← 首页
21JUL
AI前沿每日脉动
AI Frontier Pulse · 中英双语版 Bilingual Edition
2026.07.21 · 周二刊
14 位 Builder29 条推文1 期播客0 篇博客
Everything Is CodeMulti-model AgentsAI-native CompanyAI Travel
Richard Liu · Snapshot follow-builders-2026-07-21-v1
今日头条 · Everything Is Code01 / 15
今天最强信号来自 Guillermo Rauch：AI 把 slide deck、design、promo video、Excel automation 都重新解释成 code。Today's clearest signal: AI turns decks, design, promo videos, and Excel automation into code-shaped work.
@rauchgX 原文4,300 ❤ · 400 RT · 328 💬
Guillermo Rauch：一切都正在变成 codeEverything Is Code
Rauch 用一句话概括了 AI 产品形态的变化：不只是软件开发变快，而是更多知识工作、创意工作和运营工作都被转写为可生成、可修改、可验证的 code-like artifact。
Rauch captures a product shift: AI does not only speed up software. It turns more knowledge, creative, and operational work into generated, editable, testable code-like artifacts.
@rauchgX 原文4,300 ❤ · 400 RT · 328 💬
Sam Altman：OpenAI 的节奏仍在加速OpenAI Keeps Moving
Sam 的短帖和 Thibault 的“OpenAI 永远不无聊”形成互文：今天的社区情绪不是单一产品发布，而是前沿实验室持续快速迭代带来的高噪声、高兴奋度。
Sam Altman and Thibault both point to the same tempo: frontier labs are iterating in public at a pace that creates noise, excitement, and constant product motion.
@samaX 原文2,820 ❤ · 121 RT · 456 💬
01 / 15
Agent 架构 · Multi-model Routing02 / 15
Aaron Levie：复杂 agent 的核心设计模式Planner Plus Workhorse
Levie 总结 Cursor 研究：前沿模型负责规划、拆解、设计决策和关键 trade-off，便宜模型负责执行明确任务。这种 planner + workhorse 组合能让整体 token 成本显著下降。
Levie summarizes Cursor research: frontier models handle planning, decomposition, design choices, and key tradeoffs, while cheaper models execute explicit instructions. Planner plus workhorse becomes the core agent pattern.
@levieX 原文206 ❤ · 29 RT · 37 💬
15X 成本改善背后的组织含义Cost Routing Is Product Design
这不是单纯省钱技巧，而是 agent 产品设计的结构变化：把 ambiguity 压缩成 instruction，然后把大量执行分配给不同智能层级。企业落地会越来越像调度系统。
This is not just a cost trick. It is a product architecture: collapse ambiguity into instructions, then dispatch execution across intelligence tiers. Enterprise AI increasingly looks like orchestration.
@levieX 原文206 ❤ · 29 RT · 37 💬
02 / 15
公司形态 · Post-coding-agent03 / 15
Zara：coding agents 之后创办的公司不一样Companies After Coding Agents
Zara 区分两类公司：coding agents 出现前建立、正在 retrofitting 的公司；以及之后出生的公司。后者从第一天就是小团队、项目制、每个人闭环、内部会议极少。
Zara separates companies founded before coding agents from those born after. The latter start with tiny teams, project-based work, individual closed loops, and almost no internal meetings.
@zarazhangruiX 原文162 ❤ · 13 RT · 19 💬
AI 面试：一轮不用 AI，一轮必须用 AIHiring For Agent Fluency
Zara 提出的面试流程很有启发：第一轮现场且禁用 AI，测领域知识；第二轮必须用 AI 完成项目，并评估结果和 agent 聊天记录。未来招聘要同时测人和人机协作。
Her interview design tests both domain expertise without AI and agent fluency with AI. The output matters, but so does the transcript of how the candidate worked with agents.
@zarazhangruiX 原文154 ❤ · 8 RT · 36 💬
Madhu Guru：产品感进入黄金时代Product Sense Premium
Madhu 认为现在是最适合拥有 product sense 的时代。模型能力快速扩散后，真正差异越来越在于把能力放进正确任务、正确流程和正确用户体验。
Madhu says it is the greatest time to have product sense. As model ability diffuses, differentiation shifts to choosing the right task, workflow, and user experience.
@realmadhuguruX 原文100 ❤ · 9 RT · 2 💬
Dan Shipper：agent 公司也在招聘 agent 人才Agent-native Hiring
Every 招 senior engineer 做 agent，要求“must love agents”。这类岗位说明 AI-native 公司不只购买工具，而是在把 agent 作为核心产品和组织能力。
Every is hiring a senior engineer for its agent work and explicitly wants someone who loves agents. AI-native companies are treating agents as product and organizational capability.
