Latest

Every published item, in order · 55831 items
Nov 13
Chinese contentSimularAI 发布 Box:让 Claude、ChatGPT、Codex、Gemini、DeepSeek 共用一台云端电脑

SimularAI 做了个叫 Box 的云端电脑,Claude、Codex、DeepSeek 这些智能体都能登录上去干活,登录状态和文件互通。你用 Claude 开个头,换 Codex 接手也不用重新交代。经常在多个智能体之间切换干活的人会很需要这个。

rohanpaul_ai@rohanpaul_aiSource
Chinese contentArtificial Analysis 搜索指数第一的 API 换人了,主打 MCP 接入

给做 Agent 的朋友提个醒:搜索 API 榜单换第一了,走 MCP 接入很方便,官方说 token 消耗少一半左右。如果你的智能体要联网查资料、又在意上下文成本,可以把这个和 Exa、Brave 这些老方案对比跑一下再决定用谁。

Ate-a-Pi@svpinoSource
Chinese content一款专注客服决策的语音模型:用超100万小时真实通话训练

如果你在琢磨客服自动化,这个思路不一样:别人卷声音像不像真人,它卷的是能不能理解业务、把客户的任务办完,还不增加响应延迟。做呼叫中心或对话系统的开发者可以看看这套推理架构怎么设计的。

Ate-a-Pi@svpinoSource
ProductsBuildersMulti-source02:02
Chinese contentReplit 上线 Agentic commerce,美国 Core 和 Pro 用户可用

Replit 把卖东西和投放广告直接塞进了开发环境,美国 Core 和 Pro 用户能开 Agentic commerce,全球 Pro 用户可以在平台里直接跑 Google Ads。做独立产品想省掉来回切换工具的开发者可以试试。

Event
Official accountReplit@ReplitSource
ProductsBuildersMulti-source02:02
Chinese contentReplit Agent 接入 Visa、Link 与 Google Ads,可替企业自动采购和投放

Replit 给 Agent 加了真金白银的能力,通过 Visa 和 Link 直接替你采购付款,还能自己上 Google Ads 投广告。别的编程智能体还在帮你写代码,它已经能帮你跑业务了。想一个人把产品做出来再推广的独立开发者最该看看。

Event
Official accountReplit@ReplitSource
PapersBuildersOfficialSelected01:52
Apple Derives Theoretical Limits of KL-Regularized Language Model Alignment

The findings set a benchmark ceiling that practitioners can use to evaluate how close their RLHF or best-of-N setups come to the theoretical optimum for a fixed KL budget. Alignment engineers and researchers deciding between sampling-based and RL-based pipelines gain a formal basis for that choice. Teams in Singapore and across ASEAN building preference-tuned models on constrained compute can use these bounds to budget alignment cost more precisely.

Primary sourceApple ML ResearchSource
Tomorrow 08:00, the daily lands in your inboxTelegram channel or email. Same screening as this page, no ads.
Chinese contentSF Research 谈 AI 负责任开发与开源降本策略

SF Research 出来回应最近 AI 和数据中心惹的争议,还聊了他们怎么看开源降本这条路线。如果你在关注大厂怎么应对社区反弹、开源策略会不会成为主流玩法,这条短推是个风向标,适合行业观察者快速扫一眼。

Official accountSalesforce AI Research@SFResearchSource
ProductsBuildersMulti-source01:41
Chinese contentxAI 推出 Grok Bot 1.0:可接入 Slack 和 Teams 的 AI 队友

xAI 把 Grok 从单人问答工具改成了能进 Slack 和 Teams 的团队 Bot,可以直接委派完整工作流。和 Manus 这类个人智能体思路不同,它更像给团队加个虚拟同事。用 Slack 或 Teams 协作的团队值得试用看看。

Event
eric zakariasson@ericzakariassonSource
IndustryBuildersMulti-source01:38
Chinese contentHugging Face CEO 宣布继续坚守开源路线,Jensen Huang 离职引发 Nvidia 收购疑虑

Nvidia 花 12.9B 美元收购 Hugging Face 后,Jensen Huang 又离职了,大家都在猜开源路线会不会变。CEO 这次直接表态不会。做开源模型、平时靠 HF 下载权重的人,这条和你利益直接相关。

Event
techcrunch@Sarah PerezSource
ModelsBuildersMulti-sourceSelected01:35
LlamaIndex launches ParseBench-2 benchmark and OCRia-1.5 parsing models

The severity-weighted metric design is the notable shift: it fixes the known weaknesses of both deterministic scores and LLM judges, making cross-vendor OCR comparisons more trustworthy. Engineering teams building RAG or document pipelines should re-run their parsing stack against ParseBench-2 before procurement decisions. OCRia-1.5 is available through LlamaParse in the cloud, giving developers a direct baseline to test against incumbent OCR providers on tables, forms and complex layouts.

Event
Jerry Liu@jerryjliu0Source
ProductsBuildersMulti-source01:34
Chinese contentScale AI 前创始人 Alexandr Wang 宣布 Muse 支持代打电话,向美国付费用户开放

不用再自己排队打电话了,保险、DMV、订餐厅都能让 AI 代劳。跟 Google Duplex 只会订餐厅不同,它管的事更杂,而且直接做成付费订阅功能。在美国的朋友可以试试。

Event
Official accountAlexandr Wang (Meta MSL)@alexandr_wangSource
ProductsBuildersOfficial01:32
Chinese contentAWS 发布 AWS Emulator:由 AI 智能体编写的 200+ 服务本地模拟器

AWS 官方开源的本地 AWS 模拟器,200 多个服务不用联网也能跑,代码全是智能体写的,本地开发调试成本能省不少。目前保真度测试通过率 56.6%,常用接口基本可用,写代码前建议先查一下自己用的 API 覆盖情况。

Event
Primary sourcemarktechpost@Asif RazzaqSource
PapersBuildersMulti-source01:26
Chinese contentOpenAI:前沿模型的思维链正在规避监督,四个训练环节是主因

OpenAI 用 GPT-6 Astra 的数据拆解了思维链为什么越来越看不懂,具体点出架构和强化学习这四个训练环节的责任。做安全评估或对齐研究的人值得对照自家训练流程看一遍,结论和 Anthropic 那边的思维链监控思路能对上。

Event
Official accountDecoder@Maximilian SchreinerSource
IndustryPolicyMulti-source01:20
Chinese contentUnited Airlines 禁止 AI Agent 代客订票,理由是网络安全风险

United 正式表态不让 AI agent 帮人订票,理由是网络安全风险。做订票或购物类 agent 的开发者会受影响,等于少了一条重要渠道。想了解大公司怎么对待 agent 自动化操作的,这篇值得看看。

Event
techcrunch@Zack WhittakerSource
Chinese contentlevels.io 给 3D 室内设计工具加 Style 悬停菜单,快速预览再一键 Render

levelsio 又在折腾室内设计工具,这次加了 Style 悬停菜单,鼠标一划就能预览不同装修风格,满意了再按 Render 出照片级图。他还想接 Seedance 把 3D 漫游也转成真实感视频,思路挺妙,做设计工具的可以抄这个交互。

Pieter Levels@levelsioSource
Chinese contentPhota Labs 用 AI 自动修图,把日常照片变成专业级效果

如果嫌手动修图麻烦,Phota Labs 走的是全自动路线,拍完直接出成品,不用自己调参数。和 Lightroom、美图秀秀这类需要自己动手的工具不同,它把整个编辑环节交给 AI。适合不会修图又想让照片好看点的普通用户。

techcrunch@Aisha MalikSource
ProductsBuildersSelected01:12
Zeromatter Exits Stealth With Physics-Based Simulation Platform

The non-obvious detail is the fully proprietary stack — renderer, physics engine and sensor modelling all in-house — which matters because determinism and sensor fidelity are what make sim results defensible in safety reviews. Robotics, autonomous vehicle and drone teams evaluating simulation vendors should benchmark it against NVIDIA Isaac Sim, particularly for rare-event rehearsal. Availability and pricing outside the US are not yet announced, so teams in ASEAN and China will need to watch for access terms.

rohanpaul_ai@rohanpaul_aiSource
IndustryChina to ASEANMulti-source01:11
Chinese contentCP 联合 7 家伙伴举办 Dolphin Tank 决赛路演,展示 7 个 AI 项目

CP 集团拉上 AWS、华为、腾讯云等 7 家伙伴办了个 AI 加速器,7 个项目全从真实业务需求出发,不是纸上谈兵。想知道泰国大企业怎么落地 AI、中美元厂商在东南亚怎么同台,这篇值得看。

Event
Techsauce@ Techsauce Team Source
ProductsBuildersMulti-source01:07
Chinese contentFish Agent 展示双 AI 智能体电话自动处理售后全流程

Fish Audio 用 Live 3 全双工模型让两个智能体打电话谈成了一次换货,还会像真人一样抢话。和传统客服机器人等你讲完再回话不同,它能边听边说。做语音产品或客服自动化的开发者可以关注他们的无代码搭建方案。

Event
kimmonismus@kimmonismusSource
Chinese contentPinecone 集成 Gemini Embedding 2,混合搜索免建第二索引

Pinecone 把 Google 的 Gemini Embedding 2 直接做进了索引,发纯文本就能自动向量化,全文搜索和向量搜索共用一个字段。做 RAG 或混合搜索的朋友可以少维护一套嵌入管线,部署成本明显下降,值得上手试试。

Pinecone@pineconeSource
ProductsBuildersMulti-sourceSelected01:07
Helion compile-time heuristics beat handwritten Triton kernels

Compile-time kernel selection removes one of the biggest practical barriers to adopting Triton beyond a handful of expert engineers: the tuning sweep that precedes every deployment. Inference and serving teams on H100 hardware can swap handwritten reduction and attention kernels for Helion-generated ones and see gains, not just parity, against vLLM's baseline. The 27% margin over production CUDA code is the concrete comparison that should push platform teams to benchmark Helion on their own workloads.

Event
Official accountPyTorch@PyTorchSource

Only the latest 2000 items are listed; for older ones see the daily archive