每日简报

2026-10-03

← 历史归档

Panniantong/Agent-Reach

Python · ★ 88,610 · 🍴 7,800 · 📈 696 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 给 AI agent 装上一双能看遍整个互联网的眼睛。作者把 Twitter、Reddit、YouTube、GitHub、B 站、小红书等平台的读取与搜索能力封装成统一 CLI 工具,无需申请任何 API key、没有接口费用。适合需要让 agent 自主抓取多平台实时信息的开发者,比如搭建舆情监控或内容聚合类智能体时用。

JuliusBrussee/caveman

Go · ★ 109,095 · 🍴 6,315 · 📈 209 stars today

🪨 why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.

中文介绍 针对编程 agent 的省钱「原始人语」方案,核心思路是用最少 token 表达同样意思。它提供一个可安装的 skill 与代理层,通过让 AI 用类似穴居人的简短说法交流,把 token 消耗砍掉约 65%。适合频繁调用 Claude Code、Codex 等 agent 且在意 API 账单和上下文占用的开发者。

obra/superpowers

Shell · ★ 294,455 · 🍴 26,318 · 📈 556 stars today

An agentic skills framework & software development methodology that works.

中文介绍 一套可落地的 agentic 技能框架与软件开发方法论。它把「技能(skills)」组织成可复用的能力单元,让 AI 编程 agent 按既定流程完成开发任务,而非零散提示。适合希望规范 agent 协作方式、把个人经验沉淀为可共享技能的工程师与团队使用。

DietrichGebert/ponytail

JavaScript · ★ 151,787 · 🍴 8,141 · 📈 1,435 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 像团队里最懒的资深工程师那样思考:能不做就不做,最好的代码是没写出来的代码。它通过技能或提示约束,引导 agent 避免过度设计、减少无谓代码改动和新增文件。适合厌倦 agent 动不动大改一堆代码、追求最小改动落地的人。

pbakaus/impeccable

JavaScript · ★ 74,315 · 🍴 4,480 · 📈 722 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,用来提升 AI agent 在界面与视觉设计上的表现。它把设计准则、约束和判断标准结构化,让 agent 输出更符合审美的方案,而不是堆砌默认样式。适合用 AI 做前端、生成 UI 或需要让 agent 具备设计品味的开发者与设计师。

mattpocock/skills

Shell · ★ 274,700 · 🍴 23,061 · 📈 955 stars today

Skills for Real Engineers. Straight from my .agents directory.

中文介绍 来自知名 TypeScript 教育者 Matt Pocock 的 agent 技能合集,直接取自他本人 .agents 目录的实战配置。内容面向真实工程场景,涵盖日常编码协作所需的技能定义。适合使用 Claude Code 等工具、想借鉴成熟工程师如何组织 AI 技能库的开发者参考。

NVIDIA/OpenShell

Rust · ★ 14,427 · 🍴 1,660 · 📈 594 stars today

OpenShell is the safe, private runtime for autonomous AI agents.

中文介绍 NVIDIA 推出的自主 AI agent 安全私有运行时。它为 agent 提供隔离的执行环境,让 agent 在受限沙箱中运行代码和调用工具,从而控制风险并保护数据隐私。适合企业在生产环境中部署自主 agent、又需满足安全与合规要求的场景,如内部自动化与代码执行任务。

coreyhaines31/marketingskills

JavaScript · ★ 52,410 · 🍴 7,880 · 📈 140 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 及各类 AI agent 的营销技能包,覆盖转化率优化(CRO)、文案撰写、SEO、数据分析和增长工程。它把营销方法论打包成 agent 可调用的技能,让非技术营销人员也能借 AI 自动执行常规营销任务。适合增长、市场团队及独立创业者使用。

heygen-com/hyperframes

TypeScript · ★ 55,893 · 🍴 5,038 · 📈 580 stars today

Write HTML. Render video. Built for agents.

中文介绍 用写 HTML 的方式渲染视频,专为 agent 设计。开发者像编写网页一样描述画面结构与动画,工具再将其渲染成视频文件。它省去了传统视频剪辑软件的操作门槛,适合需要批量、程序化生成视频内容的团队,也便于让 AI agent 自动产出视频素材。

mksglu/context-mode

TypeScript · ★ 25,029 · 🍴 1,798 · 📈 282 stars today

Context window optimization for AI coding agents. Sandboxes tool output (98% reduction), persists session memory, and enforces routing across 17 platforms via MCP + hooks.

中文介绍 针对 AI 编程 agent 的上下文窗口优化工具。它对工具输出做沙箱隔离处理,可减少约 98% 的 token 占用,同时持久化会话记忆,并通过 MCP 与 hooks 在 17 个平台上强制路由管理。适合长会话、多工具调用的 agent 工作流,缓解上下文爆满问题。

google/skills

Python · ★ 20,744 · 🍴 1,721 · 📈 39 stars today

Agent Skills for Google products and technologies

中文介绍 Google 官方提供的 Agent Skills 集合,覆盖 Google 各类产品与技术栈。它把调用 Google 服务、使用其技术的能力标准化为 agent 可识别的技能,方便 AI 编程助手直接接入。适合在 Google Cloud、Gemini 等生态中构建 agent 应用的开发者使用。

getsentry/sentry

Python · ★ 45,031 · 🍴 4,891 · 📈 16 stars today

Developer-first error tracking and performance monitoring

中文介绍 Sentry 是面向开发者的错误追踪与性能监控平台,帮助团队实时捕获线上异常、定位崩溃堆栈并分析性能瓶颈。它支持多种语言与框架,能与主流开发流程集成。适合各类应用的后端、前端及移动端团队做线上质量监控与问题排查,是业界常用的可观测性工具。

colbymchenry/codegraph

C · ★ 72,957 · 🍴 4,682 · 📈 98 stars today

Pre-indexed code knowledge graph, auto syncs on code changes, for Claude Code, Codex, Gemini, Cursor, OpenCode, AntiGravity, Kiro, CoPilot, and Hermes Agent — fewer tokens, fewer tool calls, 100% local

