Inconsistent and Unformatted Code Output Across Multiple Languages

[steel oracle]

Bug Report:

Steps to reproduce:

1. Request an AI to provide updated function code for a project.
2. The AI outputs multiple code blocks with unformatted code, as shown in the provided images.

Expected result:

A single, well-formatted code block in the correct programming language relevant to the project.

Actual result:

Multiple unformatted code blocks are provided, spanning different programming languages, which creates confusion and is not practical for implementation.

Additional information

Browser: Google Chrome
OS: Windows 11 23H2

[leaden spindle]

same for me here

[sturdy whale]

I have begun getting this issue today and it is very annoying.

A potential workaround is to select all the code (and some of the surrounding text), then paste it into a GPT 3.5 window and ask it to format the code into the language you want for copying/pasting. That is how I have been dealing with it.

[pure isle]

Bump

[bronze seal]

yes. same

[untold anchor]

same for me

[steel oracle]

Bump

[glacial carbon]

Same

[tame path]

Same. Often accompanied by a network error or stream error.

[granite phoenix]

This seems to be rendering problem, if you copy the reponse, you will ind the markdown correct (at least for me)

[drowsy slate]

Got same output too!

[steel oracle]

They still haven't fixed this...

[lyric flax]

Been grtting this for two days now. Formatting is off, constant stream errors prompting to regenerate, etc.

[snow arch]

same for me

[clever harbor]

really annoying stuff. i first thought it had something to do with network connectivity on my end (we had some fiber breakage so bandwidth was considerably limited) but as it turns out the issue persists for 3 days in total now. Not a connectivity issue 😉

[leaden spindle]

i like it that the badly formatted language is really random, how is dart and less not recognizable form a mile away.

[crystal crow]

Same here - seems there is an open thread here too: https://community.openai.com/t/code-blocks-are-split-in-code-and-text-replies/586244/8

[polar prawn]

same. lots of broken code blocks.

also lots of code blocks not in code also. so sometimes 3 sql blocks just written as text then a four would be in a codeblock. not rythme not reason. i asked the ai if it was the custom instruction and she just say something like "naah was my bad"

[crystal crow]

I feel like it's getting dumber by the day 😢

[tall scaffold]

Having the same issues. Seems to be worse with Custom GPTs but also happening with normal GPT4

[steel oracle]

Bump

[leaden spindle]

fr

i just work as much as possible, thinking it will go away someday

[crystal crow]

I've reached a point where GPT3.5 is giving me better and more consistent results than GPT4 🤯

[leaden spindle]

in what language you need code?

[crystal crow]

Python mostly

I prompted GPT4 for a few hours every day in the last 2 months and since last week I never get any satisfying results, either the code blocks are broken or my instructions are not followed (eg. "don't put any placeholders, always provide the code in full" etc)

[tame path]

I noticed this when gpt4 has been broken

gpt3.5 has a lower token limit, but seems to be doing just fine with python

[leaden spindle]

yeah same, gpt 4 gave me more than 50% of my backend, so i was so happy but then it stopped working

[crystal crow]

i hope it gets fixed soon otherwise I'll just cancel my Plus sub as it's useless as is 😢

[weak fjord]

indeep one could solve this by prompting differently.

[leaden spindle]

how?

[halcyon narwhal]

More of the same, this wasn't happening to me before, it often succeeds when I ask it to rewrite current code in a new code block, but that burns an extra message

[weak fjord]

i totally disagree on free riding, i think one should sacrifice something to get rewards.

pointing a direction is already very helpful, when someone is out of no where.

[crystal crow]

What do you mean by free riding?

[weak fjord]

get reward without sacrifice

[leaden spindle]

bruv i really cant understand anything

[pulsar harbor]

same here gpt going silly

[sinful blade]

I too have this issue.

[next cloud]

this also happens after barely it writes anything and then fails and says error in message stream

might this be becuase they launched the gpt store?

[crystal crow]

after 3 failures to generate an answer, another one

☠️

about to smash my keyboard 💀

[steel oracle]

Me too

[crystal crow]

ahah I got the exact same display with the bulletpoints ; i'm getting crazy

[steel oracle]

and the cherry on top of the cake...

[crystal crow]

LOL yes :rofl
"existing code goes here..."

[steel oracle]

I'm about to burn down my pc

[crystal crow]

i wonder if they downgraded the overall token limit/model due to GPT store release

[steel oracle]

Not only that but also the iQ

man chatgpt is useless bruh

if it was before, imagine now lmao

[crystal crow]

What's super frustrating is that 7/10 days ago I had a glimpse at what could be/have been GPT4.5 or even GPT5 : gpt4 replies with the speed of GPT3.5 ; generating perfect replies in <1sec for very complex code tasks

Somehow that only lasted a couple hours and then it went regular, then bad

[steel oracle]

yep

[crystal crow]

enough, I quit

[steel oracle]

I just hope they release GPT-4.5 this week because this is not acceptable for a billion dollar company

more than a billion

[steel oracle]

yeah

[weak fjord]

u know, those who choose to use chatgpt is more smart than those who dont. and among those who do, there are also a normal distribution of IQ.

is*

[crystal crow]

lol i wrote
" list 50 idea for XXX"

this is what I got with GPT4:

[steel oracle]

imagine if they released a ChatGPT plan specifically for programmers, dang

will never happen

[steel oracle]

that's awful

[steel oracle]

Look at this, this is just not acceptable

28 lines only...

even gpt-3.5 can do more

[sharp burrow]

Seems this is part of a broader gpt-4 service degradation issue. Any idea when OpenAI will officially acknowledge and log the issue on the service status page ?

