每日简报

2026-10-05

← 历史归档

tester-army/e2e

TypeScript · ★ 3,111 · 🍴 117 · 📈 345 stars today

Next generation e2e testing framework for web and mobile apps.

中文介绍 面向 Web 与移动应用的下一代端到端(e2e)测试框架。相比 Playwright、Cypress 等现有方案,它尝试在真实用户交互路径上做统一测试,减少多端重复维护成本。适合需要覆盖浏览器和移动 App 关键流程的 QA 与前端团队。

pbakaus/impeccable

JavaScript · ★ 76,294 · 🍴 4,549 · 📈 1,171 stars today

The design language that makes your AI harness better at design.

中文介绍 一套设计语言规范,用于提升 AI 编程助手的设计输出质量。通过把设计原则、排版与配色规则注入 AI harness,让生成界面更符合专业水准。适合用 Claude Code、Cursor 等工具做前端或产品原型、又苦于 AI 审美一般的开发者。

coreyhaines31/marketingskills

JavaScript · ★ 53,074 · 🍴 7,917 · 📈 197 stars today

Marketing skills for Claude Code and AI agents. CRO, copywriting, SEO, analytics, and growth engineering.

中文介绍 面向 Claude Code 与 AI Agent 的营销技能包,涵盖 CRO 转化优化、文案写作、SEO、数据分析和增长工程等方向。把常用营销方法论封装成可被 Agent 调用的技能,让开发者或小团队用 AI 完成营销任务,降低对专业营销人员的依赖。

DietrichGebert/ponytail

JavaScript · ★ 154,882 · 🍴 8,322 · 📈 1,894 stars today

Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.

中文介绍 让 AI Agent 以“最懒的资深工程师”思维行事,强调能不写的代码就不写,优先复用与简化而非堆砌实现。通过约束 Agent 的编码倾向来减少过度设计和冗余代码。适合在意代码简洁、想让 AI 生成更克制方案的开发者。

earthtojake/text-to-cad

Python · ★ 16,860 · 🍴 1,742 · 📈 83 stars today

Give your agent CAD superpowers.

中文介绍 为 AI Agent 增加 CAD 建模能力,通过自然语言描述生成三维模型。把文本到几何参数的流程封装为 Agent 可调用的工具,让不会 CAD 软件的人也能快速产出可编辑模型。适合机械设计、3D 打印和工业原型场景。

Panniantong/Agent-Reach

Python · ★ 90,880 · 🍴 7,983 · 📈 980 stars today

Give your AI agent eyes to see the entire internet. Read & search Twitter, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu — one CLI, zero API fees.

中文介绍 为 AI Agent 装上“看整个互联网”的能力,用一个 CLI 即可读取和搜索 Twitter、Reddit、YouTube、GitHub、Bilibili、小红书等平台内容,且不依赖付费 API。适合做舆情监控、内容聚合和情报搜集的开发者与研究者。

getsentry/sentry

Python · ★ 45,383 · 🍴 4,904 · 📈 152 stars today

Developer-first error tracking and performance monitoring

中文介绍 面向开发者的错误追踪与性能监控平台,可捕获前后端异常、事务链路与性能瓶颈,并定位到具体代码和发布版本。支持自建部署与多种语言 SDK,广泛用于线上应用的故障排查、告警和性能优化。

calesthio/OpenMontage

Python · ★ 63,219 · 🍴 8,059 · 📈 245 stars today

World's first open-source, agentic video production system. 12 production pipelines, 100+ tools, 700+ agent skill and production-knowledge files. Turn your AI coding assistant into a full video production studio.

中文介绍 号称首个开源 Agent 化视频制作系统,内置 12 条制作流水线、100 多个工具以及 700 多份 Agent 技能与制作知识文件,可把 AI 编程助手变成完整的视频生产工作流。适合需要批量做视频的内容团队和技术型创作者。

pingdotgg/t3code

TypeScript · ★ 25,166 · 🍴 6,516 · 📈 490 stars today

中文介绍 仓库信息非常有限,缺少项目描述,无法确认其具体功能与用途。建议查看仓库 README 和源码了解它提供的能力。

caddyserver/caddy

Go · ★ 76,565 · 🍴 5,048 · 📈 24 stars today

Fast and extensible multi-platform HTTP/1-2-3 web server with automatic HTTPS

中文介绍 用 Go 编写的跨平台 HTTP/1、HTTP/2、HTTP/3 Web 服务器,最大特点是自动申请与续期 HTTPS 证书,配置简单,并通过模块化插件体系扩展能力。常用于反向代理、静态站点托管和自建服务,适合运维与后端开发者。

michael-denyer/pstack-claude

JavaScript · ★ 1,140 · 🍴 129 · 📈 232 stars today

Claude Code, Codex, Pi, OpenCode, Gemini, and Prime Agent versions of Poteto's pstack. Rigorous agent workflows with Cursor primitives translated for other harnesses.

中文介绍 把 Poteto 的 pstack 工作流移植到 Claude Code、Codex、Pi、OpenCode、Gemini 及 Prime Agent 等多个 AI Agent 环境,并将 Cursor 原语翻译为各 harness 可用形式。面向需要在不同编码助手间统一严格工作流的开发者。

addyosmani/agent-skills

JavaScript · ★ 101,216 · 🍴 10,616 · 📈 336 stars today

Production-grade engineering skills for AI coding agents.

