←搜索结果

按匹配度 · 共 3407 条结果 清除搜索

星期一

  1. The New Stack

    Andrej Karpathy:只有 5,000 人见过 AI 真正的能力

    加入 Anthropic 的 OpenAI 创始成员 Andrej Karpathy 估算,约 60 亿人几乎没用过 LLM,10 至 20 亿人偶尔使用聊天机器人,约 2,000 万人专业使用前沿模型做数学和代码,仅约 5,000 人能在实验室内部接触前沿系统。Box CEO Aaron Levie 回应称这正说明 AI 扩散仍处早期,并看好“agent swarms”在企业工作流中的前景,但指出模型能力不均、规模化成本高、多智能体管理工具尚不成熟。

    .” “Capability peaks in domains that are digital, verifiable and economically valuable.” Karpathy also thinks people aren’t giving these systems enough to do. “A question with a paragraph answer barely stresses the system,” he wrote. The real gains become apparent when agents can spend hours working through a problem, testing changes and trying again. In March, Karpathy’s AutoResearch project ran 50 experiments overnight on a single GPU, found a better learning rate and committed the change to G

    匹配字段正文

  2. Computerworld

    Google Messages 的 16 个进阶 Android 消息功能

    Google Messages 新增“回收站”分区,被删除的对话可在 30 天内查看并恢复;同时支持在 RCS 会话中编辑已发送消息,以及为所有人撤回删除消息。此外,用户还能把联系人头像设为自定义通知图标,该功能需 Android 11 或更高版本。

    does exactly that. A long-standing Messages feature called, appropriately, Nudges, will periodically move conversations to the top of your messages list and emphasize ’em to encourage you to reply. If you find that to be helpful, hey, more power to ya. Keep it on! But if you find it to be little more than a nuisance, as I do, take 10 seconds to disable it: Just head back into the Messages app’s settings (by tapping your profile picture in the upper-right corner and selecting “Messages settings”

    匹配字段正文

  3. MIT Tech Review

    火星或许是地狱,但月球也好不到哪去

    NASA 在局长 Jared Isaacman 领导下公布月球基地计划,拟于 2028 年启动首次载人登月任务,2030 年代建成长期居住舱、科研站与资源开采设施,规模达数百平方英里。月球南极永久阴影陨石坑温度可低于 –410 °F,且面临辐射、月尘与昼夜各 14 天的极端温差,基地还需自给自足并依靠自主机器人建造维护。

    hese cosmic rays have oodles of kinetic energy, meaning they can punch through thin shielding (say, a metal wall) with ease. “They also have this habit where the heavier ions are fragmenting in that shielding and making secondary particles,” says George. That means if your shield is too thin, the astronauts behind it will be blasted with a surprisingly large dose of radiation. Then you have the sun. It may sustain life on Earth, but without our planet’s thick atmosphere and magnetic bubble to ge

    匹配字段正文

  4. Data Center Knowledge

    TRG DataCenters 发布美国 AI 数据中心就绪度排名:弗吉尼亚居首,小州凭电力与水资源追赶

    TRG DataCenters 2026 年 8 月对 24 个美国州做 AI 数据中心就绪度排名,弗吉尼亚以 100 分居首,俄亥俄 88.2、田纳西 83.7、佐治亚 81.8、得克萨斯 80.8 分列其后。弗吉尼亚映射 7 个大型 AI 站点、共 1,085 MW,年碳排放约 230 万吨,低于俄亥俄的 450 万吨和田纳西的 430 万吨,但约三分之一站点位于高水压力地区。南卡罗来纳仅 1 个大型站点、142 MW 却以 80.5 分排第六,得州 13 个站点中 58.8% 位于高水压力地区。

    of those facilities were located in areas of high water stress. Hinkle said he expects those conditions to influence cooling choices and project locations. “That is likely to encourage more sophisticated siting and infrastructure strategies rather than simply slowing development,” he said. “We could see greater use of approaches such as closed-loop or air-based cooling, recycled water, and locations where local water conditions are more favorable.” TRG notes that the water-stress measure describ

