AI 资讯
来自 Twitter · Hacker News · GitHub 的 AI 行业动态
Google Chrome 发布 WebMCP 早期预览，让网站主动为 AI Agent 暴露工具接口
Google Chrome 团队发布 WebMCP 协议早期预览，网站可通过声明式 HTML 表单和命令式 JS API 向 AI Agent 提供结构化操作接口，取代传统 DOM 操作方式，适用于客服、电商、预订等场景。
@dotey
阅读 →
v0 新增智能规划功能：复杂应用和 Agent 构建更精准
Vercel 旗下 v0 推出智能规划功能，可自动检测复杂任务并进行策略规划，确保首次构建即准确到位，尤其适合复杂应用和 Agent 开发场景。
@v0
阅读 →
ElevenLabs Summit：ElevenCreative 让营销团队一天内完成从创意到完整广告
ElevenLabs 在伦敦峰会展示 ElevenCreative 平台，整合语音、音乐、音效、图像和视频生成，帮助品牌用 AI 快速制作工作室品质的广告内容。
@elevenlabsio
阅读 →
海螺 AI 创作者用 Hailuo 重新演绎花木兰
海螺 AI 展示创作者 Cat Jin 使用 Hailuo 制作的花木兰主题视频作品，铁骑金戈的中国古典风格令人惊艳。
@Hailuo_AI
阅读 →
vLLM 实测 DeepSeek R1 在 GB300 上的惊人性能：预填充提速 8 倍
vLLM 发布 DeepSeek R1 在 NVIDIA GB300 上的性能数据：单 GPU 预填充 22.5K TGS、解码 3K TGS，相比 Hopper 架构预填充提升 8 倍。DeepSeek V3.2 仅需 2 块 GPU 即可运行。
@vllm_project
阅读 →
vLLM 首日支持 MiniMax M2.5：SWE-Bench 80.2%，速度媲美 Opus 4.6
vLLM 宣布首日支持 MiniMax M2.5 模型，该模型在 SWE-Bench Verified 达 80.2%、BrowseComp 达 76.3%，基于 20 万+真实环境 RL 训练，推理速度比 M2.1 快 37%，已在 NVIDIA GPU 上验证。
@vllm_project
阅读 →
通义千问发布 AI Slides：Qwen3 Agent + Qwen-Image 2.0 驱动智能 PPT 生成
阿里通义千问推出 AI Slides 工具，基于 Qwen3 Agent 和 Qwen-Image 2.0，支持从简单想法、长文本或文档一键生成精美演示文稿，自动完成调研、结构组织和视觉设计。
@Alibaba_Qwen
阅读 →
Kimi 推出 Allegro 计划：1T 参数多模态 K2.5 模型全面开放
Kimi 发布 Allegro 计划，所有套餐均可使用完整的 1T 参数原生多模态 K2.5 模型，不再因付费等级限制模型能力，支持 Agent 重度使用场景。
@Kimi_Moonshot
阅读 →
Google DeepMind：Project Genie 构建 AI 生成的虚拟世界
Google DeepMind 展示了 Project Genie 生成的虚拟世界，用户可以通过描述梦想场景，由 AI 自动构建可探索的 3D 世界。
@GoogleDeepMind
阅读 →
Anthropic：任命前微软 CFO Chris Liddell 加入董事会
Anthropic 宣布 Chris Liddell 加入董事会，他拥有 30 年领导经验，曾任微软和通用汽车 CFO，以及特朗普政府副幕僚长。
@AnthropicAI
阅读 →
Kimi：开源推理框架 Mooncake 诞生于月之暗面与清华合作
Kimi (月之暗面) 分享 Mooncake 开源项目的起源故事，该框架源于与清华大学的合作研究，旨在解决大规模模型服务中的内存墙问题，已发展为社区驱动的项目。
@Kimi_Moonshot
阅读 →
Figure AI：展示接近人手灵活度的机器人手部能力
Figure AI 发布视频展示其机器人手部操控能力，目标是达到与人手同等的灵活性和精细度。
@Figure_robot
阅读 →
OpenAI：GPT-5.2 在理论物理中推导出新成果
OpenAI 宣布 GPT-5.2 在理论物理领域推导出新结果，与普林斯顿高等研究院、范德堡大学、剑桥和哈佛等机构合作发表预印本，证明胶子相互作用在特定条件下可以发生。
@OpenAI
阅读 →
宇树科技：具身智能模型已在自家工厂实际部署制造机器人
宇树科技展示基于 UnifoLM-X1-0 具身 AI 模型的工厂实际部署场景，让 AI 模型直接参与机器人生产制造，实现 AI 造机器人。
@UnitreeRobotics
阅读 →
Anthropic：与 CodePath 合作，将 Claude 带给 2 万+大学生
Anthropic 宣布与美国最大的大学计算机科学项目 CodePath 合作，将 Claude 和 Claude Code 引入社区学院、州立大学和 HBCU 的两万多名学生。
@AnthropicAI
阅读 →
比我能想象的最糟噩梦还要糟糕
Elon Musk 在推文中发表感慨，称当前情况比他能想象到的最坏噩梦还要糟糕，引发广泛关注，浏览量超 1290 万。
@elonmusk
阅读 →
[D] The AI training market is broken. Here's why.
