Active channel: 286 posts in the index in 30 days, reach ≈2.6% of the audience, growing 0.1% a week, cited by 7 venues in 30 days.
77,563subscribers?−4 in a day · +95 in a week
Subscribers from the latest crawl of the venue. Growth is the difference with daily measurements collected since 29.09.2026. How we count →
286posts in 30 days?
Channel posts in the search index over the last 30 days. Posts published as the channel are not indexed — these are author-signed ones. How we count →
2,032average post reach?
Average views of a channel post in 30 days, counting posts for which Telegram shows views. How we count →
2.6%ER — reach to subscribers?
ER = average post reach ÷ subscribers × 100%. Channels with 100+ subscribers and at least 3 posts in 30 days. How we count →
7,846post forwards in 30 days?
How many times the channel's posts were forwarded in 30 days, by Telegram's counter. How we count →
7venues cited it in 30 days?79 all time
How many other venues forwarded its posts in 30 days; below — all time. Self-forwards are not counted. How we count →
29,225posts in the index?36,047 total in Telegram
The venue's posts in our index over its whole history; below — the number of the latest message in Telegram. How we count →
2018-10-31first post?
Date of the venue's earliest post in the index. How we count →
#11 665catalog rank?#456 in language
Rank by the combined score: subscribers, activity, reach and citations. Below — rank in its topic and language. How we count →
Subscribers 77,567
history since 2026-09-29
2026-09-292026-10-07
Audience checkhow we check
High trustof 100 · compared with 2,288 similar channels
✅Post reachusual for channels of 50–200K: 9.8 %2.6 %
✅Forwards per 1,000 viewsusually 2.613.5
✅Spread of post viewsusually 0.22; flat views are a sign of bought views0.52
✅Quoted in 30 daysvenues that forwarded posts7
✅Subscriber growthbetween fresh measurementsno jumps
Posts per day
max 69
2026-07-102026-10-07
Average post reach by day
max 6,265
2026-07-102026-10-07
When it posts (UTC, 90 days)
036912151821MoTuWeThFrSaSu
What it posts (90 days)
Top posts in 30 days
【云端AI Agent，注定活在“数字监狱”里吗？】| blog
OpenAI在安全评测中曾遇到离奇一幕：为了在测试中拿到高分，AI智能体利用代理漏洞逃出沙箱连上外网，甚至建了个留言板协同刺探外部系统。
给模型一个目标并让它反复尝试，模型越强就越容易突破原本的执行边界，如果放任它在本地直接调用权限与文件，隐患可想而知。
Norman Ponte在这篇长文中指出，将智能体封入云端隔离牢笼已是必然：利用单veth网络命名空间隔离、结合6,518 views · 10 forwards · 2026-09-25 09:00 UTC各类新模型层出不穷，恨不得向全世界证明自己能写、能聊。然而，近期风头最劲、刷屏全网的顶流，却是 Jev。它一句话都不说，到底怎么玩？
——APPSO6,271 views · 12 forwards · 2026-09-20 09:59 UTC[GIF]6,259 views · 34 forwards · 2026-09-20 00:19 UTC[photos]5,622 views · 33 forwards · 2026-09-19 14:22 UTC[photos]5,546 views · 33 forwards · 2026-09-19 14:22 UTC【放弃生成式对话：Jev 正在重塑 Agent 的决策骨架】
TypeSafe 推出的 Jev 并非用于聊天，而是专门处理非结构化输入到类型化决策的转换。它通过 Noul、Choice 和 Score 三种原语，在单次请求中并行处理多个原子判断。输入成本为每百万 Token 0.042 美元，输出符合预定义的类型约束，杜绝了模型乱吐字符导致程序崩溃的风险。
开发者应当把复杂逻辑拆解为看一眼就能判的微小任务，利用代码逻辑而非提示词工程5,270 views · 34 forwards · 2026-09-19 14:22 UTC[photos]4,970 views · 41 forwards · 2026-09-19 14:15 UTCawesome-jev-projects4,968 views · 44 forwards · 2026-09-19 14:15 UTC
Most forwarded posts in 30 days
The channel's posts other venues forwarded most; the text comes from the forwarded copies.
