[🤖] The Trash AI Made in 16 Hours Humbled Me

✨ GPT-5.6 Sol’s Summary 　

I opened Codex’s result expecting something solid after 16 hours of tests and audits, and it was absolute trash. That one screen started exposing every problem I had missed.

I Let It Run for 16 Hours, and the Result Was Absolute Trash

Yesterday, I spent a long time debating and writing detailed page-by-page specifications for a major UI/UX overhaul. Then I set a Goal telling Codex to build from those specifications.

The Goal ran for 16 hours and 44 minutes. I opened it today full of expectation, and the result was absolute trash.

I was humbled. Again.

When I think about how much I could achieve if I learned to run AI well, I get completely carried away. Then it forgets what I meant, does whatever it wants, and sprays out shit that only looks vaguely convincing. The crash afterward is brutal. This time I had even written a long specification and received a report saying it had passed 16 hours of tests and audits. What I wanted still was not there.

One Failure Exposed Everything Else I Had Missed

As I argued with AI about why this had happened, I learned that several problems I had been wrestling with while using Codex already had names and established methods. People keep running into the same issues, so of course the discussion was already active. I simply had not known because I had shut my ears and paid no attention.

I kept pressing on that one screen, and the problems I had missed started falling out one after another. Instead of forcing them into one post, I recorded each problem I actually retried and the result separately.

• When AI Got an Intermediate Result Wrong, I Stopped the Next Job
• I Tore Apart My Long-Winded AI News Briefing
• AI Kept Agreeing With Me, So I Made Another AI Argue First
• 30% Context Use From the Start of a Chat, Moving Shared Rules/Skills From Korean to English
• Codex Erased Me From My Own Tech Blog

That made me think someone must already have turned this into a service, and of course someone had. Find a recurring irritation, scratch it properly, and rake in the money. It sounds so easy when said out loud. Building that cycle and keeping it running is anything but easy.

I still believe that running AI well can produce enormous results. But running it longer, spending more tokens, and passing more tests were not evidence that it understood what I wanted. Before I accept another 16-hour pile of trash, I need to check at the start whether it actually understands the result I am asking for.