AI前沿每日脉动

22SEP

AI前沿每日脉动

AI Frontier Pulse · 中英双语版 Bilingual Edition

数据截至北京时间 2026-09-22 14:42 · 历史 Feed 补发版。
Source updated Sep 22, 2026 14:42 China time. Backfilled historical feed edition.

2026.09.22 · 周二刊

10 位 Builder21 条推文0 期播客1 篇博客

ChatGPT WorkApplied AIOpen WeightsTokens per WattTrust & Access

Richard Liu · 2026

Curated by Richard Liu · Snapshot follow-builders-2026-09-22-v1

今天最强信号：GPT-5.6 Sol 只用两项运行设置变化，就在 ARC-AGI-3 达到 SoTA。模型能力不再能脱离 harness、compaction 与跨 context 执行单独评估。The strongest signal: two runtime setting changes put GPT-5.6 Sol at the top of ARC-AGI-3. Model capability can no longer be separated from harness design, compaction, and cross-context execution.

@thsottiaux X 原文12.5K ❤ · 760 RT · 1.7K 💬

Thibault Sottiaux：Ladies and gentlemen...Builder 动态 · Source Excerpt

Thibault Sottiaux 原文要点：Ladies and gentlemen...

Source-grounded excerpt from Thibault Sottiaux: Ladies and gentlemen...

@thsottiaux X 原文12.5K ❤ · 760 RT · 1.7K 💬

Peter Steinberger：Because this “Meta uses OpenClaw” story is going around.Builder 动态 · Source Excerpt

Peter Steinberger 原文要点：Because this “Meta uses OpenClaw” story is going around.

Source-grounded excerpt from Peter Steinberger: Because this “Meta uses OpenClaw” story is going around.

@steipete X 原文4.4K ❤ · 146 RT · 102 💬

02 / 15

Guillermo Rauch：I just gave Grok 4.7 a pretty hard problem, involving reveBuilder 动态 · Source Excerpt

Guillermo Rauch 原文要点：I just gave Grok 4.7 a pretty hard problem, involving reverse-engineering a running binary.

Source-grounded excerpt from Guillermo Rauch: I just gave Grok 4.7 a pretty hard problem, involving reverse-engineering a running binary.

@rauchg X 原文2.2K ❤ · 139 RT · 81 💬

Thariq：I now type "use big pictures and few words" several times Builder 动态 · Source Excerpt

Thariq 原文要点：I now type "use big pictures and few words" several times a day

Source-grounded excerpt from Thariq: I now type "use big pictures and few words" several times a day

@trq212 X 原文1.8K ❤ · 32 RT · 132 💬

03 / 15

AI 降低执行成本，却没有消除品味与经验：资深创始人的判断继续复利，设计工具也必须把审美、易用性和乐趣一起产品化。AI lowers execution cost without eliminating taste or experience. Seasoned judgment compounds, while design tools must productize aesthetics, ease, and fun together.

@steipete X 原文1.1K ❤ · 51 RT · 110 💬

Peter Steinberger：The beauty of running a claw yourself: they cannot block yBuilder 动态 · Source Excerpt

Peter Steinberger 原文要点：The beauty of running a claw yourself: they cannot block you.

Source-grounded excerpt from Peter Steinberger: The beauty of running a claw yourself: they cannot block you.

@steipete X 原文1.1K ❤ · 51 RT · 110 💬

Peter Steinberger：Proud to say that we did our homework.Builder 动态 · Source Excerpt

Peter Steinberger 原文要点：Proud to say that we did our homework.

Source-grounded excerpt from Peter Steinberger: Proud to say that we did our homework.

@steipete X 原文433 ❤ · 15 RT · 24 💬

04 / 15

Peter Yang：I think we're in for a rude awakening in the ad markets: IBuilder 动态 · Source Excerpt

Peter Yang 原文要点：I think we're in for a rude awakening in the ad markets: If a huge part of your business is showing targeted display ads to humans what happens when agents browse your website a…

Source-grounded excerpt from Peter Yang: I think we're in for a rude awakening in the ad markets: If a huge part of your business is showing targeted display ads to humans what happens when agents browse your website a…

@petergyang X 原文379 ❤ · 21 RT · 91 💬

Garry Tan：Not really sure how @capydotai does it, but it really is a工作产品化 · Source Excerpt

Garry Tan 原文要点：Not really sure how @capydotai does it, but it really is able to track multi-step workflow and do large PRs faster than Codex or Claude Code on its own.

Source-grounded excerpt from Garry Tan: Not really sure how @capydotai does it, but it really is able to track multi-step workflow and do large PRs faster than Codex or Claude Code on its own.

@garrytan X 原文356 ❤ · 22 RT · 56 💬

Garry Tan：If you really think about it Cluely as a realtime thought Builder 动态 · Source Excerpt

Garry Tan 原文要点：If you really think about it Cluely as a realtime thought helper and semi-adversarial assistant with ongoing context is still a good idea

Source-grounded excerpt from Garry Tan: If you really think about it Cluely as a realtime thought helper and semi-adversarial assistant with ongoing context is still a good idea

@garrytan X 原文213 ❤ · 5 RT · 37 💬

Nikunj Kothari：I have been using Codex as my primary "agent" for such a l工作产品化 · Source Excerpt

Nikunj Kothari 原文要点：I have been using Codex as my primary "agent" for such a long time that I forget the average person has never used it..

