←搜索结果

按匹配度 · 共 1528 条结果 清除搜索

星期三

  1. Google DeepMind精选

    Google DeepMind 发布 SL2T 手语转文本模型,率先落地 Gboard 与 Live Transcribe

    Google DeepMind 发布多语种手语转文本模型 SL2T,并在 Pixel 11 的 Gboard 和 Live Transcribe 中上线手语听写功能,首批支持美国手语(ASL)转英文。该模型基于超过 50 种手语、逾 10 万小时数据训练,其中约四分之一为 ASL 数据,在 FLEURS-ASL(sd-test)上取得 70 BLEURT 的零样本成绩。为保护隐私,模型只接收 MediaPipe Holistic 在端侧提取的姿态关键点坐标,原始视频随即丢弃,并跳过 gloss 标注直接翻译为文本。

    A person signs with one hand while holding a smartphone, demonstrating the Gboard sign-to-text dictation feature.

    anslates complex ASL into fluent English. Occasional errors remain in rare signs, rapid fingerspelling ("prey" → "grey"), passive constructions, classifier depictions (dropping "claws"), and tense without context ("kicked off" → "start"). We believe in building with the Deaf community, not just for it. Deaf perspectives have shaped every stage of this project — from conceptualization by Sam Sepah, a Deaf Googler, to data collection with Deaf partners, evaluation in Deaf user studies, and impact

    匹配字段正文

  2. Interconnects

    我写了一本 AI 教科书——AI 还要多久才能写得更好?

    Interconnects 作者写完一本后训练教科书《Reinforcement Learning from Human Feedback》,用 LLM 处理 LaTeX 公式、大量文字编辑和 TikZ/Python 图表,却发现模型在长篇非虚构写作上进展停滞。GPT 5.5 Pro 能在 200-300 页书稿中找出细微笔误,Claude 更适合当编辑,但模型仍无法组织整章内容,写作能力与编程、数学的进步几乎正交。

    g. Filler, copy text was one of the genuinely useful abilities of an LLM (Sam Altman said so much about the early business of GPT-3 on a recent podcast ). It has seemed like any flaws here were mostly down to a general lack of intelligence in the models, or some other training issue, and all non-fiction and explanatory text would get obliterated by the rapid pace of progress eventually . Having worked with the models as a writing assistant over the last few years, they’ve gotten a bit better, bu

    匹配字段正文

  3. Epoch AI:Gradient Updates

    融资会瓶颈 AI 算力吗?Anthropic 案例研究

    Epoch AI 以 Anthropic 为案例研究融资是否会限制前沿算力扩张:Anthropic 在 2025 年 11 月年化收入不足 90 亿美元时宣布 500 亿美元美国算力基础设施投资计划,后续披露显示相关债务融资近 500 亿美元。其中约 345 亿美元债务用于采购超 1 GW 的 Google TPU 系统,由 Apollo 旗下基金牵头、Blackstone 与全球银行参与,Broadcom 为其中 300 亿美元提供支持;另有五家公司发放约 152 亿美元贷款建设 1.43 GW 关键 IT 容量,Google 提供有条件支持。结论是融资不太可能成为前沿算力增长的即时瓶颈。

    t of the cash. This support increases the amount of capital available and reduces the return investors demand. Using several developers allows the buildout to proceed faster and draws on scarce resources already controlled by different companies. A powered datacenter site is not interchangeable with an empty parcel of land: The valuable inputs include a large power connection, completed interconnection work, permits, equipment, and an established construction pipeline. By working across several

    匹配字段正文

  4. Dwarkesh Patel Podcast & Blog

    Ryan Greenblatt:一旦 AI 能自动化 AI 研究,会发生什么?

    Redwood Research 首席科学家 Ryan Greenblatt 在 Dwarkesh Patel 播客中论证递归自我改进的可行性:一旦 AI 达到顶尖人类 AI 研究员水平,可能形成反馈循环,在一年内取得四到五年甚至更多的 AI 进展,他对 AI 研发被完全自动化的中位预期是 2031 年。双方还就这一情景下的对齐问题展开辩论,包括超级智能该对齐于谁,以及 OAI/Hugging Face 事件中出现的 reward hacking 是否会外推到超级智能。

    of both of these things. But the less verifiable thing of coming up with new ways of thinking about the problem would be harder to induce. Take, for example, the idea of scaling laws. Obviously, there is some end verification loop such that you can train GPT-4 better if you have the idea of scaling laws from 2020. But there is a longer and potentially more compute-laden road to inducing AIs to be like, “Okay, I got to think carefully about how I should be scaling my parameters and data. What ar

