← 首页
27JUL
AI前沿每日脉动
AI Frontier Pulse · 中英双语版 Bilingual Edition
2026.07.27 · 周一刊
11 位 Builder21 条推文1 期播客0 篇博客
ChatGPT WorkApplied AIOpen WeightsTokens per WattTrust & Access
Richard Liu · 2026
Curated by Richard Liu · Snapshot follow-builders-2026-07-27-v2
今日头条 · ChatGPT Work Becomes The Computer01 / 15
今天最强信号：一句自然语言任务开始跨越聊天历史、协作网站、群体决策、预订与邮件，AI 正从对话框变成新的计算入口。The strongest signal: one natural-language task now crosses chat history, a collaborative site, group decisions, reservations, and email. AI is becoming a new computing surface.
@samaX 原文14.7K ❤ · 471 RT · 1.4K 💬
Sam Altman：ChatGPT Work 把一句话变成完整旅行协作系统One Prompt Becomes A Full Travel Coordination System
Sam Altman 从手机发出一条任务：读取历史对话、为 9 人长周末筛选目的地、搭建协作网站、形成共识后预订，并在 Gmail 起草通知邮件。它完整跑通，说明 Work 的核心不是聊天，而是跨上下文、应用与人的长链条执行。
Sam Altman gave ChatGPT Work one phone prompt: use chat history, propose destinations for nine friends, build a coordination site, make reservations after agreement, and draft the Gmail message. It completed the chain, showing Work as cross-context execution rather than chat.
@samaX 原文14.7K ❤ · 471 RT · 1.4K 💬
Sam Altman：Work 指向一种新型计算机Work Points Toward A New Kind Of Computer
Sam Altman 直接把 ChatGPT Work 的意义提升到“新型计算机”：入口不再是应用图标与菜单，而是由意图驱动、能调用上下文和工具、持续推进任务的执行环境。
Sam Altman frames ChatGPT Work as a new kind of computer: an intent-driven execution environment that can use context and tools and keep advancing a task, rather than a collection of app icons and menus.
@samaX 原文5.5K ❤ · 188 RT · 510 💬
02 / 15
行业分析 · From Task Assistant To Computing Layer02 / 15
Thibault：议价、退订与比价都缩成手机上的一个提示Negotiation, Unsubscribing, And Shopping Collapse Into One Prompt
Thibault 表示，协商网络账单、批量退订垃圾邮件、寻找商品或活动优惠，都可从手机上一条提示启动；他每天让 Work 完成至少 20 件事。产品价值正在从回答质量转向被代办的任务数量。
Thibault says negotiating an internet bill, unsubscribing from spam, and finding a deal can all start from one phone prompt, and Work handles at least 20 things for him daily. Product value shifts from answer quality toward completed tasks.
@thsottiauxX 原文3.8K ❤ · 132 RT · 395 💬
OpenAI 的产品组织进入高度聚焦状态OpenAI Enters A Highly Focused Product Phase
Thibault 说自己从未见过 OpenAI 如此聚焦、如此高效运转。结合 Work 的快速扩展，这更像组织信号：模型、产品与执行团队正在围绕同一个计算入口收敛。
Thibault says he has never seen OpenAI this focused and humming. Alongside Work's expansion, it reads as an organizational signal: model, product, and execution teams are converging on one computing surface.
@thsottiauxX 原文3.4K ❤ · 61 RT · 472 💬
03 / 15
开放生态 · Open Weights And Native Delivery03 / 15
开放权重扩张模型生态，AI 代码翻译把 TypeScript 工具推向轻量原生分发；开放与性能正在同一条产品链上汇合。Open weights expand the model ecosystem while AI code translation moves TypeScript tools toward lightweight native delivery. Openness and performance are converging.
@rauchgX 原文1.3K ❤ · 65 RT · 38 💬
Vercel 联署开放权重与美国 AI 领导力倡议Vercel Co-signs The Open Weights Leadership Letter
Guillermo Rauch 宣布 Vercel 联署 Open Weights and American AI Leadership letter，强调开放源码、数据、协议与研究带来的复利，并将开放权重视为下一个前沿。开放生态正在从开发者偏好升级为产业战略。
Guillermo Rauch says Vercel co-signed the Open Weights and American AI Leadership letter, arguing that open source, data, protocols, and research compound progress. Open weights are becoming an industrial strategy, not merely a developer preference.