Source-grounded excerpt from Nikunj Kothari: I have been using Codex as my primary "agent" for such a long time that I forget the average person has never used it..

@nikunj X 原文164 ❤ · 3 RT · 23 💬

05 / 15

今天的深度信号来自官方博客：AI 工具正在把工作过程转成更可协作的界面。Today's depth signal comes from the official blog: AI tools are turning work into more collaborative interfaces.

Claude Blog 原文链接

Claude in Chrome is generally available Today's Deep Blog

今天的博客信号来自官方长文：AI 工具正在把上下文、工具调用和协作流程整合到更完整的工作界面中。

Today's blog signal points to AI tools integrating context, tool use, and collaboration into more complete work interfaces.

Claude Blog 原文链接Aug 26, 2026

06 / 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.

07 / 15

算力上线即被消耗New Compute Is Consumed Immediately

新增容量上线即被工作负载吸收，过早放慢建设通常会带来负面意外。

New capacity is consumed immediately, and slowing the build too early repeatedly creates negative surprises.

关键指标：每瓦 token 数The Metric Is Tokens Per Watt

定制芯片与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。

Custom silicon and model workloads are co-designed to maximize tokens generated per watt.

数据中心必须新增电力Data Centers Must Add Power

大规模 AI 基础设施需要同步投资发电、输电、变电站与配套容量。

Large-scale AI infrastructure requires new generation, transmission, substations, and supporting capacity.

算力策略是 All of the AboveThe Strategy Is All Of The Above

多种云、芯片、数据中心与供应路径并行推进，减少单点约束。

Clouds, chips, data centers, and supply paths advance in parallel to reduce single-point constraints.

08 / 15

Peter Yang：I think we're in for a rude awakening in the ad markets: IBuilder 动态 · Source Excerpt

@petergyang X 原文379 ❤

Garry Tan：Not really sure how @capydotai does it, but it really is a工作产品化 · Source Excerpt

Garry Tan 原文要点：Not really sure how @capydotai does it, but it really is able to track multi-step workflow and do large PRs faster than Codex or Claude Code on its own.

Source-grounded excerpt from Garry Tan: Not really sure how @capydotai does it, but it really is able to track multi-step workflow and do large PRs faster than Codex or Claude Code on its own.

@garrytan X 原文356 ❤

Garry Tan：If you really think about it Cluely as a realtime thought Builder 动态 · Source Excerpt

Garry Tan 原文要点：If you really think about it Cluely as a realtime thought helper and semi-adversarial assistant with ongoing context is still a good idea

Source-grounded excerpt from Garry Tan: If you really think about it Cluely as a realtime thought helper and semi-adversarial assistant with ongoing context is still a good idea

@garrytan X 原文213 ❤

Nikunj Kothari：I have been using Codex as my primary "agent" for such a l工作产品化 · Source Excerpt

Nikunj Kothari 原文要点：I have been using Codex as my primary "agent" for such a long time that I forget the average person has never used it..

Source-grounded excerpt from Nikunj Kothari: I have been using Codex as my primary "agent" for such a long time that I forget the average person has never used it..

@nikunj X 原文164 ❤

Guillermo Rauch：In addition to our type-safe TypeSafe @aisdk API (get it?)模型网关 · Source Excerpt

Guillermo Rauch 原文要点：In addition to our type-safe TypeSafe @aisdk API (get it?), you can now use Jev over HTTP in AI Gateway

Source-grounded excerpt from Guillermo Rauch: In addition to our type-safe TypeSafe @aisdk API (get it?), you can now use Jev over HTTP in AI Gateway

@rauchg X 原文91 ❤

Aaron Levie：AI agents will use software 100X more than people ever didAgent 安全 · Source Excerpt

Aaron Levie 原文要点：AI agents will use software 100X more than people ever did.

Source-grounded excerpt from Aaron Levie: AI agents will use software 100X more than people ever did.

@levie X 原文69 ❤

09 / 15

今日数据概览Today Stats

收录 Builder：10 位
总推文数：21 条
播客节目：0 期
博客文章：1 篇

最高互动：Thibault Sottiaux：Ladies and gentlemen... · 12.5K ❤
第二高互动：Peter Steinberger：Because this “Meta uses OpenClaw” story is going around. · 4.4K ❤

The snapshot includes 10 builders, 21 tweets, 0 podcast, and 1 blog post.

follow-buildersSnapshot follow-builders-2026-09-22-v1

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

本周之声 VOICE 01

只要允许模型持续推理，并借助 canonical compaction 跨多个 context window 工作。

“You just have to allow it to reason and work over multiple context windows.”

@thsottiaux X 原文12.5K ❤ · 760 RT · 1.7K 💬

11 / 15

本周之声 VOICE 02

最好的路径，是赋能科学家，而不是试图由我们自己解决所有问题。

“The best way to do this is for us to empower scientists.”

@rauchg X 原文2.2K ❤ · 139 RT · 81 💬

12 / 15

本周之声 VOICE 03

我总会先创建 design.md，写清颜色、字体以及其他设计准则。

“I always create a design md file with guidelines for color, typography, and more.”

@garrytan X 原文356 ❤ · 22 RT · 56 💬

13 / 15

播客之声 PODCAST VOICE

AI 是一场“50 年的一夜成功”。

“AI's fifty year overnight success.”

14 / 15

AI前沿每日脉动

AI Frontier Pulse · 2026.09.22

本期收录 10 位 Builder · 21 条推文 · 0 期播客 · 1 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-2026-09-22-v1