AI Builders Digest | 2026-09-29
X / TWITTER
Thibault Sottiaux, Codex & ChatGPT at OpenAI
Sottiaux argues that the traditional pre-release code freeze is becoming obsolete. His more radical prediction is that software may eventually be generated online for each request, subject to explicit constraints, shifting release engineering from shipping fixed artifacts toward governing dynamic generation.
Sottiaux 认为，传统的发布前 code freeze 正在失去意义。更激进的判断是，未来软件可能根据每次请求，在明确约束下在线生成。届时，发布工程的核心将从交付固定制品，转向治理动态生成过程。
Source / 原文：https://x.com/thsottiaux/status/2104108167806550046
Thariq, Claude Code at Anthropic
Thariq's concern is not that agents fail to deliver productivity gains, but that people absorb those gains by lowering their own effort. Better tools do not automatically create better output; teams need operating norms that convert saved time into more ambitious work.
Thariq 担心的不是 agent 无法带来生产力提升，而是人们会用更低的投入把这些增益消耗掉。工具变强并不会自动产生更好的结果，团队仍需建立新的工作规范，把节省下来的时间转化为更有雄心的产出。
Source / 原文：https://x.com/trq212/status/2104273243599405395
Vercel CEO Guillermo Rauch
Guillermo Rauch ported his Mini browser from Electron and Bun to Rust and Swift, using the cef crate for Chromium and ACP plus MCP to connect an embedded agent to his local fx CLI. He reports faster startup, tighter security, Liquid Glass, and more native macOS behavior, and predicts software will “nativify” faster than expected across desktop and cloud environments.
Guillermo Rauch 将 Mini 浏览器从 Electron 和 Bun 迁移到 Rust 与 Swift，使用 cef crate 集成 Chromium，并通过 ACP 和 MCP 将内置 agent 连接到本地 fx CLI。他获得了更快的启动速度、更好的安全性、Liquid Glass 和更原生的 macOS 体验，并预测桌面与云端软件的“原生化”速度将超出预期。
Source / 原文：https://x.com/rauchg/status/2104428800134013205
Box CEO Aaron Levie
Aaron Levie expects agents that optimize choices for users to compress switching costs and intensify competition in markets that profit from customer inertia. At the same time, agents could expand healthcare, travel, local services, and information markets by removing transaction friction that currently suppresses demand. An agent-mediated economy will therefore redistribute value rather than simply automate today's market structure.
Aaron Levie 认为，当 agent 开始替用户优化选择时，依赖客户惯性的行业将面临更低的转换成本和更激烈的竞争。与此同时，agent 也会消除交易摩擦，释放医疗、旅行、本地服务和信息服务中被压抑的需求。因此，agent 驱动的经济不仅会自动化现有市场，还会重新分配价值。
Source / 原文：https://x.com/levie/status/2104350592290406849
FirstMark VC Matt Turck
Matt Turck highlights a mismatch in the AI debate: many researchers who understand the systems firsthand reject both certain doom and uncontrollable acceleration, while outsiders often hold the most categorical views. His point is less a forecast than a warning that confidence and domain knowledge are diverging.
FirstMark VC Matt Turck 指出，AI 讨论中存在明显错位：许多真正理解 AI 工作机制的研究者，既不相信必然毁灭，也不认同不可控的极速加速；反而是专业知识有限的局外人最容易持有绝对化立场。这更像是在提醒我们，观点的自信程度与领域知识正在背离。
Source / 原文：https://x.com/mattturck/status/2104331402385002831
Builder Zara Zhang
Zara Zhang makes two related observations about building in public. Self-expression does not need a monetization rationale, and durable influence can be more valuable than immediate revenue. She also warns that dazzling technology can make users blame themselves when products fail, a useful corrective for builders evaluating whether an AI workflow is genuinely effective.
Builder Zara Zhang 提出两点相关判断：公开表达不需要以商业变现为理由，长期影响力也可能比即时收入更有价值。她同时提醒，耀眼的新技术容易让用户在产品失效时反过来责怪自己。对 AI builder 而言，这是一项重要校准：应判断工作流是否真的有效，而不是让技术光环掩盖问题。
Sources / 原文：
- https://x.com/zarazhangrui/status/2104253882025341231
- https://x.com/zarazhangrui/status/2104112917264126195
FPV Ventures Partner Nikunj Kothari
Nikunj Kothari says Astra performs exceptionally well when given difficult, verifiable, end-to-end tasks together with the right tools. The emphasis on verifiability is important: agent performance becomes easier to trust when success can be checked objectively rather than judged from plausible prose.