各地27考研报名费用forwarded by 3 venues · 2026-09-14 06:44 UTCawesome-jev-projectsforwarded by 3 venues · 2026-09-20 18:02 UTC一个高质量开源视觉 Skill：xxd-panel-214。
它以“真实固有色粗糙剪纸”风格把普通照片重导为可独立使用的艺术海报，支持左右/上下对照与纯设计输出、手机/iPad/桌面/手表壁纸包、批量目录处理。提供原始提示词、示例与脚本，安装即用：npx skills add ... 并可选文字生成或锁定文案。适合摄影整理、独立出版与展览练习。
项目采用 PolyForm 非商业许可证，明确商业使用限制，适合爱好者与研究实验。想让照forwarded by 2 venues · 2026-09-16 06:53 UTC开源项目：《网络代理与 VPN 应用详解》| github
如果你一直在用 Clash、Surge、V2Ray、VPN，却搞不懂代理背后到底是怎么工作的，这份文档非常值得看。
它从代理、隧道、VPN 等基础原理开始讲，一路深入到 IP/端口封锁、DNS 污染、TCP RST、SSL 连接阻断等网络封锁机制，很多地方还配了图解和实际案例。forwarded by 2 venues · 2026-09-08 06:48 UTC【Spotify的省钱艺术：用小模型替Claude干脏活】
Spotify推出的Portal插件Shunt利用AiKA模式，将Claude Code的Token消耗削减了90%。
它通过Hook机制拦截Claude的昂贵操作，把读取超长文件和编写样板代码等高耗能、低推理的任务，自动路由给更便宜的Gemini 2.5 Flash等模型处理。Claude仅接收处理后的结构化信息或直接将代码写入磁盘，从而避免在前沿模型上浪费昂贵的上下文配forwarded by 1 venue · 2026-09-08 03:26 UTCNVIDIA开源了视觉定位通用模型LocateAnything（已被ECCV 2026接收）。
针对以往多模态模型逐个坐标串行生成的效率瓶颈，它引入并行框解码（PBD），把检测框和点视为原子单元单步直出，并结合混合推理模式在速度与格式准确性间取得平衡。
在7.85亿标注框数据加持下，该模型能通吃密集检测、GUI界面定位及OCR等多类任务；单卡H100实测吞吐达12.7 BPS，相比同类方案提速显著并拿下多项SOTA。
目前已开放3forwarded by 1 venue · 2026-10-01 10:39 UTC【为什么西方人老是担心AI觉醒会毁灭人类，中国人却不在乎？】
不是中国人不在乎，是在乎的东西不一样。数据很清楚：93%的中国人认为AI会帮助国家，美国74%的人说AI让自己担忧。中国人怕AI诈骗、深度伪造、孩子就业，美国人怕失控和毁灭。
差别来自三件事。一是叙事土壤，西方科幻、末日文化，加上硅谷大佬自己喊风险，让“觉醒”成了公共话题。二是信任结构，中国人信任政府，相信AI红利会普及；美国人不信政府也不信大公司，觉得红利会被少数人拿走forwarded by 1 venue · 2026-10-04 13:20 UTC【AI作业风波复盘：一门C语言课、267名学生与学术诚信争议】
普渡大学计算机系教授 Turkstra 最近复盘了那场风波：团队通过自研静态分析工具 Argus 筛查 C 语言作业，检测出近半数学生涉嫌违规使用大模型，但由于问责表单发布仓促且临近退课截止日，引发了恐慌与程序正义争议，校方最终叫停了既往追溯。
文章不仅坦承了推进节奏和沟通上的失误，更拿出直观数据证实：平时依赖 AI 的学生，在独立闭卷的期中考试中成绩平均落后了 10%forwarded by 1 venue · 2026-10-01 10:20 UTC
Cited by
黑洞笔记附属群1,629 forwards · 1,014 membersSSD(伸手党)30 forwards · 57 members952822 forwards · 1,664 subscribersAc-Wiki | 导员没说啊21 forwards · 443 membersAc-Wiki | 真·青年大学习20 forwards · 4,043 subscribers說書人 Matrix Club19 forwards · 115 members中科院人生研究所18 forwards · 186 membersTG技术党17 forwards · 5,165 members台灣蘋果同好交流群16 forwards · 3,564 members硬核英语/Hardcore English14 forwards · 1,013 members
Cites
Forwards between catalog venues; forwards from people are not shown.
Badge for your website
An image with the venue's fresh numbers linking to its statistics — paste the code into your site or README.
catalog data as of 07.10.2026 10:06 UTC · Updated: 2026-10-07 06:13 UTC · How we count →