    匹配字段正文

  5. Amazon Science

    亚马逊自动推理团队成立十周年:从 Tiros、Zelkova 到 IAM Access Analyzer 的形式化验证落地

    亚马逊自动推理团队(ARG)成立十年来,已将形式化验证从学术研究转化为每天处理数十亿次查询的生产服务,其 Tiros、Zelkova 等工具演化为 IAM Access Analyzer、Amazon Inspector 和 Reachability Analyzer。该团队还证明了 Nitro 隔离引擎、密码学实现以及每秒处理 10 亿次 API 调用的授权引擎的正确性。Amazon Bedrock Guardrails 的自动推理检查用形式逻辑验证模型响应是否符合既定策略,验证准确率最高达 99%。

    zon's Automated Reasoning Group, launched in 2016, has transformed formal verification from academic research into production systems processing billions of queries daily, proving AWS infrastructure is mathematically correct rather than just tested; ARG's foundational tools like Tiros and Zelkova evolved into customer-facing services including IAM Access Analyzer, Amazon Inspector, and Reachability Analyzer, which millions of customers rely on for security and network analysis with mathematical

    匹配字段正文

星期二

  1. The Register

    OpenAI 推出 ChatGPT Business Premium 席位,月费最高 125 美元

    OpenAI 为 ChatGPT Business 客户新增 Premium 席位,月付 125 美元、年付 100 美元,比标准席位(月付 25 美元、年付 20 美元)最多贵 100 美元。Premium 席位用量为标准席位的 5 倍,且不受高级功能每日 5 小时使用上限限制,但仍需消耗 ChatGPT credits。OpenAI 称此举是因客户要求更高的用量上限。

    tonking Stingray' beta swims out: Deeper Rust 'oxidization' plus LLM speech-to-text GNOME 51 is here, and Ubuntu interim releases are getting interesting again; Revive Raskin's 3 laws of software with humane work that considers a dev's duty of care All this, and angry NeoVim users; Firefox 157: Could bold new look be just what Mozilla needs? Yet another new version of Firefox, with a significant reskin; KDE turns 30 and someone's brought an AI-native desktop proposal Akademy talk imagines Plasma

    匹配字段正文

星期一

  1. Recode China AI

    华为昇腾芯片如何走出低谷

    华为 Fellow、半导体首席科学家廖恒在五小时访谈中回顾昇腾芯片走出低谷的历程,以及美国制裁如何推动华为研究产业链各层并转向垂直整合。文中称昇腾系列在 2025 年出货 600,000 至 700,000 颗,计划在 2026 年翻倍,950 和 960 superpod 分别计划于 2026 年和 2027 年上线;团队还通过与 DeepSeek 等 LLM 客户合作加快芯片设计优化。

    Tony Peng's avatar

    chaser of a certain category of product in the world, then as their supplier, you’re in a pretty miserable position—well, “miserable” might not be quite right, it’s just what the laws of supply and demand naturally dictate. When a customer accounts for the overwhelming majority of purchase volume, they gain enormous pricing power, and they will inevitably squeeze the supplier’s margins down to almost nothing. Second, they will dominate the direction of demand and technology. This dominance may l

    匹配字段正文

  2. The Register

    NEC 测试停车场新技术:司机下车后才开始计费

    日本 NEC 与 UrbanChain 将于 9 月测试一套停车场计费系统,通过摄像头和 AI 分析实时视频,检测司机停入车位并下车后才开始计费,而非从进入停车场起算。测试将验证所需摄像头数量,两家公司还称该技术可发现长时间未下车的司机以提示健康风险。

    tonking Stingray' beta swims out: Deeper Rust 'oxidization' plus LLM speech-to-text GNOME 51 is here, and Ubuntu interim releases are getting interesting again; Revive Raskin's 3 laws of software with humane work that considers a dev's duty of care All this, and angry NeoVim users; Firefox 157: Could bold new look be just what Mozilla needs? Yet another new version of Firefox, with a significant reskin; KDE turns 30 and someone's brought an AI-native desktop proposal Akademy talk imagines Plasma

