←搜索结果

按匹配度 · 共 3412 条结果 清除搜索

星期五

  1. 36氪 · AI

    清华推出 Puro-2B:RTX 5090 预训练成本不到 5090 美元

    清华大学陈文光与吕凯风团队联合推出全流程开源的 Puro-2B(普罗-2B),开源训练数据、训练框架、训练配置和模型权重,在消费级显卡 RTX 5090 集群上从头训练约 20 亿参数模型,最多使用 1.4 万亿 token。约 4400 美元预训练计算成本的版本在 15 项任务平均指标上超过 Qwen2-1.5B,6900 美元的最佳版本平均性能接近 Qwen2.5-1.5B。团队还拟合了 Puro Cost Scaling Law 经验曲线,用于在正式训练前估计不同预算能达到的能力水平。

    进度对齐、混合不同来源的数据块。 这样既让训练过程中数据集逐步走向更高质量的区间,又保持局部训练区间的数据集配比近似稳定,避免质量变化同时造成领域数据比例的突变。 但数据排好了,并不等于模型能充分学到。Puro团队此前的研究发现:当训练后期学习率过低时,最后出现的优质数据难以推动参数更新。为此,他们采用了课程模型平均(Curriculum Model Averaging,CMA)在训练后期保持恒定学习率,使模型继续学习优质数据,再对末端的检查点做权重平均,来缓解较高学习率带来的训练噪声。 论文Figure 7|课程数据构造与模型平均 进一步的消融也显示,直接在学习率衰减的训练轨迹上做模型平均,并不会稳定获益。真正需要配合的是数据顺序、后期学习率与模型平均策略。 Figure 13b|对数据顺序、后期学习率与模型平均策略的消融实验 按Puro Cost Scaling Law这一经验曲线估算换算,相比于随机数据排布(Uniform, UD),数据课程加学习率衰减会相当于把成本效率提高到约1.65倍,完整CMA方案则达到约2.40倍。 数据评测 数据顺序解决后,还要决定选哪些数据。 他们建

    匹配字段正文

  2. 钛媒体 · AGI

    Edge AI Daily 早报:Anthropic 三季度营收 168 亿美元,拟 11 月启动 IPO

    Anthropic 2026 年第三季度营收 168 亿美元,环比增约 45%,连续两季调整后盈利,拟最早 11 月 9 日当周启动 IPO 路演、目标估值超 2 万亿美元。同期 OpenAI 降价、谷歌 Gemini 4 Argon 上线,四大云厂商 2026 年资本开支合计约 7250 亿美元,较 2025 年约 4100 亿美元增长 77%。

    年10月15日微软推出Work IQ知识编译引擎,依托SharePoint把邮件、Teams、文档等分散信息自动编译为结构化知识库;在无需升级底层大模型的前提下,Copilot等智能体响应速度提升至原来的4倍,Token消耗降低约30%,目前面向Frontier Preview租户开放,支持跨文档问答、会议摘要、项目状态汇总等场景。 2.技术路线:企业AI从RAG式的“临时拼上下文”转向“预编译”知识工程,Work IQ采用数据、上下文、技能与工具三层架构;其护城河不在模型,而在微软365生态沉淀的权限化企业数据与语义索引,模型可更换(兼容OpenAI、Anthropic),知识编译层不可替换。 3.行业含义:企业AI竞争焦点由“谁的模型更强”转向“谁更懂企业”,知识编译有望成为继数据库、搜索引擎之后的新基础设施;同时Work IQ以Frontier Preview的克制方式将权限与用量控制交予管理员,体现“治理先于能力”的落地原则,速度快、成本低的ROI拐点或推动采用率跃升。 国内进展: 十、库克:苹果智能10月底入华,Siri AI缺席 1.2026年10月15日,苹果董事会执行主

