模型发布/更新
-
MolmoMotion: A Language-Guided 3D Motion Prediction Model
MolmoMotion is built on the Molmo 2 backbone and takes video frames, 3D point markers on objects and textual action…
-
Grok 4.3 Now Generally Available on Amazon Bedrock
On June 17, xAI announced that Grok 4.3 is generally available on Amazon Bedrock. The model achieves the lowest…
产品发布/更新
-
Vercel Releases Open-Source AI Agent Framework Eve: Each Agent Is a File Directory
Vercel has released Eve, an open-source AI agent framework (npm package, Apache-2.0 license). Eve adopts a…
-
Google Releases $99 Gemini Smart Speaker
Google has launched the Google Home Speaker, its first smart speaker built specifically for Gemini, priced at $99.99.…
-
Alibaba Cloud Releases HappyOyster 1.0: A Real-Time Interactive Digital World from One Sentence
On June 17, Alibaba Cloud released the open world model HappyOyster 1.0. Built on a native multimodal architecture,…
-
Claude Design update: maintains brand consistency across projects, works with Claude Code
On June 17, Claude Design was updated to support a unified design system across projects and to synchronize workflows…
-
Strands Robots SDK: One Agent from Hugging Face Hub to Physical Robots
AWS's open-source (Apache 2.0) Strands Robots SDK wraps the LeRobot stack as AgentTools to build a unified agent. It…
行业动态
-
Report: OpenAI burned $3.7 billion in cash in Q1, more than half its revenue
OpenAI burned $3.7 billion in cash in the first quarter of 2026, more than half of its $5.7 billion in revenue over…
-
China steps up preparations for World AI Cooperation Organization
China is stepping up preparations for the World AI Cooperation Organization and welcomes all parties to join. On July…
-
Google Releases Agentic Resource Discovery (ARD) Open Specification
Agentic Resource Discovery (ARD) is an open specification for publishing, discovering and verifying AI tools, skills…
-
Databricks 扩大对 Snowflake 的领先优势
Databricks 年化经常性收入(ARR)达 69 亿美元,同比增长 80%;同期 Snowflake ARR 约 53 亿美元,增速 34%。两者差距从 3 月的 4.9 亿美元扩大至 16 亿美元。AI 产品年化收入 17 亿美元…
-
Claude Opus 4.8 Build Day黑客马拉松获奖项目揭晓
6月13日,Anthropic在旧金山举办12小时黑客马拉松,310名参与者使用Opus 4.8和$500 credits完成原型。第一名Tekton:输入历史建筑照片后,Claude自动搜集图纸等资料,跨339个施工步骤重建3D模型,每…
-
Anthropic 在首尔开设办公室并宣布多项韩国AI生态合作
Anthropic 正式启用首尔办公室,并宣布与韩国AI生态的多项合作。NAVER 在全公司部署 Claude Code,数千工程师用于提升编码效率;Nexon 工程团队用 Claude Code 编写游戏代码。LG CNS 将 Clau…
论文研究
-
NVIDIA GEAR Lab Releases ENPIRE: 8 Codex Agents Autonomously Control Robots to Conduct Physical Experiments
NVIDIA GEAR Lab has introduced ENPIRE, the first system to achieve autonomous research in the physical world. The…
-
Optimizing Ling-2.6-1T on TPU with SGLang-JAX: One Pallas Kernel Hides MoE Data Movement in Compute
SGLang-JAX now supports efficient inference of inclusionAI's Ling-2.6-1T (a 1T sparse MoE with 63B activated…
-
LifeSciBench released
In June 2026, OpenAI, together with 173 PhD-level life scientists, released the LifeSciBench evaluation benchmark,…
-
Google Medical Reasoning AI System AMIE: From Diagnosis to Long-Term Disease Management
A study published today in Nature shows that Google's medical reasoning AI system AMIE (Articulate Medical…
-
OpenAI 与 Molecule.one 合作:GPT‑5.4 自主优化 Chan‑Lam 偶联反应
OpenAI 将 GPT‑5.4 接入 Molecule.one 的自主化学智能体 Maria,用于优化药物化学中的 Chan‑Lam 偶联反应。GPT‑5.4 独立识别伯磺酰胺为高价值挑战性底物,并建议使用 TEMPO 等温和氧化剂。经…
技巧与观点
-
Google shares three integration architecture patterns for A2UI and MCP Apps
Google shared three architecture patterns for integrating A2UI with MCP Apps, aiming to combine the strengths of both.…
-
预训练还不够“苦涩”
Richard Sutton的“苦涩教训”通常被解读为警告不要在AI系统中编码过多人类知识,最终胜出的方法是能吸收更多算力和数据的一般性方法。现代基础模型预训练表面上是这一教训的胜利:采用通用架构、海量数据、简单的自监督目标(语言模型预测…
-
博客现状,2026年中
Nathan Lambert 在 Interconnects 博客创办约三年后更新规划。他当前三大目标:为前沿模型演进提供清晰度、创建开放模型生态、建立支撑机构。博客定位为原始、高辨识度的独立声音,避免成为全职分析平台。已披露与 Arce…
-
特朗普向Anthropic提出不可能的要求
特朗普要求Anthropic完成不可能的任务,暴露了生成式AI安全护栏的根本困境。早在2024年1月,Gary Marcus就指出任何护栏都难以在过于严格和过于宽松之间找到平衡。如今这一判断得到验证:基于next-token predic…