eggwalking

138.4K posts

eggwalking

eggwalking

@eggwalkingdog

Greater Vancouver A Katılım Aralık 2019
3.7K Takip Edilen1.2K Takipçiler
eggwalking retweetledi
Rey|判断位 x 英语自由
Rey|判断位 x 英语自由@ReyJudgementOS·
如何培养青少年的判断力和创造力? 以色列可能是全球最好的表率之一 这个人口仅1000万的小体量国家, 却大量向美国输送10亿美元初创企业创始人 (按每百万人口涌现的独角兽创始人计算, 是美国的162倍) 我曾在以色列特拉维夫大学学习, 亲身感受到他们的教育哲学: 不是教孩子“标准答案”, 而是把整个社会变成判断力和创造力的训练场: 用前沿技术学前沿知识, 10岁就开始做自己想要的东西 以色列孩子从小就被鼓励: 1. 在高度不确定性中快速做决策 2. 敢于挑战权威和质疑主流观点 3. 把失败视为正常实验成本(而非人格失败) 4. 把“为什么”和“如果”当作日常语言 他们没有把教育做成“知识灌输机器”, 而是做成了高强度判断力 + 创造力演练场。 这也是为什么以色列以极小的国家体量, 却能诞生如此多全球级创新公司。 对我们中国企业主/高认知家庭的启发: AI时代,执行力会被大幅压缩, 真正稀缺的是判断力 (在模糊环境中抓指向未来的核心变量) 和创造力(把不可能变成可能/改写约束)。 我们不需要把孩子送到以色列, 但完全可以借鉴这种第二轨道的培养逻辑: 用真实世界高阶输入(而非教材) 大量练习“开放式判断”(而非标准答案) 把失败和不确定性变成日常训练, 结合AI工具做自己喜欢的项目 判断力和创造力并非天生, 而是结构化训练的结果。 想了解如何在国内帮10-16岁孩子 建立这种“以色列式判断力 + 创造力”路径的家长, 评论「判断位」 我下一条分享具体可落地的训练框架 #AI与教育 #以色列教育 #判断力与创造力 下图:美国独角兽创始人原国别 以色列移民是美国本土独角兽创始人的 移民创始人占美国10亿美元估值初创企业的50%
Rey|判断位 x 英语自由 tweet media
Crémieux@cremieuxrecueil

Relative to native-born Americans, several countries' immigrants to America have produced vastly greater rates of unicorn founders. In particular, Israel: with shockingly few people in the U.S., Israel is the #2 origin *in total* for foreign Unicorn founders.