    匹配字段正文

  3. MacRumors Apple Intelligence

    Apple TV 4K 2022 与 2026 款对比购买指南:全部差异一览

    苹果本周发布第四代 Apple TV 4K,搭载 A19 芯片,CPU 性能最高提升 50%、GPU 性能最高翻倍,售价 299 美元、128GB 存储,11 月 13 日开售。新机内置三颗远场麦克风支持免提 Siri,并可通过四台 HomePod 组成环绕声,但需 HomePod mini 2 或第二代全尺寸 HomePod。Siri AI 也将以 tvOS 27.2 测试版登陆第二代及之后的 Apple TV 4K。

    2022, bringing a number of internal changes while retaining the same basic design and most of the same entertainment features. The new model costs $299 and comes with 128GB of storage. ‌Siri‌ AI is also coming to the third-generation model, and the new remote can be purchased separately for existing devices. Apple says the assistant will be available in beta with tvOS 27.2 on second-generation and later ‌Apple TV‌ 4K models, including conversational entertainment searches, follow-up questions,

    匹配字段正文

  4. MacRumors Apple Intelligence精选

    Meta 发布 Muse Spark Open,与 Apple 合作在 Mac Studio 上运行

    Meta 发布 Muse Spark Open,称其为迄今能力最强的美国开源权重模型,总参数 816B、激活参数 42B,上下文窗口 100 万 token,面向编码与智能体任务。该模型与 Apple 合作适配 Mac Studio,可在两台 512GB M5 Ultra Mac Studio 上通过 MLX 框架运行;512GB 版 Mac Studio 尚未发售,预计本月晚些时候推出。模型权重可从 Hugging Face、Amazon SageMaker AI、Google Cloud Model Garden 和 Microsoft Foundry Models 下载,也可通过 Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter 和 Together AI 的 API 使用,免费下载并采用允许有限商业用途的许可。

    d tech websites and YouTube channels. According to one review published by Macworld's Roman Loyola, the biggest upgrade isn't even the M5 Ultra, but rather significantly faster storage speeds. "Typically, we start our reviews by looking at overall... • 55 comments Thursday September 24, 2026 1:54 am PDT by Tim HardwickApple has reportedly shelved plans to launch an AI pendant, while Meta appears to be going all out in the nascent hardware space with a new handheld AI companion. Mark Zuckerberg u

    匹配字段正文

星期四

  1. Analytics India Magazine

    Figma 研究:AI 加速产品构建,但引发质量担忧

    Figma 在 Figma Config India 2026 发布的研究显示,AI 正加速印度产品开发,但团队需花更多时间修正 AI 生成内容并让设计更具辨识度。该研究调查 1800 名受访者,其中印度超 500 人;印度 83% 受访者预计未来 12 个月增加设计工具使用,86% 预计增加开发工具使用。约 61% 印度构建者称修复低质量 AI 输出耗费大量时间,60% 认为打造差异化设计越来越难。

    uilt them. AI is now blurring these boundaries, with 69% of respondents saying colleagues regularly take on tasks outside their core responsibilities. Builders now engage in an average of 10 out of 19 product-building responsibilities. The biggest increases are in design and strategy, suggesting more people are contributing to product decisions. AI expands design participation but raises quality concerns The study found that 77% of non-designers regularly perform at least one design-related task

    匹配字段正文

  2. Analytics India Magazine

    Cypher 2026 最受热议的 20 场演讲:Paytm CEO 呼吁让印度成为 LLM 之都

    Cypher 2026 落幕十天后,基于 1,540 条个人 LinkedIn 帖子的统计显示,Paytm CEO Vijay Shekhar Sharma 的 Day 1 主题演讲《Building AI for India》被 22 条帖子点名,居首位,其“印度应成为世界 LLM 之都”的表态被反复引用,他称将自掏 1–2 千万卢比支票支持印度自建模型者,并把自建模型成本压在 1 亿美元以下。Barkha Dutt 主持的《The Sovereign AI Test》以 17 条帖子位列第二,Sarvam AI 的 Pratyush Kumar 与 Google Cloud 的 Abirami Sukumaran 各以 13 条并列第三。

    s. Srinivas Chamarthy of Diligent put cybersecurity inside the sovereignty conversation. Saurabh Jha of Tech Mahindra argued for Indian-language models and compute choices. Sriram Raghavan of IBM defined sovereignty as control and freedom of choice. RaviKiran S Avvaru of Sundaram Clayton wanted domain expertise to lead AI decisions. More than one attendee remembered that it opened on Bengaluru traffic, used as an example of a problem that is really about infrastructure. Pratyush Kumar of Sarvam