    匹配字段正文

  5. heise online

    智能 Exosuit 如何让疲惫的身体重新站起来

    慕尼黑工业大学(TUM)智能仿生机器人系统教席负责人 Lorenzo Masia 团队展示了一款轻量 Exosuit,可辅助使用者从坐姿站起、爬楼梯和散步,动力输出与自然动作节奏相匹配。该设备采用隐藏式电机与人工肌腱,未来有望直接缝入衣物,面向行动能力受限人群提供助力。

    h zukünftig in Kleidung einnähen. Der Entwickler verspricht Hilfe, um vom Stuhl aufzustehen oder Treppen zu steigen. Ein leichter Exosuit, in diesem Fall noch über der Kleidung getragen, gibt zusätzliche Kraft zum Aufstehen, abgestimmt auf den natürlichen Bewegungsablauf. (Bild: Andreas Heddergott/TUM) Arne Grävemeyer Wie intelligente Exosuits müde Körper wieder auf die Beine bringen Exoskelette für die schwere Arbeit; KI-gesteuert passend zur natürlichen Bewegung Manchmal braucht ein Mensch nur

    匹配字段正文

  6. The Guardian Technology

    六位家长谈如何限制孩子使用社交媒体与 AI

    《卫报》采访六位家长,多数受访者表示不再希望孩子使用社交媒体,不少人也在引导孩子远离 AI。一位在 AI 领域工作的家长称自己“变得明显更加谨慎”,担忧学校过早引入这些工具且缺乏家长参与;另一位家长则因 AI 耗水耗能而全家尽量避免使用 AI 应用。

    that supports their upbringing, and the knowledge that the very work is destroying their future – whether that be on an economic or environmental level. Do I feel the need to encourage them to use AI? Yes. Do I want to? No. On the flipside, social media is out. They can use WhatsApp to connect and build real-world relationships. But otherwise no good can come of it. – Andy, 43, London, UK The eco-conscious mom who is wary of both social media and AI The recent major lawsuits against Meta, along

    匹配字段正文

  7. heise online

    语言模型如何推动 Decision Intelligence 的临界点

    语言模型正在改变企业将数据转化为决策的方式,把分析链条从“发生了什么”扩展到预测性预测和规范性行动建议,并以自然语言直接呈现给不具备 SQL 或数据模型知识的用户。文章梳理了 LLM 进入 BI 的三条路径:BI 工具内置 LLM、原生 LLM-BI 工具和外部语言模型,并强调语义模型、数据质量与治理是基础,同时需应对模型幻觉、偏见和提示词注入风险。

    n Berichte und Dashboards, die Kennzahlen visualisieren und regelmäßiges Monitoring ermöglichen. Dieses Paradigma hat Unternehmen zuverlässig bei der Beantwortung einer zentralen Frage unterstützt: Was ist passiert? Doch Entscheider brauchen mehr. Sie wollen verstehen, warum etwas passiert ist, was als Nächstes passieren wird und welche Maßnahmen den größten Hebel bieten. Genau hier können Large Language Models ansetzen. Sie erweitern die analytische Kette um prädiktive Prognosen und präskriptiv

    匹配字段正文

  8. The Guardian Technology

    iPhone 18 Pro Max 评测:苹果最大、最重、续航最长的超级手机

    iPhone 18 Pro Max 是苹果迄今电池最大的 iPhone,轻用可撑近三天、亮屏约 7 小时,比前代多约 4 小时,重 249g、起售价 £1,299(€1,599/$1,299/A$2,299)。它搭载 A20 Pro 芯片、6.9in 120Hz OLED 屏、12GB RAM 与三颗 48MP 后置摄像头,主摄新增可变光圈,并运行 iOS 27 与新版 Siri AI。

    maller camera cut out and bigger dynamic island of the regular 18 Pro. Specifications Screen: 6.9in Super Retina XDR (120Hz OLED) (460ppi); Processor: Apple A20 Pro; RAM: 12GB; Storage: 256, 512GB or 1TB; Operating system: iOS 27; Camera: 48MP + 48MP UW + 48MP 4x, 18MP front-facing; Connectivity: 5G, wifi 7, NFC, Bluetooth 6, Thread, USB-C, Satellite, UWB and GNSS; Water resistance: IP68 (6 metres for 30 mins); Dimensions: 163.4 x 78 x 8.75mm; Weight: 249g Battery life for days The 18 Pro Max ha