中文
1
12
44
5K
eggwalking retweetledi
Matt Van Swol
Matt Van Swol@mattvanswol·
Her name was Isabella Stroupe. She was 19. She loved books. Her family called her Bella. She was tied to a bed with a tow strap and tortured for months in an east Charlotte NC apartment. Multiple broken bones. St*b wounds. R*ped repeatedly. Her mother said she screamed and screamed when she found out. Thomaz Hamilton, a violent repeat offender is charged with first-degree m*rder and first-degree r*pe. Months. She was alive in there for months. Say her name. Isabella Stroupe. WE DO NOT HAVE TO LIVE LIKE THIS.
English
1.7K
13.9K
34.9K
1.7M
eggwalking retweetledi
Hänz Solo
Hänz Solo@culture_warz·
@Channel6ixNEWS This video is old and i think its from a Vietnam canal cleanup or something. You can tell its not India, there is no human feces in the streets, also its far cleaner, this picture is actually India.
Hänz Solo tweet media
English
2
2
33
516
eggwalking retweetledi
美国焦点
美国焦点@USFocus24·
华盛顿州米尔顿市的民主党人安装了新的测速摄像头。仅6个月时间,这些新的测速摄像头产生的收入就已超过了整个交通预算。近15,000张罚单共带来超过80万美元的收入。超速仅6英里/小时()
中文
8
8
35
7.5K
eggwalking retweetledi
Phairy Megan
Phairy Megan@tadgh_dc·
Boeing hates America
Phairy Megan tweet media
English
344
1.6K
8.6K
84.6K
eggwalking retweetledi
Rohan Paul
Rohan Paul@rohanpaul_ai·
Somebody just ran one trillion param model (Kimi K2.5) on a single RTX 3060 12GB GPU at over 4 tokens/sec and 768GB of second-hand Intel Optane memory. What happened is that a sparse model met an unusual memory tier that could hold its enormous body while the GPU handled the most time-sensitive organs. i.e. the bulk of the sparse expert weights live in a larger, cheaper memory tier and are pulled into the computation as needed. This worked because Kimi K2.5 is a Mixture-of-Experts model, so it has 1T total parameters but activates only 32B per token. The RTX 3060’s 12GB VRAM holds latency-sensitive parts like routing, attention, dense layers, and shared experts. The huge expert weights sit in Optane PMem, configured as RAM, while 192GB DDR4 ECC acts as cache. He is using 6 Optane PMem (DCPMM) sticks. This retired memory format was made to bridge DRAM and SSD performance. The 768GB Optane configuration, using 6x128GB modules, does beat the best NVMe SSDs on latency by a wide margin, but remains 2x to 3x slower than DRAM. llama.cpp handled hybrid GPU/CPU inference, with tensor placement tuned through flags like override-tensor. The result was roughly 4 tokens/sec, which is slow for chat but impressive for a local 1T-parameter model on cheap retired enterprise hardware. The DDR4 acted as cache, the Optane acted as a giant memory pool, and llama.cpp pushed routing and other critical tensors onto the 12GB GPU.
Rohan Paul tweet media
English
22
49
683
49.1K
NiKITa🇺🇦
NiKITa🇺🇦@NiKiTa_32156·
俄羅斯女宣傳員到薩卡特維洛被佔領區拍攝關於俄軍宣傳影片,然後墜入洪水淹死了。
NiKITa🇺🇦 tweet media
中文
17
22
334
39.1K
eggwalking retweetledi
李老师不是你老师
李老师不是你老师@whyyoutouzhele·
5月23日,袁立因疾病在上海瑞金医院住院治疗。 其丈夫梁太平在社交平台发布了她在病床上的照片及长文。 照片显示袁立颈部插管、佩戴鼻氧管,病情较为严重。 她在文中称:最近几年有许多观众问我为什么突然消失了?其实我也不知道,就像小崔老师突然消失了一样,我不懂政治,我只是一个普通的演员,我真的不明白。 他还说:魔鬼的毒钩位置刁钻,并手握十字架表示"在地上的使命尚未结束,还要继续在中国大地上演一台大戏"。 当天袁立还发视频,讲述了病中的感悟。 她说:人的一生如果活到80多岁,一共也就3万多天。 我已经过了一半多一点了,身体有恙啊需要修理,就像车需要保养,进修理厂一样这是一件非常自然的事情。 我希望,等我修养一段时间以后,可以继续的站起来,去服务在中国大地上这一块,需要安慰绝望的人群。 感谢上帝让我有机会用我的母语,来服务中国大地上那千千万万生病的绝望的受了伤害的人。 但是我们的体量太小能力也太有限。 我只是个柔软的人,我不是救世主,我能做的太有限了,但是我想在有生之年,能在我有限的日子里面,能够去尽量安慰那一个个我手能触及的人。
李老师不是你老师 tweet media李老师不是你老师 tweet media
李老师不是你老师 tweet media
中文
242
137
1.5K
296.9K
eggwalking retweetledi
🤍 𝕂𝕚𝕥𝕥𝕖𝕟 🤍
A mother is recording her children create a messy nightmare for store employees... With zero correction, how will these kids adapt to civilized society in 10-15 years? 🤔
English
3.6K
5.4K
28.9K
1M
eggwalking retweetledi
Shazi
Shazi@ShaziGoalie·
This Pearson Airport story somehow keeps getting worse. W5 literally walked in from the street into the baggage pickup area and showed how easy it would allegedly be to grab a “drug bag” and leave. They also observed workers entering without pass scans and random people using exit doors. And this is at Canada’s busiest airport.
Shazi@ShaziGoalie