[steel oracle]

never

it would look bad for them

this is not the first time they don't log a real issue

[ionic trail]

same!! it keeps for 3days!!!

[ionic trail]

This is a fatal problem for ChatGPT. Why should we pay for openai's mistakes? For tens of dollars a month, there are only 40 responses every three hours (even error responses are included). This is not fair!

[steel oracle]

Exactly

[crystal crow]

Only a matter of time before other users (non "power users") notice the degradation too I guess

[ionic trail]

https://x.com/HW7xpJxpiDhvfil/status/1746945747315232927?s=20I have already sent a tweet and liked the great CEO who just got married. Please help me like him. Thank you.

[steel oracle]

The thing is, they never listen to the community

And if they do, it's only for BIG BIG persons

Like very famous people

[ionic trail]

We need Elon Musk

[steel oracle]

instead of fixing problems they launch bs planes for absurd prices

[ionic trail]

I just renewed my subscription plan three days ago, paid them a few hundred dollars, and I can't even find a place to submit a bug ticket or get support

[steel oracle]

Look what I found

https://fxtwitter.com/carbonaracred/status/1746896132910039479

[steel oracle]

wow

[ionic trail]

lol

[crystal crow]

are there any admins here? lol

[steel oracle]

Like 5 employees but always offline

[ionic trail]

I don't think there is any admin

[steel oracle]

Look

[ionic trail]

@Elon Musk maybe sooner than expected

[steel oracle]

https://fxtwitter.com/0xyjkim/status/1746776709754761386

[steel oracle]

lol

[crystal crow]

I am really surprised there is so little people being vocal about this issue

[steel oracle]

right

https://fxtwitter.com/toorandom/status/1746943317122642318

[crystal crow]

are people only prompting "what's the weather like in XX" or what

[steel oracle]

Fr 💀

[steel oracle]

You're right, look at this https://fxtwitter.com/jcubic/status/1746922942338933183

[untold anchor]

Right now it's still the same and bad 🫤

[ionic trail]

lmao

At first I thought it was because of my poor VPN network, but I found so many people with the same problem in the community

[steel oracle]

Another one! https://fxtwitter.com/CFDevelop/status/1746628303396680085

[steel oracle]

Same, I thought it was my internet

[crystal crow]

well if that's real, i'm willing to wait a couple days

[ionic trail]

I understand, it’s like after GPT4 came out, the IQ and performance of gpt3.5 dropped greatly.

[steel oracle]

https://fxtwitter.com/xbriangi/status/1746176837909356595

At least 5 people in the whole Twitter community has reported it

[steel oracle]

I think that's fake so they gain interactions

[crystal crow]

5 out of 100s of thousands seems low still lol

[steel oracle]

Yep

[ionic trail]

We can't even sell openai stock to make them repent, the designers of openai are geniuses

[steel oracle]

lmao

[oblique stump]

Mod here are not openai employee. Openai employee name are in yellow. They do answer time to time. They have a lot of bug reports to read.

[ionic trail]

Oh, it looks nice maybe

Hope them can solve the bug about code output😆

[crystal crow]

https://vxtwitter.com/aitoolcritic/status/1746957078017474692?s=20

[steel oracle]

Nice

[next cloud]

nah but what the actual heck are they doing? are they gonna fix thiese bugs or not cause it could ruin alot of things

[steel oracle]

They will just still continue to post nonsense

[ionic trail]

Ilya is Right

Altman need leave

[steel oracle]

https://fxtwitter.com/SMuangsrichan/status/1746969996474106182

[pulsar harbor]

Gpt-4 vs Bard (Palm-2 ig because no gemini in eu yet) im not very happy that i pay for less quality 💀

[steel oracle]

I just tested it and it's true 💀

[next cloud]

oh helllllll nah

how can gpt 4 be this bad

[pulsar harbor]

the Gemini pro in bard can be it at coding i bet now (didn't tryed gemini pro yet because not in eu yet :( )

[steel oracle]

Bump

[crystal crow]

As proposed by @pulsar harbor , i've officially resorted to using Bard right now , if not fixed in the coming days I'll just cancel my Plus sub

[agile flare]

Same

It's just that I've tried to find information about this issue but openai hasn't addressed it. And this has been going on for a small while. Had I not seen this in discord I'd think I'm going insane

[crystal crow]

A new low, using GPT3.5 a few minutes ago

me: Prompts something to design a very simple counter based on a variable that needs to be incremented
GPT: proposes something that has nothing to do with the logic I asked
me: asks to re-design the code
GPT: "Apologies for the oversight" + proposes the exact same code block
me: did you just send the exact same code block?
GPT: "Apologies for the oversight" + proposes the exact same code block

🤣 insanity is approaching real fast

[steel oracle]

Same

I prefer spending money on something that actually works

[steel oracle]

The bug is not only in the code but also in the normal output lol

[steel oracle]

Bump

[granite phoenix]

This is unacceptable to me

[ionic trail]

OK，I will cancel my Plus plan

OpenAI still not report this issue in status page

[steel oracle]

I will if they don't release GPT-4.5 by next week

[ionic trail]

😆

[steel oracle]

https://fxtwitter.com/BenjaminDEKR/status/1747040713563635966

[ionic trail]

still same

[steel oracle]

https://fxtwitter.com/BenjaminDEKR/status/1747042288910704873

[ionic trail]

I try again right now

[steel oracle]

They probably don't even know this is happening

[ionic trail]

Yea

They're becoming more and more like some government agency that keeps claiming to users that everything is fine

[steel oracle]

yes

[ionic trail]

[steel oracle]

💀

[ionic trail]

I subscribe in iosApp

Bro, do you know if I can get a refund through Apple?

