←搜索结果

按匹配度 · 共 3407 条结果 清除搜索

星期六

  1. Latent Space

    Latent Space 新节目 High Taste Testers 首期试播

    Latent Space 试播新节目 High Taste Testers 第一期,把旧金山午餐桌上的闲聊搬上节目,话题从 Dev Day 延伸到最新模型与助手。同期 AI 新闻汇总中,Meta 发布 816B 总参数、42B 激活的 MoE 开放权重模型 Muse Spark Open,Artificial Analysis 智能指数 39,为美国开放权重模型最高;Reflection 以 Apache 2.0 发布 501B 总参数、23B 激活的 Beam 1.0,支持 1M 上下文。

    d. Related theory: A new paper offers a theory for why meaningful neurons appear at all (@klindt david). Minecraft completed end-to-end: Four Gemini 4 Argon agents beat the Ender Dragon in 27h14m with no human help. They wrote 2,783 programs, died 605 times and coordinated through chat and shared memory (GDM, paper). Setup and cost: Agents received numeric world state rather than screen pixels, and the run cost about $520 at list API prices (summary). Agent research papers: Forced forgetting: Pe

    匹配字段正文

  2. 邵猛

    DeepSeek 发布 V4.1-Flash-1017 并开放 Search & Fetch API

    DeepSeek 发布 DeepSeek-V4.1-Flash-1017,称其在基准表现、任务完成速度和输出长度三方面较上一版本改进,已在官网、App 和 API 上线。该版本采用带自我反馈的强化学习训练,在 Terminal-Bench 4.0 得分 42.3、DeepSWE v1.1 得分 77.0,并适配 DeepSeek Harness(需 v0.2.1-rc.1 或更高)、腾讯 CodeBuddy、WorkBuddy 等工具。DeepSeek 同时把内部搜索引擎开放为 Search API(/v1/tools/search)和 Fetch API(/v1/tools/fetch),后者可将单个网页转为 Markdown。

    长周期的工作流。这与其在 Terminal-Bench 4.0(42.3)、DeepSWE v1.1(77.0)等偏重真实工程环境的基准上的表现定位一致,模型卡的对比对象也集中在 K3、GLM-5.3、GPT-5.6 等同期的 Agent/编程向模型上。 工具生态:抢占 Agent Harness 入口 为 DeepSeek Harness(需升级至 v0.2.1-rc.1 或更高)、腾讯 CodeBuddy、WorkBuddy 等做了适配。DeepSeek API 也直接兼容 Claude Code 和 Codex 的服务端工具调用。 Search & Fetch API:搜索引擎基础设施商业化 商业层面最重要的一条,DeepSeek 把支撑自身产品的内部搜索引擎开放为两个独立 API: · Search API(/v1/tools/search):返回结构化网页结果,支持 site:、lang:、after: 等检索算子,可附带页面正文。 · Fetch API(/v1/tools/fetch):抓取单个网页并转为 Markdown,这是 RAG 和 Agent 工具链的标准需求。

    匹配字段正文

  3. Testing Catalog

    10月17日AI日报:DeepSeek-V4.1-Flash与GPT-6 Astra Pro更新

    DeepSeek 发布 MIT 许可的开源权重模型 DeepSeek-V4.1-Flash-1017,在推理、编程、智能体和视觉任务上大幅提升,输出更快更短。OpenAI 为 ChatGPT 升级 GPT-6 Astra Pro,响应更快、插件与工具调用更好,并支持多个智能体并行,已向 Pro 用户开放。Claude Code Review 面向 Team 和 Enterprise 套餐正式可用,每个仓库提供 Quick、Standard、Deep 三档审查级别。

    d visual tasks plus faster, shorter outputs. SPACEXSI 🔥: - Grok Bots are now available on Grok web and mobile. - Grok Bot got much faster, finishing computer tasks 3x sooner on average and other replies in less than half the time. ANTHROPIC 🔥: - Claude Code Review is now generally available for Team and Enterprise plans, with Quick, Standard and Deep review levels per repo. NVIDIA 🔥: - NVIDIA RTX Spark Windows laptops, pairing a Blackwell GPU with a Grace CPU and up to 128GB of unified memory, a

