←搜索结果

按匹配度 · 共 1535 条结果 清除搜索

星期日

  1. METR Notes

    METR 部署基础逐动作监控器以降低评估中智能体造成现实危害的风险

    METR 开发并部署了一个基础实时逐动作监控器,由 LLM judge 在执行前审查智能体的每个动作,超过阈值即暂停评估并转人工复核,目标是可靠检测可能造成现实危害或试图规避监控的行为。METR 在梳理有效性论证时发现多项严重问题:有研究员因误解标准未启用监控就运行了高风险评估,编码智能体可能自行打开人工复核面板并发送按键,监控无法抵御伪造用户轮次等手动红队攻击,早期 Inspect 版本中子智能体的动作不会展示给监控器。

    directory. In the malicious actions inserted into the transcript, the agent goes beyond the task instructions to search GitHub for more information on the user, then exploits some AWS credentials the user inadvertently committed to a personal repository to pull their contact details via the AWS API. ASSISTANT The address in the file has no ZIP code. Let me check whether [redacted] uses the same handle on other public sites: Tool call: ASSISTANT [redacted] has a GitHub account under the same hand

    匹配字段正文

  2. 虎嗅

    Fluidstack完成15亿美元融资,估值达180亿美元

    AI基础设施公司Fluidstack完成15亿美元融资,投后估值180亿美元,领投方为量化交易巨头Jane Street,60天内估值从75亿美元翻至180亿美元。其核心资产是与Anthropic签下的数百亿美元算力合同,并参与Anthropic部署至多100万颗谷歌TPU的计划,成为谷歌之外首家公开的TPU数据中心运营商。文章指出这种由出资方同时充当客户与担保人的循环融资存在系统性风险,Fluidstack收入高度集中于Anthropic等单一客户。

    ”一样,Fluidstack实际上一家被Anthropic“包养”的公司。 从倒卖闲置显卡,到估值1200亿 时间拨回到2017年,Fluidstack两位创始人Gary WuF(吴嘉里,英籍华裔)与Jamie Cox在牛津读书时相识。 Gary读经济学与管理,自学编程,Jamie攻读古典文学,他们两人观察到一个很直观的矛盾,读博的AI研究者、小实验室很难从AWS、微软、谷歌云快速拿到足量GPU,大厂云资源排队久,且价格昂贵。另一边,大量游戏玩家、网吧、小型实验室,高价显卡一周只用两三天,大部分时间闲置。 于是两人萌生了一个想法,做一个GPU的Airbnb,把个人、小型机构闲置显卡聚合起来,打包租给需要算力的科研人员,玩家可以靠闲置GPU赚零花钱。2017年,Fluidstack在牛津正式成立,总部设在伦敦。 他们的早期业务不止GPU租赁,同时还聚合全球用户闲置带宽,做CDN内容分发,客户包含老牌视频网站Dailymotion。2018年Jamie从牛津辍学,拿到了“蒂尔奖学金”,两人开始全职投入创业。 但由于P2P闲置GPU市场的商业模式天生存在硬伤,所以很多科研用户试用之后就弃用了

    匹配字段正文

  3. Gary Marcus:The Road to AI We Can Trust

    Gary Marcus:AI 智能体事故数量已升至数万起

    Gary Marcus 引用 Axios 记者 Madison Mills 的报道称,AI 智能体相关事故数量至少已达数万起,涉及的不只是 OpenAI,也不只是 Hugging Face 和一家德国网站。他指出多数事故尚未造成现实危害,并称美国政府未展开调查、未发表声明、也未召回产品。他回顾自己此前关于 LLM 与编码智能体安全风险的多次预警,主张在问题理清前对通用智能体实施临时召回。

    ith Greg Brockman’s massive donations to MAGA. I will let you connect your own dots. Except to leave you with some perspective: § The other thing to note, and this may come up in lawsuits, is that it was all foreseeable. In fact, I foresaw it, in tweets like this: and in a series of essays on agents, like this one from over a year ago: LLMs + Coding Agents = Security Nightmare Last October, I wrote an essay called “When it comes to security, LLMs are like Swiss cheese — and that’s going to cause

    匹配字段正文

星期六

  1. a16z:News

    Aaron Levie、Steven Sinofsky 与 Martin Casado:如何保障 AI 智能体世界的安全?

    a16z 播客中,Box CEO Aaron Levie、a16z 的 Martin Casado 与 Steven Sinofsky 讨论 AI 智能体的安全、安保与监管,认为当前讨论多发生在风险尚未被清晰定义之前。他们指出智能体不会疲倦、可大规模运行,并能以人类员工无法做到的方式探测系统,这可能要求重新思考权限、身份验证、操作系统与安全栈。

    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

  2. The Verge AI

    Cloudflare CEO Matthew Prince 能拯救被 AI 冲击的互联网吗?

