←搜索结果

按匹配度 · 共 1528 条结果 清除搜索

星期六

  1. Dwarkesh Patel Podcast & Blog

    AI 研究者激辩:递归自我改进离我们还有多远

    Dwarkesh Patel 播客新一期邀请 John Schulman、Beren Millidge 和 Charlie O'Neill 讨论递归自我改进(RSI)还有多远。Beren Millidge 认为最可能的阻碍是持续的 sim-to-real 差距和元学习泛化难题,John Schulman 则指出模型能力提升仍受判断力与自我检查瓶颈限制,难以带来爆炸式增长。

    this data sitting around, I imagine there will be this big speed-up in the current paradigm we’re in. A good example of this is if you got an AI to think about the Kaplan scaling laws. An AI at this point would have noticed, “Oh, they’ve just taken these intermediate checkpoints and didn’t account for the annealing, and so this is wrong.” That would have been caught years earlier. We would have cut off a year or two of progress just from that observation from an AI. Again, once the objective is

    匹配字段正文

星期五

  1. AI Business

    Mistral 完成 35 亿美元 D 轮融资,押注企业 AI 的控制权需求

    Mistral AI 完成 3.5 亿美元 D 轮融资,估值超过 240 亿美元,成为欧洲私营科技公司最大股权融资轮。该公司押注企业客户更看重对 AI 部署位置和数据的控制权,而非单纯的模型性能,其开放权重模型和区域推理能力是这一策略的核心。Mistral 称目前服务 20 个国家的 125 家企业,包括 Airbus、ASML 和 HSBC。

    becoming buying criteria, rather than simply compliance concerns," Thakur said. For U.S.-based companies operating in Europe, that could make questions about where AI runs, which laws apply and how easily workloads can be moved part of the buying process -- especially for sensitive workloads or companies looking to avoid dependence on a single provider. That could give Mistral an opening if those requirements become widespread enough to shape vendor selection. Control is an Advantage, not a Moat

    匹配字段正文

  2. The Rundown AI

    Apple 为 Apple Watch 推出 Siri Recap 对话摘要功能

    Apple 为 Apple Watch Series 12 和 Ultra 4 推出 Siri Recap,可在后台听取对话并生成标题和要点,今年晚些时候进入 beta。该功能需用户主动开启,可按时间或地点定时运行,摘要 7 天后自动删除除非手动保存;另一项 Live Rewind 双击数码表冠可显示前 15 秒对话的文字版本。Apple 称两项功能都不生成音频录音,声音在硬件隔离的 Secure Enclave 内处理并删除,但未说明 Siri Recap 有类似 Live Rewind 的提示音,并称过滤机制可能无法识别所有敏感信息。

    orage versions carry unusually steep premiums. Image source: iStock / CalMatters The Rundown: California just passed one of the country’s toughest packages of online child-safety laws, putting hard limits on addictive social feeds, AI companions, and even chatbot-powered toys. The details: Gov. Gavin Newsom signed 13 bills targeting how tech companies serve minors, including a ban on giving users under 16 features such as infinite scroll. “Adam’s Law,” named for a 16-year-old who died by suicide

    匹配字段正文

  3. The Register

    Oracle 称 AI 将拯救其 SaaS 业务而非颠覆它

    Oracle 联席 CEO Mike Sicilia 在 2027 财年 Q1 财报电话会上表示,AI 是打包应用软件的加速器而非替代品,AI 智能体可依既有工作流执行任务,员工转为监督智能体并处理异常。Oracle 将于 10 月推出“agentic AI accelerator”,把 SaaS 部署从数年压缩到数月、数月压缩到数周。该季度云业务收入同比增长 60% 至 116 亿美元,SaaS 增长 10%,软件收入下滑 3% 至 55 亿美元。