中文介绍 为编程 agent 预建代码知识图谱,并在代码变更时自动同步。它可接入 Claude Code、Codex、Gemini、Cursor、Copilot 等十余种工具,通过图谱检索减少 token 消耗和工具调用次数,且完全本地运行。适合大型代码库中追求低成本、高准确率的 agent 使用场景。

cursor/plugins

TypeScript · ★ 9,499 · 🍴 896 · 📈 163 stars today

Cursor plugin specification and official plugins

中文介绍 Cursor 官方发布的插件规范与官方插件集合。它定义了编辑器扩展插件应遵循的接口与结构,并提供可直接使用的官方示例,便于开发者按标准为 Cursor 开发或接入插件。适合希望在 Cursor 编辑器中扩展功能、或对接其插件生态的开发者参考。

mvschwarz/openrig

TypeScript · ★ 4,302 · 🍴 293 · 📈 683 stars today

Build your own network of agents from Claude Code, Codex and Pi: persistent teams with roles, shared context and owned work.

中文介绍 用于把 Claude Code、Codex 和 Pi 等 agent 组建成自己的智能体网络。它支持建立持久化的团队,为每个成员设定角色、共享上下文并分配各自负责的工作。适合需要多个 agent 协同完成复杂任务、而非单打独斗的开发与研究场景。

Effect-TS/effect

TypeScript · ★ 16,545 · 🍴 802 · 📈 80 stars today

Build production-ready applications in TypeScript

中文介绍 Effect 是一个用于构建生产级 TypeScript 应用的库,提供类型安全的错误处理、并发、依赖注入与资源管理等能力。它用函数式编程范式把副作用、异步和失败显式建模,减少运行时意外。适合对可靠性和可维护性要求较高的中大型 TS 后端与全栈项目。

pablostanley/yoinks

TypeScript · ★ 3,497 · 🍴 310 · 📈 623 stars today

yoink any video from your terminal. no shady ads.

中文介绍 在终端里直接下载任意视频,省去网页端满屏的诱导广告和弹窗。它把视频抓取做成命令行工具,在终端中输入地址即可完成下载,流程干净。适合习惯终端操作、需要批量或脚本化保存视频内容的用户。

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

👍 146

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 面向流式视频交互的流式视频 LLM 需在无法预知未来相关性的情况下保留证据,并在证据充分时即时响应。OneStreamer 联合学习与查询无关的可复用事实记忆,在不牺牲实时感知的前提下统一感知、记忆与主动响应,缓解实时性与长期记忆的矛盾。

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

👍 114

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 比较 on-policy 与 off-policy 学习对灾难性遗忘、参数更新稀疏性和泛化的影响时,已有 SFT 与 RL 对比混淆了多个变量。本文系统解耦 rollout 策略等因素,厘清 on-policy 带来收益的真正来源,为蒸馏动态提供受控实验依据。

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

👍 49

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)逐步解掩多个 token,但反向过程通常按位置因式分解,限制了少步采样质量,而少步正是扩散相对自回归的速度优势所在。E-MoE 提出增强型 MoE 架构用于非因式分解扩散语言模型,提升少步生成的样本质量。

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

👍 42

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 等使用的奖励级归一化虽保留组内相对信息,但不同目标学习进度不均。本文提出密度感知的奖励聚合方法,根据奖励稀疏程度调整聚合权重,改善多目标优化中的学习平衡。

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

👍 36

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 部分可观测下的去中心化多智能体路径规划依赖通信与专家数据训练的策略,但多个协调联合动作可行时易产生意图歧义。本文提出「去中心化大脑」框架,通过迭代式意图去噪联合精炼各智能体的动作,提升协调与避碰成功率。

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

👍 36

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,而是复用已有技能解决未训练过的任务,并从自身尝试中改进并保留所学。核心洞察是通用控制器已隐含大量可复用能力,InterEvolve 通过测试时进化奖励程序实现技能重组。

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

👍 33

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

中文介绍 真实第一人称视频中的具身任务多由工具中介完成,要求智能体在物理约束下跟踪物体与任务状态。EgoTools 提出以工具为中心的推理基准与评测,考察模型对工具使用流程、状态变化与约束的理解能力。

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

👍 24

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 智能体通过视觉反馈与动作原语控制机器人,但反复调用模型和冗余观察带来大量 token 开销。PyRUA-Lean 提出交互式代码执行框架,耦合反馈驱动的原语组合,在成功率提升 14% 的同时减少 65% 的 token 消耗。

Decoding Looped Transformers Better for (Almost) Free

👍 18

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,每次循环都产生可解码同一 next token 的中间表示,但标准解码丢弃早期状态。本文利用早期循环计算量较小的特性改进解码,以近乎零成本提升循环 Transformer 的生成质量。

