Top Signals
OpenAI identifies three sources of unexpected Codex usage
OpenAI's Thibault Sottiaux said the Codex team found three contributors to unexpectedly high usage: inefficiencies when images remain in long sessions through multiple compactions, high p95-plus usage for Computer History, and a conversation-title feature consuming more usage than intended. A dedicated team was combing through the system, with fixes planned for the next day and a separate efficiency approach scheduled for work the following week. Sottiaux also said all paid subscriptions would receive a full usage reset as part of the immediate fixes.
OpenAI 的 Thibault Sottiaux 表示，Codex 团队找到了三类意外高 usage 的来源：图片在经历多次 compaction 的长 session 中存在效率问题，Computer History 的 p95 以上 usage 偏高，以及一个用于生成 conversation title 的功能消耗超出预期。一个专项团队正在全面排查系统，相关修复计划于次日发布，另一个独立的效率提升方案则安排在接下来一周推进。Sottiaux 还表示，作为这批即时修复的一部分，所有付费订阅的 usage 都会获得一次完整 reset。
Enterprise AI is becoming an evals problem, not just a model problem
Box CEO Aaron Levie argues that public model benchmarks describe general progress and relative model capability, but cannot tell an enterprise whether its own workflows are ready for automation. The larger opportunity is company-specific evaluation across major workflows, because organizations cannot responsibly automate work they cannot measure. His point shifts the deployment bottleneck away from choosing the newest model and toward building evidence that an AI system improves the actual process it is meant to run.
Box CEO Aaron Levie 认为，公开 model benchmark 能描述总体进展和模型之间的相对能力，却无法告诉企业自己的 workflow 是否已经适合自动化。更大的机会在于针对主要业务流程建立 company-specific evaluation，因为企业无法负责任地自动化一项自己都不能衡量的工作。这个判断把部署瓶颈从选择最新模型，转向为 AI system 是否真正改善目标流程建立证据。
Instinct adds a deletion path after a connected-data privacy complaint
Peter Yang reported that Instinct now lets users delete external data from a Data Privacy section, including the 36 Gmail records associated with his case. He credited the team for responding and shipping quickly. The update turns the previous privacy failure into a concrete product lesson: assistants that ingest connected accounts need visible deletion controls, and rapid remediation can recover some trust only after users can directly control the retained data.
Peter Yang 表示，Instinct 现在允许用户在 Data Privacy 页面删除 external data，其中包括与他这次案例相关的 36 条 Gmail records。他肯定了团队听取反馈并快速上线修复。这个更新把此前的 privacy failure 变成了一条具体的产品经验：会读取 connected accounts 的 assistant 必须提供清晰可见的删除控制，而快速补救只有在用户能够直接管理留存数据之后，才有可能挽回一部分信任。
Zara Zhang sees AI leverage concentrating outside large organizations
Builder Zara Zhang says talented people working on their own projects can reach roughly ten times their prior potential with AI, while the same people inside large organizations may gain only about 20%, and can sometimes become less effective. She connects that gap to more talent leaving big companies, with top AI labs such as OpenAI and Anthropic as possible exceptions. This is an observation rather than measured research, but it identifies organizational coordination and process as potential limits on how much individual AI leverage becomes company output.
Builder Zara Zhang 认为，有能力的人在做自己的项目时，借助 AI 可能把潜力提升到原来的约十倍；同一个人进入大型组织后，提升可能只有约 20%，有时甚至会变得更低。她把这种差距与越来越多人才离开大公司联系起来，同时认为 OpenAI、Anthropic 等顶级 AI labs 可能是例外。这是一项个人观察，而不是经过测量的研究，但它指出了一个重要限制：组织协调与流程，可能会阻碍个人的 AI leverage 转化为公司的实际产出。
Engineering & Research
Madhu Guru turns eval improvement into a concrete hill-climbing loop
Meta Senior Director of AI Madhu Guru defines hill climbing on evals as choosing a dimension that matters and optimizing it through better harnesses, model selection, prompt and context engineering, memory, post-training, or deterministic code. A failure-mode taxonomy should determine where to work: if tool calling is weak because 20 tools are always placed in context while a task needs only three to five, the next experiment is to expose the right tools at the right stage. For cost reduction, his sequence is to launch first with the strongest model until users value the experience, then use evals to preserve quality while moving to a smaller, cheaper and faster model.