W5 investigation says organized crime groups allegedly infiltrated Pearson Airport by paying off corrupt workers to move drugs through Canada’s busiest airport. A whistleblower even claimed workers joke you could “walk out with a cruise missile and nobody would stop you.” Meanwhile passengers go through intense screening… while insiders allegedly move freely behind the scenes. 👀

English
97
500
1.9K
204.5K
eggwalking retweetledi
Macro_Lin | 市场观察员
NVIDIA从GB300开始把超级电容集成进供电架构。每个NVL72机柜的超级电容用量在300颗量级,配合5个BBU模组。储能元器件从"可选件"变成了"系统级组件",在供电链里的权重正在接近HBM。 全球能量产AI数据中心用混合超级电容(HSC)的厂商,目前只有日本Musashi一家。它的HSC方案是正极用活性炭(EDLC结构),负极用锂掺杂石墨(锂电池结构),通过专有的预掺锂工艺把两种储能机制融合在一颗电芯里。3.8V工作电压,秒级充放电,百万次循环寿命,正极没有金属氧化物所以不会热失控。Flex(伟创力)作为系统集成方把Musashi的HSC封装成CESS模组,从2025年上半年开始量产交付。整条链路的核心元器件只有一个来源。 供需缺口非常明显。供应链预估GB300单柜用量300颗量级,郭明錤提示过这个数字可能下调,因为机柜内空间有限,NVIDIA仍在考虑不同规格的储能电容方案。但即使打个折,按GB300的机柜出货预期推算,2026年仅这一个平台的需求也在千万颗量级。Musashi 2024年产能20万颗,2025年Q1扩到150万颗,2026年Q3山梨县新工厂投产后到650万颗。产能和需求之间的缺口肉眼可见。 Rubin平台单柜用量预计进一步增加,全球需求还会继续放大。Skeleton莱比锡工厂2025年底刚启用,产能爬坡需要时间。国内江海股份在做LIC和EDLC,但直流内阻指标跟Musashi差了接近一个数量级(6.5mΩ vs 0.7mΩ),进入Hyperscaler供应链还需要通过UL认证和客户验证周期。短期内没有第二个Musashi。 800V HVDC架构把整个数据中心供电从54V机柜内配电升级到800V高压直流。现有CESS模组是48V级别设计,新架构下的适配方案、绝缘设计、安全认证都需要重新验证。这个架构迁移窗口叠加Musashi的产能爬坡节奏,2026-2027年AI服务器供应链里,超级电容这个环节的供给弹性,可能会成为类似2023年CoWoS那样的结构性约束。 $7220.T
Macro_Lin | 市场观察员 tweet media
中文
7
7
65
7.1K
eggwalking retweetledi
C3
C3@C_3C_3·
Want some truth? Not releasing Police Bodycam footage is an admission of guilt. The British Police are not releasing Henry Nowak’s. What’s that tell you…
English
648
5.7K
41.8K
17.9M
eggwalking retweetledi
Kirby
Kirby@Kirbyhome·
@charlenenideya @applepiehouse @chnp101 我男的,但有几个孩子。我们夫妻都喜欢孩子。让你失望了。我最喜欢做的事情就是假期或者周末全家一起开车出去玩。 这个所谓的医生在吓唬未婚女性,并煽动他们不生孩子,而用的每一个论据都是扯淡的,你觉得这没错?!
中文
4
1
6
517
eggwalking retweetledi
Kirby
Kirby@Kirbyhome·
1、胎盘和你的内脏最不一样的地方是设计上就准备脱落的组织。 2、好的处置没有恶漏。有恶漏是因为手术方式落后。 3、喂奶不上班,上班不喂奶。除了喂奶以外全都是睡眠时间。 这个医生有病,该看就去看。我不知道ta宣扬这个的目的是什么,但是ta至少是愚蠢的, 1、简单直接的告诉病人应该怎么做才是高效有用的。 2、人们都不生孩子了,ta会失业。 说的人说,听的人信,都是蠢货。
中文
11
2
39
3K
eggwalking retweetledi
Phoenix Yin
Phoenix Yin@Phoenixyin13·
在目标检测界,一直有两个门派: YOLO 派,传统豪强,走的是天下武功,唯快不破的路线。速度极快,是工业界、无人机、监控摄像头的绝对霸主。 Transformer 派,学院派贵族,脑子聪明精度极高,但由于算力消耗太大,过去像个林黛玉,在需要实时反应的场景里跑不动。 而但是,ICLR2026中RF-DETR 的出现,意味着 Transformer 派终于练成了凌波微步,它不仅保留了高智商,速度还跟上了实时要求。这就算直接不装了,开始抢 YOLO 赖以生存的实时检测饭碗了! 私以为,现在的RF-DETR三大绝活绝对惊艳: 第一个,火眼金睛。以前最顶级的保安看监控,100 个小偷能抓到 50 多个。这个新保安直接把业务能力提到了新高度,100 个里能稳稳抓出 60 多个,而且是在实时监控的超快车速下实现的。 还有,强大的领域适应性。很多 AI 都是偏科生,在学校考满分,去工厂、去农田、去医院就抓瞎。这个模型在 100 个完全不同的真实世界场景里考试都拿了高分。不管是看农田里的害虫,还是看医院的 X 光片,它都能无缝切换。 最重要的,成本。这东西有专门给手机、边缘芯片用的Nano,也有给超级计算机用的2XL。你预算多大,它就能变多大。 以后无人机跟踪、自动驾驶避让、工业流水线质检的大脑换代了。过去因为算力不够、反应太慢而无法使用的更聪明、更精准的 AI 架构,现在终于可以真正飞入寻常百姓家了。
