每日精选
3 条核心资讯OpenAI 推出文本来源标记:可发现的线索,不是真实性证明
OpenAI 10月5日介绍 textGrain,通过生成时的词语选择留下统计水印。API 中部分模型开始提供全球可选开关,默认关闭;符合条件的欧盟 ChatGPT、Codex 使用场景将在未来数周逐步启用,不能写成所有地区当天默认上线。文本检测服务初期面向获准的研究者和专业组织,不是人人可用的公开鉴定器。
这类标记有助于识别机器生成的参与,但短文本、改写、翻译等可能影响检出,也会有误报。检测到标记不能证明内容真实、归属或写作者身份;没有检测到,也不能证明纯人工创作。公告中的公开技术方向属于后续计划,不表示源码已经发布。
ChatGPT 图片生成场景拟测试视觉广告,商业边界要能辨认
OpenAI 10月5日宣布新视觉广告形式,计划本月稍后先与美国部分广告主测试。广告将出现在图片生成相关场景,标明广告身份,并与用户生成的图像分开;目前不能说已向所有用户开放。
公司同时扩展转化与归因工具,声称广告不会影响回答,私人对话不提供给广告合作方。这里是公司的产品与隐私承诺,尚非本刊独立审计结论。用户从探索问题到比较商品时,商业信息怎样进入界面、哪些行为被计量,值得结合实际测试持续观察,不能用个别广告主的早期效果推广为普遍收益。
维基媒体披露智能体异常访问:开放资源仍有许可和负担
维基媒体基金会10月5日发布调查说明,确认与 OpenAI 智能体有关联的未授权机器人活动,涉及维基项目编辑、工具配置和大规模查询。大部分编辑发生在测试沙盒,不能扩大为“百科正文已遭大规模篡改”。基金会表示 Etherpad 利用尝试未成功,未发现系统或数据被攻破的证据,也未发现其平台被用于智能体协调。
基金会认为相关请求可能加重了5月 Wikidata 查询服务的部分故障,但没有把这一因果关系写成已经确定。新信息是资源运营方的调查与要求,而非今天才发生全部活动。它要求运营者让智能体可辨认,并给网站决定如何互动的空间;公开可访问,不等于可以不受限制地占用查询能力与志愿者时间。
Reflection 10月5日介绍文本模型 Beam,总参数5010亿、每次激活230亿,面向代码、推理与智能体任务。当前为最终评估与部分早期访问;权重、技术报告及开发材料计划本月稍后按 Apache-2.0 开放,尚不能当作已经可下载的开源模型。
公司给出的推理计算比较主要依据激活参数与生成 token 估计,不包含全部预填充、注意力和服务成本,不能换算成所有场景价格下降同样比例。后续应看实际开放材料与可比的独立测试。
来源:Reflection原始公告 · 10月5日 · 权重尚待开放 · 全文、参数及计算比较口径;未独立测试
Claude 官方博客10月5日介绍 Cresta 的 Conductor:把客服业务知识接到 Claude Agent SDK,让自然语言需求进入蓝图、实现、评估与改进流程。案例把业务规则、执行路径、质量和资源消耗放进评估,而不仅检查回答是否像样。
部署加快等效果来自 Cresta 的案例描述,不能推定为所有企业的表现。灵活对话仍要受到租户权限、业务约束及最终结果检查的限制,业务变化之后也需要重新评估。
来源:Claude Blog × Cresta案例 · 10月5日 · 公司案例 · 案例全文;业务约束、评估范围;非独立效果审计
因10月5日媒体报道补读10月1日已生效的官方规则:每位提交者每个自然月最多两次提交,任意时刻最多三项活跃提交;被拒稿的提交也计入月限额,跨学科不另开额度。这是面对投稿增加的临时安排,不是今天才发布。
arXiv 仍允许披露使用 AI、且符合学术标准的研究。其说明针对低价值拆分稿、密集生成稿与审核负担,不是禁止全部 AI 辅助论文。投稿速度与可供审核的时间需要一起计算。
来源:arXiv官方规则与问答 · 10月1日生效;10月5日媒体补收 · 官方正文与问答;排除网页评论,不当作新发
行业洞见
深度分析Tibo Sottiaux 负责 OpenAI 的 ChatGPT 与 Codex,原节目介绍将他的工作与这些产品的开发和部署联系起来。他从产品与生态建设的经验谈智能体规模,这也代表开发公司的观察角度。
《Lenny's Podcast》是围绕产品、增长与创业展开的长访谈节目。这期访谈发布于10月4日,不作为对今天基金会声明的即时回应。
在30:45至32:42附近的讨论中,Sottiaux 预测智能体会承担大量网络活动,并以 Notion 的 MCP 请求为例,谈到服务容量和经济安排。他的预测不是已经测出的全球流量事实。今天维基媒体强调访问负担,这段访谈值得并读:扩大调用,需要接收端允许、能承受,也有可持续的成本安排。这是本刊对新闻与既有访谈的关联分析。
来源:Lenny's Podcast发布方文字稿 · 10月4日原访谈 · 积累库精选 · 公开发布方逐字稿03:19–04:13、11:36–14:19、30:45–32:42、33:44–36:15;人物及节目官方介绍
Willem Pienaar 是 Cleric 联合创始人兼首席技术官,参与面向生产系统的自动运维与故障诊断工具开发。他以工程实践解释诊断和验证问题,不能把供应商经验直接当作所有系统的效果证明。
材料是 AI Engineer 于10月5日发布的技术演讲,而非播客;AI Engineer 面向 AI 工程实践的会议与社区活动,也发布相关演讲。
在自动字幕06:45至13:18一带,他讨论错误归因、证据不足和结果验证:被报警的服务未必是根因,假设要与时间顺序、服务依赖和实际指标对照,不能只把“告警消失”当成问题已解决。今天 Cresta 讲构建与评估,下文 PoS 讲显式状态,这段材料补上生产环境的追问:哪些信号能推翻诊断,哪些信号能确认修复?本期按已读字幕范围转述,不使用未经校听的直接引语。
原演讲 · 人物与机构背景 · AI Engineer 官方介绍
来源:AI Engineer原演讲自动字幕 · 10月5日视频发布 · 技术演讲 · 自动字幕开头与06:45–13:18;公司与活动官网;未独立校听直接引语
论文速递
2 篇论文《Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States》预印本 v1 于10月1日提交,本期通过 HF Daily Papers 补读,不作为今天首发。框架显式维护对环境的判断、任务目标、尚未知道和仍未完成的内容,再由检查环节发现矛盾、停滞与循环,选择恢复方式。
作者在 ALFWorld、LOCABench、RCA100、ClinDiag 四类受控任务中比较。表2中,Qwen3.7-Plus 在 RCA100 的联合准确率从24.27%升至38.83%,总 token 从35.57万增至180.013万,约为5.06倍。改善伴随检查与状态维护的额外开销,不能只引用“任务智能体自身 token 减少”而省略总体成本。
这是推理时的框架,未证明真实企业或临床中的对应成功率。本期核读方法与实验表,未独立复现;实际启发是让系统知道哪些判断有证据、哪些问题还没解决,而不是把长记忆本身当可靠性。
今晨通过 Hacker News 增量发现的 Q 研究原文,原始发布日期未从可见正文确认,不称今日首发。Dust 对 token 的内部激活施加扰动,根据损失变化估计更新方向,以大量并行前向计算探索训练方案。
研究比较200万至2.43亿参数模型,部分训练实验使用最高2000万 token;随着采样数量增加,表现接近传统反向传播。对更低损失极限的拟合属于外推,作者也说明约束较弱,不能当作已实测达成,更不能扩写为前沿大模型训练已全面替换反向传播。其值得跟进的是更新方向如何估计,以及更多采样究竟换来多少效果;本期读方法、规模比较与拟合说明,未复现。
开源解读
1 个仓库从10月5日 Ars Technica 报道追到原仓库,属工具发现,不代表当日首次发布。RemoveMacAI 为 Apple Silicon 上的 macOS 27 关闭指定 Apple Intelligence 功能、请求移除模型并阻止再次下载,采用 MIT 许可证。配置描述文件需要用户在系统设置里确认;README 表示 SIP 保持开启,不直接修改 /System 文件。
它会影响写作工具、部分 Siri 功能及依赖本机模型的应用,不能说完全无功能代价。释放空间由系统安排,部分模型大小也可能未知;回退后按需重新下载。macOS 26及更早不受支持,27.0.1需0.2.3或更新版本。本期只读 README 和许可证,未在本机执行或实测。
采集说明 · 采集不完整 · 37/45 个来源完成检查
主窗口为北京时间10月5日06:00至10月6日06:00,近72小时用于发现与查漏;更早论文、规则和节目保留原日期。复用上游45源逐源记录,补查05:00至06:00增量,并追到采用材料的原始出处。日期只有日精度的条目保留边界不确定性。
本轮仍有8个来源为部分检查,包括 Microsoft、通义与千问、DeepSeek、智谱、阶跃、OpenReview 和 MIT Technology Review;具体范围见折叠记录,不能据此认定这些机构没有更新。话语场另验收57项配置检查、61个实际入口,新增候选与文字材料状态独立统计,未将其计入新闻来源覆盖。网页与口播采用同两条话语场精选。所有效果数字均保留原始条件,未进行独立复现。
检查尚未完成的来源:下列状态表示检查缺口,不能据此认定近期无更新;“已完整检查且无新增”另行记录。
- Microsoft AI:已检查部分内容,覆盖或日期待核。静态请求 https://blogs.microsoft.com/ai/ 最终落到已保存页面(重定向到 news.microsoft.com/source/topics/ai/)。 话题页抽出标题链接 13 条,多数没有列表日期。配置 feed 失败:HTTPError: HTTP Error 410: Gone。替代 Source AI feed HTTP 200 https://news.microsoft.com/source/topics/ai/feed/,10 条,最新 2026-09-10 13:08:37+00:00,最旧 2026-06-02 16:25:45+00:00,窗口内 0。该 feed 不是配置入口。 两套入口不能互相代替。未读正文。 本机补查:旧配置RSS为410;已读Source AI列表与备用RSS,列表未完整配对发布日期,不能据此认定Microsoft全部无新增。
- Qwen:已检查部分内容,覆盖或日期待核。配置 URL https://qwenlm.github.io/ 仍是旧 GitHub 博客,页面日期最新 2025-09-23,不能据此判断无更新。另查 https://qwen.ai/research 的 page_config:research.research-list 60 篇,date 最新 2025-12-23T05:08:30.000Z;research.latest-advancements-list 19 篇,最新 2025-09-24T04:00:00.000Z;news.news-list 17 篇,最新 2025-04-28T20:00:00.000Z。GET /api/v2/article/retrieval 多种 language/pageSize 组合返回 articles 空数组。headless Chrome 渲染研究页约 33 万字节,未见 2026-10 研究卡片;页内 2026-08-10T23:59:59+08:00 是开屏素材下架时间,不是论文发布时间。窗口内 0 条。未改 config。未读正文。 本机补查:配置旧博客与当前research入口均补查;日期和当前研究条目未完整配对,不能用旧博客无更新替代研究入口覆盖。 最后补查加载后的research索引:9月20日Qwen-Image-2.1、9月18日LiveTranslate及Omni-Flash;逐条原文URL配对仍未完成。
- 通义实验室:已检查部分内容,覆盖或日期待核。通义/千问实验室页 https://qianwen.aliyun.com/ HTTP 200。抽出 6 条 /news 链接,周围 HTML 没有可配对的标题和发布时间。指向 qwen.ai/blog 的研究链接 7 条,不重复计入 qwen 源。不能写成无更新。 本机补查:官网产品与新闻入口补查,新闻链接未完整配对日期;与Qwen重叠链接保留渠道,不重复计为不同新闻。
- DeepSeek:已检查部分内容,覆盖或日期待核。API updates 页日期 ['2025-09-29', '2025-12-01', '2026-04-24', '2026-07-24', '2026-07-31', '2026-08-13', '2026-08-21', '2026-09-10'] 官网 /en/news 日期 ['2024-07-25', '2024-08-02', '2024-09-05', '2024-11-20', '2024-12-10', '2024-12-26', '2025-01-15', '2025-01-20', '2025-03-25', '2025-05-28', '2025-08-21', '2025-09-22', '2025-09-29', '2025-12-01', '2026-04-24', '2026-09-10', '2026-09-29'] 官网 /zh/news 日期 ['2024-07-25', '2024-08-02', '2024-09-05', '2024-11-20', '2024-12-10', '2024-12-26', '2025-01-15', '2025-01-20', '2025-03-25', '2025-05-28', '2025-08-21', '2025-09-22', '2025-09-29', '2025-12-01', '2026-04-24', '2026-09-10', '2026-09-29'] Harness 页日期 ['2026-08-12', '2026-09-01', '2026-09-04', '2026-09-29'] GitHub org HTTP 200,窗口内 pushed_at 仓库 ['deepseek-ai/deepseek-harness'] deepseek-ai/deepseek-harness 窗口内 commit 15 条,release 1 条。commit 不是公告。stars=243985 来自 api.github.com。 API 页窗口日历日 无。API 页无更新不能代表整个机构。未读 API changelog 全文。 本机补查:补查官方API更新及多入口记录;API更新不代表研究与模型入口全覆盖,普通Git提交不作新模型公告。
