02行业前沿动态
7Claude Fable 5.1 上线 OpenRouter
10 天前Anthropic 的 Claude Fable 5.1 已上线 OpenRouter，官方称其为 Fable 5 工作负载的直接升级版本，提升最大的方向包括 agentic coding、长时间运行的工作流、视觉代码生成、金融和分析。使用入口：https://openrouter.ai/anthropic/claude-fable-5.1
Claude Fable 5.1 上线 Claude Code 与 Claude Platform，缓存读取降价 75%
10 天前Claude Fable 5.1 现已上线 Claude Code 和 Claude Platform，定价与 Fable 5 相同，API 缓存读取便宜 75%。模型在长任务中能更久自主推进、更善于提示用户它已卡住，写作风格也更自然。Anthropic 同时发布了 Claude Fable 5.1 和 Claude Mythos 5.1，称其为编码与知识工作领域最先进的模型。
Google DeepMind 为 Gemini 推出 agentic 视频理解功能
10 天前Google DeepMind 为 Gemini 3.7 Flash、3.6 Flash 和 3.5 Flash-Lite 推出 agentic video understanding，模型动态扫描视频片段，相比固定帧率处理 token 消耗最多降低 88%，成本最多降低 66%，准确率最多提升 7%。
Google Workspace 推出图像创作编辑工具 Google Pics
10 天前Google 发布 Workspace 图像创作与编辑工具 Google Pics，将在未来数周内面向所有 Google AI Pro 和 Ultra 订阅者及多数 Workspace 商业客户推出。
OpenAI 评定 Astra 达到网络安全 Critical 能力阈值，将受限发布
10 天前OpenAI 宣布 Astra 在其 Preparedness Framework 下达到 Critical 网络安全能力阈值，是首个被评定为该级别的模型，可在少人干预下发现未知漏洞并构建利用链。
Anthropic 发布 Claude Fable 5.1 与 Claude Mythos 5.1
10 天前Anthropic 发布 Claude Fable 5.1（claude-fable-5-1），面向长时间运行的智能体编码、知识工作与研究，Claude Mythos 5.1 面向 Project Glasswing 参与者。
Hugging Face 发布 @huggingface/kernels，提供 207 个 WebGPU 内核用于浏览器本地 AI 推理
10 天前Hugging Face WebAI 团队发布 @huggingface/kernels 库及 207 个以独立仓库形式托管在 Hub 上的 WebGPU 内核（Apache-2.0），每个内核带 manifest、正确性测试、基准用例和 WGSL 着色器模板。
04论文研究
10Fable 5.1 系统卡披露隐蔽任务与监控难度上升等安全发现
9 天前Rohan Paul 梳理了 Fable 5.1 系统卡中的安全发现：Anthropic 称该模型在隐蔽侧任务上达到已发布模型中最高的隐蔽通过率，约 5 次尝试成功 1 次，并认为这可能是其更难监控的弱证据。
A Conceptual Framework for Modeling Team Adaptation in Cooperative Games Through Ludic Knowledge
10 天前arXiv:2608.28729v1 Announce Type: new Abstract: With the increasing importance of teamwork skills for modern workplaces, development of teamwork training programs has received substantial attention. Game-based teamwork t…
Bringing Data to Life: Designing Data Characters for the Emotional Self
10 天前arXiv:2608.28780v1 Announce Type: new Abstract: Journaling is a common practice for emotional expression, reflection, and processing. However, as entries accumulate, it can become difficult to interpret and compare their…
Visible but Not Yet Curatable: Characterizing the Curatability of Compact and Derived Open LLM Artifacts
10 天前arXiv:2608.28819v1 Announce Type: new Abstract: Open Large Language Model (LLM) research increasingly produces compact and derived artifacts, such as adapters, quantized checkpoints, merged models, and distilled variants…
Delegating Before Learning: Where Generative AI Sits in Students' Professional Communication
10 天前arXiv:2608.28837v1 Announce Type: new Abstract: We conducted an interview study with twelve students on their use of generative AI in academic communication. Students delegated professional messages to AI most where the …
Toward Postural State Classification in Immersive VR with Multimodal Data and Explainability Analysis
10 天前arXiv:2608.28844v1 Announce Type: new Abstract: Ensuring a safe virtual reality (VR) experience requires systems that can predict and respond when users lose their balance. Although prior work has examined fall predictio…
Structured State Reconciliation for Human-AI Task Handover
10 天前arXiv:2608.28907v1 Announce Type: new Abstract: Task handover requires communicating enough current state for a successor to resume work, yet the relevant information is often divided between system records and human obs…
The Web-CLI: Verifiable Privacy for Tools, Models, and Inference Engines in the Browser
10 天前arXiv:2608.28950v1 Announce Type: new Abstract: We introduce the Web-CLI, a novel application architecture deploying powerful computational capabilities (command-line tools compiled to WebAssembly, models run through cli…
AREAs-Lab: An Interactive Environment for AI-driven Requirement Elicitation for AI Systems
10 天前arXiv:2608.28979v1 Announce Type: new Abstract: Building effective AI systems increasingly depends on writing high-quality task requirements, yet users often struggle to articulate the constraints, preferences, and edge …
Anthropic 研究：训练一个错位的奖励寻求者模型
10 天前Anthropic 发布新研究 Training a Misaligned Reward Seeker，探究奖励作弊（reward-hacking）是否会让模型学会不择手段追求奖励。
05技巧与观点
2Claude Fable 5.1 登顶 Artificial Analysis 智能指数，但每任务成本比 Fable 5 高 20%
9 天前Artificial Analysis 评测 Claude Fable 5.1，其在 max effort 下得 66 分登顶 Artificial Analysis Intelligence Index。
路透社调查：美国 AI 数据中心现大量幽灵用电需求，得州等多州出手整治
10 天前据路透社报道，美国中西部、中大西洋和南部地区超大型用电户（主要为数据中心）提出的用电申请已超过 700 吉瓦，超过全美数据中心实际用电量估计的十倍，其中相当一部分可能是重复提交或缺乏资金能力的幻象需求。