AI Builders Digest — 2026-09-16

X / TWITTER

Josh Woodward — VP, Google Labs / Gemini

Google VP Josh Woodward says the Gemini power-user group has tested more than 20 early features in two months. Google is now opening another cohort for upcoming Daily Brief and Personal Intelligence capabilities, showing a product-development loop built around continuous user testing rather than one-off previews.

Google VP Josh Woodward 表示，Gemini 核心用户组在两个月内已测试 20 多项早期功能。Google 正在为新一批用户开放 Daily Brief 和 Personal Intelligence 的后续能力，体现出一种围绕持续用户测试，而非一次性预览的产品开发机制。

Source / 原文：https://x.com/joshwoodward/status/2099558443078365287

Boris Cherny — Claude Code, Anthropic

Anthropic's Boris Cherny announced that Claude Mods are rolling out, with the community already building experiments such as a Tetris-in-Claude mod. The update points toward Claude Code becoming an extensible runtime, not merely a fixed coding interface.

Anthropic 的 Boris Cherny 宣布 Claude Mods 正在上线，社区已经做出了 Tetris-in-Claude 等实验。这意味着 Claude Code 正从固定的编程界面，逐步演化为可扩展的运行环境。

Source / 原文：https://x.com/bcherny/status/2099551291601248485

Thibault Sottiaux — Codex & ChatGPT, OpenAI

OpenAI's Thibault Sottiaux teased a week of releases comparable to DevDay 2025, while separately asking users which obsolete Codex feature should be removed. Taken together, the posts signal both a major shipping cycle and an active effort to reduce product complexity.

OpenAI 的 Thibault Sottiaux 预告，本周发布密度将堪比 DevDay 2025；同时，他公开询问用户 Codex 中哪些过时功能应该移除。两条动态合在一起，释放出两个信号：产品将进入密集发布周期，也在主动削减复杂度。

Sources / 原文：
- https://x.com/thsottiaux/status/2099744972195131850
- https://x.com/thsottiaux/status/2099393115241300166

Thariq — Claude Code, Anthropic

Anthropic engineer Thariq recorded a technical Latent Space discussion about topics the Claude Code team has not previously covered in depth. He also reflected on building Claude Code, including the difficulty of keeping the product aligned with rapidly improving model capabilities and what has been lost from pre-AI software engineering.

Anthropic 工程师 Thariq 录制了一期偏技术的 Latent Space 节目，涉及 Claude Code 团队此前很少深入公开的话题。他还讨论了 Claude Code 的构建过程，包括产品如何艰难追赶快速提升的模型能力，以及 AI 时代的软件工程失去了什么。

Sources / 原文：
- https://x.com/trq212/status/2099671266068496802
- https://x.com/trq212/status/2099551141621329994

Guillermo Rauch — CEO, Vercel

Vercel CEO Guillermo Rauch argues that agents are only as reliable as the verifiers around them: proof-checkers, compilers, type systems, and linters. He frames “verifiers + skills” as the new frameworks, citing shadcn/lint as a way to keep agents inside design-system rules. Vercel also hired Google Cloud Run creator Steren to lead its Fluid compute products, based on the thesis that agents need a new generation of compute primitives.

Vercel CEO Guillermo Rauch 认为，agent 的可靠性取决于其周围的验证机制，包括 proof-checker、compiler、type system 和 linter。他把“verifiers + skills”定义为新一代 framework，并以 shadcn/lint 为例，说明如何让 agent 遵守设计系统规则。Vercel 同时邀请 Google Cloud Run 创始人 Steren 领导 Fluid 计算产品，核心判断是 agent 需要新一代计算原语。

Sources / 原文：
- https://x.com/rauchg/status/2099540886409695346
- https://x.com/rauchg/status/2099514906366328902
- https://x.com/rauchg/status/2099653035685445760

Aaron Levie — CEO, Box

Box CEO Aaron Levie predicts agentic workloads will become roughly 100 times larger than today's prompt-driven usage as swarms, computer use, APIs, MCPs, and vertical agents spread. That scale creates a paired governance problem: excessive access threatens enterprise data, while excessive restriction eliminates productivity gains. Box is responding with classification-based controls in Box Shield and anomaly detection for agent access.

