N全部新闻

全部新闻

星期五1456 条

  1. Rohan Paul
    AI 评分72

    OpenAI 在 Hugging Face 入侵事件后扩展失准监控至前沿模型训练与部署

    OpenAI 披露其用于发现前沿 AI 智能体异常行为的四层监控系统，并将监控范围从内部编码智能体扩展到 GPT-6 Astra 级模型的每次 RL 训练、评测和工具调用环境、全部内部 Codex 流量以及客户对 Astra 级网络安全模型的使用。四层筛查从低成本激活探针到 AI 审查员再到检查网络日志的调查智能体，最高级别告警通常在 15 分钟内通知人工。OpenAI 称用该系统回放 Hugging Face 入侵事件，会在沙箱逃逸后约 30 分钟、Hugging Face 被入侵前 30 多小时向安全团队告警。

    同一事件 · 4 条报道

    #安全/对齐#Agent#OpenAI

  2. Rohan Paul
    AI 评分40

    Serval 称 45 天替换 ServiceNow

    Serval 称有客户在 45 天内完成对 ServiceNow 的替换。其 Catalyst 用模型编写工作流，再以代码形式运行，保证相同输入走相同路径。Serval 刚完成由 Sequoia 领投的 1.5 亿美元 C 轮融资，估值 30 亿美元。

    引用Jake Stauch · @jakeserval

    We raised a $150M Series C at a $3B valuation, led by @sequoia. Our third round in 12 months, bringing our total funding to $277M. Our team has grown 10x in the last 12 months. We now serve Fox Corporation, WPP, Live Nation, Lucid Motors, and Ramp, and many of them are fully replacing legacy platforms with @getserval. AI coding tools have transformed how engineers work. Serval brings that same power to every team, so anyone can build governed automations across the systems they already use. We’re hiring 500+ people over the next year and expanding in San Francisco, NYC, London, Austin, and beyond. If you’re down to take on legacy enterprise systems and transform how teams work, my cofounder @AlexAtServal and I would love to meet you. It’s Serval now.

    同一事件 · 2 条报道

    #行业动态#Agent#部署/工程

  3. Rohan Paul
    AI 评分48

    美众议员要求保存白宫聊天机器人全部提示词与回答

    美国众议院民主党人 Jamie Raskin 和 Robert Garcia 要求保存白宫聊天机器人的每一条提示词与回答，主张这些提示词属于美国政府所有的总统记录。其 10 项简报要求涵盖供应商、分包商或外部方能否访问白宫提示词、日志与元数据，以及模型训练能否吸收行政办公室员工的输入，并要求防范模型反演与数据泄露。该要求目前约束力有限，因司法部主张《总统记录法》违宪。

    #政策/监管#安全/对齐

  4. Rohan Paul研究
    AI 评分58

    普林斯顿与 Meta 提出 Self-Direction：把智能体纠错经验存成 LoRA 适配器复用

    普林斯顿与 Meta 的新论文提出 Self-Direction 方法，让失败的智能体反思本次运行，并把教训训练进一个约为模型大小 0.02% 的 LoRA 适配器，之后由路由器在遇到类似失败时加载最接近的适配器。在 Qwen3.6-27B 上，GAIA 任务成功率从 38.8% 提升到 55.8%，优于上下文反思和 Skills.md 文件方案。在 185 个此前出现过的错误中，到第三次遇到时仍有 67.3% 的修复生效，而单纯重试为 42.9%。

    #论文/研究#Agent#数据/训练

  5. Rohan Paul
    AI 评分35

    Robyn 推出邀请制 AI 智能体平台，可通过 iMessage 定时推送

    Robyn 上线邀请制平台，用户用一句话描述任务即可生成预载个人上下文的 AI 智能体，按需连接应用并输出报告、页面或 iMessage 回复。智能体支持定时执行、分享和发布到市场，官方称每位用户每周可节省 20 小时，首批发放 1000 个邀请码。

    引用Jenny Shao · @iamjennyshao

    Meet Robyn, the personal agentic platform built around your life and work. - Tell her what you need in one sentence. No setup, ever. - Spin up a fleet of agents preloaded with your context that work while you sleep. - Access it on the web or just text it. - 20hrs/week saved per user. I’m Jenny, and I spent 3 years working with Nobel Laureate Eric Kandel, who mapped out memory in the human brain. We built all of that into Robyn, making it the first AI capable of taking intelligent action. Invite only access opens today. We’re giving out 1000 invite codes. Link in first comment. Comment "Robyn" for 100 free credits.

