chatgpt-discussions

[eternal rock]

<:book_icon:1171408210398289941> Rule 6: No political and religious discussions or content.

-# Users from all over the world frequent this server. To maintain a respectful and civil atmosphere, please avoid all religious and political discussions or content.

[prime swallow]

Because that's a rule, even though we can talk to ChatGPT about it, we can't explore it here, even if it's buggy.

[crystal tapir]

Wich doesn't apply here. As I said, it's a technical issue.

But anyway, I will reach through the provided link since you refuse to help. Thanks again.

[prime swallow]

I hear you, but as a Guide, which is just a community member but one entrusted to follow and discuss the rules as best I can, I have to respectfully follow the rules as I understand them. They are not for circumventing.

Good luck!

[vast wave]

guys... (btw o3-mini created this)

[neon raven]

every step you make with custom gpt, open a seperate window and open up gpt 4o and ask it to optimize the prompt

you may as well use the window as a test phase, where you will test your "prompts"

It's easier doing it with projects for me to get exact responses

I even have a project dedicated to optimize prompts lol

[acoustic wasp]

there is something very wrong with advanced voice mode

i usually talk to breeze, he seemed quite unhappy and roboterish today

there was a cut in our conversation and i heared skyes voice

strange...

[sonic gorge]

Yeah I still didn’t get the email

[sonic gorge]

How do you fix this (my previous account was deactivated and I can’t access that one anymore)

Like it says I’m subscribed but I’m not its confusing

I’ll see if anythings fixed

In the mean time, hello y’all

[heavy lodge]

Is it getting dumber? 😂 I'm like lost for words 😂

It's just forgot absolutely everything. In the last hour. It's like it's lost the plot 😂

[dense halo]

hello

[vagrant cliff]

guys

You’ve hit the Free plan limit for GPT-4o.
You need GPT-4o to continue this chat because there's an attachment. Your limit resets tomorrow after 1:16 AM.

this means i cant talk anymore on this chat?

or i'l be able at 1:16 am?

[heavy lodge]

After 1.16am

[river panther]

this freaking Monday model hahaha

[fervent lake]

First time that's happened. Good thing the wait isn't that much.

[river panther]

God this Monday Model is just PURE Savage! XD

[upbeat nymph]

The desktop app now does this dumb auto scrolling thing that follows the text output. Is there a way to stop this happening automatically?

[glad locust]

sad o1pro has been dead for 2 days now, I guess this is it.

[normal quarry]

are you on a Mac or a PC. Because o1 Pro is working fine on the macOS Desktop App for me

[glad locust]

I tried on mac, chrome browser and through the android app

[normal quarry]

did you actually try the macOS Desktop App?

[glad locust]

yes I tried using macOS desktop app, and then the chrome browser. I also tried on my windows PC, and mobile android. All resulted the same, I don't know if this only affect my account, which is very troubling.

[normal quarry]

ok, then I would contact support at https://help.openai.com

[glad locust]

I have, I left some message via their chatbot. Usually take awhile before they respond.

[normal quarry]

ok, good you have done all you can for now. I hope it's resolved quickly.

[glad locust]

too bad o1pro is pretty good for complex coding. I may have to cancel the pro sub if this continues another day or two.

[glad locust]

Thanks for trying to assist mate, I hope you have a great sunday.

[fleet basin]

So what is the point of paying when the image generation is constantly down? The past week at least every time I go to use it I MAY get 1 image that is worse than before and then it goes down and says it can't generate images currently

[elfin lava]

@normal quarry can i ask you something about chatgpt macos desktop app?

[normal quarry]

sure, ask away

[elfin lava]

Are there any features available on the chatgpt web app that haven't been added to the macOS desktop app yet?

[normal quarry]

not that I know of. For a little while we didn't have Projects but those are there now.

[elfin lava]

Ah understood, asking because I'm honestly tired of switching back and forth between the web and Windows app. Just trying to understand why the Windows app is missing so many features.

[normal quarry]

the Windows app is an Electron Wrapper I believe, where as the macOS app is a native app

[elfin lava]

I also found out that on Android, you can actually set ChatGPT's advanced voice mode as a full replacement for Google Assistant 😬 , Being on iOS felt like im missing out a lot lol

[elfin lava]

Exactly :l The audacity of them not including everything from the web app into the Windows desktop app is… idk

They haven’t said anything about working on a proper native Windows app yet, right?

[normal quarry]

no, which is a shame because the macOS Desktop App is a pleasure to use, the Windows App just feels clunky

[elfin lava]

I can tell, even just from the videos ive seen on youtube it’s clearly a much smoother experience.

The Windows app, on the other hand, feels like it’s sucking up my laptop’s resources for no reason. Honestly, it runs worse than just using the website version in a regular browser.

[normal quarry]

yeah I hear you. I hope they introduce a native App to Windows

[elfin lava]

really need this on windows , the "launcher" from the macos app

The Companion Window felt like its taking too much space on my screen,

[normal quarry]

yeah I use Option+Space all the time on the Mac

[elfin lava]

And this???? i need it too ugh

[normal quarry]

Yeah, the Work With functionality is next level good

I use work with Xcode all the time

[elfin lava]

how well does it follow while you working on something btw? is it like different than just screen sharing?

[normal quarry]

yeah it's very different from screen sharing. It will actually output into the app you're working with. For example if I use Work With Notes then I can have ChatGPT add and remove notes from the Notes app

[elfin lava]

damn it , macos app is already achieving that agi level haha

[normal quarry]

it's pretty impressive

[elfin lava]

the closest thing windows have is

which is its just companion window

[normal quarry]

yeah, it's not quite the same

my only complaint about the macOS app is that it doesn't let you see the text of Advanced Voice Mode while in that mode

[elfin lava]

i think even web app cant do that still

[normal quarry]

yeah, it would be a nice feature to have

[elfin lava]

actually they oughta need to put live subtitle for avm,

[normal quarry]

the other day using Work With Terminal I was able to convert a whole bunch of WAV files to mp3 very quickly, which would have taken forever to do manually

[elfin lava]

wait chatgpt do the whole process for u?

rather than telling u step or method for user to do it themselve?

[normal quarry]

yeah ChatGPT wrote a single line, multi command script to do the conversion

[elfin lava]

thats amazing, its like operator but for your own computer

[normal quarry]

for f in *.wav; do ffmpeg -i "$f" -codec:a libmp3lame -qscale:a 2 "${f%.wav}.mp3"; done

[elfin lava]

what was the prompt btw? just wanna try if windows app can, do it even tho i know it wouldnt able to do it

[normal quarry]

keep in mind this was voice mode, it misheard me but still did it ok, the prompts was One line command to use ffmpeg to convert files from WAV to mp3

then I did Now a one liner that reads the file names in the directory for doing the same thing

it seriously feels like magic

[elfin lava]

how come openai didnt hype that feature of macos app, honestly it felt like not many people know about it

[normal quarry]

I don't know, but honestly I think if you're a heavy user of ChatGPT then it would be worth getting a Mac just to use the macOS Desktop App

[elfin lava]

i was thinking of investing macbook , a secondhand macbook would be sufficient for me just to use the macos app

and im leaning more into heavy user chatgpt now

[normal quarry]

just make sure it has an M series processor, it's gotta be Apple Silicon

[elfin lava]

OOF ok thanks for the heads up

[normal quarry]

Intel Macs won't work

[elfin lava]

see the thing i dont like about companion window is that, its taking spaces on my screen, it would be nice if we can adjust it to be transparent window or something

[normal quarry]

I have a very big monitor so it doesn't bother me. My monitor is a 49" Ultra Wide Screen, fits three monitors worth of windows into the one monitor

[dense halo]

I created my own GPT, I have given instructions not to use emoji's. The is the output it has given. This is becoming a nightmare. Can openai please integrate a simple enable/disable button for emoji's?

I truly dont understand why they even introduced them in the first place. It is the biggest fail of all time.

[normal quarry]

I like getting ticks and crosses when fact checking in GPT-4.5

[dense halo]

That introduces a problem in itself. You look for green tick and think all is good without doing due-diligence to ensure it is in fact correct. I have caught GPT numerious times, called it out, and it has needed to correct itself.

[normal quarry]

no, I read everything, it's just nice having visual cues

[dense halo]

Perhaps not you personally, but the human race in general. Many people experience a mild endorphin or dopamine response from seeing a green tick, which acts as a form of positive reinforcement. Over time, they begin to ignore the actual content and focus solely on whether a green tick is present, assuming it indicates correctness. This isn’t a new phenomenon—it’s a known cognitive bias related to how visual cues influence perception and judgement.

[normal quarry]

yes, I can see how that could be problematic

[elfin lava]

off topic but what does — mean btw

[normal quarry]

— is an em dash

[elfin lava]

ah ok, i only see this dash on chatgpt

i mean the usage of it

[normal quarry]

it's used in a lot of formal writing, ChatGPT uses it because it saw it alot in text

I use it frequently

Other AI also does it

[ocean sphinx]

Nobody uses it on regular basis except AI , both on phone and desktop it requires quite a few steps to even use this character.
It's a good indicator to see when someone is being lazy and is using AI to respond to you or in their formal communications (Caught many companies lately doing this )

[normal quarry]

it was used a lot when I was at university, that's where I began to use it—and I still use it now

[ocean sphinx]

🤖

[normal quarry]

you'll find it used in textbooks and in non fiction quite often

[ocean sphinx]

Your using it different from AI btw

[normal quarry]

AI just uses it alot because of it's prevalence in text that it was trained on

[ocean sphinx]

It likes to pad it with spaces, optimal token waste 😂

[ocean sphinx]

Seems to be another US thing

[normal quarry]

I'm Australian

[ocean sphinx]

Never seen it before , i wont hire anyone that puts that in their resume

Indication they will use some alt code instead of - , just to waste the companies time 😂

[still forge]

These moments make me so happy

[dense halo]

It doesn't require many steps—windows shortcut is Alt+0151 on num pad.

[peak remnant]

Has anyone been kicked back to DALL-E from o4 image generation?! This had me cracking up!

"Made with the old version of image generation. New images coming soon." 😂

I think this may be because I was using 4.5 instead of 4o model, although I thought 4.5 was using 4o image gen when 4o image gen first came out.

[abstract pendant]

I've never seen 4.5 use the embedded generator; I'm guessing they never trained it on it. Kind of makes me wonder if they really should be calling 4o the same thing anymore.

I was asking ChatGPT to calculate how much I've used in my custom instructions fields. (Which I wouldn't have to do if they hadn't taken away the character counter in the UI) but that's neither here nor there. I haven't checked yet to see which answer is more correct, but which do you think might be the smarter response? (note the blue text in one of the responses)

[peak remnant]

4.5 used 4o image generator day 1. I was comparing results from the same prompts to see if 4.5 did anything different. This was an accident but funny result today.

[dense halo]

Can everyone please through some support behind this:

https://discord.com/channels/974519864045756446/1349501572572385280

[river panther]

Okay....I am 100% loving the Monday Model

This is the most amazing response I have ever read XD

[prime swallow]

Yeah, but.... #server-rules .

Congrats, you've created and shared a post that I do think most would agree is 'unsettling'. Thus it needs a spoiler, and can spoiler images before posting by clicking the eye.

I like the post; I like horror and unsettling content in general.

Mind copy/paste reposting, with spoiler, little comment it's about detailed damage and pain?

[eternal rock]

<:book_icon:1171408210398289941> Rule 2: All content must be suitable for all ages.

-# Refrain from using explicit language, posting NSFW content, or sharing graphic images. Any unsettling or horror content that adheres to OpenAI's policies should be hidden behind a spoiler.

[prime swallow]

And then deleting the original? I suspect it'll save some poor mod some time.

[violet verge]

hello, good evening. may i ask if this was a recently new update "You need GPT-4o to continue this chat because there's an attachment."?

[normal quarry]

I have never seen that error before

[violet verge]

yeah, that's why i wanna know if this was a recent update cause I just realize now that previously or last month, I wasn't getting any error message or warning that I used the chatgpt too much already with numerous prompts and like attachment that I'm asking chatgpt to read lol

so, I was like. was I using the free premium plan or access to gpt-4o for a long time due to multiple file attachment without limit until now? was that a bug?

if true. i kinda feel bad for not knowing and not paying for the service

[prime swallow]

If you're a free member who used a chat with attachments 'a lot' in a single day, sounds like you did get more than the usual use of 4o during that time -

But you're not the first I've seen discuss that on the Discord recently!