FPV Ventures Partner Nikunj Kothari 表示，Astra 在获得合适工具，并处理困难、可验证的端到端任务时表现突出。这里真正关键的是“可验证”：当成功标准可以客观检查，而不是依赖看似合理的文字时，agent 的能力才更容易被信任。
Source / 原文：https://x.com/nikunj/status/2104444216017637575
OpenClaw & OpenAI Builder Peter Steinberger
Peter Steinberger plans to let Codex decide which tests actually need to run, sharply reducing continuous CI load and moving broader test runs to an hourly cadence. This treats the coding agent not only as a code producer, but as an adaptive controller for engineering infrastructure and cost.
OpenClaw 与 OpenAI builder Peter Steinberger 计划让 Codex 判断哪些测试真正需要运行，从而大幅降低持续 CI 负载，并把更完整的测试改为每小时运行。这意味着 coding agent 不只是代码生产者，也开始成为工程基础设施与成本的自适应控制器。
Source / 原文：https://x.com/steipete/status/2104305554760114488
Every CEO Dan Shipper
Dan Shipper captures the economic irony of agents in one line: industries built on modeling humans as rational actors are alarmed when humans begin adopting actual rational agents. The implication is that advertising, commerce, and other persuasion-heavy markets may face structural pressure as software increasingly represents user interests.
Every CEO Dan Shipper 用一句话概括了 agent 的经济学悖论：那些长期把人类建模为理性行动者的行业，在人类真正开始使用理性 agent 时反而陷入恐慌。其潜台词是，当软件越来越能够代表用户利益时，广告、电商等高度依赖说服与摩擦的市场将面临结构性压力。
Source / 原文：https://x.com/danshipper/status/2104302924251951553
PODCASTS
Training Data: Box's Aaron Levie: On Reinventing Yourself in the AI Age and Enterprise Diffusion
The Takeaway: The trillion-dollar AI opportunity is not confined to frontier models; it lies in the operational bridge between intelligence and the messy workflows, permissions, data, and change management of real enterprises.
Box founder and CEO Aaron Levie argues that application companies are becoming “neo-labs” because domain data and workflow depth can create their own model and product flywheels. Model capability alone is insufficient: production systems must connect to legacy data, respect access controls, pause for humans, survive organizational delays, and fit how lawyers, scientists, salespeople, and government workers actually operate. That favors model-neutral application vendors capable of selecting the best model for each task while owning the customer context.
Coding agents diffused first because software work is unusually automation-friendly: inputs and outputs are digital, code is centralized in systems such as GitHub, results are frequently testable, engineers can fix integration failures, and productivity has high economic value. Most knowledge work lacks those conditions. Levie expects enterprise diffusion to take longer than Silicon Valley assumes, but believes the friction itself creates “a trillion dollars of applied layer AI value.”
For founders, faster AI-assisted building shifts advantage toward distribution and implementation. A product that once required forty people can now be built by two, but every good idea also attracts competitors almost instantly. Winning therefore requires reaching customers, mastering domain workflows, and helping organizations change, not merely producing software faster. Inside Box, the practical playbook is centralized, well-governed data; targeting high-leverage workflows; studying top internal AI users; and rapidly teaching their successful patterns across teams.
核心结论： AI 的万亿美元机会并不只属于 frontier model，更存在于智能能力与真实企业中混乱的工作流、权限、数据和变革管理之间的连接层。
Box 创始人兼 CEO Aaron Levie 认为，应用公司正在成为“neo-labs”，因为领域数据和工作流深度能够形成自己的模型与产品飞轮。仅有模型能力远远不够：生产系统必须连接遗留数据、遵守访问权限、允许人类介入、承受组织延迟，并适配律师、科研人员、销售和政府员工的真实工作方式。这有利于 model-neutral 的应用厂商，它们既能为每项任务选择最合适的模型，又掌握客户上下文。
Coding agent 最先普及，是因为软件开发天然适合自动化：输入输出均为数字形式，代码集中在 GitHub 等系统中，结果通常可测试，工程师能自行修复集成故障，而且效率提升具有很高经济价值。多数知识工作并不具备这些条件。Levie 预计企业 AI 扩散会比 Silicon Valley 想象得更慢，但正是这些摩擦创造了“应用层 AI 的万亿美元价值”。
对于创业者，AI 加速开发后，竞争优势会转向分发与落地。过去需要 40 人完成的产品，如今可能由 2 人做出，但好点子也会立即迎来多个竞争者。胜负因此取决于谁能触达客户、理解领域工作流并推动组织变革，而不仅是谁写软件更快。Box 的内部实践包括：集中且治理良好的数据、优先改造高杠杆工作流、识别最优秀的内部 AI 用户，并迅速把他们的成功用法推广给团队。
Source / 原始节目：https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders