八月二十五日

Thibault Sottiaux said OpenAI will bring back the five-hour limit for Plus accounts across ChatGPT Work and Codex on August 26. He said the window helps smooth compute load, keeps weekly usage generous, and reduces cases where casual or new users accidentally burn through a full week of allowance; Pro $100 and Pro $200 plans will remain exempt for the next few months.

Thibault Sottiaux 表示，OpenAI 将在 8 月 26 日恢复 Plus 账户在 ChatGPT Work 和 Codex 上的五小时限制。他解释说，这个时间窗口可以平滑 compute 负载、维持相对宽松的每周使用量，并减少轻度或新用户不小心耗尽整周额度的情况；未来几个月，Pro $100 和 Pro $200 方案仍将继续豁免。

Aaron Levie wrote that systems of record become more important when AI agents do far more work on those platforms than people did. In his framing, the critical layer is not just data access but the governance, reliability, security, access controls, business logic, APIs, and product experiences that let agents execute safely on and off the platform.

Aaron Levie 认为，当 AI agents 在这些平台上完成的工作量远超人类时，systems of record 的价值只会更高。在他的框架里，关键不只是数据访问，而是 governance、reliability、security、access controls、business logic、APIs，以及能让 agents 在平台内外安全执行任务的产品体验。

He also argued that Zero Data Retention has been a major driver of AI adoption because it simplifies enterprise compliance. Applied AI vendors can standardize on approved models, while many enterprises only allow ZDR models because they cannot reliably separate sensitive data before it enters the context window.

他还认为，Zero Data Retention 一直是推动 AI adoption 的重要因素，因为它显著简化了企业合规流程。Applied AI 厂商可以围绕已批准的模型做标准化，而许多企业之所以只允许 ZDR models，是因为它们无法在数据进入 context window 之前稳定地把敏感信息分离出来。

Amjad Masad said Replit Agent has replaced Claude CoWork in his daily work because it is more persistent, more fastidious, and better at using code and software to complete tasks. Peter Steinberger made the adjacent product argument: the industry should move away from software that users cannot change with a prompt.

Amjad Masad 表示，Replit Agent 已经在他的日常工作中取代了 Claude CoWork，因为它更持久、更细致，也更擅长借助 code 和 software 完成任务。Peter Steinberger 则提出相邻但更广的产品判断：行业应该摆脱那些用户无法通过 prompt 改变的软件。

Madhu Guru argued that hill-climbing evals need discriminatory power: if systems that are known to differ substantially still score almost the same, the eval is not useful. He described the target as an eval that is realistic, difficult, and sensitive to capability differences, while noting that strong evals eventually saturate and need to evolve as models and harnesses improve.

Madhu Guru 认为，适合 hill-climbing 的 eval 必须具备 discriminatory power：如果已知差异明显的系统最后得分几乎一样，这个 eval 就没有用。他把目标定义为同时具备 realistic、difficult，以及对 capability 差异敏感的评测，并强调随着 models 和 harnesses 变强，优秀的 eval 最终会饱和，因此必须继续演进。

Guillermo Rauch traced a slow terminal `reset` to a `sleep(1)` in 3BSD's 1979 `tset`, originally added so mechanical terminals could settle down. He then had `fx` write a Zig replacement that finishes in 1 millisecond instead of 1 second, turning a tiny latency annoyance into a concrete example of AI-assisted systems cleanup.

Guillermo Rauch 把 terminal `reset` 的缓慢问题追溯到 1979 年 3BSD `tset` 里的 `sleep(1)`，这段逻辑最初是为了让机械终端稳定下来。随后他让 `fx` 用 Zig 写了一个替代实现，把执行时间从 1 秒降到 1 毫秒，也把一个细小的延迟问题变成了 AI-assisted systems cleanup 的具体例子。

Peter Yang said he often stops using products that force another website or app login, or agents that are trapped inside a website or app. In a separate post, he asked for the easiest way to use voice AI to navigate customer-support phone trees, reach a human, and complete real tasks such as booking appointments or canceling subscriptions.

Peter Yang 表示，当产品要求再次登录另一个网站或 app，或者 agent 被困在某个网站或 app 内时，他往往会直接放弃使用。在另一条帖子里，他还询问，怎样才能最简单地用 voice AI 穿过客服电话的自动语音系统、接通真人，并完成预约或取消订阅这类真实任务。