中文介绍 由 Addy Osmani 整理的面向 AI 编码 Agent 的生产级工程技能集合,把代码审查、测试、重构等软件工程实践封装为可复用的技能。帮助开发者让 AI 助手在实际项目中遵循成熟工程规范,而非只做代码补全。

thedotmack/claude-mem

TypeScript · ★ 96,131 · 🍴 8,487 · 📈 628 stars today

Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More

中文介绍 为各类 AI Agent 提供跨会话的持久上下文能力:记录 Agent 在会话中的操作,用 AI 压缩提炼,再把相关上下文注入后续会话。解决 Agent 每次对话都从零开始、遗忘历史的问题,兼容 Claude 等工具,适合长期项目使用。

garrytan/gstack

TypeScript · ★ 135,164 · 🍴 20,092 · 📈 125 stars today

Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA

中文介绍 复刻 Garry Tan 本人的 Claude Code 配置,包含 23 个带明确立场的工具,分别扮演 CEO、设计师、工程经理、发布经理、文档工程师和 QA 角色。适合想借鉴成熟 AI 编码工作流、把多角色协作引入日常开发的团队。

OpenCut-app/OpenCut

TypeScript · ★ 92,138 · 🍴 9,082 · 📈 512 stars today

The open-source CapCut alternative

中文介绍 开源视频剪辑工具,定位为 CapCut(剪映海外版)的替代方案,提供时间线剪辑等常用能力,且不依赖闭源服务。适合希望自托管、注重数据隐私或想二次开发视频编辑功能的创作者与开发者。

antirez/ds4

C · ★ 23,447 · 🍴 2,254 · 📈 211 stars today

DeepSeek 4 Flash and PRO local inference engine for Metal, CUDA and ROCm

中文介绍 由 Redis 作者 antirez 推出的 DeepSeek 4 Flash 与 PRO 本地推理引擎,支持 Metal、CUDA 和 ROCm 三种后端,可在本地或自有 GPU 上运行模型。适合关注隐私、想脱离云端 API 部署大模型的开发者。

On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics

👍 175

On-policy learning has been argued to reduce catastrophic forgetting, produce sparser parameter updates, and improve generalisation. However, existing comparisons between supervised fine-tuning and reinforcement learning vary many factors simultaneously, making the contribution of rollout policy dif

中文介绍 针对 on-policy 与 off-policy 学习谁更优的争论,论文指出既有对比混淆了多个变量,无法确定 rollout 策略本身的贡献。作者系统解耦各因素,比较 SFT 与 RL 的蒸馏动态,厘清 on-policy 数据在减少灾难性遗忘、稀疏参数更新与泛化上的真实作用。

OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction

👍 165

Streaming video LLMs must retain evidence before its relevance to future tasks is known and respond when sufficient evidence becomes available. The challenge is to form reusable factual memory without compromising real-time perception. We introduce OneStreamer, which jointly learns query-independent

中文介绍 真实具身任务中的工具使用需在物理约束下追踪物体与任务状态变化。论文提出 EgoTools,面向第一人称视频中的工具中心推理,构建相关任务与评测,考察智能体在真实操作流程中对工具、状态及物理约束的理解与推理能力。

Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs

👍 64

Recent multi-agent LLM systems increasingly combine heterogeneous models for specialized agent roles. However, text-based communication requires each receiver to prefill shared context already processed by the sender. Reusing the sender's key-value (KV) cache avoids this redundancy, but prefill-free

中文介绍 异构多智能体 LLM 以文本通信时,接收方需重复 prefill 发送方已处理的共享上下文。论文研究免 prefill 的跨模型族 KV cache 传输,使不同家族的模型间可直接复用 KV cache,从而省去冗余前缀计算,降低多智能体协作的开销。

E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models

👍 60

Masked diffusion models (MDMs) generate sequences by progressively unmasking several tokens per denoising step, but their reverse process is typically factorized over positions, limiting sample quality in the few-step regime where diffusion's speed advantage over autoregressive decoding matters most

中文介绍 掩码扩散模型(MDM)每步并行解掩多个 token,但反向过程通常按位置分解,在少步采样这一速度优势最关键的场景下限制样本质量。论文提出 E-MoE,用增强的 MoE 结构对非分解扩散语言模型建模,改善少步生成的样本质量。

X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization

👍 59

Multi-step agents are trained on flat action streams: SFT and RLVR weight every token uniformly and ignore the sub-procedures that recur across tasks, the hierarchy that lets humans plan top-down from reusable routines. This structure sits unused, and flat training uses each scarce trajectory less f

中文介绍 多步智能体通常在扁平动作流上以 SFT/RLVR 训练,所有 token 等权,忽略了跨任务复现的子过程与层级结构,导致稀缺轨迹利用率低。论文提出 X-Tree,将可复用的经验分词化,实现层级化的经验复用,提升智能体的泛化效率。

EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric Videos

👍 57

Real-world embodied tasks, from everyday activities to professional procedures, require agents to act under physical constraints while tracking evolving object and task states. Tool use sits at the heart of such tasks, as many everyday and professional activities are tool-mediated. Understanding the

Make Sparse Rewards Count: Density-Aware Reward Aggregation for Multi-Reward RL

👍 57

Multi-reward reinforcement learning trains large language models to satisfy multiple behavioral objectives simultaneously. Reward-wise normalization, as used in GDPO, preserves reward-specific relative information within rollout groups, but different objectives can still exhibit uneven learning prog

中文介绍 多奖励 RL 训练 LLM 同时满足多个行为目标,GDPO 等按奖励归一会保留组内相对信息,但各目标学习进度仍不均衡。论文提出密度感知的奖励聚合方法,根据奖励分布密度动态调节聚合,使稀疏奖励也能有效贡献梯度,平衡多目标优化。

InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation

👍 55

We study test-time evolution for humanoid loco-manipulation: solving tasks that a controller was never trained for by repurposing its existing skills, improving from its own attempts, and retaining what it learns, without retraining. Our key insight is that a broad controller already holds much of t

中文介绍 研究人形机器人 loco-manipulation 的测试时进化:不重新训练,复用已有技能解决未训练任务。核心洞察是通用控制器已隐含大量能力。论文提出 InterEvolve,通过迭代进化奖励程序并保留所学,使机器人从自身尝试中持续改进。

Decentralized Master-Mind: Joint Action Refinement through Iterative Intent Denoising in Multi-Agent Pathfinding