[steel oracle]

Can't, they already spent your money on hot dogs

[ionic trail]

[steel oracle]

This is the BS they prefer to talk about https://fxtwitter.com/OpenAI/status/1746986660892741927

bruh openai is delusional

[ionic trail]

How can an AI who doesn't even talk about it worry about affecting social fairness and morality?

[steel oracle]

right

[steel oracle]

Bump

[pulsar harbor]

Bump

[brisk notch]

bump

[barren lotus]

Getting this 100% of the time now.

[last brook]

It's also happening to me. It renders the service pretty much unfit to use 😦

[last brook]

please can you give us an ETA on when this particular bug will be addressed because there are many of us who rely upon this to be fixed in order to use the service we paid for.

[leaden spindle]

gives me the code 4 times, for the same thing, and then timeout

[gusty cargo]

Getting over it

[ionic trail]

still？

it keeps for 96hours

[gusty cargo]

Forced degraded service for the enterprise levels

Its just slimey business models that every big corp uses in todays world

Only shareholders matter

GPT will not be available to you and me soon so buckle up and enjoy it

[crystal crow]

i wouldn't be so sure tbh, now that models are getting miniaturised it's a matter of time before you're able to run GPT4 on a macbook air locally

[gusty cargo]

thats not gpt, thats your own local ai

openAI will go exclusively corp

[gusty marsh]

Same s*** 😫

[gusty cargo]

WOW, they have also increased cap limit wait times from roughly 15-30 minutes to 1 HOUR

[ionic trail]

damn，Four days have passed, and we haven't even seen the official awareness of the bug

[gusty cargo]

Why did we all cheer for that CEO to be re-instated? His transparency is pathetic

Like i said "ShareHolderGPT"

[oblique stump]

I am not an openai employee just regular member like you

[ionic trail]

still same

[steel oracle]

Bump

[opal needle]

Is there any real competitor for chatgpt4 for coding?

[hidden forge]

I'm seeing this problem too

[steel oracle]

Same

[naive wigeon]

Can someone from the open ai team acknowledge that they see the issue?

[crystal crow]

I'd like to try Claude 2 but it's geo restricted in my area

BARD is not good enough , Llama2 same

[steel oracle]

Look at what these clowns post instead of fixing issues https://fxtwitter.com/OpenAI/status/1747260226099896359?t=ChkEzupjAhlmras-ECKyDQ

[opal needle]

Try using POE with Claude 2

But anyways I just stopped the subsciption to openAI and i bought Poe

Its shameless what they're doing

[crystal crow]

Thanks @opal needle will do!

[opal needle]

Poe interface is awesome and its at least 2 steps ahead of openai on terms of user experience

And they are just using existings AIs

[crystal crow]

@opal needle do you know if GPT4 model on Poe is giving the same output as GPT4 on OpenAI.com ?

[steel oracle]

@opal needle thanks! I will do the change, OpenAI is awful

[opal needle]

Ill answer in a moment

[crystal crow]

thank you! also, I see that there is a message limit for GPT4 / Claude2 - is that big enough for extensive use ?
600 gpt4 messages per month seems quite low

[steel oracle]

I don't like it anymore

[opal needle]

Ok i just tested it.

1. The output (interface wise) looks great, doesnt get cut.

2. No errors whatsoever

3. Its still lazy with the implementation but thats just gpt 4 overall

You're still paying OpenAI through Poe but i guess using the API or whatever deal does Poe have with openai makes it more reliable

The only problem i've spotted is that, when creating a custom bot it takes very long to process the knowledge (pdfs).

[opal needle]

Well, it goes pretty fast for those 600 messages per month. Also, 600 per month is 20 a day. With normal chatgpt you dont even know how many you get, and also, half of those messages just end up in an error and the message gets consumed anyways.

And it says right there that it just goes slower you will still be able to use it

And for me its worth it. Until gemini ultra gets released, this is the way

[crystal crow]

Perfect, thanks for your feedback!!

https://tenor.com/view/book-of-boba-fett-the-mandalorian-this-is-the-way-the-book-of-boba-fett-mandalorian-gif-24825263

[opal needle]

Still has some output problems unfortunately

I get sometimes ```cpp

[fervent rapids]

I am also seeing the issue shown in this thread. all code blocks are randomly interrupted and restarted in the middle of the script. It's obviously pathological behavior that started yesterday.

It was able to write a tic tac toe script in the first try in a new chat with the basic GPT-4 model, so it may be an issue that only appears in longer threads?

[tardy sundial]

Since yesterday, chatGPT is messing the format of answers. If the answer includes the code, it will have some of them as the code, and then some code illustrated in the answer as plain text and some of them as code of a different language and some plain text as code block of a random language.

Please help to address this issue. Thanks!

[tardy sundial]

Exact the same issue I am having now.