Meta AI Senior Director Madhu Guru 把 eval 上的 hill climbing 定义为：选定一个真正重要的维度，再通过更好的 harness、model selection、prompt 与 context engineering、memory、post-training 或 deterministic code 持续优化。failure-mode taxonomy 应该决定团队把精力放在哪里：如果 tool calling 表现差，是因为 context 中始终塞入 20 个 tools，而单个任务其实只需要 3 到 5 个，那么下一轮实验就应该在正确阶段只提供正确的 tools。对于 cost reduction，他建议先用最强模型发布，直到确认用户认可体验，再借助 evals 在迁移到更小、更便宜、更快的模型时维持质量。
Podcast
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 model reasoning has moved ahead of what products deliver, creating a capability overhang that the industry must close with memory, tools, permissions and open protocols. Agents remain too transactional when memory disappears after a task, and they cannot handle meaningful delegation unless they can act across diverse systems. Scott compares MCP to HTTP and NLWeb to HTML: simple, composable layers that could let existing websites and APIs participate in an agentic web. Inside Microsoft, he wants systems to speak a standard protocol to internal agents so products do not simply “ship your org chart.”
核心结论：Microsoft CTO Kevin Scott 认为，模型的 reasoning capability 已经领先于产品真正交付给用户的能力，形成了需要用 memory、tools、permissions 与 open protocols 来填补的 capability overhang。任务结束后 memory 消失，会让 agent 过于 transactional；如果不能跨多个系统采取行动，它们也无法承接真正有意义的 delegation。Scott 把 MCP 类比为 HTTP、把 NLWeb 类比为 HTML，认为这些简单且可组合的层，可以让现有 websites 与 APIs 接入 agentic web。在 Microsoft 内部，他希望所有系统都通过标准协议与 internal agents 沟通，避免产品只是简单地“ship your org chart”。
Security is the unresolved companion problem. Scott says agents need identities and entitlement systems that can explain which resources a delegated task requires, request permission from the user, and remain governable by administrators. He rejects a strict choice between open innovation and robust security, arguing that permissionless systems can coexist with agents that understand personal sharing preferences and assess risk across communication channels. He also expects many specialized agents rather than one universal winner, with differentiation coming from a nuanced understanding of user problems rather than unique infrastructure.
Security 是与之相伴、尚未解决的问题。Scott 认为，agent 需要 identity 与 entitlement systems，能够说明一个 delegated task 要访问哪些资源，向用户请求 permission，同时接受管理员治理。他不认同 open innovation 与 robust security 只能二选一，认为 permissionless system 可以与理解个人分享偏好、并能跨 communication channels 评估风险的 agent 共存。他也预计未来会出现许多 specialized agents，而不是一个包办一切的赢家；真正的差异化将来自对用户问题的细致理解，而不是独有的 infrastructure。
His advice to builders is not to wait for marginally lower prices or slightly better capability, because both are already moving in that direction. Over the next year, he expects ambition to rise and interaction to shift from synchronous prompting toward asynchronous delegation, where an agent can spend substantial time calling systems, integrating results and iterating before handing work back. On coding craft, his position is pragmatic: people should keep strong opinions and choose whether they value process or outcome, but remain curious as tools change.
他给 builders 的建议是，不要为了价格再低一点或能力再强一点而等待，因为这两个方向已经在持续改善。未来一年，他预计人们处理的问题会更有野心，交互方式也会从 synchronous prompting 转向 asynchronous delegation：agent 可以花较长时间调用多个系统、整合结果并反复迭代，最后再把工作交还给人。对于 coding craft，他的态度很务实：人们应该保留对技艺的强烈判断，也可以自行选择更看重过程还是结果，但在工具变化时要保持好奇。