👍 51

Decentralized multi-agent path finding (MAPF) with communication requires agents to reach individual goals without collisions under partial observability. Learnable policies trained on expert data provide an effective approach to this problem. However, when several coordinated joint actions are vali

中文介绍 通信式去中心化多智能体路径规划(MAPF)需在部分可观测下无碰撞到达目标。基于专家数据学习的策略易在多个有效联合动作间冲突。论文提出 Decentralized Master-Mind,通过迭代意图去噪精化联合动作,缓解多解冲突,提升协调成功率。

Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer Tokens

👍 45

Vision language model (VLM) agents can control robots through visual feedback and action primitives, but repeated model invocations and redundant observations incur substantial token overhead. We introduce PyRUA-Lean, an interactive code-execution framework that couples feedback-driven primitive com

中文介绍 VLM 机器人智能体通过视觉反馈与动作原语控制机器人,但重复调用模型与冗余观测带来大量 token 开销。论文提出 PyRUA-Lean,以交互式代码执行框架耦合反馈驱动的原语组合,在 GPT-6 Astra 上实现成功率提升 14%、token 减少 65%。

Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans

👍 43

Furnished floor plans support real-estate visualization, interior design, and architectural workflows, yet automatic furnishing remains challenged by limited real-world data and the need to satisfy interacting geometric and functional constraints. We ask whether professional furnishing knowledge can

中文介绍 自动为建筑平面图配置家具受限于真实数据稀缺与几何、功能约束交互。论文提出 Architect-Ant,探索能否从专业配置知识出发实现可编辑的自动家具布置,生成满足多重约束、可人工修改的家具布局方案。

Decoding Looped Transformers Better for (Almost) Free

👍 41

Looped Transformers achieve parameter efficiency by repeatedly executing a shared block across recurrent loops. Each loop yields an intermediate representation decodable for the same next token, yet standard decoding discards earlier states. Because earlier loops embody less computation, recurrence

中文介绍 循环 Transformer 通过重复执行共享块实现参数高效,每次循环的中间表示都可解码同一 next token,但标准解码丢弃了早期状态。论文利用早期循环计算量较少的特点,提出几乎零成本的解码策略,改进循环 Transformer 的生成质量。

SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation

👍 35

High-resolution 3D generation increasingly relies on voxel latents and multi-stage pipelines that first predict active structure and then synthesize local geometry. While effective, this design fragments continuous surfaces into many local tokens, inflates generation cost, and often weakens topologi

中文介绍 高分辨率 3D 生成常先预测活跃结构再合成局部几何,导致连续表面被切分为大量局部 token,成本高且拓扑易被破坏。论文提出 SILSA,用滑窗切片隐表示保持拓扑,在降低生成开销的同时产出拓扑完整的高分辨率 3D 内容。

CorrGRPO: Correlation-Normalized GRPO for Multi-Reward Learning

👍 30

Group Relative Policy Optimization (GRPO) is widely used to train reasoning language models, where it computes advantages by centering and normalizing rewards across rollouts of the same prompt. For multiple rewards, GRPO sums the reward components and normalizes the total reward by its within-group

中文介绍 GRPO 通过组内奖励归一计算优势,多奖励时直接把各分量求和后再归一,忽略了奖励间的相关性差异。论文提出 CorrGRPO,采用相关性归一化处理多奖励优势,使各奖励目标的优化更均衡,缓解目标间相互干扰。

Multimodal Flow: Unified Flow Modeling of Language and Vision in Embedding Spaces

👍 27

We present Multimodal Flow, a fully continuous generative model of language and vision. Most unified multimodal models either model both language and quantized images as discrete tokens or combine discrete language prediction with continuous image generation. The former introduces a visual quantizat

中文介绍 统一多模态模型多以离散 token 建模图像,引入视觉量化误差;另一类则将离散语言预测与连续图像生成拼接。论文提出 Multimodal Flow,在嵌入空间对语言与视觉做全连续流建模,避免量化损失,实现统一的多模态生成。

Argo-Bench: Evaluating Data Agents on Enterprise-Scale Workflows

👍 26

Real-world enterprise data science and analytics workflows require reasoning across dozens of tables, performing statistical analyses, and acting on the results. Established text-to-SQL benchmarks evaluate query generation alone, and audits have found their answer keys frequently wrong. Because real

中文介绍 企业级数据科学工作流需跨数十张表推理并执行统计分析,而既有 text-to-SQL 基准只评测查询生成,且答案键常被审计出错误。论文提出 Argo-Bench,面向真实企业规模工作流评测数据智能体,考察其多表推理与端到端分析能力。

Ego2Act: Evaluating Goal-Directed Manipulation in Egocentric Video Generation

👍 25

Video generation models are increasingly being explored as world simulators for embodied planning and learning. To do so effectively, these models must not only generate visually appealing frames, but also predict how environments dynamically evolve when executing goal-directed actions. While evalua

中文介绍 视频生成模型被用作具身规划的世界模拟器,但现有评测多关注视觉质量而非目标导向下的环境演化。论文提出 Ego2Act,面向第一人称视频中目标导向操作的生成评测,检验模型能否正确预测执行动作后环境的状态变化。

PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop

👍 23

Vision-Language Models (VLMs) have shown strong multimodal reasoning capabilities, yet whether they truly capture the physical consistency underlying real-world dynamics remains unclear. Existing benchmark paradigms often suffer from fragmented evaluation, focusing on isolated cognitive stages while

中文介绍 VLM 虽具强多模态推理能力,却未必掌握真实动力学背后的物理一致性;现有基准评价碎片化,只考察孤立认知阶段。论文提出 PhysVista,以感知—推理—评估闭环系统评测 VLM 的物理智能,更完整地衡量其物理理解水平。

Smaller Models, Better Rejects: Preference Distillation Scaling

👍 21

Preference distillation typically treats a teacher response as preferred and the student's own response as rejected. This assumes that self-generated failures are the most informative negatives and that rejects must come from a model at least as large as the student, making generation costly at scal

