七月二十九日

OpenAI reset usage limits for ChatGPT Work and Codex users after investigating why GPT-5.6 Sol consumed some users' allowances faster than expected. Sottiaux says typical Sol usage should now last about 18% longer, with larger gains for some power users, after improvements to long-running tool workflows, web-search waits, and programmatic tool calling. The team says Sol's willingness to keep working and its higher effective effort created a long tail of unexpectedly expensive sessions, even though median usage remained efficient.

OpenAI 在调查 GPT-5.6 Sol 为什么比预期更快消耗部分用户额度后，重置了 ChatGPT Work 和 Codex 用户的 usage limits。Sottiaux 表示，通过优化长时间运行的工具工作流、web search 等待和 programmatic tool calling，典型 Sol 使用时长预计增加约 18%，部分重度用户提升更明显。团队发现，Sol 更愿意持续工作、调用更多工具，而且同一 reasoning effort 下实际投入更高，因此少数复杂 session 的消耗远超预期，尽管中位数用户仍然相当高效。

Sottiaux also announced an open-source CLI and TypeScript SDK for finding, validating, and fixing code vulnerabilities. It can scan repositories, review changes, track findings over time, and run security checks in CI.

Sottiaux 还宣布开源一套用于发现、验证和修复代码漏洞的 CLI 与 TypeScript SDK。它可以扫描仓库、审查变更、持续跟踪 findings，并在 CI 中运行安全检查。

Swyx sees a sharp hiring split: a bull market for AI-native individual contributors and player-coaches, but a bear market for traditional "head of X" managers. His deliberately simplified comparison is that one year managing ten agents may now be more valuable than ten years managing teams of 10 to 100 people.

Swyx 观察到招聘市场正在剧烈分化：AI-native IC 和 player-coach 进入牛市，传统的 "head of X" 管理者则进入熊市。他刻意简化后的判断是，如今管理 10 个 agent 一年的经验，可能比管理 10 到 100 人团队十年的经验更值钱。

Dan Shipper says he produced a definitive history of Codex entirely through ChatGPT Work voice mode, including organizing hours of team interviews, building a timeline, writing, editing, and revising without touching a keyboard or mouse. He describes voice-driven agent workflows as especially strong for creative work and plans to publish the piece through Every.

Dan Shipper 表示，他完全通过 ChatGPT Work 的 voice mode 制作了一篇 Codex 完整发展史，包括整理数小时团队访谈、构建时间线、写作、编辑和修订，全程没有碰键盘或鼠标。他认为语音驱动的 agent 工作流特别适合创意工作，并计划通过 Every 发布这篇文章。

Amjad Masad proposed a SETI@Home-style network that would let people donate spare compute to search for mathematical proofs instead of extraterrestrial signals. The idea points toward distributed public compute as infrastructure for machine-assisted mathematics.

Amjad Masad 提议建立一个类似 SETI@Home 的网络，让人们捐出闲置算力来搜索数学证明，而不是寻找外星信号。这个构想把分布式公共算力变成机器辅助数学研究的基础设施。

Peter Yang shared the workflow behind Tastemaker, a product he built from a rough idea using Claude Design and Claude Code. His process moved from a `design.md` and HTML specification to screen prototypes in Claude Design, then implementation and iteration in Claude Code.

AI 产品教育者 Peter Yang 分享了 Tastemaker 的构建流程，他使用 Claude Design 和 Claude Code 把一个粗略想法做成产品。整个过程从 `design.md` 和 HTML spec 开始，在 Claude Design 中制作 screen prototype，再进入 Claude Code 完成实现和迭代。

Aditya Agarwal reflected on two principles that guided him through an uncertain period a decade ago: build new things in the world, and work with brilliant, intense people. He says both now feel unusually abundant, a concise signal of optimism about the current builder environment.

创业者 Aditya Agarwal 回顾了十年前穿越不确定时期时形成的两条原则：在世界上创造新东西，并与聪明且投入的人一起工作。他认为这两点如今都前所未有地充沛，传递出对当前 builder 环境的乐观判断。