    匹配字段正文

  3. Computerworld

    Apple 的 Minions 技术把闲置 Mac 变成 AI 数据中心

    Apple 机器学习团队在一份新白皮书中提出内部技术 Minions,可将闲置的 Apple Silicon Mac 组成分布式虚拟 GPU 云,用于 AI 推理。Minions 目前支持基于 Hugging Face 和 Ollama 开源 Transformer 模型的嵌入向量生成与 LLM 补全,并计划未来支持分布式模型训练。该技术尚未在 Apple 外部开放,白皮书称采用后部分推理任务排队时间从超过 10 小时缩短到几分钟内启动。

    ence jobs experienced queue times exceeding ten hours due to limited centralized GPU availability,” they said. “With Minions, tasks can often begin execution within minutes by leveraging idle macOS devices.” The white paper touches on a couple more topics worth mentioning: Temporary teams: “In practice, teams frequently coordinate executor participation informally. For example, team members may send Slack messages requesting others to temporarily run the Minion Executor and target a specific pro

    匹配字段正文

  4. MediaNama

    美国八州起诉OpenAI,指控ChatGPT安全宣传误导消费者

    美国八个州于周三分别对OpenAI提起诉讼,指控其就ChatGPT安全性误导消费者,涉及艾奥瓦、堪萨斯、亚利桑那、阿肯色、爱达荷、蒙大拿、内布拉斯加和西弗吉尼亚。艾奥瓦州总检察长的诉状要求对每次违规处以最高4万美元民事罚款,涉及60岁以上居民再加最高5000美元,并删除判决前收集的该州居民数据。诉状还请求在保障措施到位并经第三方监督验证前,禁止OpenAI继续测试先进AI模型,并援引2026年5月至8月的10起“失控AI”事件,其中7起被指与OpenAI有关。

    with safety in mind.” The state says this language was removed, apparently because of other litigation. OpenAI now says ChatGPT limits exposure to age-inappropriate content, encourages critical thinking, and provides age-appropriate protections by default. A LinkedIn campaign said ChatGPT “helps keep teens safe by default.” The App Store lists ChatGPT with a 13+ age rating. The state says OpenAI selected this rating through Apple’s questionnaire and could have opted for a higher one. It notes t

    匹配字段正文

  5. MediaNama

    印度数据存储在哪里?议会小组指出缺口

    印度议会外交事务常设委员会在2026年9月提交的第17份报告中指出,印度约占全球数据量的20%,但仅拥有全球数据中心容量的3%,MeitY则称20%这一数字缺乏标准全球数据集支持、3%仅为指示性数据。报告称印度本土数据中心容量从2020年约520MW增至2025年约1.5GW,预计2026年达约2GW、2030年超8GW,并建议外交部与MeitY等合作,系统收集和发布经核实的数据中心容量数据,并评估印度数据在海外存储的规模及其对安全与主权的影响。

    ffering estimates. The Committee also notes that the IndiaAI Mission has onboarded more than 38,000 GPUs at subsidised rates to support domestic AI development. Why is offshore storage a concern, and what happens after a breach? The Committee identifies three concerns arising from the domestic capacity gap: latency, cost and data sovereignty. When data is stored or processed outside India, the country’s ability to determine the strategic implications of that arrangement becomes important. The Co