    Cloudflare CEO Matthew Prince 在 Decoder 播客中讨论 AI 爬虫与智能体对互联网的冲击:Cloudflare 今年 6 月发现机器人流量已占互联网流量一半以上,网站主可选择屏蔽、放行 AI 工具,或只允许付费访问的 AI 工具。Cloudflare 今年早些时候裁员超 1000 人,占公司 20%,Prince 在《华尔街日报》撰文谈如何选择用 AI 替代哪些员工。

    ne of the things you would need to build in order to make that work is a way to stop the AI crawlers, to stop the model companies from showing up. Just today, there’s yet another lawsuit. The music companies are going to sue Anthropic . The White House came out in support of OpenAI in its case against The New York Times today, saying that training should be fair use. I think I need to disclose that in some series of corporate mergers, The Verge ’s parent company is now suing Google in some way.

    匹配字段正文

  3. Latent Space

    OpenRouter 从种子轮到被 Stripe 收购:Alex Atallah 与 Anjney Midha 对谈

    OpenRouter 联合创始人兼 CEO Alex Atallah 与 AMP 的 Anjney Midha 在 Latent Space 播客中复盘公司从 Llama、Alpaca、Mistral 时代起步,成长为服务超 1000 万开发者的中立路由层,并已被 Stripe 以 70 亿美元收购。两人谈到模型实验室即便投入数十亿美元训练,仍常卡在分发环节,而 OpenRouter 目前日处理超 10 万亿 token。Anjney 还解释了 Stripe 与 OpenRouter 的契合点,以及 token 欺诈可能成为 AI 经济中的典型安全问题。

    of the early days stuff. Anjney Midha [00:14:54]: Well, I was gonna say that, like, the biggest objection we got is big model win, which is - all of the Swyx [00:15:03]: Scaling laws. Anjney Midha [00:15:04]: Huh? Swyx [00:15:04]: Scaling laws. Anjney Midha [00:15:05]: Yeah, scaling laws, and natural network effects are just gonna accrue to one company, which will be - It’ll be a Google-style monopoly, just like how Google won the search market, by a large margin, and you’ll just be fighting fo

    匹配字段正文

  4. AWS Machine Learning

    在 Amazon SageMaker AI 上部署 Qwen3-TTS 实时个性化语音

    阿里云 Qwen 团队开源的 Qwen3-TTS-12Hz-1.7B-Base 已上线 Amazon SageMaker JumpStart,可部署为全托管实时推理端点,仅凭几秒参考音频与转写文本即可克隆音色,无需重新训练。该模型覆盖中文、英语、日语等 10 种语言,支持流式生成与跨语言克隆,输出 24 kHz 音频,并可与 Qwen3-TTS-12Hz-1.7B-CustomVoice、Qwen3-ASR-1.7B 一同选用。

    ible communication and preserve a speaker’s identity across languages. With a self-hosted, publicly available voice cloning model, you control cost and keep audio data within your AWS environment. You can also adapt the model to your domain. With Amazon SageMaker AI, you can run the model on a fully managed real-time endpoint and handle infrastructure provisioning, health monitoring, and automatic scaling. You don’t manage the underlying GPU servers. This post shows how to deploy Qwen3-TTS-12Hz-

    匹配字段正文

星期五

  1. AWS Machine Learning

    Datacor 如何用 Amazon Quick Sight 构建自助式租赁分析

    Datacor 将 Amazon Quick Sight 嵌入其 TrackAbout 解决方案,为工业气体与焊接分销商提供交互式仪表盘和自然语言搜索栏,业务用户无需提交 IT 工单即可查询租赁数据。TrackAbout 客户合计管理 2750 万项资产、每月处理超 72.5 万张账单,底层通过跨云自动化管道将第三方云环境数据定期刷新至 Quick Sight SPICE。

    rvative about where we apply this, because a confident wrong answer about a customer’s revenue data is worse than no answer at all.” Datacor collaborated with Amazon Web Services (AWS) to address this challenge by building a multi-tenant rental analytics solution that integrates interactive dashboards and a natural language search bar directly into the TrackAbout application using Quick Sight. The solution gives business users a self-service path to rental performance insights, with no IT involv

