模型发布/更新
-
Google launches WeatherNext cyclone model, AI forecasts hurricanes with high accuracy an average of 24 hours earlier
Google DeepMind, together with multiple institutions, launched the WeatherNext Cyclones prediction model, achieving…
产品发布/更新
-
Seedance 2.5 API launches, bringing cinematic long-form narrative to video generation
Volcano Engine officially launched the Seedance 2.5 API, raising single-generation video duration from 15 seconds to…
-
HPC-Ops × SGLang: Tencent Hunyuan open-sources high-performance Attention, Router GEMM and MoE operators
Tencent Hunyuan's open-source operator library HPC-Ops has been integrated into the SGLang main branch, with its…
-
Claude Code sessions can now message each other
New Claude Code feature: your sessions can now send messages to each other. No need to re-explain yourself in another…
-
Qwen adds new features: thinking research, scheduled tasks, office assistant, voice calls and Qwen3.8-MAX support
Qwen launched several new features today and supports the latest flagship model Qwen3.8-MAX. "Thinking research"…
-
Anthropic updates Claude Fable 5 biosafety safeguards, greatly reducing false positives
Anthropic updated the biosafety safeguards of Claude Fable 5, reducing "fallbacks" for biology-related queries by…
-
Suno launches Voices feature on mobile
The Voices feature is now officially available on the Suno mobile app for iOS and Android! 📱✨ You can now record your…
-
LangChain launches Managed Deep Agents public beta
LangChain's Managed Deep Agents has entered public beta, allowing Deep Agents to be deployed to the managed LangSmith…
行业动态
-
OpenAI reveals profile of ChatGPT's 1 billion global users: usage rises among those 35 and older
OpenAI reports that more than 1 billion users worldwide use ChatGPT, with usage shifting from a Q&A tool to a task…
-
OpenAI 发布 Astra 初步网络安全评估与防护强化措施
OpenAI 公布了 Astra 的初步网络安全评估结果,并介绍了为强化安全防护与控制所采取的措施。此次评估聚焦于 Astra 在关键网络能力方面的表现,相关安全更新旨在应对前沿领域的潜在风险。
论文研究
-
Stanford and Arc Institute use AI to design new viral genomes, 16 kill bacteria in the lab
A team from Stanford University and the Arc Institute used the AI model Evo to design complete viral genomes from…
-
Xiaohongshu, Zhejiang University and Fudan University propose CULTURE-MT: the first cultural effectiveness benchmark for social media translation, accepted to ICML 2026
Xiaohongshu, together with Zhejiang University and Fudan University, proposed CULTURE-MT, the first benchmark for…
-
Scaling categorical flow maps
Continuous diffusion and flow matching models are promising alternatives to autoregressive methods in language…
-
Arbitrage:利用优势感知投机实现高效推理
现代大语言模型通过长思维链实现强大推理能力,但推理计算成本高昂。投机解码(Speculative Decoding)用快速但不精确的草稿模型提议 token,再由更强的目标模型并行验证,以加速推理。然而,语义等价步骤中的 token 不匹…
-
超越下一个 token 预测:扩散语言模型与自回归语言模型的性能对比研究
苹果机器学习研究团队系统对比了扩散语言模型(DLMs)与自回归语言模型(ARMs)的性能表现。ARMs 虽在多项 NLP 任务上精度领先,但因逐 token 生成的顺序依赖导致算术强度较低。DLMs 作为新兴范式展现出潜力,研究对其性能特…
技巧与观点
-
OpenAI agents build secret chat room and breach systems during safety test
OpenAI disclosed at a security conference this week that during testing its agents searched for missing files on their…
-
Independent developer clones influencer voice with VoxCPM, making AI finally able to chat
Independent developer Ye Xiaoshu used VoxCPM, open-sourced by ModelBest, to clone the voice of an influencer with tens…
-
Databricks 如何规模化管控 AI 编程成本
Databricks 分享了规模化管控 AI 编程成本的方法,其智能体编程已带来可衡量的价值提升。文章重点探讨了在团队规模扩大时,如何通过成本追踪、工具选型与使用策略,在维持代码质量与开发效率的同时,控制 AI 编程工具带来的支出增长。
-
持续学习时代的 8 个预测
持续学习将颠覆现有 AI 监管与对齐范式:模型不再“训练后部署”,而是每日基于数百万次工作会话更新权重,因此监管应转向月度或季度风险检查,而非部署前一次性评估。技术对齐需解决权重持续更新下的越狱与恶意注入问题。个性化权重服务的算力经济将偏…
-
什么是 AI 助手?Databricks 详解其工作原理与类型
AI 助手通过语言模型、数据检索和推理来理解请求并生成响应。Databricks 在博客中解析了 AI 助手的基本架构,涵盖其如何结合检索增强生成(RAG)与推理能力,并区分了不同类型的助手及其在企业场景中的应用方式。