AI Builders Digest — 2026-08-26

2026-08-26

AI Builders Digest — 2026-08-26

X / TWITTER

Thibault Sottiaux, Codex & ChatGPT at OpenAI

OpenAI's Thibault Sottiaux says DevDay 2026 will be the company's strongest yet. More immediately, OpenAI will restore a five-hour usage window for Plus accounts across ChatGPT Work and Codex to smooth compute demand and reduce accidental exhaustion of weekly allowances; the limit will remain disabled for Pro $100 and Pro $200 plans for the coming months.

OpenAI 的 Thibault Sottiaux 表示，DevDay 2026 将成为公司史上最强的一届。更直接的变化是，OpenAI 将为 Plus 账户恢复 ChatGPT Work 与 Codex 的五小时使用窗口，以平滑算力负载，并减少用户意外耗尽每周额度的情况；未来几个月，Pro $100 和 Pro $200 套餐仍不启用该限制。

Madhu Guru, Senior Director of AI at Meta

Meta AI Senior Director Madhu Guru argues that a useful hill-climbing eval must discriminate between meaningfully different systems. The target is not arbitrary difficulty, but a realistic, difficult test that remains sensitive to capability differences; once a good eval saturates, the benchmark itself must evolve.

https://x.com/realmadhuguru/status/2092058332735693264

Meta AI Senior Director Madhu Guru 认为，有效的 hill-climbing eval 必须能区分能力确实不同的系统。目标不是盲目提高难度，而是构建真实、困难且对能力差异敏感的测试；当优秀 eval 被模型能力追平后，benchmark 本身也必须继续演进。

https://x.com/realmadhuguru/status/2092058332735693264

Aaron Levie, CEO of Box

Box CEO Aaron Levie says systems of record become more important, not less, when agents perform 100 times more work than humans on enterprise platforms. Governance, reliability, security, permissions, business logic, agent-friendly APIs, and viable business models become core competitive infrastructure. He also argues that zero data retention has materially accelerated enterprise AI adoption by simplifying subprocessor compliance and internal governance; without ZDR, diffusion would slow sharply.

Box CEO Aaron Levie 认为，当 agent 在企业平台上完成的人类工作量达到 100 倍时，system of record 不会被削弱，反而更加重要。治理、可靠性、安全、权限、业务逻辑、agent-friendly API 与可持续商业模式，将成为核心竞争基础设施。他还指出，zero data retention 通过简化 subprocessors 合规与企业内部治理，显著加速了 AI 的企业采用；没有 ZDR，AI 扩散速度会明显下降。

Amjad Masad, CEO of Replit

Replit CEO Amjad Masad says Replit Agent has replaced Claude CoWork in his daily workflow because it is more persistent, more meticulous, and better at using code and software to finish tasks. The signal is less about a single benchmark and more about agent reliability over long, tool-heavy execution loops.

https://x.com/amasad/status/2091962601907638352

Replit CEO Amjad Masad 表示，Replit Agent 已在他的日常工作中取代 Claude CoWork，原因是它更具持续性、更细致，也更善于利用代码和软件完成任务。这个信号的重点不在单一 benchmark，而在 agent 面对长链路、重工具任务时的执行可靠性。

https://x.com/amasad/status/2091962601907638352

Guillermo Rauch, CEO of Vercel

Vercel CEO Guillermo Rauch highlights software that becomes faster, cheaper, more capable, and smaller instead of accumulating bloat. In a second example, he used fx to build a Zig replacement for the Unix terminal reset path, cutting a one-second legacy delay inherited from mechanical-terminal assumptions to roughly one millisecond.

Vercel CEO Guillermo Rauch 强调，软件应当在演进中变得更快、更便宜、更强且更小，而不是持续膨胀。另一个案例中，他借助 fx 用 Zig 重写 Unix terminal reset 路径，将源自机械终端时代假设的一秒遗留延迟，降至约一毫秒。

Peter Steinberger, OpenClaw and OpenAI

OpenClaw's Peter Steinberger makes a sharp product claim: software that cannot be changed with a prompt is becoming obsolete. The implication for builders is that prompt-level adaptability may become a basic product property, not merely an AI feature layered onto fixed software.

https://x.com/steipete/status/2091923535513928015

OpenClaw 的 Peter Steinberger 提出一个鲜明判断：不能通过 prompt 改变的软件正在过时。对 builder 而言，这意味着 prompt 级可塑性可能会成为产品的基础属性，而不只是叠加在固定软件之上的 AI 功能。

https://x.com/steipete/status/2091923535513928015

Garry Tan, President and CEO of Y Combinator

Y Combinator CEO Garry Tan says Conductor Cloud has improved his productivity by removing the need to keep a MacBook running. His broader builder loop is concise: form a view, turn it into an artifact or experiment, expose it to reality, read the result honestly, then revise and repeat.

Y Combinator CEO Garry Tan 表示，Conductor Cloud 提升了他的生产力，让他不必持续开着 MacBook。他给出的 builder 闭环也非常凝练：形成判断，把它变成作品或实验，让它接触现实，诚实读取结果，然后修正并再次运行。

Nikunj Kothari, Partner at FPV Ventures

FPV Ventures partner Nikunj Kothari says unconventional venture deals often die many times before funds are wired. Founders should identify the internal champion who will steer fund dynamics, then equip that person with the information needed to make the case; associates and non-voting partners can be especially useful allies.

https://x.com/nikunj/status/2092079149028716877

FPV Ventures partner Nikunj Kothari 表示，非传统 venture deal 在资金正式到账前，往往会经历多次“死亡”。Founder 应识别真正能推动基金内部决策的 champion，并向其提供充分信息；associate 和无投票权 partner 往往是尤其有价值的内部盟友。

https://x.com/nikunj/status/2092079149028716877

Peter Yang, AI Educator and Product Thinker

Peter Yang points to two product gaps for agents: users increasingly resist products that require another login or trap agents inside a website, while voice agents still lack an obvious, low-friction way to navigate automated phone trees, reach humans, book appointments, and cancel subscriptions. Both observations favor agents that cross application boundaries and complete outcomes without adding interface burden.

AI 教育者与产品观察者 Peter Yang 指出两个 agent 产品缺口：用户越来越抗拒要求再次登录、或把 agent 困在网站内部的产品；与此同时，voice agent 仍缺少一种足够简单的方式，去穿越自动电话菜单、接通人工客服、预约服务或取消订阅。这两个观察都指向同一方向：agent 应跨越应用边界完成结果，而不是增加新的界面负担。

Nan Yu, Product Enthusiast

Nan Yu shows a practical glimpse of natural-language computer setup: asking Codex to download and install Handy, Slack, Chrome, CleanShot, and Rectangle on a new machine. The experience makes desktop automation feel less like scripting and more like delegating intent.

https://x.com/thenanyu/status/2092048044502192374

产品实践者 Nan Yu 展示了自然语言配置电脑的一个实用场景：直接让 Codex 在新机器上下载并安装 Handy、Slack、Chrome、CleanShot 和 Rectangle。这种体验让 desktop automation 从“写脚本”进一步转向“委派意图”。

https://x.com/thenanyu/status/2092048044502192374

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