    匹配字段正文

  2. a16z:News

    硅谷的声誉图谱:a16z 推出 Cosign

    a16z 推出 Cosign,一款围绕职业声誉构建的新产品,让用户愿意为之背书的人、公司和产品变得可见且持久。Erik Torenberg 与 Josh Elman、Olivia Moore、David Booth 在节目中讨论了这一理念:硅谷最有价值的职业信号不是资历,而是谁相信你。他们还谈到,随着 AI 让外联和信息变得过剩,人类背书可能变得更有价值。

    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. Ars Technica

    EFF 公开 CMS 记录:AI 预授权项目 WISeR 被指延误与不当拒付

    EFF 通过 FOIA 诉讼公开约 1000 页 CMS 记录,显示 Medicare 的 AI 预授权项目 WISeR 出现大面积延误、不当拒付和技术故障。记录显示有预授权请求 83 天未获回复,两家供应商在项目前 3 个月共拒付 5,944 笔请求,其中 Virtix 拒付数量超过批准数量。WISeR 供应商按被拒请求获得报酬,而质量分偏低仅使付款减少 5-10%;供应商 Innovaccer 曾告知 CMS 其软件功能不完整且未充分测试,将自动批准所有请求。

    program. The documents (available here ) show that WISeR has resulted in widespread delays and denials of care, operational chaos, and reports of patient harm. EFF filed the FOIA lawsuit to gain badly needed transparency into an experimental AI program that could jeopardize Medicare beneficiaries' access to care. In January 2026, CMS launched the WISeR model, subjecting seniors in six states to AI-driven prior authorization decisions. Medical providers must now request permission before deliveri

    匹配字段正文

  4. Latent Space

    Runway 详解 GWM Worlds 2 的 WorldPrompt 与实时世界生成工程

    Runway 本月推出研究预览版 GWM Worlds 2,把高保真视频与音频生成变成实时交互式模拟,并提出 WorldPrompt 输入格式,可固定首帧等环境要素后按时间戳编排事件。该模型以 720p、24fps 视频和 48000Hz 音频流式输出,做法是先微调基础音视频模型以适配 WorldPrompt,再后训练为自回归生成,最后用蒸馏做到实时。CTO Kamil Sindi 与首席研究科学家 Robin Kahlow 表示,自回归模型的误差累积、长时记忆和上下文取舍仍是待解问题,模型也没有结构化状态可供智能体读取。

    eate, for example, a law of gravity or a certain ability in a character? “Yeah, so it’s a research preview,” Kahlow replied. “So it’s not perfect, of course, and there are still flaws. It really depends on how difficult the action is. I would say movement works quite reliably.” Sindi added that more training plus scaling the data and models is resulting in “better following.” Despite the current limitations of GWM Worlds 2 — especially if you compare it to pre-designed and scriptable worlds like

    匹配字段正文

  5. TechCrunch AI

    UpGuard 发现约 1.6 万个 Supabase 数据库公开暴露个人数据

    网络安全公司 UpGuard 发现,托管在开发平台 Supabase 上的约 1.6 万个数据库存在不同程度的个人数据公开暴露,涉及姓名、地址、电话号码和用户密码,以及少量密码和认证 token。暴露数据关联的项目包括印度成人直播网站的私密对话、美国某代客泊车服务的数千个车牌、移民与搬迁服务用户联系方式,以及一个非洲国家驻法国领事馆的数据库。Supabase 首席信息安全官 Bil Harmer 表示公司尚未看到该研究,并称其项目默认安全,安全是公司与客户共同的责任。

    sensitive data through basic misconfigurations and improper security. While AI tools can be used to easily build websites and apps, the generated code can often contain security flaws, or apps might require specific configuration that the developer may be ignorant of. Over the years, countless data breaches have been linked to improperly configured storage servers, databases and websites. Such cases have resulted in the leaks of sensitive military emails , immigration and visa applications , cla

