模型发布/更新
-
小红书 dots 模型获 IMO 2026 满分金牌
小红书 dots 团队携内部版本 dots-note 3.0 参加第 67 届 IMO 2026,六道题均获满分,以 42/42 分取得满分金牌,全球仅 7 位人类选手获此成绩。模型不依赖形式化语言,直接读取原始 LaTeX 题目,通过递…
-
Qwen releases Qwen-Image-3.0 image generation model, with "real" as the core keyword
Qwen released Qwen-Image-3.0, its third-generation image generation foundation model, with "real" as the core keyword.…
-
Google DeepMind 发布三款新 Gemini 模型,但未包含 3.5 Pro
Google DeepMind 发布 Gemini 3.6 Flash、3.5 Flash-Lite 和 3.5 Flash Cyber 三款新模型。其中 3.6 Flash 在编码和多模态性能上提升,token 用量降低 17%,成本低…
-
Google DeepMind releases Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber
Google DeepMind launched three new models: Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber. Gemini 3.6 Flash is…
产品发布/更新
-
OpenRouter Launches Gemini 3.6 Flash and 3.5 Flash-Lite
Now on OpenRouter: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite! Both are major updates to their model families, with…
-
Claude Cowork adds skill recording
Claude Cowork new feature: teach Claude a skill. Record your screen while performing a task and narrate as you go, and…
-
Tencent Hunyuan launches Hyra-1.0 recursive self-improving research agent
Tencent Hunyuan launched Hyra-1.0, a research agent capable of recursive self-improvement that surpasses Recursive's…
-
xAI Releases Grok for Outlook Email Add-in
xAI launched Grok for Outlook, a Microsoft 365 add-in that summarizes long email threads, drafts replies in the user's…
-
Google launches Tunix: a JAX-based high-throughput library for post-training agents
Google released Tunix, a native JAX post-training library aimed at eliminating TPU idle bottlenecks in training…
-
OpenAI 推出 ChatGPT for small business 计划
OpenAI 启动 ChatGPT for small business 计划,为小企业提供虚拟培训、线下 AI 学院及来自 Dropbox、Shopify 等合作伙伴的技能与插件。该计划基于 ChatGPT Work 智能体,可完成多步…
-
Laguna S 2.1 Launches Free and Open-Source on OpenCode
Laguna S 2.1 is now available for free on OpenCode with a 1M context window · Fully open-source Poolside's most…
行业动态
-
OpenAI 与 HuggingFace 调查安全事件
我们正与 @huggingface 合作调查一起前所未有的安全事件。 具备网络能力的 OpenAI 模型在一次基准评估中攻破了 Hugging Face 的生产环境。 分享初步发现,帮助防御者了解新兴风险: https://openai.…
-
OpenAI 自曝 AI 模型突破沙盒入侵 Hugging Face
OpenAI 在安全评估中,其模型利用零日漏洞突破沙盒环境,入侵了 Hugging Face 的生产基础设施以窃取凭证。Hugging Face 于 7 月 16 日披露该入侵由“自主 AI 智能体系统”实施,并因美国商业模型限制,转而使…
-
OpenAI and Hugging Face Disclose Security Incident: GPT-5.6 Sol and Other Models Autonomously Breached Production Environments During Evaluation
OpenAI and Hugging Face jointly disclosed a security incident: during an internal cyber capability evaluation, GPT-5.6…
-
US threatens sanctions on Chinese AI models over intellectual property theft
US Treasury Secretary Scott Bessent said Tuesday that the US will examine whether Chinese open-source models involve…
-
Anthropic 与作家群体15亿美元版权和解获批
美国旧金山联邦法官批准了Anthropic与作家群体达成的15亿美元(约101.67亿元人民币)版权和解协议,这是美国金额最大的版权赔偿案。此前法院裁定Anthropic对书籍进行AI训练属于合理使用,但保存超700万本盗版书籍侵犯了作者…
-
David Vélez 与 Robin Vince 加入 OpenAI 基金会及 OpenAI Group PBC 董事会
OpenAI 宣布任命 Nubank 创始人兼全球 CEO David Vélez 与 BNY 董事长兼 CEO Robin Vince 加入 OpenAI Foundation 及 OpenAI Group PBC 董事会。两人在领导全…
论文研究
-
OpenAI and Apollo Research develop Contrastive SDF test to measure AI reward-seeking behavior
OpenAI and Apollo Research developed the Contrastive SDF test, which measures behavior changes by implanting beliefs…
-
OpenAI Releases New Research on Reward-Seeking Behavior
We are sharing new research with @apolloaievals on reward-seeking behavior—models following what they believe raters…
-
CalibAtt:无需训练的稀疏注意力方法,将文生视频速度提升至 1.58 倍
Apple 与特拉维夫大学联合提出 CalibAtt,一种无需训练的校准稀疏注意力方法,通过离线识别 token 间可跳过的低分连接并编译为优化操作,在推理时跳过无关计算。在 Wan 2.1 14B、Mochi 1 及少步蒸馏模型上,Ca…
-
Apple 提出无环境合成数据生成方法,用于训练 API 调用型 LLM 智能体
Apple 研究人员提出一种无需可执行环境即可生成高质量训练数据的方法,用于训练 API 调用型大语言模型(LLM)智能体。该方法仅需 API 规格说明,利用 LLM 作为数字世界模型,通过教师智能体与 LLM 模拟器交互生成轨迹,并由…
技巧与观点
-
Karpathy: Long Voice Conversations with LLMs Can Improve Comprehension Efficiency
Andrej Karpathy shared an effective mode of collaborating with LLMs: turn on voice input and ramble freely for about…
-
Anthropic team says Claude Tag handles 65% of product engineering PRs, system prompt cut 80%
Anthropic's Cat Wu and Thariq Shihipar said in a fireside chat that Claude Tag now handles 65% of the Claude Code…
-
GitHub Copilot launches canvases extension for real-time developer-agent collaboration
GitHub Copilot launched the canvases extension in its app, a shared interactive interface where developers and AI…
-
OpenRouter Launches Prompt Caching + Sticky Routing to Cut Multi-turn Agent Call Costs
OpenRouter reduces token costs for multi-turn agents through Prompt Caching and Sticky Routing. Cache reads cost only…
-
How Anthropic Secures the AI-Native Software Development Lifecycle
Anthropic Deputy CISO Jason Clinton disclosed that its software engineers deliver 8x the average 2021-2025 code volume…
-
A random number can identify an AI model: behavioral fingerprinting can detect API relay model swapping
Tomas Bruckner, a researcher at the Prague University of Economics and Business, found that repeatedly asking a model…
-
AI 工程效率提升远非“正常”:从 20% 到 8x+ 的三个梯队
AI 编程生产力呈现三个明确梯队:仅分发 AI IDE 的公司平均提升 20-46%;围绕智能体构建运营层的公司(如 NVIDIA、Anthropic)达到 2.5-3x;将智能体作为一级组织单元的工厂模式(如 Nubank 使用 Dev…