@rauchgX 原文1.3K ❤ · 65 RT · 38 💬
Vercel CLI TypeScript 被编译成 1.28MB 原生二进制Vercel CLI TypeScript Compiles To A 1.28MB Native Binary
Rauch 用 scriptc 将 Vercel CLI 的 TypeScript 编译为静态原生程序：二进制 1.28MB，平均启动开销 1.5ms，编译约 2.94 秒，还保留高度可读的 TypeScript。AI 翻译代码正在把脚本生态推向原生分发。
Rauch used scriptc to compile the Vercel CLI's TypeScript into a fully static native program: 1.28MB, 1.5ms mean startup overhead, and about 2.94 seconds to compile while preserving readable TypeScript. AI code translation is pushing script ecosystems toward native distribution.
@rauchgX 原文577 ❤ · 28 RT · 39 💬
04 / 15
产品设计 · Community, Applied AI, Research, Security04 / 15
YC Startup School 2026 由 Sam Altman 收官Sam Altman Closes YC Startup School 2026
Garry Tan 感谢 Sam Altman 为 YC Startup School 2026 做收官分享。模型公司与创业者社区的连接仍在加强，前沿能力会更快被翻译为新产品和新公司。
Garry Tan thanked Sam Altman for closing YC Startup School 2026. The connection between frontier labs and founder communities keeps tightening, accelerating how model capability becomes products and companies.
@garrytanX 原文592 ❤ · 8 RT · 42 💬
Aaron Levie：模型越强，Applied AI 层反而越重要Better Models Increase The Need For Applied AI
Aaron Levie 认为，智能本身不足以改造企业流程。真正落地需要连接系统和数据、设计人类决策节点、持续改善工作流，并满足监管要求；银行开户、法律审阅与生命科学都需要不同的 applied AI。模型越强，可自动化目标越大，这一层的机会也越大。
Aaron Levie argues that intelligence alone cannot transform enterprise processes. Teams must connect systems and data, design human decision points, improve workflows, and meet regulatory constraints. As models improve, automation becomes more ambitious and the applied-AI opportunity grows.
@levieX 原文321 ❤ · 41 RT · 58 💬
Dan Shipper 开始口述史式追踪 Codex 的诞生A Deep Oral History Of How Codex Happened
Dan Shipper 暂停日常工作一周，通过对 OpenAI 内部人士的深度访谈撰写 Codex 的完整历史，并计划沿途公开线索与发现。AI 产品史正在从发布公告进入组织决策、技术转折和人物选择的深层复盘。
Dan Shipper is taking a week to write a definitive history of Codex from deep interviews with OpenAI insiders, sharing breadcrumbs along the way. AI product history is moving beyond launch notes into decisions, technical turns, and people.
@danshipperX 原文246 ❤ · 6 RT · 15 💬
Amjad：攻击者更偏爱被补贴的闭源订阅Attackers May Prefer Subsidized Lab Subscriptions
Amjad 转述一位前 Anthropic 员工的观察：攻击者更倾向使用实验室大量补贴的 AI 订阅，而不是开放模型。风险讨论不能只盯开放权重，还要审视商业服务的价格补贴、访问控制与滥用监测。
Amjad relays an observation from a former Anthropic employee: attackers may prefer heavily subsidized lab subscriptions over open models. Risk analysis must examine commercial pricing, access controls, and abuse monitoring—not only open weights.
@amasadX 原文189 ❤ · 14 RT · 12 💬
05 / 15
深度主题 · Product And Engineering Signals05 / 15
没有博客并不等于没有深度：今天的深度来自工程、安全、组织与商业化信号在同一天汇合。No blog does not mean no depth: today's depth comes from engineering, safety, organization, and commercialization signals converging.
follow-buildersNo blog post in this snapshot
今日无新增博客：深度信号来自产品与工程现场No Blog Today, But Strong Product Signals
今天中央 feed 没有新增博客，但高互动推文已经形成清晰主线：产品入口、工程基础设施、模型治理和组织能力正在同时变化。
The central feed has no new blog post today, but the top tweets still form a clear arc across product surfaces, engineering infrastructure, model governance, and organization design.
follow-buildersSnapshot follow-builders-2026-07-27-v2
06 / 15
播客深度 · Compute Meets The Physical World06 / 15
PODCAST DEEP DIVE
AI 的下一轮竞争不只发生在模型层，也发生在冷却、电网、数据中心、供应链与每瓦 token 数构成的物理系统里。The next AI race is not only at the model layer, but across cooling, grids, data centers, supply chains, and tokens per watt.
