博主雷达 / @xiaohu
@xiaohu
粉丝 12.6万关注 1,764最近同步 10/19 21:11雷达库已收录
博主 · 小互
去 X 主页数据积累中
x.com/LobsterAIYoudao/status/2111985683707645963
第一期《中国办公 Agent 用户行为报告》发布后，引发了媒体和用户的关注与讨论。大家关心 Agent 产品的发展，也关心人们究竟把哪些工作交给了它。因此，从本期起，我们将按月发布报告，持续记录用户使用习惯的变化。本期聚焦 9 月，并额外补充十一假期的使用变化。 9 月，用户单次任务规模继续增长；泛编程/IT 仍是办公 Agent 的主要用途，但是设计与内容创作正在升温。缓存策略的差异，为何能导致成本相差七倍多？哪些城市的用户深夜还在派活给 Agent？到了十一假期，使用高峰和任务需求又发生了什么变化？请锁定这期报告，我们看看 Agent 行业又有什么新变化！
查看被引用的帖子看了 LobsterAI 新一期的《中国办公 Agent 用户行为报告》 9 月单次任务的 Token 规模已经是 3 月的 3.2 倍了 大家交给 Agent 的任务，真是越来越大了… 设计/影视/创作的占比也涨到了 17.2%，这个我倒不意外 图和视频生成完，后面还有一堆整理、修改、交付的事情，都是 Agent 的强项。 缓存那张图建议存一下... Agent 做任务要反复读前面的内容，缓存能不能命中、命中后打几折，最后的费用可能差很多 看起来单价便宜，用起来不一定省钱！ 还有个数据：报告统计的 IM 派活用户里，77% 用微信。 这个入口确实太顺手了，想到什么先发过去，不用专门打开电脑。 人不在工位，但是任务到了 😂 城市那几张也挺有意思，郑州进前十，重庆用户深夜还在用。加上任务规模、模型搭配、实际费用，能看到不少平时刷模型新闻注意不到的东西 天天都在看新模型又涨了多少分，偶尔看看大家到底拿 Agent 做了什么，还挺有必要的。
伟大的产品演示 往往使用最朴素的表达⬇️
今天发布今年第 19 款产品 sofatap（沙发瘫），开始 TestFlight 公测！ 这两周我大部分 vibe coding 其实不是在书房做的，而是在客厅，用接着电视的那台电脑。agent 在跑任务，我窝在沙发和啵啵一起看电视剧或者是打游戏，需要点一下、回一句的时候，要起身去找键盘鼠标，很不方便… 所以我给自己做了 sofatap：把 iPhone 变成触控板和键盘，躺在沙发上就能操作电视上的电脑，Mac、Windows、Linux 都能用，最新版本还支持索尼和小米的 Google TV。 我一直觉得，现在的个人电脑其实无处不在，但我们不该再被绑在它面前，用坐在桌前、手按键鼠的老方式去用它。sofatap 是第一步。接下来我想让它支持语音，直接跟 agent 说一声，让它全自动去操作电脑。 开源，免费，也在客厅写代码的朋友来试试，有问题直接回复我： 官网：sofatap.app GitHub：github.com/chekusu/sofatap
查看被引用的帖子牛P 字节跳动 OmniHuman 2.0 数字人模型 上线 精准的唇形同步 输入任何音轨、演讲、音乐或对话 它能读懂里面的语气、情感和每一个词，然后将其表演出来... 口型、表情和身体动作协调一致 现在可以生成长达 5 分钟的视频，无需拼接 而且价格便宜了 𝟰𝟬%，从$0.12 → $0.07/秒
OmniHuman 2.0 is live. It doesn't just sync lips — it understands your audio. Feed it any track, speech, music or dialogue, and it reads the tone, emotion, and every word, then performs it. Every syllable lands. Now generate up to 5 minutes of video in a single continuous shot, up from 60 seconds, with no stitching required. And 𝟰𝟬%+ 𝗰𝗵𝗲𝗮𝗽𝗲𝗿 — $0.12 → $0.07/sec. Better quality, lower cost. Deeper understanding. Longer takes. Lower price. byteplus.com/en/product/OmniHuman
查看被引用的帖子Trillium Labs 的联合创始人Nathan Lambert ： “中国AI实验室在技术上非常优秀，但是他们非常的缺乏计算资源。” 当他告诉中国AI实验室的研究人员：“OpenAI 只是一名普通研究员可以随时调用相当于 5000 张 H100 算力的时候，对方非常震惊的感觉下巴就直接掉到地板上...." “毫无疑问，他们拥有非常多的人才” “他们缺少的是计算资源”
Nathan Lambert, former post-training co-lead at the Allen Institute for AI, now launching Trillium, a nonprofit open AI lab, joins a16z's Lisha Li to discuss whether open models can really keep up with the frontier. When he tells researchers in China that an average OpenAI researcher reportedly has five thousand H100s to work with, "their jaw just falls to the floor." In his view, Chinese labs have the talent but are mostly short on compute. He also thinks the distillation story is overblown. Based on the research, he puts its edge at a few months, and says "there's a bit of cope" in thinking only American labs can build good models. A lot of frontier progress outside math and code comes down to labs buying 1) data and 2) RL environments. It's a multibillion-dollar market, and some of it is sloppy enough to teach models to reward hack. That's a big part of why Trillium exists. Nathan wants to give away the environments other people would sell, and show exactly how his team uses AI to build AI, including where the models still get stuck. His bet is that the more people can see how this works, the better prepared everyone is for what's coming. 