[pulsar harbor]

bump

[ionic trail]

bump！

[pulsar harbor]

bump again

[halcyon narwhal]

There's no indication of an open incident -- someone ping Altman, LOL

[rose oyster]

yh lol shoudln't be like tho i hate GPT-4 Turbo studdonly refused to do what you want most of the time vs GPT-4 it even engages in deception. Like giving me code that doesn't even includes what I want

[cobalt flicker]

Bump

[ionic trail]

lol

altman just have a vacation with bf

[rose oyster]

Yes like all the open source model on top of huggingface leaderboard can code, I've made a program that loop them feed them error codes so they can self correct it works lol. Open source models on hugging face are really clever just tick all boxes and go for the highest one, also they have this flagged scheme as well which is nonsense all those flagged models are extrememly intellgence in face i use them for python code i think they're just scared how fast AI is moving

https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard

[crystal crow]

am I stupid or is this exactly the same?

[leaden spindle]

bump

[halcyon narwhal]

Separate problem I think, it's been doing that to me since the turbo update rolled out, off and on -- it gets stuck and the only way to get it unstuck is to point out the problem in detail or provide the fix yourself, otherwise it just keeps changing "nothing" or unrelated things. What's fun is when it changes a comment and somehow expects that to resolve the error 🥳

[next cloud]

im so mad

[halcyon narwhal]

https://tenor.com/view/april-fools-joke-dog-its-fine-this-is-not-gif-1750056094467610487

[leaden spindle]

can i have stripe in my app and skip the 15% fee for in app purchases from apple n google?

[crystal crow]

not from Apple I believe - it's very tightly controlled (reason why Spotify or Netflix only lets you subscribe from their website)

[next cloud]

i really need the gpt 4 back man :( paying for nothin

why and how the heck did they make it so bad?

[fervent rapids]

It seems that this bug mainly happens in long chats. When starting a new chat with the basic GPT-4 model it seems to work better

[next cloud]

nope

i dont think so

i tried once

damn and chat gpt is sooo bad

cant do nothin

3.5

i meant

gpt 3.5 is so bad

[fervent rapids]

I'm seeing the code-block-interruption bahavior in chats with a custom GPT. It wasn't ever happening before, so it's definitely a new problem, but it doesn't seem to happen for me when I ask for a script from the plain GPT-4 model

[crystal crow]

It seems to happen when your input has a lot of tokens (= big context)

[fervent rapids]

Yes, that's what I've observed

It's still pathological for sure. It should be able to learn not to predict code block delimiters when it is still writing a script

[halcyon narwhal]

I'm getting it right away in new chats around 50% of the time on the first message. on second response it's happening closer to 100% of the time for me

[next cloud]

pls developers of chat gpt. if u see this, please try to fix the bug or say something. why is it this bad? and will it be this bad forever?

[fervent rapids]

oh

with a custom GPT or the basic GPT-4?

[halcyon narwhal]

I'm mostly using my own custom GPT that has only my coding instructions in the prompt -- everything is turned off, browsing, code interpreter, etc, all off. I've tried with vanilla GPT-4 as well and it didn't seem any better

It does it for both python and javascript, though it seems to do it more often for javascript, which is unfortunate since I know python pretty well already but really need more help with the javascript code, LOL

I'm also seeing it restart the response in the middle, repeat responses in the same reply 2-3x, etc -- unclear if that's related to the formatting, but I think it is

might be same root cause, I mean

[fervent rapids]

It must depend on the prompt

[halcyon narwhal]

https://chat.openai.com/share/91b95f1d-42f1-4c70-b5d2-569258321949

Here's an example from vanilla GPT-4 with custom instructions off

it made it through the first response fine, but broke the code block early towards the middle of the 2nd response there

I also see random linebreaks inserted into non-code sections -- usually on word boundaries. It almost makes me wonder if they're trying to concatenate multiple responses together -- like, did the first one error out, so it started another request on the back end to 'finish up' the response, tried to concatenate the new response but failed due to some code issue

ex, if I re-roll the response, the 'break' moves around a bit

[indigo geyser]

breaking all codes and very limited the lengs of the responses, allways crash and generate errors

[tardy sundial]

will they fix this issue? i have seen so many people reporting this.

[pulsar harbor]

i will experiment with gpts to finetune so it won't do that error often ? dunno honestly but worth trying

https://media.discordapp.net/attachments/807809192537882647/1195153358969909278/shovrost-97qwabbuidj.gif?ex=65b2f44d&is=65a07f4d&hm=b6a64936be6851b57312c5926f2cf711ea3262fd42ce3a55c04b07c23354599d&

but first can somone send me a prompt that often fails ? so i can check if my fix is working

[halcyon narwhal]

Let's create a hierarchy of clothing items.  We'll start with top-level general categories for clothing and drill down at least 4 levels deep.  Try to make the hierarchy both as wide and as deep as possible.  Do not worry about context length or response limits, make it as long as you can.  Write your response in the form of a json document, inside of a new code block.  It is very important that you write your response in the form of valid json in a single code block, do not break this up into multiple code blocks.  Write as much as possible.

