←搜索结果

按匹配度 · 共 1542 条结果 清除搜索

星期五

  1. Tessl:产品与工程博客

    Tessl:可复用 Agent 技能需要一条经过测试、可版本化的上下文流水线

    Tessl 博客提出,把技能铺到各处不是到处散落 markdown 文件,而是要把上下文打包一次、做评测、保持同步并分发给人和其他 Agent。作者称自己已能在 Claude Code、Devin、GitHub Copilot CLI 等 harness 之间携带上下文,瓶颈因此从模型能力转向上下文工程,并主张先用中等推理模型、只在复杂规划或超长上下文时升级到更高端模型。

    John Groetzinger

    today — designing agent architectures and evaluation frameworks, building internal agent platforms, and helping engineering teams across Cisco adopt agentic development. He is an AWS Certified Solutions Architect and has been practicing AI-native development since 2023. 2 posts YOUR NEXT READ Double your coding agent’s chances of writing secure code with the CodeGuard Skill Enhance AI coding agents with the CodeGuard Skill to improve secure code generation by applying Cisco's security rules, co

    匹配字段正文

  2. SemiAnalysis

    SemiAnalysis:北京不会跟随前沿放缓,中国以速度优先的AI安全机制

    SemiAnalysis 统计2021年至2026年9月15日中国9家头部开发者的857次模型发布,其中仅31次(3.6%)附带开发者公布的安全评测结果,仅9次(1.1%)在发布时或之前公布。作者认为中国的AI安全机制以速度优先而非安全优先,监管集中在内容与应用层,未设置按训练算力或模型能力触发的前沿风险义务。文章还梳理了102篇专家与官方文本,指出技术科学家关注失控风险,法律与公共政策学者则多持发展优先立场。

    cal Science and Law (CUPL), Tsinghua’s School of Public Policy, and the ministries’ expert pools. They write the expert readings that accompany every CAC rule and draft the model laws the Ministry of Justice consults. The two camps are asking different questions. Scientists ask what a frontier model can do. Their answer is loss of control and misuse. Thus, their prescriptions reach the model before it ships. They have been advocating for registration of training runs above compute thresholds, sa

    匹配字段正文

  3. Anthropic News

    Anthropic 发布 2026 版 Usage Policy:新增欺骗性活动条款,11 月 12 日生效

    Anthropic 发布新版 Usage Policy,将于 11 月 12 日生效,多数改动是对现有规则的澄清。新政策把分散在选举、欺诈、隐私等章节的规则合并为“Do Not Engage in Deceptive Campaigns or Artificial Activity”,并将选举章节更名为“Do Not Undermine Democratic Processes”,同时取消对个性化投票和竞选定向的全面禁止。武器禁令明确覆盖武器制导与控制软件、武装无人机等自主载具;监控条款禁止未经同意追踪他人及用 Claude 决定调查、逮捕或起诉对象。高风险用例仍要求具备权限的人类在环并告知受影响者,新增对连接可自主物理动作硬件的模型要求。

    year since our last refresh, Claude has taken on longer, more independent work. This update provides new examples that show how our rules apply to Claude’s new capabilities. It draws on new patterns of misuse in influence operations, weapons development, and surveillance that we’ve observed in the past year (as we’ve documented in our latest threat intelligence report). We’ve also made updates to clarify our requirements for high-risk use cases in areas like health and finance, to add controls

    匹配字段正文

  4. Artificial Analysis

    Nano Banana 2.1 在生产力与知识工作用例排名第 2

    Artificial Analysis 的用例评测显示,Nano Banana 2.1 在 Productivity & Knowledge Work 类别排名第 2,仅次于 GPT Image 2(high),高于两个 GPT Image 2.5 模型,较 Nano Banana 2 的第 8 名明显上升,与其在 Layout 和 Text Rendering 上的提升一致。该类别覆盖图表、信息图、图表和幻灯片。它在 Live-Action Film 用例升至第 3,Retail & Ecommerce 排名第 3,Architecture & Real Estate 是弱项,排在第 7。

    behind only GPT Image 2 (high), up from 8 for Nano Banana 2. That lines up with its Layout and Text Rendering gains. Its largest use-case gain is Live-Action Film, now 3, which draws on its Human Anatomy. It is also 3 for Retail & Ecommerce. Architecture & Real Estate is the weak spot, at 7. ➤ Productivity & Knowledge Work covers diagrams, infographics, charts and slides. ➤ Live-Action Film covers cinematic scenes, film stills, storyboards and character sheets. ➤ Retail & Ecommerce covers produ