Box CEO Aaron Levie 预测，随着 agent swarm、computer use、API、MCP 和垂直 agent 普及，agent 工作负载可能达到今天 prompt 驱动使用量的约 100 倍。这会带来一对相互牵制的治理难题：权限过大会威胁企业数据，限制过严又会消灭生产力收益。Box 正通过 Box Shield 的文档分级控制和 agent 异常访问检测应对这一问题。

Sources / 原文：
- https://x.com/levie/status/2099739019517235618
- https://x.com/levie/status/2099550035239424465

Matt Turck — Partner, FirstMark

FirstMark partner Matt Turck argues that AI progress will not meaningfully slow because too many competitors, economic incentives, and national-level prisoner’s dilemmas are pushing in the same direction. His point is less a product forecast than a structural claim: coordination to pause is unlikely even after periods of public concern.

FirstMark 合伙人 Matt Turck 认为，AI 进展不会实质性放缓，因为大量竞争者、经济激励和国家层面的“囚徒困境”都在推动同一个方向。这与其说是产品预测，不如说是结构性判断：即便公众阶段性担忧加剧，各方也很难协调暂停。

Source / 原文：https://x.com/mattturck/status/2099589199104033031

Nikunj Kothari — Partner, FPV Ventures

FPV Ventures partner Nikunj Kothari announced a Series A investment in Piston, which replaces fleet fuel cards with a closed payment rail. Piston ties each transaction to a driver, time, location, and fuel type; Kothari says payment volume grew 8x, the merchant network 40x, retention remains above 98%, and the service now reaches 2,000 stations across 48 states. The core insight is that fraud cannot be fixed by making the card better when the card itself is the weak point.

FPV Ventures 合伙人 Nikunj Kothari 宣布领投 Piston 的 Series A。Piston 用闭环支付网络取代车队燃油卡，并把每笔交易绑定到司机、时间、地点和燃油类型。Kothari 称，其支付规模增长 8 倍、商户网络增长 40 倍、留存率超过 98%，目前已覆盖美国 48 个州的 2,000 个加油站。核心判断是：当卡本身就是薄弱环节时，仅仅改良卡无法解决欺诈。

Source / 原文：https://x.com/nikunj/status/2099631145268969840

PODCASTS

AI & I by Every — How a Professional Writer Writes With AI

The Takeaway: Professional writer Katie Parrott treats AI as a compounding system: preserve context, codify each useful correction once, and bring fresh evidence and lived experience that the model cannot supply.

Parrott moved from using ChatGPT as an inexpensive career coach to operating a Codex-based “chief of staff” containing her role, company positioning, OKRs, performance evidence, reader feedback, and a Kanban board. The system helps her prioritize work and is increasingly operated by voice. Her writing workflow follows the same logic: durable context files capture audience, positioning, style, and process, while each new piece receives fresh research and personal insight. As she puts it, “the model is the kitchen,” but the writer must provide good ingredients.

Her Compound Writing plugin adapts Compound Engineering into brainstorming, outlining, drafting, substantive editing, line editing, and final review. Feedback is codified so the next output improves, while reviewer skills based on Vonnegut, Hitchcock, Sorkin, and Sedaris supply distinct critical lenses. The larger lesson is that AI does not remove the hard work of judgment; it makes expert processes reusable and gives writers more energy for argument, taste, and originality. Parrott's final warning is distributive: if access, education, time, and experimentation remain concentrated, AI's compounding gains will also concentrate among early adopters.

核心结论：职业作家 Katie Parrott 把 AI 当作一个复利系统：长期保留 context，把每次有效纠正只编码一次，再持续加入模型无法自行获得的新证据和真实经验。

Parrott 最初把 ChatGPT 当作低成本职业教练，如今则运行着一个基于 Codex 的“chief of staff”。其中包含她的岗位信息、公司定位、OKR、业绩证据、读者反馈和 Kanban 看板，能帮助她排序任务，并越来越多地通过语音操作。她的写作流程遵循同一逻辑：持久化 context 文件负责记录受众、定位、风格和流程，每篇新文章则加入最新研究与个人洞察。正如她所说：“模型是厨房”，但作者必须提供优质食材。

她的 Compound Writing plugin 把 Compound Engineering 改造成适用于写作的 brainstorming、outline、draft、substantive edit、line edit 和 final review 流程。反馈会被编码，使下一次输出自动改进；基于 Vonnegut、Hitchcock、Sorkin 和 Sedaris 的 reviewer skills，则提供不同的批评视角。更大的启示是：AI 没有消除判断的艰苦工作，而是让专家流程可以复用，并把更多精力释放给论证、品味与原创性。Parrott 最后的警告关乎机会分配：如果 access、教育、时间和实验条件仍高度集中，AI 的复利收益也只会集中在少数早期采用者手中。

Source / 原文：https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL

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