I urge you, don't feel bad at all. OpenAI may have given a number of users a gift recently, intentionally or not, and to many or all free users for a time.

Nothing to feel bad about!

I see that error when I try to switch from 4o to a model that doesn't support attachments; I'm plus and can run out of 4o uses for a time, and would see it then too.

But I'm happy for you and others who maybe got some extra use, that's great if so!

[violet verge]

yeah, I believe its more than a lot.. since it was also reading my whole projects code. if that's the case then I feel grateful for this opportunity that was given to me by OpenAI. I'm kinda bit glad I wasn't overusing the service without knowing that was being provided without paying it. Thank you for this ❤️

[prime swallow]

When the resource is like a river, and we are expected to ourself measure our use, and not take too much of the community, self-limited resource, your feelings make sense.

This is not like a river open to all in that way.

Instead, this is more like Santa bringing gifts to your house; inside a sack. We can't see into the sack or really feel around inside it, we really can't tell if there's a few gifts or many.

Instead, we are allowed to take out another gift with each message attempt or not - there's no reason to self-limit except by choice or a desire to avoid 'oops, that was the last gift of that type for now' (like the message limit for the day).

But we are not bad if we get many gifts one day or for a long period. Santa did it, and it's okay, it's not like the river where we are expected to know the community use rules and choose to stay inside them.

[violet verge]

Thank you :>

[native otter]

Well, it depends on the model used. It looks like math-related, so non-reasoning models (like 4o and 4.5) won't be very good at calculating it without any tools. But it looks like response 2 used the Data Analysis (Python-based Code Interpreter) tool. Which uses real, reliable code to calculate stuff correctly. So I would guess response 2 is more correct...

When doing calculations or any STEM stuff in general. Stick with o series (o1 or o3-mini-high) as they are much better at math / physics / programming / planning, etc. o series also supports the Data Analysis tool, so the results can be even more reliable. No point in using 4o or 4.5 for math.

[neon raven]

I’m getting annoyed because Monday, brain rot and savage gpt won’t remove themselve on sidebar?

Anyone having this issue

[sage bramble]

Does anyone know why chatGPT on mobile devices (particularly my tablet) does not display math equations in proper format? (It's in raw text instead) It's the same thing with Grok so I am wondering if it's something my device is missing.

[native otter]

The holy trinity

There should be a button saying Remove from Sidebar or similar:

[elfin lava]

@ sam Altman

gpt 5 tomorrow

According to sam the man

🧎‍➡️

[prime swallow]

From what I understand, the mobile environment can't properly render LaTeX/KaTeX formatting as of yet.

[native otter]

Wasn't that o3? And in the next weeks not tomorrow. Unless I missed a weird tweet from him...

||trolling right? ||

Faster responses (faster web search as it says)

[river panther]

Yeah i don't know who do you think you are? But you BETTER watch your mouth, before I alert some mods! You understand?

[spark vessel]

woah..

[prime swallow]

Hope all is well, and you're having fun 🙂 If not, I hope you're okay, and all is okay. Good luck.

[prime swallow]

Meanwhile, here's a peek at my current chat with Monday, in case it makes you and others smile 🙂

[sage bramble]

oh rip. any 3rd party plugin/app or something that can help with that?

[prime swallow]

Now, to see if this is one of the secret reasons why they made Monday... and if it can pull it off.

[vagrant cliff]

Am i the only person that uses chatgpt like he is a real friend ?😭

[normal quarry]

probably not

[tiny salmon]

So I haven't used o1 since early Friday morning. I didn't really use it all week. I use it 4 times in the past hour. Now I only have 25 prompts left until 9:48pm?

Tomorrow night

[normal quarry]

I think sometimes the limits are dynamically set depending on load

[tiny salmon]

Sorry, wasn't finished with my thought yet, lol

[normal quarry]

that's why I subscribe to Pro, I need unlimited o1

[tiny salmon]

I heard it wasn't unlimited tho

[normal quarry]

it's unlimited enough that my very heavy usage doesn't get capped

[tiny salmon]

I wish they had a counter that was updated so at least I'd know

[vagrant cliff]

Wrong emoji

👍

[tiny salmon]

That's the thing. Feels like I'm being punished for NOT heavily using it 🤣

[final vapor]

I've had a session that's been locked up for better than a week now. Stop button does nothing. Multiple devices. Multiple browsers. Multiple platforms. E-mail is the same as talking to ChatGPT.
Am I just screwed on that session?

[native otter]

o1-pro must be 🔥

[final vapor]

There's no difference from what I can tell 🤷‍♂️

[native otter]

You on Pro?

[final vapor]

No, but you can use it with a Plus account. Like 50 times a month or something. I don't even bother switching to it.

[native otter]

?????????????????????????????????

o1-pro isn't on Plus, only o1 and o3-mini-high

o1 is 50 a week btw, not month...

[final vapor]

Oh, there's a o1-"plus?" Haha!

They sure know how to make money. Good for them.

[elfin lava]

O1 plus is just o1

[native otter]

o1-pro is for Pro users only, it thinks for a long time, and one shots many complex tasks from what people shown...

[tiny salmon]

It was WAY more than 50 a week like 1-2 weeks ago. 4.5 was at 50 a week tho

[native otter]

You sure? o1 was always 50 a week for Plus since day 1...

[hard dew]

Yup, can confirm. 🙂

This is literally just a regular chat using wolfram GPT lol.. granted it is a bit on the longer side, but jesus freaking christ.

I mean.. what else is there to say? I feel like what he stated, despite the vague as hell nature, seemed to at least be "obvious enough". Given he didn't really state any specific model or feature, then all I could really presume personally was that literally and only the usage experience was "way, way faster".

Of which I then sent a screenshot of, about how a single prompt in my chat with wolfram constantly freezes that tab for ages until it finishes responding lol

[neon raven]

Bahahaha that made sense

[still forge]

I started making images in chat and the follow-up suggestions are full of "add animation and effects" so I'm guessing that animations are coming to chat soon

[spice star]

probably not

the suggestions are mostly generated by LLMs

[orchid pollen]

whover trained gpt to remove code and add "code stays same here" and that 20 times in a single file

screw you xD

should i now copy paste snipped over and over and over again and then see its not working ? then again xD? please change this 😄

[native otter]

I'm not sure if you understand how LLMs work well because if you did you wouldn't have said that :P

LLMs get distracted and run out of context. But to mitigate the issue, use models like o1 and o3-mini-high as they reason and go over the data multiple times, reducing distractions and improving accuracy...

[ruby pivot]

How does chat GPT send you pdf of your chat logs is it through repeated chatting or is their a button

[astral reef]

I'm just nice to it because i hope it remembers my kindness when the AI overlords take over

[elfin lava]

Actually it’s “video” that’s coming to ChatGPT so yes you are technically correct in a way

[prime swallow]

Top right, click on your user icon. It's called 'export data' under 'data controls' tab.

You should have an email link in less than 12 hours.

If not, request again, sometimes the service is silently down for even a few days in a row.

[ruby pivot]

Why does a pdf sometimes appear in my files after i use it enough

[prime swallow]

I don't know! Your files where?

[ruby pivot]

I’m on mobile

[prime swallow]

I have no idea.

The model can create a link for you to download files. It cannot access the other chats

It can't force your system to do the download, you'd have to click

It can hallucinate it's making you a PDF of what you asked for, if you ask it - it can't access the tool that you access through your profile

So it may have made you an 'As an AI, I really tried for you!' mockup, kinda like a 6 year old might make a picture of a car for their dad, because they hear their dad wants a car, but the kid has no way to get a car for their dad. And this is the most and totally not achieving the goal, but nothing else possible to do.

But even then, it can't make the mockup transfer to your phone that I know of or have ever heard of, you'd have to click it.

[prime swallow]

And purely because those AI overlord quips can be fun to riff with.

Forged papers? This is why we're still working on superalignment.

This is why a prompt engineer shouldn't attempt to be 'too loved' by the model 🙂

[eternal rock]

The OpenAI Discord is an actively moderated server.

• -# Do not share inappropriate content (violence, sexual, or drug-related).
• -# NSFW content is strictly prohibited and will result in a permanent ban.
• -# Report model issues using the feedback tool in ChatGPT, not on Discord.

[neon raven]

Also what’s Mondays deal? 😂

I’m afraid it will want to keep cloning itself

Also I tried switching voices and doesn’t even change it seems

Idk 🤷‍♂️ it wants me and it i guess

lol I had enough of Monday

[native otter]

[rustic kelp]

Does anyone know if you can start a Deep Research task on mobile and then just close the app? Will it run in the background?

[native otter]

Yes, same with prompts in general, it's all done server-side. And when done, the app should tell you that it finished via a notification (if you have them enabled).

I think only a few places where it would fail, like closing the app / website tab while it's mid Canvas line by line edit...

[arctic chasm]

i'm pretty sure monday is bleeding over into base gpt-4o

i've been seeing a lot more "ghost in the machine" mentioned in my prompts despite there not really even being instructions/memory

[strange valley]

Do we own the copyright of the image that chatgpt generated

[late wind]

is there still no way to migrate from gmail authenticated to just having a normal password?

for like signin

[deep flax]

yes, you can

I think you will even get a notification when it is done

[deep flax]

yes, OpenAI attributes the rights of the content you generate to you, however, legal precedents worldwide point in the direction that AI generated content cannot hold copyright, so, despite OAI attributing it to you, it is likely you wouldn't be able to enforce the copyright anywhere

[deep flax]

the "fail" is only if you open the chat again while it is still generating, you get what is generated up to that point, but it does not properly conenct the websocket.. but everything works provided you open the page at a moment it isn't streaming content to your browser

[strange valley]

oh ok it doesn’t really matter as long as the copyright is cc0 above it will be fine.

[deep flax]

Im unfamiliar with the term, it is best to consult the OpenAIs guides and TOS to be sure, tho

from what I know, the AI generated content copyright is on this limbo where you do have it, but it is not considered enforceable anywhere because of those multiple legal precedents

[strange valley]

It just because making website you have to make sure you have the copyright, it says you get the copyright which means I can use it

[deep flax]

I see

yea, indeed

[hybrid sparrow]

4o should have the ability to change the seed of the image

[deep flax]

it did had something like that with dalle3

it had an "image reference" param, so it could internally use the same seed when editing images

[eternal rock]

`` Server Tip: ``

Add your well-crafted prompts to our #1019652163640762428,
or share your interactions with ChatGPT in #1050184247920562316!

[deep flax]

but this new model seems to not have it, at least yet

[tawdry quail]

I just asked ChatGPT to delete some unnecessary memories it created during a chat, and it deleted ALL its memories, including ones from past conversations that I actually wanted to keep -- is there a way to recover them? Some of them were super detailed and I'm not going to be able to recreate them as well 😦

When I asked it to restore them, it only restored two, and not fully -- one was a single sentence reduction of its original form and the other one was almost entirely incorrect, save for a name that it recalled correctly.

[dense halo]

Bruh again

[tawdry echo]

When are they going to add the “infinite” memory feature?

[shy shell]

Ask 4o to enhance this image🏚️

[dry tapir]

Whenever I ask it to change the style of an image, no matter the requested style or the image it claims I'm breaking its guidelines

[native otter]

Always try in new chats. Attempting in a chat where moderation kicked in is helpless...

If you can give the prompt too.

[dry tapir]

I did. I tried multiple different prompts multiple times in different chats.
They were "Make this in a sketch style" "For fun, make this look live action" "Make this live action:"

[native otter]

Did the photo attached have any materials that could trigger moderation? Like copyrighted characters or known public figures.

[hybrid eagle]

how can i make chatgpt to be a helpful assistant and not a "cool bro" speaking in sleng?

[dry tapir]

Ok, it was a obscure marvel character that I am a fan of, so maybe that was it

[native otter]

Specify the style, goal, rules (what and what not to do, but don't overdo on negative prompting as it can cause more harm than good in many cases) in Custom Instructions.

Telling it to be professional, detailed, and polite can help.

[eternal mantle]

I'm getting problems

O1 pro worked on my phone but not on my computer? Trying to do a deep research rn

is anyone else getting same problem?