    匹配字段正文

  4. heise online

    智能手表评测:Pebble 回归与 Pixel Watch 5 | c’t uplink

    c’t uplink 播客本期评测 Pebble 回归之作 Time 2 与 Google Pixel Watch 5。Pebble Time 2 主打按键操作,所有敏感数据仅存于手表本地、不上云也不同步到手机,续航以周或月计;Pixel Watch 5 相比 Watch 4 增加内存并改进 GPS,新增若干 AI 功能,手势控制被评价为实用,但续航最多约一天半,且充电底座历代 Pixel 型号均有变化。

    xel Watch 5 ein. Gegenüber dem Vorgänger Watch 4 hat sie mehr Speicher und ein besseres GPS. Wobei wir vermuten, dass die Verbesserungen bei der Positionierung eine reine Softwarefrage sind, die per Firmware-Update – oder sogar per Update der Companion-App auf dem Smartphone – auch für ältere Modelle umsetzbar wäre. Die Watch 5 kommt mit ein paar KI-Funktionen, über deren Praxisnutzen Alex berichtet. Hilfreich fand er vor allem die Gestensteuerung, die er ausführlich erklärt. Allerdings schafft

    匹配字段正文

  5. heise online

    开发者快讯:Postgres 19、Python 3.16 与 Symfony 8.2 有哪些新变化

    Postgres 19 发布首个 Release Candidate,主要包含一批 bug 修复,正式版计划于 11 月 19 日推出;Python 3.16 Alpha 1 已发布,iter() 新增 stop_exception 参数,Beta 阶段定于 2027 年 5 月 4 日、正式版定于 2027 年 10 月 5 日。Symfony 8.2 支持对敏感操作要求硬件安全密钥等更强登录方式,并记录用户登录所用方法,预计 2026 年 11 月发布。Eclipse Theia 1.76 让 GitHub Copilot 订阅用户可在 Theia AI 中使用当前 Copilot 模型。

    n und schließt sechs Sicherheitslücken: CVE-2026-86471, CVE-2026-90797, CVE-2026-90798, CVE-2026-90800, CVE-2026-90799 und CVE-2026-107755. Details dazu will das Team laut Blogbeitrag am 19. Oktober veröffentlichen und empfiehlt, zügig zu aktualisieren. Der Schwerpunkt von List & Label 32 liegt in der Integration von KI in alltägliche Reporting-Aufgaben von Entwicklern. Diese fordern mit dem neuen KI-Formelassistenten Berechnung oder Logik in natürlicher Sprache an und lassen sich automatisch pa

    匹配字段正文

  6. heise online

    enterJS 2027 征稿启动,KI 线上主题日 enterJS Agentic 最后报名

    enterJS 2027 将于 6 月 9 日至 10 日在科隆 Komed 举办,演讲与工作坊提案需在 11 月 3 日前提交。在此之前,10 月 22 日的线上主题日 enterJS Agentic 聚焦 JavaScript 和 TypeScript 项目中的智能体 AI,含 6 场各 45 分钟演讲,门票 299 欧元(另加 19% 增值税),11 月 6 日另有 Claude Code 全天线上工作坊,票价 549 欧元。

    Die Enterprise-JavaScript-Konferenz enterJS 2027 wird am 9. und 10. Juni im Komed in Köln stattfinden. Entwicklerinnen und Entwickler, die einen Vortrag oder Workshop halten möchten, können nun ihre Vorschläge bis zum 3. November einreichen. Doch zunächst findet am 22. Oktober der nächste Online-Thementag zur Heise-Konferenz statt: enterJS Agentic. Dort erfahren Teilnehmende, wie sie agentische KI speziell in ihren JavaScript- und TypeScript-Projekten einsetzen können. Weiterlesen nach der Anzei

    匹配字段正文

  7. SiliconANGLE

    微软发布 2 万亿参数基座模型 MAI-Base-2

    微软 AI 研究团队公布基座模型 MAI-Base-2,采用全 MoE 架构,总参数 2 万亿、每次提示激活 800 亿。微软称其在科学、编程和数学的内部基准上超过 DeepSeek-V4-Pro-Base,部分指标上硬件效率最高提升 12 倍。该模型用 16000 块 NVIDIA GB200 训练 38 天,数据集为 50 万亿 token,其中约五分之一用于训练集成的计算机视觉模块;权重将通过 Azure 以授权方式提供给可信合作伙伴。

    f industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations. LATEST STORIES ServiceNow AI Control Tower: Selling a penthouse while building the frame Reflection AI open-sources Beam’s weights, post-training recipes What to expect during