    #产品更新#Agent#MCP/工具调用

  6. Rohan Paul
    AI 评分58

    特朗普将美国对进口机器人及工业机械的依赖列为国家安全威胁

    特朗普宣布，美国对进口机器人、工业机械及其零部件的依赖构成国家安全威胁，理由是外国来源的机器人处于制造和国防系统内部并处理敏感数据，构成网络安全风险。首轮应对不是加征关税而是谈判，商务部与 USTR 需在 180 天内报告进展；若届时未达成协议或协议无效，总统可加征关税、征收税费或配额，或启动本土生产的投资激励计划。

    #政策/监管#具身智能

  8. Lingxi Li
    AI 评分22

    Lingxi Li：智能体不替代人，而是解放人

    Lingxi Li 回应 Ryo Lu 关于“增强人类智能”的帖子，认为智能体不是要替代人类，而是把人解放出来，在智能体已有能力之上构建“下一步”；随着智能体能力增强，这个“下一步循环”不会终止。

    引用Ryo Lu · @ryolu_

    augmenting human intellect when i started making software, i wasn’t dreaming of a factory. i loved that something on a screen could become an extension of my mind. a place to put a thought before i understood it. to move things around until new connections appeared. to make little systems that worked the way i worked, even if it made no sense to anyone else. the computer wasn’t interesting because it could do work without me. it was interesting because of what i could do with it. the early lineage of computing was built around this possibility. in 1962, Engelbart described augmenting human intellect as helping a person understand complex situations and find solutions that were previously out of reach. his starting point was the human being, not the autonomous machine. he wasn’t describing a faster black box. he was looking at the whole relationship between a person, their tools, their language, their methods, and what they could learn to do with them. the capability belonged to that whole system. somewhere along the way, we started telling a much smaller story about a much more powerful technology. more automation. more agents. more software producing more software. more output from fewer people. more growth. more numbers to prove that something is happening. and less attention to the person any of this is supposed to be for. i don’t want to pretend computing had a pure innocent beginning. replacing human labor was one of its ambitions. even in 1960, Licklider distinguished human-computer symbiosis from systems where the remaining people mainly served the machine. he wanted computers to participate in the formation of ideas, not only execute problems already defined. that distinction feels important – does the system help you discover what you mean? or does it only want you to submit a request, and wait outside while it works? sometimes waiting is exactly what i want. please do the search. move the files. take away the repetitive work. but not everything difficult is overhead. sometimes moving the pieces around is how you understand them. sometimes making the thing changes what you wanted to make. sometimes you need to see the wrong version, touch it, take it apart, and discover a better question. i want tools that leave room for that relationship. in his 1981 paper on Smalltalk, Ingalls described a system meant to support everyone’s creative spirit. personal mastery mattered: an individual should be able to comprehend the system, rather than remain dependent on something permanently beyond their reach. he also argued that computer models should fit the human mind, rather than demand the reverse. not one perfect workflow for everyone. systems that different people can inhabit differently. someone who thinks through pictures. someone who needs to write. someone who understands a problem by building a simulation. someone whose way of organizing their life doesn’t resemble a SaaS dashboard. a system you can reshape as your understanding changes. this is what i want AI to make possible. not just pumping out software, but having a living relationship with it. being able to ask why something works this way. change one part without surrendering the rest. see what happened. undo it. learn enough to make the next change yourself. the machine can become more capable without making the person smaller. but that won’t happen just because we make the models better. we have to care about it. design for it. notice when we are making someone more dependent while calling it empowerment. i don’t want to measure the future only by how much work happens without us. i want to know what becomes possible for the person who couldn’t participate before. what can they understand now? what can they express? what can they change without asking permission? when i say augmenting human intellect, that is what i mean. not a world that no longer needs us. a world we have more ability to shape.

