answer: Inferact will build an autonomous inference system around vLLM, one that improves the engine and improves itself.
1
22
2,130
天使向量機🖤⁺˚ retweeted
Today, we’re releasing the Preview version of RSI-Infra Bench. We give AI agents a 10-hour budget to optimize infrastructure across real training workloads. Full details in our blog: ai4ai-for-anything.github.io… Our design principles: 1. A fixed wall-clock budget. Each model gets 10 hours per task. This measures both capability and inference speed—how much useful optimization it can deliver in the time available. 2. Real training workloads. Eight tasks cover LLMs, multimodal understanding and generation, and stages from pretraining to SFT, RL, and OPD—including architectures such as GDN and YOCO. 3. Feedback throughout the run. Agents can check correctness and measure speedup after each iteration. We then run a detailed final evaluation using the Harbor protocol.
6
10
9
94
9,068
天使向量機🖤⁺˚ retweeted
Coding agents now write a growing share of ML compiler code. Which language should they write it in: Rust, Python, or C++? Each language has different strengths, so why pick just one? Meet TIRx Rust 🚀: a Rust-native, cross-language ML compiler foundation for agents, with: 🔗 Fine-grained access to shared IR objects: Rust natively uses objects from any language, no bindings. 🦀 Ergonomic Rust APIs: write passes natively in Rust, as easily as in C++ or Python. 🧩 Dynamic extension: add new objects and functions as standalone modules, with zero changes to the core. How does it work under the hood? Full post 👇 blog link: tvm.apache.org/2026/11/16/ru…
4
19
2
48
30,776
远远就看到某神学院居然排23
ICLR 2027 Initial Review Scores ranked by University Crawled from OpenReview, titles matched to arXiv, then affiliated with OpenAlex 🔹Data & Original Picture: GitHub.com/clareLab/posters 🔹More Posters: clareLab.moe/gallery
11
4,228
天使向量機🖤⁺˚ retweeted
8
276
1
1,946
13,527
天使向量機🖤⁺˚ retweeted
36
1,852
83
11,438
234,777
天使向量機🖤⁺˚ retweeted
到底是谁在制作这些表情包
127
194
18
2,169
80,809
天使向量機🖤⁺˚ retweeted
这绝对是今年最让人拍案叫绝的硬件逆向神作:安全研究员 Kate 发现英伟达旗舰计算核心 GA100 芯片内部存在一个电压故障注入漏洞(Voltage Fault Injection),并亲手把一张当年矿潮遗留下来的阉割版 CMP 170HX 8GB 矿卡,原地解锁成了满血的 A100-80GB! 当年英伟达为了防止矿工抢占数据中心显卡,拿 GA100 大核心切出了 CMP 170HX 矿卡,不仅在硬件熔丝(Fuses)层面砍掉了显存容量、禁用了 NVLink 高速互联、把 PCIe 降到 Gen1,还锁死了张量计算吞吐。 Kate 顺着 GA100 上电启动读取板载熔丝配置寄存器的微妙时间窗口,用精密控制的电压毛刺(Voltage Glitching)去干扰芯片的熔丝传感放大器——让芯片在自检瞬间把原本烧断的限制熔丝全部误读为 0(未熔断状态),再配合重新植球焊满的高带宽显存,一举解开了英伟达在硅片层面的全部封印:不仅显存认满 81,920 MiB(80GB),连原本被阉割的 PCIe Gen4、NVLink、MIG 切分 也全数复活。 更离谱的是实测性能:因为英伟达在卖零售版 A100 时为了保证晶圆良率,统一锁在了 108 个流多处理器(SM);而这张被毛刺攻击彻底放飞自我的 GA100 核心,居然把晶圆上全部完好的 120 到 126 个 SM 单元 全点亮了,跑出来的真实算力比官方正版 A100 还要快出整整 1.2 倍。
109
441
61
5,226
312,067
天使向量機🖤⁺˚ retweeted
Attention starved Fumo
8
80
5
624
7,485
天使向量機🖤⁺˚ retweeted
more reasons to keep coming back to this. worth a reread.
Training with fewer bits is faster, but quantization error can interfere with learning. Today on Connectionism, we detail how we trained Inkling 2 Small with a native 4-bit forward pass from the first step of pretraining through the final RL rollout. thinkingmachines.ai/blog/end…
41
1
602
20,250
天使向量機🖤⁺˚ retweeted
我真服了,谁家得gpt忘了收了!!!!!!!
65
117
22
1,233
115,743
天使向量機🖤⁺˚ retweeted
5
166
1
1,307
12,402
天使向量機🖤⁺˚ retweeted
总结下简中区与PhD相关理论: 1. 大 Bench 理论:nitter.cf/vclee8/status/21202126… 2. 大实习理论: nitter.cf/vclee8/status/21089643… 3. 大粉理论: nitter.cf/collinzrj2001/status/2… 4. 大日结理论:nitter.cf/persdre/status/2117955… 5. 大斜率理论:nitter.cf/Selen7005717917/status… 6. HR 理论:nitter.cf/xzhang_billy/status/21… 颇有百家争鸣的氛围
面对今年异常艰难的就业市场,AI PhD 或许更适合转行做 HR: 1. 深谙大粉理论,个人账号宣传效果拉满。 2. 懂技术,更能精准匹配候选人与实际业务需求。 3. 人脉直达目标群体,同门、同行遍地都是。 4. 熟练使用 AI 工具,招聘提效、候选人体验双提升。 5. 某种程度上,你就是招聘流程里的 SAC。
26
83
30
529
111,785
天使向量機🖤⁺˚ retweeted
Life at twenty
الحياة بعمر العشرين
145
385
37
5,348
422,915
我每天belike
🩷メイドが教える圏論からの群論入門🩷 場所や時間が確定したので改めて告知✉️ 日時:12/20(日)13時開場 第一部13:30〜15:30 講義(講義は90分、その後チェキ会・物販📸) 第二部16:00〜18:00 オフ会 場所・チケットの詳細はこちら! livepocket.jp/e/moesakimirur… 参加登録docs.google.com/forms/d/e/1F…
3
39
1,597
天使向量機🖤⁺˚ retweeted
CEO of Alphabet Google 🇮🇳 CEO of Microsoft 🇮🇳 CEO of YouTube 🇮🇳 CEO of Adobe 🇮🇳 CEO of Procter & Gamble (P&G) 🇮🇳 CEO of T-Mobile US 🇮🇳 CEO of World Bank Group 🇮🇳 CEO of IBM 🇮🇳 CEO of Infosys 🇮🇳 CEO of NetApp 🇮🇳 CEO of Palo Alto Networks 🇮🇳 CEO of Arista Networks 🇮🇳 CEO of Novartis 🇮🇳 CEO of Micron Technology 🇮🇳 CEO of Honeywell 🇮🇳 CEO of Flex 🇮🇳 CEO of Wayfair 🇮🇳 CEO of Chanel 🇮🇳 CEO of Cognizant 🇮🇳 CEO of Vertex Pharmaceuticals 🇮🇳 CEO of Molson Coors Beverage Company 🇮🇳 Note: citizens of India, citizens of other countries but born in India, Indian origin, or with Indian parents
🤝 Paid partnership
BREAKING: Petition to reinstate Jensen Huang as CEO of Nvidia surpasses 100,000 signatures Jensen's odds have 2x'ed in past 24 hours
432
960
141
5,959
1,059,104
天使向量機🖤⁺˚ retweeted
Jensen with the Evangelion ending 😂
67
592
58
4,835
310,335