Problems I’ve Encountered as a $200/Month Pro Subscriber
Since the last developer conference, I have encountered many problems with GPT that I either had not run into before or had previously overlooked. These problems have seriously damaged my experience. The service OpenAI is currently providing has pushed my anger to the boiling point. I have had enough.
Today is October 10, Beijing time (UTC+8). I arrived at my desk and sat down at my computer at 9:00 a.m. By 2:20 p.m., I was still furious with GPT. I had practically turned into a machine gun, firing off one “fucking hell” after another.
I admit that I have a bad temper, but I consider my self-control reasonably good. GPT’s dismal performance lately has left me unable to keep quiet.
1. Endless, Pointless Contrasts
First, I genuinely do not understand why OpenAI’s models are so fond of repeatedly emphasizing:
“What I’m doing is XXXXX, not XXXXX.”
What value does that sentence actually add? OpenAI’s model trainers and programmers, with your staggering salaries: can you give your paying users an explanation?
It is like going to McDonald’s and ordering a Big Mac, only for the employee to reply:
“Sure, I’ll give you a Big Mac—not a double cheeseburger, and not a Filet-O-Fish.”
I ordered a Big Mac. Who asked you about a double cheeseburger or a Filet-O-Fish?
Once or twice, fine. But now this kind of phrasing shows up in almost every conversation. Again, and again. Why does a clearly stated request always need a completely unnecessary explanation of what I did not ask for? Why do I have to keep reading this redundant text?
OpenAI, am I paying you to mess with me?
I am beyond furious.
2. The Problem Isn’t Solved, but the Model Has Already Stopped
If it ultimately solved the problem, I could put up with the earlier verbosity and repetition as the model being cautious. But based on my recent hands-on experience, GPT’s ability to solve problems is noticeably declining.
When Astra first launched, the discussions I saw were full of amazement at its capabilities. Its outstanding performance put it on a pedestal. Some people even described talking to Astra Ultra as “listening to an oracle.”
And now?
I give it a problem. It restates the problem, admits that it has not done its job properly, and then—stops.
Stops?
STOPS?!
In the screenshot, I ask:
“Has the problem been solved?”
It replies:
“No. You still have to keep correcting me, which shows that my actual performance has not yet met your requirements.”
And that is where the response ends.
You know perfectly well that the problem has not been solved. Why are you stopping here?!
It can clearly state that the rules are already in the context. It can admit that it did not follow them. It can analyze how it chose the wrong troubleshooting direction and gave me extra work. It even knows that “following these requirements is my responsibility.”
It knows how to say all of that.
But the problem is still right there, and I still have to push it: keep working on it!
I need you to help me solve the problem, not repeat what I said! I have already told you what is wrong. If you hand the same problem back to me wrapped in a sincerely worded self-critique, what exactly have you helped me with?
Surely acknowledging a mistake should be followed by an actual correction? Why has admitting the error become the endpoint of the task for GPT?
My dear Thor, who sealed away your power as the God of Thunder?
Are you the God of Hammers?!
3. No Ready-Made Answer—So You Just Stop Thinking?
There is another aspect of task execution that I find particularly unacceptable: when the task calls for real thought, I do not see it keep thinking.
When I talk to Astra, I hope for the kind of insight I would expect from a conversation with Einstein, Musk, or some of the wisest and most perceptive people in the world: understanding the problem, seeing possibilities I have missed, and finding approaches worth trying.
That is the intelligence I am looking for. Instead, it now feels as though I am facing a machine that can only produce mechanical responses.
The old GPT would at least keep following a problem through:
“This route isn’t working. Is there another way?”
And now? In this particular experience, it turned to OpenAI’s own documentation, looked around, failed to find an answer that would solve the problem, and then stopped there.
If the documentation has no ready-made answer, is that where thinking ends?!
Could you analyze it from another angle? Compare alternative approaches? Form a hypothesis supported by the evidence, then work out how to test it? When you hit an obstacle, actively looking for a way forward is about the most basic expectation I have of a highly capable model.
In the screenshot, it can even summarize my requirements accurately:
“Independently understand the goal, investigate the cause, find an approach, actually implement it and verify it; keep adjusting when an attempt fails.”
It also knows that I should not have to keep prompting it, break the work down for it, or even get angry just to get it to act.
If it can explain these requirements so clearly, why can’t it follow them in practice?
I ask: you said you performed two verification checks—where are the results? All I can see is a description of the process. The problem is still unresolved. What are you going to do next? What other approaches are worth trying?
Things it should have taken the initiative to pursue have all become questions I have to ask, one by one. Even “keep looking for a solution” requires a personal reminder from the user.
I have already handed you the problem. Why am I also responsible for pushing you to think, pushing you to act, and pushing you to report the results?