中文介绍 偏好蒸馏通常以教师回答为正例、学生自身失败为负例,假定规模不小于学生的模型才能提供有效负例,导致大规模生成成本高昂。论文研究偏好蒸馏的规模规律,发现更小模型也能提供优质拒绝样本,从而降低数据生成成本。

Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation

👍 19

Extending a text embedding model to new modalities typically degrades text retrieval quality, and existing omni-modal embedders compensate with multi-billion parameters. We present Omni-Embed-Mini, a 0.9B-parameter model that maps text, speech, audio, images, video, and visually-rich documents into

中文介绍 将文本嵌入模型扩展到新模态常导致文本检索性能下降,现有全模态嵌入模型靠数十亿参数弥补。论文提出 Omni-Embed-Mini,以 0.9B 参数经稠密蒸馏将文本、语音、音频、图像、视频与富视觉文档映射到同一空间,且不遗忘文本检索能力。

Designing MCP Gateway Uber's MCP Management Platform

@UberEng · 66.9K 粉丝 · 54.0K 阅 · 529 赞 · 59 转

Introduction Uber’s rapid adoption of AI agents has fundamentally changed how teams interact with code, data, and operational systems. Early ad-hoc integrations with the MCP (Model Context Protocol)

中文介绍 Uber 工程团队撰文介绍其内部 MCP Gateway 平台的设计。随着 AI agent 在团队内快速普及,早期零散的 Model Context Protocol 集成方式暴露出治理与扩展问题,文章从背景讲到如何构建统一管理平台,适合关注企业级 agent 基础设施的工程师参考。

VIBE MANUFACTURING IS HERE

@gregisenberg · 724.2K 粉丝 · 44.1K 阅 · 569 赞 · 53 转

For the FIRST TIME in HISTORY, you can describe a physical product in one sentence and, a week later, hold it in your hands, cut from metal and built to your exact measurements. That's VIBE

中文介绍 Greg Isenberg 提出「VIBE 制造」概念:用一句话描述一个实体产品,一周后就能拿到按精确尺寸用金属切割出来的成品。他将其描述为历史上首次出现的按需实体制造模式,属于对 AI 驱动硬件/制造新范式的前瞻观察。

Evals: how to know whether an AI system actually works

@sermakarevich · 549 粉丝 · 33.0K 阅 · 520 赞 · 64 转

An article for everyone who ships, buys, or signs off on software built on large language models (LLMs): engineers, product managers, and the CEO. Written in plain language, from the big picture down

中文介绍 一篇面向 LLM 应用相关从业者的评测入门文章,写给工程师、产品经理和 CEO——凡是交付、采购或签字确认 LLM 软件的人都算。作者用平实语言从整体图景讲到具体方法,讲清楚怎么判断一个 AI 系统是否真的有效。

How to build a content machine with Opus 5.5 (Full Course)

@VibeMarketer_ · 37.8K 粉丝 · 31.4K 阅 · 504 赞 · 59 转

i'm going to show you how to build a content machine with Opus 5.5 that turns your ideas, research and expertise into articles, posts and newsletters in your own voice. from finding your next topic to

中文介绍 一份完整教程,演示如何用 Opus 5.5 搭建内容生产流水线,把个人想法、调研和专业知识转成保持本人语气的文章、社交帖子与 newsletter,覆盖从选题到成稿的全流程。属于可直接套用的内容自动化工作流分享。

LIST OF HACKATHON FOR OCTOBER.

@onlyonealexia · 1.6K 粉丝 · 26.5K 阅 · 509 赞 · 53 转

A List of Web3, AI & Web2 Hackathons Compiled by @onlyonealexia There are a lot of hackathon lists that look impressive until you actually open the links. I compiled a list of ongoing and upcoming

中文介绍 博主整理了一份 10 月的 Web3、AI 与 Web2 黑客松清单。他吐槽很多黑客松列表看着唬人、点开链接就废,于是自己汇编了正在进行和即将开始的赛事。适合想找比赛练手或组队的开发者按需取用。

All games are open source now

@rehan_shei · 11.9K 粉丝 · 24.9K 阅 · 520 赞 · 49 转

A week ago I pointed Opus 5.5 at Terraria before bed. I woke up to a fully working, reverse engineered version of the game. Then I added a homing missile launcher, a tactical nuke and a bunch of new

中文介绍 博主睡前把 Opus 5.5 指向游戏 Terraria,醒来就得到一个可运行的逆向工程版本,之后还自己加了追踪导弹发射器、战术核弹等新内容。案例直观展示了当前模型在逆向与改游戏上能做到什么程度,话题性大于教程性。

Reinforcement Learning from scratch #1

@goyal__pramod · 11.4K 粉丝 · 24.7K 阅 · 504 赞 · 59 转

Now I am aware you must have been looking forward to the second part of my CUDA blog (if you haven't read it yet, check it out!), but hey! A man can have varied interests. And I believe if you are

中文介绍 博主开启新系列「从零实现强化学习」第一篇,插在 CUDA 博客系列之间。他自称兴趣广泛,默认读者会感兴趣,准备从基础讲起手写 RL。属于面向想打牢原理底子的开发者的连载入门教程开篇。

The Agent Said It Was Done. The Database Disagreed.

中文介绍 文章指出代理声称任务已完成但数据库状态并不一致,疑似微软相关技术博客探讨人工智能代理在实际执行任务时存在结果与真实系统状态不符的问题。

[AINews] not much happened today

a quiet day.

中文介绍 Latent Space 的 AINews 栏目当日发布简短更新,称当天人工智能领域新闻较少,整体较为平静,未提供具体事件或数据。

A model guide for the GPT-6 family

Learn how startups can choose GPT-6 models, tune reasoning effort, improve prompts and skills, coordinate tools, and prepare workflows for production.