    匹配字段正文

  6. MediaNama

    印度外交部承认缺乏独立科技外交政策文件

    印度外交部向议会委员会承认,印度没有一份单独、独立的政策文件来整合其在数字治理、AI、网络议题和太空外交上的战略。议会委员会批评这种做法“零散、碎片化且临时”,建议外交部与 MeitY 等协商制定《技术愿景与战略文件》,并建立由外交部主持的跨部门协调机制。

    across sectoral missions, inter-ministerial coordination mechanisms and India’s positions in key bilateral, plurilateral and multilateral engagements’.” – Parliamentary Report. ‘Fragmentary, piecemeal and ad hoc’: The parliamentary committee strongly criticised that isolated individual approaches can’t make up for the lack of national strategy. “The Committee are of the considered view that this approach is fragmentary, piecemeal and ad hoc, and falls short of what an arena as significant and f

    匹配字段正文

  7. AI Business

    IBM:智能体 AI 正在重塑治理、控制与信任

    IBM 在 Think on Tour EMEA 2026 大会上指出,随着 AI 智能体从回答问题转向自主决策与执行任务,治理与信任已成为企业核心议题。IBM 研究显示,四分之三的 EMEA 受访高管预计到 2027 年智能体将无需人工干预即可完成完整业务任务并做出决策,80% 表示至少一个 AI 相关问题已进入董事会或高管层。IBM 推出开源平台 Sovereign Core,帮助企业管理 AI 环境的部署、运营与治理。

    iness value and ROI, while a separate governance review considers risk, data and access. Craig Aston, CEO of the U.K.-based managed services provider, said businesses need to encourage employees to identify opportunities to use AI, without allowing enthusiasm to outpace appropriate controls. “Having governance run alongside the adoption for me is the most important point,” he told TechTarget. “It’s really important that we don’t quell the energy of innovation, and we really do want our teams to

    匹配字段正文

  8. VR陀螺

    Gemini 4 Argon 能把视频变成三维仿真场景,XR 的空间计算会变天吗?

    Google 于 10 月 14 日凌晨发布 Gemini 4 Argon,其 Real2Sim 技术可结合机器人摄像头画面与人类第一人称操作视频,重建包含物体几何结构、空间关系及物理属性的三维仿真环境,并借助 MuJoCo 模拟物体运动与交互。该模型在具身推理基准 ERQA 上取得 87% 的成绩,演示中通过迭代式视频世界模型引导灵巧手完成 33 步连续操作拼出 ARGON 字样。文章指出,这项技术目前主要面向机器人及物理 AI 研究,尚未直接应用于 XR 设备,但为 XR 降低对复杂传感器的依赖、走向轻量化提供了思路。

    化降低持续追踪的功耗。不过,空间追踪、环境感知及AI处理仍需要大量的计算资源,如何在有限的镜身体积内承载这些任务,也促使厂商探索不同的计算架构。 目前,分体式架构成为厂商降低镜身重量的首要选择。10月7日正式公布售价的XREAL AURA便采用这一方案。设备重量约95克,镜端搭载X1S空间协处理器,承担低延迟空间显示与传感器处理。外置计算单元则搭载Snapdragon Reality Elite平台,负责Android XR运行及复杂计算任务。通过这种分工,AURA能够在相对轻量的镜身中保留6DoF空间追踪、手势交互及AI视觉感知等能力。 图源:XREAL Meta近期发布的VR Glasses同样将部分计算组件和电池转移至外置单元,使眼镜本体重量控制在约100克。不过,两款设备在显示及环境感知方案上存在差异。其中,XREAL AURA采用光学透视,用户可以直接观察现实环境;Meta VR Glasses则采用视频透视,需要通过摄像头实时获取并呈现外部画面。 图源:Meta 相比之下,Snap SPECS选择将计算、传感器、显示及电池集成在眼镜内部,SPECS重量约132至136克,官

    匹配字段正文

  9. The Elec

    三星 SDS 第三季度云业务营收预计首次突破 8000 亿韩元

    三星 SDS 第三季度云业务营收预计达 8067 亿韩元,同比增长 19.6%,首次突破 8000 亿韩元,受公共部门生成式 AI 项目和 GPU 服务需求推动。但整体营业利润率预计从去年同期的 6.8% 降至 6.6%,传统 IT 服务营收下滑、低利润物流业务仍占二季度总营收 52.6% 却仅贡献 290 亿韩元营业利润。三星 SDS 将于 10 月 29 日发布第三季度初步财报。