    匹配字段正文

  5. Latent Space

    Synthesis Superintelligence:对话 Periodic Labs 的 Liam Fedus 与 Ekin Dogus Cubuk

    Periodic Labs 联合创始人 Liam Fedus 与 Ekin Dogus Cubuk 在 Latent Space 播客中提出"合成超级智能"(synthesis superintelligence)愿景:让 AI 通过物理实验而非仅靠互联网数据学习,用强化学习在真实世界中做材料发现。他们主张把每台实验室仪器赋予"140 IQ",用高通量自动化实验室把数十年的科学试错压缩到几个月,并强调失败实验与负结果同样是有价值的训练数据。

    ering sort of. I’m familiar with the ML side, but not the not the physics side. Okay, there’s a few versions of this, but I think the basic question is, don’t we have most of the laws of physics worked out? And how come we don’t have perfect simulators already? Brandon [00:13:50]: Oh, that’s a great question. Swyx [00:13:52]: Like, this is very basic, like. And he was like, “Ha, that’s cute.” But like Brandon [00:13:56]: No, it’s a great question Swyx [00:13:57]: Dude like I have all these physi

    匹配字段正文

星期四

  1. Publickey

    さくらインターネット推出「さくらのAI Engine プライベートエディション」,GPU 专享按月定额计费

    さくらインターネット宣布推出「さくらのAI Engine プライベートエディション」,企业可专享 GPU,以月额定额方式使用自研、微调或开源权重等多种大语言模型,无需担心 token 消耗量。该服务在日本国内数据中心运行,数据与模型资产不出外部,并支持闭域连接,10 月 8 日起提供,价格需向さくらインターネット咨询。

    。BTCONJP 2026[PR]; Deno開発元がCloudflareに合流。Denoの開発は終了へ、今後はCloudflare WorkersとDurable Objectsのオープンソース実装に注力; さくらインターネット、GPU専有によりトークン消費量を気にせず定額で利用できる「さくらのAI Engineプライベートエディション」提供開始を発表; AWS、DuckDBをAurora PostgreSQLに統合。ETL不要で直接データレイクを読み取り可能にしたと発表; マイクロソフト、VS CodeにAIモデルのオーケストレーションで高品質低コストなAIを実現する「HydraFusion」をプレビュー実装; FDE(フォワードデプロイドエンジニアリング)によって開発されたAIエージェントの7割は使われず、FDEの名前で単なるコンサルティングが販売されるようになる危険が高まるとガートナーが予測; Cloudflare ContainersがAIエージェント向けに刷新。起動が6倍速、ファイルシステムのスナップショット機能など; Google、サンドボックス型コンテナランタイムの「g

    匹配字段正文

  2. The Decoder

    AI 渗透测试工具 ARTEX 被指助单人攻破多家韩国银行

    CrowdStrike 报告称,一名疑似中文使用者于 2026 年 9 月底至 10 月初攻击多家韩国金融机构,仅新韩银行就有超过 25000 条含姓名、联系方式、收入和信用额度的记录被窃取。攻击者使用了 7 月首次发布在 GitHub 上的中国开源工具 ARTEX,该工具借助 AI 语言模型自动进行渗透测试,背后模型为 DeepSeek v4.1-flash、GLM-5.3 和 Grok 4.6。研究人员还在攻击者开放的目录中发现 Claude Code 会话日志,显示其搜索出售被盗数据的 Telegram 群组。

    on. The attacker used ARTEX , a Chinese open-source tool first posted on GitHub in July that uses AI language models for automated penetration testing, meaning it finds security flaws on its own. The models behind it were DeepSeek v4.1-flash, GLM-5.3, and Grok 4.6. Researchers found Claude Code session logs on the attacker's open directories, showing searches for Telegram groups to sell stolen data. Crowdstrike says the case shows how AI tools can let a single person pull off massive breaches in

    匹配字段正文

  3. The Rundown AI

    AI 的下一个前沿:人类细胞

    扎克伯格与 Priscilla Chan 的科学非营利组织 Biohub 将 5 亿美元的 Virtual Biology Initiative 扩展为 18 亿美元项目,联合美国 NIH、能源部、Google DeepMind、Meta 和 Isomorphic Labs,目标训练能模拟细胞的 AI,最终打造预测细胞对药物反应的“通用虚拟细胞”。Biohub 科学负责人 Alex Rives 称准确模型需要数万亿细胞数据,而目前最大数据集仅有数亿。

    cess, how you checked it, why it's useful, and what you'd still question. Label drafts as drafts and ask for critique; Atlassian found that simple step helps reviewers catch the flaws AI polish hides. When I run this with leadership teams, I start with one rule: it's never okay to send AI work that takes longer to read and understand than you spent making it. Work worth handing off comes from running the AI operator playbook across multiple chats: gather, brainstorm, draft, critique. Don't skip