    匹配字段正文

  8. SiliconANGLE

    ServiceNow AI Control Tower 分析:治理工作流强,运行时管控仍待验证

    对 15 家 ServiceNow 大企业客户的调研显示,AI Control Tower 在治理工作流与问责上表现最强,但尚未被广泛验证为覆盖异构企业的完整技术治理与运行时执行平面。ServiceNow 称该平台支持约 1000 亿条工作流和近 8 万亿次事务,已有超过 500 家客户上线 AI Control Tower,Q2 AI ACV 超过 10 亿美元,全年目标 15 亿美元。

    bed runtime control as external, custom, human mediated, limited to selected paths or not proven at scale. What has not been validated – and appears to be a gap – is universal coverage across Azure OpenAI, custom agents, SaaS copilots, third-party models, shadow AI, multiple clouds and non-ServiceNow workflows. One security practitioner estimated that only 30 to 40 percent of total AI activity could realistically pass through ServiceNow control paths today. Our assessment is the kill-switch clai

    匹配字段正文

  9. Analytics India Magazine

    Wipro 大单签约下滑,AI 增长故事面临考验

    Wipro 2026 财年第二季度大额订单签约额环比下降 58.6% 至 6.74 亿美元,CEO Srinivas Pallia 称两笔已确认订单错过签约截止日期,将在 12 月季度完成签署。当季 IT 服务收入 26.1 亿美元,同比仅增 0.6%,12 月季度指引为环比下滑 1.5% 至增长 0.5%。Wipro 未披露 AI 业务收入,而 TCS 公布 AI 年化收入 31 亿美元、HCLTech 当季 Advanced AI 收入 2.19 亿美元。

    t AI initiatives had increased productivity equal to the output of 20,000 employees, who were redeployed within the company. Iyer said Wipro’s contracts run two to 2.5 years on average and that the deflationary cycle is about 1.5 to two years old. She expects an “inflection curve in a few quarters.” Shubham Singh Rathore, Principal Analyst at research firm Gartner, describes where client budgets are moving. “Current enterprise AI adoption reflects a strategic budget shift away from traditional t

    匹配字段正文

  10. The Guardian Technology

    英国为何应拒绝廉价而令人担忧的中国 AI

    《卫报》专栏作者 Phillip Inman 认为,英国应拒绝 DeepSeek、小米等中国开发者每月发布的“开放权重”模型,这类模型可免费获取核心软件、无需支付持续授权费。牛津大学 AI 教授 Michael Wooldridge 称,输入聊天机器人的数据可能被共享给中国官方;美国 NIST 的评估报告也在本地运行的 DeepSeek 模型中发现政治审查。作者主张英国转而更好地约束美国 AI。

    itarian, command and control state apparatus and how it could trigger social unrest, much as it did in the last century in Tiananmen Square. A property crash that still acts as a drag on the economy, rising unemployment, poor health services in many regions, a mismatch of men and women, an ageing population – they could all be the spark that threatens Communist party rule. Then came AI and the state discovered complete control. Now Beijing has huge surveillance capabilities, using AI to capture

    匹配字段正文

  11. The Guardian Technology

    AI 权力引发英国工党高层担忧,也担心政府“落后于公众

    英国工党高层正将 AI 安全视为最紧迫的担忧之一,一位内阁大臣称其担心的是高度集中的超级智能权力落入恶意行为者手中,而非“杀手机器人”。AI 大臣 Kanishka Narayan 表示恶意智能体攻击英国政府“很可能,甚至不可避免”,MI5 负责人 Ken McCallum 则警告极端组织正利用 AI 制造生物武器。商业与贸易委员会主席 Liam Byrne 认为,从 Keir Starmer 到 Andy Burnham 的政府更迭带来了姿态转变,但政府仍需在模型访问与安全监管之间做出抉择。

    But a number of senior Labour figures were privately starting to name the power of AI as the thing worrying them most, alongside the immediate pressures on the cost of living and fragile economy. One cabinet minister – with a brief unrelated to technology – told the Guardian that dangerous, out-of-control AI was the issue keeping them up at night, fretting about the threat to humanity. It was “not killer robots” causing the worry, they explained, but the idea of a huge amount of concentrated, su

    匹配字段正文

  12. The Guardian Technology

    加州最大 AI 数据中心之争:Rucci 的 Imperial 项目为何引发社区强烈反对

    律师 Sebastian Rucci 计划在加州 Imperial 县建设近 100 万平方英尺的 AI 数据中心,已投入约 500 万美元,并争取到 330 兆瓦电力。项目引发社区强烈反对,已出现六起以上诉讼、罢免官员行动和 45 天数据中心暂停令;县法官上月裁定项目须先通过全面环境审查才能推进。

    t behind the proposed site for a 950,000 sq ft data center on 12 March in Imperial. Photograph: Gina Ferazzi/Los Angeles Times/Getty Images Major projects in California take an average of seven years to get completed, said Michael Thomas, the founder of Cleanview, which tracks datacenter development. For comparison, the average in Texas is about three years. Thomas said he had seen a surge of new players, like Rucci, get into AI datacenter development nationwide. These people, he said, were will