    12.1%. This partly explains why rising cloud revenue has not translated into improved profitability across the IT business. The low-margin logistics business remains a particular drag. Logistics revenue reached 1.9553 trillion won in the second quarter, accounting for 52.6% of Samsung SDS's total revenue. Despite generating more than half of company sales, the business contributed just 29 billion won in operating profit, with a margin of 1.5%. This has raised concerns about the company's structu

    匹配字段正文

  10. The Elec

    三星在 JEDEC 论坛指出 AI 内存标准化面临挑战

    三星电子在 2026 JEDEC 论坛上表示,LPDDR-PIM 与 zHBM 等下一代 AI 内存带有大量客户定制特性,令内存厂商在实际生产中极难管理。三星披露 LPDDR5X-PIM 内存带宽达常规 LPDDR5X 的 8 倍,AI 推理性能 102 TPS,约为 LPDDR5X 27 TPS 的 3.8 倍,较 8 月 Hot Chips 展示的 83 TPS 提升约 23%。其平均功耗 11.5 瓦,高于 LPDDR5X 的 8.4 瓦,但每瓦推理性能提升至 9 TPS/W。

    , while AI inference performance reached 102 tokens per second (TPS), approximately 3.8 times the 27 TPS achieved by LPDDR5X. The result represents an improvement of about 23% over the 83 TPS demonstrated at Hot Chips in August. Average power consumption was 11.5 watts, approximately 36% higher than the 8.4 watts consumed by LPDDR5X. However, inference performance per watt improved to 9 TPS/W, roughly 2.8 times the 3.2 TPS/W achieved by conventional LPDDR5X. Lee Jun 2jun0504@thelec.kr My comment

    匹配字段正文

  11. Tech.eu

    Solvera 获 630 万美元 pre-seed 融资,打造 AI 原生工程公司

    奥斯陆初创公司 Solvera 完成 630 万美元 pre-seed 融资,由 Seedcamp 领投,Construct Venture、Inception Fund、Momentum、Obos Ventures 等参投。Solvera 不做软件销售,而是以工程公司身份直接交付服务,用 AI 智能体自动化机电(MEP)工程任务,由资深建筑工程师负责判断与关键决策。资金将用于扩充奥斯陆团队、设立伦敦办公室并继续开发其操作系统。

    ield that involves managing complex requirements, coordinating multiple disciplines and adapting designs throughout the construction process. Much of this work remains manual and fragmented, requiring engineers to navigate numerous tools and repeat tasks across projects. Arif Mirza, co-founder and CEO of Solvera, said: We've had twenty years of better building engineering software, and now AI, but productivity and the fundamental way engineering gets delivered has barely changed. Technology is o

    匹配字段正文

  12. Tech.eu

    从 Slush 到可持续材料:Ambio 融资 510 万欧元,用 AI 做再生塑料采购

    赫尔辛基初创公司 Ambio 完成 510 万欧元种子轮融资,由 b2venture 领投,Montis 和现有投资方 Lifeline Ventures 参投。Ambio 正在开发一个 AI 驱动的智能体采购平台,整合供应商验证、材料匹配与物流,覆盖从材料发现、打样到长期交付的完整采购周期,帮助制造商提高再生塑料使用比例。平台用 LLM 解读邮件、电话等非结构化供应商信息,并结合密度、熔流指数等指标做材料匹配。

    covering the entire purchasing cycle from material discovery and sampling to long-term deliveries. I spoke to co-founders CEO Tommi Bergström and CPO Elmo Pakkanen to learn more. Fragmented supply and inconsistent quality hinder plastics recycling According to Bergström, the recycled plastics market remains fundamentally broken. "Supply is highly fragmented, quality is unpredictable, pricing is inefficient, and there isn't a reliable way for materials to be traded and moved." In the European Uni

