jalen.cool
全部内容
ai builders

June 22, 2026

11 builders27 posts1 podcast

Google DeepMind's Logan Kilpatrick argues on Training Data that 'the model is not the model anymore' but an expanding system that will eat the agent harness within a year, with Antigravity becoming Google's new through line and AI Studio users shipping ~350K Android apps in a single week; Box CEO Aaron Levie reads Sakana's Fugu mixture-of-models router and an agents-use-software-100X-more future as proof the routing and headless-platform layers will capture enormous value; Vercel CEO Guillermo Rauch warns coding agents 'will squeeze every ounce of IKEA effect out of you'; YC's Garry Tan pitches open-sourced GBrain as the personal-context unlock for usable AGI; Linear's Nan Yu argues 'quality is irrational'; Zara Zhang says input context should run 3-5x longer than output to avoid AI slop; OpenClaw's Peter Steinberger reports his strongest week yet as a non-profit while doubting multi-model routing; Peter Yang unpacks why HTML is the LLM's native language for agentic video; and Cursor's Ryo Lu ships a hand-tuned EPUB reader built in Cursor mobile.

X / TWITTER

Peter Yang (AI educator and newsletter writer) — petergyang on X

Peter Yang highlighted why HTML turned out to be the foundation for agentic video-making, quoting builder Liu Bin (liu8in): teams trying to build a video agent "learned the hard way that agents have no visual intelligence," so they turned to code. "HTML is the LLM's native language," letting models express not just information but visual aesthetics through HTML, CSS, and JavaScript, with footage, images, and SVGs all sitting on top. As an example, he showed how to turn any website URL into a product video.

AI 教育者、newsletter 作者 Peter Yang 解释了为什么 HTML 最终成了 agentic 视频制作的基础,并引用了 builder Liu Bin(liu8in)的话:那些想做 video agent 的团队"用很痛的方式学到了一件事——agent 没有视觉智能",于是他们转向了代码。"HTML 是 LLM 的母语",让模型不仅能表达信息,还能通过 HTML、CSS 和 JavaScript 表达视觉美感,而画面素材、图片、SVG 都可以叠在 HTML 之上。作为例子,他展示了如何把任意一个网站 URL 变成一段产品视频。

Nan Yu (Head of Product, Linear) — thenanyu on X

Linear's Head of Product Nan Yu endorsed the framing that "quality is irrational": shipping truly great products takes an irrational level of commitment to constantly choose quality, and an irrational level of self-belief that controlling things top to bottom will yield better results than leaning on common frameworks.

Linear 的产品负责人 Nan Yu 认同"quality is irrational(质量是非理性的)"这个说法:做出真正出色的产品,需要一种近乎非理性的执念去持续选择质量,以及一种近乎非理性的自信,相信从头到尾自己掌控一切,会比套用通用框架做得更好。

Guillermo Rauch (Vercel CEO) — rauchg on X

Vercel CEO Guillermo Rauch shared a line that resonated widely: "Coding agents will squeeze every ounce of IKEA effect out of you, if you let them" — a caution that the pride of building something yourself can keep you over-involved when the agent could just do the work. He also detailed a deep performance pass in which the team scrutinized every frame (painting, layout, WebGPU shaders, blocking scripts) and said they will fold the lessons learned back into their public docs.

Vercel CEO Guillermo Rauch 分享了一条引发广泛共鸣的话:"如果你放任它们,coding agent 会把你身上'宜家效应'的每一丝都榨干",提醒大家:那种亲手做出东西的成就感,可能让你在 agent 本可以代劳时仍然过度参与。他还详细讲了一次深度性能优化,团队对每一帧都做了打磨(绘制、布局、WebGPU shader、阻塞脚本),并表示会把这次学到的经验更新进他们的公开文档里。

Aaron Levie (Box CEO) — levie on X

Box CEO Aaron Levie flagged two threads. On Sakana's newly released Fugu model, he noted it uses a mixture of models behind a single API, farming each task out to the model that performs it best and handling "model selection, delegation, verification, and synthesis automatically" — and argued the layer that can route to the best model will produce enormous value as both closed and open models proliferate. Separately, he predicted "agents will use software 100X more than people," which creates huge demand for guardrails, authoritative sources of truth, and logging and auditing; the platforms that can power these headless, agent-driven interactions will be best positioned.

Box CEO Aaron Levie 提到了两条线索。关于 Sakana 刚发布的 Fugu 模型,他指出它在单一 API 背后用了一套 mixture of models,把每个任务分派给最擅长它的模型,并"自动完成模型选择、分派、验证和综合"。他认为,随着闭源和开源模型不断涌现,那个能把任务路由到最佳模型的层会产生巨大价值。另外,他预测"agent 使用软件的频率会是人类的 100 倍",这会催生对 guardrails、权威事实来源、以及日志和审计的巨大需求;那些能够支撑这种无人值守、由 agent 驱动的交互的平台,将处于最有利的位置。

Ryo Lu (Designer, Cursor) — ryolu_ on X

Cursor designer Ryo Lu shipped "Books," a wooden-shelf e-reader for his ryOS project that works with any EPUB and syncs reading progress to a ryOS account. He started it in Cursor mobile, then hand-tuned the animations and textures until they felt right.

Cursor 的设计师 Ryo Lu 上线了"Books",一个为他的 ryOS 项目做的木质书架风格阅读器,支持任意 EPUB,并能把阅读进度同步到 ryOS 账号。他先在 Cursor mobile 上起的头,然后手动打磨动画和材质质感,直到感觉对了为止。

Garry Tan (Y Combinator President & CEO) — garrytan on X

Y Combinator's Garry Tan argued that an underestimated unlock at the "dawn of usable AGI" in 2026 is personal and company context: "AGI gives you the intelligence. You still have to collect your personal context to get the real unlock." He pointed to GBrain, the personal/company "brain" tool he built and open-sourced, as his answer.

