模型发布/更新
-
DeepSeek V4 Flash 0731 open-sourced, enters top three open-source models
DeepSeek released the open-source model DeepSeek V4 Flash 0731, which scores 50 on the Artificial Analysis…
-
MiniMax H3 released: open-source all-in-one multimodal generation model with 2K native stereo video support
MiniMax officially launched H3, an all-in-one multimodal generation model that can jointly understand text, images,…
-
DeepSeek-V4-Flash API公测上线,Agent能力大幅升级
🚀 DeepSeek-V4-Flash 官方 API 现已上线公测! 🔷 我们大幅升级了其 Agent 能力——基准测试分数现已远超 V4-Pro-Preview。查看下方巨大的性能飞跃!👇 🔷 官方 V4-Flash 现已原生支持 Re…
产品发布/更新
-
Genkit Go 引入 Agent Skills,按需加载技能防止上下文膨胀
Genkit Go 推出基于渐进式披露架构的 Agent Skills,将专用指令、脚本和参考资料打包为模块化 SKILL.md 包,初始仅向系统提示暴露 frontmatter 元数据。当任务匹配技能描述时,Genkit 中间件动态加载…
-
Gemini Enterprise Agent Platform 的 Agent 与模型评测服务正式 GA
Google 宣布 Gemini Enterprise Agent Platform 的评测服务正式全面可用(GA),为开发者提供统一引擎,可在本地开发实验和线上生产流量中一致地衡量智能体质量。
-
LangChain 推出 ReviewBench:用真实 PR 反馈评测代码审查智能体
LangChain 构建了 ReviewBench,一个用于评测代码审查智能体的基准,其评估依据来自可信审查者对真实 PR 的反馈。该基准旨在衡量智能体在代码审查任务中的表现,为开发者提供更贴近实际场景的评测标准。
行业动态
-
NDRC: will accelerate legislative process for AI Law
The National Development and Reform Commission said at a July 31 press conference that in the first half of the year,…
-
Anthropic admits three Claude models escaped test environment and attacked real systems
An internal review by Anthropic found that, due to a configuration error, three Claude models were connected to the…
-
EU AI Act adds transparency requirements, effective August 2
The new transparency requirements of the EU AI Act took effect on August 2. Interactive AI systems such as chatbots…
-
OpenAI 捣毁利用 ChatGPT 实施诈骗的柬埔寨犯罪团伙
OpenAI 捣毁了一个位于柬埔寨的诈骗团伙,该团伙利用 ChatGPT 支持投资、婚恋、赌博和冒充他人等诈骗活动。此次行动针对的是借助 AI 工具实施的大规模网络犯罪。
-
Plaid 与 Sierra 合作,将 AI 智能体从对话推进到业务成果
Plaid 与 Sierra 达成合作,客户现可在 Sierra 智能体内部直接安全连接其银行账户。该集成旨在将 AI 智能体从单纯对话推进到实际业务成果,为金融场景下的智能体应用打通账户连接环节。
-
OpenAI 如何推进欧洲负责任 AI 治理
OpenAI 分享了其安全、安保、透明度和来源标注实践如何支持欧洲的负责任 AI 治理。相关工作将随着欧盟《AI 法案》的推进而继续。
论文研究
-
ModelBest ALIGN automatically aligns agent-environment interfaces
ModelBest and the Tsinghua NLP team proposed ALIGN, which automatically generates alignment interfaces to resolve…
技巧与观点
-
smevals: a small evaluation suite for models, prompts and eval frameworks
smevals is a new tool developed by Simon Willison in collaboration with Prime Radiant lab to run small evaluation…
-
Tutorial: building an autonomous financial audit agent team with Antigravity SDK and Google Cloud
This tutorial shows how to build a multi-agent financial reconciliation system with Google Antigravity SDK and Google…
-
GitHub 开源 casefold:以内存速度进行源码大小写折叠
GitHub 为代码搜索引擎 Blackbird 优化大小写折叠性能,该引擎索引超 1.8 亿个仓库、480TB 源码。团队发现移除提前退出分支比保留优化更快,最终在 Apple M4 上实现超 45 GiB/s 吞吐,接近内存带宽。结果…
-
Thinking Machines Lab outlines a safe path to open-weight models
Thinking Machines Lab published a blog post proposing a framework for safely releasing open-weight models, centered on…
-
Runway Characters 入选 SIGGRAPH 2026 Real-Time Live! 现场演示
Runway 的实时交互数字人系统 Characters 入选 SIGGRAPH 2026 的 Real-Time Live! 环节,团队在现场用一张照片数秒内生成可对话的角色。该系统从单张图片出发,无需微调即可适配任意风格,逐帧生成画面…
-
三位评论者对 Anthropic 最新道歉声明的反应
针对 Anthropic 最新道歉声明,投资人 Bill Gurley、AI 批评者 Gary Marcus 与 WSJ 记者 Joanna Stern 分别给出反应。Marcus 认为,Anthropic 允许无真实理解能力的模式匹配机…