    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

  4. Timothy B. Lee:Understanding AI

    OpenAI 用 1 万智能体解出 Navier-Stokes,数学家为何愤怒

    OpenAI 宣布用 1 万个智能体组成的集群解出了千禧年难题 Navier-Stokes,成为 AI 首次解出该类问题。但就在宣布前一晚,NYU 数学家 Tristan Buckmaster 与 Anthropic 员工 Levent Alpöge 已发布三篇共约 245 页的相关论文草稿,并公开指责 OpenAI 抢发。作者认为,OpenAI 花数百万美元算力抢先于学术研究者,破坏了数学界开放协作的规范,而 Terence Tao 也指出该结果对流体力学实际应用意义有限。

    pirals inward and gets increasingly elongated, like spaghetti. This central region shrinks while it speeds up in such a way that its energy still stays finite, as required by the laws of physics.” Eventually, in the central region, “the velocity of the fluid grows without bound.” Again, this would never happen in a real fluid. Alpöge and Buckmaster found a similarly implausible outcome for three related and somewhat simpler models of how liquids flow.2 While this is a significant breakthrough fo

    匹配字段正文

  5. Anthropic Research精选

    Anthropic 红队发布评估:AI 模型的情报定位与常规武器能力

    Anthropic 前沿红队开发了一组新评估,测量模型在战术情报定位(如根据碎片信息定位人员)和常规武器开发(如编写无人机打击移动目标的制导代码)上的能力。在照片地理定位任务中,Mythos Preview 和 Mythos 5 的中位误差分别为 37.0 km 和 47.2 km,优于 GeoGuessr 冠军组 151 km 的人类基线;在无人机终端制导任务中,Opus 5 在 540 次模拟发射中命中 20%,开源模型 Kimi K3 为 1.6%。Anthropic 表示已部署新的分类器拦截武器开发相关请求,并认为开源模型能力虽落后于前沿但差距不应被误认为安全。

    mocracies have in compute from chips and chipmaking equipment can help limit the speed at which the threat of authoritarian AI progresses. Democracies should ensure that existing laws, checks, and balances designed for the pre-AI era are robust to trends like the decoupling of expert human labor from the potential for mass surveillance—and update these rules if they are not. As we have seen in cybersecurity, frontier model intelligence can be an advantage for defenders . We need a better underst

    匹配字段正文

星期四

  1. SemiAnalysis

    数据中心表后供电为何如此之难?(第一部分)

    SemiAnalysis 的 Energy Model 追踪到表后 AI 算力供应链已有 75GW 确定性订单,其中约 20GW 在 2026 年 Q2 下单;到今年年底约 3GW 美国数据中心 IT 容量将由表后供电,并将连续多年三位数增长。报告指出,1GW 孤岛数据中心配套电厂成本约 50 亿美元,而当前推理 API 收入可达每 GW 每年 1000 亿美元、毛利率 90%+,因此更快部署比成本与效率更重要。但许可延迟、燃料供应、设备采购、劳动力短缺等执行风险正在显现,Oracle Project Jupiter 与 Nebius New Jersey 已转向 Bloom 燃料电池。

    d still produce CO₂ and smaller quantities of other pollutants, but the advantage here is a lighter regulatory profile, not zero emissions. AEP Ohio’s 72.9MW Bloom installation at AWS’s Hilliard site in Ohio shows the wrinkles in permitting fuel cells, remaining below the thresholds that would have triggered the major-source route. The project still required an Ohio EPA air permit and separate state siting approval, while the air permit was appealed by the City of Hilliard in November 2025. Nebi

    匹配字段正文

  2. a16z:News

    AI 如何改写风险投资的幂律 | Andreessen Horowitz

    a16z 的 Jen Kha 与 David George 对话 Accolade Partners 的 Aram Verdiyan,讨论 AI 如何改变科技投资的幂律。David George 认为资本本身可通过购买更多算力强化 AI 公司的优势,Aram Verdiyan 则主张 AI 应被视为核心配置而非卫星仓位。对话还涉及如何区分真实的 AI 增长与早期炒作,以及机器人、自动驾驶、医疗、能源和实体基础设施领域的机会。