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

👍 18

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 生成常用体素潜变量与多阶段流程,先预测结构再合成局部几何,但会将连续表面碎片化为大量局部 token,抬高成本并削弱拓扑一致性。SILSA 提出滑窗切片潜表示,以保持拓扑的方式生成高分辨率 3D 资产。

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

👍 17

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 现有统一多模态模型要么把语言与量化图像都建模为离散 token,要么将离散语言预测与连续图像生成结合,前者引入视觉量化误差。Multimodal Flow 提出在嵌入空间中对语言与视觉进行统一连续流建模的全连续生成模型。

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

👍 17

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 真实企业数据分析需跨数十张表推理、做统计分析并据结果行动,而既有 text-to-SQL 基准只评查询生成,且答案键常被审计发现错误。Argo-Bench 面向企业级工作流评测数据智能体,覆盖多表推理与端到端任务执行。

CorrGRPO: Correlation-Normalized GRPO for Multi-Reward Learning

👍 16

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 在多奖励场景下将各奖励分量求和后按组内总奖励归一化,忽略奖励间相关性,导致优势估计偏差。CorrGRPO 提出相关性归一化的 GRPO,利用奖励间协方差结构校正优势,提升多奖励推理训练的稳定性与效果。

X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization

👍 12

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 和 RLVR 训练,均匀加权每个 token,忽略跨任务复现的子过程与层次结构。X-Tree 将可复用经验 token 化,让智能体自上而下从可复用例程规划,提升轨迹利用效率与泛化能力。

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

👍 12

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 虽有多模态推理能力,却未必掌握真实世界动态背后的物理一致性,现有基准评估碎片化、只考察孤立认知阶段。PhysVista 提出感知—推理—评估闭环的物理智能基准,系统检验模型对物理规律的理解与运用。

Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans

👍 11

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 自动家具布置可用于房产可视化与室内设计,但受限于真实数据稀缺,且需同时满足相互耦合的几何与功能约束。Architect-Ant 探索能否将专业布置知识注入模型,实现可编辑的户型图自动家具配置。

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

👍 10

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质量下降,现有多模态嵌入器靠数十亿参数弥补。Omni-Embed-Mini 以 0.9B 参数将文本、语音、音频、图像、视频和富视觉文档映射到同一空间,通过稠密蒸馏在绑定模态时不遗忘文本能力。

Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes

👍 7

On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains larg

中文介绍 on-policy 自蒸馏已有效提升语言模型推理,但在 MLLM 上应用仍有限。Where-OPD 提出空间引导的 on-policy 自蒸馏,利用合成场景为多模态大模型提供空间特权信息监督,改善其视觉空间推理能力。

Better Supervision Is Nearby: Neighborhood On-Policy Self-Distillation

👍 7

On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises student-sampled prefixes. Standard OPSD uses one fixed parameter setting at every state, but nearby settings may offer additional supervision. We find that

中文介绍 on-policy 自蒸馏(OPSD)用可见参考解的特权教师监督学生采样前缀,但标准做法在每个状态使用固定参数设置。本文发现邻近参数设置可提供额外监督,提出邻域 on-policy 自蒸馏,以「更好的监督就在附近」提升数学推理训练效果。

Benchmarking and Enhancing Skill-Level Memory for Partially Observable Robotic Manipulation

👍 7

Recent advances in robot learning have enabled manipulation policies to perform increasingly diverse tasks and generalize across environments. However, reliable execution often depends on hidden task states that cannot be determined from current observations alone, making interaction history essenti

中文介绍 机器人操作策略日益多样且可跨环境泛化,但可靠执行常依赖当前观测无法确定的隐藏任务状态,交互历史因而关键。本文提出面向部分可观测操作任务的技能级记忆基准与增强方法,评测并提升策略对历史信息的利用。

The Rise of the Automation Engineer

@jakeserval · 1.1K 粉丝 · 406.0K 阅 · 505 赞 · 1.1K 转

There is a lot of talk about the jobs AI takes away and very little about the ones it creates. While pessimism might be easier (and certainly gets more clicks), I’m more optimistic. From where I sit,

中文介绍 作者从乐观视角谈 AI 创造的岗位,而非取代的岗位,认为自动化工程师这一新职业正在崛起。文章批判了以悲观叙事博取点击的风气,从自身所处行业出发分析 AI 带来的新机会。

Why LLMs can't write well and how Sherpa fixes it

@RohanNayak2 · 13.7K 粉丝 · 230.7K 阅 · 514 赞 · 142 转

As we pivoted to AI-written shows, countless people told us this would kill Pocket FM. After six long months, I almost believed them. But then our ecosystem exploded, and not despite AI writing but

中文介绍 Pocket FM 转型 AI 撰写剧本后遭遇大量质疑,作者称熬过六个月后生态反而爆发。推文探讨 LLM 写作能力不足的原因,并介绍内部工具 Sherpa 如何补足短板,属于结合实操的产品案例复盘。

Safeguarding the AI Era of Biology: Watermarking the Building Blocks of Life

@pushmeet · 29.0K 粉丝 · 209.0K 阅 · 502 赞 · 88 转

Today, we announced SynthID Bio, a method to watermark and detect AI-designed proteins, the building blocks of life. We believe this is the first time humanity has successfully synthesised AI-designed

中文介绍 Google 发布 SynthID Bio,一种为 AI 设计蛋白质添加水印并实现检测的方法。作者称这是人类首次成功合成可被追踪的 AI 设计蛋白质,聚焦生物安全与 AI 治理方向,属于研究成果发布类推文。

History Rhymes: Large Language Models Off to a Bad Start?

@michaeljburry · 2.1M 粉丝 · 120.2K 阅 · 519 赞 · 84 转

This 145-year-old case study – presented at the Smithsonian Institute no less - provides a potentially devastating critique of today’s Large Language Models and the spending behind them. Michael Burry

中文介绍 Michael Burry 引用一份 145 年前曾在美国史密森尼学会展示的案例,类比当下大语言模型及其背后巨额投入,提出可能具有颠覆性的批评。属于借历史周期论证 AI 泡沫风险的看空观点。

here's how to make viral videos with Seedance 2.5

@0xROAS · 54.6K 粉丝 · 118.6K 阅 · 502 赞 · 56 转

Seedance 2.5 is one of the craziest "nano banana moments" of AI VIDEO in 2026. It's the model behind every viral video that you are seeing right now... and it's pretty simple to generate those video,

中文介绍 作者称 Seedance 2.5 是 2026 年 AI 视频领域的「nano banana 时刻」,目前大量爆款视频都由该模型生成,并给出用它制作病毒视频的操作方法。属于工具教程类分享,强调上手门槛低。

Own your AI, don’t let it own you

@0xSigil · 56.0K 粉丝 · 79.1K 阅 · 509 赞 · 72 转

6 months ago, the #1 AI was Claude Opus 4.6. Our model, Underdog 27B beats it while running on your Mac or iPhone. Most AI will run on your device, not in data centers. Intelligence that recently

中文介绍 作者称自家 27B 模型 Underdog 在 Mac 或 iPhone 本地运行即可超越六个月前的头号模型 Claude Opus 4.6,主张未来多数 AI 会跑在个人设备而非数据中心,属端侧智能方向的产品宣传。

We can and must solve alignment

@eric_ho · 4.4K 粉丝 · 66.2K 阅 · 505 赞 · 72 转

In San Francisco, it feels like the eve of the singularity. Yet, a walk through the city looks surprisingly mundane, with driverless cars and rolling fog equal parts of the scenery. The world feels

中文介绍 作者在旧金山感受到「奇点前夜」的氛围,但街头景象依旧平常,无人驾驶汽车与雾气并存。借此论证 AI 对齐问题必须且可以被解决,属于对齐议题的观点性短文。

how to build agentic systems for knowledge work

@arscontexta · 13.6K 粉丝 · 41.4K 阅 · 510 赞 · 45 转

the goal is to build software around your agents that gives them the context, tools and methods for a particular kind of work @garrytan put it like this: but what are people really trying to achieve

中文介绍 讲如何为知识工作构建 agent 系统:核心是围绕 agent 搭建软件,为特定类型的工作提供上下文、工具和方法。引用 Garry Tan 的观点,强调先想清楚人们真正要达成什么目标,再谈 agent 设计。

Why Instinct feels Magical

@yashhsm · 27.2K 粉丝 · 35.7K 阅 · 516 赞 · 24 转

how instinct, a $10b startup works - by reverse-engineering its harness: single-threaded chat, memory, computer use, background tasks etc One month ago, i got an invite for instinct my first

中文介绍 作者以逆向工程方式拆解估值 100 亿美元的初创公司 Instinct 的 harness 架构,涵盖单线程对话、记忆、computer use、后台任务等设计。基于其拿到内测邀请一个月后的使用体验,属产品拆解类推文。

DeepSeek Harness for Desktop Is Here

@DeepSeekHarness · 989 粉丝 · 35.4K 阅 · 1.0K 赞 · 69 转

Today, we’re releasing the DeepSeek Harness v0.2 preview, with a desktop app available for macOS and Windows. You can install it here and follow the instructions to get started. This release comes

中文介绍 DeepSeek Harness 发布 v0.2 预览版,提供 macOS 与 Windows 桌面应用,用户可按指引安装上手。属于官方产品发布公告,推文未展开具体功能细节。

Creative prompting with Gemini Omni in Google Flow

@FlowbyGoogle · 73.2K 粉丝 · 23.0K 阅 · 539 赞 · 51 转

To help get you started exploring Gemini Omni Flash in Google Flow, here is a comprehensive guide featuring prompt templates and detailed ideas for your next creative generations. Whether you are

中文介绍 Google Flow 官方发布 Gemini Omni Flash 的创作提示词指南,包含成套 prompt 模板与具体创意方向,帮助用户上手新一代生成能力。属于官方教程类内容,面向创意生成场景。

September at OptimAI: Building the Next Layer of the Agentic Internet

@OptimaiNetwork · 98.4K 粉丝 · 10.7K 阅 · 513 赞 · 138 转

September marked an important step in OptimAI's evolution. Over the past month, we have been expanding the products people interact with while continuing to strengthen the infrastructure underneath

中文介绍 OptimAI 发布月度进展总结,称九月是关键一步:一边扩展用户可交互的产品,一边继续强化底层基础设施,定位为构建 agentic internet 的下一层。属于项目方例行月报性质的内容。

OpenAI Withheld GPT-6.1 Astra. It Exposed AI’s Permission Problem.