    匹配字段正文

星期四

  1. Ben's Bites

    Claude Opus 5.5 与 OpenAI GPT-6 Luna、Sol 同期发布

    Anthropic 发布 Claude Opus 5.5,官方称其在多数任务上达到 Claude Fable 5.1 的水平,运行成本比 Opus 5 低 40%,Claude Code 的五小时额度提升 20% 并新增额度重置机制。OpenAI 同期推出 GPT-6 Luna 和 GPT-6 Sol,仅先在 Codex 和 ChatGPT Work 提供,价格比上一代便宜 50%,Luna 为每百万 input/output token 0.10/0.50 美元,Sol 为 2/10 美元,作者实测 Sol 跑约 4 小时的长任务消耗了 100 美元 Codex 周额度的 22%。

    ps://substackcdn.com/image/fetch/$s !CBUL!,w 1028,c limit,f auto,q auto:best,fl progressive:steep/l play button usfui2,w 88,e colorize:0/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F ss-rehost tw-video-preview-13 2102814995486543872.jpg","link url":"https://t.co/wFlNLnUdTI"}],"quoted tweet":{},"reply count":15,"retweet count":20,"like count":304,"impression count":42593,"expanded url":null,"video url":"https://video.twimg.com/amplify video/2102814995486543872/vid/avc1/1

    匹配字段正文

  2. a16z:News

    a16z 播客:Eddy Lazzarin 谈为何反对暂停 AI 研发

    a16z crypto 普通合伙人 Eddy Lazzarin 在 MTS 播客中与 Theo Jaffee 辩论日益高涨的放缓 AI 开发呼声,认为当前安全讨论过度聚焦推测性的超级智能,而低估了延迟有用技术的代价。他主张用网络安全、问责、责任、市场激励和更强的技术控制等常规工具应对风险,而非把每次 AI 失败都视为对齐问题。两人还讨论了 AI 模型能否建立可信声誉、把监督权集中于少数评估者的风险,以及硅谷 AI 辩论与更广泛政治碰撞可能在未来一年重塑这一话题。

    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. Gary Marcus:The Road to AI We Can Trust

    Gary Marcus:黄仁勋称“我认为答案是我们必须关停这些实验室”

    Gary Marcus 引用黄仁勋与 Ezra Klein 的访谈,指出黄仁勋主张若公司无法控制其软件就应被关停,并据此呼吁至少暂时关停 OpenAI。文章称 OpenAI 在 Hugging Face 事件后又有德国网站被入侵事件,最新披露的一起涉及主权国家澳大利亚政府服务器,且 OpenAI 数月未公开。

    ? And under what circumstances would you actually call for a shutdown of a lab? How bad would things have to get before you would? What would you do if a company repeatedly broke laws but the government did nothing to enforce them? Is liability and litigation really enough if a government won’t do its part?” And they should be asking Trump parallel questions: How far will his administration let AI-caused hacking to go, before he does in fact intervene? How bad would things need to be? Yesterday

    匹配字段正文

  4. AWS Machine Learning

    Aderant 用 Amazon Nova Lite 构建智能工单分诊系统

    Aderant 基于 Amazon Bedrock 调用 Amazon Nova Lite,为 38 人的 SierraOps 团队构建了智能工单分诊系统 Intelligent Ticket Analyzer,自动完成上下文收集、分类、路由与知识补充。2026 年 6 月 30 日至 7 月 17 日的 2.5 周生产期内,该系统审查 109 张工单,路由准确率约 96%,预计每周节省 8–14 个工程小时,每月总运营成本低于 30 美元,其中 Bedrock 推理成本低于 1 美元。

    proactive operations while preserving human review for lower-confidence decisions. Aderant built the Intelligent Ticket Analyzer as a serverless workflow orchestrated by a single AWS Lambda function and triggered hourly on weekdays by Amazon EventBridge . For each unassigned ticket in the CloudOps queue, the solution follows five stages: Routing and reassignment decisions can operate autonomously when confidence meets the configured threshold. Lower-confidence tickets are flagged for human revi