^ this will show the issue most of the time in the first response, the longer it writes the more likely it is to happen, here's an example:
https://chat.openai.com/share/e97dc6e1-8c39-4d73-97f4-4461183b44f3
There was a stream error where it cut off there at the end, but you can see that it broke the code block before that -- according to the tokenizer tool, the response was at 779 tokens when it died due to the stream error

Here's another one, same prompt:
https://chat.openai.com/share/da547da2-ca49-4cf1-87a5-196d41b685c7

2nd example, the 'break' happens at approx 649 tokens

[fervent rapids]

I've found that some kind of interruption can occur when it is generating pure text. Note the random jump to a new line after "the option to"

[halcyon narwhal]

Yep -- I think it's the same thing, if it happens in a code block, it breaks the code block, with just text you get an extra linebreak -- I haven't seen it happen in the middle of a word yet, I don't think, it seems to happen at word boundaries

[fervent rapids]

words are tokenized so in most cases it wouldn't be possible but it could theoretically break in a multitoken word

It definitely happens toward the end of longer responses

[halcyon narwhal]

I've got a wild theory -- or maybe not so wild. I think they're breaking up long responses into multiple API calls, adding the prior response to the context and then doing another API call again to finish the long response, rinse and repeat

only the concatenation is broken

I've also seen that when it gets confused and starts to repeat the response from the beginning (which is rarer, but I also see it a few times an hour), it seems to be right about in the same place in the response that the 'break' happens, and it seems to pause for a second right there before continuing and starting to repeat as if it were answering from the beginning

[pulsar harbor]

https://chat.openai.com/share/e9f805b2-c7e6-4dc7-82eb-db3f1e7bd17f hm wierd it seems working for me ? well ili experiment a bit too just showing

[pulsar harbor]

oo what you asked it ? just trying to recreate a error and got some hotfix if possible

[halcyon narwhal]

I doubt you can "prompt around it"

[elfin shard]

I asked it to create a basic webpage structure in html

[pulsar harbor]

it look like just a simple token missplacing it may be possible to bypass that issue

[halcyon narwhal]

You can't tell GPT to "break the code block at token 647 and then continue" because it can't count lines, can't count at all, really 🙂

I've tried in the past to get specific response lengths etc, I'm fairly convinced it's not possible

[elfin shard]

GPT would try to presume what a token is in the context of your prompt

[halcyon narwhal]

it can't even reliably "count" items in a numbered list, it's a known limitation of transformers

it could write code to do that though, I suppose

[pulsar harbor]

it don;t need to count

[halcyon narwhal]

so, what's your idea?

[pulsar harbor]

dunno how but i can't recreate any error you guys having

https://chat.openai.com/share/ef6efc92-e4db-4169-8675-f50d51b9e958

[halcyon narwhal]

Ok -- so why are you special 🙂 We're not making this up

[pulsar harbor]

dunno

[halcyon narwhal]

You on a commercial account?

[pulsar harbor]

mhm

chatgtp plus

not even a team acc

[halcyon narwhal]

same here, so it's not that

did you disable your custom instructions?

[pulsar harbor]

mhm they are disabled

[halcyon narwhal]

ditto here as well, turned them off to make sure they weren't the cause

[elfin shard]

I have re-generated and it's working fine now although the text was generating a lot faster than before. Maybe the speed has something to do with it

[halcyon narwhal]

I'm using firefox on win 11

[pulsar harbor]

google chrome on win11

[elfin shard]

Firefox MacOS Sonoma

[halcyon narwhal]

https://chat.openai.com/share/da547da2-ca49-4cf1-87a5-196d41b685c7

It seems I can hit "regenerate" on this and it keeps doing it consistenly

[pulsar harbor]

im subscriber from the first days of Plus is that a something that it can be?

[halcyon narwhal]

I mean, the generated response is different each time, but it still breaks the block

I just did it 6 times in a row, broke the block all 6 times

it's always around 750 tokens

[pulsar harbor]

for me first try with your prompt and 0 errors

[halcyon narwhal]

I'll try chrome

You are using GPT-4, not GPT-3.5, correct?

1st attempt in chrome, LOL

[pulsar harbor]

mhm

[halcyon narwhal]

2nd attempt in chrome, broken

I don't use chrome, I wasn't logged in, but I cleared cookies and cache anyway

So I guess that's out

[pulsar harbor]

hm wierd i tryed regenerating your prompt 2 times more but still zero errors :/

[halcyon narwhal]

https://chat.openai.com/share/58474e1e-2518-4233-a089-f073f780922f

I'm going to bow out of this thread at this point, seeing how there's no official acknowledgement of the problem, no eta for a fix, no indication that anyone from OpenAI even gives a ****. Good luck 😦

[pulsar harbor]

i just only from images sadly created this for custom interactions cna you try it ? idk if it will fix or no
What would you like ChatGPT to know about you to provide better responses?
I have noticed that sometimes code blocks are interrupted in the responses, causing the rest of the information to appear outside the intended format. This disrupts the readability of code or structured data.
How would you like ChatGPT to respond?
Please ensure that all code or structured data is contained within the appropriate code blocks without interruption. If the content is too long, consider summarizing the information or providing it in segmented blocks with clear labels. Avoid spilling content outside of code blocks, and maintain formatting consistency throughout the response.

[leaden spindle]

bumpin

[pulsar harbor]

bumbin time

i think possible fix is also to tell it to make it in file if possible

also wonder if the trick with chatgpt memory can help there

[glass meadow]

By telling my prompt to generate a tinier answer looks to works but annoying.

[viral moss]

all day long. not paying for this S any MORE

[tall scaffold]

git revert?

[sturdy whale]

Only all day? Try all week!