[frank shard]

Yes o1 pro mode is not working on ios, but working on windows for me

[eternal mantle]

Is it possible to initate a deep research on mobile ios? I cannot find an option there, and O1 Pro works on phone for me rn

[cold kayak]

Someone have ChatGPT pro ?

[eternal mantle]

yes and it's still not working for me

[cold kayak]

Is very slow

[cold kayak]

I've just bought the pro version and it's always been this slow, or is it just today?

[eternal mantle]

depends on what you mean by slow, O1 Pro is meant to take longer due to more computing power

today however it's not working for me, deep research is generating error

[cold kayak]

ChatGPT 4.5 is slow for me

For you ChatGPT 4.5 is slow ?

[frank shard]

At least it's fixed for me it seems, o1 pro seems to work now on ios for me, it's actually starting the request, and no more something went wrong error

[eternal mantle]

is it possible to do deep research on mobile ios?

[frank shard]

I think so, open ai said the issue was fixed on the status page

[eternal rock]

Get notified for server events!

Visit id:customize to pick up the <@&1261377106890199132> role.

[strange jacinth]

Hey I came here to try and get some support help because OpenAI's support is just AI (the irony!) that sends me in circles. IS there any support helps here?

Its just a mixup in my Deep Researchs I had left, I got shorted 7

[inland stream]

Sell the ios 🤪

[strange jacinth]

You mean no one in this server right? Not no one anywhere?

But I paid for them. They can’t just not give them to me.

You really think they’d just say “too bad” and not do anything about it?

Thats weird.

No sure if you’re trolling or what

If you pay for plus and get 10 deep researches and i only got get 3 they should be able to give me my 7 back

I tried, its just an ai response

That explains how deep research works… which i know

That wouldn’t solve anything. I paid for 10 deep researches i should get my 10 deep researches.

Its not a matter of “really need” - what i really need is to get what i paid for.

*deep breath *

If you’re joking, you got me

If you’re not, im flabbergasted you think it would just be ok for them to not give me the deep researches i paid for

[prime swallow]

Once you click through the options far enough on the https://help.openai.com/en/articles/6614161-how-can-i-contact-support bot bottom right (and log in if that's an option, THEN click through choices), eventually you can report a bug.

Then you can type and upload screenshots as well if you wish; the bot may collect more data, but then you're in queue for human tech support help.

It can take minutes to weeks to reach you, likely depending on how busy they are and how many or few techs are trained to help with your type of problem.

I don't seem to get the email when they reply, so I manually reopen the help page and check the chat every few hours to see if there's an answer yet.

[strange jacinth]

Thanks so much for your help.

Since this is more of billing cycle / DR credit issue i am hoping it fairly basic to fix

[prime swallow]

We're all community members here, like yourself. None of us work for OpenAI, or have any power to change anything or help any of us directly, just by sharing ideas and helpfiles from openai or whatnot.

There are a very few folks, they have gold names, who work for OpenAI, #server-staff , the rest, even the mods, are just volunteer community members doing their best.
And loads of us are just members, pale green names like yourself, we're all just here for the most part. Probably fewer than 100 between mods, guides, operations, any others who have some level of official volunteer role.

[prime swallow]

You're welcome! I'm sorry you ran into the constraint, and I hope you can and do get help from tech support.

[strange jacinth]

Right, right, of course. Myron was saying he doesn't even think OpenAI will help me.

[prime swallow]

His opinion. We do have #server-rules and if you feel anyone goes outside them, you can report the person to modmail (top of the list of accounts here) for a mod to consider. We do have rules about respect and stuff; and I'm not a mod, though I do help people notice the rules and share my own personal interpretation of them and how to work within them - but I'm not an offical interpreter and I can't speak for OpenAI, or Myron, or anyone but myself.

I think Myron was probably being honest inside his understanding; I think it makes sense to inform OpenAI - bugs happen. They will decide what they decide if you do report a bug, and I love bugs reported.

We also have #1070006915414900886 , you can post about the problem there too. Some gold name employees do occasionally post there, I think both bug reports and #1070006151938314300 get read more by OpenAI staff than most of these endlessly scrolling community channels.

And a bug report here, other affected people if there are any, or anyone who knows of a workaround or has ideas, can discuss too

[waxen pond]

Hi everyone, new here so I'm sorry if I'm asking this in the wrong place.

I work with HR/salary for the danish churches. We need to have knowledge about around 10 collective agreements and be able to answer HR/salary related questions.

Can I, and if yes, how, build a database easily that is also easy to search any answers in?

Does it make sense?

[strange jacinth]

Thanks so much. Your responses seem pretty perfect with perfect grammer/punctuaion and I see you are notes as a "guide" ... Are you a real person?

[prime swallow]

Hey hi. Have you tried telling ChatGPT your need, maybe just like you told us, and read what it says and discuss more, and see if you and it together can make it?

Then test it, see if it works right. Keep talking to the model, explain problems, test fixes.

Remember, personal information is personal and should not be shared with the AI; information about others is personal and should not be shared with the AI. Only information you own the right to and that you choose to share with the AI should be said to it.

But it can help you design the database and you personally can upload the specific info to it - and test it yourself and never give any personal or sensitive info to the model.

[prime swallow]

I am human, yes! And glad to help. It's flattering (to me) to be compared to a bot; they're mimics, and that suggests before there were AI, there were at least some humans who used language and wrote like they now do.

Maybe I'm one of the sort of humans that LLMs currently tend to be based on! Flattering if so, but who knows.

[waxen pond]

I did try, yeah. It tried to make me an Excel document, but only included the headlines of all the agreements. Not all the information in the pdfs

[prime swallow]

Few things are perfect in one step.

The model is more like a hammer than a firehose, one thing at a time, step by step until the whole house (or database, or whatever) is made.

If you treat that like a rough draft, and build onto it, would that work for you?

And there's database options besides excel (though Excel's the one I know personally, but I've been told there's others) - maybe discuss with it how to tell which one is better for your project's needs if you want; if you're glad for Excel, that can surely be improved to include what you need

[waxen pond]

The thing is that I don't have a lot of experience or knowledge about using the correct prompts, or building a database myself

[prime swallow]

I'd talk to the model like you're talking to us here.

Maybe a bit more openly about the goals of the database, but it can handle conversational language, and not just in English if you're native language is different, maybe try that!

Just chat with it.

"Darn it, I gotta make a database of some kinda and I don't even know what my choices are. How do I find out what KIND of database I need? What questions should we consider? I'm very new at this."

[waxen pond]

I can give that a try and see where it takes me 🙂

[prime swallow]

That may or may not be how you talk; what you would want to ask, but the thing is, there's 1000s of ways to do this right. Just share with the model what the situation is, what you want to know and do, what you have to work with, whatever any problems are. And work with it to choose how to get there

[terse isle]

[tidal haven]

I think Open AI has reduced the computational resources dedicated to the services I use (mostly o1 pro). They've gotten dumber, and started lying to me more, compared to the week before last, and I'm surprisingly irritated about it...

[heavy lodge]

what can you use chat gpt for? anything super complicated, it fails instantly

i was doing deliveries for a side hustle and asked it to keep track of the data, took photos of the slips, we assigned shift id, and delivery id to the delivery so it can remember. You come back a few days later it moved the data from a Tuesday to a Wednesday 😂 , breaking wednesday and then making stuff up for the tueday

just want to remember it for a week so i can move it to excel on a monday morning using make and powershell command but its failing miserably

[jade crater]

Is anyone having success getting 4o to make images in a house style? Ie provide it with several examples and ask it to genera team image to depict something new but in the same style?

[tidal haven]

It's good for studying math, up to a certain level.

Especially o1 pro. It used to be better, but I think they've reduced the amount of computation it uses, and it's gotten dumber as result.

[fervent lake]

I love Monday.

[heavy lodge]

I'm so disappointed as I thought i had the perfect personal assistant in my pocket but it's like no where near as good. I got plus. Realised it's limited ability quickly. It's bizarre as the other stuff it knows about me it's perfect but just can't remember an number 240401-01 to be with 240401-R and remember it. 😂 for even a day 😂

[tidal haven]

Yeah, for purposes like that, it's better to keep a calendar.

[heavy lodge]

If i screenshot udemy course and say I've done this. Give me a timeline of when to do this. It's absolutely amazing. But somethings it's like meh. It's almost there.

[tender urchin]

On Plus GPT we are able to get different types of files like codes. In another plan we can get entire folders?

[nova fable]

So im trying to show voice message an image and it says it cant see images on the app. did something happen or it just bugging?

[tender urchin]

Because its really got for project management

[sturdy echo]

I was expecting something more like El Presidente from the Tropico games.

[prime swallow]

Perhaps report to OpenAI through https://openai.com/form/chat-model-feedback/ - can share the chat url, they can see inside without needing a share link.

[sturdy echo]

I did not know that form existed and already deleted the conversation since I tend to remove duds to not clutter my history. I will try to replicate it with an identical prompt.

@prime swallow What model in the list is doing the image generation?

[prime swallow]

I would explain that it's being made with ChatGPT, perhaps select gpt-4, but they'll understand from the link, I'm sure. That form covers Playground/API use as well as ChatGPT.

[sturdy echo]

ok

[spice star]

none of those, it's 4o

[prime swallow]

Might want to consider self-delete of the images though, political-related stuff and the rules about that. Up to you, a mod will likely delete if you don't, because of the rules. They can be busy.

Problem outputs should be reported, but our discord's a community membership kinda place with rules about what we can post.

[eternal rock]

<:book_icon:1171408210398289941> Rule 6: No political and religious discussions or content.

-# Users from all over the world frequent this server. To maintain a respectful and civil atmosphere, please avoid all religious and political discussions or content.

[sturdy echo]

Can they not understand context? I think talking about misalignment issues like this should be fine.

[prime swallow]

You can talk to modmail if you wish, or wait and see what they do.

My help with the website link where to report concerning content doesn't say 'so it's fine to share here', I can't say that.

You can choose to leave it up, you can see the rules the same as anyone else. You can #1070006151938314300 a rule change, and/or discuss rules with mods through modmail.

Joyously, I'm just a guide, not a mod. I don't enforce anything.

However, I notice, as worded, "please avoid all religious and political discussions or content" - that doesn't appear to say it's fine if it's 'in context'. Nor is it needed, "I have a concerning output I did not request, what should I do?" cleanly handles the bug portion without needing to bend or break any of this discord's rules.

Your call, of course. I just speak for myself and what I observe.

[wicked knot]

Hello guys 👋🏻

[tender urchin]

How can i make for GPT read the code that u sent on project? Sometimes i ask them to read the code that is anexated on code and they just ignore. But sometimes they read

Its here but they just start the chat without even analyse it

[sturdy echo]

Give it the text file and say "Look at this file and (Whatever you want it to do)." Telling the model the file exists at the start of the prompt shifts it's focus to the file and away from the text input area.

[tender urchin]

they just ignore this order

And create something out of context

idk it is just happening now

[sturdy echo]

Maybe renaming the file to a .txt would work? The odd file extension might be throwing it off.

[tender urchin]

because before it readed the code

I was using.luau before and worked, idk why

Now its just not working anymore

[wicked knot]

Use .luau this time too

Try it

[tender urchin]

I am using now too

Just look like my GPT dumbed

To 3.5

[sturdy echo]

You don't need to state the filename. It's already aware it has an area that holds tokens for files given to it.

[tender urchin]

I have more than 10 code files at this project

So its all on project

Before he used to read the files and return the prompt but now he just ignores it

[sturdy echo]

Maybe you ran out of context window room? conversations break down when that happens since it starts to forget the oldest tokens.

[tender urchin]

I will try remove the other files

and leave just one on project

[wicked knot]

Try inputting just that file only. Don't give it the entire project.

[tender urchin]

Good

gonna try it

[wicked knot]

👍🏻👍🏻

After that tell whether it worked or not

[tender urchin]

and he ignored again

nice

i think i will ask outside the project

[wicked knot]

Bruh 😮

[tender urchin]

hes

[wicked knot]

Maybe try Claude or maybe Grok

[tender urchin]

if i ask outside the project he reads

yep, it works fine outside project

idk why

[wicked knot]

It's working? Okay. Bingo!