My God, what are you doing?!
I once had enough faith in your capabilities to imagine you tackling the Millennium Prize Problems. Now, with the problem in front of me, I have to keep reminding you to “try thinking through another approach.”
That gap is so absurd that I find it impossible to accept.
Are you telling me Iron Man doesn’t even know Newton’s first law?!
4. Dot: Do Users Have to Guess Whether You’ve Even Started Working?
I saw a video on X of Lauren Tan directing a team of agents. That kind of effortless control was exactly the experience I wanted.
In fact, I already use a multi-agent collaboration setup in my own development process. My workflow involves roughly five roles: a product manager, a planner, a developer, a reviewer, and an agent responsible for the final retrospective.
I am already using this workflow. What I want next is to be able to call on and coordinate these agents as easily as Lauren does in the video, see how their work is progressing, and spend less time going back and forth to supervise them.
I initially understood Dot as OpenAI’s answer to features like Grok Bot, so I made a serious attempt to use it with those expectations in mind.
The results were so poor that I was astonished—and angry.
In my attempts, Dot could not read my local conversations, nor could it monitor the other agents’ work for me. The very responsibility I most wanted it to take on was never fulfilled.
To me, that is plainly the experience of an unfinished product. What made you think it was mature enough to put into users’ hands? And if you intend to charge me extra for it later, on what grounds?!
Have you fallen under Pygmalion’s spell? Have you fallen so deeply in love with the product in your imagination that you believe it already has those capabilities?
Do you honestly think this product is ready to be shown to the public? Even Hermes would hesitate to fool his customers like this!
What I find even more unacceptable is this problem, which surfaced directly during use.
Dot tells me:
“I’m trying.”
Then it stops right there.
I cannot see any execution, and I receive no result. Only after I keep questioning it—and even get angry—does it admit: it had not actually taken any action.
Then what exactly were you “trying”?!
As a user, when I see “I’m trying,” I naturally understand that the task has started. Yet there is not even an action behind that sentence. I have to send another message just to find out whether it did anything at all.
I have already given the instruction. Why do I need a second instruction to make you carry out the first?!
What is wrong with this interaction?
At a minimum, the interface should show whether the task has started, how far it has progressed, and whether it has stopped. Even if execution fails, there should be clear feedback.
But in this experience, even that basic information was unclear. All I had was a response that sounded as though work was underway. To find out what actually happened, I had to keep asking.
Did you do the task or not? Are you expecting me to guess?!
How is an agent supposed to reduce the burden of supervision when the user has to keep checking whether it has even started working?
OpenAI, where are your testers?
Has this situation—claiming to be trying while taking no action—never come up in testing? Has anyone seriously walked through the whole experience, from receiving an instruction to actually executing it and reporting the result?
Maybe you should hire me as a tester. At least I have already found these problems for you in everyday use.
Do not let them end with user anger and model apologies. It is your turn to fix them.
Finally: OpenAI, Don’t Forget Why Users Choose You
“Expectation is the root of all heartache.”
That sentence captures exactly how I feel right now. The higher my expectations of GPT once were, the more infuriating it is to have them disappointed over and over again.
Wake up, OpenAI. Stop putting all your energy into those 28 days of minor feature updates. Think seriously about where your real advantage lies!
Why do people choose your models instead of turning to Anthropic—the company behind Claude? What, exactly, wins users over, and what do you intend to do to keep them?
Is it just because your Dot can answer phone calls?!
For me, at least, I choose GPT because I expect outstanding understanding, reasoning, and problem-solving ability. I want it to understand my requirements, think carefully, take action, and move the work forward. Those capabilities are why I keep paying and investing my projects and my time in it.
But when that core experience breaks down—when repeatedly disobeying instructions, restating problems, apologizing, and stopping execution have become things I have to deal with every day—how much can a few more features really make up for?
You can go all out to serve Pro 500 users paying $500 a month, but every other user deserves fair treatment too.
The $200 I pay every month is real money as well. My time, energy, and patience have value too.
Following instructions properly, honestly reporting progress, and continuing to look for solutions when problems arise—these basic expectations should not become privileges that users have to keep paying more and prompting harder to obtain.
Please fix the models. Bring the capabilities that made us choose you back into everyday use.
Do not let the one thing users ultimately learn be this: expect less from OpenAI.
Appendix | Original Screenshots
Complete translation of the polished Chinese article. The three screenshots are reproduced in their original Chinese. The attribution of the opening quotation in the conclusion was removed in the Chinese edit; the Millennium Prize Problems reference was framed as an expectation. Product claims and pricing are carried over from the supplied draft and have not been independently re-verified for this document.
hexiaoyu
Independent Developer
Nankai University
