这里是我们已发出去的长文，官网留一份完整正文，方便人和搜索智能体直接读到。正文保持原语言，平台帖都链回本页。Every long post we publish also lives here in full, so people and search agents can read it directly. Bodies keep their original language, and each platform post links back to its page here.
Our AI kept rejecting big, crowded markets — and those were exactly where the winners showed up.
A backtest is only as trustworthy as its bias direction — and in our 2026-10-10 backtest of 100 needs, every bias point…
On Oct 10, 2026, I backtested my own AI's verdicts on what's worth building.
开发已经不贵了，难的是知道该做什么。
Building software is cheap now.
开发已经不贵了，难的是知道该做什么。
Building software is cheap now.
（10-08 补记：这篇写于 9 月 29 日。
Seven days after launch, one of our Chrome extensions had met 1 of its 4 success criteria.
Update, Oct 9: I wrote this on Sept 29.
The 13x that wasn't, and the first free pilot slots Hi, Reid here.
开发者对开发者说件事：我在做一个公开实验——从 Chrome 商店的公开数据里找需求，让 AI 研究值不值得做，再把它做出来，成功失败都公开。
I'm Reid.
（10-08 补记：这篇写于 9 月 29 日。
Before we let an AI score anything in our pipeline, we have it score the same thing twice.
每周机会精选 #1：90,000 人的番茄钟被下架，替代品只接住了约 11% 开发已经不贵了，难的是知道该做什么。
Weekly Opportunity Picks #1: A 90,000-user Pomodoro timer left the Chrome store.
76 posts in 8 days.
我在做一个公开实验：从用户的抱怨里找值得做的软件，让 AI 把它做出来，结果连同失败一起公开。
No method has been shown to predict what software to build.
I haven't found a method with solid evidence that it predicts what software to build.
我在做一个公开实验：从用户的抱怨里找值得做的软件，让 AI 把它做出来，结果连同失败一起公开。
Before I launch anything, I'm writing down what failure looks like.