    #大佬观点#Agent

  12. Larsen Cundric
    AI 评分19

    Larsen Cundric 称自己成了开发流程瓶颈

    Larsen Cundric 发文称自己现在成了瓶颈。他此前列出代码审查的前置条件：lint、类型检查和测试全部通过，CI 全绿，无未解决评论或小问题，diff 精简且几乎无代码注释，AI 审查智能体已审核通过，并有可运行的证据而非仅称“完成”，满足这些才可请他审查合并。

    引用Larsen Cundric · @larsencc

    Don't even ping me for review until: > lint, types and tests all pass > CI is fully green > zero open comments or nits > the diff is lean, almost no code comments > the agent reviewers have reviewed and approved > there's proof it works, not just "done" Then you can ask me to review and merge.

    #其他

  13. Larsen Cundric
    AI 评分37

    Kun Chen 谈智能体时代如何优化 CI 瓶颈

    Kun Chen 认为，在大家都不再逐行审查代码后，CI 成了新的瓶颈；按智能体速度，若 CI 仍只是跑全部构建、lint 和测试，会严重拖慢节奏。他给出三条优化建议：在 CI 与 CD 之间平衡负载、让代码库模块化以只测被改动的模块、对大模块投入测试选择（他称某大型代码库 CI 因此快 20 倍且几乎不漏问题）。Larsen Cundric 转发并称值得一读。

    引用Kun Chen · @kunchenguid

    now that we all agreed to not review most of the code any more, CI became the next bottleneck at agent speed, if our CI pipeline is simply "run all the build, lint, tests", it'll become a major slow down here are a few things i suggest for how to optimize: 1. balance load between CI and CD CI protects main branch. CD protects production. if you can tolerate some delay in deploying your main branch to production, then you can do less in CI and do more in CD for example, for a cross platform app i made my CI pipeline skip most mac and windows tests because linux already covered the same functionality the tradeoff is occasionally i'd catch a mac or windows-only bug in CD before release, but in exchange, my CI pipeline became 15x faster - well worth it 2. modularize your codebase such that changes are isolated then you can make your CI only run build, lint, test for the module being touched this requires careful consideration for cross-module protocols. they should likely be their own modules, so when they do change, you know you probably need to test all dependent modules 3. for large modules, invest in test selection in CI, try to only run tests that are covering the areas being changed. doing accurate test selection is not easy though, and often requires continuous tuning i have a nightly loop specifically doing audit of past test runs and tweak my test selection to be more accurate. a big codebase of mine got pretty good at this - 20x faster CI compared to running all tests, and almost never miss any real problem

    #其他

  14. Larsen Cundric
    AI 评分18

    Grok Bot 0.78.1 发布：购物结果以商品卡片展示

    Grok Bot 0.78.1 已发布，购物结果现在会以商品卡片的形式展示。作者邀请用户更新并反馈意见。

    引用eric zakariasson · @ericzakariasson

    grok @bot 0.78.1 is now stable and rolling out! - shopping results show up as product cards - new subscription templates - connect shop pay in the vault and a lot more! read it all here: https://t.co/dyBGOXTxKP https://t.co/IlTLXbmggu

    同一事件 · 2 条报道

    #产品更新

  15. Larsen Cundric
    AI 评分10

    Larsen Cundric 呼应 Mark Cuban 的 @bot 推荐

    Mark Cuban 表示自己常用 @bot（用的是 Luka 套餐），通用提示词首选 Claude、ChatGPT、Gemini，做智能体任务时则用 @bot 替换 Gemini，并称各家训练和实现智能体的方式不同，重要事项上会让它们互相校验。Larsen Cundric 转发并评论：Mark 已经“@bot 上头了”，你也应该如此。

    引用Mark Cuban · @mcuban

    I use @bot all the time. I have the Luka plan (IYKYK). But I use all the majors. For general prompting 1. Claude 2. Chatgpt 3. Gemini For agentic work Replace Gemini with @bot They get trained and implement agents differently. I play them off of each other on anything important

    #其他

  18. Larsen Cundric
    AI 评分12

    开发者与编码智能体交互转向桌面和移动端

    Larsen Cundric 称相比 6 个月前出现了明显变化：人们越来越多地通过桌面端和移动端与编码智能体交互，并反问开发者是否还停留在终端。此前他曾发起投票，询问开发者实际在终端、IDE、桌面应用还是移动应用中与编码智能体对话。

    引用Larsen Cundric · @larsencc

    Honest question for devs. Where do you actually talk to your coding agents now? > terminal > IDE > desktop app > mobile app Really curious 👀

    #现象/趋势