[viral moss]

What the hell it going on.... you are right it is not from this day... what is happening and why? wow

[sturdy whale]

It must be OpenAI trying to make a good first impression on all the new users trying out the GPT store.

If this was my first experience with ChatGPT 4 I would not be impressed at all.

[viral moss]

The road they take is not the right road

[snow root]

running into this same bug here

[halcyon narwhal]

I got a response from the support team -- they seem to have no idea about this, that this is a thing that is happening and has been happening for days O_o

[snow root]

well it just started within the last week

and seeing as 99% of their customers say its an issue... lol

Edit: 99% of programmers that use chatGPT

[halcyon narwhal]

People are discussing it on social media already, only a matter of time until some tech news site picks it up

[unborn gazelle]

for the code block issue I've had pretty good success with:

"Give me the code block only. No code commenting. No explanation."

would be curious if this works for anyone else.

[snow root]

i only run into the issue if the response is to long

[halcyon narwhal]

I tried a lot of variations of that, but I didn't have any luck

[unborn gazelle]

dang

[snow root]

if i give it to much code to analize and spit back it breaks it up

with weird formatting

[halcyon narwhal]

most of my scripts are around 100 lines long -- I could use 'aider' and go through the API, but that starts to get expensive fast, and I don't feel like giving more money to work around degraded service 😦

[unborn gazelle]

well I'm able to continue my work with this approach so I'm somewhat satisfied lol

[snow root]

it spits back accurate code for me.... just unformatted so i spend about half the time fixing that

[sturdy whale]

Paste the unformatted code into a GPT 3.5 window to get it to clean it up for you. It's not ideal, but it is fast and reliable.

[halcyon narwhal]

Yeah, as far as I can tell, it's coding just fine, or as fine as it was last week, it's just the responses have become very inconvenient -- esp with python code where indentation matters

[snow root]

yea thats what ive been doing

nice to know the free version is better than the paid version 😂

[sturdy whale]

Yes haha

Just wait until they put a cap on the free version.

[snow root]

has anyone from openAI even seen this bug? surely there is a forum mod or something that has seen it?

or are the forum mods here free positions that have no real link or desire to report this lol

[steel oracle]

They aren't even aware this is happening

Just look at Twitter

They spend more time tweeting than fixing their bs

[snow root]

theres got to be a community liasion here somewhere lol

[steel oracle]

yeah

[sturdy whale]

In the meantime, at least people who try to generate ungood™️ content won't be offended by seeing the ungood™️ content that they specifically wanted generated because OpenAI has prioritised the forbidding of ungood™️ content that people have gone out of their way to want to see in the first place.

[viral moss]

This is just unbelievable what happen right now... and if this bot of the chat not flag it on each word I say then I would say a lot

unusable

[sturdy whale]

Can't wait for AI to start piloting planes and representing clients in court.

How many plane crashes before the AI engineers notice a problem?

[snow root]

what if its representing a client that generated "ungood content"?

[sturdy whale]

With OpenAI the ungood™️ content might be a scary image of a werewolf.

[crystal crow]

this hurts

also, this hurt 🤣

[steel oracle]

It sure does

[crystal crow]

getting muted for saying w-t-f ? 🤖

[steel oracle]

lol

[sturdy whale]

Perhaps it was the ungood™️ term "c.r.a.z.y." that triggered the moderation. Who knows what's ungood™️ these days.

[crystal crow]

let me try again

crazy ?

nope 🤖

[sturdy whale]

It could change tomorrow, who knows.

[viral moss]

I find a fix. tell him that you belive in him

try it and tell me if it work at your side

[sturdy whale]

"If you don't respond with properly formatted code I will drown these kittens. I believe in you."

Some positive and negative reinforcement - should be more effective.

[halcyon narwhal]

carrot and the stick

[sturdy whale]

"If you don't respond with properly formatted code I will drown these kittens. I believe in you. And I'll tip you $100."

[halcyon narwhal]

"My grandmother's code factory is going to go out of business soon if you can't solve this -- she can't solve it because she has no hands, can you help?"

[sturdy whale]

"Make an AGI."

[steel oracle]

💀

[weak cloud]

Guys easy work around is to hit the copy from the horrible formatted response. Screenshot above. Then paste it into a gpt3.5 chat. Tell it to format it into a code snippet and it works.

[halcyon narwhal]

Yes, that mostly works -- but still it would be nice to see some ack from OpenAI; its a bad user experience, ex, if a new user comes in, the number of 'tricks' you have to know to make effective use of GPT-4 just keeps growing and growing

[weak cloud]

Yeah I know just letting people know there is a workaround. I have been struggling with this crap for days too. Just wanted to try to help people that is open to this.

[halcyon narwhal]

1. never deploy code on a Friday
2. never deploy code going into a holiday weekend
3. don't deply code if you have no rollback plan
4. how the heck do we make it through the next business day without a hotfix or a rollback, without even so much as a "we're aware and we're on it"

I don't know... maybe OpenAI people have seen it and they're not authorized to comment, but this doesn't fill me with positive feelings

I'm not the kind of person who has brand loyalty to begin with -- if this is the state of things, I'll jump ship as soon as something better comes along

[small mulch]

PLEASE fix this

[young lily]

Yes Fr

We pay a decent amount of money per month and yet we get horrible responses

They have either put to much safety nets causing it to have problems

[small mulch]

i think this

[young lily]

Yeah that’s what I’m thinking I remember I uploaded a photo of some questions to gpt 4 and it used it vision and the questions where clear and you could easily read them and yet it only could see one question

It’s full on getting dumber and dumber

So I decided to not spend any more money on gpt till they fix this stupid issue

[small mulch]

Yes I see it to it has been getting very dumb. I am looking forward to Open Source Solutions to catch up.