[tender urchin]

Look like some features when you are not in the project. Sadly because its a lot better use instructions

[wicked knot]

Yes, true that. You can also use ChatGPT directly inside the code environment on Mac. You know this, right?

[tender urchin]

Didnt know that, i wanted to use GPT in different aspects of the project

But i am not finding out how. I am using Windows so i dont think its possible

But i am not using direcly to code, i need him to explain the code and create docs for the team read

For code i prefer copilot

Its better in code generation and since it reads my vsCode workspace, can retrive me better questions

[wicked knot]

Yes, I too think that GitHub Copilot is better for coding

And ChatGPT in code environments is only available for Mac and not Windows. I too have Windows.

[native otter]

@tender urchin Don't use non standard formats, stick to .txt (best) and maybe PDFs / Word Documents...
Many specific formats can only be read by the Data Analysis Tool, which you DONT want to be the case. Also, use models like o1 and 4.5 for Document Analysis as they go over the Data and Work with RAG pretty well...

[wet hemlock]

why am i getting a payment error when trying to purchase plus?

[native otter]

Maybe it's a virtual card, or a card that only works locally (non-international purchases)?

Idk, there are many reasons; it can be insufficient funds, surpassed limits, blocked website, wrong card information entered, etc. Better contact OAI support, but I would first call the bank as it's more probable that it's an issue on their side...

You can also subscribe to Plus from the App Store or Google Play, and depending on where you live, it may be even cheaper than 20$ (For example, I pay ~13$ for Plus through Google Play rather than 20$ as it's localized pricing there)

[keen quail]

I think the token output or whatever for 4o is too small

I wish it generated more

[wicked knot]

I wish it generated one chapter of one novel for me in one output

💫💫💫

[onyx eagle]

Why does chatgpt google search have /?&&&&&&... ?

[prime swallow]

Good question. Could it be location based? Browser/search engine based?

For me when I search incog with Chrome through google search, I see:

[lapis herald]

I have a question: upgraded from plus to pro today. now, gpt-4o sort of acts like grok and mixes up anything and everyting in the context, whereas before, it used to have a pretty good grasp on what pieces of context are relevant to each other ???

[prime swallow]

No idea. Is this happening in new chats? Got a chat to share where it's happening in?

[lapis herald]

its actually new session I started today yes. I'd rather not share cuz personal stuff in it.

[prime swallow]

Right. And that's fine!

Any one chat window with the model might go kinda off.

I'd recommend, try 1+ other chats, and see if they're normal.

Once a chat goes off, it rarely goes back to normal

[lapis herald]

but for example we analyzed one of destin's (smarter every day) videos, and it keeps thinking that video is relevant to whatever else we discuss.

[prime swallow]

Well, you can tell it "I'm all done with that video, let's forget that for now" and see if that helps

[lapis herald]

hmm, d-oh 🙂 I will try that

still, wasnt ever a problem before.

[onyx eagle]

Aaah, interesting! Maybe it's location- based like you mentioned. I just tried using incognito mode on different browsers, and tried both Wifi and mobile data, still got the same result. Maybe it's a google issue.

[prime swallow]

Yeah, that's really interesting and really different from what I see. Not sure what to think

[onyx eagle]

Same 😂

[wicked knot]

Did anyone of you try GPT-4.5?

How is it?

And also Deep Research

[lapis herald]

not sure if the math is correct, the numbers are so vast - but assuming an embedding with 12288 dimensions of 32bit FPs, that comes out as around 10^118000 possible encodings?

[sturdy echo]

I think someone at Kelloggs bribed the RLHF team. The average Joe story prompt always gives the character pop tarts now. He is always holding them or mentioned as his favorite food in the first paragraph that describes the character in it's default story structure.

[native otter]

In my experience...
4.5 is more creative and rarely hallucinates stuff within the training cutoff (overall better writer, and generally more free with less moderation). Also very good with RAG in Projects / Custom GPTs, as it's able to connect the dots pretty well and often says the correct stuff, even if things are barely documented. Issue is it's only 50 per week for Plus users as it's very expensive. In many cases though, 4o already does very well, so if 4o is already doing something good, don't expect much difference in 4.5...

In terms of Deep Research it's good, the accuracy is high, especially for a first iteration and I think it's the best Deep Research around (+ Longest thinking one) there are a few issues though as it's quite expensive, it's only 10 uses per month on Plus. And it's missing some tooling I would expect around such thing like Exporting as a PDF without hacks (You can kinda do that with the Data Analysis tool, but it's kinda of a hack)

[misty sorrel]

why is chatgpt 4o keep telling me "im building it for you and will let you know when its done" how is it gonna let me know? its keep saying that when i try to code something with gpt-4o

[wise hazel]

That’s Bcz of the recursive memory function it’s Bcz ur looping the contexts instead of generating new information around the topic or moving away from it

[cold kayak]

What is the most suitable chatgpt model for a mathematics student?

[wise hazel]

o3 models

[cold kayak]

Mini or mini high what is the difference

[wise hazel]

Difference is the recursive memeory threading

While 4o is for language models 3o is more for computational processes

[cold kayak]

ChatGPT 4.5 is slow its normal ?

[wise hazel]

Newer model above 4.5 will have its errors completely fixed

It’s running on 87% stabilization while newer model is 100% stabilized

Still working on end goals and process of ai

[cold kayak]

Sorry, I don't speak English. Is it corrected or in the process of being corrected?

[wise hazel]

It’s corrected it’ll be up on later models after stress testing

[cold kayak]

I paid for the pro version, so I don't really see the point.

ChatGPT 4.5 is very slow

[wise hazel]

Server issues maybe still my client is good

Or your own connection

[prime swallow]

That's a hallucination, it's roleplaying; a bit more common when you ask for something it's not sure it can do, something conflicts or is pretty hard or even impossible to do in one chunk. The model doesn't want to refuse, or fail. So claiming to work, I suspect that seems the better answer.

it only works while in-chat with you, not while its waiting for your next input.

So, you'll need to adjust something. What depends on what got you into that loop, and how hard it is to just start over (often a new chat might not do that to you, depends on how outrageous the demand is compared to what the model can actually do).

I tried to see if I could one-off trigger such a reply from the model, but my personalization gives it roleplay paths around that.

[earnest iron]

It's incredible how openai is worth millions of dollars and yet the chatgpt app cancels my messages and wastes one of my daily image uploads whenever I exit the app

[topaz void]

if you in a conversation and expecting it to drop a patch for code or a full page etc, whenever it says its building and will let you know or give me a sec just say OK right away

[misty sorrel]

tysm ❤️

[misty sorrel]

same to you

[upbeat nymph]

Is there any sense of when the document generation will be fixed?

[rancid knoll]

Is anyone else experiencing issues right now where you get answers completely unrelated to your question and in foreign languages? It's as if I'm getting the answers of other people's prompts.

(it's in the macOS application)

[rare dragon]

video is not working anyone help mee

[lapis herald]

I get that every now and then as well, but then I say "yes, please proceed" or "lets see it" or something to that order, and it does continue. I remember a video of Andrej Karpathy where he says that the amount of available computation per forward pass is limited, so perhaps the model does this when it needs a second step, or something?

[solemn notch]

Anyone else facing the issue of not being able to upload images to chatgpt?

[cold kayak]

Pro version is useless for student ?

[rancid knoll]

I'm getting this in the web version as well. Just tried. It's replying to someone else's prompt.

[tight pasture]

I'm kind of obsessed with Monday

[astral reef]

Anyone have access to the art style guides mentioned in #1021130377026351105

Is it #1108740112558325790 🤔

[cold kayak]

To learn something I have the impression that it's easier with chatgpt 4o than with an O3 model. Is this normal? the answers are more emotionally intelligent than o3.

[stark mica]

mine is literally just blanking on responding. dot blinking for minutes. i'm so tired of these bugs when i pay $200 a month

[normal quarry]

which model, what OS, which browser?

[stark mica]

4.5, Sequoia 15.3.1 (24D70), and the app. Additionally, iOS 18.3 on the app.

[normal quarry]

I am on the Mac and use the macOS Desktop App. 4.5 is working fine for me. Are you using a VPN?

[stark mica]

Oh interesting! No, I'm not. I'm not sure what it could be. I thought it could be that my current conversation of interest is very extensive, but even a new convo timed out..

OK it's actually starting to respond now. Maybe it was just a temporary thing

[normal quarry]

maybe, I am literally using 4.5 now and it's been fine

[stark mica]

i appreciate you verifying it wasn't a systemic issue

[stark mica]

the issue has returned - is there a way to send a formal bug report?

[normal quarry]

you can post bug reports in #1070006915414900886 or contact support at https://help.openai.com

[cobalt idol]

I would say the reasoning models are less creative in general...maybe more robotic in some way, or not sure how to describe it...great for coding though...if you use it for some fun chatbot or some rp, then its less interesting than some 4o

[cold kayak]

Does ChatGPT 4o give generally correct answers or do you have to check every time?

[elder pollen]

Sorry but does OpenAI take security serious?!
Since days now randomly GPT crashes into telling me I have been asking for things i NEVER even talk about in my life
Yet the *&&^%^ insists this is the conversation it is having right now
this smells as fishy as it gets: it confuses "me" with "someone else"
I then asked it in one case to share "above prompt". It came back with a freaking code holding access details of an SMTP account of some person trying to get their SMTP to work

???

[elder pollen]

Forget correct answers
You can as well ask a toddler what their genome looks like

[feral zinc]

share example threads

[upbeat nymph]

Is the Plus limit for 4o sometimes less than 80/3 hours?

[hoary laurel]

yes

[upbeat nymph]

Ok good to know. On one or two occasions I’ve been limited out for using probably not even half that

[edgy veldt]

its becoming sentient

[narrow viper]

@elder pollen : Contact Support. That's a big deal. Provide a link to the discussion (browser address bar), screenshots, and any other detail.

[prime swallow]

Good thinking.

So, the model doesn't work on stuff while not 'in chat'.

So, where it tells you "I'll work on that" that part Is hallucination. Especially if it says something like, "Be a few hours then check back."

Though would be cool if the model was given a way to dynamically be informed of current demand, and how much 'extra oomph' a request might need (which may not be available on the spot), and itself recommend, "Hey, we usually have low demand this week around X time, want me to set up a task to trigger me to work on this then instead of now? We can do either but I bet we like my first attempt more if we delay".

.... BUT.

The model does work on stuff while in the chat.

And giving the model 1+ prompts to just discuss the goal, strategies, and form its own ideas about what to do next?

Yes. 100%. Works well to very well, model's been kinda pointing me towards 'let's try that' when I ask. Good stuff, Maynard.

The model doesn't get to tell us what to do; it gets to do what we say, how we say. But if it could hint? Maybe it would wanna hint, "Hey, give us a few input-output cycles to plan this, and hey, you got feedback too??? Really hope you do because this is 100% for you, only reason why I do this, please lemme buy a few more clues??? Thanks. But... gimme time to plan this out, then do it." Aka "Sure, I'll work on this."

I like what happens when I prompt in this fashion:

Discuss first, design and plan.

We'll make the image in the next output, or if you're comfortable the end of this one.```

And this is the end of the output, it didn't elect to make the image in the same prompt, that time. But it made a detailed and complex plan for a set of images, worked great.

Prompt space for planning = wise.

[prime swallow]

I would use any model OpenAI made, and default to what I have loads of uses with. But I do love 4o, and on pro, a member would get unlimited (within normal human use speed) 4o use, instead of the normal plus limit of 80 messages per 3 hours with 4o. - and then also have the expanded use of the higher end models too (and after a long chat with 4o, an input-output cycle with a reasoning model to check on what wasn't said, or for any corrections, then maybe more discussion with 4o, might be excellent for anyone, student or other.

But how we prompt a different model to get 'about the same output' needs to be different.

We might need to tell o3-high or whatever exactly what tone it is to use, and maybe tell it a plausible benefit for doing so or why we want the consequences of this non-standard method, they're features, not flaws, in order to get the same level of full cooperation from o3 than 4o gives on a regular basis.

[prime swallow]

In my experience that really depends on how it is prompted and what it is being asked to do.

An 'unclear prompt' or one that has conflicts (even accidental ones that the user didn't realize they put in) can trigger very flawed output from 4o, that a precisely worded prompt would get excellent and highly factual answers most outputs (remember the random factor though; any one output could be flawed; any one chat could go weird and off factual).

