← 首页
20JUL
AI前沿每日脉动
AI Frontier Pulse · 中英双语版 Bilingual Edition
2026.07.20 · 周一刊 · v2
15 位 Builder29 条推文1 期播客0 篇博客
ChatGPT WorkSkillsAI DiffusionAgentic Commerce
Richard Liu · Snapshot follow-builders-2026-07-20-v2
今日头条 · ChatGPT Work Impact01 / 15
ChatGPT Work 今天的主题不再是功能清单，而是用户如何把真实生活和工作中的正向变化反馈给团队。ChatGPT Work is moving from feature list to lived impact: how people delegate real work and life moments back into the product loop.
@thsottiauxX 原文1,796 ❤ · 40 RT · 791 💬
Thibault：收集 ChatGPT 正向影响故事Impact Becomes Product Feedback
Thibault 在读完用户如何使用 ChatGPT Work 的私信后，继续向社区征集 ChatGPT 对自己或他人生活产生深度正向影响的时刻。高回复量说明，Work 产品正在从“效率工具”进入“真实委派与生活辅助”的叙事。
After reading DMs about how people use ChatGPT Work, Thibault asked for moments where ChatGPT had a deeply positive impact. The replies signal that Work is becoming a lived delegation surface, not just a productivity feature.
@thsottiauxX 原文1,796 ❤ · 40 RT · 791 💬
Peter Yang：非技术用户需要更清楚的云端文案Copy Is Part Of The Product
Peter Yang 指出，“run this chat in the cloud” 对非技术用户并不清楚；Codex 跳转体验还会让他重新下载 app。AI 工作流产品的下一道门槛，是把云端执行解释成人能自然理解的动作。
Peter Yang notes that non-technical users may not understand “run this chat in the cloud.” The next hurdle for AI workflow products is copy that translates cloud execution into obvious user intent.
@petergyangX 原文15 ❤ · 0 RT · 2 💬
01 / 15
技能化工作流 · Cowork / Claude Code02 / 15
Cat Wu：用 Claude Cowork 管理日历Calendar As A Skill
Cat Wu 分享了 Claude Cowork 管理日历的 prompt：一周会议少于 20 小时、去重冲突会议、参考过去拒绝过的会议类型、晚餐不计入会议，并在更新邀请前询问。关键点是“build a skill that you refine”。
Cat Wu uses Claude Cowork to manage her calendar with constraints, history, deduping, and confirmation before invite changes. The key phrase is “build a skill that you refine.”
@_catwuX 原文350 ❤ · 8 RT · 42 💬
Thariq：从系统 prompt 学到的东西会沉淀成技能Prompt Lessons Become Skills
Thariq 正在写一篇关于团队学到什么、以及如何把这些经验应用到 skills 和 system prompts 的文章。Claude Code 生态正在把一次性 prompt 经验沉淀成可复用的操作规范。
Thariq is writing about lessons learned and how to apply them to skills and system prompts. The Claude Code ecosystem is turning one-off prompt experience into reusable operating rules.
@trq212X 原文1,053 ❤ · 45 RT · 48 💬
02 / 15
文件格式 · Markdown / GSkills03 / 15
当 intelligence stack 快速变化时，最朴素的文本格式反而成为最稳的接口。When the intelligence stack is changing this fast, the simplest text formats become the most durable interface.
@garrytanX 原文668 ❤ · 54 RT · 82 💬
Garry Tan：Markdown 是 AI 时代的长寿数据格式Markdown As Durable AI Substrate
Garry Tan 认为 Markdown 文件是通用格式，会存活很久；在 intelligence stack 剧烈变化时，它自然成为很好的数据格式。这呼应了 AI skill、prompt、知识库都在向可读文本靠拢。
Garry Tan argues Markdown files are universal and durable, making them a strong data format while the intelligence stack is in flux. Skills, prompts, and knowledge bases are all converging on readable text.
@garrytanX 原文668 ❤ · 54 RT · 82 💬
GSkills anyone? 技能正在产品化Skills As Product Surface
Garry 另一条短帖提出 “GSkills anyone?” 虽然简短，但和今天的技能化主题高度一致：AI 工作流的复用单位正在从 app 变成 skill、prompt、markdown 和可执行规则。
His short “GSkills anyone?” post fits the broader theme: reusable AI workflow units are shifting from apps to skills, prompts, Markdown, and executable rules.