中文介绍 OpenAI 发布 GPT-6 系列模型指南,介绍初创企业如何选择 GPT-6 模型、调整推理强度、优化提示与技能、协调工具,并为生产环境准备工作流。

Redefining enterprise intelligence with autonomous AI

Enterprise AI is no longer a future ambition. It is in full operational flight. Model capabilities are advancing faster than most organizations can absorb, while the cost of performance continues to fall. Globally, AI investment is set to reach $2.5 trillion in 2026, up 44% from the previous year. F

中文介绍 MIT 科技评论文章称企业人工智能已进入实际运营阶段,模型能力提升快于多数组织吸收速度,性能成本持续下降。文章提到 2026 年全球人工智能投资预计达 2.5 万亿美元,同比增长 44%。

Inside-Out AI: Rebuilding Airbnb Behind the Scenes and Across the Guest Experience

After leading Meta’s Llama models, Ahmad Al-Dahle is now transforming Airbnb with AI — from how its teams develop products to how it serves guests.

中文介绍 Latent Space 报道称,曾领导 Meta Llama 模型的 Ahmad Al-Dahle 现正以人工智能改造 Airbnb,涉及团队产品开发方式及面向房客的服务体验。

Don’t be fooled—LLMs don’t reason

On an afternoon in Seoul in March 2016, I watched a program I helped build put a stone on the fifth line of a Go board in what looked like a gift to its human opponent. Move 37 in game two of the five-game match looked so absurd that some commentators thought it was a…

中文介绍 MIT 科技评论文章认为大语言模型并不具备推理能力。作者回忆 2016 年 3 月在首尔观看一场围棋比赛第二局第 37 手,该落子曾被评论员视为荒谬之举。

[AINews] Pi 1.0, Pi Durable, and AIE NYC

the minimalist harness goes stable... and TypeScript!

中文介绍 Latent Space 的 AINews 报道称,极简式代理框架 Pi 1.0 与 Pi Durable 发布并进入稳定版本,同时支持 TypeScript,另有 AIE NYC 相关动态。

Academia is for Ambition — Alex Zhang, MIT

We catch up with RLM first author Alex Zhang, MIT PhD, on Jev, PhD masxing, and the future of harnesses.

中文介绍 Latent Space 采访 RLM 论文第一作者、麻省理工学院博士 Alex Zhang,谈及 Je 相关话题、攻读博士的体验以及代理框架的未来。

Chatham scales its capital markets expertise with OpenAI

Chatham Financial uses Codex and GPT-5.6 to build technology and redesign workflows, cutting trade validation from 30 minutes to under 4.

中文介绍 Chatham Financial 使用 OpenAI 的 Codex 和 GPT-5.6 构建技术并重塑工作流程,将交易验证时间从 30 分钟缩短至 4 分钟以内。

The eternal complement

Advanced AI may matter most for the routine work behind breakthrough ideas. Explore why execution could shape the next economy and the pace of progress.

中文介绍 OpenAI 发文讨论先进人工智能对突破性创意背后常规工作的作用,认为执行环节可能决定下一阶段经济形态与进步速度。

How Albertsons Companies is reimagining retail from the inside out

Albertsons Cos. is using ChatGPT Enterprise and the OpenAI API to help teams work faster and make grocery shopping easier for millions of customers.

中文介绍 Albertsons 公司正使用 ChatGPT Enterprise 和 OpenAI API,帮助团队加快工作速度,并为数百万顾客简化食品杂货购物流程。

[AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output

... but you can’t try it yet unless you are “government users and trusted cyber defenders in the Fairwind Program”

中文介绍 Latent Space 的 AINews 报道称,谷歌 DeepMind 发布 Gemini 4 Argon,作为对 Astra 和 Fable 的回应,支持 100 万输出,但目前仅限 Fairwind 计划中的政府用户与受信任网络防御者试用。

市场总览

美股主要指数偏强:QQQ 1 日涨 1.02%、距 52 周高仅 -0.66%,NVDA、MSFT 维持多头排列并分别出现 MACD 金叉与价格创新高邻近;但 SPY 与 AAPL 已现 MACD 死叉,Rsi14 均在 54 附近,属高位钝化。加密方面,恐慌贪婪指数 70 处贪婪区间,总市值 2.92T USD(24h -2.17%),BTC 主导率 59.3%、ETH 11.4%;BTC 现价 86509.36 涨 2.06%,RSI14 68.6,ETH 63.9、SOL 65.7,三者均在 SMA20 上方维持多头排列,但 ETH、SOL 的 MACD 已转死叉。中概股整体承压:BABA、PDD、0700.HK、JD 均收于 SMA20 之下,RSI14 落在 30.4 至 40.6 的弱势区,BABA 与 PDD 为空头排列叠加 MACD 死叉,JD 距 52 周低仅 5.19%。商品外汇分化:黄金期货 4180.7 低于 SMA20、SMA50、SMA200 呈空头排列,RSI14 36.4;WTI 原油 90.71 虽 5 日跌 2.04% 但仍在多头排列;美元指数 101.95 与 10Y 美债收益率 5.28% 双双 RSI14 超买(71.9、74.6)且接近 52 周高;VIX 15.31 走低并呈空头排列,风险偏好读数整体不弱。

今日关注

^TNX 10Y 美债收益率 (%)
偏上行

现价 5.28%,RSI14 达 74.6 处于超买区,且高于 20 日均线 5.05、50 日均线 4.83 与 200 日均线 4.45,多头排列明确;MACD 0.1343 上穿信号线 0.1163 维持金叉,5 日涨幅 1.79%,距 52 周高点仅 -1.22%。多项趋势与动量指标同向偏强。

DX-Y.NYB 美元指数 DXY
偏上行

现价 101.95,RSI14 71.9 同处超买区,价格站上 SMA20 100.5、SMA50 100.01 与 SMA200 99.29 呈多头排列,MACD 0.6025 高于信号线 0.431;5 日涨 0.74%,距 52 周高点仅 -0.26%,属高位强势形态,需留意超买读数。