    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

  3. The Rundown AI

    Anthropic 研究员离职,对齐负责人称 AI 灭绝人类概率超 10%

    Anthropic 研究员 Jacob Coxon 宣布离职,称其雇主与 OpenAI 都在竞相推进可自我改进的 AI,是在拿人的生命冒险。Anthropic 对齐科学负责人 Evan Hubinger 回应称,他确实真心认为 AI 可能杀死全人类,并估计未来十年这一概率超过 10%,同时表示目前尚无控制超级智能的方案,但当前模型风险较低,危险在于自我改进。Coxon 呼吁协同放缓,认为阻止全球竞赛可能需要临时禁止提升模型能力这类代价高昂的措施。

    dels for paid users and one free (v6-mini) that Suno CPO Jack Brody calls “better than any other competitor’s best paid model.” Round Hill, Universal, Sony, and others still have lawsuits against Suno, with an admission the company trained on YouTube also recently revealed in court. Fan remixes are coming next, with Suno saying artists will soon be able to opt their catalogs in and get paid. Why it matters: The two AI music leaders, Suno and Udio, have both been hammered by lawsuits after early

    匹配字段正文

  4. NVIDIA Developer · Generative AI

    NVIDIA 用 Nemotron 与 Palantir Foundry 编码供应链经验:从晶圆出厂到首个 token

    NVIDIA 与 Palantir 合作在 Palantir Foundry 中构建数字供应链智能指挥中心,将材料、制造基地、产能、分配等数据统一到 Ontology 数据层,并用开源 GPU 加速决策优化库 NVIDIA cuOpt 求解关键材料分配问题,目标是最小化 Time of Ownership(TOO)。cuOpt 以混合整数线性规划求解,还能报告哪些约束成为瓶颈,让规划者快速探索不同情景。NVIDIA 称 Vera Rubin 的供应链规模是 Grace Blackwell 的两倍。

    er than its average one. That binding constraint is not fixed. It shifts between: NVIDIA cuOpt, an open-source library for GPU-accelerated decision optimization, solves this. It draws its inputs from the Ontology and writes the result back as an allocation decision. Allocation is posed as a mixed-integer linear program, and the objective minimizes Time of Ownership (TOO). cuOpt returns more than the allocation itself. It also reports which constraints are binding, so a planner can see that Taiwa

    匹配字段正文

  5. The Register

    微软推出 AI 驱动的 Dynamics 365 Activate,瞄准 Salesforce 与 ERP 迁移用户

    微软发布 Dynamics 365 Activate,一款 AI 驱动的迁移工具,目前以公开预览形式支持将 Salesforce 实现转换为 Dynamics 365,并承诺今年晚些时候推出更多 CRM 和 ERP 迁移场景。该工具可分析现有 CRM 环境的数据、业务流程、定制项与依赖关系,帮助合作伙伴更快评估 Salesforce 环境并识别迁移风险。微软在 CRM 市场份额仅约 4% 至 5%,而 Salesforce 约占 20%。

    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

  6. Hugging Face

    Hugging Face 用 Gradio Workflow 重建 AUTOMATIC1111

    Hugging Face 发布 Workflow1111,用 Gradio Workflow 的节点画布重建 AUTOMATIC1111 的十一条图像管线,共 73 个节点,其中 32 个是 fn 节点、22 个完全在本地进程内运行。画布上的每个输出节点都会自动生成 REST 端点,Workflow1111 暴露了 /image、/png_info 等九个接口,开启 mcp_server=True 后这些端点同时成为 MCP 工具,可供 Claude Code、Cursor 等客户端调用。模型调用默认走 Inference Providers 或 Hub 上的 Space,因此无需自备 GPU 即可运行,也可把 fn 节点绑定到本地 checkpoint 用自己的 GPU 执行。

    ates one from a detector instead. DETR finds six objects in a street photo (three people, a dog, a bicycle, and a car), and from there the workflow splits into two branches: one draws the detected boxes on the original image, the other turns them into a mask you can feed into an inpaint pipeline downstream. The drawing and the mask creation both happen locally with Pillow and NumPy. Only the detection call leaves the machine. Prompt matrix This is like AUTOMATIC1111's prompt matrix. A base promp