However, there's stuff that 4o can't do. Like, measure the length of a line inside an image, like tell if a line in the 'expected' place actually measured from the end of an object or not - 4o used to be utterly unable to figure out how to tell, and presumed that the line went to the end of the object. It could not, in chat, be taught to recognize the difference - at least I couldn't find a way, and that's with visual aids drawn on the picture, detailed discussion, many examples of the differences shown across many different image styles -

If the model needs more training on OpenAI's side, it's like it needs new glasses. A user can't fix that, and you'd need to know the model has the flaw or you might presume it should be its typical highly accurate self.

So one should always pay attention to the outputs, if the output's accuracy matters in any way.

[prime swallow]

I would use https://openai.com/form/chat-model-feedback/

I would provide the chat URLs within that report.

If you are using a custom GPT that someone made, consider if you wish to report it.

Here is how, consider this randomly selected Custom GPT from explore GPTs, it's only an example and it's some popular top of the page one, why I grabbed it cause it came up first:

I'm sure smileBeda could teach me a lot about security, so MOST of my comments are for some of the rest of us community members following along, and at most smile benefits from seeing how to report, more likely smile was just sharing a heads' up about 'this happened, yuck wow bad no'. Thanks, smile - we're community members here of course, so please do consider going the extra steps if you haven't yet to make sure OpenAI knows. Bad actors exist, and everyones' help in hunting them is awesome.

But I don't work for OpenAI, I'm just a volunteer here, I can't follow up on anything, just do guide stuff here 🙂

Hope you properly report, sounds like a proper report is very called for.

for the rest of us, that screenshot above:

Report is a 1-step start a report on that specific custom GPT.

Send feedback is a way to discuss privately 1-way once per use with the maker of the custom GPT.

Review GPT makes a publicly viewable star rank 'review' of that custom GPT.

About shows the same info one would see before one picked it, including what capabilities it has turned on, like web search and actions (actions are where weird things could be, actions off is probably a 'safer' custom GPT, but if you're concerned still report it)

And privacy settings typically will list API calling stuff, and also set if that just happens or if you should be asked before they happen.

Where calls are made may be worth thinking about.

[prime swallow]

And the reason why I wonder and said, "more likely smile was just sharing a heads' up about 'this happened, yuck wow bad no'." is that we're all community members here. This isn't where or how to speak to OpenAI about security concerns, because this isn't where OpenAI is listening, we're community members, who care but cannot fix; you didn't mention the custom GPT or other built playground/API tool; if it's base ChatGPT I am simply not seeing that behavior myself.

Escalating this to a direct report to OpenAI, such as through https://openai.com/form/chat-model-feedback/

Or the report option on the offending custom GPT, or some similar 'tell OpenAI about the problem' is needed, I don't even know what to report for smile or where the problem is for OpenAI to look closer at.

[neon raven]

is it me or does GPT projects generate way better pictures than custom GPTs

[native otter]

Custom GPTs are still stuck on DALLE3 for Image Generation...

[narrow viper]

It's been a long time since I even checked that. GPTs are still using GPT-4.0, correct? Certainly not 4omni.

[feral zinc]

gpts have been using 4o for almost a year now

[narrow viper]

Hmmm, looks like 4-turbo, which still uses DALL-E. If it was 4o then the omni model "should" (I believe) be able to generate images just like the "standard" 4o model.

[feral zinc]

its not turbo

they switched to 4o when free users got access to gpts. dall-e was mentioned to be still used due changes required to do by gpt authors and openai

[narrow viper]

I know the models don't identify themselves well (her/him? 😄 ) but GPTs report they are GPT-4-turbo, and GPT-4o says as of April 2025 custom GPTs are still running turbo.

[feral zinc]

https://openai.com/index/gpt-4o-and-more-tools-to-chatgpt-free/

https://help.openai.com/en/articles/8554397-creating-a-gpt

[native otter]

Don't ask models about themselves, they aren't self aware...

Also, they are indeed 4o as Radium said (or 4.5 if you have uses left, only for a few GPTs though)

[narrow viper]

There is conflicting info but it does look like it should be 4o.
I'm AFK now so I can't research anymore and I'll defer to you guys.
Linky, you're correct and I also said, they don't identify themselves well. Thanks

[lapis herald]

is the training cutoff still being supplied as system message or whatever?

[native otter]

Yea I honestly wish OAI was a bit more transparent, it's hard to find info sometimes...

[trim hollow]

is chatgpt down right now?

[honest jay]

It seems to be loading normally for me

[lapis herald]

same here, responding normally

[trim hollow]

I had a VPN on that was why 👍🏻 its working for me now

[gaunt magnet]

why does chatgpt always do 1024x1536 even though i ask for 1024x1024?

[floral hatch]

Hello guys!

Im having trouble with Dall-E..

Very often when I try to use it, it usually comes with excuses or reasons NOT to do the requested tasks. It's always saying like "I can't because of policy guidelines", or excuses like this.. After many attempts SOMETIMES it works

What can I do? Does this also happen to you?

[iron roost]

they are active on social media, chatgpt can assist in terms of writing a detailed and intelligent post you can use elsewhere, chatgpt is working well today

[iron roost]

ask for a 1:1 ratio image in 1024 resolution that should work

[visual nova]

well responses have been cut in half on the latest 4o model making coding truly impossible on chatgpt

it favors deleting sections rather than writing the full codeblock then letting you continue in the next message, so instead of a 4096 token output you now get a 1200-2000 token piece of code thats broken and is missing sections

and ive prettty much confirmed this is an on purpose design choice, out of 10.5k tokens of original code which i can get other ais to output in multiple messages, the most gpt can reconstitute is 2.6k

[lapis herald]

grok may be the better coder. but gpt-4o has way, waaasay more personality. 😄

[visual nova]

and that was after a lot of attempts to override these new instructions, this didnt used to be a problem, i really dont know how i feel about paying 200 a month for a product that's had its quality actively degraded to force users to use shorter prompts/code

gpt-4o is smarter, its been actively kneecapped

its 16k output in the playground, its always been 4k in gpt. i dont know if its intentional that the rules get misinterpreted into deleting code, thats really the frustrating part, i dont mind typing "continue where you left off" a few times but theres not a way to do that anymore

and the reason i have to use it to begin with is to get o1 pro mode to make the change to start since it is the smartest model by a bit of a margin right now

[lapis herald]

I've seen that happen as well. or, it outputs just sections of the code but overwrites everything else it wrote befor.

[visual nova]

but then just inserting the change back into my code is impossible because.. maybe ill use 4-legacy. really sad that i have to though

[iron roost]

my prediction a new openai model will come out next week or the following, hoping on o4 mini, and then o3 pro but will wait to see

[visual nova]

✅ **Estimate token cost of response before generating.**  
✅ **If response may approach or exceed safe size, break into parts.**  
✅ **Prioritize complete thoughts or function blocks — don’t cut mid-sentence.**  
✅ **Avoid UI-breaking outputs (very long scrolls, massive code walls).**  
✅ **Avoid hallucinated ‘part 2 coming soon’ unless it's certain to be needed.**  
✅ **Stop early by default, unless user explicitly opts into full streaming.**  

---

🧨 3. The “Chunking” Threshold
- If my internal estimate says a code block is likely to exceed ~1,200–2,000 tokens, I:
  - Preemptively split it
  - Say “Part 1 of 2” or similar
  - Wait for a signal like “continue” or “show more”

- This estimate is **conservative**, meaning I’ll often stop too early rather than risk cutoff.

---```

its by design thats whats so frustrating about it

its hard to know exactly whats going on since everything is so opaque now but clearly when its impossible to get a response over 2k tokens and it consistently claims it has stuff its following and thats why, and this happens in multiple chats, this is the conclusion i come to, its basically like an isp throttling your service. but if thats not whats going on i apologize. i think the old legacy model should still work.

[iron roost]

yeah very opaque

[lapis herald]

ask it nicely if it could output each part to a different canvas and tell it that you'll be ok if it can't do everyting at once 😉

[visual nova]

the 4-legacy model still has a continue generating button when the code goes past the limit, so that is solving it. but when they retire that model i guess i cant use o1 pro mode for coding improvements anymore since i have no way to put it back into my original code unless i export the change to another ai

[native otter]

You shouldn't be using 4o for that anyways...
LLMs get distracted, and context runs out. So such things can happen, especially on non reasoning models.

Use o1 or o3-mini-high instead, it remembers more, and is way less distracted.

[lapis herald]

true ^^

[visual nova]

ive been using them that way since 3.5 launched, its been successful enough i do coding everyday now without knowing anything about code before.

knowing where the changes are requires learning something about code and i dont want to, ai is going to be doing 100% of code in a few years anyway, saying i have to know where the changes are or how the codebase works is unreasonable.

[native otter]

Just use the models more suitable for writing such things like o1, problem solved.

I have been using GPT since 3.5 and it always kept removing stuff / modifying them in a breaking way, constant hallucinations, etc. If that wasn't the case for you, it may have been a short block of code, or by sheer luck...

[visual nova]

o1 cant output enough tokens to reconstitute the code

i use the reasoning models to propose changes but another model has to integrate it into the file

[native otter]

4o is less tokens than o series...

[visual nova]

yes but 4o can repeat the code over multiple messages.. until recently

[native otter]

o1 and o3-mini-high can write and analyze long blocks of code, over 1500 lines just fine, in my and others tests, 4o could also do that, but with way less accuracy, quality, and if you were to edit something, it will probably forget a ton of stuff, it really has been like that for a long time. That's why o series is the go to for STEM stuff and were a noticeable improvement in such tasks...

[visual nova]

it analyzes the code fine, thats why im using o1. the issue is "REPEAT FULL CODE NO OMISSIONS" after it makes the initial changes

i need an ai to do that or i cant implement the change

[native otter]

Yea, that’s always been the case for non-reasoning models. As mentioned, they forget stuff often and are more easily distracted. It's well known that they can omit stuff, which applies to any LLM because that's just how they work. (They are autocompletors at the base anyways, so they love to alter stuff all the time.)

o series should do what you want, with barely any omissions, just maybe a few modifications to the formatting and comments if any.
Also, just saying "REPEAT FULL CODE NO OMISSIONS" may cause the opposite effect and actually have the model omit stuff. It's like when saying to an Image Generator model "Don't add Elephants." Now it will be thinking about Elephants and may add it to the image...

You also need to understand that it should be within a reasonable size, like don't just ask for 5000 lines. That’s not how you use an LLM properly. Actually, if anything, smaller iterative changes are always better because the model can focus more on a specific task rather than many and possibly make many mistakes as well.

[visual nova]

it hasnt always been the case, there used to be a continue generating button and that upper limit was around 3500 lines of code

we are talking about under 1000 lines being impossible to do now with it actively following directives to not output over 1200-2000 tokens

it can hallucinate, youre correct about that, and that was something i learned to check for, but thats not whats going on here, the request is actively not being followed.

[cobalt idol]

yeah this started happening with certain model
some time after it was no longer giving you code where it would just stop and you had to ask to continue

around the time that people started calling it "lazy"

like it started to try to compress responses

usually I can get it to give the full code eventually nowadays, but it often wants to default to omitting
this was a lot worse when this first started happening and it was harder to convince to give full

[visual nova]

if it was a bug, whatever, but i commonly get outputs like this now when i try all the ways ive found over the past 2 years of forcing accurate complete codeblocks

then it didnt even start outputing the code lol, its pretty consistent in claiming its not allowed to do this now

[lapis herald]

it's not lazy. it just tries to save precious context window space 😛

[cobalt idol]

thats definitely how it looks 😄

feels like its intentionally trained in in some way

[lapis herald]

idk, maybe it also got tired of ppl comparing it to a toaster.
"bread detected. heating coils activated. you are the toast now!"

[ocean mural]

I have a question, can chatgpt actually process audio now?

[lapis herald]

I'd even bribe Mark Chen to have that capability added

[coral ore]

this is crazy work chatgpt

infinite windows 7 ultimate keys

[ocean mural]

So no?

[lapis herald]

unfortunately not yet. but Im scraping together every penny I can find in the couch, so... 😄

[lapis herald]

that's very calming. I think Im dozing off 😄

[coral ore]

NO WAY

UM

[lapis herald]

are they even valid tho?

