Latest
Sourcegraph 的 Agentic Batch Changes 能连 Linear、Confluence、Buildkite 这些 MCP 服务器了，写代码的 agent 可以自己跑完整工作流，做批量代码变更的可以看看。
Synthesia 把视频生成搬进了 ChatGPT，从写脚本到出片一条聊天搞定，还能设置成自动跑，手机上就能用。
瑞典法律科技公司 Legora 上线了本体驱动的法律检索，Agent 能顺着法条之间的关系查，比关键词搜索靠谱，做法律工作可以试试。
The shift is from tribal analyst knowledge to certified infrastructure: once business definitions are codified, sales, finance and operations agents share one auditable source of truth instead of each inventing its own metrics. Teams building text-to-SQL or data agents should evaluate it against hand-written dbt metric layers, since the Certify and Validate steps directly target hallucinated numbers. NVIDIA released it open-source, so enterprises can self-host the semantic layer alongside their warehouses.
Alibaba Cloud 把 WAN 视频生成和 Qwen 智能体搬进了达喀尔青奥会现场，观众能现场生成鼓乐视频，塞内加尔风味挺独特。
Mid-sentence voice switching and word-level tone control push speech synthesis beyond whole-clip generation, which matters for dubbing, audiobooks and dialogue-heavy content where one model per line is too rigid. Developers building voice products on Alibaba Cloud get a new expressive TTS option alongside existing offerings. Fish Audio is a China-based audio AI startup, so its distribution through Alibaba Cloud shapes where teams in the region can access it.
KV cache reuse is becoming the bottleneck for agent serving, and engine-level cache sharing across parallelism setups is a practical lever for cutting latency and GPU cost. Teams running SGLang inference at scale can apply HiCache to avoid recomputing prefixes in multi-parallel, distributed deployments. Alibaba shaping an open-source inference stack also signals how Chinese cloud vendors compete at the serving layer.
阿里把 Qwen 系列模型放进 QwenWork 里并排跑同一提示词，速度和 credit 消耗一目了然，选模型不用再瞎猜了。
Alibaba Cloud 的 Wan3.0 能一口气生成 30 秒视频，现在还能免费试用，做视频的可以去试试效果。
阿里云把 Qoder 搬到了阿姆斯特丹的 Apsara 大会现场，Qoder 国际负责人 William Yao 实际演示了怎么用它写代码、跑智能体，想入门 agentic 编程的可以看看。
阿里云在新加坡给电信运营商画了条新路线：用 Qwen 和 AgentCore 帮运营商卖自有品牌 AI 服务，国内厂商出海东南亚的典型打法。
Real-world robot evaluation has lacked shared standards, and publishing failed rollouts alongside successes gives the field a reproducible benchmark rather than cherry-picked demos. Labs and hobbyists can apply for free eval hours to test policies on heterogeneous hardware. eaic_sg's involvement ties the initiative to the Singapore robotics ecosystem.
Vercel 出了个新工具，每次 git push 自动扫漏洞，发现问题用 Vercel Agent 一键修复，写代码的可以试试。
Runway 给 Claude Code 做了个插件，你说一句想生成什么，它就能把视频和音频的 API 集成代码直接写好，做 AI 应用可以试试。
The bigger win is not just restore speed but that indexes come back pre-built, so recovered collections are query-ready without a rebuild window that can stretch much longer on large datasets. Teams running disaster-recovery drills or migrations on Milvus 3.0.1 or later should switch to snapshot backups via the milvus-backup tool. Compared with binlog restores, snapshots eliminate the post-restore index rebuild wait entirely.
一个开源的本地富文本编辑器，用起来跟飞书几乎一样，100 万字打字都不卡，还专门做了让 Claude Code 这类 Agent 直接读写文件的设计。
品牌在 Muse、Dots 这类购物智能体里看不见自家被推荐情况，Enrich Labs 的 ACO 专门补这个盲区，只先开放 100 个品牌。
用 Codex 存的知识能在 Claude Code 里读出来，多个智能体共用一个记忆库，不用重复存，做智能体的可以看看。
Soniox 出了个语音转文字模型，最妙的是能把智能体说的话当上下文传进去，识别用户回复更准，还能自动滤掉智能体自己的声音。
OpenAI 又在 ChatGPT 里埋了个 Family Basic 套餐，被网友从结账配置里挖出来了，可能是全家共用的新方案，价格还没公布。
OpenAI 的 Codex 在 computer use 上把 Claude Code 甩开了，一个开发者的真实使用感受，做编程工具选型可以参考。
Parag Agrawal 创办的 Parallel 和 CoreWeave 打通了，智能体不用自己去搜，被信息唤醒就能干活，一个设置就能开。
Using one shared embedding space lets teams trade off cost and quality per workload without re-embedding an entire index when they switch models. Search and RAG engineers can index corpora once with Pro and serve high-traffic queries on Fast. Compared with single-tier embedding offerings from OpenAI or Voyage, this two-tier setup reduces operational coupling between indexing and serving.
Listing Embed 5 and Rerank 5 on OpenRouter means teams can combine embedding and reranking with models from other providers under one bill, simplifying vendor management for RAG pipelines. Developers evaluating retrieval stacks can benchmark Cohere's models against alternatives like Voyage or Jina without new contracts. Pricing follows OpenRouter's existing Pro and Fast tier structure.
Living Brain 做了个跨模型共享记忆层，ChatGPT 和 Claude 能共用同一个记忆库，LoCoMo 跑到 92.2%，token 还省 22 倍多。
Kling 4.0 Flash 这次全量开放了，网站和 API 都能用，还能装进 Premiere Pro、After Effects 里直接生成视频，做剪辑的可以试试。
Writing to Figma canvases, not just reading them, turns coding agents into design-editing participants and removes a manual handoff between design and code. Developers with view-only access benefit too, since the extension shows CSS, variables and layout without Dev Mode. Anyone using MCP-compatible agents can configure it in minutes and manage edits through task anchors and status bars.
Hermes Agent 桌面版自带浏览器，AI 自己点网页，你直接在页面上框选留言，它带着截图和选择器去改，调前端 bug 挺顺手。
The setup shows a division of labour — a large agent fleet for discovery and verification, a separate model for patching — that others could replicate for security audits of their own codebases. European labs like Mistral are positioning large-scale agent swarms as a differentiator, a useful comparison point against OpenAI and Anthropic security demos.
妙言这个纯本地 Markdown 应用加了 miao CLI 命令，AI 能直接帮你建笔记、搜内容，还能本地版本回滚，免费去玩玩。
Open WebUI v0.13.0 上线 Open Terminal，模型能直接在你桌面上开应用、开浏览器干活，一条 Docker 命令就能试。
Karpathy 用 Opus 5.5 加 Sonnet 分工流水线，两小时没写一行代码就上线了有 6454 个视频的网站，工作流细节都写出来了。
Only the latest 2000 items are listed; for older ones see the daily archive