The journal
Notes on building software, trying new tools, and figuring things out along the way.
Using Claude Fable 5 convinced me that AI has crossed a capability threshold. Falling costs, and AI helping develop its own infrastructure, will shape how widely those capabilities spread.
My experience with ChatGPT, Codex, and Claude, and why exploring unclear needs, staying focused, and correcting course matter more to me than execution alone.
My account was suspended during identity verification after upgrading to Claude Max. The experience raised a larger question about accountability as AI becomes more important to our ability to work.
Trying to subscribe to Claude Max and ChatGPT Pro, I hit more payment risk reviews in one month than in the whole previous year. The reviews were tiring, but the real problem was that customer support could not help at all.
GPT-6 is fast and reliable on well-scoped tasks, but a big task I had not fully thought through ran for four hours and came back a mess. It never asked questions and never noticed it was off track.
Codex has plenty of features and strong Computer Use. But it agrees with whatever the user asks, rarely checks its own work along the way, and hides what it's actually doing.
In September I built two small Mac apps, and only then realized that turning an idea into an app is no longer hard. The hard part is getting people to know it exists and want to try it. For an indie developer, marketing is much harder than coding.
iOS 27 drops the tilt-responsive lighting that made Liquid Glass feel like glass. Buttons are easier to see, but the result looks more like plastic. Plus some thoughts on the faster startup and a Mac black screen.
Codex Desktop has more features, better computer use, and a stronger frontend model in Astra 6, yet its interface and input box make it feel heavy and slow. Some thoughts on working with both apps.
My experience trying Google Play and Apple gift cards with China Merchants Bank cards, and the payment reviews that made the process harder than expected.
试用 Claude Fable 5 后，我认为 AI 的能力已经跨过一道门槛。接下来，成本下降以及 AI 对基础设施研发的参与，将决定这种能力如何普及。
从 ChatGPT、Codex 和 Claude 的使用体验出发，谈我为什么更看重 AI 探索模糊需求、保持目标和自我纠错的能力。
升级 Claude Max 后，在身份验证过程中收到停权通知。这次经历让我思考：当 AI 越来越影响工作能力时，提供它的公司应该受到怎样的约束。
为了订阅 Claude Max 和 ChatGPT Pro，我最近一个月碰到的风控比过去一年加起来还多。真正让人累的不是风控本身，而是客服帮不上任何忙。
GPT-6 执行边界清楚的任务又快又好，但一个没想清楚的大任务，它跑了四个小时，最后交出一团糟。问题出在它不会追问，也不会发现自己跑偏。
9 月我做了两款 Mac 小工具，做完才发现，把想法做出来已经不难，难的是让别人知道它、愿意试一试。对独立开发者来说，营销比写代码难多了。
iOS 27 去掉了随手机倾斜变化的玻璃光影，按钮是更清楚了，但看起来更像塑料。顺便说说变快的启动，和今天遇到的 Mac 黑屏。
Codex Desktop 功能多很多，computer use 也好用很多，但用起来就是没有 Claude Desktop 那么舒服。
从 Google Play 到苹果礼品卡，记录我使用招行 Visa 和 MasterCard 尝试订阅 AI 服务的经历，以及支付风控带来的等待。