NVDA 英伟达 (NVDA)
偏上行

现价 233.95,距 52 周高点仅 -1.65%,1 日涨 1.34%、5 日涨 3.95%;价格位于 SMA20 223.79、SMA50 218.28、SMA200 200.71 之上,多头排列,RSI14 62.9 偏强但未超买,MACD 3.4886 高于信号线 2.6614,动量与趋势一致向上。

BABA 阿里巴巴 (BABA)
偏下行

现价 105.85,跌破 SMA20 110.07、SMA50 116.87 与 SMA200 130.93,呈空头排列;RSI14 35.6 偏弱,MACD -2.3886 低于信号线 -2.1837 维持死叉,1 日跌 1.49%、5 日跌 3.54%,距 52 周高点 -44.17%,下行趋势指标较为一致。

0700.HK 腾讯控股 (0700.HK)
偏下行

现价 421.2,低于 SMA20 433.7、SMA50 447.51、SMA200 498.88,空头排列;RSI14 40.6 偏弱,MACD -3.9952 低于信号线 -3.9536 呈死叉,1 日跌 2.27%、5 日跌 3.92%,距 52 周高点 -38.29%,并标注接近 52 周低,弱势特征集中。

全部资产

^VIX

VIX 恐慌指数

$15.31 -6.59%
5 日
+2.96%
距 52w 高
-56.6%
RSI(14)
47.6
趋势
空头
SMA 20 / 50 / 200
15.92 / 15.78 / 18.06
MACD / 信号
0.005 / -0.027
MACD 金叉 (3 天前)空头排列

^TNX

10Y 美债收益率 (%)

$5.28 +0.76%
5 日
+1.79%
距 52w 高
-1.2%
RSI(14)
74.6
趋势
多头
SMA 20 / 50 / 200
5.05 / 4.83 / 4.45
MACD / 信号
0.134 / 0.116
RSI 超买接近 52 周高多头排列

DX-Y.NYB

美元指数 DXY

$101.94 +0.01%
5 日
+0.74%
距 52w 高
-0.3%
RSI(14)
71.9
趋势
多头
SMA 20 / 50 / 200
100.50 / 100.01 / 99.29
MACD / 信号
0.602 / 0.431
RSI 超买接近 52 周高多头排列

SPY

S&P 500 ETF

$769.64 +0.74%
5 日
-0.22%
距 52w 高
-1.2%
RSI(14)
54.5
趋势
多头
SMA 20 / 50 / 200
764.83 / 763.70 / 720.47
MACD / 信号
0.836 / 0.893
MACD 死叉 (2 天前)接近 52 周高多头排列

QQQ

Nasdaq 100 ETF

$749.58 +1.02%
5 日
+0.68%
距 52w 高
-0.7%
RSI(14)
65.4
趋势
多头
SMA 20 / 50 / 200
727.78 / 716.82 / 668.60
MACD / 信号
8.205 / 6.488
接近 52 周高多头排列

AAPL

Apple

$333.69 +1.02%
5 日
-2.16%
距 52w 高
-3.4%
RSI(14)
54.7
趋势
多头
SMA 20 / 50 / 200
331.89 / 322.42 / 289.16
MACD / 信号
3.807 / 4.820
MACD 死叉 (3 天前)多头排列

MSFT

Microsoft

$517.53 +0.92%
5 日
+0.26%
距 52w 高
-6.5%
RSI(14)
62.6
趋势
多头
SMA 20 / 50 / 200
501.67 / 487.72 / 432.83
MACD / 信号
7.804 / 7.445
MACD 金叉 (1 天前)多头排列

NVDA

Nvidia

$233.95 +1.34%
5 日
+3.95%
距 52w 高
-1.7%
RSI(14)
62.9
趋势
多头
SMA 20 / 50 / 200
223.79 / 218.28 / 200.71
MACD / 信号
3.489 / 2.661
接近 52 周高多头排列

GOOGL

Alphabet

$343.50 +1.56%
5 日
-0.12%
距 52w 高
-15.9%
RSI(14)
50.0
趋势
中性
SMA 20 / 50 / 200
342.62 / 344.37 / 338.92
MACD / 信号
-0.546 / -0.397
MACD 死叉 (1 天前)

TSLA

Tesla

$370.59 +4.65%
5 日
-0.41%
距 52w 高
-25.7%
RSI(14)
55.0
趋势
中性
SMA 20 / 50 / 200
364.87 / 347.58 / 393.18
MACD / 信号
1.834 / 3.332
MACD 死叉 (4 天前)

META

Meta

$728.08 +0.30%
5 日
-3.14%
距 52w 高
-6.6%
RSI(14)
61.5
趋势
中性
SMA 20 / 50 / 200
695.92 / 625.04 / 628.38
MACD / 信号
34.313 / 35.083
MACD 死叉 (今天)
加密恐慌贪婪
70
贪婪
加密总市值
$2.92 T
-2.17% / 24h
BTC 主导率
59.3%
ETH 11.4%
24h 成交量
$82.8 B
活跃币 21,849

BTC-USD

Bitcoin

$86,509.36 +2.06%
5 日
+3.45%
距 52w 高
-31.4%
RSI(14)
68.6
趋势
多头
SMA 20 / 50 / 200
82,859.16 / 79,028.36 / 71,501.76
MACD / 信号
2,016.513 / 2,078.171
多头排列

ETH-USD

Ethereum

$2,725.73 +1.43%
5 日
+1.83%
距 52w 高
-42.7%
RSI(14)
63.9
趋势
多头
SMA 20 / 50 / 200
2,647.90 / 2,500.99 / 2,118.42
MACD / 信号
65.945 / 75.657
多头排列