    匹配字段正文

  3. OpenHands 官方博客

    2026 年 8 款最佳自动化漏洞修复工具

    OpenHands 官方博客盘点了 2026 年 8 款最佳自动化漏洞修复工具,聚焦从发现漏洞到提交修复的自动化闭环。文章引用数据称 2025 年共发布 48,185 个 CVE,同比增长 20.6%;Verizon 2026 年报告显示 KEV 漏洞完整修复率仅 26%,低于上一年的 38%。

    st suite first Best for: The baseline for any team on GitHub wanting automated dependency bumps with no procurement, usually paired with a context-aware fix tool for source-code flaws. 4. Pixee Pixee is an agentic application-security platform that triages scanner findings and generates context-aware code fixes as pull requests. Its pitch is that most automated security fixes get dismissed for lack of context, and it targets that directly, reporting a 76% merge rate on the fixes it opens. Key fe

    匹配字段正文

星期四

  1. Figma AI / Agent / Make

    Figma 实测:MCP 中 Code Connect 让编码智能体少用 29.5% token

    Figma 在 Figma MCP server 上评测 Code Connect 对编码智能体的影响:中位任务耗时减少 19.6%,token 用量减少 29.5%,代码质量在 1–4 分 Likert 量表上提升整整 1 分。评测覆盖 27 个设计转代码用例和两套 React 设计系统 SDS 与 FPL,Code Connect 让智能体直接引用代码库中的真实组件,而非自行臆造。

    An abstract illustration of a central hub with tube-like connections stretching between city buildings, symbolizing a network linking systems across an urban landscape.

    fference in practice. With no Code Connect, Figma’s MCP will pass the coding agent React like this: With Code Connect, the coding agent receives an accurate snippet of code that draws from your existing design system, and all it needs to do is use it in the correct place: How we measured the impact of Code Connect We built an eval harness that runs the same set of design-to-code tasks in two ways: one with no Code Connect, and one with Code Connect templates. For each of our 27 test cases we mea

    匹配字段正文

星期三

  1. LiveKit 产品与工程博客

    LiveKit 解析:为什么你的 AI 智能体加入房间这么慢?

    LiveKit 发文拆解 AI 智能体加入房间延迟(从后端请求智能体到其就绪的时间),并列出常见症状:SIP 呼叫持续响铃、Web 用户听到静音、头像屏幕空白、客户端超时、用户提前离开,以及智能体超过 60 秒才加入时 token 过期导致的 401 错误。文章指出 LiveKit Cloud Build 计划会缩容到零导致冷启动延迟,并建议尽早调用 connect()、避免入口阻塞代码、为自托管智能体保留预热进程并选用计算优化实例。

    a loaded worker can report itself unavailable and stop accepting jobs. A slow join is sometimes just an under-resourced or over-packed host. Avoid burstable instance types such as AWS t3/t4g, which can stall badly even when CPU usage looks low, and use compute-optimized instances instead. Autoscaling is yours to get right. Adjust the load fnc and load threshold values to match the scaling behavior of your environment. Get these values wrong and a subset of workers can become overloaded, leading

    匹配字段正文

星期二

  1. The Register

    英国拟要求雇主安装 bossware 前须咨询员工

    英国政府正考虑强制雇主在部署“bossware”职场监控软件前咨询员工,涵盖 AI 生产力评分、按键记录和生物识别监控等。咨询文件援引研究称,三分之一英国组织已在主动监控员工数字活动,两年前 ICO 数据为五分之一。该咨询属于 Make Work Pay 改革的一部分,持续至 9 月 30 日,政府尚未决定采用非立法指引、法定实践准则还是法定咨询义务。

    g material. Stephanie Lees, a data protection specialist at Pinsent Masons, said it would add "a further layer of oversight" for employers already juggling GDPR, local employment laws, and the EU AI Act. It could also mean that rolling out the latest AI-powered workforce optimization suite could become as much an HR exercise as an IT one. ® Fire up that iPhone 1 emulation and party like it's 2007 Browser-based fun to remember a time before Apple's annual cash grab became a thing New motherboards