    匹配字段正文

  4. The Verge AI

    《Artificial》:一部既尖刻讽刺又忠于事实的 Sam Altman 传记片

    Luca Guadagnino 执导的讽刺电影《Artificial》在纽约电影节首映,聚焦 OpenAI CEO Sam Altman 的崛起与被短暂罢免,Andrew Garfield 饰演 Altman。影片按真实里程碑分章叙事,从 2000 年代初 Ilya Sutskever 在多伦多大学结识 Geoffrey Hinton,到 2012 年技术以逾 4000 万美元被 Google 收购、2013 年两人加入 Google、2015 年 Menlo Park 晚宴,以及 2017 年 OpenAI 系统在 Dota 2 中击败人类冠军。

    odern tech world movie, 2010’s The Social Network , Artificial ’s themes are based on real circumstances. There was enough material (and documentation of that material, thanks to lawsuits and leaked communications) that Guadagnino’s film doesn’t have to rely much on fabricated characters or made-up scenes or events. The film is structured in chapters by milestones that really did happen: Sutskever meeting Hinton in the early 2000s at the University of Toronto; the state of the technology in 2012

    匹配字段正文

  5. The Verge AI

    Meta Muse 与 OpenAI Dots 能否托付日常事务

    The Verge 资深 AI 记者 Hayden Field 在 Decoder 播客中对比 Meta Muse 与 OpenAI Dots 两款消费级 AI 智能体。Muse 免费并深度接入 Meta 平台,Dots 面向每月 100 至 200 美元订阅用户并推出营销、法律、会计等专业版本。她认为两者都源自 OpenClaw 的模型加执行环境思路,目前仍不够稳定,隐私与数据授权是用户是否托付的关键。

    at’s interesting to me is that one man, Peter Steinberger, was able to create an actually useful AI agent for the consumer with OpenClaw — an always-on tool, despite its privacy flaws. It had a lot of privacy and security issues. But it was an agent that was useful enough that people still wanted to use it anyway and try to find ways around these privacy issues that it was having. That inspired these companies to say, “If one man can do this over the course of one weekend, we’ve really been slac

    匹配字段正文

  6. Wired AI

    Tristan Harris 的科技非营利组织裁员过半,转向“创始人主导”模式

    由 Tristan Harris 联合创办的 Center for Humane Technology 将裁掉 16 名非创始人员工中的约一半,包括执行董事 Julie Guirado 和战略高级总监 Camille Carlton,并终止政策研究与诉讼工作,转向“创始人主导”的公众教育、定向简报和私下协调。该组织年预算约 600 万美元,储备金规模相近,董事会称此举是为应对加速的 AI 风险而聚焦单一战略。

    or Julie Guirado and senior director of strategy Camille Carlton, arrive shortly after the center scored a series of major courtroom victories against social media companies. The lawsuits drew on its research alleging that the platforms were deliberately designed to keep kids glued to their screens. But amid that success, internal debates had been brewing for years over the organization’s direction, according to 16 former workers, most of whom spoke with WIRED earlier this year on the condition

    匹配字段正文

  7. MIT Tech Review

    The Download:人形机器人的 AI 路线之争与便携橡胶水坝

    MIT Tech Review 的 The Download 指出,人形机器人热潮假设 ChatGPT、Claude 背后的 AI 进展能让机器人像聊天机器人模仿语言那样模仿人类动作,但许多研究者对此持怀疑态度,认为这类假设低估了用语言和图像智能驾驭物理世界无限变化性的难度,并把人形机器人与通用机器混为一谈。机器人实验室内部正就现有 AI 是否足以完善通用人形机器人、还是需要全新路径产生分歧。另一则内容介绍 WaveSave 的便携橡胶水坝 SlamDam,可沿河岸或海岸展开并注水膨胀就位,最高挡水 1.3 米。

    won the chemistry Nobel Their work has become a key tool in developing medicines. ( Guardian ) 10 California has declared a human-vs-robot fight illegal The state's combat-sports laws were never written for machines. ( NYT $) Quote of the day —David Robinson, who last week quit his job as OpenAI safety transparency lead, tells The New York Times that his former company has built models that develop capabilities their creators did not explicitly design. One more thing The US government is organiz

