2 905
Subscribers
-124 hours
+47 days
+730 days
Data loading in progress...
Similar Channels
Tags Cloud
Incoming and Outgoing Mentions
---
---
---
---
---
---
Attracting Subscribers
September '26
September '26
+7
in 1 channels
August '26
+133
in 4 channels
Get PRO
July '26
+116
in 7 channels
Get PRO
June '26
+63
in 2 channels
Get PRO
May '26
+52
in 3 channels
Get PRO
April '26
+86
in 10 channels
Get PRO
March '26
+73
in 2 channels
Get PRO
February '26
+85
in 11 channels
Get PRO
January '26
+88
in 7 channels
Get PRO
December '25
+87
in 11 channels
Get PRO
November '25
+119
in 9 channels
Get PRO
October '25
+51
in 2 channels
Get PRO
September '25
+25
in 7 channels
Get PRO
August '25
+63
in 9 channels
Get PRO
July '25
+36
in 9 channels
Get PRO
June '25
+64
in 3 channels
Get PRO
May '25
+59
in 9 channels
Get PRO
April '25
+36
in 8 channels
Get PRO
March '25
+60
in 2 channels
Get PRO
February '25
+50
in 6 channels
Get PRO
January '25
+31
in 8 channels
Get PRO
December '24
+71
in 14 channels
Get PRO
November '24
+59
in 8 channels
Get PRO
October '24
+144
in 16 channels
Get PRO
September '24
+41
in 11 channels
Get PRO
August '24
+126
in 13 channels
Get PRO
July '24
+71
in 6 channels
Get PRO
June '24
+87
in 11 channels
Get PRO
May '24
+103
in 11 channels
Get PRO
April '24
+117
in 11 channels
Get PRO
March '24
+123
in 8 channels
Get PRO
February '24
+54
in 4 channels
Get PRO
January '24
+117
in 23 channels
Get PRO
December '23
+101
in 8 channels
Get PRO
November '23
+113
in 18 channels
Get PRO
October '23
+64
in 10 channels
Get PRO
September '23
+39
in 0 channels
Get PRO
August '23
+73
in 0 channels
Get PRO
July '23
+87
in 0 channels
Get PRO
June '23
+108
in 0 channels
Get PRO
May '23
+44
in 0 channels
Get PRO
April '23
+119
in 0 channels
Get PRO
March '23
+261
in 0 channels
Get PRO
February '23
+155
in 0 channels
Get PRO
January '23
+53
in 0 channels
Get PRO
December '22
+521
in 0 channels
| Date | Subscriber Growth | Mentions | Channels | |
| 06 September | 0 | |||
| 05 September | +1 | |||
| 04 September | +3 | |||
| 03 September | +1 | |||
| 02 September | +1 | |||
| 01 September | +1 |
Channel Posts
Finite time blowup for an averaged three-dimensional Navier-Stokes equation | What's new
https://terrytao.wordpress.com/2014/02/04/finite-time-blowup-for-an-averaged-three-dimensional-navier-stokes-equation/
| 2 | https://fxtwitter.com/tene1la/status/2096198106857832844 | 240 |
| 3 | 葫芦娃其实是最早的 MoE 架构。
七个兄弟是挂在同一个系统里的七个 specialized experts：大娃负责力量和体型扩展，二娃负责视觉与听觉感知，三娃负责高防御，四娃、五娃分别处理火系和水系能力，六娃负责隐身，七娃则直接持有高权限法器。
这套系统的问题从来不是 expert 不够强，而是 router 太差。
每次遇到任务，系统都会根据当前问题激活一个看起来最匹配的 expert，然后让他单独进入副本。结果就是单专家能力很强，但跨专家协同极差：二娃明明已经完成环境感知，却没有把 observation 写入 shared memory；三娃有防御能力，却没有承担 tank；四娃和五娃本可以做组合控制，偏偏始终串行调用。
这就是典型的 MoE 利用率问题：expert specialization 做得很好，orchestration 几乎没有。
而且蛇精很快发现了这个漏洞。她根本不需要正面打赢整个系统，只需要针对当前激活的 expert 做 adversarial routing，把每个葫芦娃分别诱导进自己最不擅长的任务空间，就能逐个击破。
所以真正完成系统升级的，不是某一个葫芦娃继续 scaling，而是最后七个 expert 融合成葫芦小金刚，第一次把 perception、planning、defense、tool use 和 execution 压进了同一个统一策略里。
大家小时候以为《葫芦兄弟》讲七兄弟救爷爷。
长大以后才发现，它讲的是一个 MoE 系统，因为 router 和 agent orchestration 没做好，被 prompt injection 打到反复团灭。 | 363 |
| 4 | 超级新鲜刚刚出炉的冷知识，Nvidia 收购 Hagging Face 掏的钱
$12,930,300,000.00
它的前六位数字 129303 所代表的 Unicode 代码是 U+1F917
也就是 Emoji 🤗。。。 | 316 |
| 5 |  现在的每个开发者： | 498 |
| 6 | 知乎内容已发布到 Telegraph:
https://telegra.ph/Gemini-38-Flash-发布有什么值得关注的-09-03-2 | 493 |
| 7 |  🤣 | 518 |
| 8 | 知乎内容已发布到 Telegraph:
https://telegra.ph/DLSS5意外泄露从目前泄露版表现来看你觉得实际效果如何对业界会有什么影响-09-02 | 505 |
| 9 |  Cat 🐈 @Cutecatpost
🔗 查看原推 | 490 |
| 10 |  外籍个人分红免征个税，终结！
@慕有枝613:
发布视频
播放量:9681 弹幕:16 评论:80
点赞:572 投币:27 收藏:127 转发:94
发布日期:2026-09-01 09:47:44 | 525 |
| 11 | GPU世界
原文：https://www.gpuworld.org/
阅读时间：2 分钟
分数：157 | 497 |
| 12 |  旱 地 拔 葱 | 540 |
| 13 |  この感じ | 22 |
| 14 |  codex突然出现了超快?!
wow~ ⊙o⊙ | 778 |
| 15 |  omp配好role之后太爽了，已忠诚
#ohmypi | 770 |
| 16 | 有一天，你最后一次从 Stack Overflow 复制粘贴，却没有意识到这一点。
X | 1 217 |
| 17 |  GLM-5.3-Flash
该模型拥有 3200 亿个总参数，但仅有 180 亿个活跃参数。在各项基准测试和实际应用场景中，其性能均优于 GLM-5.2。
而且，其成本仅为 GLM-5.2 的十分之一。在编程和智能体相关任务上，其性能则可与 Claude Opus 4.8 相媲美。
https://huggingface.co/zai-org/GLM-5.3-Flash
#GLM #GLM-5.3-Falsh
支持多模态，1M上下文 | 1 022 |
| 18 |  +1No text... | 860 |
| 19 |  +3西藏吉隆口岸发生泥石流 | 800 |
| 20 | 感觉这几年新科幻看的少了，
是不是之后的科幻里宇宙战舰一半的舱位都是 GPU/ASIC 啊👻🙈
大机柜会飞 | 738 |