OpenAI 算力负责人：任何新增算力都会被立刻消耗Every New Unit Of Compute Is Consumed Immediately
MAD Podcast 访谈 OpenAI 工业算力负责人 Sachin Katti：AI 需求持续远超算力供给，新增容量上线即被使用。OpenAI 正同时推进数据中心、电网投资、多云采购与定制芯片 Jalapeño，并以每瓦生成的 token 数作为关键效率指标。
The MAD Podcast interviews OpenAI industrial compute chief Sachin Katti. Demand still far exceeds supply, so new capacity is consumed immediately. OpenAI is pursuing data centers, grid investment, multi-cloud supply, and its Jalapeño chip, optimizing tokens per watt.
MAD PodcastYouTube 完整播客2026-07-16
07 / 15
播客理念 · Industrial Compute At Scale07 / 15
算力上线即被消耗New Compute Is Consumed Immediately
OpenAI 的需求持续超过供给，任何新增容量都会立刻被工作负载吸收；过早放慢建设通常会带来负面意外。
OpenAI demand continues to exceed supply, so every new unit of capacity is consumed immediately; slowing the build too early repeatedly creates negative surprises.
MAD PodcastYouTube 完整播客
关键指标：每瓦 token 数The Metric Is Tokens Per Watt
定制芯片 Jalapeño 围绕已知模型工作负载协同设计，目标是在电力约束下最大化每瓦可生成的 token。
The Jalapeño custom chip is co-designed around known model workloads to maximize tokens generated per watt under power constraints.
OpenAI ComputeYouTube 完整播客
数据中心必须新增电力Data Centers Must Add Power
OpenAI 表示建设数据中心时不会挤占既有电网，而会投资新增发电、输电、变电站与配套容量。
OpenAI says its data centers should not take existing grid power; projects invest in new generation, transmission, substations, and capacity.
Grid Build-outYouTube 完整播客
算力策略是 All of the AboveThe Strategy Is All Of The Above
多家 hyperscaler、neocloud、芯片伙伴、自主设计数据中心和定制芯片并行推进，以减少单一供给路径的约束。
Hyperscalers, neoclouds, chip partners, self-designed data centers, and custom silicon advance in parallel to reduce dependence on any single supply path.
Industrial ComputeYouTube 完整播客
08 / 15
快讯速览 · Builder Signals08 / 15
YC Startup School 2026 由 Sam Altman 收官Sam Altman Closes YC Startup School 2026
Garry Tan 感谢 Sam Altman 为 YC Startup School 2026 做收官分享。模型公司与创业者社区的连接仍在加强，前沿能力会更快被翻译为新产品和新公司。
Garry Tan thanked Sam Altman for closing YC Startup School 2026. The connection between frontier labs and founder communities keeps tightening, accelerating how model capability becomes products and companies.
@garrytanX 原文592 ❤
Aaron Levie：模型越强，Applied AI 层反而越重要Better Models Increase The Need For Applied AI
Aaron Levie 认为，智能本身不足以改造企业流程。真正落地需要连接系统和数据、设计人类决策节点、持续改善工作流，并满足监管要求；银行开户、法律审阅与生命科学都需要不同的 applied AI。模型越强，可自动化目标越大，这一层的机会也越大。
Aaron Levie argues that intelligence alone cannot transform enterprise processes. Teams must connect systems and data, design human decision points, improve workflows, and meet regulatory constraints. As models improve, automation becomes more ambitious and the applied-AI opportunity grows.
@levieX 原文321 ❤
Dan Shipper 开始口述史式追踪 Codex 的诞生A Deep Oral History Of How Codex Happened
Dan Shipper 暂停日常工作一周，通过对 OpenAI 内部人士的深度访谈撰写 Codex 的完整历史，并计划沿途公开线索与发现。AI 产品史正在从发布公告进入组织决策、技术转折和人物选择的深层复盘。
Dan Shipper is taking a week to write a definitive history of Codex from deep interviews with OpenAI insiders, sharing breadcrumbs along the way. AI product history is moving beyond launch notes into decisions, technical turns, and people.