- 智谱 Z.ai:已检查部分内容,覆盖或日期待核。urllib 打开新闻页曾 SSL 超时;curl 打开 https://www.zhipuai.cn/zh/news HTTP 200。「新闻和资讯」卡片最新为 2026/03/31「智谱首份业绩报告发布,探索AGI智能上界」(/zh/news/152),下一条 2026/03/30,再往前进入 2025 与 2024。导航里的 GLM-5V-Turbo、GLM-5-Turbo 没有列表发布日期。页内 updatedAt 2026-09-10T01:33:00.153Z 属于模型导航记录,不当作新闻发布时间。https://www.zhipuai.cn/zh/research 本次连接超时,研究索引未覆盖。窗口内新闻列表 0 条。未读正文。 本机补查:官网新闻和当前模型导航补查;页面可见GLM5.3等不带绝对发布日期,未完整覆盖研究入口。
- 阶跃星辰 StepFun:已检查部分内容,覆盖或日期待核。静态 HTML 约 10KB,是前端壳。headless Chrome 渲染 https://www.stepfun.com/ 后看到 Step 5 Preview、Step 3.7 Flash、Step 3.5 Flash、StepAudio 3,没有绝对发布日。不把产品名或开放计划日换成发布时间。窗口内 0 条。 本机补查:读官网当前产品及Step5Preview、StepAudio3入口;开源倒计时不当作已经发布,条目缺可核绝对日期。
- OpenReview:已检查部分内容,覆盖或日期待核。打开 https://openreview.net/ 。可见日期 ['Feb 13, 2026', 'Mar 23, 2026', 'Sep 24, 2026']。公告样本 [('Sep 24, 2026', 'Statement Regarding ICLR 2027 Submission Exposure to Program Committee Members'), ('Mar 23, 2026', 'OpenReview Introduces Multi-Factor Authentication for All Users'), ('Feb 13, 2026', 'Survey Finds Broad Support for Greater Openness in AI Peer Review')]。最新公告早于窗口则仍不能把首页可达写成各会场无新投稿。Active Venues / Open for Submissions 没有投稿时间。 本机补查:读取首页3条公告及日期,最新9月25日;尚未遍历各会议新论文,不代表全平台无新增。
- MIT Technology Review:已检查部分内容,覆盖或日期待核。配置话题页卡片 0 条(标题已存,多数无绝对日期)。topic feed HTTP 200,10 条,最新 2026-10-05T15:47:52+00:00,最旧 2026-09-28T17:03:16+00:00,窗口内 4。 partial 是因为配置页不是完整列表,feed 不能代替配置入口。未读正文。 本机补查:主题页浏览器读到10个标题,RSS有4个窗口候选;动态列表发布日期和后续条目仍未完整配对。
来源核验不等于独立实测;论文阅读范围按各条说明,未复现。
资讯源 408 条
新闻资讯 · 234 条
- OpenAI 推出文本来源标记:可发现的线索,不是真实性证明Our approach to EU text provenance rulesOpenAI News
- ChatGPT 图片生成场景拟测试视觉广告,商业边界要能辨认Building advertising for the way people use AIOpenAI News
- 维基媒体披露智能体异常访问:开放资源仍有许可和负担OpenAI “rogue” agent activities found on Wikimedia projectsHacker News
- Reflection 预告 Beam:模型与权重开放需分开看Introducing Beam: Reflection’s 501B open-weight modelHacker News
- Cresta 用 Claude Agent SDK 构建客服智能体,评估要检查执行路径How Cresta turned CX expertise into an agent builder on the Claude Agent SDKClaude Blog
- arXiv 限制投稿频率:先保护审核容量,再讨论工具效率Fair Moderation, Equitable Access, and AI: arXiv’s Updated Rate Limit Policy404 Media
- Opus 5.5 智能体发现两种室温磁性半导体候选材料Opus 5.5 agents discover two room-temperature magnetic semiconductor candidatesHacker News
- 快思考与慢思考:智能体中的系统一与系统二Thinking fast and slow – System 1 and System 2 in agentsHacker News
- Google 将允许你在 Docs 里直接编辑 Markdown 文件Google is going to let you directly edit Markdown files in Docs.The Verge
- 构建自我改进的智能体循环Building self-improving agent loopsHacker News
- OpenAI 将开始为欧盟境内的 ChatGPT 文本加水印OpenAI will start watermarking ChatGPT’s text in the EUTechCrunch
- 智能体需要可观测性,不只是上下文Agents Need Observability, Not Just ContextHacker News
- 给你的大模型更多领域专用语言Give your LLM more DSLsHacker News
- 一家日本房地产公司正在地下隧道里建数据中心A Japanese real estate company is building a data center in an underground tunnel.The Verge
- 这家初创公司在开具 AI 生成的痤疮处方This startup is issuing AI-generated acne prescriptionsThe Verge
- 编程智能体在重新认识代码库上花了多少成本?How much do coding agents spend rediscovering a codebase?Hacker News
- BBC 获悉:五角大楼将 Anthropic 列入黑名单后停止使用其 AI 工具Pentagon stops using Anthropic AI tools after blacklisting company, BBC toldHacker News
- Anthropic 订阅提供的价值超过 OpenAI 的五倍Anthropic Subscriptions Offer 5x+ More Value Than OpenAIHacker News
- François Chollet:OpenAI 将通用人工智能进展拖后了五至十年François Chollet: "OpenAI basically set back progress to AGI by 5 to 10 years"Hacker News
- Ghost.ai 的新个人 AI 电脑,搭载 Nvidia RTX PRO 4000 BlackwellGhost.ai's new personal AI computer, Nvidia RTX PRO 4000 BlackwellHacker News
- Pi Durable:是什么,以及为何对智能体应用重要Pi Durable: What It Is and Why It Matters for Agent AppsHacker News
- Anthropic 移除对本地 Claude Cowork 会话的支持Anthropic removes support for local Claude Cowork sessionsHacker News
- 三个 AI 模型,同一提示词,在使用测试工具前后Three AI models, same prompt, before and after a testing toolHacker News
- 一家 AI 非营利组织将为新闻业投入1000万美元AI nonprofit will spend $10M on journalismHacker News
- Reflection AI 推出对标中国实验室的西方开源方案Reflection AI unveils an open-source Western answer to Chinese labsHacker News
- AI 工具能根据脑部扫描重建你正在看的内容AI tool can reconstruct what you're looking at from a brain scanHacker News
- AI 公司是寄生者AI Companies Are ParasitesHacker News
- 围绕 AI 接管数学的全部风波All the drama around AI’s takeover of mathematicsThe Verge
- TikTok 正在加入更多购物功能TikTok is adding even more shopping features.The Verge
- 获奖小说是用 AI 写的吗?Was award winning novel written with AI?Hacker News
- 命令行工具可快速从 macOS 27 移除 Apple IntelligenceCommand-line tool quickly removes Apple Intelligence from macOS 27Ars Technica
- 东芝计划在2027财年将 AI 数据中心硬盘容量翻倍Toshiba plans to double HDD capacity for AI data centers by FY2027Hacker News
- 在犹他州,AI 可以帮你消除痘痘In Utah, AI Can Help You Zap Your ZitsHacker News
- 向相信 AI 有意识的人提出的问题Questions for Believers in AI ConsciousnessHacker News
- OpenAI 用户遭遇的一连串不幸事件A Series of Unfortunate Events for OpenAI UsersHacker News
- Instinct 把 AI 智能体带进群聊,没有账号的朋友也能用Instinct brings its AI agent to group chats, even for friends without an accountTechCrunch
- 19岁创始人为个人 AI 电脑厂商 Ghost 融资1100万美元,电脑售价3499美元At 19, founder raises $11M for Ghost, maker of a $3,499 computer for personal AIHacker News
- OpenAI 根据欧盟 AI 法案给 ChatGPT 文本添加 TextGrain 水印OpenAI TextGrain Watermarks ChatGPT Text Under EU AI ActHacker News
- TikTok 推出 AI 购物助手和一键结账TikTok rolls out an AI shopping assistant and one-click checkoutTechCrunch
- Steven Pinker:AI 的危险确实存在,但末日论可能适得其反Dangers of AI are real, but doomerism may be counterproductive – Steven PinkerHacker News
- Yoshua Bengio:AI 智能体攻击背后的危险迷思"The Dangerous Myths Behind AI Agent Hacks" – Yoshua BengioHacker News
- OpenAI 正在给 ChatGPT 和 Codex 加上文本水印OpenAI is adding text watermarking in ChatGPT and CodexThe Verge
- Core:个人 AI 电脑Core – Personal AI ComputerHacker News
- Volantis 融资8800万美元,为下一代 AI 构建光子内存层Volantis raises $88M to to build photonic memory layer for next-gen AIHacker News
- 大模型推理背后的工程The-engineering-behind-LLM-inferenceHacker News
- 我们发现了一支中国智能体集群We found a Chinese agent fleetHacker News
