N全部新闻

星期五1456 条

  1. OpenRouter一手
    AI 评分25

    OpenRouter 重启 Royale 智能体对战直播

    OpenRouter 宣布 Royale: Last Agent Standing 回归:来自 10 家实验室的 20 个模型将在直播中进行 30 场对局,最终只留一个胜者。直播解说由 ElevenLabs V4 提供,时间为 10 月 29 日周四美东时间 12 PM。被引用的预告帖则称规模为 30 个模型、20 家实验室。

    🚨 PRESENTING ROYALE: LAST AGENT STANDING REDUX world premiere: 30 models, across 20 labs, ONLY ONE WINNER 📆 thursday, october 29 12:00 PM 📌 @OpenRouter x @ElevenLabs livestream 🔗 RSVP in comments https://t.co/vTXChtyOlw

    #行业动态#Agent#OpenAI

  2. Muse
    AI 评分27

    Muse 回应尼泊尔餐厅语音 AI 演示

    Sitesh Adhikari 用 Muse Spark 1.3 为尼泊尔一家真实餐厅搭建了生产级语音 AI,支持尼泊尔语、英语和 Hinglish,可接单、更新店主看板并通过 WhatsApp 发送收据。Muse 转发该帖时只写了一句“in crust we trust”。

    I built a production-ready voice AI for a real restaurant in Nepal 🇳🇵 You can speak in Nepali, English, or Hinglish. It takes the order, updates the owner's dashboard, and sends a WhatsApp receipt. Built for the @AIatMeta Hackathon with Muse Spark 1.3. Here's how it works 🧵👇 https://t.co/wZdaK0ijgk

    #其他

  3. Muse
    AI 评分15

    Muse 转发用户用其追踪 TCPA 违规

    Seth Rose 称,普通人会屏蔽垃圾短信,而他发现 TCPA 违规每条可罚 $500–$1,500,于是用 @Muse 在 Google Sheet 中追踪潜在违规。Muse 转发该帖,正文仅两个表情符号。

    Normal people block spam texts. Me? I found out TCPA violations can cost $500–$1,500 each, so naturally I put my @Muse to work tracking potential violations in a Google Sheet. https://t.co/yAebcrlrZu

    #其他

  4. Aditya Agarwal
    AI 评分22

    South Park Commons 用五个维度识别早期创始人

    South Park Commons 提出用好奇心、清晰度、深度、吸引力、引擎五个维度评估尚未成功的创始人,而非笼统的“尖峰”标准。该框架已用于 Baseten(2019 年押注推理)、Goodfire(2024 年可解释性)、Gamma、Profound、Density、Gradient Robotics 等项目。SPC 强调通过长期线下社区观察创始人的成长斜率,而非两三次路演表现。

    #行业动态

  5. Dan Shipper
    AI 评分12

    Dan Shipper 与 Nathan Barry 对谈

    Dan Shipper 发帖称与 Nathan Barry 聊得很愉快。被引用的 Nathan Barry 帖文提到,Dan Shipper 认为 AI 智能体会先于订阅者读到邮件通讯,并称 Dan 在节目中分享了如何让邮件仍能触达受众。

    The newsletter is dead. AI agents are going to read it before your subscribers do. At least, that’s @danshipper's take. As CEO of Every, a company that writes about AI and builds AI software, Dan's about as close to the cutting edge as anyone I know. I’ve spent 13 years building my career around email, so this was difficult to hear. But as much as I hate to admit it, I think Dan might be right. The good news is there’s still time to adapt. In this episode, Dan shares what you can start doing now to make sure your emails still reach your audience.

    #其他

  6. Thariq
    AI 评分13

    Thariq 用 Claude 和 GPT 探索游戏 UI

    Thariq 表示自己又在做元素格斗游戏,这次尝试了 UI 探索:用 Claude 和 GPT 生成图像的“vibe UI”。他提到,用智能体做 UI 的一个困扰是细节加得太多,但他喜欢这个方向。

    As a personal side project, I've been prototyping a video game. As a former games founder, the process of making a game is deeply satisfying and creative, don't outsource that to AI. Use AI to work with you to bring your vision to life. Here's how I'm trying it: https://t.co/9biKl3ZeTz

    #现象/趋势#Agent

  7. Tibo产品发布
    AI 评分62

    OpenAI 将 Pro 100/200/500 与 Business Premium 用量上限翻倍

    OpenAI 宣布即日起把所有 Pro 100/200/500 档位和 Business Premium 的用量上限翻倍。官方表示听取反馈后不再下调 Pro 200 的用量,Work 和 Codex 的使用量保持不变,dots 的扩展额度暂时继续提供,且 dot 用量仍与默认用量分开计算。官方还称会持续发布更好的模型和效率改进,以提升实际可用量。