Phoenix Yin tweet mediaPhoenix Yin tweet mediaPhoenix Yin tweet mediaPhoenix Yin tweet media
中文
8
79
487
41.5K
eggwalking retweetledi
Crazy Vibes
Crazy Vibes@CrazyVibes_1·
July 26, 2020. A beach near Collingwood, Ontario. Sixteen-year-old Jamey Ruth Klassen was supposed to be enjoying a quiet family vacation beside the icy blue waters of Georgian Bay. Farther out on the lake, a man named Christopher Robertson had taken his kayak out alone for a peaceful paddle. Then the kayak filled with water and flipped. Suddenly, he was stranded in the freezing bay, clinging desperately to the overturned hull while shouting for help. Jamey didn’t hear him directly. What she heard instead were strangers nearby calling 911, panicking about a kayaker who had disappeared beneath the surface and wasn’t coming back up. Most teenagers would’ve stayed on shore. The water was brutally cold. The distance looked impossible. Lifeguards and paramedics were already being called. Waiting would’ve been understandable. Jamey never waited. She ran toward the water and dove in. Alone, she swam nearly 600 feet through Georgian Bay — the distance of two football fields — pushing herself farther and farther from shore toward the empty kayak floating in the distance. By the time she reached it, Christopher Robertson was gone. Then Jamey looked down. Through the clear Canadian water, she could see him lying motionless twelve feet below on the lake floor. She took one breath. And dove. The cold tightened around her body instantly as she reached the bottom. She grabbed Robertson beneath both arms and forced herself upward, dragging his unconscious body back toward the surface. He wasn’t breathing. His body hung limp in the water. Jamey refused to let go. She turned him onto his back, balanced his head against her shoulder, wrapped one arm across his chest, and began swimming him toward shore using only one arm and her legs. Every second became harder. Her muscles burned violently. Her lungs screamed. She had no formal lifeguard certification because the pandemic had canceled the courses she planned to take that summer. Still, she kept kicking. Then fear hit her. Jamey realized she might drown beside him before reaching shore. Exhausted and losing strength, she used the last thing she still had left: Her voice. She screamed for help. A nearby paddleboarder heard her cries and rushed across the water. Together, they lifted Robertson onto the board while Jamey, shivering and exhausted, swam the