    匹配字段正文

  9. MarkTechPost

    TII 发布 Falcon-ASR:1.6B 阿拉伯语语音模型,平均 WER 20.92%

    阿布扎比技术创新研究院(TII)发布 1.6B 参数的阿拉伯语语音识别模型 Falcon-ASR,在 6 个阿拉伯语测试集上平均词错误率(WER)为 20.92%,比 TII 榜单中最佳已发布结果低 2.25 个百分点。该模型以阿联酋方言为重点,同一套权重覆盖阿拉伯语、英语、法语、西班牙语和葡萄牙语,并输出词级时间戳。目前仅提供 Hugging Face 托管 demo(支持最长 60 秒、20 MB 音频),未发布权重、许可证或 API。

    Arabic speech recognition model. It focuses on the Emirati dialect and also transcribes English, French, Spanish and Portuguese. In TII’s evaluation, Falcon-ASR posts a 20.92% average word error rate (WER) across 6 Arabic test sets. That beats the best published entry in TII’s leaderboard snapshot by 2.25 points. For developers building Arabic voice products, the catch is access: today it runs only as a hosted demo. TL;DR Size: 1.6B parameters. Max audio context: Not disclosed (the demo accepts

    匹配字段正文

  10. Andrej Karpathy

    Karpathy:LLM 能力认知差距正扩大为漏斗

    Karpathy 称 LLM 能力认知差距正从“两群人各说各话”扩大为漏斗:约 60 亿人(75%)几乎没接触过 LLM,10-20 亿人(20%)只用免费版聊天机器人,约 2000 万人(0.2%)在数学与代码领域专业使用前沿模型,仅约 5000 人(0.00006%)能内部访问前沿系统。漏斗由三因素驱动:影响随问题规模与时间跨度放大、系统能力在数字化可验证领域呈锯齿状峰值、以及免费/付费/内部访问层级差异。

    nuscripts on frontier mathematics for a while. The funnel is driven by a combination of factors: 1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel. 2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, veri

    匹配字段正文

  11. OpenClaw

    OpenClaw v2026.10.1 发布

    OpenClaw v2026.10.1 发布 🦞 🔎 已安装技能发现 🧩 聊天旁的 MCP Apps 💾 完整备份至外部存储或 Cloudflare R2 💻 更简单的 Mac 设置 🛠️ 大量 bug 修复与调整 来自 402 位贡献者的 2,425 个 PR。谢谢!https://t.co/PbCnvJ2QEx

    OpenClaw v2026.10.1 is out 🦞 🔎 Installed-skill discovery 🧩 MCP Apps beside chat 💾 Full backups to external storage or Cloudflare R2 💻 Simpler Mac setup 🛠️ MANY bug fixes and tweaks 2,425 PRs from 402 contributors. Thank you! https://t.co/PbCnvJ2QEx

    匹配字段正文

星期日

  1. Recode China AI

    DeepSeek 拟融资超 120 亿美元,Manus 分拆后获 5 亿美元,可灵筹备超 10 亿美元 IPO

    中国 AI 周报汇总本周动态:DeepSeek 正融资至少 120 亿美元,腾讯参投,CNBC 消息称其考虑将规模翻倍至 150 亿美元,中文媒体则称这是千亿元人民币级融资。Manus 从 Meta 分拆后首轮融资超 5 亿美元,重启北京办公室并大举招聘,转向个人 AI 助手方向。彭博社报道快手旗下 AI 视频业务可灵已选定投行,计划 IPO 募资超 10 亿美元。此外,生数发布 Vidu Q4 Preview,支持最多 15 个图像参考和 3 个音频参考;字节将豆包 2.1 Lite 接入豆包 Work 并新增无限画布;路透报道一项研究称中国模型发布中仅 3.6% 公布了安全测试。

    Tony Peng's avatar

    more competition abroad. Shengshu released Vidu Q4 Preview, with its official announcement describing support for up to 15 image references and three audio references. Chinese coverage also reported the launch; the advertised improvements in expressiveness still deserve hands-on testing. (爱范儿/@ViduAI official); ByteDance put Doubao 2.1 Lite, a lightweight model tuned for office work, into Doubao Work alongside a new infinite-canvas feature. (雷峰网); Reflection unveiled an open AI model positioned

    匹配字段正文

  2. The Rundown AI

    微软 Copilot 负责人 Jacob Andreou 谈 AI 产品缺口

    微软 Copilot 负责人 Jacob Andreou 表示,新版 Copilot 围绕 Home、Code、Autopilot 三项能力重建,定位为“工作操作系统”。他称微软已从多个 Windows 应用中撤下 Copilot,因为高流量入口并未带来实际价值,减少入口后单用户使用量反而上升。他还指出没有单一实验室的模型能通吃所有任务,Copilot 因此推出 Auto 模式自动选模型,同时保留手动选择。

    , customize, and scale AI at work. Home: Your starting point in Copilot, with chat, Cowork, and full Office capabilities in one place; Code: Build apps, dashboards, and workflows in a safe environment by describing them in plain language, no coding experience needed; Autopilot: An always-on digital teammate with its own identity, memory, and workspace that keeps working even when you're not See new Copilot capabilities in action. That's it for today! ⭐️⭐️⭐️⭐️⭐️ Nailed it; ⭐️⭐️⭐️ Average; ⭐️ Fail

    匹配字段正文

  3. Rohan Paul

    Amazon 论文:LLM 可作研究世界模型预测训练收益

    Amazon 新论文让 LLM 充当研究世界模型,在消耗 GPU 前预测训练改动是否有效。基于 9 个场景的 2653 条真实实验记录测试,加入同场景历史记录后,5 个场景的平均排名相关性从 0.506 升至 0.774;OLMo3-100M 场景下低推理强度加记录得 0.892,最高强度无记录仅 0.648。作者建议研究智能体记录包括失败在内的每次实验,并喂给挑选下一次运行的模型。

    dicts an experiment's gain before it runs. They tested it on 2,653 real experiment records from 9 setups, from pretraining to inference. Past records from the same setup raised average ranking correlation with actual results from 0.506 to 0.774 across 5 setups. On the OLMo3-100M setup, adding records at low reasoning effort scored 0.892, while max effort without records reached 0.648. If you run research agents, log every experiment, including failures, and feed those records to whatever model p

    匹配字段正文

  4. The Verge AI

    在 Mac Studio 上试用本地 AI:Hermes 与 Qwen 的体验日记

    The Verge 编辑在 256GB 统一内存的 M5 Ultra Mac Studio 上安装开源 Hermes Agent,并选用 1250 亿参数、约 105GB 的 Qwen 3.8 Flash Next 模型,通过 Telegram bot 远程控制。他用它生成每日晨间简报、整理 400 多款 Steam 游戏库、分析财务记录,并尝试自动化笔记本跑分脚本,但坦言本地 AI 仍令人兴奋、不知所措且沮丧。

    on is a script to run a series of laptop benchmark tests. I run all our benchmarks manually, and tending to a laptop to run a dozen or so tests — three times over, to take the average — is incredibly time-consuming. I’d love to automate at least some of it. Even walking Hermes through all of the required procedures and getting it to spit out vibe coded Python scripts has been a bit of an endeavor, and it remains a work in progress. But I’m hoping to make some positive headway. It’s still early

    匹配字段正文

  5. Fortune AI

    Aurora:年费 1 万美元的邀请制 AI 礼宾服务,服务超高净值人群

    邀请制 AI 礼宾应用 Aurora 将 AI 与真人专家结合,管理会员的餐饮、旅行、奢侈品和健康服务,年费 1 万美元外加 5000 美元入会费。会员可连接日历及 Whoop、Oura 等可穿戴设备,Aurora 据此提前预判需求,例如在会员长途飞行后推荐并预订桑拿、静脉输液或运动按摩。上线 8 个月已签约数百名会员,超半数每日使用、逾 85% 每周使用,公司已融资约 1200 万美元并准备启动 A 轮。

    ties arms race,” and ultrawealthy residents now expect hotel-level amenities at home. The key, though, is that Aurora blends AI insights with the human touch. New members are encouraged to start with an onboarding call so the team can learn their preferences, lifestyle, and details as small as their partner’s favorite flowers. “You have the speed and the memory and the accuracy of technology, but then you’ve got the taste and the relationships and the expertise of humans,” Lowenstein said. ‘Surv