    匹配字段正文

  13. Latent Space

    Latent Space 新节目 High Taste Testers 首期上线

    Latent Space 推出新节目 High Taste Testers 首期试播,以旧金山午餐闲聊形式讨论 Dev Day 与最新模型和助手。本期 AI 新闻覆盖 Meta 开源 816B 总参数、42B 激活的 MoE 模型 Muse Spark Open,以及 501B 总参数、23B 激活、Apache 2.0 许可的 Reflection Beam 1.0。Anthropic 的 Haiku 5.5 定价为每百万 token $0.10/$0.50,比 Haiku 4.5 便宜约 90%。

    d. Related theory: A new paper offers a theory for why meaningful neurons appear at all (@klindt david). Minecraft completed end-to-end: Four Gemini 4 Argon agents beat the Ender Dragon in 27h14m with no human help. They wrote 2,783 programs, died 605 times and coordinated through chat and shared memory (GDM, paper). Setup and cost: Agents received numeric world state rather than screen pixels, and the run cost about $520 at list API prices (summary). Agent research papers: Forced forgetting: Pe

    匹配字段正文

  14. Unite.AI

    Replit CEO Masad 称公司已不再与 Anthropic 合作

    Replit CEO Amjad Masad 在 TechCrunch Disrupt 2026 台上访谈中确认,公司已不再与 Anthropic 合作,理由涉及成本、伙伴关系和性能,并称 GPT 现在更适合 Replit 的工作负载。他提到 Replit 曾是 Anthropic 前五大客户之一,并引用内部 VBench 及公开基准称 OpenAI 的 Sol 和 Astra 模型在从零构建 vibe-coded 应用上表现更好。Masad 还表示 2026 年企业合同约占 Replit 收入的 15%,2025 年约为 2%,目前超过 50% 的收入来自 B2B。

    dels into scalable, market-ready products. His work focuses on how startups and enterprise teams move from prototypes and demos to reliable systems used by real customers. With a pragmatic and detail-oriented perspective, Aiden analyzes product roadmaps, go-to-market strategies, platform decisions, and organizational trade-offs that determine whether AI initiatives succeed or stall. He pays particular attention to deployment realities, user adoption, infrastructure constraints, and the alignment

    匹配字段正文

  15. SiliconANGLE

    Dell 谈数据保护与网络韧性如何支撑企业 AI

    Dell 在 AI Cybersecurity and Resilience Summit 预热中表示,数据保护与网络韧性已成为企业 AI 的基础控制层,其 AI Data Platform、PowerProtect、CyberSense 与自动化栈被整合为更统一的架构。Dell 产品管理副总裁 Simon Jelley 称,公司已用 Claude Mythos Preview 等前沿模型主动发现并修复产品漏洞,并通过随机化测试场景挖掘架构弱点。该峰会 Dell 旗舰专场将于 10 月 27 日播出。

    hly unpredictable test scenarios can uncover hidden architectural weaknesses, according to Jelley. “The reality is you need it to be randomized, and again, that’s where you’re leveraging those on demand environments to be able to spin up quickly — randomize the test set,” Jelley explained. “AI is going to actually be a contributor to helping there, in terms of randomizing the vulnerability. It’s definitely both a risk and a tool in terms of how we help our customers.” Dell’s approach highlights

    匹配字段正文

  16. SiliconANGLE

    Reflection AI 在 Hugging Face 开源 Beam 权重与后训练配方

    Reflection AI 将其首个大语言模型 Beam 的权重发布到 Hugging Face,同时开放后训练配方、代码模板以及用于衡量定制后能力与安全护栏变化的工具。Beam 约 5000 亿参数,官方称在部分基准上接近四倍规模模型的表现,训练历时两个月、使用数千块高端显卡,并提供标准版与多个量化版本。此次还发布了用于投机解码的 drafter 模型、多个主流 serving engine 的即用配置和 fused kernels,后续将补充基于自蒸馏的训练配方。

    f industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations. LATEST STORIES What to expect during Dell's 'Cyber Resilience' event: Join theCUBE Oct. 27 RapidCanvas raises $29M to develop custom AI applications for enterprises The agenti