<!-- SC_OFF --><div class="md"><p>$10.5B industry, yet 94% of companies say employees lack AI skills (Gartner 2025).</p> <p>Why are we selling courses when we need assessments?</p> <p>On one hand there&#39;s providers that offer courses for up...
[D] Opinion required: Was Intelligence Just Gradient Descent All Along?
<!-- SC_OFF --><div class="md"><p>In medieval philosophy, thinkers debated whether intelligence came from divine reason, innate forms, or logical structures built into the mind. Centuries later, early AI researchers tried to recreate intelligence through symbols and...
[P] Graph Representation Learning Help
<!-- SC_OFF --><div class="md"><p>Im working on a Graph based JEPA style model for encoding small molecule data and I’m running into some issues. For reference I’ve been using this paper/code as a blueprint: <a...
The Evolution of Categorization During the era of AI Programming [D]
<!-- SC_OFF --><div class="md"><p>TL;DR -</p> <p>Hypothetically If the majority of code written is eventually generative, does this mean that the field of categorization will stagnate? If yes, does this have real implications; what if the future bottle neck...
[D] CVPR Score stats
<!-- SC_OFF --><div class="md"><p>Are the stats for the scores in paper copilot weighted by confidence?</p> <p>FYI - current CVPR stats: <a...
[D] Is a KDD publication considered prestigious for more theoretical results?
<!-- SC_OFF --><div class="md"><p>I do work at the intersection of ML and exact sciences and have some quite technical results that I submitted to KDD because they had a very fitting new AI for science track and all other deadlines were far away. Slightly hesitating now...
[P] A library for linear RNNs
<!-- SC_OFF --><div class="md"><p>Hi everyone, in the past few months, a few of my friends and I have developed this library containing implementation of several popular Linear RNNs, with accelerated kernels for inference and training (similar to mamba). All in PyTorch....
[D] Conformal Prediction vs naive thresholding to represent uncertainty
<!-- SC_OFF --><div class="md"><p>So I recently found out about conformal prediction (cp). I’m still trying to understand it and implications of it for tasks like classification/anomaly detection. Say we have a knn based anomaly detector trained on non anomalous samples....
[P] ML training cluster for university students
<!-- SC_OFF --><div class="md"><p>Hi! I&#39;m an exec at a University AI research club. We are trying to build a gpu cluster for our student body so they can have reliable access to compute, but we aren&#39;t sure where to start.</p> <p>Our goal is to...
[D] We scanned 18,000 exposed OpenClaw instances and found 15% of community skills contain...
<!-- SC_OFF --><div class="md"><p>I do security research and recently started looking at autonomous agents after OpenClaw blew up. What I found honestly caught me off guard. I knew the ecosystem was growing fast (165k GitHub stars, 60k Discord members) but the actual...
Launch HN: Omnara (YC S25) – Run Claude Code and Codex from anywhere
Hey y’all, Kartik, Ishaan, and Christian from Omnara (<a href="https://www.omnara.com/">https://www.omnara.com/</a>) here. We’re building a web and mobile agentic IDE for Claude Code and Codex that lets you run and interact with coding agents from anywhere. Omnara lets...
Anthropic raises $30B in Series G funding at $380B post-money valuation
Anthropic raises $30B in Series G funding at $380B post-money valuation
Tell HN: Ralph Giles has died (Xiph.org| Rust@Mozilla | Ghostscript)
It's with much sadness that we announce the passing of our friend and colleague Ralph Giles, or rillian as he was known on IRC.<p>Ralph began contributing to Xiph.org in 2000 and became a core Ghostscript developer in 2001[1]. Ralph made many contributions to the royalty-free media ecosystem,...
An AI agent published a hit piece on me
Previously: <i>AI agent opens a PR write a blogpost to shames the maintainer who closes it</i> - <a href="https://news.ycombinator.com/item?id=46987559">https://news.ycombinator.com/item?id=46987559</a> - Feb 2026 (582 comments)