@ActionModelAI · 60.2K 粉丝 · 6.8K 阅 · 501 赞 · 439 转

OpenAI has chosen not to release GPT-6.1 Astra after the model failed to meet its safety standards. To be clear, GPT-6.1 Astra is not GPT-6.1 Sol, the model OpenAI released at DevDay. GPT-6.1 Sol has

中文介绍 OpenAI 因未通过安全标准而选择不发布 GPT-6.1 Astra,作者澄清 Astra 并非 DevDay 上发布的 GPT-6.1 Sol。推文借此讨论 AI 发布中的「许可问题」,即模型能否上线由谁决定,属安全治理话题。

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 系列模型选型指南,介绍初创公司如何选择模型、调节推理强度、优化提示词与技能、协调工具,并为生产环境准备流程。

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 科技评论指出企业 AI 已进入全面运营阶段,模型能力提升快于企业吸收速度,性能成本持续下降。全球 AI 投资预计 2026 年达 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 在领导 Meta Llama 系列模型后,Ahmad Al-Dahle 加入爱彼迎(Airbnb)推动 AI 转型,涉及团队产品开发方式及对房客的服务体验。

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 科技评论刊文认为大语言模型并不具备推理能力。作者以 2016 年 3 月首尔人机围棋赛第二局第 37 手为例,说明看似荒谬的落子曾被误读为推理。

[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:极简 AI 框架 Pi 发布 1.0 稳定版,同时推出 Pi Durable,并支持 TypeScript;另有 AIE 纽约活动消息。

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 对话 RLM 论文第一作者、MIT 博士生 Alex Zhang,谈及 Jev、博士阶段的挣扎与工具链(harness)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 Codex 和 GPT-5.6 构建技术并重塑工作流程,将交易验证耗时从 30 分钟缩短至 4 分钟以内,以扩展其资本市场专业能力。

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 撰文提出「永恒的互补」观点:先进 AI 对突破性想法背后的常规工作可能最为重要,执行环节或将决定下一轮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 美国连锁超市 Albertsons 使用 ChatGPT Enterprise 和 OpenAI API 提升团队效率,并让数百万顾客的购物体验更便捷,从内部重塑零售业务。

[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:谷歌 DeepMind 发布 Gemini 4 Argon,输出长度达 100 万,作为对 Astra 与 Fable 的回应;目前仅限「Fairwind 计划」中的政府用户和可信网络防御人员试用。

The Den frees up 10-15 hours a week to grow with ChatGPT Work

As it opens a new location, the social club prepares grant applications in 2 hours instead of 3 days and liquor-license materials in 3 hours instead of 4 days.

中文介绍 社交俱乐部 The Den 借助 ChatGPT Work 每周节省 10 至 15 小时,拨款申请耗时从 3 天缩短至 2 小时,酒类执照材料从 4 天缩短至 3 小时,以支持其新址开业。

市场总览

美股四大指数与权重股整体呈偏强技术结构:QQQ 现价 749.58 距 52 周高仅 -0.66%,MACD 红柱扩张,NVDA 与 MSFT 维持多头排列并触发金叉,SPY 虽现 MACD 死叉但仍在 SMA20/50/200 上方且距高点 -1.25%。中概板块技术面明显分化走弱:BABA、PDD、0700.HK 均处空头排列,RSI14 分别 38.4 / 36.1 / 40.6,0700.HK 更逼近 52 周低点,仅 JD 无 MACD 信号但 RSI 35.3 偏弱。加密方面,恐慌贪婪指数 67 属「贪婪」,总市值 2.89 万亿 USD、24h -3.34%,BTC 主导率 58.7%、ETH 11.3%;BTC、ETH、SOL 均维持多头排列但同步出现 MACD 死叉,RSI14 在 59-63 区间,价格距 52 周高回撤 33%-50%。商品外汇与宏观层面,黄金期货 RSI 34.7 偏弱并处空头排列,WTI 原油 RSI 47.7 中性而维持多头排列;美元指数 RSI 71.6、10Y 美债收益率 RSI 74.6 双双超买且逼近 52 周高,USDCNY 6.7 贴近 52 周低。整体看,股强、中概与贵金属弱、加密高位动能钝化、美元与美债收益率超买。

今日关注

QQQ 纳斯达克100 ETF (QQQ)
偏上行

现价 749.58 站上 SMA20 726.18 与 SMA50 715.66 之上,距 52 周高仅 -0.66%,呈现多头排列;RSI14 读数 61.5 仍属常态偏强,MACD 7.5833 高于信号线 6.0583,红柱延续,近一日 +1.33%、近五日 +1.14%,动量与技术结构同步向上。

NVDA 英伟达 (NVDA)
偏上行

现价 233.95 位于 SMA20 223.51、SMA50 217.77 与 SMA200 200.43 之上,多头排列完整;距 52 周高仅 -1.65%,近一日 +2.44%、近五日 +4.17% 为样本中较强;RSI14 60.1 未过热,MACD 3.0687 显著高于信号线 2.4546,动量柱为正。

^TNX 10Y 美债收益率 (^TNX)
中性

读数 5.28,已在 SMA20 5.05、SMA50 4.83、SMA200 4.45 之上形成多头排列,距 52 周高仅 -1.22%;但 RSI14 达 74.6 已进入超买区,MACD 0.1343 高于信号线 0.1163,近一日 +0.76%、近五日 +1.79%,动能强但超买读数提示短线钝化风险,方向偏震荡。