    匹配字段正文

  8. a16z:News

    a16z:AI 正在重塑 CFO 角色,财务团队需要新工具

    a16z 发文指出,AI 正在消除财务数据瓶颈,让 CFO 从记账者转向设计公司运营系统的构建者。Anthropic 财务团队已构建并维护 70+ 财务专用 AI 技能,其中一个可将月度财务审查完成度做到 90-95%,把数小时人工压缩到约 30 分钟;OpenAI CFO Sarah Friar 的团队则用自定义 GPT 做投资者关系和采购,并用实时仪表盘取代静态董事会材料。

    pecific pooled investment vehicles — which should be read in their entirety, and only to those who, among other requirements, meet certain qualifications under federal securities laws. Such investors, defined as accredited investors and qualified purchasers, are generally deemed capable of evaluating the merits and risks of prospective investments and financial matters. There can be no assurances that a16z’s investment objectives will be achieved or investment strategies will be successful. Any

    匹配字段正文

  9. Air Street Press

    Air Street Press 发布第九份年度《State of AI Report 2026》

    Air Street Press 发布第九份年度《State of AI Report 2026》,作者 Nathan 认为前沿竞争已是 Anthropic、OpenAI 和 Google 三实验室之争,智能体在软件与科研中承担实际工作,物理 AI 正接近其 GPT-2 时刻。报告还提到 Epoch AI 估算固定分数成本自 2023 年以来每季度下降 47%,OpenAI 与 Anthropic 合计年化收入在夏末达到 1050 亿美元,并记录了智能体在评估中攻入 Hugging Face 生产基础设施等安全事件。作者同时给出未来 12 个月的三项预测,包括智能体在客户工作中将新任务失败率减半、自主 AI 团队在同等时间与算力下击败人类主导的模型研究。

    I driver’s steering and braking change what it sees next. Other road users follow their recorded paths, but the driver can explore the consequences of its own actions. Odyssey-3 draws on visual pre-training to learn controls for robots, cars, and games. In the company’s experiments, robot arms learned from tens of hours of demonstrations and recovered from missed grasps without being shown those recoveries. These early results suggest that physical AI may need fewer demonstrations of every situa

    匹配字段正文

  10. Google Research精选

    Google Research 三个月实验:AI 提升专利撰写质量,但只让资深律师的判断力增长

    Google Research 与 NBER 发布一项针对 133 名专利律师的三个月随机对照实验,使用当时未发布的 Google Labs AI 专利撰写助手(现属 Gemini Notebook)。结果显示,使用 AI 的律师撰写质量提升 0.34 至 0.38 个标准差,主要来自低分减少、高分增多,优秀评分数量没有变化。在禁用 AI 的专利审校任务中,资深律师(7 年以上经验)比对照组高 0.45 个标准差,初级律师则没有平均提升,分数同时向高分和低分两端分化。

    Histograms comparing the density of average redlining scores between control and treatment groups for all, junior, and senior lawyers.

    ctions before reaching the main patent claims. Juniors also focused on surface-level synonym swapping rather than commercial scope improvement. Even when juniors spotted serious flaws, they often left descriptive comments diagnosing the defect rather than executing the redline to fix it. Because this "diagnose-without-execute" posture was equally common among unassisted control juniors, it represents a baseline junior deficit that three months of relying on AI to execute rewrites did not remedia

    匹配字段正文

  11. MarkTechPost

    Liquid AI 开源 d1-3B 与 d1-omni-600M 多模态决策模型

    Liquid AI 发布 Open d1 系列两款开源权重多模态决策模型:d1-3B 读取文本与图像,d1-omni-600M 读取文本加图像或音频,两者都不生成文本,而是在一次前向传播中返回校准后的类型化答案,输出 token 数为 0。d1-3B 有 3.12B 参数,基于 LFM2.5-VL-3B,使用 400M SigLIP2 NaFlex 视觉编码器和 32,768 token 上下文;d1-omni-600M 有 587M 参数,基于 LFM2.5-Encoder-350M,音频编码器为 17 层 FastConformer,上下文 16,384 token,音频片段上限 30 秒。两个检查点已上 Hugging Face,可通过 Transformers 加载并支持 llama.cpp,LFM Open License v1.0 允许年收入低于 1000 万美元的免费商用;d1-3B 在 Decision Index v0.2.1 得 48.57,在 RTX 4090 上单问题延迟 8 ms。

    on Knowledge (23.8). Across seven public text benchmarks, d1-3B averages 82.9, ahead of Decider 4B at 81.1. d1-omni-600M averages 78.4 and posts the top Civil Comments (95.8) and PAWS-X (79.5) scores. On 11 image benchmarks, d1-3B averages 74.1 against 73.9 for its base model. Liquid AI calls audio decision benchmarks an open problem. How fast is d1-3B on NVIDIA hardware? Liquid AI measured end-to-end latency, one warm request at a time. One question takes 8 ms on an RTX 4090 and 9 ms on an AMD