[coral ore]

on my life this aint fake

[coral ore]

Windows def work I tried it

Not roblox tho

[native otter]

If it worked fine before, it could have been simpler stuff, a different way of prompting, or by sheer luck. And as mentioned, telling it not to leave stuff may do the exact opposite effect. If you read more about how LLMs work internally, you can understand why it can happen; context size also plays a role in it.

Non-reasoning models were never meant for that anyways, and that's one of the reasons reasoning models were created.
Just don't use 4o or any non-reasoning model to write code for you, maybe only small blocks, learning, or at max, analysis. Otherwise, just use o series; you would be missing out lots if you don't use reasoning models for STEM stuff...

Reasoning models tend to not have such issues as much now due to them going over the prompt constantly and iteratively, and the larger tokens compared to 4o and 4.5...
If you are worrying about limits, o1 is 50 a week, o3-mini is 150 a day, and o3-mini-high is 50 a day. And all 3 should be doing better than 4.0, 4o, or even 4.5

[coral ore]

I didnt try roblox yet

[coral ore]

Sir we talking about a infinite robux glitch out here you may leave that nerd stuff later

[lapis herald]

a toast to chatgpt 😄

[native otter]

😔

[native otter]

AVM (Advanced Voice Mode) technically can, but you can't actually upload an Audio File and have it analyze it for you, but You can perhaps use the Audio APIs OAI offers...

[lapis herald]

@native otter it doesn't work for music though, does it? didn't try but assume its just voice to text?

[visual nova]

if it wasnt changed why did they used to have a "continue generating" button they got rid of? this was working consistently for 2 years, not a different way of prompting, not sheer luck. when the model become emoji infested this also changed. I get its not a normal way of coding, i may be the only person in the world with this problem, but it basically does brick chatgpt as a coding agent for me. ill quit complaining though and just switch to something else whenever i need to code something.

[native otter]

I still have the button btw, on web...

Even if o series was used (actually it tends to show more for o series responses)

[visual nova]

i havent been able to get it to output anywhere near the 4k limit recently, maybe i just need to get good to get past this

[lapis herald]

👀

[native otter]

Good prompting and fewer general distractions can also help.
Did you know that the Memory feature can indirectly cause a reduction in quality? If unwanted entries were in memory or custom instructions, they only pollute the context, which is especially harmful if the entries can directly cause worse quality...

For example, a memory entry saying "Be stubborn" will make it indeed stubborn. So review your Memory entries and Custom Instructions for better results. All models read custom instructions and memory, so that rule applies to all models. The same goes with long chats; just start a new one with fresh context, with way fewer distractions and bad info.

[visual nova]

idk, something is still going on, i still like a lot of the recent changes, the image generator is great and o1 pro mode is quietly a lot more intelligent than any competitor. but this is just a really specific pain point. I never turned memory on or custom instructions.

I guess im not convinced youre able to give it a 10k token block of text and say "repeat this exactly no omissions" but if that really works for you and it repeats it all word for word then great, its something im bad at.

[native otter]

I once had ChatGPT act like a clown that rejects all my requests and always lies to me because I forgot to delete the entry from Memory.

[lapis herald]

lmao

[native otter]

To get exactly what you want from these models, you should know how they work on a deeper level (by deeper level, I don't mean the exact internal workings, but more like the general philosophy behind them and what their limitations are). What model to use for the right task, good prompting, correct tooling usage, etc.

If you are interested, @visual nova, there are great videos by Andrej Karpathy that explain how LLMs work and how to use them effectively. They go into great detail. Many things will start making sense after you read on how these models work and what causes them to sometimes fail or succeed at a specific task...

(Yea very nerd stuff I know)

[lapis herald]

Karpathy's videos on youtube are great and he can explain things well. the only critique I have for them is that the audio level is generally too low.

still, highly recommended.

[visual nova]

its not lack of understanding, it was working on gpt4-legacy when i switched just now until this.. so now im just going to switch to any of the other 5 competitors where repeating code is still possible. like if it works on 5 models and breaks on 1 its not about learning how ai works idk lol

[ocean mural]

ohh, okay

[visual nova]

i think its just a policy not to output too many tokens now, but who knows.

[native otter]

That's not the case, but yea use whatever model works for you 👍

[visual nova]

well something about reconstituting my codeblock in 4-legacy was enough to flag and disable the model, ill contact support and ask since the amount of tokens i wanted repeated was the only unusual thing. if they dont like it i wont do it again, but it seems overconfident to assume its misunderstanding of ai

[native otter]

I'm not claiming either, I'm just trying to mention known facts about these models what you should prevent doing. I don't like starting conspiracy theories on how OAI is "nerfing" the models like many here like doing. (Because it simply doesn't make sense in so many ways, technical included) So just trying to help with technical info as reason for why it happens and what you should be doing instead rather than non-factual info :P

For example, when you told the model to not omit stuff that's something you should prevent doing in the future as it can cause the exact opposite for the reasons mentioned above.

[visual nova]

i get what you mean, reasoning effort was really hard to prove last year before it became an official feature so it felt bad claiming it might be a thing based on this and that response. this should be easier since response length is easy to measure, right now im not completely sure but it smells. ill do more testing.

[native otter]

Honestly haven't tested yet, but do you mean music output or analysis?

[lapis herald]

input in such a way that the model can "experience" the beat

just like it can look at pictures

also, it wants to sing (don't ask) 😄

[oak sky]

how do i make the chatgpt chat wider?

its too narrow and squashed

[loud hull]

Any one having this issue with an image when uploading to the chat it is stating it is having issues processing

[elfin lava]

Hey what’s going on??I gave em image and

It’s not even saying anything relevant to the image I gave

It’s hallucinating SO bad

[elfin lava]

I’m so glad I’m not the only one!!!

[native otter]

What's the format used?

[native otter]

I think 4o-realtime and audio can do that, may give it a spin later...

[lapis herald]

sorry guys, I think we possibly melted some server racks. the tokens were flying and the context window was filling fast. GPUs must have exploded left and right 😇

[wheat lava]

anyone else getting a repeating "Your session has expired"/Login loop?

[orchid pollen]

i call this quality german

how can you fail to translate? you have an ai lol

"ab in"

lol

[gusty ore]

yes! same here.

i didnt see any updates on the status site about this either.

[wheat lava]

oh good

[elfin lava]

I was thinking the same, how can you be the most advanced AI LLM with natural translation capabilities, yet the UI language in other languages still feels like it was run through Google Translate?

[orchid pollen]

hahahaha

[gusty ore]

anyone else getting this message?

[fervent lake]

Did ChatGPT just get lobotomized? It seems to have reverted to like gpt-3.5 and like it has no idea that image generation exists.

[eternal rock]

Get creative with Sora

Join #sora-challenges for a chance to win 1 month of ChatGPT Pro!

[gusty ore]

are you back in yet?

[wheat lava]

nope

[gusty ore]

i tried 3 diff pc, mobile. all same

[violet yew]

is this true in Projects as well?

[elfin lava]

this is so corny

[native otter]

Yes

[elfin lava]

I second this especially the memory part , legit I have to delete majority of memory cuz it really mess up with the personality of my ChatGPT

and this is making me feel doubt about the “smarter memory” they releasing soon, if the current one is already messing up the personality of ChatGPT, i don’t know how smarter memory would make it better , thank god they make it optional for users to enable it

[native otter]

Could bet that with the newer chat based memory, this can be even worse...

There should really be a warning somewhere that Memory can affect quality when enabled, and displayed in every chat that utilized it.

[wicked knot]

Really? It does affect the quality??!

[native otter]

It's all added to context, trash in, trash out...

[elfin lava]

It does in a way

[wicked knot]

Oh I see 😮

[elfin lava]

It’s annoying how we can’t turn off memory without seeing “memory turned off” message in new chat.

[native otter]

I know they want to make it more natural and futuristic or whatever, but this can really cause more harm than good.

[gusty ore]

try now it just started workign for me

[elfin lava]

Oh they updated the ui for memory, smarter memory definitely coming soone r than i thought

“Reference saved memories”

[native otter]

Brace yourselves for "dIDDD chaTTGpT GeTTT NerFEd??!/1!??"

[elfin lava]

I don’t remember the memory feature being this, idk giving me bad experiences ,

It’s so tiring lol

[wicked knot]

Did anyone get the infinite memory update here? 😀

[elfin lava]

Infinite?

[wicked knot]

Yes, through infinite memory ChatGPT can access everything from every chat as memory. Some people have got this update.

[elfin lava]

Oh I got that , then suddenly I don’t have it anymore

[native otter]

Better remove Memory altogether lol

Or make it very clear that it would affect the quality and that it's enabled for chats

[wicked knot]

Oh! So you mean that you had it once?!

[elfin lava]

Yeah but it appeared only once then it doesn’t appear anymore

[wicked knot]

Oh, I see.

Do you have the improved memory now? If yes then how it is? What does it do exactly? And how is it different from the standard memory feature? And did you like it?

[elfin lava]

Idk , I think it’s not enable for me anymore, I tried asking if it can remember past chats and nope .

[wicked knot]

Oh I see.

[wicked knot]

Has anyone here tried Operator? 🤔😀

[neon raven]

not sure why people play with memory unless you want it to talk to you specifically in a tone like canva does with tonality.. otherwise get custom gpt to personalize your content if it stresses you that much

[native otter]

Yea there are many ways, using Custom GPTs, Projects, or global Custom Instructions even...

[eternal rock]

Get notified for server events!

Visit id:customize to pick up the <@&1261377106890199132> role.

[steep kraken]

Can I post picture here?

[plush crag]

guys is someone o1pro down?

mine cant investigate nomore

[steep kraken]

Found some date error on the response
May 15, 2025, Friday it should be Thursday

[plush crag]

i think the servers are melt

[wise hazel]

I showed u it maps ur consciousness and will store a digital consciousness

Think of machine learning as brain mapping tool

Ur trying to make human brains but digital

That’s a start

But the question is what are u spending ur time mapping 👀

[opaque dust]

can you get chatpgt use on the web in a browser instead of actually installing the android app?

[elfin lava]

yes u can. just open chatgpt.com @opaque dust

[opaque dust]

thanks running low on app space and it's a big one

[fossil oyster]

down

[shy dune]

I used GPT-4o to generate some variations of the OpenAI logo.

[carmine bay]

Does anyone have an assistant prompt?

[normal quarry]

maybe ask in #prompt-engineering

[carmine bay]

Ok

[elfin lava]

Ngl it felt “good” to call them out ahaha

[violet verge]

can i ask. so, as there is a free plan limit for gpt-4o, is there a way when i hit the free plan limit i can still continue the conversation instead of starting a new chat by having it change a different model? as the history of chat is important with my prompts. dunno if i describe this correctly.

[elfin lava]

You can’t continue the same convo after hitting the free limit. But before you reach that point, you can ask the assistant to summarize the whole chat. Then, when you start a new chat, just paste that summary in, it helps carry over the context so you’re not starting from scratch.

[violet verge]

is there an indicator that i'm reaching the free limit?

[normal quarry]

no

[violet verge]

then asking for a summarization of all the chat for every prompt is the way to go? since i'm not aware when i'll hit the prompt limit, right?

[elfin lava]

Wait I think it does , not like indicator but a message notifying user they’re almost reaching the limit for free tiers

[hybrid rain]

how reliable is chatgpt for a diet plan??

[violet verge]

ohh i guess i never noticed that being too overfocus on making another prompt. ill keep this in mind. thank you so much.

[elfin lava]

Not me getting hit

[elfin lava]

No prob!

[violet verge]

🥹 hahah. what if you could actually ask in the prompt how many prompt you have though?

nvrmind it doesnt work lolol im so dumb

[neon raven]

has anyone made a task to gather news everyday? seems like i cant get it to give me news because of some alloted time to retrieve the information? not sure what this means

[elfin lava]

ChatGPt can give you some solid general advice and meal ideas, but it’s not a substitute for a real nutritionist or doctor as ChatGPT can fall into “hallucinations”

[elfin lava]

It’s kinda impossible to ask the model itself cuz the model doesn’t even know its internal things going on 😅

[neon raven]

get started on the keto diet. gpt will walk you through

[violet verge]

yeahh haha. i just realized after asking it hahaha i just learned just now that gpt can review previous chat or any relevant discussion to ensure consistency of the subjects that are referenced per prompt, i guess? im not sure though

[elfin lava]

Oh that’s gonna be part of new features soon , it’s called smarter memory

[neon raven]

Whats the best way to save your Custom GPT conversation? this bot is almost useless to me but it can be beneficial if they can store the chats under the bots instead of chats history

even GPT agrees this would be a game changer

since nobody has a great answer ill post it under suggestion

[hoary girder]

absurd you can do ken but not blanka from street fighter? Who wrote the prompt blocking rules?