@danshipperX 原文47 ❤ · 5 RT · 3 💬
03 / 15
开放模型 · Regulation / Competition04 / 15
Peter Yang：禁中国模型会成为自伤Model Ban As Self-own
Peter Yang 把“禁中国模型”类比成“禁中国电动车”：短期看像保护，长期可能让本土生态失去与强竞争者共同进化的机会。
Peter Yang compares banning Chinese models to banning Chinese EVs: it may look protective short term, but can weaken the local ecosystem by removing strong competitive pressure.
@petergyangX 原文1,019 ❤ · 65 RT · 71 💬
Matt Turck：开源模型给前沿实验室制造压力Open Models Pressure Frontier Labs
Matt 用梗图评论中国开源模型冲击 OpenAI 和 Anthropic。虽然轻松，但它反映了今天的严肃议题：开源能力的速度正在改变监管和商业策略。
Matt jokes about top Chinese open-source models pressuring OpenAI and Anthropic. Under the joke is a serious point: open capability changes regulation and business strategy.
@mattturckX 原文70 ❤ · 4 RT · 6 💬
Madhu：真正 tokenomics 是开放、成本和路由The Tokenomics That Matter
Madhu 指出，后 web3 时代真正重要的 tokenomics 不是币，而是 open vs closed weight、inference costs 和 model routing。这和 Levie 的多模型 agent 观点正好合流。
Madhu reframes tokenomics: the important debate is open vs closed weights, inference cost, and model routing. That converges with Levie’s multi-model agent architecture.
@realmadhuguruX 原文24 ❤ · 0 RT · 2 💬
Garry Tan：compute 仍是底层约束Compute Rules
Garry 的 compute 梗延续一个朴素现实：无论模型开放还是闭源，推理成本、容量和调度能力仍然决定许多产品边界。
Garry’s compute post points back to a basic constraint: whether models are open or closed, inference cost, capacity, and scheduling still shape product boundaries.
@garrytanX 原文216 ❤ · 23 RT · 30 💬
04 / 15
创业判断 · Moats / Unique Insight05 / 15
AI 时代没有永恒 moat，不等于 scale 和 capital 自动成为 moat。真正难的是找到值得十年投入的独特洞察。No permanent AI moat does not mean scale and capital become the moat. The hard part is a unique insight worth ten years.
@nikunjX 原文107 ❤ · 5 RT · 13 💬
Nikunj：资本与规模不能替代洞察Capital Is Not A Substitute
Nikunj 提醒创始人读一点历史：Webvan、Groupon、MySpace、Yahoo、Nokia 等都曾有结构和资本优势，但最终被更强洞察或自身规模拖累击败。
Nikunj reminds founders that many companies once had scale and capital advantages but lost to better insight or collapsed under their own scale.
@nikunjX 原文107 ❤ · 5 RT · 13 💬
Booking CEO：moat 会被创新持续侵蚀No Such Thing As A Permanent Moat
No Priors 访谈中，Booking.com CEO Glenn Fogel 的核心态度是：没有永久受保护的位置，长期胜利来自持续开发新服务、理解客户痛点并改进业务。
Booking.com CEO Glenn Fogel argues there is no permanently protected position. Long-term advantage comes from continuously building new services and understanding customer pain.
No PriorsYouTube 完整播客2026-07-09
05 / 15
产品设计 · Verification / Rubrics06 / 15
Peter Yang：用一个 agent 做事，另一个按 rubric 审查Agent Reviewer Pattern
Peter Yang 引用 Thariq 的观点：对视频短片这类非确定任务，应让独立 verification agent 根据 rubric 审查，而不是让模型偏爱自己的输出。
Peter Yang shares Thariq’s reviewer pattern: for non-deterministic tasks, use a separate verification agent with a rubric instead of letting a model judge its own work.
@petergyangX 原文47 ❤ · 5 RT · 3 💬
Swyx：benchmark 可被 test lookalikes 影响Benchmark Lookalikes
Swyx 讨论 RLM paper 中的 trajectory comparison：即使没有直接训练测试集，也可能通过 test lookalikes 逼近目标分数。开放权重发布如果不给数据和环境，就很难判断。
Swyx highlights a trajectory comparison issue: even without training on the test set, models can be steered by test lookalikes. Without datasets and environments, claims are hard to inspect.