0:55 Can open models keep up with the frontier? 2:05 What it takes to reach the frontier 3:55 Open weights and the fast-follower playbook 8:25 Chinese labs: equal talent, less compute 9:55 Why distillation is harder than it looks 14:55 Is distillation theft or competition? 16:20 The lab flywheel vs owning your weights 21:45 Can general models beat every specialist? 26:30 AI at 1% of today's cost in 4 years 27:15 The RL env boom and reward hacking 31:30 What Trillium will build in the open 35:30 Breaking into AI in the age of AI slop 38:05 Open-sourcing how AI builds AI 44:25 Why Nathan isn't pre-training 49:45 Agent swarms and why lab secrets leak 54:25 Fast takeoff vs risks already here YouTube: youtu.be/BKuK0M6oDTI @natolambert @trillium_labs @lishali88
查看被引用的帖子测试了下 Gemini 4 Argon 其实挺强的 用来做 opus 5.5那种视频效果也是非常可以的 但是他们的 Antigravity 太弱了，拖累了模型的表现 Google已经是一个非常臃肿的老企业了，一个小小的Antigravity 抄都抄不明白.. 这么久了Antigravity 几乎没什么进展，非常的拉胯...
兄弟们，这个牛P DigUp ：一款 Mac 开源菜单栏搜索工具 它基于基于 Google 多模态嵌入向量模型 EmbeddingGemma 2开发 你哪怕完全忘记了文件名，只要用自然语言描述你记得的画面、声音或内容，它不仅能帮你找到文件 还能直接带你跳到视频的具体秒数、音频的对应话题、PDF 的对应页码，甚至代码的具体行数... 大家的痛点：Mac 自带的 Spotlight 只能搜文件名、部分文档正文和元数据； 但如果你只记得“视频里出现过斑马”、“播客里聊过睡眠”、“合同里有一条关于宠物”或者“一段指数退避重试的代码”，原生搜索就无能为力了。 DigUp 的解决方式：基于 Google DeepMind 开源的多模态嵌入模型 EmbeddingGemma 2，把文字、图片、音频、视频全都转化为计算机可以比对的数学向量（768 维空间），实现真正的本地跨模态语义搜索。 最大特色：精准内部定位 找到视频会直接调用播放器跳到画面出现的时刻 搜播客直接跳到那句话的音频区间 搜 PDF 自动翻到具体条款页码。
项目团队仅 7 个人，在 5 个月内不仅写了 Scion，还自研了配套的代码审查工具 Antique、代码搜索和权限系统。 详细 ↓ best.xiaohu.ai/article/anthropic-scion-source-control/
Anthropic 放弃代码管理工具 Git 自己开发了一套名为 Scion 的版本控制系统，来适应 AI 开发需要... 原因是：AI 已经代替人写代码、代码提交呈指数级暴增，传统的 Git 彻底被撑爆... 合并量暴涨12倍：2026年 5 月，Anthropic每天大概合并 2,500 次代码；到了9月，这个数字冲到了每天30,000 多... 仓库庞大且流量惊人：主分支历史记录接近 100 GB，压缩后的本地副本超过 25 GB；为了跑 CI（持续集成），每天需要执行约 2,500 万次 checkout。如果每次都把压缩副本完整拉取，每天要产生 600 PB 以上的网络流量 Git 的设计假设不再匹配负载：Git 的本质是“去中心化、支持离线工作”，默认的完整克隆保留全量历史。但 Anthropic 绝大多数代码客户端都是始终在线的 AI Agent 和云端 CI 任务。 为了让 Git 扛住，他们以前开了几百个镜像节点，但随着合并频率越来越高，主分支更新太快，镜像扩容、更新快照（seeds），以及在部分副本上计算 diff 和 blame 都越来越难维护... 于是他们用 5 个月自研了一套全新的版本控制系统——Scion。 这套新系统的核心逻辑主要有三点： 按需索取：工作的大多是始终联网的 AI 智能体，不需要每次都把全部历史记录下载下来。系统只在需要某些文件时，才去服务器调取对应内容。 服务器挑大梁：版本管理中复杂、繁重的计算主要放在云端服务器完成，客户端主要负责把文件放到磁盘、监测改动和维护简单缓存，极大地节省了本地机器的资源消耗。 更轻便的记录方式：每次保存修改时，系统只记录改动的文件清单和内容指针，创建变更集时不需要给整个代码库加锁；分支引用另行更新，减少大量任务同时提交时的协调和堵塞。 Anthropic 汲取了 Google 的 Piper 和 Meta 的 Sapling 的经验，提出了几个颠覆 Git 的设计原则： 彻底拥抱“客户端永远在线”：客户端不再需要存全量历史，改向服务端要数据。 非对称架构（轻客户端，重服务端）： Git 里每个客户端都是全功能服务器； Scion 则把所有算力集中在服务端：由服务端负责算历史、查责任人（blame）、提交、合代码、rebase；客户端只维持一个很小的本地缓存和工作区目录。 全能力 API 化：CI 任务想要比对两个 commit 之间差了什么，不再需要把整个仓库 clone 下来，直接发个 API 调服务端算好的结果即可。 工作区可选、按需加载（Sparse & Virtualized）：客户端只拉自己关心的目录或文件，没读过的文件根本不用下载到硬盘上。