remaining distance alone. Onshore, police officers and paramedics immediately began CPR. Moments later, Christopher Robertson started breathing again. He survived. Nearly a year later, Jamey Ruth Klassen received the Carnegie Medal — North America’s highest civilian honor for heroism. Out of millions of people, only eighteen recipients were chosen that year. But Jamey barely spoke about herself afterward. Instead, she used the scholarship money from the award to attend nursing school at McMaster University, quietly continuing the same instinct that had driven her into the freezing water that day: If someone needs help, you go. No hesitation. No spotlight. No waiting for someone braver. Just a sixteen-year-old girl who saw a stranger drowning… and decided his life mattered more than her fear.
Crazy Vibes tweet media
English
228
1.8K
12.5K
411.6K
eggwalking retweetledi
AI Dance
AI Dance@AI_Whisper_X·
苦涩教训第二弹:只要你算力够,最好的数据过滤器就是不过滤。 看完这篇 paper 最大的感受是,rich 老爷子的苦涩教训,这是要到数据侧了? 斯坦福的 Hashimoto 发了一篇《A Bitter Lesson for Data Filtering》,核心结论一句话:只要你算力够,最好的数据过滤器就是不过滤。他们的意思是,业界花了好几年打磨的数据清洗 pipeline,在足够大的 scaling 面前,优势可能就不在了。至少在这篇 paper 的设定里,很多小算力阶段看起来合理的过滤策略,放大以后反而会输给最粗暴的方案:直接用完整池子。 实验做法其实不复杂。把 Common Crawl 和它的各种过滤版本(轻滤、重滤)同比例缩小,然后看随着模型变大、训练步数增加,哪个池子最终能训练出最好的模型。结果:在 670M token 的 CC 子集实验里,未过滤的完整池子胜过了他们测试的所有过滤版本。后面他们把 pool size 放大两个数量级继续看,至少在 CC vs RefinedWeb 这组对比里,这个趋势仍然稳定。不过他们最多只做了 10B token 的实验,仍然是个非常小的尺度 他们还做了更极端的测试:往训练池里注入低质量数据。 ① 先构造一个由 1 万个随机词组成的词表,再从中随机采样拼成文档 ② 把 CC 文档的词序完全打乱 其中,词序打乱文档的注入量最高做到原池的 8 倍。结果是,足够大的模型对这类低质量数据表现出惊人的鲁棒性。 最反直觉的一个结果是:打乱词序的文档,在 330M 模型上不仅没拖累,反而帮模型超过了纯 CC 池的表现(除了 +800% 那组还没训够)。 他们还建了一套 scaling law 来预测:DCLM-Pool 完整的 240T tokens CC 池,最早在 1e30 FLOPs 时就会成为最优选择。而且 1e30 倒也不是那么无法想象。现在前沿模型的预训练算力大约在 5e26 FLOPs 量级;而到 2030 年,已有预测认为单次训练可能到 1e29 FLOPs 换句话说,我们距离“不过滤反而更好”的临界点,可能没有想象中那么远。 这其实呼应了 Sutton 原文里的那个核心观察:试图把你对领域的知识编码进算法,长期看往往会被更简单、随算力优雅扩展的方法击败。 但有个前提必须说清楚:当算力还是瓶颈的时候,过滤仍然重要。而且更重要的是,随着模型增大,对于算力需求是越来越大的,所以我们可能永远到不了算力不是瓶颈的那一天 hhhh 作者也列出了适用边界:他们讨论的是dense 模型的标准预训练,没有数据课程、数据权重和 post-training;MoE、合成数据、训练后期的数据策略,可能都会是另一回事。 而且从另一个角度说,如果 filtering 本身是完美的,我们当然可以 filter #S6" target="_blank" rel="nofollow noopener">arxiv.org/html/2605.1940…
AI Dance tweet mediaAI Dance tweet media
中文
2
9
40
4.8K
eggwalking retweetledi
Right Angle News Network
Right Angle News Network@Rightanglenews·
U.S. Forest Service law enforcement is now asking for the public’s help identifying a group of Indian nationals seen defacing Cathedral Rock in Sedona, Arizona, a sacred Native American site, with furious Americans demanding their immediate deportation.
English
4.1K
23.5K
72.8K
2.7M