@swyxX 原文62 ❤ · 7 RT · 13 💬
Amjad：coding agent 发货实体产品Physical Products From Agents
Amjad 转发“coding agent 发货实体产品”的案例。AI 从软件 artifact 进入物理商品，是 everything is code 叙事向现实世界延伸的一小步。
Amjad points to a physical product shipped by a coding agent. It is a small sign that everything-is-code can leak from software artifacts into physical goods.
@amasadX 原文270 ❤ · 10 RT · 27 💬
Thariq：短暂线上 bug 与快速修复文化Fast Feedback Culture
Thariq 提到一个只上线几分钟的 bug。对 AI coding 工具体系来说，快速反馈、快速修复、本地夜间使用都构成产品学习回路。
Thariq mentions a bug that was live for only minutes. Fast feedback, fast fixes, and real personal use become part of AI coding product learning loops.
@trq212X 原文124 ❤ · 0 RT · 21 💬
06 / 15
播客深度 · AI Travel07 / 15
No Priors
AI travel 不是简单把聊天机器人放到订票页，而是理解库存、供应、需求、服务、信任和复杂线下约束。AI travel is not a chatbot on a booking page. It touches inventory, supply, demand, service, trust, and offline constraints.
No PriorsYouTube 完整播客2026-07-09
Booking.com CEO：旅行平台的护城河必须持续重建Travel Moats Are Temporary
Glenn Fogel 强调，旅行行业没有不会被创新侵蚀的永久护城河。AI 时代的 Booking 不只是搜索结果排序，而是围绕用户需求、供应关系和服务体验持续建新能力。
Glenn Fogel argues travel has no permanent moat. In the AI era, Booking must keep building new capabilities around user demand, supplier relationships, and service experience.
No PriorsYouTube 完整播客2026-07-09
AI 旅行 agent 需要真正懂业务复杂性Travel Agents Need Domain Depth
访谈提醒创业者：旅行是复杂业务，不是“做一个 agent 就打掉大玩家”。真正难点在供给、履约、价格、信任、售后和全球化运营。
The interview warns that travel is complex. Building an agent is not enough to beat incumbents; the hard parts are supply, fulfillment, pricing, trust, support, and global operations.
No PriorsYouTube 完整播客2026-07-09
07 / 15
快讯速览 · 6 条精选动态08 / 15
Rauch：Everything is codeEverything Is Code
Deck、design、video、Excel automation 都开始被 AI 转译成 code-like artifact。
Decks, design, videos, and Excel automation become code-like artifacts.
@rauchgX 原文4,300 ❤ · 400 RT · 328 💬
Levie：多模型 agentModel Routing
前沿 planner + 便宜 workhorse 成为复杂 agent 的成本优化模式。
Frontier planner plus cheaper workhorse becomes a cost pattern for agents.
@levieX 原文206 ❤ · 29 RT · 37 💬
Zara：AI-native 公司Post-agent Company
coding agents 之后出生的公司会更小、更项目制、更少会议。
Post-agent companies are smaller, project-based, and lighter on meetings.
@zarazhangruiX 原文162 ❤ · 13 RT · 19 💬
Zara：AI 面试流程Agent Interview
面试要同时测无 AI 的领域力和用 AI 完成项目的协作力。
Interviews should test both domain skill without AI and agent collaboration.
@zarazhangruiX 原文154 ❤ · 8 RT · 36 💬
Peter：禁模型是自伤Model Competition
禁止强竞争模型可能削弱本土生态学习速度。
Banning strong competitor models may weaken local learning speed.
@petergyangX 原文1,019 ❤ · 65 RT · 71 💬
Madhu：企业 AI 是 AGI 前沿Enterprise Tasks
经济有价值任务聚集在企业流程中，企业 AI 是重要前沿。
Economically valuable tasks live in enterprise workflows, making enterprise AI a frontier.
@realmadhuguruX 原文24 ❤ · 0 RT · 2 💬
08 / 15
数据洞察 · Snapshot09 / 15
今日数据概览Today Stats
收录 Builder：14 位
总推文数：29 条
播客节目：1 期（No Priors × Booking.com）
博客文章：0 篇

最高互动：Everything is code · 4,300 ❤
第二高互动：OpenAI 工作节奏 · 3,672 ❤
The snapshot includes 14 builders, 29 tweets, 1 podcast, and 0 blogs. Top signal: everything is code, multi-model agents, and AI-native company design.
follow-buildersSnapshot follow-builders-2026-07-21-v1
5 条关键洞察5 Key Takeaways
AI 正在把更多媒介转成 code-like artifact，生成、修改、验证成为统一操作。
AI turns more media into code-like artifacts: generate, edit, verify.
复杂 agent 会采用多模型路由：高智能负责规划，低成本模型负责执行。
Complex agents will route work across models: frontier planning, cheaper execution.