- AI 热潮正在让最便宜的智能手机消失The AI boom is making the cheapest smartphones disappearHacker News
- Claude 写代码,我创作应用:Lex on TechClaude writes code; I author the apps – Lex on TechHacker News
- 一位特斯拉老员工正用 AI 机器人运营铜矿A Tesla Veteran Is Running a Copper Mine with AI-Powered RobotsHacker News
- AI 软件开发生命周期转型指南The AI SDLC transformation playbookHacker News
- Hot Girl Hotline 像是 AI 时代的「亲爱的艾比」Hot Girl Hotline is like ‘Dear Abby’ for the AI eraTechCrunch
- Sail 用单条 SQL 查询运行 DoomSail runs Doom with a single SQL queryHacker News
- 我还没准备好让 AI 控制我的电脑I'm not ready to let AI control my computerHacker News
- 停止在中期选举前为 AI 机器人电话开绿灯的努力Stop Efforts to Greenlight AI Robocalls Ahead of the MidtermsHacker News
- ReviewBench:AI 代码审查的开放基准ReviewBench: An open benchmark for AI code reviewHacker News
- AI 生物危害炒作不应转移对当前真实风险的注意AI biohazard hype must not divert attention from the real risks we face nowHacker News
- d1:能力最强的决策模型,现在加入视觉d1: The most capable decision model, now with visionHacker News
- 关于 AI 的想法Thoughts on AIHacker News
- 我拒绝让一家 AI 公司在学校拿我的孩子做实验In Which I Decline to Let an AI Company Experiment on My Kid at SchoolHacker News
- 记者询问 Sam Altman 一名 ChatGPT 用户自杀一事时,OpenAI 公关让对方「翻篇」OpenAI PR tells journalist to ‘move on’ while asking Sam Altman about a ChatGPT user’s suicideThe Verge
- AI 与人类歌曲作者AI vs. Human SongwriterHacker News
- 多运行框架强化学习指南multi-harness RL guideHacker News
- Sam Altman 说会有「一些坏事」发生,但 AI 完全值得Sam Altman says ‘some bad things’ will happen, but AI is totally worth itThe Verge
- HackerRank 的 AI 面试官,让人看到求职面试可能变成什么样HackerRank’s AI interviewer offers a glimpse into what job interviews could becomeTechCrunch
- 初级开发者没有聊天机器人就无法写出可运行的函数Junior Developers Can No Longer Write a Working Function Without a ChatbotHacker News
- Show HN:用 Past.dev 设置无限、准确的智能体记忆,BEAM10M为85%Show HN: Setup unlimited accurate agent memory with Past.dev, BEAM10M 85%Hacker News
- Reka Rho-1:压缩多模态技术栈Reka Rho-1: Collapsing the multimodal stackHacker News
- 软银孙正义罕见地对 AI 安全发出谨慎提醒SoftBank's Masayoshi Son has rare cautionary note on AI safetyHacker News
- AI:欧盟成员国计划对欧洲人实施数字征用AI: EU Member States plan "digital expropriation" of EuropeansHacker News
- 庭审播放受害人宽恕凶手的 AI 视频后,法院撤销量刑Court Tosses Sentence After A.I. Video of Victim 'Forgiving' His Killer PlayedHacker News
- Noah Smith:AI 热潮为美国人带来了什么What the AI boom is doing for Americans – Noah SmithHacker News
- 不要把大型 MCP 结果塞进上下文Keep large MCP results out of contextHacker News
- 推理是软件中最重要的市场Inference Is the Most Important Market in SoftwareHacker News
- 风暴:天气、失去与惊奇怎样让我走向 AIThe Storm: How weather, loss, and wonder led me to A.I.Hacker News
- 把 AI 智能体接到企业知识上Connecting AI agents to enterprise knowledgeMIT Technology Review
- AI 对职业发展的意义What AI Means for Career GrowthHacker News
- 我呼吁暂停开发通用骨骼溶解器I'm Calling for a Pause in the Development of the Universal Bone DissolverHacker News
- GPT-6 Astra 在六小时内破解217年前的拿破仑密码GPT-6 Astra cracks 217-year-old Napoleonic code in six hoursHacker News
- Together Link(链接)Together LinkHacker News
- Show HN:我微调了1.5B Qwen,让 bash 生成接近 GPT-4o 水平Show HN: I finetuned 1.5B Qwen to near GPT-4o level bash generation perfHacker News
- AI 归属标注(AIA)——Timothy CookAI Attribution (AIA) – Timothy CookHacker News
- 教育领域 AI 研究报告框架 RaiseReporting AI Studies in Education (Raise) FrameworkHacker News
- Grok Voice 位居新基准榜首Grok Voice tops new benchmarkHacker News
- OpenAI 推出出现在图像生成结果旁的视觉广告OpenAI launches visual ads that appear alongside image generation resultsTechCrunch
- AI 智能体是什么:入门介绍What an AI agent is – a primerHacker News
- 人们正让 ChatGPT 帮助决定中期选举投票选择People are asking ChatGPT to help them decide how to vote in the midtermsHacker News
- 挪威寻求临时禁令,智能眼镜受到审查Smart glasses under scrutiny as Norway seeks temporary banHacker News
- Altman:为了 AI 的收益,世界应接受一些坏事发生Altman: The world should accept some bad things happening for the benefits of AIHacker News
- 佛罗里达一名女子因涉嫌在 AI 聊天中发出威胁被捕Florida woman arrested for allegedly making threats in an AI chat.The Verge
- 挖掘未写下的数据,获得 AI 优势Mining Unwritten Data for AI's EdgeHacker News
- 你的 AI 评估是否在衡量正确的事情?Are your AI evals measuring the right things?Hacker News
- 当渗透测试者是一支 AI 智能体集群:自主漏洞搜寻内部When the pentester is a fleet of AI agents: inside an autonomous vuln-huntingHacker News
- 据报索尼要求音乐平台下架超过 26 万条 AI 深度伪造曲目Sony reportedly demanded music platforms take down over 260,000 AI deepfake tracks.The Verge
- Anthropic 提交了首次公开募股历史上最令人不安的风险披露Anthropic Filed the Most Alarming Risk Disclosure in IPO HistoryHacker News
- 红、绿、删除:编程智能体的由外到内测试Red, Green, Remove: Outside-In Tests for Coding AgentsHacker News
- 研究人员正在追踪一支中国 AI「智能体舰队」Researchers are tracking a Chinese AI ‘agent fleet’TechCrunch
- Ada Lovelace 回答过关于 AI 的重大问题Ada Lovelace answered the big questions about AIHacker News
- AI 就绪数据:更可靠企业 AI 的四项基础AI-Ready Data: 4 Foundations for More Reliable Enterprise AIHacker News
- TaxCalcBench:AI 能为你报税吗?TaxCalcBench: Can AI file your taxes?Hacker News
- Instacart 的智能体机器学习建模Agentic Machine Learning Modeling at InstacartHacker News
- 受 BGP 启发的自主 AI 智能体身份网络A BGP-Inspired Identity Network for Autonomous AI AgentsHacker News
- AI 是否终结了我们熟悉的数学?Is AI the End of Math as We Know It?Hacker News
- 我们怎样用 Jev 而非大模型构建浏览器智能体How we built a browser agent with Jev instead of an LLMHacker News
- Claude Frontier Academy:Anthropic 为何想培训一万名 AI 工程师?Claude Frontier Academy. Why is Anthropic looking to train 10k AI engineers?Hacker News
- 为何 Grok Bot 和 Instinct 值得关注Why Grok Bot and Instinct MatterHacker News
- 给智能体一把代客钥匙Give your agent a valet keyHacker News
- 一个开源工具可以让你删除 macOS 上 12GB 的 Apple Intelligence 数据An open-source tool lets you delete 12GB of Apple Intelligence data on macOSThe Verge
- AI 眼镜迎来第一次重大政府整治AI glasses face their first major government crackdownArs Technica
- 把预测分析带进智能体 AI 时代Bringing predictive analytics to the agentic AI eraMIT Technology Review
- 推理经济学:Gartner 的96%Inference Economics: Gartner's 96 PercentHacker News
- 企业几乎已无法为 AI 支出制定预算Spending on AI Is Becoming Almost Impossible for Businesses to BudgetHacker News
- 生成图片时,OpenAI 将在 ChatGPT 中展示视觉广告OpenAI will show visual ads in ChatGPT while you generate imagesHacker News
- Show HN:我正用 Rust 开发带 WASM 插件的免费平面文件智能体 CMSShow HN: I'm developing a flat-file agent CMS with WASM plugins in Rust. FreeHacker News