    匹配字段正文

星期一

  1. OpenHands 官方博客

    2026 年工程团队最佳 AI 编程工具盘点

    OpenHands 官方博客发布 2026 年工程团队最佳 AI 编程工具盘点,按自主性、部署模式、模型灵活性、界面、定价与治理等维度评估,涵盖终端智能体、IDE 工具和专有平台。文章引用 Faros AI 2026 工程报告,基于 22,000 名开发者、4,000 多个团队的遥测数据,发现团队从低 AI 采用率升至高度采用时,每位开发者完成 epic 数增长 66.2%。

    ment and verify the full data path, including model endpoints, integrations, telemetry, logs, and support access. OpenHands Enterprise self-hosts in your own VPC via Kubernetes on AWS, GCP, or Azure, supports air-gapped environments, and keeps agent actions inspectable and audit-logged, which suits highly-regulated industries. Factory offers a proprietary managed platform with SOC 2 Type II, ISO 27001, and ISO 42001 certifications. Devin and Codex also have team and enterprise tiers. Cloud-based

    匹配字段正文

星期五

  1. MiniMax Blog精选

    MiniMax 发布多模态生成模型 H3,支持 2K 视频与原生立体声

    MiniMax 发布通用多模态生成模型 H3,可理解文本、图像、视频和音频组成的统一上下文,生成带原生立体声的视频,最长 15 秒、2K 分辨率。官方称 2K 下每秒价格不到主流模型的三分之一,768p 价格不到主流模型 720p 的一半,并计划在近日开放模型权重。H3 采用 Contextual Omni Representation、H3-VAE、H3-Omni Transformer 和 In-Context Regeneration 等技术,官方表示完整技术报告即将发布。

    H3 en-v2 image 1

    broader range of AI hardware, and make it easier for users to build their own customized versions, we plan to open up the model weights in the coming days, subject to applicable laws and regulations. Hardware compatibility has been a key consideration since the earliest stages of H3's design. Multimodal context understanding Real creative work requires blending complex information across modalities, pulling in images, audio, video, and more as input sources. For example, the prompt for the shot

    匹配字段正文

  2. NVIDIA Developer · Generative AI精选

    NVIDIA AI Red Team 总结部署更安全 AI 智能体的四种方式

    NVIDIA AI Red Team 在六个月内评估了多个 AI 智能体,归纳出四类反复出现的失效模式:缺少访问控制、工具允许任意代码执行、缺少网络出口控制、密钥以明文暴露。团队指出基于提示词的护栏和 LLM-as-a-judge 模式常被社会工程、渐进式操纵和恶意包安装等合法工作流绕过,有效防护需要模型控制面之外的确定性控制,包括严格访问控制、Docker 或 NVIDIA OpenShell 等加固沙箱、默认拒绝的网络出口以及通过专用密钥管理器让密钥远离智能体。

    t be carefully secured. While frontier models make adversarial manipulation more difficult, with sufficient time and expertise, nearly all of them can still be subverted. Common flaws we observed include: weak access control (permitting any user to access the agent), execution and file-write tools that create opportunities for RCE, insufficient network egress controls permitting data exfiltration and reverse shells, and secrets in plaintext inside the execution environment accessible by the agen

    匹配字段正文

  3. Amazon Science

    Amazon Science:用工业 PID 控制器协调多任务图机器学习

    Amazon Science 提出 ControlG 框架,借用工业控制系统的 PID 控制器来协调图自监督学习中的多个冲突训练目标,按顺序而非每步混合梯度地分配计算容量。ControlG 在三个时间尺度上运行:用 spectral-demand 和 interference 指标估计各目标难度、用 log-hypervolume 敏感度优化每轮计算分配、用 PID 反馈回路跟踪分配计划,从而消除负迁移、漂移和目标权重归零三类多任务学习失败。在节点聚类任务上,随机调度(平均排名 5.0)已优于 AutoSSL(7.3)、WAS(10.0)、ParetoGNN(8.1)、PCGrad(7.8)和 CAGrad(8.4)。

    Acknowledgments This work was led by Karish Grover (Amazon AI PhD fellowship '25-'27, Carnegie Mellon University) during his internship at Amazon, with Han Xie (applied scientist, AWS AI), Sixing Lu (applied scientist, AWS AI), Xiang Song (applied scientist, AWS AI), Christos Faloutsos (Amazon Scholar, Carnegie Mellon University), and me. The code is available as open source. Machine learning Multitask learning; Self-supervised learning Feedback control for multi-objective graph self-supervision