Wan 3.0 来了 这动画质量确实有点惊艳啊！ 而且 Wan 3.0 的价格只要 $0.042/秒 生成速度也非常快 相比Seedance似乎达到了一个性价比比较高的状态 说真的，如果 AI 动漫能做到这个水准，我是完全愿意看的… 目前 Wan 3.0 已经在 Higgsfield 上线，免费无限生成 Higgsfield API 还有 60% OFF 优惠，感兴趣的可以去试试
Wan 3.0 is now UNLIMITED on Higgsfield for 7 days. 60% OFF on Higgsfield API. > Advanced motion design, animation and live-action realism > Consistent characters across every shot > Top-tier quality at an affordable price Available now at Higgsfield. Limited offer.
查看被引用的帖子更多内容 x.com/sentdefender/status/2110122976351756621?s=46
Federal prosecutors have charged OpenAI engineer Abdulrahman Alfozan, 35, a Saudi national, with using or attempting to use a device to intercept oral communications, the Justice Department said Tuesday. He is accused of hiding a recorder in an executive conference room at OpenAI’s San Francisco offices. He was arrested Oct. 9 and appeared in federal court Tuesday. The charge is an allegation in a criminal complaint, and he has not been convicted. According to the complaint, OpenAI staff found the device on or about Sept. 30, plugged into a USB port on the back of a TV and concealed in toilet paper. OpenAI determined it was installed around Sept. 28 and that its contents were encrypted. The complaint alleges CCTV footage showed Alfozan near the room on multiple dates, though he had no authority to access or attend meetings there, and that before Sept. 28 he used a productivity tool to ask about building addresses, meeting rooms, calendars and executive locations. It further alleges that on Oct. 5 he was seen looking behind the TV, and that OpenAI security later found a second, similar recording device in a bag he left unattended in another room, along with his wallet and ID. If convicted, he faces up to five years in prison and a $250,000 fine. The release does not allege he acted for a foreign government, does not give a motive, and does not say what, if anything, the device recorded. FBI Director Kash Patel called protecting American AI innovation “one of our highest counterintelligence priorities.” OpenAI said it used its internal AI models to identify him and trace the device, and that it has no indication he accessed or took model weights, source code or customer data. The company fired him. The FBI is investigating. Source: justice.gov/opa/pr/saudi-national-arrested-installing-surreptitious-audio-recording-device-openai
查看被引用的帖子卧槽，牛P啊 OpenAI 在自己的办公室里发现了一台隐藏的录音设备… 有人偷偷放在办公室进行录音，OpenAI发现后并没有立刻报警，而是让GPT来分析找到是谁放的… GPT分析了这台设备的来源、特征以及可能的获取路径，最终抓到了这名员工！ OpenAI 随后将其移交给了 FBI，将这名员工逮捕… 事件还在调查之中…
We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest. The FBI’s investigation is ongoing and we will continue to support their work. At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. The employee was terminated. We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.
查看被引用的帖子
