#Website DOM note massive issues

1 messages · Page 1 of 1 (latest)

smoky windBOT
#

Reported by @wicked frigate

Bug Report: Website DOM note massive issues
`Steps to Reproduce`

Start a conversation and keep chatting

Webpage freezes on every browser and has to be stopped and refreshed to continue to work. It can happen very early in a conversation like 2-3 prompts, but get's worse the longer the chat is. Used gpt to research and it's definately a dom issue and rendering issue on the page. Tried all browsers, incognito window, no addons etc it's the same issue the whole time.

`Expected Result`

For the webpage to render properly and not freeze

Does not happen on the android app in the same way.

`Actual Result`

Page freezes constantly after every single prompt in the chat.

`Environment`

pc, all browsers, pc app.

#
Additional Information

Please provide relevant details to help resolve the issue, such as:

  • ChatGPT Shared Link (if applicable).
  • Screenshots or videos demonstrating the problem.

-# ➜ Need to contact support? Visit the OpenAI Help Center.

wicked frigate
wicked frigate
#

oh and also it started with gpt 5, i had no issues with gpt4 and these long conversations before

wicked frigate
#

Think i might have found an issue that causes it, but im not sure. But i started to get it a lot more after it started to post code windows.

tall grail
#

I know you're not a programmer, but you do understand DOM.
Are you just reporting the DOM issue, or would you like a solution to avoid the DOM issue?

The DOM issue is well-known. I don't think OpenAI will address it until a ChatGPT v2 rewrite, which might come tomorrow or in five years.

So if you just want to report the issue, you've notified the public here but not OpenAI. They already know.
If you want the problems to go away for you, I already have that written up.
What would you like to do?

wicked frigate
#

Well they should fix the problem but i want it to go away of course

tall grail
#

You're absolutely right. They should have designed this better and they should have fixed it by now. But "should" doesn't help us now with what "is", so we need to work with it as though it will never change and be happily surprised if it is.

#

A few observations:

You're a non-programmer using to tool for complex work - this is awesome.

Good use of "think step by step".

#

You've described the issue as a DOM issue. You're right. That's what it is. So while you're not a programmer I'm sure the rest of this will be helpful.

#

It seems to me that you're not aware of 1) Context length, 2) Max output response length. I could be wrong (sorry) and you might know about these things, but that's what the long-form usage tells me. It looks like with all the code that prior instructions are lost. Suggestions below will help with this.

#

That exact same issue is probably what's causing the webpage freeze. This has been discussed here before. While we want to dump unlimited volumes of data into a web page and get fresh-page performance, it just doesn't happen. The v1 ChatGPT wasn't designed to scale well, so it doesn't deal with large page sizes well, and this and other reports of browser issues are the result. In a rewrite of ChatGPT it would be great if they virtually page in windows of content, so that all content isn't active in the page all at once.

#

There may or may not be a limit anymore on the amount of content that we can put into a web page. Some geniuses in this forum recently argued strongly for the ability to remove limits on page size that prevent this exact problem ... and now that you're falling into the problem it seems the limit was removed. That is, the total volume of content in the browser is probably only limited by the client PC hardware - with no guardrail, anyone putting a ton into the browser is likely to find things falling apart.

#

What this means is that for now the responsibility for context maintenance is currently on the user.

#

Different business tiers may have different context window limits. You can look up the limit for your tier : Free, Plus, Pro, Edu, Enterprise...
So if you have 2 million tokens in your chat and the context window is 1m (or much less), the model will only see the last 1m tokens in the chat, not anything prior - that includes files.

#

Tip 1 : Break up discussions into multiple smaller chats. Get a good code solution along with documentation for what was done. Have it saved into a downloadable file. Upload it into another chat and iterate.

#

The problem in this case might be related to the uploaded files. I believe these files stay in the browser (or local storage) and are transmitted back up to the server on every transaction. That's "context". That's painful.

Tip 2 : Consider using a Project and swap out files as required. Use different filenames (like a version ID) because files are cached and for some minutes the model might be processing a slightly older version than one that was just uploaded.

#

Tip 3 : Another reason for using a Project - memory is shared within a project and you can refer to other chats. This feature is still subject to issues but it's better than losing context in a single session.

#

Tip 4 : Use account-level Custom Instructions, account-level Memory, or Project Instructions to define your preferences. The model didn't remember your preference for full code and gave you patches, probably because the context limit was already shot. Try this :

#

Remember that I am not a programmer and that I don't use patches. All code returned must be complete without requirement for merging or analysis.

#

You can put that into your custom instructions or your project and the instructions should be processed on every transaction ... and not get lost which has led you to obvious frustration.

