AI Builders Digest — 2026-09-25

2026-09-25

AI Builders Digest — 2026-09-25

X / TWITTER

Thibault Sottiaux, Codex & ChatGPT at OpenAI

Sottiaux says next Tuesday's DevDay follows OpenAI's most ambitious sprint yet, with Astra enabling products that could materially change how people work. He also describes voice as the connective layer across ChatGPT, email, coding, calendars, and third-party plugins, pointing toward a conversational operating interface rather than a standalone chatbot.

Sottiaux 表示，下周二的 DevDay 建立在 OpenAI 迄今最激进的一轮冲刺之上，Astra 让一批可能实质改变工作方式的产品成为可能。他还把 voice 描述为贯穿 ChatGPT、邮件、coding、日历和第三方 plugins 的连接层，这指向的是一个对话式操作界面，而不只是独立 chatbot。

Claude / Anthropic

Claude launched a marketplace for discovering connectors, plugins, agents, products, and expert partners. The initial ecosystem spans tools such as Slack and Notion, products from Cursor and CrowdStrike, and service partners including Accenture and Deloitte, signaling a shift from model distribution toward a full commercial agent ecosystem.

Claude 推出了 marketplace，用于发现 connectors、plugins、agents、产品和专家服务伙伴。首批生态覆盖 Slack、Notion 等工具，Cursor、CrowdStrike 等产品，以及 Accenture、Deloitte 等服务商，说明竞争正从模型分发转向完整的商业 agent 生态。

Guillermo Rauch, Vercel CEO

Rauch frames successful agents as three separable layers: brain (model and harness), hands (tools, browser, computer), and files (memory, skills, repos). Vercel's new Drives product decouples persistent storage from agent compute, which he argues cuts cloud cost while improving security and auditability; he also reports Opus 5.5 finding shell-startup optimizations missed by other models.

Rauch 将成功的 agent 拆成三个可分离层：brain（model 与 harness）、hands（tools、browser、computer）和 files（memory、skills、repos）。Vercel 新推出的 Drives 将持久化存储与 agent compute 解耦，他认为这既能降低云端成本，也能提升安全性和可审计性；他还提到 Opus 5.5 找到了其他模型遗漏的 shell 启动优化。

Boris Cherny, Claude Code at Anthropic

Cherny explains Claude's practical use of formal methods: it models tricky state machines or race-prone code, searches for counterexamples, reproduces suspected bugs, and fixes them. This is targeted verification of the riskiest code paths rather than proof of an entire codebase; he also points engineers to a detailed account of recent Claude web and desktop performance improvements.

Cherny 解释了 Claude 如何在工程中实际使用 formal methods：针对复杂状态机或容易出现 race condition 的代码建模，寻找反例，复现疑似 bug，再修复代码。这并非验证整个 codebase，而是集中检查风险最高的路径；他还分享了近期 Claude Web 与桌面端性能优化的详细工程复盘。

Thariq, Claude Code at Anthropic

Thariq cautions that claims such as “Claude one-shot this” often hide substantial scaffolding: long prompts, strong examples, skills, and API access. Anthropic is responding with a new format that publishes replicable prompts and techniques, making the harness and workflow visible instead of attributing the outcome solely to the model.

Thariq 提醒，所谓“Claude 一次生成完成”的案例背后，往往隐藏着大量 scaffolding：长 prompt、优质示例、skills 和 API access。Anthropic 正尝试发布可复现的 prompts 与技巧，让 harness 和 workflow 透明化，而不是把成果全部归因于模型本身。

Garry Tan, Y Combinator President and CEO

Tan identifies two simultaneous go-to-market shifts: builders must make software that agents want to use, while also using agents to make people want software. The compact formulation suggests that agent compatibility is becoming both a product-distribution surface and a customer-acquisition capability.

Tan 提出两条同时发生的 go-to-market 变化：一方面要让软件成为 agents 愿意使用的对象，另一方面要用 agents 激发人们对软件的需求。这意味着 agent compatibility 正同时变成产品分发入口和获客能力。

Madhu Guru, Senior Director of AI at Meta

Guru argues that consumers do not have one universal preference for “doing things.” Some journeys are enjoyable to browse, but high-friction tasks such as hiring a roofer are ideal for near-zero-effort agents; the opportunity lies in distinguishing experiences users want to explore from chores they want delegated.

Meta AI Senior Director Madhu Guru 认为，消费者对“亲自完成事情”并不存在统一偏好。有些购物过程本身有乐趣，但寻找屋顶维修商这类高摩擦任务非常适合交给近乎零操作的 agent；真正的机会在于区分用户想探索的体验与想彻底委托的杂务。

Peter Yang, AI educator and interviewer

Yang says the market currently splits excellence between labs: one has the strongest model, another the strongest harness. His product implication is direct: Claude Code needs first-class live voice plus stronger browser and computer use to turn model quality into a complete execution experience; he praises Opus 5.5 for speed, writing, conversational quality, and usable rate limits.

Peter Yang 认为当前市场的优势分裂在不同 labs：一家拥有最强 model，另一家拥有最强 harness。他给出的产品判断很直接：Claude Code 需要一流的实时 voice，以及更强的 browser 和 computer use，才能把模型能力转化为完整执行体验；他同时肯定 Opus 5.5 的速度、写作、对话质量和可用的 rate limits。

Ryo Lu, former Cursor, Notion, and Stripe designer

Lu challenges the industry's worship of speed, scale, and automated output. His central warning is that AI may not make people lazy but endlessly busy, creating infinite execution before intention and taste; the more valuable frontier may be discernment, including knowing what not to build and when to stop.

曾任职于 Cursor、Notion 和 Stripe 的设计师 Ryo Lu 质疑行业对速度、规模和自动化产出的崇拜。他的核心警告是：AI 可能不会让人懒惰，却会让人永远忙碌，在意图和品味成熟之前就带来无限执行力；更有价值的前沿或许是 discernment，即知道什么不该做、何时停手。

Aditya Agarwal, SPC General Partner and former Dropbox CTO

Agarwal argues that team longevity is a better predictor of throughput and resilience than team size. A group that has worked together for more than three years can outperform a newly assembled team, so investors and employees should evaluate the “area under the curve” of shared operating history rather than headcount alone.

SPC General Partner、前 Dropbox CTO Aditya Agarwal 认为，团队共同工作的时长比人数更能预测吞吐量和韧性。协作超过三年的团队可能显著优于新组建团队，因此投资人和员工应评估团队共同经历的“曲线下面积”，而不只是 headcount。

Aaron Levie, Box CEO

Levie argues that lower production costs will expand creative opportunity rather than eliminate it: more films, more risk-taking, and new storytelling forms. His broader claim is that technology changes techniques and media, but creative skill and taste remain scarce; Box's pairing with Muse hints at how content systems may become part of AI-native creative workflows.

Box CEO Aaron Levie 认为，制作成本下降会扩大而非消灭创作机会：更多影片、更多风险尝试，以及新的叙事形式。他更广泛的判断是，技术会改变技法和媒介，但创造力与品味仍然稀缺；Box 与 Muse 的结合也暗示内容系统将进入 AI-native 创意 workflow。

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders