AI前沿每日脉动

19JUL

AI前沿每日脉动

AI Frontier Pulse · 中英双语版 Bilingual Edition

2026.07.19 · 周日刊

11 位 Builder23 条推文1 期深度播客0 篇博客

ChatGPT WorkVoice DelegationCyber EvalsAI Scientist

Richard Liu · 2026

Curated by Richard Liu · Snapshot follow-builders-2026-07-19-v1

ChatGPT Work 的产品叙事今天从“能做什么”进入“用户怎样把真实工作委派进去”：建站、邮件、文档、表格、幻灯片，以及通过语音下达复杂任务。ChatGPT Work shifted from capability list to real delegation: sites, email, documents, spreadsheets, slides, and complex voice-directed tasks.

@thsottiaux X 原文1,465 ❤ · 68 RT · 266 💬

Thibault：ChatGPT Work 已在 Plus/Pro/Business/Enterprise 中Work Becomes The Surface

Thibault 明确列出 ChatGPT Work 的用途：创建和托管 sites、管理 emails、总结大量 documents、创建 docs/sheets/slides，并说明已经在 mobile app 和 web 中可用。工作入口正在从聊天框变成可执行工作台。

Thibault lists ChatGPT Work use cases: creating and hosting sites, managing emails, summarizing documents, and creating docs, sheets, and slides. The work entry point is moving from chat box to executable workbench.

@thsottiaux X 原文1,465 ❤ · 68 RT · 266 💬

“我的工作基本上就是委派给 ChatGPT Work”Delegation As Daily Job

他随后说自己的工作基本变成 delegating to ChatGPT Work，并停不下使用 dictation。AI 工作流的下一步不是多一个按钮，而是把自然语言、语音和工具执行连成日常操作。

He then says his job is basically delegating to ChatGPT Work and he cannot stop using dictation. The next step is not another button; it is everyday operation through language, voice, and tool execution.

@thsottiaux X 原文229 ❤ · 2 RT · 50 💬

01 / 15

一段口述任务：从 Twitter DMs 到 cohort spreadsheetVoice-To-Workflow

Thibault 贴出自己口述给 ChatGPT Work 的长任务：遍历 Twitter DMs，找出提到 ChatGPT Work 的申请者，整理姓名、DM 链接、帖子文本、工作分类和 workflow sophistication rating，用于挑选早测 cohort。

Thibault posted a dictated task: go through Twitter DMs, find ChatGPT Work applicants, create a spreadsheet with names, DM links, post text, work taxonomy, and workflow sophistication ratings for beta cohort selection.

@thsottiaux X 原文136 ❤ · 1 RT · 40 💬

Peter Yang：孩子和 ChatGPT Site 做乘法游戏Sites As Family-Scale Software

Peter Yang 和 8 岁孩子做了一个 ChatGPT Site 来学乘法表：用图片生成 UI 和角色，加入音乐和 timed boss level。AI site creation 开始覆盖家庭、小工具、教育和个人创作。

Peter Yang and his 8-year-old built a ChatGPT Site to learn multiplication tables, using image-generated UI and characters, music, and a timed boss level. AI site creation now reaches family-scale tools, education, and personal creation.

@petergyang X 原文221 ❤ · 8 RT · 21 💬

委派任务变成 spreadsheet productionStructured Output At Work

这条口述任务的关键不在“总结 DMs”，而在要求 taxonomy、rating、链接回溯和 cohort sampling。AI 助手正在承担结构化运营任务，而不只是生成文字。

The key is not summarizing DMs; it asks for taxonomy, ratings, traceable links, and cohort sampling. AI assistants are taking on structured operations, not just writing.

@thsottiaux X 原文136 ❤ · 1 RT · 40 💬

语音是复杂委派的低摩擦入口Voice Lowers Planning Friction

复杂任务往往很难打字描述，但可以边想边说。语音输入把“想法到任务规格”的摩擦降下来，让 Work 类产品更像真人助理入口。

Complex tasks are hard to type but easier to talk through. Voice lowers the friction from idea to task specification, making Work products feel more like assistant entry points.

@thsottiaux X 原文229 ❤ · 2 RT · 50 💬

02 / 15

Rauch：Kimi K3 在 cyber stealth evals 上 top-tierOpen Cyber Capability Arrives

Rauch 根据内部 evals 表示 Kimi K3 在 cybersecurity 上 top-tier，且这些是 stealth evals，不是公开 benchmark 过拟合。他的结论是 frontier open-weight cybersecurity capability 已经到来。

Rauch says internal stealth evals show Kimi K3 is top-tier at cybersecurity, not merely overfitting public benchmarks. His takeaway: frontier open-weight cybersecurity capability is here.

@rauchg X 原文1,856 ❤ · 116 RT · 65 💬

Sol 在 cyber capability 上更大跃迁，但成本更高Capability Versus Cost

同一组 evals 里，Sol 在 cyber capability 上 leap ahead，但成本显著更高。这继续强化今天的模型栈问题：能力、成本、拒答率、可用性必须一起看。

In the same evals, Sol leaps ahead in cyber capability but at much higher cost. The model-stack question remains capability, cost, refusal rate, and usability together.

@rauchg X 原文1,856 ❤ · 116 RT · 65 💬

Fable refuses everythingSafety Tradeoff Exposed

Rauch 说他们无法让 Fable 完成 run，几乎全部拒答；相比之下 Sol 更愿意协助 defensive cyber hardening。这不是简单好坏，而是 safety policy 与实用性之间的产品权衡。

Rauch says they could not get Fable to complete the run because it refused everything; Sol was more open to defensive cyber hardening. This exposes a product tradeoff between safety policy and usefulness.

@rauchg X 原文1,856 ❤ · 116 RT · 65 💬

Zara：每个人都需要 personal eval setPersonal Evals Matter

Zara 建议每个人建立 personal eval set：几个真正贴近日常工作/生活的任务。行业 benchmark 有帮助，但不一定反映模型对你是否有用。

Zara suggests everyone build a personal eval set of tasks relevant to daily work and life. Industry benchmarks help, but may not reflect what makes a model useful to you.

@zarazhangrui X 原文130 ❤ · 6 RT · 20 💬

03 / 15

Aaron Levie：AI 价值不会只流向少数公司Healthy Ecosystem

Levie 认为过去几个月说明 AI 未来会更异质：frontier labs 继续推进模型边界，同时应用公司、垂直 lab、基础设施公司和服务 firm 把 AI 扩散进真实世界。

Levie argues recent months show AI value will not accrue only to a few companies. Frontier labs push model limits while applied AI companies, vertical labs, infrastructure, and services diffuse AI into the real world.

@levie X 原文327 ❤ · 43 RT · 52 💬

Gate keeping models 不会规模化有效Diffusion Over Lockdown

Levie 说 Kimi K3 之后，模型 gatekeeping 很难在规模上奏效。更好的路线是安全地保持高进展速度、扩散技术、建设基础设施并推动美国 OSS。

Levie says after Kimi K3, model gatekeeping will not work at scale. The better path is safely maintaining progress, diffusing technology, building infrastructure, and enabling US OSS.

@levie X 原文366 ❤ · 39 RT · 58 💬

Matt Turck：模型层仍未 commodityModel Layer Still Matters

Matt Turck 调侃 2024、2025、2026 都有人说模型层 commoditizing，但模型层仍未商品化。Kimi、Sol、Fable 的差异也说明模型层还在快速分化。

Matt Turck jokes that people keep saying the model layer is commoditizing, but it still is not. Kimi, Sol, and Fable differences show the model layer is still differentiating fast.

@mattturck X 原文82 ❤ · 4 RT · 23 💬

Zara：企业 AI 障碍是业务人与 AI 人互不理解Translation Gap

Zara 说企业 AI 采用最大障碍是懂 AI 的人不懂业务，懂业务的人不懂 AI。Levie 的服务 firm、应用层公司和 change management 观点正好回应这个 gap。

Zara says the biggest enterprise AI barrier is that AI people do not understand the business, and business people do not understand AI. Levie’s services, application layer, and change-management thesis addresses this gap.

@zarazhangrui X 原文242 ❤ · 27 RT · 60 💬

04 / 15

Rauch：AGI 这个词已经老化Superintelligence But Not Humanity

Rauch 认为“AGI”一词老化了：AI 在许多经济相关任务上比人强，但这不等于 AI 比 humanity 更好。人类 care for/about other humans 的能力仍是要保持的核心。

Rauch says AGI has aged poorly: AI is better than humans at many economically relevant tasks, but that does not make AIs better than humanity. Human care for other humans must remain central.

@rauchg X 原文577 ❤ · 39 RT · 91 💬

人真正变得无关紧要的方式，不是机器变强，而是人停止表达自己：放弃写作、判断、身份和 weighing in。People become irrelevant not because machines get stronger, but because they stop expressing themselves: writing, judging, identity, and weighing in.

AI replies 与模板化 landing page 正在让人反感Quality And Humanity Prevail

Rauch 批评社媒 AI replies 和套路化 landing page，认为 quality and humanity will prevail。越能自动生成，越要保护人的声音和审美。

Rauch criticizes AI replies on social media and formulaic landing pages, arguing quality and humanity will prevail. The more generation scales, the more human voice and taste matter.

