AI Builders Digest — 2026-08-24

X / TWITTER

Thibault Sottiaux, Codex & ChatGPT at OpenAI

OpenAI's Thibault Sottiaux traced Codex rate-limit pressure to image handling across long, repeatedly compacted sessions, high-p95 Computer History usage, and an unexpectedly expensive conversation-title feature. A tiger team is shipping fixes, all paid subscriptions will receive a usage reset, and the team has identified a separate efficiency approach for further work next week.

OpenAI 的 Thibault Sottiaux 将 Codex 的 rate limit 压力定位到三处：长会话多次 compaction 后的图片处理、Computer History 的高 p95 用量，以及消耗超出预期的会话标题功能。专项团队正在发布修复，所有付费订阅的 usage 将被重置；团队还发现了一条独立的效率提升路径，将于下周继续推进。

Source / 原文：https://x.com/thsottiaux/status/2091407991736332689

Peter Yang, AI educator and builder

Peter Yang showed a practical privacy workflow: open Google's third-party connection page in Chrome, let Codex or Claude Code inspect the open tab, then choose obsolete integrations for the agent to disconnect. He also highlighted Instinct's newly shipped external-data deletion control and is designing an AI skill to help cancer patients and families navigate care while staying informed.

AI 教程作者 Peter Yang 展示了一套实用隐私工作流：在 Chrome 打开 Google 第三方连接页面，让 Codex 或 Claude Code 检查当前标签页，再选择并断开不再需要的集成。他还关注到 Instinct 新增了外部数据删除入口，并正设计一个帮助癌症患者及家属理解治疗流程、持续获取信息的 AI skill。

Sources / 原文：
- https://x.com/petergyang/status/2091331251211059468
- https://x.com/petergyang/status/2091187611507499321
- https://x.com/petergyang/status/2091239339204415969

Madhu Guru, Senior Director of AI at Meta

Madhu Guru frames eval improvement as hill climbing: choose one dimension that matters, use a failure-mode taxonomy to locate the bottleneck, and iterate on harnesses, model choice, prompt/context engineering, memory, post-training, or deterministic code. For tool-calling failures, reduce a 20-tool context to the 3–5 tools needed at each stage; for cost, first prove demand with the strongest model, then seek equivalent quality with a smaller, faster, cheaper one.

Meta AI 高级总监 Madhu Guru 将 eval 优化概括为“爬山”：选定一个重要指标，用 failure-mode taxonomy 找到瓶颈，再围绕 harness、模型选择、prompt/context engineering、memory、post-training 或确定性代码迭代。若主要问题是 tool calling，就把一次塞入 20 个工具改为按阶段只提供 3–5 个；若目标是降本，则先用最强模型验证用户价值，再迁移到更小、更快、更便宜的模型，同时用 eval 保证质量。

Source / 原文：https://x.com/realmadhuguru/status/2091278653435072523

Amjad Masad, Replit CEO

Replit CEO Amjad Masad compressed his operating cadence into one line: seven days should mean seven ships. The signal is less about literal daily releases than treating shipping frequency as a core product discipline.

Replit CEO Amjad Masad 用一句话概括了产品节奏：一周七天，就应有七次 shipping。重点未必是机械地每日发布，而是把高频交付视为产品团队的核心纪律。

Source / 原文：https://x.com/amasad/status/2091346778746757204

Aaron Levie, Box CEO

Box CEO Aaron Levie argues that enterprise AI adoption is constrained less by general model capability than by workflow-specific evals. Public benchmarks show the broad shape of progress, but companies need assessments tailored to their own high-value processes because “you can’t automate what you can’t assess”; enterprise automation cannot run on vibes.