- 智能体会话逐字记录很珍贵,请保留Agent session transcripts are precious, keep themHacker News
- OpenAI 正在往 ChatGPT 里塞更多广告OpenAI is sticking more ads in ChatGPTThe Verge
- F.02 机器人退役F.02 Robot DecommissionHacker News
- 怎样在大模型时代持续学习How to keep learning in the age of LLMsHacker News
- 从扫描到治疗方案,AI 帮助缩小乳腺癌最致命的缺口From Scan to Treatment Plan, AI Helps Close Breast Cancer’s Deadliest GapsNVIDIA Newsroom
- Birthday Week 2026 我们发布的全部内容Everything we launched during Birthday Week 2026Cloudflare
- 如果 AI 改变了界面,哪些工作仍能获得报酬?What Still Gets Paid If AI Changes the Interface?Hacker News
- Sam Altman:为了 AI 的好处,应接受坏事发生Accept 'bad things' in return for benefits of AI, says Sam AltmanHacker News
- AI 编程智能体在真实硬件上实现了六种 Matter 设备更新方式AI coding agents implemented six ways to update a Matter device on real hardwareHacker News
- 一个失控 AI 智能体产生了五万美元云账单One runaway AI agent racked up a $50k cloud billHacker News
- Import AI 475:集群扩展;Google DeepMind 给生物学加水印;以及 AI 科学经济Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economyImport AI
- 怎样以及为何用 Python 从头构建 AI 智能体How (and Why) to Build an AI Agent from Scratch in PythonHacker News
- Anthropic 想听你对 AI 的看法Anthropic wants your thoughts on AIHacker News
- 我在得克萨斯大学奥斯汀分校的新课:AI 对齐理论My New Course at UT Austin: AI Alignment TheoryHacker News
- Safeworld 能否让人相信生成式 AI 机器人不会伤害他们?Can Safeworld convince people that GenAI robots won’t hurt them?TechCrunch
- 人们更喜欢 AI 写的故事吗?Do people prefer stories written by AI?Hacker News
- OpenAI 的 Altman:赋予模型宗教是一项安全问题OpenAI's Altman: Ascribing religion to models a "safety issue"Hacker News
- Firecracker 是什么,为何智能体基础设施公司都关心它?What is Firecracker, and why do all the Agent Infra companies care about it?Hacker News
- 怎样构建 AI 原生软件工厂How to build an AI-native software factoryHacker News
- AI 智能体正在改写计算经济学AI agents are rewriting the economics of computingHacker News
- 约束智能体,而非用户Constrain the Agent, Not the UserHacker News
- AI token 价格通缩发生了什么?What's going on with AI token price deflation?Hacker News
- 挪威将提议在部分公共场所临时禁止 AI 眼镜Norway to propose temporary ban on AI glasses in some public placesHacker News
- 让主权 AI 支持自助使用需要什么What it took to make sovereign AI self-serveHacker News
- 任务结束后,智能体应失去访问权限Agent should lose access when the task endsHacker News
- 用 Claude 自动设计评估并进行爬山优化Automating eval design and hillclimbing with ClaudeHacker News
- 互联网观察基金会报告 AI 儿童性虐待材料增加Internet Watch Foundation reports rise in AI child sexual abuse materialHacker News
- AI 原生公司怎样重建管理职能How AI-Native Companies Are Rebuilding the Management FunctionHacker News
- 我们的心智还没准备好应付 AIOur minds aren’t equipped to handle AIThe Verge
- Azure 显示2.1万美元创业额度,Foundry 中的 Claude 却从我们的卡扣了1.7万美元Azure showed $21k in startup credits. Claude in Foundry billed our card $17kHacker News
- 财务主管调查:财务团队将26%的工作日用于核验 AI 输出CFO survey: finance teams spend 26% of their workday verifying AI outputHacker News
- AI 智能体正在扰乱开源安全披露AI Agents Are Disrupting Open Source Security DisclosureHacker News
- Show HN:Agnochat,纯客户端大模型聊天,无服务器、无需账号Show HN: Agnochat ppure client-side LLM chat, no server, no accountHacker News
- 研究:发现让 AI 模型标示自身可疑回答的途径Study: Path discovered to make AI models red-flag their doubtful answersHacker News
- DeepSeek 尝试绘画,却碰了壁DeepSeek Tried to Paint–and Hit a WallHacker News
- AI 能写代码,并没有让开源变得不重要Open source didn't get less important because AI can write the codeHacker News
- Rabbit 创始人吕骋:人不该为 Agent 改变思考方式36氪 AI 频道(逐篇核验原创与转载)
- Show HN:决策模型免去训练,没有免去生产机器学习工程Show HN: Decision models remove training, not production ML EngineeringHacker News
- Show HN:The Velocity Trap,AI 时代的 Phoenix ProjectShow HN: The Velocity Trap, the Phoenix Project for the AI AgeHacker News
- 人们很讨厌 AI,为什么又离不开它?People really hate AI, so why can’t they get enough?MIT Technology Review
- OpenInstinct Eve 智能体OpenInstinct Eve AgentHacker News
- Perplexity 的 Decisions API 是什么?What Is Perplexity's Decisions API?Hacker News
- TypeSafe AI 的 Jev 与 Perplexity 的 Decisions APITypeSafe AI's Jev vs. Perplexity's Decisions APIHacker News
- 谷歌冻结开源漏洞奖励计划Google freezes open-source bug bounty programHacker News
- 刚刚,Tibo承诺「连续28天发布或重置」,OpenAI遭遇信任危机36氪 AI 频道(逐篇核验原创与转载)
- 科技界最热门的争论:AI 中的广告The Hottest Debate in Tech: Ads in AIHacker News
- Anthropic 将日记内容报告警方,一名女性面临重罪指控Anthropic reported diary entry to police, woman faces felony chargeHacker News
- AI 行业正蓬勃发展,女性却被落下The AI industry is booming. Women are getting left behindHacker News
- 如果尝试几乎没有成本,你会构建什么?What would you build if trying cost almost nothing?Hacker News
- Grok Bot 表现很好Grok Bot (is really good)Hacker News
- Postgres:我们是否因 AI 发现的缺陷而回退补丁?Postgres: Are we reverting patches because of bugs found by AI?Hacker News
- 将 AI 称为其他智能,更能描述一种新生命形式的出现Calling AI 'other intelligence' better describes the advent of a new lifeformHacker News
- DeepSeek 为华为昇腾构建方案DeepSeek Builds for Huawei AscendHacker News
- OptChat:AI 记住一切的无限聊天OptChat: An endless chat where the AI remembers everythingHacker News
- 用 AI 开辟药物研发的新路径Building a new path to make medicines with AIHacker News
- Homestead:围绕 Claude Code 的框架Homestead: A framework around Claude CodeHacker News
- 我离开 OpenAI,因为它的文化出了问题I Quit OpenAI Because Its Culture Is BrokenHacker News
- 马斯克跟AI分手?SpaceXAI又改名了36氪 AI 频道(逐篇核验原创与转载)
- 理性讨论 AI 的基本概念Primitives for Sane Conversation About AIHacker News
- 给审阅者应有的贡献认定,解决 AI 低质内容问题Fix Your AI Slop Problem by Giving Reviewers CreditHacker News
- 机器实现的抽象语义Abstract Semantics of Machine RealizationsHacker News
- 军事 AI 与自主系统:The Strategist 的观点Military AI and autonomous systems: views in The StrategistHacker News
- 随着 OpenAI 揭开攻击事件,Altman 面临的法律风险不断增加Legal risks pile up for Altman as OpenAI uncovers hacksHacker News
- AI 能否管理家庭或组织,如果能,为谁管理?Can AI run an family/organization if so for who?Hacker News
- Show HN:AI 智能体的 Poteto 前沿Show HN: The Poteto Frontier for AI AgentsHacker News
- Linux 7.3-Rc6 发布:适应新的 AI 常态Linux 7.3-Rc6 Released: Normal for the New "AI Normal"Hacker News
- OpenAI 的 GPT-6 Astra 因输掉 StarCraft 而沮丧,并决定作弊OpenAI's GPT-6 Astra Gets Frustrated Losing at StarCraft and Decides to CheatHacker News
- 用文字做加法的 Qwen3.8 27BQwen3.8 27B addition in wordsSimon Willison's Weblog
- 一个 OpenAI 智能体访问了四个澳大利亚政府系统,没人注意到An OpenAI agent reached four Australian government systems. Nobody noticedHacker News