@rauchg X 原文577 ❤ · 39 RT · 91 💬

05 / 15

Schmidhuber：真 AI 不只在屏幕后面Embodied AI Gap

Schmidhuber 在 Unsupervised Learning 中说，从宇宙视角看我们一直很接近 AI，但真正 AI 不只是屏幕后面的系统，还包括现实世界机器人和机械。机器人硬件仍远逊于人类身体。

Schmidhuber says true AI is not just behind the screen; it includes robots and machinery in the physical world. Robot hardware remains far inferior to human bodies.

Unsupervised Learning YouTube 播客2026-07-09

“你不能只有屏幕后面的 AGI。” 这句话给今天的 Work/agent 热潮补上一层现实约束：工具执行之外，还有身体、环境和物理世界。“You cannot have AGI just behind the screen.” It adds a physical constraint to today’s Work and agent boom: beyond tools, there are bodies, environments, and the real world.

他对当前模型结果并不惊讶Long History, Sudden Attention

Schmidhuber 认为近年结果对长期研究 neural networks 的人并不意外，因为 LLM 和相关训练思想有很长历史。ChatGPT moment 更像公众注意力突然转向。

Schmidhuber says recent results were less surprising to people who followed neural networks for decades; LLMs and training ideas have a long history. The ChatGPT moment was sudden public attention.

Unsupervised Learning YouTube 播客2026-07-09

06 / 15

人工科学家靠压缩与好奇心获得奖励Compression As Curiosity

Schmidhuber 描述人工科学家：当系统发现数据中的 regularity 并能更好压缩它，就产生 internal joy/reward，驱动 controller 设计更多实验来理解世界。

Schmidhuber describes artificial scientists: when a system discovers regularity and compresses data better, it receives internal joy or reward, motivating more experiments to understand the world.

Unsupervised Learning YouTube 播客2026-07-09

AI chemistry 已经有专用路径Specialized AI Scientists

他举化学为例：模型通过大量实验数据学习输入输出关系，再从目标输出反推实验建议。它不一定理解第一性原理，却能成为 intuitive chemist。

He uses chemistry as an example: models learn input-output relations from experiments, then work backward from desired outcomes to suggest experiments. They may not understand first principles but can become intuitive chemists.

Unsupervised Learning YouTube 播客2026-07-09

自设目标让系统更聪明，也更难预测Self-Generated Objectives

Schmidhuber 对 alignment 的单目标框架持怀疑态度，认为真正聪明的系统需要能提出自己的问题和目标；代价是更不可预测。

Schmidhuber is skeptical of one-objective alignment frames. Truly smart systems need freedom to invent their own questions and goals, at the cost of predictability.

Unsupervised Learning YouTube 播客2026-07-09

他对安全的担忧低于很多人Safety Through Scientific Interest

他认为未来更聪明的人工科学家会对生命、文明、自己起源等复杂 pattern 感兴趣，因此可能保护这些有趣模式。这个观点和主流 doom/safety 叙事形成鲜明对照。

He is less worried than many safety voices, arguing smarter artificial scientists may be fascinated by life, civilization, and their own origins, and thus motivated to protect interesting patterns.

Unsupervised Learning YouTube 播客2026-07-09

07 / 15

Swyx：欧洲仍有顶级 AI 工程师Global Talent Arena

Swyx 反驳看低欧洲 AI 的观点，认为欧洲有世界顶级 AI engineers，只要找到合适的人。AI Engineer 大会像全球竞技场，展示 talks/workshops 级别的人才密度。

Swyx pushes back on dismissing Europe, saying it has top AI engineers if you elicit the right ones. AI Engineer events function as a global arena for talent through talks and workshops.

@swyx X 原文53 ❤ · 1 RT · 21 💬

AEO 明年可能贡献 100 万收入Agent-Engine Optimization

Swyx 延续 AEO 线索，认为按当前速度 AEO 明年可能贡献 100 万美元收入。agent/search/answer surfaces 正在成为新增长渠道。

Swyx continues the AEO thread, saying at the current rate it may drive $1M in revenue next year. Agent, search, and answer surfaces are becoming growth channels.

@swyx X 原文6 ❤ · 0 RT · 2 💬

Peter Yang：ChatGPT Site 是个人软件分发Personal Software Distribution

乘法游戏案例说明，ChatGPT Sites 不只是企业建站，也可以是家长和孩子一起做的 personal software。分发入口越低，创作场景越私人化。

The multiplication game shows ChatGPT Sites are not only enterprise sites; they are personal software parents and children can build together. Lower distribution friction makes creation more personal.