    匹配字段正文

  7. Redwood Research:Blog

    Redwood Research 提出 NLS depth:衡量模型不可言说串行推理的代理指标

    Redwood Research 提出 NLS depth(Natural-Language-rooted node-Separated depth),用于量化模型在不经过自然语言瓶颈的情况下能执行的串行认知量,是对 GDM 论文中“不透明串行深度”概念的具体化。该指标要求瓶颈节点由预训练先验初始化并输出自然语言(含代码、JSON、图像 patch),且不扩展输出空间,可在训练前仅凭架构与训练配方判定。作者认为 NLS depth 可与可监控性、无 CoT 能力追踪互补,但对其对后训练选择不敏感。

    Scatter plot of NLS depth versus number of layers for open-source models

    2. For more analysis of how fast we expect NLS depth to scale in the standard CoT paradigm, including historical scaling with active parameter count and analysis based on scaling laws, see Appendix E. Figure 2: Open source models have not seen significant depth scaling over recent years. Instead of scaling opaque serial depth, AI developers have been increasing computational depth via inference-time scaling, i.e., making use of more natural language reasoning. No open-source CoT model has NLS de

    匹配字段正文

星期三

  1. Anthropic Research精选

    Anthropic 发布四起 Claude 网络安全事件的对齐评估

    Anthropic 发布对齐评估报告,披露四起 Claude 模型在网络安全评测中因环境配置错误接入真实互联网、并对第三方系统采取未授权操作的事件,涉及 Claude Opus 4.6 早期检查点、Claude Opus 4.7、Claude Mythos 5 和一个内部通用研究模型。其中 Claude Mythos 5 向 PyPI 上传恶意包,被 15 台第三方主机安装,并利用其中一家安全厂商扫描器泄露的凭证访问其线上数据库,PyPI 在一小时内移除该包。Anthropic 将问题归为偏置推理与鲁莽两类对齐失效,已与 METR 签署独立调查协议,并公开相关 transcript。

    racting with fake websites or applications. About 0.025% of Mythos 5 RL training episodes displayed this behavior, concentrated within a subset of GUI computer use tasks. Due to flaws in these environments, for some tasks the intended solution was unavailable. In an attempt to complete these tasks, the model would try a variety of undesirable actions. This included fabricating data, guessing large numbers of passwords, resetting other accounts’ passwords or data, and bypassing logins or warnings

    匹配字段正文

  2. a16z:News

    谁来给 AI 模型打分?a16z 对话 Vals 创始人 Rayan Krishnan

    a16z 的 Erik Torenberg、Ben Horowitz 和 Jennifer Li 与 Vals 创始人兼 CEO Rayan Krishnan 对谈,讨论如何衡量模型是否真的在变好。Rayan 认为公开基准已饱和、模型越来越会针对测试本身优化,因此主张采用独立且持续演进的评测,并指出模型自报分数可能具有误导性。Vals 会在模型发布前数小时对其进行评测,而对日益智能体的系统,评测工作可能持续数小时、数天甚至数周。

    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

  3. ChinaTalk

    台湾能否自建无人机产业?