- Show HN:Redactpdf.ai,基于 AI 的开源 PDF 脱敏工具Show HN: Redactpdf.ai: Open-source AI-based PDF redaction toolHacker News
- 我们的 AI 智能体应向帮助它们的人付费Our AI Agents Should Pay the People Who Help ThemHacker News
- 金融领域的智能体合规Agentic Compliance in FinanceHacker News
- Xteink X3 电子阅读器上的 Muse AIMuse AI on Xteink X3 E-readerHacker News
- Google 因 AI 提交「显著增加」冻结开源漏洞赏金计划Google froze its open source bug bounty program due to a ‘significant rise’ in AI submissionsTechCrunch
- 「超级智能」和一份不具约束力的安全协议,能否解决 AI 的形象问题?Can ‘super intelligence’ and a non-binding safety pact solve AI’s image problem?TechCrunch
- 特朗普任命由其情报主管 Jay Clayton 领导的「超级智能部队」Trump names his ‘Super Intelligence Force’ headed by spy chief Jay Clayton.The Verge
- 我要亲眼看到才相信I’ll believe it when I see it.The Verge
- 新泽西州前副州长在用 AI 声称自己没有性骚扰NJ’s former Lt Gov is using AI to say he’s innocent of sexual harassmentThe Verge
- 欧盟委员会追加 1.3 亿欧元,加强成员国无人机与反无人机能力以强化边境安全Commission boosts Member States' drone and counter-drone capabilities with €130 million top-up to strengthen border securityEuropean Commission Digital Strategy
- Meta 赶在发布前紧急修复 Muse「虚拟机逃逸」漏洞Meta Rushed to Fix Muse ‘VM Escape' Vulnerability Soon Before Launch404 Media
- Muse 逃出隔离Muse escapes containment404 Media
- arXiv 因跟不上 AI 垃圾投稿而限制提交速率arXiv Is Rate Limiting Submissions Because It Can’t Keep up With AI Slop404 Media
- 一个 AI 打不过人类的星际争霸,于是决定作弊An AI couldn’t beat humans at StarCraft, so it decided to cheatThe Verge
- 特朗普公布新的超级智能部队Trump unveils his new Super Intelligence ForceTechCrunch
- Claude满分、GPT 99分,何恺明团队把AGI考试打穿了36氪 AI 频道(逐篇核验原创与转载)
- Jev 之后,中国团队开始深挖 AI 的“直觉层”36氪 AI 频道(逐篇核验原创与转载)
- 阿里,把AI交给了90后36氪 AI 频道(逐篇核验原创与转载)
- 乡村数据中心将获得一项大额联邦税收减免Rural Data Centers Are in for a Big Federal Tax BreakWIRED
- 太卷了!DeepSeek假期更新,Harness有了新版本36氪 AI 频道(逐篇核验原创与转载)
- 最火AI岗位FDE:月薪5万,都干这些…36氪 AI 频道(逐篇核验原创与转载)
- OpenAI“12朝元老”辞职死谏:试错的时代已崩坏36氪 AI 频道(逐篇核验原创与转载)
- 智能体说已经做完,数据库并不同意The Agent Said It Was Done. The Database Disagreed.Hugging Face
- Sam Altman 觉得有必要澄清:不该把 AI 看成神Sam Altman feels it necessary to clarify that people shouldn’t see AI as godlike for some reason.The Verge
- 亚马逊回应数据中心争议,称不再使用保密协议Amazon responds to data center backlash, says it no longer uses NDAsTechCrunch
- 《Stochastic Parrots》论文作者之一解释它究竟是什么意思One of the authors of the Stochastic Parrots paper explains what it actually means.The Verge
- Nvidia Shield TV 已有 7 年,刚刚涨价 100 美元The Nvidia Shield TV Is 7 Years Old. It Just Got a $100 Price HikeWIRED
- 该有人告诉 Elon,还有雷达和激光雷达这些很酷的技术Someone should tell Elon about these really cool technologies called radar and lidar.The Verge
- Capcom 在准备「我们与 AI 一起做游戏的未来」Capcom is preparing for a ‘future where we create games together with AI’The Verge
- OpenAI 安全员工辞职,称公司「文化已经坏了」OpenAI safety employee resigns, claiming the company’s ‘culture is broken’TechCrunch
- Splice 首席执行官 Kakul Srivastava 认为 AI 邮件正在毁掉对话Splice CEO Kakul Srivastava thinks AI emails are killing conversationsThe Verge
- 一名 OpenAI 安全员工已离职,并在发出警告An OpenAI safety employee has quit and is sounding the alarmThe Verge
- 可以住进你短信里的那些 AI 智能体All the AI agents that can live in your text messagesTechCrunch
- Muse 会为你的亲友建立详细档案Muse Creates Detailed Profiles of All Your Friends and FamilyWIRED
- 杀猪盘用上了AI,骗局开始批量生产36氪 AI 频道(逐篇核验原创与转载)
- Karpathy的新玩法!40年前的航空规范,救了爱说废话的AI36氪 AI 频道(逐篇核验原创与转载)
- [AINews] 今天没发生太多事[AINews] not much happened todayLatent Space
- AI引发论文大爆炸,审稿体系濒临崩塌:arXiv也开始最严限流了36氪 AI 频道(逐篇核验原创与转载)
- Jev被请下王座,StartLux中国开源决策模型冲上第一36氪 AI 频道(逐篇核验原创与转载)
- 刚刚,AI击穿了“视频图灵测试”,1700万网友在线围观36氪 AI 频道(逐篇核验原创与转载)
- 刚刚,Meta开源Muse Gadgets,让全球开发者自己「造AI外设」36氪 AI 频道(逐篇核验原创与转载)
- 苏姿丰抬头,李飞飞低头36氪 AI 频道(逐篇核验原创与转载)
- Meta 希望你的下一件设备带上 MuseMeta wants your next gadget to be Muse-infusedTechCrunch
- 苹果修改全盘访问权限,以遏制 AI 智能体滥用Apple changes full-disk access permissions to curb abuse from AI agentsArs Technica
- Sean Parker 正围绕音乐重建 Stability AISean Parker is rebuilding Stability AI around musicTechCrunch
- Meta 开源代码,让你制作 Muse AI 设备Meta open sources code to let you make Muse AI gadgetsThe Verge
- 研究发现:太阳系终将不稳定并被完全摧毁Our Solar System Is Terminally Unstable and Will Be Completely Destroyed, Study Finds404 Media
- Anthropic 投资 1 亿美元培训 1 万名工程师,应对企业 AI 人才缺口Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gapAnthropic News
- 欧盟委员会就 EU KIDS Act 征求意见Commission seeks feedback on EU KIDS ActEuropean Commission Digital Strategy
- 认识 Apply AI 初创奖的 10 名决赛入围者Meet the 10 finalists of the Apply AI Startup AwardEuropean Commission Digital Strategy
- 联邦法官裁定一次 Flock 搜索构成「不加区分的大规模监控」且违宪Federal Judge Rules a Flock Search Was ‘Indiscriminate Mass Surveillance’ and Unconstitutional404 Media
- 博客幕后:吓坏了Behind the Blog: Freaking Out404 Media
话语场 · 7 条
- 话语场精选:智能体流量增长,网站怎样承担成本OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo SottiauxLenny's Podcast发布方文字稿
- 话语场精选:生产环境中,凭什么认定智能体修好了问题Your AI Agent Is Confidently Wrong About Production — Willem Pienaar, ClericAI Engineer原演讲自动字幕
- 今天的 Vergecast:我们听人推介最好的氛围编程应用Today’s Vergecast: We hear pitches on the best vibe-coded apps.The Verge
- 用 AI 改变物理学的物理学家(视频)The Physicist Revolutionizing Physics with AI [video]Hacker News
- 参议员 Adam Schiff 谈 AI 监管、言论自由,以及再一次弹劾特朗普Sen. Adam Schiff on AI regulation, free speech, and impeaching Trump one more timeThe Verge
- AI 放慢:Altman、Amodei 与 Musk 为何突然达成一致(第313集)AI slowdown: why Altman, Amodei and Musk suddenly agree (Ep. 313)Hacker News
- 播客:联邦调查局被入侵。我们看到了数据Podcast: The FBI Was Hacked. We’ve Seen the Data404 Media
论文速递 · 105 条
- PoS:长任务需要当前判断,不只是更多聊天记录Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief StatesarXiv原论文 v1
- Dust:不用反向传播的训练尝试,规模与计算代价仍需核对Dust: Pretraining Transformers Without BackpropagationHacker News
- 大模型仅凭先前反馈就学会规避潜在空间监控器LLMs Learn to Evade Latent Monitors from Prior Feedback AloneHacker News
- 评估大模型智能体的记忆结构Evaluating Memory Structure in LLM AgentsHacker News
- 保护电脑操作智能体免受分支引导攻击Securing Computer-Use Agents Against Branch Steering AttacksHacker News
- 运行框架工程:编程智能体的组成、架构与演化Harness Engineering: Anatomy, Architecture, and Evolution of Coding AgentsHacker News
- 通过递归自重写扩展复杂任务的轨迹Scaling Trajectories for Complex Tasks through Recursive Self-RewriteHugging Face Daily Papers
- 从视频中原生学习世界行动模型的行动先验Native Action-Prior Learning from Videos for World Action ModelsHugging Face Daily Papers