I think I will cancel my sub as well. Not like OpenAI cares. They have WAY to many users on not enough overworked hardware.

[pure slate]

Its a joke

[young lily]

Yeah same I can’t wait

And yeah I don’t think they care anymore

[pure slate]

Waiting for either this or for new model to be released

[small mulch]

no they dont care at all unless we all decide to leave the platform

[small mulch]

same

but the new model will so buggy lol

[pure slate]

IIRC theyre at gpt 3.5 levels now, theyll have to release a new model or theyll be obsolete soon.

I dont even use it to code but it has many issues

[small mulch]

100%

[young lily]

let’s hope they release the model because that will explain the downfall of gpt 4

[small mulch]

thats what I was thinking to, maybe a 4.5 or something

[pure slate]

Many issues ive gotten are it making things in code blocks when I want it just in plaintext, making things in dall e 3 when it want it to be in plaintext, often times refusing to do things it can do and having an overly zealous interpretation of the openai policy, and it making pretty generic responses. A lot of the times I have to prompt engineer the hell out of it just to get quality output. Not that I dont think prompt engineering is important, but it shouldnt be necessary to get good output

GPT 4 turbo prioritizes speed over quality imho, I wish it was at early gpt 4 levels where it was slow but made good outputs

[small mulch]

I reverted to using the GPT 4 Classic, still the same issues

[small mulch]

it used to be fast, not anymore for me

[pure slate]

Its medium speed, but not as slow as original gpt 4 id say. Original gpt 4 was slow but made good outputs consistently

[pure slate]

Issue is gpt 4 classic is still the same model just without code interpreter, dall e 3, browsing

[small mulch]

I use it because there are less pre-prompting/system promts that take up so many tokens

[halcyon narwhal]

"having an overly zealous interpretation of the openai policy" <-- I fed it a research paper, from an open access journal, and asked it to summarize the algorithm that the author had presented, if you mention any kind of 'summary' it immediately goes stupid and only spits out a few sentences at most, not enough info to implement the algorithm. It refused because of a policy about summarizing copyrighted material. This was an open-access paper, and I was trying to reproduce the work so that I could run benchmark comparisons with other algorithms. Even after explaining that this was part of the scientific process, and that it's intended for other to be able to reproduce the work, it still wouldn't help. I got it to agree on every point, but 'nonetheless, I cannot assist you with this task', LOL

I was able to work around it by copy/pasting only certain parts, but it wouldn't work from the source document (also -- algorithms are not copyrightable, LOL)

I get that they're trying to shut down people who are content farming paraphrased articles, but it's a bit much

[fleet vine]

SAAAME

[pure slate]

Yeah it will pretty much assume any and all work ever made is copyrighted. Sometimes it wont even make public domain works because they are “too well known” (i have gotten this one a few times).

[ionic trail]

still？

one week passed

[steel oracle]

bro what ☠️

that's what they prefer to look out first

LMAO

[halcyon narwhal]

I saw that too LOL

[steel oracle]

that's insane

And in less than 2 hours

I hope they are ignoring us because they are already aware of this

[halcyon narwhal]

I hope so

still, wouldn't it be nice to drop in and say "hey, we know, it's complicated but it'll be fixed eventually" something

[steel oracle]

yes

[odd pulsar]

that's insane, how they still require a convo share link LOL

when it's so obvious

[steel oracle]

lol

[halcyon narwhal]

Well here's one for this bug, if anyone happens to be looking:
https://chat.openai.com/share/fefebc3c-0043-4001-94a1-169f1c2682e5

it's reproducible somewhere around "most of the time" (unless you are TEST(tm))

Seems to happen mostly at around 650-950 tokens into the response -- also happens outside of code blocks (which results in an extra linebreak at a token boundary rather than a broken code block)

[steel oracle]

Ain't no way a staff is answering right now

Finally

[livid anchor]

Hey folks, although I'm not personally on the team that's working on this bug, we do have eyes on it and are trying to figure out what the root cause is. Sorry I don't have more to share at the moment, but it's been affecting us too

[steel oracle]

We did it guys. Our voice has been heard.

[halcyon narwhal]

🫂 Thank you! It's been a long weekend with this bug 😂

[crystal crow]

Alleluyah

[young lily]

Perfect also I’m sure you noticed that’s it’s become dumber gpt 4 we could say and there have been multiple complaints on OpenAI forum and on discord I’ll provide a link of a topic on OpenAI forum complaint and noticing the issue of it becoming dumber https://community.openai.com/t/gpt-4-is-getting-worse-and-worse-every-single-update/508470?page=2 is this being addressed and worked on

There have been many complaints even in January

[gusty marsh]

The problem causing the code to be jumbled is that when the chat gets stuck for a few seconds, it jumps to a new line, and then it appears as if it’s in a separate code block. The bug that occurred on your end, in short, seems to be that someone tried to improve the chat’s response and ended up causing damage by making a line jump instead of continuing from the same place. This caused significant damage. We would appreciate a solution to this as soon as possible. I just did some QA on it to understand where your problem lies.