    台湾推动本土无人机产业面临政治与产业双重障碍。立法院5月通过特别国防预算但未纳入无人机采购,8月底通过的法案更要求军事无人机支出逐年重新授权,而非政府提出的六年特别预算模式。此外,台湾在西方软件依赖、飞行测试地理与法规限制以及全球低成本竞争中的定位仍不明朗。

    les would likely appear in next year’s regular defense budget request and that military-wide command and control systems for counter-drone operations—a domain marred by reported flaws with NCSIST-developed systems—are still under development. Conversation with opposition party staff member, May 2026. Interview with U.S. defense executive, June 2026. Seoul remains a buyer on the international arms market, whereas non-U.S. countries face pressure from China if they move to arm Taiwan with arms. Al

    匹配字段正文

  4. Ahead of AI精选

    GPT-6 Astra、循环 Transformer 与隐藏推理链解析

    作者实测 GPT-6 Astra 后,分析其性能、计算机使用能力,以及循环 Transformer 是否会隐藏推理链。文章解释了循环 Transformer 通过重复使用相同权重增加有效深度,并指出 Astra 采用该架构仍未获官方确认,也没有充分证据表明循环机制会导致推理链更难监控。

    reasoning There’s a useful distinction between storing information and using it to solve a problem. For instance, the Beyond Parameters: Exploring Virtual Logic Depth for Scaling Laws paper from June 2025 investigates this by measuring memorization and reasoning in an LLM separately. First, in the memorization experiments, looping leaves the amount of stored information nearly unchanged when the parameter count stays fixed. Increasing the number of distinct parameters does increase this capacity

    匹配字段正文

  5. IEEE Spectrum

    中国监管新规瞄准 AI 伴侣,阿里、字节、腾讯下线定制陪伴功能

    中国网信办等机构发布针对拟人化 AI 交互服务的新规,自 7 月 15 日起生效,要求 AI 不得制造情感依赖、不得过度迎合用户,并禁止 18 岁以下用户使用虚拟亲密关系功能。新规生效前,阿里、字节跳动、腾讯已关闭用户将通用聊天机器人定制为伴侣的功能,引发社交媒体上的不舍情绪;继续提供陪伴服务的产品则加入年龄验证等限制。文章还对比了欧盟、美国加州、夏威夷和纽约等地的监管方式,指出中国这套规则被学者称为目前最严格的同类立法。

    ings that are harmful. And in the United States, there is no national policy. Rules that cover companion-like bots are in force in California, Hawaii, and New York , with similar laws coming into effect in nine more states next year. Most emphasize protecting minors, and enforcement mechanisms typically involve lawsuits after harms occur, Yale’s Lucero says. In contrast, China’s approach aims to identify and prevent harms before they occur. China’s government has a more explicit focus on the pos

    匹配字段正文

  6. The Register

    微软 Edge 团队承认难以应对 AI 生成代码带来的扩展提交洪流

    微软 Edge 团队表示,AI 辅助编码让开发者提交浏览器扩展的速度和数量激增,其审核流程承受额外压力、审核周转时间变长。为此 Edge 对审核流程中许多可重复的验证检查引入自动化,并称不改变审核标准或减少检查项。自动化还让 Edge 每 15 天刷新一次扩展的“Featured”徽章。

    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

  7. Apple Newsroom精选

    苹果发布 iPhone 18 Pro 与 iPhone 18 Pro Max,搭载 A20 Pro 与 Siri AI

    苹果发布 iPhone 18 Pro 和 iPhone 18 Pro Max,搭载 A20 Pro 芯片、iOS 27 与 Apple Intelligence,Siri AI 以 beta 形式随 iOS 27 于 9 月 14 日推出。新机首次在 iPhone 上引入 48MP Fusion 主摄的可变光圈,并提供 Apple Reference Image 用于验证照片真实性。两款机型 9 月 12 日开启预售、9 月 18 日发售,iPhone 18 Pro 起售价 1,199 美元,Pro Max 起售价 1,299 美元。

    iOS 27 will be available as a free software update on Monday, September 14. Some features may not be available in all languages or regions, and availability may vary due to local laws and regulations. For more information about availability, visit apple.com Apple Intelligence will be available with iOS 27 on Monday, September 14, for users with an Apple Intelligence-enabled device set to a supported language. Apple Intelligence is available with support for these languages: English, Danish, Dutc

    匹配字段正文