GPT4 / Custom repeating itself and deleting its messages | OpenAI | Page 1

woven narwhal Feb 15, 2024, 6:31 AM

#

Bug Report:

Whenever I start a new conversation, it works fine the first two or three prompts and answers, but after that nothing works - as soon as I refresh the page, the last answer(s) will be gone and all of the prompts I've sent up to that point will be crammed together into my first message.

Also, the GPTs will start repeating their answer from the previous prompt - and then everything vanishes as soon as I refresh the page. This happens on the standard GPT4 model as well as all of the custom ones I've tried.

I've archived all my other chats, no change. Deleted them, no change. I have tried Firefox, Opera and Safari so far, all without any extensions and in incognito mode. I have tried on my computer and on my phone. Nothing seems to work.

Steps to reproduce:

None, I don’t think this issue can be reproduced.

Expected result:

Ability to have a conversation with any CustomGPT or GPT4 without it repeating the answer to the previous prompt and then deleting half the conversation.

Actual result:

I’m unable to hold any conversation with more than two prompts without GPT repeating itself and deleting its answers.

Additional information

Browser: FireFox / Opera / Safari
OS: Windows 10 / iOS

azure ravine Feb 19, 2024, 10:38 PM

#

I can't comment on anything else, but I came here to report specifically on this:

Also, the GPTs will start repeating their answer from the previous prompt

This isn't just happening with Custom GPTs. It's happened a couple times in my Plus sessions.

I would provide a link to the most recent and very brief and obvious occurrence, but on attempt to get a link I get an internal server error, which may be related to the issue.

Here is a link to the session from the browser (Firefox/Windows) address bar. This is probably of use only to OpenAI staff :
https://chat.openai.com/c/d79aef76-308d-4a4c-ad02-48cd64eb1bbd

I've attached a screenshot to verify the Internal Server Error on attempt to share the session.
I've attached another screenshot that shows my very first prompt, the first GPT response, my second and different prompt, and GPT responding again to the first prompt.
Note, especially helpful for our purposes here, that in my Custom Instructions I have the GPT numbering responses so that we can refer back to prior prompts and responses (nice, eh?). We see in this exchange that it repeated response [R01], confirming that it lost the context after the first response.

I hope that helps.

#

I'm glad I got the screenshots. On refreshing the page everything after the first prompt was deleted, including the first response. This verifies the original report.

Note also that it combined the first and second prompts into a single prompt.

I've seen that happen before too (reported in #chatgpt-discussions , so this is my third experience with this.

#

Follow-up. This just happened again in a new session. Context is broke.

ivory cairn Feb 20, 2024, 2:58 PM

#

Can confirm that i have the very same issue. Posts fine for while, and then it suddenly starts to respond to the past prompt in addition to the current one, and then when you refresh it folds the prompts together as if it never answered the previous one

past ember Feb 20, 2024, 3:20 PM

#

Similar with me

woven narwhal Feb 21, 2024, 12:20 AM

#

azure ravine Follow-up. This just happened again in a new session. Context is broke.

It happens to me in every single session now, no matter if it's a custom or raw GPT4. Both accounts in my workspace are affected. I'm still waiting for a reply to my ticket...

azure ravine Feb 21, 2024, 12:20 AM

#

Happened again today.

woven narwhal Feb 21, 2024, 12:21 AM

#

Do you have sessions in between where it doesn't happen? I did in the beginning, but now it's just all of them. It's completely unusable like this.

ivory cairn Feb 21, 2024, 1:02 AM

#

Yeah it happen for all of mine too now. It varies in how long it takes before it kicks in, but it inevitably does.

woven narwhal Feb 21, 2024, 1:52 AM

#

I wonder if it's the account that's just somehow compromised. Do you guys use Plus or Teams?