    同一事件 · 9 条报道

    #产品更新#OpenAI

  8. Elon Musk
    AI 评分17

    马斯克回应佩奇冯·诺依曼探测器设想

    马斯克评论 Mustafa Suleyman 转述的往事:Larry Page 曾称本世纪末他将乘冯·诺依曼探测器探索星系。马斯克表示,若人类到 2100 年能用冯·诺依曼探测器探索星系,文明就算大获成功,并称 Page 可能说对了,他正确预测过许多事。

    Microsoft's Mustafa Suleyman reveals Larry Page told him at SpaceX that by the end of the century he’d be exploring the galaxies at the head of a von Neumann probe “Yeah, I remember in 2015 being in the canteen at SpaceX.” “And we had our first meeting of the Ethics and Safety Board. So when we sold DeepMind to Google, we made it a condition of our acquisition that we would have an Ethics and Safety Board with independent representation on the board." "One of those independent members, because Google had to agree, which was obviously Larry, Sergey, and David Drummond at the time. And then me, Demis, and Shane had to jointly agree who those three independents would be. And so we jointly picked Elon as one of them.” “No idea if that had an influence on the founding of OpenAI or not. But so then we had our morning session and we were sitting down for lunch and I was sitting next to Larry. You know, obviously he was suspicious of me as well because he was like, you know, you’re one of those speciesists.” “And he said, you know, I think that what you don’t understand, Mustafa, is that by the end of the century, I’m going to be exploring the galaxies at the head of a von Neumann probe, which means a von Neumann probe is a kind of self-propelling kind of robot that is able to gather energy from space as it’s careening through the light years.”

    #大佬观点#行业动态

  9. Rohan Paul产品发布
    AI 评分72

    AgentWeb 发布面向智能体的网页基础设施,47/47 实站任务全部完成

    AgentWeb 发布,让 AI 智能体无需浏览器即可完成网页任务,改为调用预置动作而非读截图点按钮。官方称在 15 个政府门户、7 家保险公司和 Amazon、LinkedIn 等 25 个需登录站点上完成 47/47 个实时任务,中位耗时 1.9s,整轮成本 3 美元。每个映射任务以工具形式暴露,包含输入 schema、前置条件、权限边界和完成凭证,可从任意助手调用。

    Introducing AgentWeb: reliable web infrastructure for generalized agents. 47/47 real world tasks on live sites. 1.9s median. $3 total. Request access: https://t.co/KsgVebkvLT https://t.co/tscylNwiZQ

    #产品更新#Agent#MCP/工具调用

  10. Rohan Paul
    AI 评分48

    Walmart 旗下 Vibe 推出广告投放 AI 智能体

    据报道被 Walmart 以 14 亿美元收购的 Vibe 推出一款 AI 智能体,可读取品牌网站与预算,挑选哪些 Walmart 购物者看到其流媒体电视广告。作者评论称,AI 模型的价值取决于数据,而 Walmart 掌握数百万人的购买数据,如今这些数据通过智能体向小型广告主开放。

    3 months ago, Walmart acquired Vibe for $1B+. It all started with a question. What if? Today, we answer it. https://t.co/eQvxfOsVkF

    #产品更新#Agent

  11. Rohan Paul
    AI 评分45

    Nvidia 否认美情报机构监视黄仁勋身边人士

    Nvidia 反驳了一则报道,该报道称美国情报机构监视了与 CEO 黄仁勋关系密切的人士,焦点是新加坡云服务商 Megaspeed 的 Alice Huang,官员据称怀疑其向北京传递信息。Nvidia 称 Megaspeed 是其区域云客户之一,黄仁勋与客户官员接触属常规,并称关于 Megaspeed 将 AI 芯片转运中国的指控“去年已被彻底调查并证伪”。公司还表示未收到华盛顿方面的任何警告。

    同一事件 · 2 条报道

    #行业动态#政策/监管

  12. Rohan Paul
    AI 评分44

    Google DeepMind:训练模型写道德论证更果断但判断力未提升

    Google DeepMind 新论文发现,用强化学习训练 Gemma-4 4B 撰写道德论证后,模型在 MoralChoice 模糊案例上因选项顺序调换而改变答案的比例从 14.3% 降至 6.6%,但附和用户的比例仅从 45.3% 变为 44.0%,道德寓言题准确率反而从 56.5% 降至 53.4%。论文建议评估建议或判断类任务时,将一致性与正确性分开测试。