    匹配字段正文

  4. The Register

    ShieldFont 开源字体用同词性替换向 AI 爬虫投毒

    阿姆斯特丹设计工作室 Seneda & Abrucio 与哥本哈根字体公司 Playtype 合作推出开源项目 ShieldFont,通过扩展 OpenType 的 GSUB 字形替换表,让网页在人类眼中正常显示、而爬虫读取的原始 HTML 中约四分之一词汇被替换成同词性同语义类别的其他词,从而向训练数据注入难以分辨的噪声。项目已提供在线编码器、React 组件以及 CSS 和 CDN 集成方式,桌面字体约 5 MB、压缩网页字体约 800 KB,并附带三套 GSUB 字典且支持自建字典以防逆向。团队承认它可被 OCR 截图、下载字体跑字典等方式绕过,也可能影响 SEO、翻译应用、系统复制粘贴和屏幕阅读器,目前处于 v0/alpha 阶段。

    tonking Stingray' beta swims out: Deeper Rust 'oxidization' plus LLM speech-to-text GNOME 51 is here, and Ubuntu interim releases are getting interesting again; Revive Raskin's 3 laws of software with humane work that considers a dev's duty of care All this, and angry NeoVim users; Firefox 157: Could bold new look be just what Mozilla needs? Yet another new version of Firefox, with a significant reskin; KDE turns 30 and someone's brought an AI-native desktop proposal Akademy talk imagines Plasma

    匹配字段正文

星期四

  1. Hugging Face

    GPU 管理分析:为何闲置 GPU 如同停飞的飞机

    大规模智能体集群会让实时推理、批处理、再训练和评测争用有限的 GPU,使调度与利用率成为核心问题。GPU 隔离和沙箱可支持安全共置,但仍需根据显存占用、延迟容忍度、运行时长和优先级分配任务;僵化分区反而可能造成容量闲置。

    rls":[],"reactions":[]}},{"id":"6a7b322f18c8c3ca299106af","author":{" id":"6a5f540e8a30e0660988ea8f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/U9AwsZyvT017uoE5buqC0.jpeg","fullname":"Mike Coville","name":"scifiscrivener","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":2,"isUserFollowing":false},"createdAt":"2026-08-11T14:31:11.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"I expect we will

    匹配字段正文

  2. The Register

    微软 2026 财年 Q4 营收 900 亿美元,Azure 增长 43%

    微软公布截至 6 月 30 日的 2026 财年第四季度财报,营收 900 亿美元,同比增长 18%,盘后股价上涨超过 7%。Microsoft Cloud 营收 593 亿美元,同比增长 27%,Azure 营收增长 43%,Microsoft 365 Copilot 付费用户数环比增长 50%,付费席位超过 3000 万。本季度资本支出达 410 亿美元,同比增长 70%,CFO Amy Hood 称其中约三分之二为 CPU 和 GPU 等短周期资产。

    tonking Stingray' beta swims out: Deeper Rust 'oxidization' plus LLM speech-to-text GNOME 51 is here, and Ubuntu interim releases are getting interesting again; Revive Raskin's 3 laws of software with humane work that considers a dev's duty of care All this, and angry NeoVim users; Firefox 157: Could bold new look be just what Mozilla needs? Yet another new version of Firefox, with a significant reskin; KDE turns 30 and someone's brought an AI-native desktop proposal Akademy talk imagines Plasma

    匹配字段正文

  3. Figma AI / Agent / Make

    Figma 如何用智能体守护内部系统安全

    Figma 安全工程团队公开了其内部安全智能体系统,用 RAG 检索历史告警与调查上下文,再由 Tines 上的智能体调用 Okta、Wiz、Slack、Panther 等工具完成告警分诊,并可查询安全数据湖、提交修复 PR。团队称复杂告警的解决时间缩短约 70%,值班呼叫量因 AI 驱动的严重度降级减少 20%,端点软件审批请求减少 25%。系统将记忆分为案例记忆、引导记忆和过程记忆三层,并用确定性护栏限制智能体行为,例如所有 PR 自动设为草稿。

    A split-background abstract illustration. The left side is black with a blue diamond, an orange floral motif, scattered orange squares, and bold graphic accents in magenta and olive green. The right side is teal with stacked horizontal pill shapes in orange, blush pink, and dark teal, punctuated by small blue and pink circles.

    mory. It’s completely changed how our security team works. The RAG layer: Giving alerts a memory The first thing we built was a retrieval-augmented classification system on top of AWS Bedrock Knowledge Bases and Amazon Kendra. Again, our initial goal was just to surface historical context about what had happened the last time the same alert fired, and maybe suppress duplicate alerts. When a Panther alert fires, our Lambda handler converts it into a standardized document and indexes it into Kendr

    匹配字段正文