Y Combinator 总裁兼 CEO Garry Tan 认为,在 2026 年"可用 AGI 的曙光"里有一个被低估的关键解锁点,那就是个人和公司的 context:"AGI 给了你智能,但你仍然需要去收集自己的个人 context,才能拿到真正的解锁。"他提到了自己打造并开源的个人/公司"大脑"工具 GBrain,作为他给出的答案。

Zara Zhang (AI builder) — zarazhangrui on X

Builder Zara Zhang offered a rule of thumb for preventing AI slop in writing and design: ask whether your input (the context) is longer than the output. She finds that for AI to produce quality results, her input is often 3-5x the length of the output — and if the input is much shorter than the output, the result is almost certainly slop.

Builder Zara Zhang 给出了一条避免 AI slop(无论写作还是设计)的经验法则:问问自己,你的输入(也就是 context)是不是比输出更长。她发现,要让 AI 产出高质量结果,她的输入往往是输出长度的 3 到 5 倍;如果输入比输出短得多,那产出几乎注定是 slop。

Peter Steinberger (OpenClaw founder) — steipete on X

Peter Steinberger gave a state-of-the-project update on OpenClaw: after the initial hype faded, the team focused on quality and grew, and deliberately structured itself as a non-profit, in contrast to VC-funded competitors he says have "other agendas" — calling it their strongest week so far. He also voiced skepticism about multi-model routing, saying his hunch that it underdelivers seems to have been right.

Peter Steinberger 给出了 OpenClaw 的项目近况:在最初的热度退去之后,团队把精力放在了打磨质量和扩充人手上,并刻意把自己做成了一个非营利组织,与那些他认为"另有所图"的、靠 VC 投资的竞争对手形成对比,他称这是迄今为止他们最强的一周。他还表达了对 multi-model routing 的怀疑,说他"它并不好用"的直觉看来是对的。

PODCASTS

Training Data — Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness

The Takeaway: The agent harness everyone is racing to build is a temporary moat; within a year, the model itself will swallow it.

Logan Kilpatrick, who runs Google AI Studio and the Gemini API, makes a counterintuitive case about where value lives in AI. His core claim: "the model" no longer means a set of weights. "It's an entire expanding sprawling system that's built around the weights" — agentic tool calling, hosted search, code execution, containers, a harness. The pattern repeats relentlessly: scaffolding gets built a few steps ahead of the model, then the model eats it and the alpha moves elsewhere. So today's obsession with proprietary harnesses, he predicts, won't matter much in twelve months.

For Google, this rewires the company. Gemini was the "through line" connecting its 50-plus products; now Antigravity — Google's agent harness, spanning an IDE, web app, CLI, and SDK — is becoming the new through line, powering agentic features across Search, the Gemini app, and Cloud. Yet he's blunt that most Google products are still at "crawl," not "walk" or "run," because billions of users aren't ready to just let AI run loose.

The specifics land hardest. AI Studio users built roughly 350,000 Android apps in a single week — apps "that probably no one was going to build before." Coding already feels like "narrow superintelligence," and he expects the next breakthroughs in verticals with clear verifiability: math, finance, science. He also reframes the personal cost of all this capability: "I'm kicking around an idea, and I'm like, I could probably make this even more ambitious." Rather than fearing the model eats everything, he insists focus is still the startup superpower, and that there has never been more room to build.

核心要点: 所有人都在拼命打造的 agent harness 只是一道暂时的护城河;不出一年,模型自己就会把它吞掉。

Logan Kilpatrick 负责 Google AI Studio 和 Gemini API,他对"AI 的价值究竟在哪里"给出了一个反直觉的判断。他的核心观点是:"模型"早已不再只是一堆权重。"它是一整套围绕权重不断膨胀、不断蔓延的系统",包括 agentic 的工具调用、托管的 search、代码执行、容器,以及一层 harness。这个规律反复上演:脚手架总是比模型领先几步被搭出来,然后模型把它吃掉,价值高地就转移到别处。所以他预测,如今大家对自研 harness 的执着,再过十二个月就没那么重要了。

对 Google 来说,这件事重塑了整家公司。过去 Gemini 是把它五十多个产品串起来的"主线";现在 Antigravity,也就是 Google 的 agent harness,横跨 IDE、网页端、CLI 和 SDK,正在成为新的主线,为 Search、Gemini app 和 Cloud 里的 agentic 功能提供支撑。但他也直言,大多数 Google 产品仍处在"爬"的阶段,而不是"走"或"跑",因为面对数十亿用户,大家还没准备好任由 AI 撒手去干。

最有冲击力的是那些具体数字。AI Studio 的用户在短短一周里就做出了大约 35 万个 Android app,而这些"大概本来没人会去做"。他说,coding 现在已经有点像"narrow superintelligence(窄域超级智能)",并预计下一批突破会出现在那些可验证性强的垂直领域:数学、金融、科学。他还重新诠释了这种强大能力带来的个人代价:"我脑子里冒出一个想法,然后我会想,我大概可以把它做得更有野心一点。"比起担心模型会把一切吞掉,他坚持认为专注仍然是创业公司的超能力,而且现在可以构建的空间,比以往任何时候都更大。

Generated through the Follow Builders skill — daily bilingual signal from the people actually building AI.