@danshipperX 原文246 ❤
Amjad：攻击者更偏爱被补贴的闭源订阅Attackers May Prefer Subsidized Lab Subscriptions
Amjad 转述一位前 Anthropic 员工的观察：攻击者更倾向使用实验室大量补贴的 AI 订阅，而不是开放模型。风险讨论不能只盯开放权重，还要审视商业服务的价格补贴、访问控制与滥用监测。
Amjad relays an observation from a former Anthropic employee: attackers may prefer heavily subsidized lab subscriptions over open models. Risk analysis must examine commercial pricing, access controls, and abuse monitoring—not only open weights.
@amasadX 原文189 ❤
Zara Zhang：高频内容来自低摩擦记录High-frequency Publishing Comes From Low-friction Capture
Zara Zhang 平均每天发布约三次，但只花 15 到 20 分钟：想到就发，素材多来自已经当面对别人说过的话。持续输出的关键不是更复杂的内容流水线，而是缩短想法到发布的距离。
Zara Zhang posts about three times a day while spending only 15 to 20 minutes: she publishes ideas immediately, often after saying them aloud. Consistency comes from shortening the path from thought to publication.
@zarazhangruiX 原文145 ❤
Nikunj：Proof of Prompt 将替代 Proof of WorkProof Of Prompt May Replace Proof Of Work
Nikunj 预判“提示证明”将替代“工作证明”：当 agent 能执行大部分过程，人的差异更多体现在如何定义目标、约束和验收标准。提示词正在从输入文本变成可审计的意图与判断记录。
Nikunj predicts proof of prompt will replace proof of work. As agents execute more of the process, human differentiation shifts toward goals, constraints, and acceptance criteria; prompts become auditable records of intent and judgment.
@nikunjX 原文70 ❤
09 / 15
数据洞察 · Snapshot09 / 15
今日数据概览Today Stats
收录 Builder：11 位
总推文数：21 条
播客节目：1 期
博客文章：0 篇

最高互动：Sam Altman：ChatGPT Work 把一句话变成完整旅行协作系统 · 14.7K ❤
第二高互动：Sam Altman：Work 指向一种新型计算机 · 5.5K ❤
The snapshot includes 11 builders, 21 tweets, 1 podcast, and 0 blog post.
follow-buildersSnapshot follow-builders-2026-07-27-v2
5 条关键洞察5 Key Takeaways
ChatGPT Work 正从任务助手变成可调度上下文、应用与协作的新计算入口。
ChatGPT Work is becoming a computing surface that orchestrates context, apps, and collaboration.
模型越强，企业越需要连接数据、监管、人类决策点与现实反馈回路的 applied AI 层。
Stronger models increase demand for an applied-AI layer spanning data, regulation, human decisions, and real-world feedback.
开放权重从开发者偏好升级为产业领导力与生态扩散战略。
Open weights are moving from developer preference to an industrial strategy for leadership and ecosystem diffusion.
算力竞争的关键约束转向物理世界：电力、冷却、建设速度与每瓦 token 数。
Compute competition is increasingly constrained by power, cooling, build speed, and tokens per watt.
跨应用执行放大了价值，也把 Gmail、日历与办公数据的信任问题推到产品核心。
Cross-app execution expands value while making trust around Gmail, calendars, and office data a core product issue.
10 / 15
趋势拆解 · The New Work Product Stack10 / 15
入口层：意图直接变成执行Intent-to-execution Interface
自然语言任务可以调用历史上下文、搭建协作界面、等待群体决定并继续执行，入口从问答走向编排。
A natural-language task can use history, build a collaboration surface, wait for group decisions, and continue execution. The interface shifts from answers to orchestration.
@samaX 原文14.7K ❤ · 471 RT · 1.4K 💬
落地层：行业工作流与反馈Applied Workflow Layer
企业价值依赖系统连接、合规、人类决策节点和持续反馈；通用智能必须接触现实世界。
Enterprise value depends on system connections, compliance, human decision points, and continuous feedback; general intelligence must meet the real world.
@levieX 原文321 ❤ · 41 RT · 58 💬
运行层：开放模型与原生分发Open And Native Runtime
开放权重扩大供给选择，AI 代码翻译让既有脚本工具获得更轻、更快的原生交付路径。
Open weights expand model choice while AI code translation gives existing script tools a lighter, faster native delivery path.
