I say how it feels to be a little retarded today with GPT 5. 5.

English | Chinese Original

English translation of the Chinese original. This version is generated for international readers and may be refined over time.

Date: 2026-05-17

I say how it feels to be a little retarded today with GPT-5. 5.

The reason is here.

The developer of GPT, Tibo, sent out an X today saying, "We found and repaired two problems that can explain the deterioration of GPT-5. 5 in the last 48 hours in Codex."

!!!!!!

Why do I think GPT-5/5 has a certain deterioration in capacity?

The reason is that I've been doing something in the Goal mode for two days.

But I found this stuff, it's token black hole, no matter how many tokens you have, it's insignificant in Goal mode.

And it's hard to control what's made in the goal mode.

Two days ago, especially yesterday and today, I've had a lot of trouble making things in the Goal mode.

It turns out Tibo said that there were two Bugs, and it is hoped that when these Bugs are finished, GPT-5. 5 will regain its intellectual level.

It's the same thing you told Codex that you made me an xx thing out of it, and I'm only looking at the final result.

This end result is often unsatisfactory.

It's basically like opening a blind box.

I think we should do it.

Step by step.

AI is first forced to produce a "minimum feasible plan" and is not allowed to write codes.

Under this step, you can output as much as possible: your understanding of needs, technical options and rationale, project module dismantling (tree structure, with one sentence for each module), core data stream / interface design (line description is sufficient), list of documents expected to be generated

Following confirmation of the plan, do not allow AI to start full construction and proceed to the second stage.

• For every module completed, stop and display the code and brief description immediately.
• We have to wait for my confirmation to get into the next module.
• If I say "problem," you have to fix the current module before proceeding.

In the event of an output that is "hard to say" an AI self-diagnostication using solid questions.

For example, check your output and answer:

1. Need conformity: What are the points where you actually met the original demand? What's missing?
2. Technical soundness: Are there excessive design, safety hazards or performance hazards at present?
3. Maintainability: What is the proportion that can be understood in three minutes if other developers take over?
4. Improvement programme: use of 3 amendments to bring it up to production availability standards.

In other words, the clearer you give the step-by-step standard, the more amazing it gets.

You leave it to its own full power, and it will be free to play at certain points. It's not only token's bills are flaming, but it's hard to say.