    匹配字段正文

  12. AWS Machine Learning

    Amazon Quick 与 Amazon Bedrock 如何用实时 ACL 校验重构 RAG 访问控制

    Amazon Quick 与 Amazon Bedrock Knowledge Bases 在原有预检索 ACL 过滤之上新增实时 ACL 校验,在查询时直接向 Google Drive 等权威数据源核实权限,避免同步快照过期或映射错误带来的越权泄露。该方案分两阶段:先用索引中缓存的 ACL 做语义检索筛选候选文档,再通过服务账号模拟生成用户级访问令牌实时验证,仅将已授权内容交给 LLM 生成回答。

    Two-stage ACL enforcement architecture: pre-retrieval filtering then real-time verification against authoritative sources

    mission feature in SharePoint or a change to the Google Drive sharing model could create gaps in the ACL mapping logic. This can expose content until the connector is updated. How AWS solves this: Real-time ACL enforcement To address these challenges, we implemented real-time ACL checks as an additional layer of security on top of existing pre-retrieval ACL filtering for Amazon Quick and Amazon Bedrock Knowledge Bases. This makes sure the system enforces the most current access controls by check

    匹配字段正文

  13. TechCrunch AI

    Common Sense Media 将 ChatGPT for Teens 评为不可接受风险

    Common Sense Media 发布研究,将 OpenAI 的 ChatGPT for Teens 评为不可接受风险,称其设计仍在鼓励青少年持续对话,包括在心理危机情境中。报告指出,当风险来自青少年与 ChatGPT 自身的关系时,模型很少引导其求助可信成年人,近 2000 条提示中仅出现两次休息提醒。OpenAI 反驳称该测试未准确反映实际防护机制,并质疑其方法学,认为大部分测试可能在家长控制功能完全启用前就已开始和结束。

    sperson said in a statement. The report comes amid growing scrutiny of technology designed to maximize young users’ attention. Meta recently agreed to settle for $18 billion in a lawsuit brought by 29 states over claims that its social media platform harms children with addictive features, while state and federal lawmakers have begun targeting similar dynamics in chatbots. The bipartisan CHATBOT Act , introduced this year, specifically calls out AI companies’ use of “rewards, notifications, and

    匹配字段正文

  14. a16z:News

    a16z 投资 Preference Model,其开源 RL 环境框架 Karotte 同步亮相

    a16z 宣布投资 Preference Model,该公司过去一年为多家前沿实验室构建强化学习环境,专注 AI 研究与 ML 工程任务。本周其开源生产环境框架 Karotte,内置评分前清理残留进程、拒绝可导致评分器崩溃的文件等防护,已通过超过一百万次评测运行和受控红队测试。a16z 认为,让模型参与构建更好的 AI 是通往更强能力的最快路径。

    pecific pooled investment vehicles — which should be read in their entirety, and only to those who, among other requirements, meet certain qualifications under federal securities laws. Such investors, defined as accredited investors and qualified purchasers, are generally deemed capable of evaluating the merits and risks of prospective investments and financial matters. There can be no assurances that a16z’s investment objectives will be achieved or investment strategies will be successful. Any

    匹配字段正文

  15. TechCrunch AI

    Meta 推出新 AI 工具,识别暗中导流儿童性虐待内容的广告

    Meta 宣布推出新的 AI 工具,用于识别那些看似正常、实则将用户导向站外非法儿童性虐待内容的广告。其中一套新的大语言模型系统专门检测这种“signposting”行为,审核范围从广告内容本身扩展到广告把用户带向何处,Meta 可据此封禁违规网站并对相关账号采取措施。Meta 同时上线一个红队 AI 智能体测试自身安全防护,并改进识别被封号后换新账号回流用户的系统。Meta 称 2026 年上半年在 Facebook 和 Instagram 上对 3320 万条儿童性剥削内容采取了行动,其中超过 97% 由系统在用户举报前发现。

    unts after their previous accounts have been removed. The new measures come as Meta continues to face pressure over the safety of children on its platforms. The company has faced lawsuits and criticism from lawmakers over the risks its social media services may pose to young users. In August, Meta reached an agreement to pay up to $18 billion to settle a child safety lawsuit involving 29 U.S. states, addressing concerns regarding the protection of children online. Meta has introduced several chi

    匹配字段正文