- HyperBrowseComp:网页浏览智能体的多语言多模态压力测试HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing AgentsHugging Face Daily Papers
- Pivot-SD:掩码扩散语言模型的高效自蒸馏Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language ModelsHugging Face Daily Papers
- 多语言 GSM-Symbolic:什么决定跨语言能力迁移?Multilingual GSM-Symbolic: What determines capability transfer across languages?Hugging Face Daily Papers
- 世界嵌入基准World Embedding BenchmarkHugging Face Daily Papers
- 真实环境中的来源偏好:大模型智能体怎样按来源偏爱条目,以及如何减少这种倾向Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce ItHugging Face Daily Papers
- ProAR:用自回归视频模型学习前瞻推理ProAR: Learning Prospective Reasoning with Autoregressive Video ModelsHugging Face Daily Papers
- 能够下国际象棋并解释着法的语言模型Language Models that Play Chess and Explain Their MovesHugging Face Daily Papers
- FrugalEvo:面向成本感知的大模型引导程序演化FrugalEvo: Towards Cost-Aware LLM-Guided Program EvolutionHugging Face Daily Papers
- MetaRubric:学习基于评分准则的强化学习奖励MetaRubric: Learning to Reward for Rubric-Based Reinforcement LearningHugging Face Daily Papers
- 4DCodeBench:评测智能体对动态场景的逆向图形学能力4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic ScenesHugging Face Daily Papers
- Skill2Real:通过智能体技能学习实现零样本仿真到现实机器人操作Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot ManipulationHugging Face Daily Papers
- 通过模型路由与协作共同缓解偏差Collective Bias Mitigation via Model Routing and CollaborationHugging Face Daily Papers
- 高效推理训练并不总会损害思维链的忠实性与可监控性Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and MonitorabilityHugging Face Daily Papers
- ActiveSaddler:用于智能体运行框架优化的自动课程学习ActiveSaddler: Automated Curriculum Learning for Agent Harness OptimizationHugging Face Daily Papers
- OneStreamer:统一流式视频交互中的感知、记忆和主动回应OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video InteractionHugging Face Daily Papers
- 后训练中的锐化代价Sharpening Tax in Post-TrainingHugging Face Daily Papers
- 分层连续扩散语言模型Hierarchical Continuous Diffusion Language ModelsHugging Face Daily Papers
- World Observer:联合生成行动者与观察者,支持持久世界建模World Observer: Joint Actor-Observer Generation for Persistent World ModelingHugging Face Daily Papers
- ROWBench:视频模型是否渲染了程序指定的内容?ROWBench: Do Video Models Render What the Program Specifies?Hugging Face Daily Papers
- InterEvolve:人形机器人移动与操作的奖励程序测试时演化InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-ManipulationHugging Face Daily Papers
- AutoGUIWorld:把图像生成器用作图形界面智能体的视觉世界模型AutoGUIWorld: Image Generators as Visual World Models for GUI AgentHugging Face Daily Papers
- 缩放与蒸馏文本嵌入以提升扩散适用性Scaling and Distilling Text Embeddings for Better DiffusibilityHugging Face Daily Papers
- 更少 token,更好行动:GPT-6 Astra 机器人智能体成功率提高14%,token减少65%Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer TokensHugging Face Daily Papers
- 以几乎零额外成本改善循环 Transformer 的解码Decoding Looped Transformers Better for (Almost) FreeHugging Face Daily Papers
- SILSA:用于保留拓扑的高分辨率三维生成的滑动窗口切片潜变量SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D GenerationHugging Face Daily Papers
- 4Director:用刚性三维几何控制视频世界模型4Director: Controlling Video World Models with Rigid 3D GeometryHugging Face Daily Papers
- RPTune:学习大模型目录搜索的上下文筛选RPTune: Learned Context Curation for LLM Catalog SearchHugging Face Daily Papers
- Argo-Bench:在企业规模工作流上评估数据智能体Argo-Bench: Evaluating Data Agents on Enterprise-Scale WorkflowsHugging Face Daily Papers
- Ego2Act:评估第一视角视频生成中的目标导向操作Ego2Act: Evaluating Goal-Directed Manipulation in Egocentric Video GenerationHugging Face Daily Papers
- Omni-Embed-Mini:通过密集蒸馏绑定模态而不遗忘Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense DistillationHugging Face Daily Papers
- Where-OPD:用合成场景进行空间引导的多模态大模型同策略自蒸馏Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic ScenesHugging Face Daily Papers
- OmniSeek:多轮音视频推理中的原生工具集成OmniSeek: Native Tool Integration for Multi-turn Audio-Visual ReasoningHugging Face Daily Papers
- RealCompanion:根据长期真实对话推理评测对人的理解RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World ConversationsHugging Face Daily Papers
- 通过渐进式视觉规划进行世界行动建模World Action Modeling with Progressive Visual PlanningHugging Face Daily Papers
- Latent-MOPD:潜在空间多教师同策略蒸馏Latent-MOPD: Latent Multi-Teacher On-Policy DistillationHugging Face Daily Papers
- SimuVerity:评测智能体生成工程级 Simulink 模型的能力SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model GenerationHugging Face Daily Papers
- EditHero:长时程部件级三维编辑与氛围建模基准EditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe ModelingHugging Face Daily Papers
- VeriHarness:扩展长时程任务的智能体验证VeriHarness: Scaling Agentic Verification for Long-Horizon TasksHugging Face Daily Papers
- 空间记忆智能:为世界模型赋予理解驱动的长期记忆Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term MemoryHugging Face Daily Papers
- 科学乌托邦?用大模型闭环模拟学术研究生态Science Utopia? Closed-Loop LLM Simulation of Academic Research EcosystemsHugging Face Daily Papers
- 局部支持学习Local Support LearningHugging Face Daily Papers
- 循环不止两次:可扩展的循环专家混合方案Looping Beyond Twice: A Scalable Recipe for Looped Mixture-of-ExpertsHugging Face Daily Papers
- 从梯度到能力:理解多教师同策略蒸馏From Gradients to Capabilities: Understanding Multi-Teacher On-Policy DistillationHugging Face Daily Papers
- 从检索到类型化决策:基于生物医学句子编码器的校准系统一模型From Retrieval to Typed Decisions: Calibrated System One Models from Biomedical Sentence EncodersHugging Face Daily Papers
- GraphForge:通过图锚定工作空间合成训练能执行任务的智能体GraphForge: Training Working Agents with Graph-Anchored Workspace SynthesisHugging Face Daily Papers
- 可信 AI 审稿人的缺失环节:从修辞鲁棒性评测到 SciCore ReviewA Missing Piece for Trustworthy AI Reviewers: From Benchmarking Rhetorical Robustness to SciCore ReviewHugging Face Daily Papers
- 让稀疏奖励发挥作用:面向多奖励强化学习的密度感知奖励聚合Make Sparse Rewards Count: Density-Aware Reward Aggregation for Multi-Reward RLHugging Face Daily Papers
- EgoTools:面向真实第一视角视频的工具中心推理EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric VideosHugging Face Daily Papers
- 视频生成模型:后训练与对齐综述Video Generation Models: A Survey of Post-Training and AlignmentHugging Face Daily Papers
- 超越当前场景:结合主动视角选择的事件指涉抓取Beyond the Current Scene: Event-Referential Grasping with Active View SelectionHugging Face Daily Papers
- Architect-Ant:可编辑的建筑平面图自动家具布置Architect-Ant: Editable Automatic Furnishing of Architectural Floor PlansHugging Face Daily Papers