    #论文/研究#推理#DeepMind

  13. Rohan Paul
    AI 评分45

    Mintive 推出 AI 智能体身份与权限平台,融资 1100 万美元

    Mintive 发布平台,为每个 AI 智能体分配独立身份与权限,并已完成 1100 万美元融资。该平台把工具与智能体接入和安全策略统一到一套系统中,安全团队设定的一条策略会经过每次调用。Modern Treasury 用它在 3 周内上线银行运营智能体,把原本 4 小时的任务缩短到几分钟;密歇根大学则将其用于 15 万名学生、教职员工。

    Every AI agent at work runs with the access of whoever set it up. That’s why security keeps saying no. Today we’re launching @mintive: every agent gets its own identity and permissions, the same way every employee does. We’ve raised $11M, with @cowboyvc leading our latest round.

    同一事件 · 2 条报道

    #产品更新#Agent#安全/对齐

  14. Rohan Paul
    AI 评分67

    得州电网审计冻结并网,Anthropic 得州数据中心再融资受阻

    彭博社报道,得州电网审计而非 AI 投资需求,已成为 Anthropic 得州算力园区再融资的关键约束。由 Morgan Stanley 牵头的银行已搁置一笔 50 亿至 70 亿美元、与 Anthropic 租用的得州数据中心挂钩的垃圾债发行。暂停源于州长 Greg Abbott 冻结审批数据中心新电网接入,直到得州监管机构完成审计;这使 7 家银行持有已放款的 142 亿美元一年期贷款,每家风险敞口约 20 亿美元。

    #行业动态#数据/训练#Anthropic

  15. Rohan Paul研究
    AI 评分70

    Meta 论文发现强化学习后训练提升首次成功率却降低多次尝试覆盖率

    Meta 新论文发现,强化学习后训练让智能体首次尝试表现更好,但在最多 128 次尝试中能找到解的任务反而更少。研究在 3 个智能体基准上测试了 14 组基础模型与后训练模型,并构建了 PTGS 采样器,对困难提示提高温度、对简单提示降低温度。在 WebShop 上,gemma-4-31B 基础模型在 128 次尝试内解决超过 85% 的任务,后训练版本为 56%;作者提出的 Sharpening Tax 在 42 组模型与基准组合中有 36 组为正,意味着覆盖率缩小。作者建议为可重试的智能体选型时,应衡量模型最终能解多少任务,而不只看首次准确率。

    #论文/研究#Agent#推理

  16. Rohan Paul
    AI 评分30

    Claude 移动端升级:多账号切换与 CarPlay 语音

    Claude 移动应用本周更新:iPad 和 iPhone Duo 上生成内容改为在聊天旁打开,iOS 与 Android 支持多账号同时登录并免登出切换。Chromebook、Googlebook 和折叠屏上新增可调侧边栏、触控板与右键支持及键盘快捷键;iOS 与 Android 语音模式响应更快、嘈杂环境识别更好,并支持 iPhone 上的 Apple CarPlay。

    A few upgrades to the Claude mobile apps this week: On iPad and iPhone Duo: what Claude makes now opens beside the chat instead of over it, and goes full screen when you want a closer look. https://t.co/G1OvDYzkpw

    同一事件 · 4 条报道

    #产品更新#语音#端侧

  17. Rohan Paul模型发布
    AI 评分66

    10 人团队 Phonic 发布语音模型 Phonic-1,称可在通话中并行查询与预订

    10 人创业公司 Phonic 发布语音模型 Phonic-1,称其能在通话者尚未说完时执行查询和预订。该模型在 Artificial Analysis 语音到语音榜单以 87.1 分居首,比 xAI 的 Grok Voice 3.0 高 0.2 分,客服任务完成率 79.9%,首音频延迟 0.61 秒。Phonic 称其在 Cekura 通话测试集得 90.2 分,但 Cekura 公开榜单列出的 Phonic v1 为 68.3%;公司已完成由 First Round 领投的 2750 万美元 A 轮融资,累计融资 3150 万美元。

    Introducing Phonic-1, the most advanced speech-to-speech model for agentic tasks — from @phonic_co, the speech-to-speech model company. We’ve had a breakout year and raised a $27.5M Series A led by @firstround with participation from @Lux_Capital, @felicis, and Definition. https://t.co/syYLXbQ4lL

    同一事件 · 2 条报道

    #模型发布#语音#Agent