ivory cairn Feb 21, 2024, 12:39 PM

#

I use Plus personally

woven narwhal Feb 21, 2024, 5:40 PM

#

Hello,

Thank you for reaching out to OpenAI support.
We understand the steps you've taken to troubleshoot the issue, and we acknowledge your efforts and despite trying different approaches, the problem still persists. We apologize for any inconvenience this may have caused you.
Let's see if we can get this sorted out. Can you try refreshing your browser and clearing its cache and cookies? This often fixes the issue. If you're still stuck, we'd appreciate it if you could share the following details with us:

The exact error message and code you're seeing
A screenshot to show us what's happening
Any other info you think might help us figure out what's going wrong

This info will really help us get to the bottom of the problem and make our service better for everyone. Thanks for your help!
Best,
Annah
OpenAI Support

Got this reply from support. I have sent screenshots, but since I don't receive any error messages I couldn't send those. Did anyone receive one?

nova latch Feb 22, 2024, 11:25 PM

#

I seem to be having a similar issue. In my case, things broadly work fine but have recently started throwing Network Errors towards the end of a generation, and tends to be losing a chunk of context when I try to reprompt it. In my case, it's only with a custom GPT I'm using seemingly

#

In my case, generation happens and then stops, basically near the end and errors in both Chrome and the Android app

woven narwhal Feb 22, 2024, 11:53 PM

#

nova latch In my case, generation happens and then stops, basically near the end and errors...

I've never had that happen before. Does it also delete messages and put all of your messages into a single one when you refresh the page?

nova latch Feb 22, 2024, 11:54 PM

#

No, but it does delete all of its botched responses on refresh, which is pretty normal

#

When I reprompt it, it often starts with the same discussion of a file I uploaded generally, so you can see it keeps talking about Chapter 8 even though it should be continuing the conversation

woven narwhal Feb 23, 2024, 3:07 AM

#

Very strange. I'm still waiting for another response from support, the last one I got said that a specialist team needs to take a look at this and they'd get back to me. If there's no response within the next few days I'll have to cancel my subscription before it renews, I'm not paying for something I can't even use.

nova latch Feb 23, 2024, 8:42 AM

#

I can confirm there are definitely issues. Now the custom GPT is struggling to read uploads, which is an issue I had a few days ago

azure ravine Feb 24, 2024, 5:18 PM

#

Happened again today: Query1, Response1, Query2, Response2, Query3, Response2. As recommended by OpenAI Support above, I've cleared browser cache for the openai.com domain - though that's a common "punt" and I don't think it has anything to do with this specific issue.
I'm using Plus with GPT4.

I can visually see the Response1, Reponse2 discrepancies because my Custom Instructions explicitly tell GPT to preface responses with a sequentially indexed number - so in this case it's actually repeating the numbers, so I know which prompt it's responding to. 🙂

woven narwhal Feb 25, 2024, 7:43 AM

#

Yeah, that didn't work for me either. I'm still waiting for the response from their "specialist team", it's been four days by now.

nova latch Feb 27, 2024, 11:56 AM

#

I'm in a similar boat. I'm having similar issues like file read errors even in "normal" ChatGPT now, so I think they either tweaked something somewhere, or it's only affecting a select few users

woven narwhal Feb 28, 2024, 3:02 PM

#

woven narwhal Yeah, that didn't work for me either. I'm still waiting for the response from th...

A week now and my messages are left on not seen yet. Canceling subscription now - this is terrible customer service.

obsidian vigil Mar 31, 2024, 8:47 PM

#

Hey! I have a fix for this. You need to turn off QUIC in the browser your using. This protocol doesn't play nicely with chatgpt and will constantly have issues. QUIC is an experimental protocol not yet support by all routers and it often fails.

To disable QUIC protocol on Google Chrome or Microsoft Edge, you need to
1: Open Google Chrome or Microsoft Edge.
2: In the URL address field, type chrome://flags or edge://flags and press Enter.
3: Type quic into the Search flags field to filter the list of flags.
4: Select Disabled from the drop-down menu next to Experimental QUIC protocol.
5: Restart Google Chrome or Microsoft Edge for your changes to take effect.

If you're using firefox the instructions are a bit longer so please look those up to disable QUIC.

#GPT4 / Custom repeating itself and deleting its messages