    匹配字段正文

  17. AWS Machine Learning

    AWS 如何用 Amazon Bedrock AgentCore 构建 AI 智能体生产级评估框架

    AWS 销售运营团队基于 Amazon Bedrock AgentCore Evaluations 搭建了覆盖全生命周期的 AI 智能体评估框架,用于其服务全球 AWS 现场销售的多智能体系统。上线前评估 Customer Insights Agent 时发现其把 410.08 美元三个月营收报成“4.1008 亿美元”等错误,而传统监控始终显示正常;一次预生产测试中,数据、工具与提示词变更错位更让通过率从 78% 跌至 38%。

    Figure 1 Workflow diagram showing 9-stage AgentCore evaluation journey from Define to Improve

    ity of Amazon Bedrock AgentCore, is the connective tissue powering it, alongside Amazon Bedrock Guardrails (safety), Amazon CloudWatch (metrics, alarms, tracing), Amazon Simple Storage Service (Amazon S3) and Amazon Redshift (evaluation persistence), Amazon Quick (dashboards), and AWS Lambda (custom evaluators). In this post, we share how we tackled the hardest operational challenges in running AI agents at enterprise scale, and how Amazon Bedrock AgentCore Evaluations became the backbone of age

    匹配字段正文

  18. Unite.AI

    Anthropic 称 Claude Mythos 5 设计的量子电路超越已发表最优方案

    Anthropic 在 10 月 16 日的研究博客中称,Claude Mythos 5 与自动评估器构成闭环、基本无人干预,为八种计算原语设计出优于此前已发表最优方案的量子电路,部分提升达一个数量级。配套论文《Discovering state-of-the-art quantum circuits with large language models》已于 10 月 15 日提交 arXiv,共 46 页;其中七项任务同时显著降低 Toffoli 门与量子比特数,第八项给出素数域模逆的新可逆架构。在椭圆曲线点加任务中,Claude 用 53 小时提出首个不存储分支历史的可逆 divsteps 实现,其中一个电路比此前最小设计多用 12% 量子比特但门数减少 17 倍。

    ’s standard benchmarks; most of its cost comes from a single step, modular inversion, the wraparound version of division used in elliptic-curve arithmetic. Claude received a one-paragraph brief explaining the mathematical rule and how the evaluator would check and score each design, then worked for 53 hours with minimal human input, limited to the brief and occasional prompts to think more broadly, before arriving at a completely new approach, Anthropic reported. Classical computers perform modu

    匹配字段正文

  19. SiliconANGLE

    Vinci4D 用物理模型变革芯片设计

    Vinci4D 推出基于物理方程的预训练物理模型,用于加速芯片设计仿真,可同时分析热、机械等多种物理现象,替代传统有限元分析。其模型基于物理方程而非客户数据训练,减少专有设计信息暴露。Vinci4D 近期完成 2.5 亿美元 B 轮融资,估值 15 亿美元,由 Advent International、Temasek 和 Xora 联合领投,并计划从半导体扩展至汽车和航空航天等硬件行业。

    etary design information during model training. Its technology also aims to reduce the time and cost of simulations by analyzing multiple physical phenomena simultaneously. “An average customer today in the whole semiconductor ecosystem tries to solve for a certain physics,” Aravamudhan explained. “It could be thermal, it can be mechanical and it can by any[thing]. They try to solve one physics at a time because that’s the only way it has been known to be possible. That takes a whole long time.

    匹配字段正文

星期五

  1. Replit Blog

    Replit 推出 Bring Your Own Cloud,支持 Azure、Google Cloud 和 AWS

    Replit 推出 Bring Your Own Cloud(BYOC),提供 Full BYOC、BYOC for Published Apps 和 Image Exports 三种方案,支持 Microsoft Azure、Google Cloud 或 Amazon Web Services。Full BYOC 将 Replit 平台与已发布项目托管整体部署进企业自有云租户,已面向企业客户开启 private preview;BYOC for Published Apps 由 Replit 管理平台、项目托管在客户云中,处于 early access。Image Exports 则输出 Docker 镜像、Terraform 文件和 SBOM,供企业自托管 Replit 构建的应用。

    Bring Your Own Cloud to Replit — cover image

    project hosting into your cloud tenant. This lets you safely develop with core company data without any of it ever leaving your premises. The Replit platform includes compute, storage, the file system, integrations, and the Agent harness. This powers project building, connectors, scheduled agents, publishing, all knowledge work, and more. Replit’s cellular architecture groups infrastructure into isolated units called cells, supporting reliability, security, and a consistent product experience.

    匹配字段正文