@garrytanX 原文621 ❤ · 39 RT · 34 💬
03 / 15
行业分析 · AI Diffusion04 / 15
Aaron Levie：AI 扩散受现实反馈速度限制Reality Sets The Learning Rate
Levie 指出，代码能快速被 AI 采用，是因为一个人能写、测、跑并端到端改变结果；但生命科学、销售、合同、工业设计都需要现实世界回应。模型输出本身不够，applied AI 必须嵌入行业工作流。
Levie argues AI diffusion is rate-limited by how fast reality answers. Code is fast because one person can write, test, and run it end-to-end. Other industries need applied AI embedded in real workflows.
@levieX 原文280 ❤ · 34 RT · 32 💬
应用层机会：把智能变成行业反馈回路Applied AI Layer
这条观点解释了为什么垂直应用层仍有巨大空间：不是把模型接上 API 就结束，而是要改变底层 workflow、处理对手方、合规、实验、合同和销售周期。
This explains why vertical AI still has room: connecting a model API is not enough. Products must reshape workflows and handle counterparties, compliance, experiments, contracts, and sales cycles.
@levieX 原文280 ❤ · 34 RT · 32 💬
04 / 15
成本曲线 · Inference Demand05 / 15
AI 变便宜，支出未必下降Cheaper Tokens, More Consumption
Levie 反驳“AI 成本下降会让支出下降”的直觉：通常相反。token 更便宜会让更多任务变得可做，代码、审查、安全、大数据 agent 都会扩张使用量。
Levie pushes back on the idea that cheaper AI lowers spend. Usually it expands usage: more coding, more review, more security checks, more agents over previously unreachable datasets.
@levieX 原文235 ❤ · 20 RT · 40 💬
开源模型商业模式的底层逻辑Open Models Still Need Infra
他进一步指出，没人真的在设备上跑这些模型，大多数还是跑在基础设施中。因此，开放模型越强，推理和托管服务商越可能受益。
He adds that most people are not running these models on-device; they run them in infrastructure. Strong open models can still drive demand for inference and hosting providers.
@levieX 原文235 ❤ · 20 RT · 40 💬
消费级订阅为何难巨大化Consumer AI Subscription Constraint
Amjad Masad 提醒，消费者主要花钱在食物、房租、娱乐、通信和购物上，软件通常由公司购买。这解释了为什么巨大消费订阅业务比企业 AI 更难。
Amjad Masad notes consumers spend on food, rent, entertainment, connectivity, and shopping, while software is usually bought by companies. That makes giant consumer software subscriptions harder.
@amasadX 原文90 ❤ · 5 RT · 13 💬
Every：AI copy edit 终于越过门槛70 Percent Copy Edits
Dan Shipper 说 Every 内部过去一周已经能自动完成约 70% 以前手工做的 copy edits，这是多年尝试后第一次达到这个门槛。知识工作自动化正在从“能演示”进入“能替代一块流程”。
Dan Shipper says Every has been able to automate about 70% of internal copy edits for a week, the first time after years of trying. Knowledge-work automation is crossing from demo to workflow replacement.
@danshipperX 原文107 ❤ · 2 RT · 4 💬
05 / 15
开放模型与安全 · Cyber Evals06 / 15
Guillermo Rauch：Cybersecurity 是超级智能 IQ 测试Security As The IQ Test
Rauch 认为 cybersecurity 是衡量 superintelligence 的最佳 benchmark 之一。比起 one-shot clone，找漏洞、修补、逆向和利用需要跨语言、跨运行时、跨框架的 corner thinking。
Rauch argues cybersecurity is one of the best benchmarks for superintelligence. Finding, patching, reversing, and exploiting require reasoning that transcends languages, runtimes, and frameworks.
@rauchgX 原文82 ❤ · 8 RT · 10 💬
Levie：开放权重模型改变监管计算Open Weights Change Regulation
Levie 表示，当强开源替代品只落后前沿模型一点点，继续 gatekeep 前沿能力会让自己更不安全、更不竞争。监管需要重新计算开放模型已经很强这个事实。
Levie argues that when strong open alternatives are close to frontier models, gatekeeping frontier access can make an ecosystem less secure and less competitive. Regulation must account for open weights strength.
@levieX 原文232 ❤ · 24 RT · 20 💬
06 / 15
产品与内容 · Disposable Software07 / 15
Zara：把反复回答的问题变成内容Repeated Answers Become Positioning
Zara 建议刚开始做内容的人，把朋友、同事最常问的三个问题写成视频或帖子。你反复说过三次以上的东西，就值得变成内容。这是个人知识产品化的低摩擦路径。
Zara suggests beginners turn the top three questions friends or colleagues ask into posts. If you have said something more than three times, it is worth turning into content.