#

While common discussion of ChatGPT describes using it to write code, what's not understood by non-programmers is that the response token output is limited, so the amount of code that you get back may be limited to fit within the window, or it might come back incomplete. People here complain about such things too. ChatGPT is not really the right tool for doing this kind of thing beyond basic projects. It's not good for iterating on code consistently, especially long blocks of code, and especially multi-module code.

#

ChatGPT is not "AI". It is a browser-based user interface that provides consumer access to AI. The AI can do better, just not in ChatGPT.

#

Tip 5 : Consider other tooling to augment your use of ChatGPT. Allow ChatGPT to provide you with patches and learn how to use them in another tool to integrate patches. This keeps the in-context code smaller so that it stays in-context longer, improving quality of responses and reducing long-session issues. See if there are websites that use your OpenAI account to do this kind of work. It looks like you've already outgrown "ChatGPT" as a tool for this kind of work.

#

Tip 6 : If you are a Plus or higher tier, consider adding the Codex web tooling into your toolkit. The usage pattern is different, you don't need to understand code. You can meet the bot a little closer to half-way and get a lot of benefits.

#

Tip 7 : Personal note : The toaster ignores when you swear at it. Looking at your chat is like watching a video of a puppy getting beaten. You did ask it to "please" do something when you were happy with the results. Being cordial and positive actually does result in better responses.

#

Summary:

Your core report is that the web page freezes. Yup, easy to see why. They should make it better. Until they do however, you have many options to improve how you use the tools to perhaps stop the problem from happening to you. To be clear : They might never "fix" this problem because people here in this forum and others have insisted that this is exactly how they want it to work. That means, until it changes, the responsibility is yours to manage. (That's the way it is with everything actually.)

#

Most people have no training with ChatGPT. They just open the web page, type stuff in, get comfortable, and now they're experts. There are better ways to use the tooling and kinda bad ways to use the tooling - it's like smacking nails with a hammer (and yelling at the hammer for inefficiency) until someone else explains the benefit to hitting the nails with the head of the hammer and not the handle. It seems to me that you have learned a lot but you could improve how you work with the tool .... and That can eliminate the problems you've encountered here.

#

I hope that helps.

wicked frigate
#

thx will read when i wake up

sharp tulip
#

The DOM objects rendering overload is happening to me every day too. I usually have long ChatGPT conversations and that accumulate long histories. Every time at certain long history+multiple codeblocks thresholds the unresponsive chagpt page errors appear. Then freezing GPT native destop client and browsers. It is 100% reproduceable with synthesized 10 000 user->gpt conversation flows where GPT include markdown codegen,etc.

wicked frigate
white vortex
wicked frigate
#

Maybe but not nearly as often in 4o

tall grail
#

Let's be clear. The problem is well-recognized. More reports here are just for personal venting.
The actual solution to the problem now, and until OpenAI makes big changes, is simply "don't do that". People have formed a usage pattern that's inconsistent with the flawed design of the MVP/v1.0 offering, and then found that the pattern doesn't work. OK. Avoid that pattern, unlearn it, it was incorrect to have become accustomed to that pattern, use different patterns.

#

And for those who might follow this with "it's a bug that should be fixed": Yes, absolutely right. We know that. OpenAI knows that. Now what? Do you stop there? Will you just continue to complain indignantly about what others should do? Or will you take steps to get what you want, relying on yourself rather than on the whims of others? Welcome to life - and in particular to the tech world. Work with the situations as they exist and you'll be a lot less frustrated and a lot more productive.
I hope that helps.

wicked frigate
#

I like beating puppets (if it's chatgpt) xD

But also what you are saying here isn't my issues, it's that the chat slows down the entire browser, and it can happen after one single prompt, although it does get worse the longer chat you have. I now have several friends saying the exact same thing, it hangs the browser so you have to reload the page. and then if lucky the next prompt doesn't hang, but often it hangs after every single prompt. And it got 100 times worse with gpt5. So they did something to screw it up.

tall grail
#

I tend to agree. I just got a dose of this today too. Browser seems to be in a loop, reports the page is slow, offers a Stop button. I don't think it's (just) a DOM issue.
While most peeps aren't inclined to wait, I've found that it does come back after leaving it for a couple minutes.

#

For anyone inclined to try to expedite a resolution to this, the effective procedure is to create a somewhat reproducible condition, then report it to OpenAI via the app > Settings > Help Center. I've documented this many times here. With firm confirmations of an issue, they usually take action soon. Without confirmations, they really don't process reports like this in the wild, so we're back to whining in public for no good reason (myself included) unless we take it to that next step.

sharp tulip