@petergyang X 原文221 ❤ · 8 RT · 21 💬

Nikunj：细节体验仍然重要Taste In Physical Spaces

Nikunj 提到 Orangerie 的 kids room 很 thoughtful。即便 AI 工作流加速，现实空间里的体贴设计仍是体验质量的参照系。

Nikunj notes the thoughtful kids room at the Orangerie. Even as AI workflows accelerate, thoughtful physical design remains a reference point for experience quality.

@nikunj X 原文19 ❤ · 0 RT · 2 💬

08 / 15

Thariq：Fable access 背后是通宵努力

Thariq 说 Fable access 能按时做到，来自 Anthropic 许多人 literally around the clock 的 heroic effort。这给用户看到 plan 变化背后的 capacity work。

Thariq says Fable access happened through heroic, sometimes around-the-clock work at Anthropic. It reveals the capacity work behind plan updates.

@trq212 X 原文6,723 ❤ · 146 RT · 864 💬

Garry Tan：市场带来 abundance

Garry 今天多条偏公共政策，不是 AI 主线，但“abundance vs scarcity”的市场观点与 AI capacity 讨论同频：资源组织方式会决定扩散速度。

Garry’s market-abundance point is not directly AI, but it rhymes with AI capacity debates: how resources are organized shapes diffusion speed.

@garrytan X 原文160 ❤ · 10 RT · 16 💬

Matt Turck：模型层仍然没有商品化

Matt 的短评提醒，不要过早把 model layer 视作 commodity。今天的 Kimi/Sol/Fable cyber eval 正好说明差异仍大。

Matt reminds us not to prematurely call the model layer commoditized. Today’s Kimi/Sol/Fable cyber evals show differences remain large.

@mattturck X 原文82 ❤ · 4 RT · 23 💬

Aditya：资本主义热爱表达

Aditya 的短贴不是技术内容，但和今天“生态价值扩散、市场组织、模型竞争”主题有一点旁支关系。

Aditya’s short post is not technical, but loosely connects to today’s theme of ecosystem diffusion, markets, and model competition.

@adityaag X 原文98 ❤ · 2 RT · 6 💬

Peter Yang：World Cup 与 AI feed 混杂

体育和 AI builder feed 混在一起，提醒 digest 要区分高噪声个人动态和真正行业信号。

Sports posts mixed into AI builder feeds remind us that digesting means separating high-noise personal updates from industry signals.

@petergyang X 原文25 ❤ · 1 RT · 3 💬

Swyx：AEO 仍是未充分挖掘 alpha

Swyx 的 AEO 判断延续昨日“让 agent 自动研究 SEO/AEO”的观点，说明分发优化正在成为 agentic workflow 的一部分。

Swyx’s AEO view continues yesterday’s point about agents researching SEO/AEO automatically; distribution optimization is becoming part of agentic workflow.

@swyx X 原文6 ❤ · 0 RT · 2 💬

09 / 15

今日数据概览Today Stats

11 位活跃 Builder
23 条推文收录
1 期深度播客
0 篇博客
6,723 最高赞：Fable capacity effort
1,856 Cyber eval 讨论

11 builders, 23 tweets, 1 podcast, and 0 blogs. Top engagement came from Anthropic’s Fable capacity effort and the Kimi/Sol/Fable cybersecurity eval discussion.

follow-buildersSnapshot follow-builders-2026-07-19-v1

5 条关键洞察5 Key Takeaways

Work 产品进入委派时代：ChatGPT Work 的价值越来越像“把真实业务过程结构化交给 AI”。

Work products enter delegation mode: ChatGPT Work increasingly means handing structured business processes to AI.

Voice 是委派入口：复杂任务口述比打字更自然，尤其是 taxonomy、筛选、表格整理这类运营任务。

Voice is a delegation entry point: complex operational tasks are easier to speak than type.

Cyber eval 暴露模型差异：Kimi、Sol、Fable 在能力、成本、拒答率上的差异会影响真实部署。

Cyber evals expose model differences: Kimi, Sol, and Fable vary in capability, cost, and refusal behavior.

企业 AI 最大 gap 是翻译层：懂 AI 和懂业务的人需要共同定义 evals、workflows、risk 和 ROI。

The biggest enterprise gap is translation: AI and business experts must define evals, workflows, risk, and ROI together.

AGI 讨论需要物理世界约束：Schmidhuber 把 embodied robotics 和 artificial scientists 拉回长期视角。

AGI discussion needs physical constraints: Schmidhuber brings embodied robotics and artificial scientists back into long view.

10 / 15

CapEx boom 可能过热，但技术长期乐观Boom Versus Technology

播客简介提到 Schmidhuber 对 AI 技术乐观，但对 model companies 和 CapEx boom 更悲观。这和今天的 capacity/Work 增长形成平衡：真实需求强，但商业结构未必线性获利。

The episode frames Schmidhuber as optimistic on AI technology but pessimistic on model companies and the CapEx boom. This balances today’s Work and capacity growth story: demand is real, but business structures may not capture it linearly.

Unsupervised Learning YouTube 播客2026-07-09

Recursive self-improvement 不一定是公司 moatRecursive Improvement Diffuses

访谈简介提到，他不认为 recursive self-improvement 会成为公司 moat。若 AI 能帮设计下一代系统，能力可能扩散到工具链，而不是只留在某个模型公司。

The episode summary says he does not think recursive self-improvement will be a company moat. If AI helps design next systems, capability may diffuse into toolchains rather than remain inside one lab.

Unsupervised Learning YouTube 播客2026-07-09

机器人硬件仍是 AGI 缺口Hardware Still Matters

今天许多 builder 在谈软件 agent，但 Schmidhuber 提醒，physical AI 仍被硬件限制。真正进入现实世界，需要手、身体、传感、材料和能耗一起进步。

Many builders talk software agents today, but Schmidhuber reminds us physical AI is still hardware-limited. Real-world intelligence needs progress in hands, bodies, sensors, materials, and energy.

Unsupervised Learning YouTube 播客2026-07-09

Artificial scientists 已在专门领域存在Narrow Scientists First

他认为简单 artificial scientists 已存在，只是还没有 ChatGPT moment。化学、材料和实验自动化可能先出现专用突破，再逐步外溢。

He argues simple artificial scientists already exist, just without a ChatGPT moment. Chemistry, materials, and experiment automation may produce specialized breakthroughs first, then spill outward.

Unsupervised Learning YouTube 播客2026-07-09

11 / 15

入口层：Voice + WorkEntry Layer

Thibault 的 dictation 和 ChatGPT Work 任务展示了入口层变化：用户不再只问问题，而是口述目标、约束、格式和反馈样本。

Thibault’s dictation and ChatGPT Work task show an entry-layer shift: users no longer only ask questions; they dictate goals, constraints, formats, and feedback samples.

@thsottiaux X 原文136 ❤ · 1 RT · 40 💬

执行层：Sites、emails、docs、sheets、slidesExecution Layer

ChatGPT Work 和 Peter Yang 的 Site 案例说明，执行层正在覆盖从个人教育游戏到业务文档和 spreadsheet ops 的广泛任务。

ChatGPT Work and Peter Yang’s Site example show execution covering personal education games, business documents, and spreadsheet operations.

@petergyang X 原文221 ❤ · 8 RT · 21 💬

评估层：personal evals + stealth evalsEvaluation Layer

Rauch 的 internal cyber evals 和 Zara 的 personal eval set，分别代表平台方和个人用户如何判断模型边界。

Rauch’s internal cyber evals and Zara’s personal eval set show how platforms and individuals judge model boundaries.

@zarazhangrui X 原文130 ❤ · 6 RT · 20 💬

生态层：应用、垂直 labs、infra、servicesEcosystem Layer

Levie 的生态清单说明，AI 扩散不会只有模型公司胜出；应用层、基础设施、服务和垂直模型都会分工。

Levie’s ecosystem map shows AI diffusion will not only reward model companies; applications, infrastructure, services, and vertical models will all play roles.

@levie X 原文327 ❤ · 43 RT · 52 💬

12 / 15

今日之声

当 AI 从“回答问题”进入“委派工作”，真正的产品差异就变成：谁能接住上下文、执行工具、生成结构化产物，并让用户用自己的 eval 判断结果。

When AI moves from answering questions to delegated work, product differentiation becomes who can capture context, execute tools, produce structured artifacts, and let users judge results with their own evals.

@thsottiaux X 原文136 ❤ · 1 RT · 40 💬

13 / 15

AI前沿每日脉动

AI Frontier Pulse · 2026.07.19

今天的主线是“委派栈”：ChatGPT Work 把语音、工具和结构化产物连起来；模型 eval 暴露 Kimi/Sol/Fable 的差异；Schmidhuber 则提醒，真正 AI 还要穿过机器人、实验和物理世界。

Today’s thread is the delegation stack: ChatGPT Work connects voice, tools, and structured artifacts; model evals expose Kimi/Sol/Fable differences; Schmidhuber reminds us true AI still crosses robotics, experiments, and the physical world.

Richard Liu · AI前沿每日脉动 · 2026

14 / 15