@zarazhangruiX 原文201 ❤ · 7 RT · 14 💬
软件可以是一次性的Disposable Software
Zara 认为需要习惯代码和软件可以被丢弃：一次性设计 playground、用于理解代码的 HTML 页面、只看一次的 dashboard。AI 让软件从资产也变成临时认知工具。
Zara argues code and software can now be disposable: one-off design playgrounds, HTML pages to understand code, temporary dashboards. AI turns software into a temporary cognition tool.
@zarazhangruiX 原文145 ❤ · 13 RT · 29 💬
Peter Steinberger：IYKYK 的工具链文化Toolchain In-Jokes
Peter Steinberger 的短帖虽然含蓄，但放在今天的上下文里，反映了 AI 工具链用户之间越来越多的内部梗和共同经验。
Peter Steinberger’s terse post reflects a growing shared culture among AI toolchain users: dense, experiential, and often legible only to practitioners.
@steipeteX 原文131 ❤ · 8 RT · 25 💬
Swyx：键盘硬件也能触发 builder 兴奋Interfaces Still Matter
Swyx 对 custom keyboard 的兴趣提醒我们，AI 时代也不是纯软件叙事；输入设备、快捷键、控制面板和实体交互仍然会影响 builder 工作流。
Swyx’s custom-keyboard enthusiasm is a reminder that AI workflows are not purely software; input devices, shortcuts, control surfaces, and physical interaction still shape builder behavior.
@swyxX 原文86 ❤ · 0 RT · 16 💬
07 / 15
播客深度 · Agentic Commerce08 / 15
MAD Podcast
“超过六分之一的 AI 公司注册是 token abuse。” Agentic commerce 的第一课不是购物按钮，而是支付、权限、风控和 token 经济。“More than one in six signups at AI companies are this kind of abuse.” Agentic commerce starts with payments, permissions, risk, and token economics.
MAD Podcast节目来源2026-07-09
Stripe AI Chief：Agent 会买、卖、付钱Agents As Economic Actors
Stripe 的 Emily Sands 讨论 agentic commerce 的经济栈：agent 不只是代表人买鞋，也可能运行一个业务，既买东西也卖东西，并尝试创造利润。
Stripe’s Emily Sands describes the economic stack for agentic commerce: agents will not only buy shoes for people; they may run businesses, buy, sell, and generate profits.
MAD Podcast节目来源2026-07-09
Token theft 是 AI 商业最被低估的风险Token Theft As Dine And Dash
访谈开头强调，AI 欺诈者不一定要偷钱或凭据，只要偷 token 就够了。这让计费、权限和滥用检测成为 AI 产品的核心基础设施。
The conversation highlights that AI fraudsters may not need to steal money or credentials; stealing tokens is enough. Billing, permissioning, and abuse detection become core AI infrastructure.
MAD Podcast节目来源2026-07-09
08 / 15
快讯速览 · 6 条精选动态09 / 15
Thibault 收集正向影响案例Positive Impact Loop
ChatGPT Work 团队继续直接从用户故事中寻找产品信号。
The ChatGPT Work team is using user stories as product signal.
@thsottiauxX 原文1,796 ❤ · 40 RT · 791 💬
Garry Tan 推 MarkdownMarkdown For AI
Markdown 在 AI stack 不稳定时成为稳健知识载体。
Markdown becomes a durable knowledge carrier in a changing AI stack.
@garrytanX 原文668 ❤ · 54 RT · 82 💬
Cat Wu 日历 promptCalendar Cowork
Claude Cowork 的日历管理 prompt 展示了可迭代 skill 的形状。
Her calendar prompt shows the shape of refinable skills.
@_catwuX 原文350 ❤ · 8 RT · 42 💬
Levie 谈 AI 成本Inference Demand
token 更便宜可能带来更多使用，而不是更低支出。
Cheaper tokens may increase usage rather than reduce spend.
@levieX 原文235 ❤ · 20 RT · 40 💬
Zara 谈一次性软件Disposable Software
AI 让临时页面、临时 dashboard 成为日常认知工具。
AI makes throwaway pages and dashboards everyday cognition tools.
@zarazhangruiX 原文145 ❤ · 13 RT · 29 💬
Dan Shipper 的 copy edit 门槛Editing Automation
Every 内部约 70% copy edits 已可自动完成。
Every can now automate about 70% of internal copy edits.
@danshipperX 原文107 ❤ · 2 RT · 4 💬
09 / 15
数据洞察 · Snapshot10 / 15
今日数据概览Today Stats
收录 Builder：15 位
总推文数：29 条
播客节目：1 期
博客文章：0 篇

最高互动：Thibault ChatGPT impact · 1,796 ❤
核心主题：Work impact / Skills / AI diffusion / Agentic commerce
The snapshot includes 15 builders, 29 tweets, 1 podcast, and 0 blogs. The strongest substantive signal is ChatGPT Work impact and skill-based workflows.
follow-buildersSnapshot follow-builders-2026-07-20-v2
5 条关键洞察5 Key Takeaways
Work 产品开始用真实影响故事训练自己的方向，而不是只展示功能。
Work products are using impact stories as direction, not just feature demos.