    匹配字段正文

  6. Fortune AI

    麦肯锡研究:AI 创造的岗位多于取代的岗位,但未来十年 1100 万美国人需要帮助

    麦肯锡研究显示,到 2035 年美国经济岗位总数将多于今天,但岗位结构变化将迫使约 1100 万美国人离开萎缩职业、跨职业转行,平均每年 77 万人,是历史水平的 3.6 倍。办公与行政支持岗位受自动化冲击最大,80% 工时可能被替代;增长岗位中 85% 要求资质认证,其中 47% 属雇主偏好而非法律强制。

    Roughly 11 million Americans may need to leave a shrinking occupation for an entirely different one. Many will find work within the occupational group they already know, but an average of 770,000 workers a year will need to cross occupational lines over the next decade, 3.6 times the historical rate. COVID-19 produced a similar surge, but it lasted a year or two. This one will last at least a decade. Consider an administrative assistant who has spent a decade keeping a department humming along.

    匹配字段正文

  7. The Decoder

    Odyssey 开放世界模型 Odyssey-3 免费研究预览

    Odyssey 推出世界模型 Odyssey-3 的公开研究预览,用户可用文本提示生成可实时探索的交互环境,开发者可申请 API 访问。免费在线演示基于 Odyssey-3 Flash,基础模型 14B 参数、生成 832×480 视频,Pro 版支持 1280×720。公司称 Odyssey-3 Pro 在 Physics-IQ Verified 视频到视频基准得 66.1 分,但该成绩来自单次测试并采用八选一筛选,未满足基准要求的四次运行规则,四次运行平均为 63.37 分。

    hmark rules require four test runs with standard deviation reported for any record claim. The reported record doesn't meet that bar. Without the selection method, Odyssey-3 Pro averaged 63.37 points across four runs. Both results appear on the official leaderboard but were submitted by Odyssey itself. On the WorldMark benchmark , Odyssey-3 ranks first in three of four categories based on its own evaluation : First-Person Stylized (77.2), Third-Person Real (79.0), and Third-Person Stylized (76.3)

    匹配字段正文

  8. Rohan Paul

    Stability AI 创始人 Emad Mostaque:资本将雇佣 GPU 而非人力

    Stability AI 创始人 Emad Mostaque 称,资本不再需要人力,会直接雇佣 GPU,因此人们需要掌握生产资料。他预言 2 年后,若你的工作只在屏幕另一端,你将成为团队中最笨、拖累他人的人,因为同事都是 AI,你的认知劳动价值为负,他们甚至会付钱让你不加入团队。

    eed to own the means of production." - Emad Mostaque, founder of Stability AI: "In 2 years' time, if your job is on the other side of a screen, you will be the dumbest person on the team that drags down everyone else because all your colleagues will be AIs. And so the value of your cognitive labor is negative. They would pay you not to be a part of the team. As of this week, we are no longer the smartest things on this planet. " --- (Full video on 'Risk Takers' Youtube channel, linke in comment)

    匹配字段正文

  9. Exponential View

    AI 与数学 2.0、巴基斯坦的太阳能对冲,以及谁在掌控算力

    OpenAI 发布包含数百份数学手稿与证明的代码库后,Terence Tao 提出“Math 2.0”愿景,主张不再以抢先解题为核心,而更重视阐释、社区建设与新方向探索;Fields 奖得主 Hugo Duminil-Copin 称其博士生和博士后“陷入全面恐慌”。作者与 WEF 发布《下一代计算框架》,建议对供应链自主性做压力测试,并先解决电力、数据或技能约束再增加算力投入。

    OpenAI. Here I tend to agree with Rohit Krishnan that there isn’t a “way to unring the bell, nor do I think we would ever go back to the old ways of doing maths.” Terence Tao is pragmatic in his vision for Math 2.0: “Math 1.0” placed a premium on being the first to solve an open problem, even if the solution was not initially well understood. Now that this goal has been optimized to the point of unsustainability, “Math 2.0” will need to decenter the role of raw problem solving and value mathema

    匹配字段正文