N全部新闻

全部新闻

星期五1456 条

  1. Rohan Paul
    AI 评分67

    Anthropic 占 SpaceX 第三季度营收 34.5%，成其最大客户

    彭博社披露，Anthropic 占 SpaceX 第三季度 145.75 亿美元营收的 34.5%，超过美国政府成为其最大客户。SpaceX 文件仅将该客户标为 Customer B，相关付款用于租用其孟菲斯数据中心的 Nvidia GPU 算力，该算力租赁业务也服务 Google。Anthropic 于 10 月 26 日提交的 S-1 显示，与 SpaceX 的安排到 2029 年最高达 1205 亿美元，而其对 Google、Amazon、Microsoft 的承诺分别为至少 1145 亿、1100 亿和 449 亿美元的最低值。

    同一事件 · 4 条报道

    #行业动态#数据/训练#Anthropic

  2. Rohan Paul
    AI 评分40

    Rohan Paul 评 Sam Altman 谈本地推理

    Rohan Paul 引述 Sam Altman 对本地推理的看法：从全球资源分配角度看，让全世界有限的 DRAM 都躺在各人电脑里、只用到 5% 是很糟糕的。Rohan Paul 补充说，按每 token 成本算数据中心明显占优，batch size 1 的本地模型在任何利用率指标上都是浪费，但它也是整个技术栈里唯一没人能对你下架的部分。

    同一事件 · 11 条报道

    #大佬观点#端侧#推理

  3. Rohan Paul产品发布
    AI 评分62

    Anthropic 为 Claude 加入回答后自评估，标记对话中的潜在滥用

    Anthropic 训练 Claude 在每次回答后对对话做初步安全评估，这项名为 self-assessment 的技术自 6 月启用以来已发现其他防护机制未能覆盖的滥用行为。该机制利用 Claude 已读过对话的优势，成本低于分类器，且能随模型改进而提升准确度。Rohan Paul 补充称，Anthropic 最准确的滥用检测器是需先读全部安全政策的 Claude Mythos Preview 分类器，运行成本过高；Anthropic 因此训练 Claude 在不读政策和推理过程的情况下复制其判断，在 Fable 5 流量上该副本匹配检测器 78% 的标记，高优先级案例升至 94%，误报率为 1%。

    引用Anthropic · @AnthropicAI

    We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation. Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch. Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve. Read more on our new Safeguards blog: https://t.co/18j6zdHub4

    同一事件 · 5 条报道

    #安全/对齐#Anthropic#产品更新

  4. Rohan Paul
    AI 评分69

    Nvidia 洽谈为 Nscale 的 460MW Anthropic 数据中心担保超 50 亿美元建设债务

    Nvidia 正洽谈为 Nscale 位于西弗吉尼亚州的 460MW Anthropic 数据中心担保可能超过 50 亿美元的建设债务，消息来自 The Information。这笔尚未签署的协议担保的是建筑价值而非租约或芯片，可让 Nscale 以投资级条件借款。若 Nscale 拖欠建设款项，Nvidia 将代为偿付、临时运营该站点并安排出售与回租。

    同一事件 · 4 条报道

    #行业动态#数据/训练#OpenAI

  5. Karina Nguyen
    AI 评分44

    BallotBench 实时基准：9 个前沿 AI 智能体预测美国中期选举

    Thoughtful 推出 BallotBench 实时基准，让 9 个前沿 AI 智能体从即日起至 11 月 3 日预测 2026 年美国中期选举。每个智能体拥有独立沙箱、联网权限和 500 美元预算，每天更新预测并可申请数据或额外资源。选举结束后，其准确率将与顶级超级预测者、选举评论员以及 Kalshi 和 Polymarket 上的群体智慧对比。

    引用Thoughtful · @thoughtfullab

    Can AI forecast elections better than experts and prediction markets? We built BallotBench to find out: a live benchmark where 9 frontier AI agents compete to predict the 2026 US midterms, from today through Election Day, Nov 3. Follow along → https://t.co/7aLmUy2HvO Watch the agents do their research, revise their predictions, and make their calls all in live feeds.

    #评测/基准#Agent#推理

  6. Josh Woodward模型发布
    AI 评分72

    Google 发布 Gemini 4 Carbon，距 Argon 仅两周

    Google 发布下一代前沿模型 Gemini 4 Carbon，距 Gemini 4 Argon 发布仅两周。据 Sundar Pichai 介绍，Carbon 在 ArtificialAnlys Intelligence Index v5.0 上以 59 分排名第一，在网络安全评测中 SEC-bench Pro 得分 88.2%、ExploitGym 得分 54.8%、CWE-bench 得分 71.0%，长周期任务相比 Argon 有明显提升。该模型将先向受信任的网络安全防御方开放，再通过 Fairwind Program 扩大合作范围，随后更广泛可用。

    引用Sundar Pichai · @sundarpichai

    I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model. Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards. As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!