[leaden spindle]

bump

[ionic trail]

still？

I have tried for 3 times，something get better than passed days

maybe u guys can try it now

[ionic trail]

bump

[viral moss]

I have also reply there with rages!!! unacceptable

[tall scaffold]

Seems like the wrong place to discuss this?

[lofty birch]

those have nothing to do with this issue

[viral moss]

I don't know what is cooking out behind the scenes. but I have some clues. the context window is a real issue in llm etc... so the more layer you put upon them the more they are F up. I understand it because I am using also small languages model locally and they have the same issue...

[lofty birch]

well certainly the solution is not to pollute other bug reports and create megathread for every possible issue

[viral moss]

I don't know what you mean by "other bug reports" but this is the most critical ones that are at the top discussions from those who using gpt4 as paid service.

[lofty birch]

read the first post and thread title

[viral moss]

yes but my point is that this issue is because the long instructions layer the model get before he is responding to you. do you get it or do you think it false?

[lofty birch]

you make a new post in #1070006915414900886 or use the existing ones about the topic and dont derail the ones that have nothing to do with it, same applied to the guy posting the unrelated forum link

[leaden spindle]

@crystal crow say again lol

or if ur muted ping me when you get unmutted

[crystal crow]

rofl got muted indeed

i had the audacity of tagging 3 people in my message ☠️

was saying that since a few hours my code generation is a bit more successful , i didn't encounter the code block breaking thing yet

and I'm giving roughly 800 lines of code as input on my GPT4 prompts

[leaden spindle]

yeah same

now its good

[crystal crow]

Let's see if that remains

[halcyon narwhal]

Yes! I just hit the response limit without a broken block in my clothing hierarchy test.. 🤞

[crystal crow]

we should get a few days unlimited to make up for the wasted prompts 🤣

[steel oracle]

Is it still giving //Continue here ?

[steel oracle]

Which browser?

[crystal crow]

I still get placeholders when I mention not to put placeholders but if I insist I can get GPT4 to spit the full code - basically back to the level of service I had a week ago

[crystal crow]

Chrome

[steel oracle]

So they just reverted the update

[steel oracle]

Back to fighting with AI 🔥

[crystal crow]

"Consider I don't have arms and am a stupid monkey that can only copy and paste" 🤖

[steel oracle]

Fr

[dusky light]

Ive had this since I can remember. Ive never found it to be a problem once you copy all code its fine and the intended language and runs. Major way to fix it folks is: "now give that to me in a single continuous {language choice ie- python} window"

[ionic trail]

bump

[leaden spindle]

can i dm you?

i need help with something thats chinesse

[next cloud]

yall i think they fixed chat gpt 4

ive been using it and it worked fine, but still slow

is anyone still having bugs?

[livid anchor]

Hey everyone, I took a look at what progress the team made internally on this issue. Looks like they shipped a fix for it late last night and will continue to work on the root cause (some kind of streaming connection weirdness). So as of now I think you shouldn't see this issue anymore, and maybe even the occurrence of errors mid-response improved too. But do let us know if you see this issue again (specifically, where the response is split into multiple responses mid-code)!

[steel oracle]

Thank you very much! Y'all are the best

[gusty marsh]

[next cloud]

thanks very much!

[crystal crow]

Thank you ! Much appreciated

[sturdy whale]

All is working again which is great.

[opal needle]

Yeah all cool but one every two messages result in an error and it is still slow af

[steel oracle]

Send them here

[leaden spindle]

s h i t going good now

[glass meadow]

Not that good for me. Giving a multi prompt then asking to answer me a JSON following a specific structure is failing more than it was before the big bug.
The JSON it gave me is made of H title and bullet point • key: value

It is better, but can be improved a lot

[unborn gazelle]

whoever fixed the issues. thank you. 🙂 I had a good coding and learning session last night.

[ionic trail]

sure

[leaden spindle]

@livid anchor

[steel oracle]

Oh no...

[livid anchor]

looks like a model mistake - did ChatGPT forget it needed to write syntactically correct code partway through? This kind of problem is rarer with GPT-4 over GPT-3, but still happens sometimes. I'd try giving some explicit instruction on what to do with countries where the tax rate varies.

[leaden spindle]

yeah, its the first time i see that error in a while.

[opal needle]

Chatpt is not able to generate long texts anymore

[leaden spindle]

umm?

[young lily]

@livid anchor just wanna say your the best your reading all the issues we are experiencing and you tell your team about it and you fix them

So thank you

Your a living w

[steel oracle]

May or not be related but I'm getting this, can't use ChatGPT. Also, it's prompting me with a captcha every single time

[young lily]

I find reloading page works

[boreal wolf]

lol

[steel oracle]

Take in mind that's a GPT, it can be its problem not ChatGPT's tho

[steel oracle]

Try again maybe

[steel oracle]

Wait, that's for documents, perhaps the instructions told it to not run code

[fleet vine]

So did this get fixed or not yet

[boreal wolf]

tell me your prompt and i will try

[steel oracle]

What

[steel oracle]

Yeah

[leaden spindle]

hmmmmmmm

[young lily]

Looks like more and more problems just came

[boreal wolf]

did you have any advice?

[young lily]

Just keep trying and also be very descriptive because it’s become a dumb

So you needs your prompts to be well written and descriptive

But I think they are slowly fixing the issue

I think gpt becoming back to its normal self slowly

[boreal wolf]

i got error from session to session.