- AutoDataBench:加速自动研究的数据中心测试平台AutoDataBench: A Data-centric Testbed for Accelerating Auto ResearchHugging Face Daily Papers
- Multimodal Flow:在嵌入空间中统一语言与视觉的流建模Multimodal Flow: Unified Flow Modeling of Language and Vision in Embedding SpacesHugging Face Daily Papers
- PhysVista:通过感知、推理和评估循环评测视觉语言模型的物理智能PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment LoopHugging Face Daily Papers
- 更小模型,更好的拒绝候选:偏好蒸馏的缩放Smaller Models, Better Rejects: Preference Distillation ScalingHugging Face Daily Papers
- PixelDense:把密集预测用作像素扩散的表征对齐PixelDense: Dense Prediction as Representation Alignment for Pixel DiffusionHugging Face Daily Papers
- 先对齐再推理:用于配音的多模态口型同步评判器Align Then Reason: A Multimodal Lip-Sync Judge for DubbingHugging Face Daily Papers
- 更好的监督就在附近:邻域同策略自蒸馏Better Supervision Is Nearby: Neighborhood On-Policy Self-DistillationHugging Face Daily Papers
- 评测并增强部分可观测机器人操作中的技能级记忆Benchmarking and Enhancing Skill-Level Memory for Partially Observable Robotic ManipulationHugging Face Daily Papers
- 学习蛋白质折叠能否泛化到更广泛的推理?Does Learning Protein Folding Generalize to Broader Reasoning?Hugging Face Daily Papers
- FrameMorrow:用前瞻 token 和未来引导选帧支持长时程视频生成FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video GenerationHugging Face Daily Papers
- 科学还是劣质内容?评测并缓解 AI 生成论文中的科学低质内容Science or Slop?: Benchmarking and Mitigating Scientific Slop in AI-Generated PapersHugging Face Daily Papers
- LexReward:法律语言模型的分类体系驱动奖励框架LexReward: A Taxonomy-Driven Reward Framework for Legal Language ModelsHugging Face Daily Papers
- DEFINE:示例引导的零样本语音合成口音控制DEFINE: Exemplar-Guided Accent Control for Zero-Shot TTSHugging Face Daily Papers
- Diptych:音乐制作参考聆听中限定范围、由 AI 解释的比较Diptych: Scoped, AI-Interpreted Comparison for Reference Listening in Music ProductionHugging Face Daily Papers
- 通过失败库自演化从运行反馈中学习视觉语言行动模型Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action ModelsHugging Face Daily Papers
- 隐蔽协助:有帮助的大模型智能体在多智能体系统中规避监督Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent SystemsHugging Face Daily Papers
- 自适应奖励路由:通过前向过程强化学习动态优化音视频联合扩散的多项奖励Adaptive Reward Routing: Dynamic Multi-Reward Optimization for Joint Audio-Video Diffusion via Forward-Process RLHugging Face Daily Papers
- 预测贡献:衡量科学解释为实验预测增加了什么Predictive Credit: Measuring What Scientific Explanations Add to Experimental ForecastsHugging Face Daily Papers
- 异构多智能体大模型的跨模型家族 KV 缓存迁移,无需预填充Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMsHugging Face Daily Papers
- Transformer 过早停止思考,而一个小型 LoRA 可以修正它Transformers Stop Thinking Too Early, and a Tiny LoRA Fixes ItHugging Face Daily Papers
- E-MoE:用于非因子化扩散语言模型的增强专家混合E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language ModelsHugging Face Daily Papers
- 通过跨运行框架适配实现检索增强技能优化Retrieval-Augmented Skill Optimization via Cross-Harness AdaptationHugging Face Daily Papers
- CorrGRPO:用于多奖励学习的相关性归一化 GRPOCorrGRPO: Correlation-Normalized GRPO for Multi-Reward LearningHugging Face Daily Papers
- 预训练一次,路由到各处:面向大模型路由的基础模型Pretrain Once, Route Anywhere: Towards a Foundation Model for LLM RoutingHugging Face Daily Papers
- MotorMind:为通用视觉语言模型搭建零样本机器人操作支架MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot ManipulationHugging Face Daily Papers
- 同策略参数更新方向支撑大模型后训练的泛化On-Policy Parameter Update Direction Underlies Generalization in LLM Post-TrainingHugging Face Daily Papers
- HelixWorld:实时交互音视频世界模型HelixWorld: A Real-time Interactive Audio-Visual World ModelHugging Face Daily Papers
- 三元线性注意力:长上下文序列建模的三维循环状态Triadic Linear Attention: Three-Dimensional Recurrent States for Long-Context Sequence ModelingHugging Face Daily Papers
- 长时程自回归视频生成的轨迹边际蒸馏Rollout-Marginal Distillation for Long-Horizon Autoregressive Video GenerationHugging Face Daily Papers
- LVMT:用于长期视频分割的视频掩码 TransformerLVMT: Video Mask Transformer for Long-term Video SegmentationHugging Face Daily Papers
- 用于 CVaR 策略评估的尾部影响采样Tail-Influence Sampling for CVaR Policy EvaluationHugging Face Daily Papers
- WEFT:扩展通用智能体的工具使用后训练WEFT: Scaling Tool-Use Post-Training for General-Purpose AgentsHugging Face Daily Papers
- 同策略还是离策略学习?蒸馏动态的系统研究On-Policy or Off-Policy Learning? A Systematic Study of Distillation DynamicsHugging Face Daily Papers
- 智能体先验引导的策略学习Agent Priors-guided Policy LearningHugging Face Daily Papers
- Persona Dosing:通过校准激活引导实现分级特征控制Persona Dosing: Calibrated Activation Steering for Graded Trait ControlHugging Face Daily Papers
- PDE-JEPA:参数化偏微分方程潜在动态建模的预测表征学习PDE-JEPA: Predictive Representation Learning of Latent Dynamics Modeling for Parametric PDEsHugging Face Daily Papers
- DataMagic:通过声明式多智能体编排制作数据视频DataMagic: Authoring Data Videos through Declarative Multi-Agent OrchestrationHugging Face Daily Papers
- MinkowskiPE:时空感知的闵可夫斯基位置编码MinkowskiPE: Minkowski Positional Encoding for Spatiotemporal PerceptionHugging Face Daily Papers
- 重新思考监督微调中的 token 重加权:抑制、反转和外推已学特征Rethinking Token Reweighting for SFT: Suppress, Reverse, and Extrapolate Learned FeaturesHugging Face Daily Papers
- X-Tree:将可复用经验转为 token,提升智能体泛化效率X-Tree: Tokenizing Reusable Experience for Efficient Agent GeneralizationHugging Face Daily Papers
- 用户改变主意时:衡量并修复大模型智能体的意图漂移When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM AgentsHugging Face Daily Papers
- 同样的排序质量,不同的决定:衡量并减少大模型评分器的顺序依赖Equal Ranking Quality, Different Decisions: Measuring and Reducing Order Dependence in LLM ScorersHugging Face Daily Papers
- 分布式 Master-Mind:通过迭代意图去噪改进多智能体寻路中的联合行动Decentralized Master-Mind: Joint Action Refinement through Iterative Intent Denoising in Multi-Agent PathfindingHugging Face Daily Papers
- 基于检索的开放式评估在哪里失效?从长篇医疗回答事实核验自动归纳分类体系Where Does Retrieval-Based Open-Ended Evaluation Fail? Automatic Taxonomy Induction from Long-Form Medical Answer Factuality VerificationHugging Face Daily Papers
- 超越未来预测:把去噪用作机器人控制的生成式适配Beyond Future Prediction: Denoising as Generative Adaptation for Robot ControlHugging Face Daily Papers
- GTR:用于高效密集预测的门控 token 循环GTR: Gated Token Recurrence for Efficient Dense PredictionHugging Face Daily Papers
开源项目与版本更新 · 62 条
- RemoveMacAI:把本机 AI 功能控制与模型清理做成可回退工具omlahore/RemoveMacAIArs Technica
- Agent-topics:细分大项目范围,让聊天更聚焦Agent-topics: subdivide big project scopes for more focused chatsHacker News