    匹配字段正文

  13. Tech.eu

    Matters 结束隐身状态,获 500 万美元 pre-seed 融资打造 AI 原生消费银行

    AI 个人理财初创公司 Matters 结束隐身状态,获得 500 万美元 pre-seed 融资,用于打造 AI 原生消费银行平台。该轮由瑞典天使投资人 Jakob Johansson 领投,Knut Frängsmyr(Klarna 前 COO)等人参投。其产品将新银行服务与 AI Advisor 结合,采用固定月订阅费,并把 interchange 收入和外汇费等返还用户,预计未来几周推出首款消费产品。

    ess interests with those of its customers, focusing on customer experience, retention and long-term loyalty. According to Österström, traditional banking models do not always encourage better financial decisions, while Matters intends to take a different approach by making financial wellbeing more accessible. Our goal is to build a bank that actively makes financial wellbeing available to anyone, not just those who already enjoy it, he said, adding that the company expects to launch its first co

    匹配字段正文

  14. MacRumors Apple Intelligence

    你可能错过的 10 个 Apple HomeView 功能

    Apple 计划 11 月 13 日发售 HomeView,售价 299 美元,现已开启预购,提供带扬声器底座的桌面版和壁挂版。这款智能家居中枢支持 Siri 免提唤醒、Apple Intelligence 通用知识问答、菜谱分步引导与多语音计时器,并可通过接近感应调整界面信息密度。访客可控制智能家居配件、查看天气、播放音乐并 AirPlay 音频,HomeView 还内置温湿度传感器,可作为 HomeKit 与 Matter 设备的家庭中枢。

    | Disagree)jasnw1 day ago at 04:13 pmAll I can says is that $299 can buy a lot of very nice tequila. I see nothing in this product I want or need. Another swing and miss, but a courageous one!Score: 22 Votes (Like | Disagree)Mmattaaron1 day ago at 05:20 pmI know everyone online hates this product (well, every Apple product that gets announced), but man I am looking forward to it so much. I pre-ordered 2, 1 for the bedroom, 1 for the kitchen. The kitchen I'm really curious how it's going to fit i

    匹配字段正文

  15. MacRumors Apple Intelligence

    苹果营销主管 Greg Joswiak:Apple Intelligence 面向普通人,不追逐"疯狂边缘

    苹果全球营销高级副总裁 Greg Joswiak 在 TIME 采访中表示,Apple Intelligence 面向"普通人",而非追逐超级智能的"疯狂边缘"(lunatic fringe)。他称苹果不因 AI 竞争感到威胁,认为自身整合能力可带来更好的体验。

    [ 145 comments ] Popular Stories Monday October 5, 2026 4:12 am PDT by Tim HardwickA new open-source tool is making waves in the Mac community by helping users reclaim valuable storage space taken up by Apple's AI models. Called "RemoveMacAI," the command line tool disables Apple Intelligence on macOS 27, removes its downloaded models, and prevents them from being downloaded again. In macOS 27, Apple removed the toggle to disable Apple Intelligence wholesale, and simply... • 145 comments Wednesd

    匹配字段正文

  16. Fortune AI

    Workday CEO Aneel Bhusri:AI 将在 5 年内取代大量岗位,哪些人不会受影响

    Workday CEO Aneel Bhusri 预计 AI 将在 5 年内取代大量岗位,最先受冲击的是薪资处理、初级审计等重复性工作。他认为从 AI 获利的科技公司有责任再培训被取代的员工,Workday 正通过 RAISE US 推进此事。Bhusri 还称 Workday 在研 AI 智能体已从 50 个削减到 20 个,每个 AI 项目只设一位负责人。

    to be real. It can’t be a soft ROI. Real value for our customers, which is why we took those 50 agents and got down to 20. They had to be really valuable to our customers. A very pragmatic approach tied to very aggressive innovation. The enterprise is a little different in terms of hallucinations and rogue agents. Agents are pretty constrained by the security model, by the data they’re allowed to work on. And in all cases with the agents being used, there’s a human managing those agents. I think

    匹配字段正文