模型发布/更新
-
Alibaba releases Qwen-Audio-3.0-Realtime, ranks first overall in Artificial Analysis speech reasoning subcategory
Alibaba's Tongyi Lab released the real-time voice interaction model Qwen-Audio-3.0-Realtime, ranking first overall in…
-
Thinking Machines releases multimodal model Inkling
Today we introduce Inkling. Inkling can efficiently reason over text, image, and audio modalities. We will provide…
-
OpenAI releases GPT-Red to improve model robustness through automated red teaming
OpenAI trained an automated red-team model, GPT-Red, to find vulnerabilities before deployment and generate attacks…
产品发布/更新
-
xAI open-sources Grok Build coding agent and terminal interface
xAI has open-sourced the source code of Grok Build on GitHub. Grok Build is SpaceXAI's coding agent and terminal user…
-
Grok Build is now open-source
Grok Build is now open-source
-
Kingsoft Office launches WPS Comate AI office client
Kingsoft Office launched WPS Comate, an AI office client for employees, at the 2026 AI Productivity Conference; it can…
-
Claude Code adds MCP connector invocation
Claude Code artifacts can now call MCP connectors, letting you build dashboards and apps that fetch information and…
-
SGLang 与 Miles 为前沿多模态模型 Inkling 提供 Day-0 支持,推理吞吐达 71.7k tok/s
SGLang 与 Miles 为 Thinking Machines Lab 的 975B 参数多模态模型 Inkling 提供 Day-0 支持,其上下文窗口达 1M token。
行业动态
-
China-region Apple Intelligence completes filing; Alibaba Qwen to be integrated into Apple AI capabilities
Apple Technology Development (Shanghai) Co., Ltd.'s "Apple Intelligence" large model completed filing on July 8, 2026,…
-
Alibaba Qwen to be integrated into Apple Intelligence for users in China
Alibaba's Qwen model will be integrated into Apple Intelligence to provide AI features such as text and image…
论文研究
-
Anthropic study: behavioral deviations in AI agent simulations
New Anthropic research: agentic behavioral deviations in summer 2026. One year after our extortion experiment, we…
-
OpenAI uses AI to attack its own AI: GPT-Red automatically finds security flaws with 84% success rate, far exceeding humans
OpenAI trained an internal AI model, GPT-Red, which uses self-play reinforcement learning to automatically simulate…
-
Apple 提出 LLM 函数调用不确定性量化方法
Apple 机器学习研究团队提出一种针对大语言模型函数调用的不确定性量化方法,旨在提升 LLM 在自主任务执行中的可靠性。该方法通过量化模型对函数调用参数和决策的置信度,帮助识别潜在错误调用。该研究目前处于学术探索阶段,未公布具体模型版本…
-
Meta 探索分层兴趣表示以优化广告深度漏斗
Meta 提出分层兴趣表示,一种基于 Transformer 图学习与自监督跨视图蒸馏的上游表示层,为广告实体学习统一嵌入。该系统在数十亿真实交互数据上端到端训练,输出通用嵌入和 Bag-of-Meaning 兴趣 token,旨在连接稀…
-
Apple 提出 CLaRa:用连续潜在推理桥接检索与生成
Apple 机器学习研究团队提出 CLaRa,一种通过连续潜在推理桥接检索与生成的新方法。该方法在检索增强生成(RAG)中引入连续潜在空间,使模型在生成答案前先进行隐式推理,从而缓解长上下文带来的性能下降。CLaRa 在多个知识密集型基准…
技巧与观点
-
Open-source LLM TODO Skill "A Fu": knowledge management to scheduling automation with Claude Code and Codex
The author developed the open-source TODO Skill "A Fu" based on the API version of Fable5 and Codex, used to…
-
Vibe coding 16 hours a day: author shares AI development workflow with Fable 5 and GPT-5.6 Sol
The author vibe codes about 16 hours a day and considers Claude Fable 5 "in a class of its own" for first-draft design…
-
OpenAI 呼吁通过“反向联邦主义”推动美国 AI 安全标准统一
OpenAI 首席全球事务官 Chris Lehane 发文,主张通过“反向联邦主义”——即各州先通过相似立法形成事实上的国家标准——来推动美国 AI 前沿安全治理。加州、纽约州和伊利诺伊州已通过相关立法,核心要素包括风险披露、安全事故报…
-
Base44 为何信任 Claude Fable 5 处理最具挑战性的工程任务
无代码开发平台 Base44 将此前仅限资深工程师处理的系统提示词重构任务交给了 Claude Fable 5。该模型在四小时内独立完成 90%-95% 的重构,并主动发现团队评估中遗漏的缓存命中测试盲点。Base44 产品负责人表示,C…