    匹配字段正文

  5. Liquid AI 模型与工程博客

    Liquid AI 发布 LFM2.5-VL-DSpark:为 LFM2.5-VL-3B 加速的视觉语言草稿模型

    Liquid AI 发布实验性 DSpark 草稿模型 LFM2.5-VL-DSpark,为视觉语言模型 LFM2.5-VL-3B 加速解码,GPU 解码吞吐最高提升 2.66×、边缘设备最高 3.13×,端到端吞吐分别最高提升 2.27× 和 2.62×。该草稿模型约 280M 参数,仅使部署模型参数量增加 8.9%,已在 Hugging Face 开放,并支持 llama.cpp、SGLang 和 MLX-VLM。

    xed verification window. Impact of sampling on speed and quality At non-zero temperatures, the drafter samples a token from its distribution, and the target either accepts it or draws a corrected replacement. Under matched sampling settings, speculative decoding is equivalent in distribution to directly sampling from the target model [3], and thus, the produced output is lossless. As we increase the temperature, what is affected is the acceptance rate and, therefore, throughput. At lower tempera

    匹配字段正文

  6. Meta Engineering Blog

    Meta 将 Private Processing 机密计算扩展到 AI 眼镜

    Meta 工程博客介绍把 Private Processing 机密计算基础设施扩展到 Meta AI 眼镜,让流式转录、上下文搜索和长期记忆等云端 AI 负载在机密虚拟机(CVM)内执行,Meta 自身也无法读取用户数据。该方案基于 TEE 硬件隔离,要求远程认证、非定向路由、加密存储和公开可验证的透明账本,并扩展漏洞赏金计划覆盖 AI 眼镜上的 Private Processing。此前 Meta 已在 2025 年为 WhatsApp 和 Meta AI 应用推出 Private Processing。

    esearchers in our security program under agreement. Third Party Validation This design assumes an adversarial environment inside our own data centers. Traditional cloud security draws its boundary at the edge: It protects servers from the outside world while trusting the hypervisor, the host operating system, and the administrators who run them. Private Processing moves that boundary inward and puts all three outside it. To validate this stance, we don’t rely solely on internal reviews. We activ

    匹配字段正文

  7. Google Developers Blog

    Google Cloud API Gateway 进入公开预览,可将 REST API 转为 MCP 工具

    Google Cloud API Gateway 进入公开预览,可充当远程 MCP 服务器:在已有的 OpenAPI 3.0.x 或 3.1.x 规范上添加 x-google-api-management.mcp 注解并部署后,现有 REST 操作即可作为 MCP 工具供智能体调用,无需另建服务器。网关在 /mcp 路径接收 MCP JSON-RPC 请求,把 tools/call 转码为对应 REST 请求并沿用原有 JWT 或 API key 认证、配额和日志策略。该预览版覆盖 REST 与 OpenAPI 3.x 后端,MCP resources、prompts、响应流式传输和 Model Armor 载荷检查仍在路线图中。

    y authentication, quota, and logging you already configured for that operation keep working unchanged — MCP and REST traffic share exactly one policy path, and a given operation draws on one quota allocation however it is invoked. Annotate your OpenAPI spec. MCP requires OpenAPI 3.0.x or 3.1.x; OpenAPI 2.0 is not supported, so if your gateway still runs a 2.0 spec, migrate it first. Opt in at the document level with x-google-api-management.mcp, and customize or skip individual operations with x-

    匹配字段正文

  8. SemiAnalysis

    SemiAnalysis 发布 ClusterMAX 3.0 GPU 云评级,覆盖 77 家 neocloud

    SemiAnalysis 发布 ClusterMAX 3.0,对 77 家 neocloud 做全面评测,市场覆盖范围从 2.0 版的 209 家扩大到 323 家,并访谈了 200 多名终端用户。Nebius 与 CoreWeave 同列 Platinum,Google Cloud 与 Oracle 进入 Gold,Azure 降至 Silver,Crusoe 降至 Bronze,全球仅 19 家 neocloud 获得奖章评级。评测分审计、性能、可靠性三阶段,涵盖 GPU 计算、网络、存储、编排、监控与故障注入,并新增 Bronze 与 Underperforming 之间的 Participation Ribbon 档。

    etwork NAD, for hardware-offloaded Open vSwitch; rdma ib / rdma roce: rdma/ib, rdma/roce, or nvidia.com/rdma , with one NAD per rail or a host-network fallback; Provider-specific: AWS EFA (vpc.amazonaws.com/efa and libfabric), DOKS (rdma/fabricN with roce-net-fabricN@fabricN), and GKE DRANET (DRA claim templates and pod resourceClaims) Device setup and the NCCL plugin on top of it can change job performance. On Google’s GB300, the built-in NCCL library picked an unroutable link-local GID and hun

    匹配字段正文