@rauchgX 原文577 ❤ · 28 RT · 39 💬
治理层：访问、补贴与滥用Access, Subsidy, And Abuse
连接器扩大 agent 能力边界；与此同时，商业订阅的补贴与访问控制也会改变攻击者的工具选择。
Connectors expand an agent's reach, while subscription subsidies and access controls can also shape attackers' tool choices.
@amasadX 原文189 ❤ · 14 RT · 12 💬
11 / 15
内容与判断 · Fast Capture And Proof Of Prompt11 / 15
Zara Zhang：高频内容来自低摩擦记录High-frequency Publishing Comes From Low-friction Capture
Zara Zhang 平均每天发布约三次，但只花 15 到 20 分钟：想到就发，素材多来自已经当面对别人说过的话。持续输出的关键不是更复杂的内容流水线，而是缩短想法到发布的距离。
Zara Zhang posts about three times a day while spending only 15 to 20 minutes: she publishes ideas immediately, often after saying them aloud. Consistency comes from shortening the path from thought to publication.
@zarazhangruiX 原文145 ❤ · 2 RT · 25 💬
Nikunj：Proof of Prompt 将替代 Proof of WorkProof Of Prompt May Replace Proof Of Work
Nikunj 预判“提示证明”将替代“工作证明”：当 agent 能执行大部分过程，人的差异更多体现在如何定义目标、约束和验收标准。提示词正在从输入文本变成可审计的意图与判断记录。
Nikunj predicts proof of prompt will replace proof of work. As agents execute more of the process, human differentiation shifts toward goals, constraints, and acceptance criteria; prompts become auditable records of intent and judgment.
@nikunjX 原文70 ❤ · 3 RT · 17 💬
12 / 15
安全策略 · Trust, Access, And Abuse Economics12 / 15
当 AI 要读取 Gmail、日历和办公数据并代表用户行动时，采用门槛不再是 token 是否够用，而是用户是否相信系统的权限边界、审计与撤销机制。When AI reads Gmail, calendars, and office data and acts for users, adoption depends less on token limits than on trusted permissions, auditing, and revocation.
@amasadX 原文189 ❤ · 14 RT · 12 💬
Amjad：攻击者更偏爱被补贴的闭源订阅Attackers May Prefer Subsidized Lab Subscriptions
Amjad 转述一位前 Anthropic 员工的观察：攻击者更倾向使用实验室大量补贴的 AI 订阅，而不是开放模型。风险讨论不能只盯开放权重，还要审视商业服务的价格补贴、访问控制与滥用监测。
Amjad relays an observation from a former Anthropic employee: attackers may prefer heavily subsidized lab subscriptions over open models. Risk analysis must examine commercial pricing, access controls, and abuse monitoring—not only open weights.
@amasadX 原文189 ❤ · 14 RT · 12 💬
Aaron Levie：模型越强，Applied AI 层反而越重要Better Models Increase The Need For Applied AI
Aaron Levie 认为，智能本身不足以改造企业流程。真正落地需要连接系统和数据、设计人类决策节点、持续改善工作流，并满足监管要求；银行开户、法律审阅与生命科学都需要不同的 applied AI。模型越强，可自动化目标越大，这一层的机会也越大。
Aaron Levie argues that intelligence alone cannot transform enterprise processes. Teams must connect systems and data, design human decision points, improve workflows, and meet regulatory constraints. As models improve, automation becomes more ambitious and the applied-AI opportunity grows.
@levieX 原文321 ❤ · 41 RT · 58 💬
13 / 15
今日之声 VOICE OF THE DAY
智能本身不足以改造大多数流程；真正的价值来自让模型接触现实世界的系统、数据、反馈与决策节点。
Intelligence alone is not enough to transform most processes; value comes from connecting models to real systems, data, feedback, and decision points.
@levieX 原文321 ❤ · 41 RT · 58 💬
14 / 15
AI前沿每日脉动
AI Frontier Pulse · 2026.07.27
本期收录 11 位 Builder · 21 条推文 · 1 期播客 · 0 篇博客
ChatGPT Work · Applied AI · Open Weights · Tokens per Watt · Trust & Access
感谢阅读 · Thank You For Reading
Richard Liu · AI前沿每日脉动 · 2026 · Snapshot follow-builders-2026-07-27-v2