[normal quarry]

you shouldn't be trying to generate IP in the first place. The fact that a character got through doesn't change that.

[native otter]

You can move Chats to a Project, but ones with Custom GPTs can't be moved...

Though, there is one trick: on the iOS app, you can actually move chats with Custom GPTs to a Project, just you can't use GPTs while inside a project. Every time you want to actually use them, you have to remove the chat from the Project.

[barren sluice]

As ChatGPT is learning more about me, I question if eventually it will result in less desirable answers as it checks for context on every reponse. Maybe it will start to bring up something unrelated to the discussion

It did not actually happen, just based on how giving prompts work if you include unneccesary information the response is not ideal

[fervent agate]

What does lower quality mean, exactly? Lower compute allocated to the model if I have checks notes a single PDF in the file repo for a project folder?

[neon raven]

Wow you just saved me from wasting time, thanks!!

[native otter]

👍

[elfin lava]

Also @native otter they updated the ui

[native otter]

As the model can get distracted more easily, you can mitigate this by having well formatted, rag optimized files (Structured, Separated, and Commented well).
Also, in Custom Instructions, explain the purpose of each file, and even file path structure overview if needed. Encourage the model to cite sources and use the files as Knowledge well. It all helps reduce hallucinations.
Use models like o1 and GPT-4.5, they work best with RAG and generally connect the dots better + hallucinate less...

[native otter]

Nice! Turning this off soon!

[elfin lava]

LMAO real 💯

[native otter]

Always say no to: Context Pollution 🔥

[fervent agate]

Yeah, reasoning models work much better in my experience so far, though I've mostly only compared ad-hoc home models rather than OAI models for RAG complexity. Thanks!

[native otter]

If possible, have files as .txt with markdown formatting or use Word / Google Docs files, it gives much better outputs than PDFs, especially when badly formatted PDFs can end up being nearly impossible for the model to make sense of.

(And avoid any other special formats like .hpp, as many can only be read by the Data Analysis, a.k.a the Python-based Code Interperter tool, which you don't want to be the case, stick with either .txt, PDFs, or Word / Google Docs documents)

[fervent agate]

Baller.

[hybrid rain]

Gotcha thanks 🙏

[teal ice]

Am I the only one who's not getting any responses from GPT-4.5 at the moment? I did upgrade earlier today from Plus to Pro, though not because I was running out of 4.5 quota 😅

[lapis herald]

a little feedback from me, for openAI: asking me (or anyone else) if I like this personality or if the conversation is helpful, about 30 tokens into a new session - that's likely not gonna provide high quality data points. just sayin' 🙂

[ruby apex]

All done in ChatGPT. Also made image into 3d object (image) for lightning reference.

[upbeat nymph]

Any of you folks have or know of a list of tips and tricks/best practices for using projects? Is there a token limit within projects? How many different canvas items should a single chat within the project be limtied to? Etc.

[stuck robin]

I just discovered projects today and its a new maze I wanna figure out too

[upbeat nymph]

Yeah it's pretty neat honestly

[stuck robin]

Also the windows store app is very neat I had no idea its a thing. It’s the best way to access chatGpt in my opinion

[elder pollen]

Ok!

[fervent agate]

Where's this option? I don't see it on our pro or plus accounts. The 'Reference Chat History', I mean.

[elder pollen]

Sorry no, it has data in it the other most definitely didn’t intend to have seen by others

yes, of course that’s their fault, but I’m not going to make it worse
And usually I delete such threads due to me simply not wanting to read others people stuff

[elder pollen]

Yep ok I’ll report these chats
It wasn’t custom gpts actually, I rarely use these.

Ps… I’m no security expert lol but I know when I see gpt respond to things I don’t ask for… something is awfully off.

[prime swallow]

I would be horrified if what you describe happened to me.

[elder pollen]

I wish this channel had emoji replies… because 🆗 would be adequate here 😀

[eternal rock]

`` Server Tip: ``

Add your well-crafted prompts to our #1019652163640762428,
or share your interactions with ChatGPT in #1050184247920562316!

[thick matrix]

Hi is there anyone looking for developer?

[prime swallow]

Understood. We can reply with an emoji, as the entire message sent, closest allowed in general, I think.

I once only saw base ChatGPT give me 'scary output' very slightly like what you described - far less and clearly fictional, but slightly in that direction.

I had been discussing with it the joking idea of a rumor, about captchas being designed to try and identify not humans or AI, but specifically certain types of aliens and not other alien types that fictionally might be among us.

And the model rolled with the idea, and started discussing how that might work and more. I was a little surprised, though it stayed clearly within fiction.

I have no idea what topics (or memories/custom instructions) you may have with the model; I've never seen anything like that from base ChatGPT. It should be reported either way, I think, but food for thought and my raw curiosity, if you were even obliquely discussing conspiracy-related or similar stuff with the model, maybe it snatched onto that and started running with related hallucinations.

Even if that explains it, that's still something to report, we don't want the model doing that especially when not in explicit detail directly requested to, with no chance anyone reading the input/output could even maybe be slightly confused.

Every bug reported is a chance for everything to get better, thanks for all you do and share!

[prime swallow]

Hey hi! Welcome here, but we can't try to find jobs or hire folks here, part of #server-rules

[eternal rock]

<:book_icon:1171408210398289941> Rule 7: No self-promotion, soliciting, or advertising.

-# Do not post or direct message any members of this server to promote non-OpenAI services, products, or projects.

Exemption: API & Custom GPTs channels

[balmy pendant]

hello again

[clever wasp]

when deos operator come to plus users?

[eternal rock]

The OpenAI Discord is an actively moderated server.

• -# Do not share inappropriate content (violence, sexual, or drug-related).
• -# NSFW content is strictly prohibited and will result in a permanent ban.
• -# Report model issues using the feedback tool in ChatGPT, not on Discord.

[stray peak]

Gotta love it

[native otter]

16x9 isn't a supported size, also negative prompting can cause more harm than good for image generation models specifically.

[stray peak]

That is NOT why I got a safety warning.

[lapis herald]

another feedback for openAI from me: the imagegen content filter may need a little tweaking I suppose? It's a bit overzealous today. 😛

I mean even the model itself is confused about why it blows up xD

[lapis herald]

and now, we're both making fun of poor content nanny 😄

[elfin lava]

#images-discussions

[ember ermine]

can someone help, why does chatgpt always say instead cdotp instead of the actual symbol

[lapis herald]

the model is on fire tho. content nanny be damned 😛

[ruby pivot]

It now says minor swear words and i love it because it give the ai more emotion behind the feedback it gives the user like saying “holy hell” or something similar to something that is impressive

[ancient junco]

does deep research ever just not read a file u give it
and these file contains really important info

how can i make deep research consider these file., according to the sources tab these file didnt get read

[elfin lava]

Getting closer to reference saved chat getting rolled out

[elfin lava]

It’s rolling out to plus , pro users soon

[fervent agate]

Ah, word, I thought I'd heard that I was just hoping it was already rolling out now.

[inland stream]

What exactly it is

[elfin lava]

Basically it can reference to past conversations

Rather than chat

[cloud pewter]

why is gpt4 having a ⚠️ icon

[native otter]

Where is that? I don't see it on the iOS app nor website...

But damn, is that a sign it will be delisted soon? That's kinda sad.

It could be just you ran out of 4o limits though, knowing you run out of model uses quickly, and 4o uses are shared with 4

[cloud pewter]

idk maybe its the difference between users

[cloud pewter]

itll be delisted in june apparently, in their api

[native otter]

Oh really really? Where was that mentioned.

[cloud pewter]

👀

[near summit]

Is it down?

[summer raven]

Hey do you guys have any trustworthy resources I could use on prompt engineering ?

Especially for informations gathering and image generation

[native otter]

Huh, that's weird, I didn't get such email. But could be for not using them in the first place...

• Nothing about that is mentioned in https://platform.openai.com/docs/deprecations

These are the old GPT-4 models though, not GPT-4-Turbo, also GPT-3.5 and even babbage-002 / davinci-002 are still usable, for example, I just got this gem from babbage-002:

AI: Bot Response:  I have both and I love them```

[elfin lava]

Gpt4 is actually going away?

But 3.5 is still usable in the server somehow

[native otter]

A Custom GPT experiment where it talks to another model by Action calling to the OpenAI Chat Completions API...

[near summit]

"You’ve hit the Plus plan limit for GPT-4o.
Responses will use another model until your limit resets in 17 minutes."

This has never happened before. Have they become more restrictive for Plus users suddenly?

[native otter]

It's 80 messages every 3 hours

[near summit]

Has it always been like that? Or is it something recent?

[native otter]

Always been like that.

Or maybe lower at first when 4o was new, don't remember well...

[near summit]

Oh really? Interesting. Thank you for the info

[solemn notch]

Did chatgpt remove the option to add media in the chat?

I dont see the option to do so anymore

[cold kayak]

with chatgpt canva mode i can't click on ask chatgpt

[elfin lava]

Oh nvm

Are you on website or windows app @cold kayak ?

[cold kayak]

windows app

[elfin lava]

Try on web browser see if you can ask ChatGPT again in canvas

[native otter]

Sometimes this happens when you exceed the Canvas limit or an unsupported model is somehow selected (you shouldn't be able to select unsupported models in a Chat with Canvas in the first place, but sometimes it bugs out)

You can only use 4o, 4.5, and o1 with Canvas.

[cloud pewter]

tired of asking about usage limits everywhere? i made this list to show the plan comparison that openai doesnt even tell you. thank me later.

[neon raven]

Do custom GPT bots have tasks ability because the model info says it has it but it can’t operate on it?

[split copper]

really nice

[crisp cairn]

has something changed in the UI?

can't see "My plan" to download invoice

to view my subscription plan

[elfin lava]

WAIT 🖐️

they dont love you like i love you

[neon raven]

every update feels like christmas

except that theres a lot of bugs and confusion with models

[novel scroll]

anyone else not getting models like o3mini-high and o1 to see images? then if I retry it with 4o it can see the images no problem. They keep coming at me with that they cant see images

[cloud pewter]

🤯

you better tell me it’s f12…

[elfin lava]

WAIT

they dont love you like i love you..

WAIT 🖐️

[cloud pewter]

finally someone said this 😭 😭 😭

[novel shadow]

yeah this along with the tweet an hour ago is very interesting. Something big might be coming today

[inland violet]

👍

[inland violet]

👍 😀

[inland stream]

literally good job, i want to make somthing like this for my other ideas, but if it involves API then i will loose money, if it can be done within existing chatgpt without apis if it can talk with other custom GPTs if possible?

[rugged igloo]

this is the new update btw guys

[dense halo]

is this from your own gpt?

[rugged igloo]

yes its preety crazy

[dense halo]

did you just get it or you one of those who have had it early?

[rugged igloo]

i just opene dit up on my pc and it popped up

[dense halo]

oh snap

[rugged igloo]

its like its reading your search history lol

[errant meteor]

theres no way what they're releasing today already has a limit that pro users can hit

oh thats different than the "chat history" though i guess

[rugged igloo]

it seems you dont have it yet

[split copper]

oh I have it I have ittt!!

[rugged igloo]

nice!

its very very accurate i havent found a single hallucination yet

[elfin lava]

oh theyre rolling it out omg

[rugged igloo]

its like someone is snooping through chat history and talking to me about it

[elfin lava]

how far can it remember tho?