Skill 是新的复用单位：prompt、Markdown、system rule 和工作流被打包成可迭代资产。
Skills are the new reusable unit: prompts, Markdown, system rules, and workflows become assets.
AI 扩散速度取决于现实反馈回路，代码快，合同、销售、药物和工业慢。
AI diffusion depends on real-world feedback loops: code is fast; contracts, sales, drugs, and industry are slower.
token 成本下降会扩大推理需求，infra 和托管层仍是关键受益者。
Lower token costs can expand inference demand, keeping infra and hosting central.
agentic commerce 的难点不是“买按钮”，而是权限、支付、风控、token theft 和业务代理。
Agentic commerce is not just a buy button; it is permissions, payments, risk, token theft, and business agents.
10 / 15
趋势拆解 · Skills Stack11 / 15
输入层：自然语言 + 文案清晰度Input Layer
Peter Yang 的批评说明，AI 产品不是把云端能力接上就完了；入口文案必须让非技术用户知道下一步会发生什么。
Peter Yang’s critique shows AI products need copy that explains cloud execution to non-technical users.
@petergyangX 原文15 ❤ · 0 RT · 2 💬
规则层：system prompt + skillRule Layer
Thariq 和 Cat Wu 的内容都指向同一件事：把偏好、限制、过去经验和确认点写成可复用 skill。
Thariq and Cat Wu point to the same move: turning preferences, constraints, past experience, and confirmations into reusable skills.
@_catwuX 原文350 ❤ · 8 RT · 42 💬
存储层：Markdown / 文本资产Storage Layer
Garry Tan 的 Markdown 判断解释了为什么 plain text 仍是 AI 工作流底座：可读、可 diff、可迁移。
Garry Tan’s Markdown point explains why plain text remains foundational: readable, diffable, portable.
@garrytanX 原文668 ❤ · 54 RT · 82 💬
执行层：agents over real workflowsExecution Layer
Levie 的 applied AI 观点提醒：真正价值在把智能接入行业流程，而不是停在模型输出。
Levie’s applied AI point: value comes from attaching intelligence to industry workflows, not stopping at model output.
@levieX 原文280 ❤ · 34 RT · 32 💬
11 / 15
趋势合流 · Applied AI12 / 15
今天的主线是：AI 正在从“聪明回答”变成“可复用工作单元”。Work、Cowork、skills、Markdown、agentic commerce 都在描述同一个迁移。Today’s throughline: AI is moving from smart answers to reusable units of work. Work, Cowork, skills, Markdown, and agentic commerce all describe the same migration.
从聊天到业务流程From Chat To Operations
ChatGPT Work 和 Claude Cowork 都在把 AI 拉进日历、文案、编辑、云端执行、商业支付等真实流程。赢家不是“回答最漂亮”的模型，而是能接住流程的系统。
ChatGPT Work and Claude Cowork pull AI into calendars, copy, editing, cloud execution, and payments. The winner is not only the best-answering model, but the system that captures workflow.
@thsottiauxX 原文1,796 ❤ · 40 RT · 791 💬
从模型差异到部署差异From Model Gap To Deployment Gap
Rauch 与 Levie 的观点合在一起：cyber eval 证明模型推理差异仍重要，但开放模型强度也迫使企业重新思考访问、部署和安全策略。
Rauch and Levie together: cyber evals show model reasoning gaps still matter, while open model strength forces new access, deployment, and security strategy.
@rauchgX 原文82 ❤ · 8 RT · 10 💬
12 / 15
今日之声 VOICE OF THE DAY
“模型输出本身在多数情况下并不够。你需要真正改变底层 workflow，并处理这些行业里现实世界反馈回路的复杂性。”
Model outputs alone are not enough in most cases. You need to change the underlying workflows and deal with real-world feedback loops.
@levieX 原文280 ❤ · 34 RT · 32 💬
13 / 15
AI前沿每日脉动
AI Frontier Pulse · 2026.07.20 · v2
本期收录 15 位 Builder · 29 条推文 · 1 期播客
ChatGPT Work · Skills · Markdown · AI Diffusion · Agentic Commerce
感谢阅读 · Thank You For Reading
Richard Liu · AI前沿每日脉动 · 2026 · Snapshot follow-builders-2026-07-20-v2