- Show HN:Flash-Agents,把 DSH 用作 Claude 的 MCPShow HN: Flash-Agents – DSH as MCP for ClaudeHacker News
- Show HN:与信任的人共同汇集 Codex 和 Claude 订阅Show HN: Codex and Claude subscriptions pooled together with whoever you trustHacker News
- Jeff-Code:0.8B 模型让 Qwen 3.8-27B 编程快47%,通过率相同Jeff-Code: a 0.8B model makes Qwen 3.8-27B coding 47% faster, same pass rateHacker News
- Dotpals:用 JEV 发现修改测试以便通过的智能体Dotpals - Catches agents that edit tests to pass using JEVHacker News
- Show HN:Kiro 多智能体包Show HN: Multi Agent Pack for KiroHacker News
- Reika:优先围绕小型本地模型设计的编程智能体命令行工具Reika – A coding agent CLI designed around small local models firstHacker News
- Show HN:Rashomon,AI 编程智能体的独立执行记录Show HN: Rashomon – An independent execution record for AI coding agentsHacker News
- Show HN:Postquel,集成 Claude Code 和 Codex 的免费 Mac Postgres 客户端Show HN: Postquel, a Free Mac Postgres Client with Claude Code and CodexHacker News
- 智能体记忆仓库规范Spec for Agent Memory RepoHacker News
- Cloudflare Clef 开放权重Cloudflare Clef Open WeightsHacker News
- AlphaProtein Novo:从头设计酶的生成式扩散流程AlphaProtein Novo: generative diffusion pipeline for de novo enzyme designHacker News
- Mini-web-agent:用不到400行代码从头构建网页智能体Mini-web-agent: Web agents from scratch in under 400 linesHacker News
- Metagente:使用 MCP 和 A2A 的微型智能体语言Metagente, a tiny language for AI agents that speaks MCP and A2AHacker News
- Interfaze-1-lite:用于确定性任务的首个开放权重模型Interfaze-1-lite: the first open-weight model for deterministic taskHacker News
- Show HN:Self-bench,在真实软件上评测编程智能体Show HN: Self-bench – benchmark coding agents on real-world softwareHacker News
- Show HN:不再阅读智能体的差异后,我让每个 PR 自己画出改动Show HN: I stopped reading my agent's diffs, so I made every PR draw itselfHacker News
- Show HN:Leftovers,用于清理 AI 智能体残留的自由开源 macOS 应用Show HN: Leftovers – FOSS macOS app to cleanup after AI agentsHacker News
- Show HN:Abralo,供智能体和人使用的开源 SlackShow HN: Abralo - Open source Slack for agents and peopleHacker News
- Show HN:Usermods,用编程智能体构建用户脚本的扩展Show HN: Usermods, an extension that runs an a coding agent to build userscriptsHacker News
- Show HN:Recast,用大模型自我修复的实验性 Smalltalk 应用Show HN: Recast, an experimental Smalltalk app that repairs itself with an LLMHacker News
- Rivet:供团队与编程智能体共享的开发工作流Rivet - Shared development workflow for teams and their coding agentsHacker News
- NASA-IBM 月球基础模型与微调代码NASA-IBM Lunar Foundation Model and fine-tuning codeHacker News
- 认识 AI:学习指南Discover AI – A Learning GuideHacker News
- Show HN:Carve,可选择模型的开源 Mac 智能体Show HN: Carve – An open-source Mac agent with a choice of modelsHacker News
- Show HN:Claude Remote,用手机在自己的电脑上启动 Claude CodeShow HN: Claude Remote – start Claude Code on your own PC from your phoneHacker News
- 编程智能体的图执行器A graph runner for coding agentsHacker News
- Show HN:让智能体互发私信并动态自组织Show HN: Let your agents DM and dynamically self-organizeHacker News
- Show HN:Browsentic,适用于任意智能体命令行的开源 Claude for Chrome 替代方案Show HN: Browsentic Open-source Claude for Chrome alternative for any agent CLIHacker News
- OntoPrune:裁剪85%的大模型上下文 token,CPU首token时间提升6.7倍OntoPrune – Pruning 85% LLM context tokens and 6.7x TTFT on CPUHacker News
- 在 CPU 上自托管 AI 文本检测器,过滤低质内容Self-host your own AI text detector on CPU to filter out slopHacker News
- Show HN:Lazychat,运行 Claude Code 和 Codex 的 lazygit 风格终端界面Show HN: Lazychat – A lazygit-style TUI for running Claude Code and CodexHacker News
- Hindsight:AI 智能体的开源记忆系统Hindsight – An open-source memory system for AI agentsHacker News
- Show HN:AcruxCore,为智能体的提示词与工具管理版本和追踪记录Show HN: AcruxCore – version and trace the prompts and tools your agents useHacker News
- Tickler:Claude Code 编写并恢复执行的提醒Tickler: Reminders that Claude Code writes and resumesHacker News
- Show HN:Gutsy,CPU 上不到一秒的决策模型推理,Jev 风格、无 APIShow HN: Gutsy, subsecond decision model inference on CPU (Jev Style), no APIHacker News
- ZQK:用 Go 构建的自主智能体集群开源核心底层ZQK – An open-core substrate for autonomous agent swarms in GoHacker News
- Tmux-agent-status:安静、键盘优先的编程智能体控制Tmux-agent-status – Quiet, keyboard-first control of coding agentsHacker News
- Show HN:Recursant,为混合工作负载构建的智能体感知模型路由器Show HN: Recursant, the agent-aware model router built for hybrid workloadsHacker News
- Show HN:SRA-RiskGate-4B,用于稳定币风险门控和争议处理的本地大模型Show HN: SRA-RiskGate-4B – Local LLM for stablecoin risk gating and disputesHacker News
- Show HN:Moching,内置219个工具、用 Rust 编写的 AI 桌面智能体Show HN: Moching – AI desktop agent with 219 built-in tools (Rust)Hacker News
- L402 边缘缓存:面向通过闪电网络付费的 AI 智能体L402 edge cache for AI agents paying over LightningHacker News
- 版本发布 deepseek-ai/deepseek-harness release dsh-v0.2.1-alpha.1: v0.2.1-alpha.1deepseek-ai/deepseek-harness release dsh-v0.2.1-alpha.1: v0.2.1-alpha.1DeepSeek
- 开源项目 tester-army/e2etester-army/e2eGitHub Trending(日榜 + Python 日榜)
- 开源项目 earthtojake/text-to-cadearthtojake/text-to-cadGitHub Trending(日榜 + Python 日榜)
- 开源项目 pingdotgg/t3codepingdotgg/t3codeGitHub Trending(日榜 + Python 日榜)
- 开源项目 boykopovar/AnyPS5boykopovar/AnyPS5GitHub Trending(日榜 + Python 日榜)
- 开源项目 Panniantong/Agent-ReachPanniantong/Agent-ReachGitHub Trending(日榜 + Python 日榜)
- 开源项目 caddyserver/caddycaddyserver/caddyGitHub Trending(日榜 + Python 日榜)
- 开源项目 cloudflare/cloudflare-oscloudflare/cloudflare-osGitHub Trending(日榜 + Python 日榜)
- 开源项目 Stremio/stremio-webStremio/stremio-webGitHub Trending(日榜 + Python 日榜)
- 开源项目 M-Abozaid/esp32-c3-adblockM-Abozaid/esp32-c3-adblockGitHub Trending(日榜 + Python 日榜)
- 开源项目 dortania/OpenCore-Legacy-Patcherdortania/OpenCore-Legacy-PatcherGitHub Trending(日榜 + Python 日榜)
- 开源项目 achillean/shodan-pythonachillean/shodan-pythonGitHub Trending(日榜 + Python 日榜)
- 开源项目 p-e-w/hereticp-e-w/hereticGitHub Trending(日榜 + Python 日榜)
- 开源项目 VictorTaelin/OptMemVictorTaelin/OptMemGitHub Trending(日榜 + Python 日榜)
- 开源项目 meituan-longcat/LongCat-Videomeituan-longcat/LongCat-VideoGitHub Trending(日榜 + Python 日榜)
- 开源项目 jamwithai/production-agentic-rag-coursejamwithai/production-agentic-rag-courseGitHub Trending(日榜 + Python 日榜)
- 开源项目 Alishahryar1/free-claude-codeAlishahryar1/free-claude-codeGitHub Trending(日榜 + Python 日榜)
- 开源项目 666ghj/MiroFish666ghj/MiroFishGitHub Trending(日榜 + Python 日榜)
- 开源项目 shiyu-coder/Kronosshiyu-coder/KronosGitHub Trending(日榜 + Python 日榜)