Box CEO Aaron Levie 认为，企业 AI 普及的主要限制并非通用模型能力，而是缺少针对具体 workflow 的 eval。公开 benchmark 只能说明整体进展，企业还需要贴合自身高价值流程的评测，因为“无法评估，就无法自动化”，企业自动化不能靠感觉推进。

Source / 原文：https://x.com/levie/status/2091359223368315050

Zara Zhang, builder

Zara Zhang sees an organizational AI leverage gap: talented people working independently on their own products can reach 10× their prior output, while the same people inside large organizations may gain only 20%, or even regress. Her explanation for talent leaving incumbents is therefore structural: AI amplifies autonomy faster than it improves organizational coordination, with top AI labs as possible exceptions.

Builder Zara Zhang 观察到一种组织层面的 AI 杠杆差：优秀人才独立做自己的产品时，产出可能提升到 10 倍；进入大型组织后，同一个人往往只提升 20%，甚至倒退。因此，人才离开大公司的原因可能是结构性的：AI 放大个人自主性的速度，远快于它改善组织协同的速度，顶级 AI lab 或许是例外。

Source / 原文：https://x.com/zarazhangrui/status/2091379220257603593

Nikunj Kothari, FPV Ventures partner

FPV Ventures partner Nikunj Kothari warned founders that rage-baiting investors and then sending unsolicited SAFE documents is not fundraising. His broader advice is to be selective about whose guidance you follow: attention tactics can destroy trust even when the underlying founder has real substance.

FPV Ventures 合伙人 Nikunj Kothari 提醒创始人：先用挑衅内容激怒投资人，再直接发 SAFE 要求打款，并不是真正的融资。他更广泛的建议是谨慎选择听谁的意见：即便创始人本身有实力，追逐注意力的套路也会迅速摧毁信任。

Source / 原文：https://x.com/nikunj/status/2091381756012511244

PODCASTS

AI & I by Every: Microsoft’s Vision for an Internet Made for Agents With CTO Kevin Scott (Best of the Pod)

The Takeaway: Microsoft CTO Kevin Scott believes the model layer has developed a “capability overhang”; the urgent work is building memory, access, protocols, identity, and permissions that let agents turn reasoning into reliable action.

Scott argues that useful agents must operate across diverse tools, data sources, websites, and APIs. That requires an agentic web built from simple, open, composable standards. He casts MCP as the rough equivalent of HTTP for agent-to-system connections and NLWeb as an HTML-like layer, while pushing Microsoft’s internal systems to expose a common protocol so its agents do not merely “ship the org chart.”

Security is central but, in Scott’s view, not an argument for a closed ecosystem. Agents need identities and entitlement systems that can explain which resources a task requires, request the necessary permissions from users, and let administrators enforce policy. Openness preserves permissionless innovation, while AI itself may help continuously assess context and personal risk.

For software makers, agents are another major tool shift rather than the death of craft. Scott’s practical stance is: “Be curious. Try. And if it works for you, use it. And if it doesn’t, don’t.” The enduring skill is judgment about tools, materials, and outcomes, not loyalty to a particular way of typing code.

核心结论： Microsoft CTO Kevin Scott 认为，模型层已经出现“能力盈余”；当前最紧迫的工作，是补齐 memory、访问能力、协议、身份与权限体系，让 agent 把推理能力转化为可靠行动。

Scott 指出，真正有用的 agent 必须跨越多种工具、数据源、网站与 API 执行任务，这需要由简单、开放、可组合标准构成的 agentic web。他把 MCP 类比为 agent 连接系统时的 HTTP，把 NLWeb 看作近似 HTML 的一层；在 Microsoft 内部，他推动所有系统使用统一协议，避免 agent 只是“把组织架构直接做成产品”。

安全是核心问题，但在 Scott 看来，这并不意味着生态必须封闭。Agent 应拥有身份和 entitlement 系统，能够说明任务需要访问哪些资源、向用户申请必要权限，并让管理员执行策略。开放协议保留了无需许可的创新空间，而 AI 本身也可能持续评估上下文与个人风险。

对软件创造者而言，agent 是又一次重大工具迁移，而不是 craft 的终结。Scott 的实用态度是：“保持好奇，试一试；有效就用，无效就不用。”真正长期有效的能力，是对工具、材料和结果的判断，而不是执着于某一种敲代码方式。

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

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