0700.HK 腾讯控股 (0700.HK)
偏下行

现价 421.2 处于 SMA20 433.7、SMA50 447.51、SMA200 498.88 之下,空头排列明确;RSI14 40.6 偏弱,MACD -3.9952 低于信号线 -3.9536 并维持死叉,近一日 -2.27%、近五日 -3.92%,距 52 周高 -38.29%,弱势结构尚未出现扭转信号。

PDD 拼多多 (PDD)
偏下行

现价 75.38 低于 SMA20 78.79、SMA50 83.93 与 SMA200 94.68,空头排列;RSI14 仅 36.1,接近超卖边缘但未触发反转,MACD -1.7665 略高于信号线 -1.8322 属弱化死叉状态,近一日 -3.28%、近五日 -3.61%,距 52 周高 -45.93%,下行趋势结构完整。

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

$101.92 -0.17%
5 日
+0.94%
距 52w 高
-0.3%
RSI(14)
71.6
趋势
多头
SMA 20 / 50 / 200
100.36 / 100.00 / 99.27
MACD / 信号
0.579 / 0.388
RSI 超买接近 52 周高多头排列

SPY

S&P 500 ETF

$769.64 +0.92%
5 日
+0.32%
距 52w 高
-1.2%
RSI(14)
49.2
趋势
多头
SMA 20 / 50 / 200
765.00 / 763.07 / 720.03
MACD / 信号
0.536 / 0.907
MACD 死叉 (1 天前)接近 52 周高多头排列

QQQ

Nasdaq 100 ETF

$749.58 +1.33%
5 日
+1.14%
距 52w 高
-0.7%
RSI(14)
61.5
趋势
多头
SMA 20 / 50 / 200
726.18 / 715.66 / 667.91
MACD / 信号
7.583 / 6.058
接近 52 周高多头排列

AAPL

Apple

$333.69 +0.20%
5 日
-0.66%
距 52w 高
-3.4%
RSI(14)
51.4
趋势
多头
SMA 20 / 50 / 200
331.62 / 322.18 / 288.86
MACD / 信号
4.083 / 5.073
MACD 死叉 (2 天前)多头排列

MSFT

Microsoft

$517.53 +0.90%
5 日
+3.94%
距 52w 高
-6.5%
RSI(14)
60.2
趋势
多头
SMA 20 / 50 / 200
501.30 / 485.01 / 432.62
MACD / 信号
7.416 / 7.356
MACD 金叉 (今天)多头排列

NVDA

Nvidia

$233.95 +2.44%
5 日
+4.17%
距 52w 高
-1.6%
RSI(14)
60.1
趋势
多头
SMA 20 / 50 / 200
223.51 / 217.77 / 200.43
MACD / 信号
3.069 / 2.455
接近 52 周高多头排列

GOOGL

Alphabet

$343.50 -0.17%
5 日
+0.33%
距 52w 高
-15.9%
RSI(14)
45.4
趋势
中性
SMA 20 / 50 / 200
342.57 / 343.85 / 338.74
MACD / 信号
-0.648 / -0.360
MACD 死叉 (今天)

TSLA

Tesla

$370.59 +4.45%
5 日
-1.94%
距 52w 高
-25.7%
RSI(14)
45.6
趋势
中性
SMA 20 / 50 / 200
365.15 / 346.57 / 393.71
MACD / 信号
1.278 / 3.707
MACD 死叉 (3 天前)

META

Meta

$728.08 +0.40%
5 日
-6.37%
距 52w 高
-6.6%
RSI(14)
61.1
趋势
中性
SMA 20 / 50 / 200
690.05 / 622.60 / 627.98
MACD / 信号
36.037 / 35.275
加密恐慌贪婪
67
贪婪
加密总市值
$2.89 T
-3.34% / 24h
BTC 主导率
58.7%
ETH 11.3%
24h 成交量
$122.2 B
活跃币 21,783

BTC-USD

Bitcoin

$84,527.13 -0.38%
5 日
+0.08%
距 52w 高
-33.0%
RSI(14)
63.1
趋势
多头
SMA 20 / 50 / 200
82,047.09 / 78,123.50 / 71,371.39
MACD / 信号
1,993.599 / 2,132.533
MACD 死叉 (3 天前)多头排列

ETH-USD

Ethereum

$2,670.19 -1.31%
5 日
-0.62%
距 52w 高
-43.8%
RSI(14)
59.1
趋势
多头
SMA 20 / 50 / 200
2,626.93 / 2,468.00 / 2,113.98
MACD / 信号
70.364 / 81.010
MACD 死叉 (4 天前)多头排列

SOL-USD

Solana

$118.85 +0.38%
5 日
-2.63%
距 52w 高
-49.9%
RSI(14)
62.7
趋势
多头
SMA 20 / 50 / 200
113.07 / 102.89 / 85.67
MACD / 信号
5.218 / 5.539
MACD 死叉 (1 天前)多头排列

BABA

阿里巴巴 (BABA)

$105.85 -1.57%
5 日
-4.32%
距 52w 高
-44.6%
RSI(14)
38.4
趋势
空头
SMA 20 / 50 / 200
110.37 / 117.04 / 131.16
MACD / 信号
-2.239 / -2.132
MACD 死叉 (1 天前)空头排列

PDD

拼多多 (PDD)

$75.38 -3.28%
5 日
-3.61%
距 52w 高
-45.9%
RSI(14)
36.1
趋势
空头
SMA 20 / 50 / 200
78.79 / 83.93 / 94.68
MACD / 信号
-1.767 / -1.832
空头排列

JD

京东 (JD)

$25.78 -3.08%
5 日
-3.63%
距 52w 高
-28.8%
RSI(14)
35.3
趋势
中性
SMA 20 / 50 / 200
27.03 / 29.09 / 28.90
MACD / 信号
-0.627 / -0.660

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

$4,172.10 -0.72%
5 日
-3.45%
距 52w 高
-25.3%
RSI(14)
34.7
趋势
空头
SMA 20 / 50 / 200
4,334.82 / 4,366.47 / 4,553.41
MACD / 信号
-62.715 / -38.467
空头排列

CL=F

WTI 原油期货

$91.26 -1.73%
5 日
-1.24%
距 52w 高
-23.6%
RSI(14)
47.7
趋势
多头
SMA 20 / 50 / 200
96.05 / 88.58 / 82.39
MACD / 信号
0.564 / 1.823
多头排列

USDCNY=X

美元 / 人民币

¥6.70 -0.09%
5 日
-0.21%
距 52w 高
-6.8%
RSI(14)
34.5
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

以上均为公开行情数据计算出的技术指标读数复述,过去走势不代表未来表现,任何金叉、死叉、超买或超卖信号均不构成买卖建议,仅供技术指标解读参考。

How FlyDubai Passengers and Crew Averted Disaster After Pilot Was Stabbed During Flight

A badly wounded Indian pilot managed to unlock the cockpit door. Four Israeli men teamed up to overwhelm an attacker and avert disaster.

中文摘要 迪拜航空一航班在飞往特拉维夫途中发生袭击,一名印度籍飞行员受重伤后设法打开驾驶舱门,四名以色列男子合力制服袭击者,避免了灾难。

Cornell frat house rape accuser 'under siege' online, says lawyer

A lawyer for Jane Doe says she is "not doing well" amid attempts to uncover her identity.

中文摘要 美国康奈尔大学兄弟会强奸案原告的律师表示,当事人因网上有人试图查明其身份而「受到围攻」,目前状况不佳。

Dutch royal house helped slave trade to survive, study finds

In-depth study also says the royal house made huge profits from Dutch colonial policies in past centuries.

中文摘要 一项深入研究显示,荷兰王室曾帮助奴隶贸易存续,并在过去几个世纪的荷兰殖民政策中获得巨额利润。

Could El Niño mean there are no Atlantic hurricanes this year?

For the first time in three decades there has not been a single Atlantic hurricane in September, normally the most active month.

中文摘要 通常最活跃的9月,今年大西洋未出现任何飓风,为三十年来首次。报道探讨厄尔尼诺现象是否意味着今年大西洋不会出现飓风。

Iran Jails Hard-Line Lawmaker, Exposing Deep Rifts in Its Leadership

Hamid Rasaei, a vocal opponent of a diplomatic deal with the United States to end the war, was convicted of publishing false information with the intent to harm the Parliament speaker.

中文摘要 伊朗强硬派议员哈米德·拉萨伊因发布虚假信息、意图损害议会议长而被判入狱。他曾公开反对与美国达成结束战争的外交协议,此案暴露伊朗领导层内部深刻分歧。

Women given shorts at Oktoberfest to prevent upskirting

The step at the Munich festival comes after thousands of voyeuristic videos emerged on social media.

中文摘要 慕尼黑啤酒节向女性发放短裤以防止偷拍裙底。此前社交媒体上出现数千条偷窥视频。

The vanishing world of India's street magicians

India's street performers are struggling to survive as audiences shrink and public spaces become harder to access.

中文摘要 印度街头魔术师因观众减少、公共空间使用难度增加而难以维持生计,这一传统表演群体正逐渐消失。

Flydubai pilot recounts cockpit stabbing in call with Indian PM Modi

Smit Machchhar recalls midair disaster after co-pilot's attack on Flydubai flight to Tel Aviv.

中文摘要 迪拜航空飞行员斯米特·马奇查尔在与印度总理莫迪通话中,回忆了飞往特拉维夫航班上副驾驶袭击事件后的空中险情。

Bridal Shop Owner Who Put Child Onto a Coat Hook Cleared by Court

The “horseplay” was misguided, said the Canadian judge, but not malicious.

中文摘要 加拿大纽芬兰一名婚纱店店主将一名儿童挂在衣帽钩上,法院裁定其无罪。法官称这一「嬉闹」行为欠妥,但并非恶意。

What to Know About Fighting Between Ethiopia and Rebels in Tigray

The conflict is raising fears of a regional war in the Horn of Africa, drawing in foreign powers including Eritrea, Egypt the United Arab Emirates and Saudi Arabia.

中文摘要 埃塞俄比亚与提格雷叛军再度爆发战斗,引发对非洲之角地区战争的担忧,厄立特里亚、埃及、阿联酋和沙特等外部力量卷入其中。

Wall Street Week | Your Brain on AI, Rare Earth Race, College Enrollment Cliff

This week, National Economic Council Director Kevin Hassett says AI’s productivity boost may be much larger than official economic data currently indicate. And, AI companies are largely in agreement that the technology needs guardrails, but the ethical implications of AI could also create new opport

中文摘要 本期「华尔街一周」聚焦AI对大脑的影响、稀土竞赛与大学入学人数断崖。美国国家经济委员会主任哈塞特称,AI带来的生产率提升可能远高于官方经济数据目前显示的水平;AI企业普遍认同该技术需要护栏,但伦理影响仍有待探讨。

Former CEA Chair on Jobs Report, GDP Growth

Former CEA Chair Cecilia Rouse believes that we may be 'entering a new normal' of the jobs report, with the break even potentially settling in around 50,000. She speaks with Romaine Bostick on Bloomberg's "The Close." (Source: Bloomberg)

中文摘要 前白宫经济顾问委员会主席塞西莉亚·劳斯表示,就业报告可能正进入「新常态」,盈亏平衡的月度新增就业人数或稳定在约5万人左右。她就就业报告与GDP增长发表上述看法。

Morgan Stanley's Nike Analyst on quarterly earnings

Morgan Stanley's Alex Straton sees Nike sales and earnings softening as the company deals with inventory gluts in North America and China. She speaks with Romaine Bostick on Bloomberg's "The Close." (Source: Bloomberg)

中文摘要 摩根士丹利分析师亚历克斯·斯特拉顿认为,受北美和中国市场库存过剩拖累,耐克的销售与盈利正在走弱。

Pam Bondi’s Secret DOJ Email Address Revealed

A FOIA request has uncovered another little-known feature of life at the top of the Justice Department: attorneys general have used alternate email addresses to conduct government business for over a decade. This time around, Bloomberg News Senior Investigative Reporter Jason Leopold obtained the al

中文摘要 一项信息自由法申请揭露,美国司法部长帕姆·邦迪使用另类邮箱地址处理公务。彭博新闻调查报道显示,过去十多年间多位司法部长均曾使用替代邮箱开展政府事务。

EM Assets Find Relief as US Jobs Data Eases Rate Fears

An index tracking emerging-market currencies climbed to a session high after weaker-than-expected US jobs data prompted traders to scale back bets on another Federal Reserve interest-rate hike this month. Nonfarm payrolls increased by 29,000 in September after downward revisions to the prior two mon

中文摘要 美国9月非农就业仅增加2.9万人,且前值遭下修,弱于预期,交易员因此削减对美联储本月再度加息的押注。追踪新兴市场货币的指数升至盘中高点,新兴市场资产获得喘息。

CEA Chairman: Inflation coming down sufficiently fast

Chris Phelan CEA Chairman, sees inflation clearly easing and thinks it's doing so quickly enough. He speaks with Romaine Bostick on "The Close." (Source: Bloomberg)

中文摘要 美国白宫经济顾问委员会主席克里斯·费兰表示,通胀明显正在回落,且下降速度足够快。

Lukoil Loses Bid to Dismiss Suit Over Franchisee Fuel Pricing

The owners of almost 30 Lukoil PJSC filling stations in Pennsylvania and New Jersey can proceed with a lawsuit accusing the Russian oil giant of overcharging franchisees.

中文摘要 俄罗斯石油巨头卢克石油(Lukoil)请求驳回一项诉讼的动议未获支持。宾夕法尼亚州和新泽西州近30家卢克加油站的业主指控其向加盟商收取过高费用,该诉讼将继续进行。

Former Labor Secretary Rob Reich on Sept. Jobs Report

Robert Reich, former US Labor Secretary, discusses persistent inflation, very slow hiring, and wages that trail prices. He speaks with Romaine Bostick on Bloomberg's "The Close." (Source: Bloomberg)

中文摘要 美国前劳工部长罗伯特·赖克评述称,通胀持续存在,招聘非常缓慢,工资涨幅落后于物价。

Using Decision Science in Investing | Masters in Business with Omar Aguilar

Barry speaks with Omar Aguilar, CEO and CIO of Schwab Asset Management. They discuss his time at Financial Engines, where he worked under Nobel laureate William Sharpe before joining Schwab Asset Management in 2011. They discuss how his doctorate in decision sciences helps him build trust with clien

中文摘要 嘉信理财资产管理公司CEO兼首席投资官奥马尔·阿吉拉尔在访谈中回顾其职业经历:曾任职于Financial Engines,在诺贝尔奖得主威廉·夏普手下工作,2011年加入嘉信理财。他谈到决策科学博士学位如何帮助其构建投资决策。

Wall Street Tries to Live With 5% Yields as Market Cracks Grow

Wall Street has spent weeks trying to make peace with the great bond selloff. Friday offered some short-lived relief — along with a warning about the damage from stubbornly high yields across investment strategies of all stripes.

中文摘要 华尔街数周来试图适应大规模债券抛售。周五市场获得短暂喘息,但同时也收到警示:持续高企的5%收益率正对各类投资策略造成损害,市场裂痕不断加大。

Treasury Yields Climb, Adding to Weekly Gains | Closing Bell

Comprehensive cross-platform coverage of the U.S. market close on Bloomberg Television, Bloomberg Radio, and YouTube with Romaine Bostick, Katie Greifeld, Carol Massar and Tim Stenovec. (Source: Bloomberg)

中文摘要 美国国债收益率上升,进一步扩大本周涨幅。彭博电视、彭博电台与YouTube对美股收盘进行跨平台综合报道。

FAA Says Boeing 737 Max Computer Glitch Is Not Safety Issue

A Federal Aviation Administration review board says a software issue in certain Boeing 737 Max flight computers “is not a safety concern." Boeing shares rose on the news. George Ferguson of Bloomberg Intelligence has more. (Source: Bloomberg)

中文摘要 美国联邦航空管理局(FAA)审查委员会表示,波音737 Max部分飞行计算机的软件问题「不构成安全担忧」。消息公布后波音股价上涨。