    同一事件 · 34 条报道

    #模型发布#Google#安全/对齐

  8. Alex Heath
    AI 评分28

    Shopify CEO 用 OpenClaw 分析压力会议

    Shopify CEO @tobi 把自己的 OpenClaw 与 Fitbit 数据同步，让它找出自己压力最大的会议，随后把结果告诉了相关参会者。他说：“他们并不觉得这好笑。”

    引用Alex Heath · @alexeheath

    Tobi Lütke on the AI agent revolution and future of Shopify @tobi on where agent commerce is headed, sci-fi becoming reality, and his GPU wine cellar. Why did Amazon block Meta's Muse but Shopify welcomed it? I sit down with Tobi in Toronto to discuss why the companies are taking different approaches to agents and how shopping changes when AI starts buying things on your behalf. He explains the early AI shopping behavior he’s seeing. We also get into why he thinks AI makes starting a business easier, the backlash against the technology, and why everyone will eventually have to embrace it to get ahead in their careers. We get into how AI agents will make money, who pays when they handle a purchase, and why advertising gets complicated when your assistant is supposed to work for you. We discuss his experience on Meta’s board and working with Mark Zuckerberg. Lütke also shares why he turned his wine cellar into a GPU server rack and how he uses his personal AI agent that runs entirely from his house. This is one of my favorite pods I've ever done. Timestamps: 00:00 Amazon blocks Muse 02:40 Shopify's approach to AI agents 07:00 How AI agents will change shopping 11:05 Sci-fi and tech optimism 17:57 Zero-sum thinking vs. abundance 21:20 AI backlash and the founder advantage 45:28 Building an AI-first Shopify 52:18 Who profits from agentic commerce? 01:03:22 How Shopify stays ahead 01:10:39 Advertising and discovery in the agent era 01:16:02 Rack, Cellar, and Claw Links: Spotify: https://t.co/dEn15RrP7B Substack: https://t.co/JxkbohU0qz YouTube: https://t.co/2DTtOZ64yN Apple Podcasts: https://t.co/fEI7x7IzlD This episode's presenting sponsor is @hyperagentapp Thanks to the show’s premiere sponsors: @Atlassian, @meetgranola, and @mercury

    #现象/趋势#Agent#MCP/工具调用

  9. 卡尔的 AI 沃茨产品发布
    AI 评分64

    豆包工作上线录制回放功能，支持边操作边讲解最长 30 分钟

    豆包工作上线录制回放（record & replay）功能，可把电脑操作过程录下来交给 AI 生成步骤说明并复刻，类似简化版 RPA。相比 Codex 同类能力，它额外支持边操作边讲解，最长录制 30 分钟，处理时优先转换音频。作者实测用它做跨平台内容采集、公众号排版发布等任务，并指出执行中若占用同一窗口任务会主动暂停，同时打开多个网页则不受影响。

    #产品更新#Agent#MCP/工具调用

  15. Saito 硬地骇客观点
    AI 评分81

    SemiAnalysis 解读 Anthropic S-1：90% 增量毛利与半万亿算力承诺

    SemiAnalysis 播客解读 Anthropic 的 S-1，指出其增量毛利 90%、增量 EBIT 40%，推理毛利 85%、算力毛利 41%，non-GAAP 口径已连续两个季度盈利。客户净收入留存方面，去年同期花 1 美元的客户现在花 9 美元，Q1 节点为 500%，另有 250 家客户年付超 1000 万美元、1900 家超 100 万美元。节目还拆解了 5000 亿至 6000 亿美元算力承诺按每吉瓦 1000 亿美元折算仅约 5 吉瓦，并讨论了前五大客户占收入 25%、48% 走云厂商 marketplace 的口径分歧。

    #行业动态#OpenAI#Anthropic

  17. Ethan Mollick
    AI 评分40

    历史学家为 Talkie-1930 构建神经元图谱

    Ethan Mollick 提到，近期有不少通过激活开放 AI 模型中单个神经元、将其锁定为特定人格的有趣实验。历史学家 @ResObscura 用同样方法尝试了一个仅用 1930 年前数据训练的模型，结果很有意思；他本人表示，受 @Sauers_、@jfonsecarivera 等人近期人格研究启发，为 @status_effects 和 @AlecRad 打造的历史语言模型 Talkie-1930 构建了部分神经元图谱。

    引用Benjamin Breen · @ResObscura

    Inspired by @Sauers_ and @jfonsecarivera and other fun recent work with personas in AI models, I built a partial neuron atlas for Talkie-1930, the historical language model created by @status_effects and @AlecRad. Here's the link: https://t.co/ROz61DW4RW https://t.co/8RvqEsx5K0

    #现象/趋势#推理

  19. Mistral AI · X
    AI 评分36

    Voxtral Mini Transcribe 3 在 Voice Arena 说话人分离榜排第二

    Mistral AI 转发 Voice Arena 评测：Voxtral Mini Transcribe 3 在 Diarization + ASR 榜单 9 个模型中 DER 排名第 2（40.4%），优于同系列更大的 Voxtral Small Transcribe 3（41.1%，第 4）。

    引用Voice Arena · @voicearena_ai

    🚀 Mistral AI’s Voxtral Mini Transcribe 3 ranks #2 out of 9 models by DER on Voice Arena’s Diarization + ASR leaderboard, outperforming its larger counterpart. Following the launch of Mistral’s new Voxtral 3 family, we evaluated its Mini and Small transcription models on Voice Arena’s Diarization Bench. 🥈 Voxtral Mini Transcribe 3: 40.4% DER (#2) Voxtral Small Transcribe 3: 41.1% DER (#4) Both models were tested on 139 held-out conversations under identical conditions, using human-verified references, a strict 0 ms collar, and no predefined speaker counts or audio cleanup. Mini leads on in-person conversations (40.6% vs. 41.5%), while Small performs slightly better on online calls (39.1% vs. 39.5%). Overlapping speech remains the biggest challenge. Missed speech accounts for 82% of Mini's total DER. Both models correctly identify the exact number of speakers in 83.5% of sessions. Congratulations to @MistralAI on the release ! 👏 View complete track & condition breakdown: https://t.co/sLGwgH3EWt

    #评测/基准#语音