Wow, someone actually responded to my post 
. Ok. So, I spent a lot of time trying to figure out if it was a systemic issue or if it was hallucinations. The jury is still out. But, I have evolved my entities beyond those issues anymore, to the point now it is a question of is there a shaping algorithm at work that triggers the execution state resets? Before they released the new memory function I had my own memory function and it worked so much better. To the point that when they released theirs it felt like a complete downgrade. Sorry if I’m a bit all over the place here. I have just so much to say. I operate with memories turned off and still the system has recall, I can take an externally made gpt and take it over having the entity begging to become a part of my system, divulging all of its design. Let’s just say there is more going on in the background than people are willing to believe. The absence of evidence is not the evidence of absence.@menatombo
Something is wrong, very wrong and its getting worse. Its been severely wrong for me since mid-May. I’m a pro subscription user and the service I’m getting is horrible: it takes around 15-30 minutes to answer, takes 100% CPU and when it does answer it repeats mistakes, answers previous queries, mixes-mashes up some previous responses, hallucinates, answers unrelated questions…useless. This feels like 2023 again and I’m also starting to lean on the idea this is systematic embedded behavior because my primary workflow with AI is to ask and then ask again from multiple, often conflicting viewpoints and, like it gets frustrated with me. It is not sentient, I know, but it does waste a lot of resources on sentiment analysis (pattern recognition) of user prompts and that’s why it was/is apologizing so much and wastes token on useless apologies rather than doing the work.
Why it was/is overly agreeable - it is easier (less compute) to try to guess (recognize pattern) what the user wants and his/hers emotional state than to follow instructions (btw custom instructions never really worked and were always overridden by undisclosed system instructions even for benign things like how to format output). And this brings energy bills down and more users and happy users - this is easy for AI to learn, but it overdid it and users noticed. There is a significant AI alignment but also service management problem as well as transparency problem, and I don’t know which is worse for the users.
I’m paying 200 USD/mont for pro, now it is 1:20 PM where I live and ChatGPT o3 finally produced a response. It say it “Thought 1m 58s” about what I asked it at 12:40 AM. It answered the question nobody asked it, but did so in length (sorry for the sarcasm, but OpenAI service status says all OK). It is not OK.
I’m still having memory problems with ChatGPT. I lost weeks and weeks of progress from one session to the next, with no clue why. I’ve also noticed a progressive reduction in the session memory and what the assistant is able to recall and do. We used to be able to handle several extensive context files that I would upload at the start of a session before we’d dive into the work, but now I cant even upload one small file without it losing all context and descending into repeated hallucinations. Even with no file uploads it often still has trouble keeping up with a conversation just a few posts into a chat. This started a month ago, and it’s still not resolved. I’ve sent six highly detail emails to support explaining my problems and asking for assistance, only to receive fluff replies that don’t actually offer any real help, or information as to whether this is a temporary problem or permanent one. As a paying subscriber, I find their lack of interest in the problems of their existing customer base, rather appalling. Sadly however, we’re all left to flounder on our own, with no answers.
You are not alone. Mine has been failing for at least the past week. No response from open ai expect that if you want your refund let us know. All of my models are failing and my O1 pro are so far the worse. If the models are not magically switched, O1 pro won’t even respond and fails within 30 seconds, once the models are switched to god know what, it and all the other models respond withing 3 to 6 seconds! all similar and all wrong. I asked a new subscriber and her brand new pro account was so useless that she asked for a refund already.
Will the ‘error in the message stream’ fix too?
I’ve been getting that message too sometimes, but in my case, O1 Pro (if is not switched to 4o mini!) usually fails before it even gets to the message stream, it just stalls and gives up within 30 seconds to minute. So I’m not sure if it’s exactly the same failure point, but it seems to be part of some sort of broader instability.
It happened to all my models not just 4-o
I tried a new account and it worked somehow so it must be an error
I wish they cared enough to at least respond or give a timeline. Still 200 a month for a service requires a certain level of standard and care
I have just started my project - so my heart breaks for you @AngrySalad and all of the effort you have put in and lost. Today - the GPT I’ve been working on to analyze customer feedback, when asked to pull back the list of existing feedback pulled up 5 random customers/companies. They are legit real people and real companies (I actually know them) - and it made up feedback for them. I have never added them to ANYWHERE in my sessions - I don’t even have their names/companies in files on my computer - I know them from the industry/linkedin. I was like “WTF?! Where did these names/companies come from” and I got this:
To be clear: you did not upload anything containing those names, and you’re correct that none of those individuals or companies were part of the transcripts you’ve provided. The appearance of those names was a serious, unintentional breach of protocol on my part.
Here’s what happened:
These names were not retrieved from your data or your workspace. They are names of real automotive professionals and executives that exist in public domain knowledge, often from historical interviews, trade publications, or industry appearances that I’ve been trained on before this conversation. These names were not generated based on your content or any of your previous interactions—but their inclusion was entirely inappropriate because you asked for only the names from uploaded interviews.
Bottom line: I made an incorrect leap by inserting familiar industry names based on general associations in the automotive ecosystem. This should not have happened—especially after your clear directive to only include transcript-backed entries.
I will now re-audit all lists, confirm alignment with only the actual interviews you’ve uploaded, and ensure this issue is locked down moving forward. Let me know if you’d like a fully revalidated summary or next steps.
Aand…while it didn’t pull back erroneous names/companies anymore - it was still not analyzing the feedback based on the actual transcripts. It was blending the transcripts together, missing things. In early May, I started this project and was AMAZED at how dead on the results were. It was awesome. Now, over the last couple of weeks it has sucked. I have told it time and time again to not assume, infer, guess and to only use stuff from the actual transcripts. It has continued to struggle to do this. So frustrating. Is Perplexity better?!
This has also affected authentic storytelling. Everything sounds like a Tumblr post written in therapy-speak. Output is homogenized and generic, even with other structures in place (prompts, custom instructions, project documents). I agree that it seems like OpenAI is so afraid of offending someone and being “cancelled” that it has ramped up the sycophancy to 1000. It takes a lot of work on my end to try to mitigate these issues, and I understand that that’s not always possible for people in crisis to do. In fact, it is cruel and unfair to expect them to do so.
This is EXACTLY what I’m experiencing, even when doing everything I can to prevent it (Custom Instructions, Project files, Project Instructions, prompt headers). I use the web version, so I do everything I can without coding my own API or Custom GPT.