AI-native 公司从组织结构上就不同：小团队、项目闭环、少会议。
AI-native companies differ structurally: small teams, project loops, fewer meetings.
开放模型竞争改变监管、安全和商业策略，禁用强模型可能是自伤。
Open model competition changes regulation, security, and strategy; bans can backfire.
AI travel 说明现实行业不只是 agent UI，还需要供应链和服务深度。
AI travel shows real industries need supply and service depth, not only agent UI.
09 / 15
趋势拆解 · Everything Is Code Stack10 / 15
表达层：deck / design / videoExpression Layer
AI 把表达物转成可编辑的中间表示。用户想要的不是代码本身，而是能快速成形和修改的 artifact。
AI turns expression into editable intermediate representations. Users want artifacts that can be shaped and revised quickly.
@rauchgX 原文4,300 ❤ · 400 RT · 328 💬
执行层：agent 做项目Execution Layer
coding agent 不只写软件，也开始连接实体产品、业务项目和特殊任务。执行范围继续外溢。
Coding agents do not only write software; they start touching physical products, business projects, and special tasks.
@amasadX 原文270 ❤ · 10 RT · 27 💬
验证层：rubric + reviewer agentVerification Layer
当任务没有确定答案，rubric 和独立审查 agent 会成为质量控制核心。
When tasks lack deterministic answers, rubrics and separate reviewer agents become quality control.
@petergyangX 原文47 ❤ · 5 RT · 3 💬
路由层：planner / workhorseRouting Layer
多模型系统把不同认知难度的 token 分配给不同模型，成本和质量一起优化。
Multi-model systems allocate different cognitive loads to different models, optimizing cost and quality together.
@levieX 原文206 ❤ · 29 RT · 37 💬
10 / 15
行业趋势 · Enterprise AI11 / 15
通向 AGI 的路由不是抽象 benchmark，而是经济有价值任务。企业流程是这些任务最密集的地方。The road to AGI runs through economically valuable tasks, and many of those tasks live inside enterprise workflows.
@realmadhuguruX 原文24 ❤ · 0 RT · 2 💬
Madhu：企业 AI 是最重要前沿之一Enterprise As Frontier
Madhu 的判断把“AGI”拉回商业现实：很多有经济价值的任务存在于企业流程中，因此 enterprise AI 不是慢热应用层，而是能力落地的前沿。
Madhu reframes AGI through business reality: many economically valuable tasks live in enterprise workflows, making enterprise AI a true frontier.
@realmadhuguruX 原文24 ❤ · 0 RT · 2 💬
产品感成为稀缺能力Product Sense As Leverage
当模型能力持续外溢，能识别高价值任务、设计正确体验、组织反馈回路的人会更有杠杆。
As model ability diffuses, people who identify valuable tasks, design the right experience, and organize feedback loops gain leverage.
@realmadhuguruX 原文100 ❤ · 9 RT · 2 💬
11 / 15
合流判断 · What Changed Today12 / 15
从“AI 能写代码”到“一切可被代码化”From Coding To Codification
今天不是又一次 coding agent 讨论，而是 coding agent 逻辑外溢到 deck、design、video、spreadsheet、实体产品、旅行服务和组织结构。
Today is not just another coding-agent thread. The logic spills into decks, design, video, spreadsheets, physical products, travel services, and company structure.
@rauchgX 原文4,300 ❤ · 400 RT · 328 💬
从“一个大模型”到“智能调度系统”From One Model To Orchestration
Levie、Madhu、Swyx 和 Peter 的内容合在一起，指向下一阶段：模型不再单独竞争，系统会围绕成本、验证、开放性和任务难度做调度。
Levie, Madhu, Swyx, and Peter point to the next phase: models compete inside systems that route around cost, verification, openness, and task difficulty.
@levieX 原文206 ❤ · 29 RT · 37 💬
12 / 15
今日之声 VOICE OF THE DAY
AI 的大课题不是让代码生成更快，而是让更多工作变成可表达、可执行、可验证的 code-like artifact。
The big AI lesson is not only faster code generation, but turning more work into expressible, executable, verifiable code-like artifacts.
@rauchgX 原文4,300 ❤ · 400 RT · 328 💬
13 / 15
AI前沿每日脉动
AI Frontier Pulse · 2026.07.21
本期收录 14 位 Builder · 29 条推文 · 1 期播客
Everything Is Code · Multi-model Agents · AI-native Company · AI Travel
感谢阅读 · Thank You For Reading
Richard Liu · AI前沿每日脉动 · 2026 · Snapshot follow-builders-2026-07-21-v1