SOL-USD

Solana

$121.17 +1.27%
5 日
+1.78%
距 52w 高
-48.9%
RSI(14)
65.7
趋势
多头
SMA 20 / 50 / 200
115.02 / 104.69 / 85.95
MACD / 信号
4.940 / 5.333
MACD 死叉 (3 天前)多头排列

BABA

阿里巴巴 (BABA)

$105.85 -1.49%
5 日
-3.54%
距 52w 高
-44.2%
RSI(14)
35.6
趋势
空头
SMA 20 / 50 / 200
110.07 / 116.87 / 130.93
MACD / 信号
-2.389 / -2.184
MACD 死叉 (2 天前)空头排列

PDD

拼多多 (PDD)

$75.38 -1.46%
5 日
-2.82%
距 52w 高
-45.9%
RSI(14)
33.4
趋势
空头
SMA 20 / 50 / 200
78.47 / 83.77 / 94.50
MACD / 信号
-1.874 / -1.841
MACD 死叉 (今天)空头排列

JD

京东 (JD)

$25.78 -2.24%
5 日
-2.75%
距 52w 高
-28.7%
RSI(14)
30.4
趋势
中性
SMA 20 / 50 / 200
26.93 / 29.01 / 28.89
MACD / 信号
-0.661 / -0.660
MACD 死叉 (今天)

0700.HK

腾讯控股 (0700.HK)

HK$421.20 -2.27%
5 日
-3.92%
距 52w 高
-38.3%
RSI(14)
40.6
趋势
空头
SMA 20 / 50 / 200
433.70 / 447.51 / 498.88
MACD / 信号
-3.995 / -3.954
MACD 死叉 (今天)接近 52 周低空头排列

GC=F

黄金期货

$4,180.70 +0.44%
5 日
+0.30%
距 52w 高
-25.2%
RSI(14)
36.4
趋势
空头
SMA 20 / 50 / 200
4,319.53 / 4,368.47 / 4,552.61
MACD / 信号
-65.361 / -43.971
空头排列

CL=F

WTI 原油期货

$90.71 -0.44%
5 日
-2.04%
距 52w 高
-24.1%
RSI(14)
46.7
趋势
多头
SMA 20 / 50 / 200
96.01 / 88.60 / 82.56
MACD / 信号
0.315 / 1.520
多头排列

USDCNY=X

美元 / 人民币

¥6.70 -0.01%
5 日
-0.13%
距 52w 高
-6.8%
RSI(14)
38.9
趋势
空头
SMA 20 / 50 / 200
6.71 / 6.73 / 6.82
MACD / 信号
-0.006 / -0.007
接近 52 周低空头排列
风险提示

本报告仅为基于公开行情数据的技术指标解读,过去走势不代表未来表现;任何指标读数均可能随价格变化而失效,仅供技术指标解读参考,不构成任何投资建议或交易依据。

Yemen’s Saudi-backed government launches military drive to retake all Houthi-held territory

The announcement came hours after Iranian-backed Houthis reportedly cut a vital supply route to besieged city of Taiz Yemen’s Saudi-backed government has announced the start of a new military operation to retake all territory held by the Houthis, even as the militant group made fresh gains on the gr

中文摘要 也门沙特支持的政府宣布启动新一轮军事行动,目标夺回胡塞武装控制的所有领土。此前数小时,伊朗支持的胡塞武装据报切断了被围困城市塔伊兹的一条重要补给线。

Brazil presidential election 2026: first-round live results

Luiz Inácio Lula da Silva is standing against Flávio Bolsonaro in the first round of the presidential race The incumbent, Luiz Inácio Lula da Silva, is up against Flávio Bolsonaro, the son of the former president Jair Bolsonaro, who is in prison for plotting a military coup. The president in Brazil

中文摘要 巴西2026年总统选举首轮投票展开,现任总统卢拉·伊纳西奥·卢拉·达席尔瓦对阵前总统雅伊尔·博索纳罗之子弗拉维奥·博索纳罗。雅伊尔·博索纳罗因策划军事政变正在服刑。

Brazil Elections Head to Runoff Between Lula and Bolsonaro Son

Neither President Luiz Inácio Lula da Silva nor his top challenger, Flávio Bolsonaro, won a majority of votes in the first round. But for now, Mr. Bolsonaro is leading the race.

中文摘要 巴西总统选举进入第二轮,现任总统卢拉·达席尔瓦与其主要挑战者弗拉维奥·博索纳罗均未在首轮赢得多数票,目前博索纳罗暂时领先。

Iran’s Top Security Official Warns of Dire Economic Crisis

The rare admission from Mohsen Rezaei, the security chief, came weeks after the U.S. tightened sanctions and its naval blockade.

中文摘要 伊朗最高安全官员穆赫辛·雷扎伊警告该国面临严峻经济危机。此番罕见表态出现在美国收紧制裁并实施海军封锁数周之后。

Brazil's presidential race heads to Lula–Bolsonaro run-off as right gains ground

Now there are two: Brazil heads into a runoff after a first round that failed to fully energize voters, setting the stage for a more polarized contest that Washington is watching very closely.

中文摘要 巴西总统选举首轮未能充分激发选民热情,现进入卢拉与博索纳罗之子弗拉维奥的第二轮对决,选情更趋两极分化,华盛顿密切关注。

Brazil’s presidential race to go to second round as Flávio Bolsonaro leads incumbent Lula

Far-right challenger and son of disgraced ex-president Jair Bolsonaro will now face leftwing president in runoff vote Brazil’s presidential race will go to a second round after the far-right candidate, Flávio Bolsonaro, beat the incumbent leftwing president, Luiz Inácio Lula da Silva, but failed to

中文摘要 巴西总统选举将进入第二轮,极右翼候选人弗拉维奥·博索纳罗在首轮中击败现任左翼总统卢拉·达席尔瓦。弗拉维奥是身陷囹圄的前总统雅伊尔·博索纳罗之子。

Iran war live: Yemen fighting intensifies; Tehran says ready for US attacks

Yemeni government announces major offensive against Houthis, as Iran says it is ready to defend itself if US attacks.