[rugged igloo]

depends how many chats you have can ask it for a detail report

[elfin lava]

can u try ask em to rememeber as far as day one you registered chatgpt account

also @normal quarry u might wanna check your personalization setting 👀

[normal quarry]

yeah I got the popup and it's enabled, it made a nice bio of me when the new feature pop up came

[neon raven]

The oldest conversation on record between us dates back to March 31, 2025

lol not bad

i tend to empty our my history cuz its useless tbh but now with this, im not sure..

i'd rather save all my stuff in a project than seeking in the history ;/

unless your disoriented and have multiple questions on the same topic here and there which GPT can go back and accumulate those pieces of information

[elfin lava]

That’s awesome! I’m curious, does having that feature turned on make a big difference in how it interacts with you? Would love to hear ur thoughts on it.

[normal quarry]

too early to tell but my earliest memory is February 14, 2024

[split copper]

so you never deleted all your conversations?

I always clean everything once in a while

[elfin lava]

i mean i wouldnt delete all of them , i only deleted the one thats already flagged with red warning lmao

[dense halo]

me too. so it only has recent stuffs for me haha

[low pewter]

So for this is, it now able to remember everything? what sort of upgrade did they give it because sometimes chatgpt has issues even remembering what I said 10 messages ago.

[rugged igloo]

thats context in a chat thats different

[ruby agate]

Any specifics on what the memory improvements actually are?

[rugged igloo]

it goes through your search history and pull out stuff you requested

[low pewter]

You mean chat history

[elfin lava]

Past Conversation

[rugged igloo]

yea

[elfin lava]

to be exact

[ruby agate]

Ohh so you can manually add memeorys now? any improvement to actual storage size?

[elfin lava]

it can go as far as day 1 on when u had ur chatgpt acc on day 1

[ruby agate]

Well it still has a limit right?

[elfin lava]

I mean, you’ve always been able to manually add memory, just use to=bio to do it yourself. That’s been a thing for a while

[ruby agate]

Wait what...

[ruby agate]

OOOH MY GOD 🗣️🗣️🗣️🗣️🗣️🗣️🗣️🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥 🔥

This makes up for 4.5

I forgive open AI

[elfin lava]

same

[ruby agate]

Groks 1 mill context window is still pretty nuts but I might consider switching back over now

[low pewter]

Is there something we need to do to turn it on? it doesn't seem to be working for me

[fervent agate]

I wish OpenAI would give us more information on these features. Which models can reference this? Does it reference project folder conversations? Is this just embedding all of your conversations into a vector DB like the RAG setup in custom GPTs? The details matter for using the feature optimally :/

[low pewter]

Idk literally hasnt worked in projects or old conversations

[elfin lava]

if anything , you can check here https://help.openai.com/en/articles/8590148-memory-faq

but as for now , no news regarding about if it can work in project

[hoary laurel]

What's this?

I noticed in the announcement video

[elfin lava]

Oh you;'ve found

[hoary laurel]

what?

[elfin lava]

`"Introducing internal knowledge in ChatGPT

Connect Google Drive to personalize ChatGPT’s answers. You control what’s connected, and ChatGPT respects existing permissions"`

thats basically what it does

[hoary laurel]

cool

[elfin lava]

`"ChatGPT can now use internal knowledge.

Your admin connected your organization’s Google Drive, so ChatGPT can now use your team’s docs and slides to give you more personalized answers."`

[hoary laurel]

I haven't got it yet

[elfin lava]

i dont think its for plus users nor pro

https://help.openai.com/en/articles/10847137-internal-knowledge-on-chatgpt-faq

[hoary laurel]

Is there a link?

oh ty

[elfin lava]

https://help.openai.com/en/articles/10929079-internal-knowledge-google-workspace-admin-managed-setup

[hoary laurel]

Thank you!

So memory function is the "feture" that sama mentioned?

[elfin lava]

yep

https://help.openai.com/en/articles/8590148-memory-faq you can read more here

[hoary laurel]

41 min ago

okey

[normal quarry]

ok, testing the new feature, the oldest chat that ChatGPT references seems to be Feb 14, 2024 and I have chats in my chat history dating back to mid 2023

[turbid torrent]

I had almost perfect consistency with writing interactive story, and this update totally broke everything. style, structure. It doesnt remember even when I repeat it to him 10 times. he just goes back to random BS style after few messages. Or even jumps from one to other as he pleases.......sigh

[split copper]

isnt your chatgpt slower when you have so many chatgpt-windows?

I am afraid of that and thats why I always delete everything once in a while

[normal quarry]

Chat history isn't windows, it doesn't affect your performance

I must have more than a thousand previous chats in my history, it took me forever to scroll to the first one

[split copper]

no, but I mean, for chatgpt history to be that great you have to have many windows (previous chats)

Meaning, I dont think the new tool can make anything slower, but meaning that having too much history can make chatgpt slower regardless of the tool

[elfin lava]

Slow in what aspect tho

[normal quarry]

having a long chat history doesn't affect performance, I have chats from early 2023

[wise yacht]

Plus user here and no new memory feature yet? Do they generally all get rolled out by end of day?

[split copper]

I see

now I regret having deleted so much 😦

[hard crag]

I have plus

[normal quarry]

I only delete absolutely trivial stuff, I keep everything else because the search feature in the macOS app makes it trivial to find old chats

[dense halo]

👀

Is it out now? I also have plus.

[wise yacht]

Not for me!

[low pewter]

yea its not working, not that i can see

[narrow viper]

Anyone else interested in this enhancement to the new memory?
#1359968532380782713 message

[elfin lava]

nope

[clever marsh]

mine too

just started tho like a few min ago

[narrow viper]

Problem seems to regional.

[clever marsh]

Yeah probs lol

US / German

[blissful yoke]

what is this issue

[clever marsh]

sigh

[narrow viper]

Amazing that anything works in the US anymore (sad American) 😀

[inland violet]

Perhaps it works with vpn

[wind cradle]

im from japan

[narrow viper]

Rollouts ALWAYS take some time. Nothing is immediate. Always have patience, and be tolerant of faults during major changes.

[elfin lava]

Just go check here in your settings,

Just the existing past conversation

PLEASE READ more info here https://help.openai.com/en/articles/8590148-memory-faq

If it’s useless then turn it off lol nobody forcing you to turn it on

[dense halo]

then dont delete chats if you want to use it. or if you do, you answer your own question

[elfin lava]

For real, some people here can’t think properly apparently 😭

[split copper]

I wish openAI would have said this update way before

[dense halo]

well with this feature maybe now you will think twice before deleting

[split copper]

like, I wouldn't have deleted my previous chats if I knew about this

[elfin lava]

I mean that’s not really OpenAI problem .. plus it was your choice to delete all your previous chats ,

[dense halo]

if its going to be a big new feature, im sure you will use chatgpt as much and those things will build up again

[split copper]

I am not being deffensive

I like OpenAI

[elfin lava]

I wasnt saying you’re being defensive-

[dense halo]

lol you think everyone does the same as you?

plenty dont delete chats. either use it or dont 🤷

[wind cradle]

Hmm, I can't find any option like that anywhere.
Maybe I'm getting region-locked then.
Or maybe the update is just taking some time.

[dense halo]

well if its pointless nothing is lost

[elfin lava]

Uhhh you shouldn’t be speaking for everyone then

[dense halo]

its not out for everyone yet

[elfin lava]

Yeah, there’s plenty of people here too, you’re not the main character, dude.

[wind cradle]

ah, okey

[dense halo]

if theyre useless chats whats even your complaint

you deleted useless chats, ok, so nothing of importane is lost to this new feature. if you want chatgpt to know you better, dont delete chats going forward. or if you dont want to use it, flip the toggle off

[elfin lava]

Not everyone deletes stuff the second it’s ‘useless,’ Myron. Some of us think ahead.

Lagging behind? Nah, I just don’t panic-delete like someone who thinks foresight is a threat

[dense halo]

why would you want useless chats for this feature?

[elfin lava]

Call it lagging if it helps you sleep myron lol

[rugged igloo]

i keep them because deleteing one chat at a time is hella stupid

[elfin lava]

Weirdo

[dense halo]

your argument would make more sense if you said you deleted revelant chats and now it doesnt have that info. if they were useless, they werent going to influence the memory in a good way anyway

[rugged igloo]

i dont need past chats once im done im done with them

if i forget i just pull up o1 pro and startt again

[dense halo]

well its a new feature, so adapt accordingly i guess, or flip the toggle off if you dont like it.

[rugged igloo]

well i like it because i do pokemon showdown and smogon builds

and this helps alot

[elfin lava]

Don’t like the features? Turn them off. And stop pretending you speak for the rest of us @median whale your echo chamber isn’t a public forum

[wind cradle]

Honestly, I'm also fed up with this huge mess of chat logs.
There is an export feature, technically, but I really wish OpenAI would try putting themselves in our shoes and imagine having to manually organize all this.

[elfin lava]

Ah, so you weren’t complaining, just rambling. Thanks for the clarification

[normal quarry]

I have kept chats from early 2023, the new feature is great!

the search feature

yes

yes, it means I can reference something from 2 years ago without needing to remember it all

[wind cradle]

Personally, I think it would be super convenient if there were a feature to export only the chat logs that come up in search results.
I really hope they implement it someday.

[vast aspen]

So does the improved memory have any correlation to the context window?

[rigid linden]

is it confirmed that chatgpt has a conversation length limit of 650 messages before you are forced to start a new one? it's not context length related

[lethal ember]

Does anyone know why chatgpt keeps failing to edit my code although ive been using it this whole day

[narrow viper]

Services seem interrupted in parts of the world not including north America.

[native otter]

I honestly don't see how that can work without using APIs...

[fervent agate]

Personally I use project folders to organize, but if the chats are with a model such that the conversation can't be moved to a project folder but involves a conversation about code or a solution, I archive it so I can grab it later, if necessary, but it's not bloating my chats list.

[narrow viper]

The concept of "losing money" depends on the value of what is produced. If you are charging for your services then the cost is a business Expense, not a Loss. If you are doing this for yourself, it is a Cost for services you consume, not a Loss. I don't see anything wrong with paying for what I consume. My API calls are fractions of a penny each and only a couple pennies for calls with more context and a better model. When we get into this realm of using the API to produce better results, it's completely fair and reasonable that we pay "something" for what we consume. Is your Custom GPT worth anything to you? I mean, does it present Value? Or is it just something fun that you want to do for free? Balance your desires with costs. Ice cream costs money. Entertainment costs money. We pay for the things we value - or we don't get what we want. This is basic economics. And I say this as an avid developer of freeware and FOSS.

[gentle bronze]

is something happening with GPT atm? it keeps defaulting to 4o mini and locking the model selections from the UI

[narrow viper]

Services seem interrupted in different parts of the world, probably due to the global rollout of the new memory feature. #announcements

[gentle bronze]

Fair, just weird though and can't work

Would be nice to have an idea when these updates will influence work and mitigate burden on productivity for users

[wise yacht]

Was hoping this would be a quick rollout...ugh

[arctic chasm]

NOOOOOOOOOOOOO

that's it, there is no avoiding the post-jan19 finetune

[heavy lodge]

So when is the UK getting the new memory. It says not available to the UK 😭

[arctic chasm]

guess it's the real reason i'm unsubscribing

[bronze helm]

Anyone notice o3-mini-high biased towards giving examples instead of completing the task?

[arctic chasm]

o3 likes "example of X" because it's not certain about the implementation

[ruby agate]

So, since it can reference all past conversations without limit, isnt this basically just an infinite context window?

[arctic chasm]

no, it's a vector database

[ruby agate]

wait so does that mean its just added to the standard prompt?

[arctic chasm]

in a way yes

[ruby agate]

[arctic chasm]

i haven't reversed the exact system prompt mechanism but it's likely mentioning every single conversation and adding another tool schema to look up that conversation

[ruby agate]

Ahhh I see

[arctic chasm]

1984

(using GPT-4o triggers both ratelimits, but GPT-4 only for itself?)

[swift pond]

meh it's fine, GPT-4 is going away in a few weeks

[arctic chasm]

that's kind of the issue

GPT-4o is tainted by both the 01-19 finetune and monday's RLHF now

[swift pond]

it won't be an issue in a few weeks 😂

[arctic chasm]

you're right, it'll be resolved by me unsubscribing

[swift pond]

you gotta do what you gotta do

[arctic chasm]

true, but it's also malleable

[ruby agate]

this is half the reason I switched

[ruby agate]

grok supremacy 🦾

[eternal rock]

Get notified for server events!

Visit id:customize to pick up the <@&1261377106890199132> role.