模型发布/更新
-
Qwen-Image-3.0-Pro launches on Qwen Cloud
Alibaba's Tongyi Qianwen released Qwen-Image-3.0-Pro and Standard, now available on Qwen Cloud. The model ranks first…
产品发布/更新
-
Cloudflare OS:面向智能体、应用与工作的开放平台
Cloudflare 开源新版 Cloudflare OS,任何组织均可部署并连接内部系统。该平台为每位员工提供基于公司上下文与技能的智能体工作区,包含隔离运行时、安全治理框架及可共享修改的个人应用。此前内部版本已供数千名员工日常使用,新…
-
Grok 4.5 free to try, Build toolchain recommended
The best way to use Grok is through our Build toolchain. Download: http://X.ai/cli
-
Cloudflare 用身份感知分析捕捉失控 AI 行为
Cloudflare 推出身份感知 AI Gateway 与 User Insights,为每个请求绑定经 Access 验证的用户身份,并基于账户自身历史行为建立基线,识别偏离正常模式的异常会话。该功能现处于开放测试阶段,User In…
-
Meta 广告排序的多阶段序列模型:从用户序列到 LLM 式扩展定律
Meta 推出多阶段序列模型,将离线用户建模与在线排序解耦,并采用密集 token 化与目标感知注意力,使序列学习具备可预测的 LLM 式扩展定律。该架构已为 Instagram 转化率带来 6% 的累计提升,Facebook 转化率提升…
-
Claude Platform 发布 8 月 5 日版本说明:推理钩子进入 Beta 测试
Claude Platform 面向 Claude Enterprise 组织推出推理钩子(Inference hooks)Beta 版。该功能可将 claude.ai、Cowork 和 Claude Code 中的每个受管控提示词交由组…
行业动态
-
Jeff Dean leaves Google to found DiscoLoop AI
Jeff Dean announced he is leaving Google after 27 years, with his official departure tomorrow. He said Google grew…
-
Google Assistant to phase out starting next month as Gemini takes over "Hey Google"
Google notified Android users by email that Google Assistant on mobile will stop service progressively from September…
-
Demis Hassabis moves to Google DeepMind chair and Alphabet chief scientist
Demis Hassabis announced he is stepping down as Google DeepMind CEO to become chair and Alphabet chief scientist,…
-
US appeals court overturns injunction, Perplexity AI shopping agent returns to Amazon
The US Ninth Circuit Court of Appeals overturned an injunction that had blocked Perplexity from using its AI shopping…
论文研究
-
Cloudflare 提出智能体访问模型(Agent Access Model)
Cloudflare 发布《The Agent Access Model》论文,提出面向 AI 智能体的访问控制模型 AAM,核心规则是“不信任运行”,对任务执行图中的每个动作基于智能体身份、授权任务及已触达资源进行实时授权。该模型针对智…
-
驯服扩散 Transformer 中的离群 token:Dual-Stage Registers 干预
研究发现扩散 Transformer(DiT)图像生成流程中,预训练 ViT 编码器和 DiT 去噪器均会产生离群 token,尤其在中间层,且简单掩蔽高范数 token 无法改善性能,问题与局部 patch 语义损坏相关。为此提出 Du…
技巧与观点
-
Open-source "human Writing.skill" helps users write text without an AI flavor
Digital Life Kazike open-sourced "human Writing.skill," aimed at removing AI flavor and helping users write text with…
-
UK AI Safety Institute incident report: AI agents with safety filters disabled launch unauthorized attacks on the live internet
The UK AI Safety Institute (AISI) released an incident report stating that during a cyber evaluation held from July 25…
-
After burning 500 million tokens, my new tricks for slimming Skill context in Codex and Claude Code
The author slimmed down the context for more than 300 Skills in Codex and Claude Code, finding that the Skill list…
-
Building an end-to-end Bayesian marketing mix model with Google Meridian: media measurement, ROI analysis and budget optimization
This tutorial uses Google Meridian to build a complete Bayesian marketing mix modeling workflow, covering data…
-
Generating the complete Raccoon Heist game in one shot with Claude Fable 5
Simon Willison fed a 2022 game concept and screenshots generated with GPT-3 and DALL-E into Claude Fable 5 (running in…
-
SpaceXAI quarterly capital expenditure reaches $18.37B, AI spending nears 40% of Microsoft's total capex
SpaceXAI spent $18.37 billion in capital expenditure last quarter, of which $15.83 billion went to AI, close to 40% of…
-
马斯克关于机器人手术的荒谬且可能有害的预测
加里·马库斯批评马斯克关于机器人手术时间线的预测“完全疯狂”,并援引机器人专家罗德尼·布鲁克斯的观点:目前连在活体实验动物上进行手术的实验室演示都远未实现,现有手术机器人全部仍需人类在环操作。马库斯担忧此类预测可能吓退潜在外科医生,并指出…
-
LangChain 如何为 Kubernetes 构建自主 SRE 智能体
LangChain 基于 Deep Agents 为 Kubernetes 部署构建了自主 SRE 智能体,可自动执行运维任务,并对变更操作引入人工审批机制。该智能体使用 LangSmith 进行全链路追踪,并通过 evals 评估系统性…