中文摘要 也门政府宣布对胡塞武装发动大规模攻势,也门战事加剧;同时伊朗表示若美国发动攻击,已做好自卫准备。

Israel face spitting allegation as Ireland boss ‘delighted’ ties after over

Ireland manager says 'disciplinary committee will look at' allegations Israel player spat at one of his coaching staff.

中文摘要 爱尔兰队主教练表示,纪律委员会将调查有关以色列球员向己方教练组成员吐口水的指控。此前爱尔兰队对比赛以平局收场表示满意。

Rescuers suspend search for survivors after plane went missing off US coast

The aircraft lost communication with flight controllers after significantly dropping in altitude, according to flight data.

中文摘要 一架飞机在美国海岸附近失踪,飞行数据显示其高度大幅下降后与空管失去联系,救援人员已暂停搜寻幸存者。

How RAF Fairford in a Quiet Corner of England Became a U.S.-Iran Flashpoint

A mysterious call last Sunday near a military base used by the U.S. set in motion a huge counterterrorism response and days of uncertainty in Britain.

中文摘要 英国皇家空军费尔福德基地附近上周日接到一通神秘电话,引发大规模反恐响应,英国数日陷入不确定状态。该基地被美军使用,成为美伊关系的一个引爆点。

Mysterious luminous object burns in the skies above Libya

A mysterious object has been filmed flying over Tripoli and western Libya, followed by reports of a loud explosion.

中文摘要 一个神秘发光物体被拍到飞越利比亚首都的黎波里及西部上空,随后有报道称听到巨大爆炸声。

Man’s miraculous catch saves child after balcony fall in Egypt

Dramatic video shows a man in Egypt rushing and catching a child who fell from the second floor.

中文摘要 埃及一名男子在儿童从二楼阳台坠落时冲上前将其接住,戏剧性视频记录下这一救人瞬间。

Japan Should Keep Focus on Cheap Mideast Oil, Local Refiner Says

Japan should continue to rely on the Middle East for the bulk of its crude as that’s cost-effective, processor Taiyo Oil Co. said, pushing back against a government drive to diversify triggered by the US-Iran war.

中文摘要 日本炼油商太阳石油公司表示,日本应继续从中东进口大部分原油,因其具成本效益,反驳了美伊战争后政府推动的原油来源多元化政策。

The job interview question you don't have to answer

Experts explain what you should do if you are asked for your current salary during a job interview.

中文摘要 专家指出,求职面试中被问及当前薪资时,求职者并无义务回答,并给出应对建议。

Asian Shares Rise as Fed Bets Ease, Bonds Edge Up: Markets Wrap

Asian stocks rose after softer US jobs data reduced pressure on the Federal Reserve to keep raising interest rates. Oil erased early gains, while bonds crept higher.

中文摘要 美国就业数据走软缓解美联储加息压力,亚洲股市上涨,债券小幅走高,油价回吐早前涨幅。

Schneider Electric nears deal to buy software group PTC for $20bn

Acquisition would be French conglomerate’s largest and enhance its products focused on manufacturers

中文摘要 施耐德电气接近以200亿美元收购软件集团PTC,这将是该法国企业集团最大一笔收购,并强化其面向制造商的产品。

Schneider Said to Near Deal to Buy PTC for More Than $20 Billion

Schneider Electric SE is nearing a deal to acquire US-based engineering software developer PTC Inc. for more than $20 billion, according to a source familiar with the matter.

中文摘要 据知情人士透露,施耐德电气接近达成协议,以超过200亿美元收购美国工程软件开发商PTC公司。

Dealmaking slowdown threatens early end to M&A boom

Interest rate rises, AI anxiety and midterm elections cool animal spirits after record-breaking start to the year

中文摘要 在年初创纪录的并购开局后,利率上升、人工智能焦虑及中期选举令市场情绪降温,交易放缓或使并购热潮提前结束。

AkzoNobel agrees deal to sell its SE Asia decorative paint unit to Nippon Paint

Dutch group on verge of handing Japanese rival a small victory as it seeks to exit non-core markets ahead of Axalta merger

中文摘要 荷兰阿克苏诺贝尔同意将其东南亚装饰涂料业务出售给日本涂料,在寻求退出非核心市场并筹备与艾仕得合并之际,让日本竞争对手取得小胜。

The world is facing an almighty LNG crunch

In time it could bring about an equally disruptive glut

中文摘要 全球面临巨大的液化天然气供应紧张,未来可能演变为同样具有破坏性的供应过剩。

Supreme Court Opens Term With Climate Fight

Bloomberg Supreme Court Reporter Greg Stohr is on Bloomberg This Weekend previewing a new term shaped by the court’s growing emergency docket and a major climate case over whether Boulder, Colorado, can pursue state-law claims against oil companies. Speaking with hosts David Gura and Christina Ruffi

中文摘要 美国最高法院新任期开庭,紧急案件数量增加,并将审理科罗拉多州博尔德市能否依州法律起诉石油公司的重大气候案件。

Bloomberg This Weekend 10/4/2026

The news doesn’t stop when markets close. Hosts David Gura, Christina Ruffini and Lisa Mateo bring clarity, context and a bit of humor to the weekend’s biggest headlines, LIVE from New York. Joined by Ambassdor Alexander Yui, Taiwan's Representative to the United States, Kevin Gordon, Schwab, Head o

中文摘要 彭博周末节目由David Gura、Christina Ruffini和Lisa Mateo在纽约直播,梳理周末头条,嘉宾包括台湾驻美代表俞大㵢等。

The taxpayer-funded ADHD boom

The growing private market for assessments is leading to concerns over misdiagnoses and providers’ soaring revenues

中文摘要 纳税人资助的注意力缺陷多动障碍评估市场扩张,私人评估市场增长引发误诊担忧,相关服务提供商收入飙升。