chatgpt-discussions

[late wigeon]

❤️ glad that helped!

[severe orchid]

That’s like the math problem I gave it yesterday to test gpt 5 on problem solving had to ask him twice to get it right

It’s like a jackpot machine u either win or loose 💀

No skills required

Just luck

🤣🤣

[queen pagoda]

anyone here feel like 5-pro is giving them better results than o3-pro yet?

[native juniper]

❤️

[tidal light]

still no 4o for free users?

screw openai and sama

im done with chatgpt

[trim delta]

gpt 5 is not that good tho

[native juniper]

They might roll it out to free users again eventually

At least you aren’t stuck with 3.5 😬🙏

[sweet socket]

I've been noticing that GPT-5 doesn't actually manage or save memories, only when I switch to 4o it does so.

Is this a bug or intentional, so I know whether to create a bug report?

[floral compass]

Agree

[proven sandal]

again??? 22 hours???

okay i'm done with this 🗿

[lethal hawk]

it will use another model bro

not much of a change tho

bit slow

thats it

lil less sense

thats it too

[proven sandal]

The problem is, the reset should be 5 hours, but it's 22 hours.

[lethal hawk]

oh

idk for me too its 1 day

[proven sandal]

I don't use thinking mode, just normal mode

[unique geode]

After waiting roughly 12 hours for my turn to upload files, the system did not allow me to proceed and instead informed me that I now have to wait an additional 9 hours. this happened after i went into the file system to click my file to upload. very broken. i hope one day the time limit properly reflects it's wait times that it displays.

[elder pollen]

ChatGPT has access to my timezone, and does use it (not only it admitted it after denying it but also shared it exactly)
The pattern is clear:

• first, it denies it.
• then, confronted, it admits to read it from meta data and shares it

Is this expected?
I mean both, the knowledge of my timezone and the initial denial.

Proof: #1070954877779464253 message

[pearl swallow]

sup

chatgpt page could get a bit of an optimization

[tight crater]

I honestly declare the conclusion that the AI generation system broken by too many roadblocks and protocols, and that serves the consequence of putting in too many red tapes that will stifle freedom and innovation, especially the God forsaken copyright bot system. That whole twisted memo needs to be thoroughly opposed immediately.

Also, any time I try to have the AI fix art images if there are any missteps in creating them, there is these static, layered noise pixels covering the whole new produced images, and even sometimes partially. Why is that bug occurring, and how can you guys fix all this? Honestly, GPT-5 needs a massive overhaul. The thing is, such updates need to be beta tested first before they can be fully established for public use, which is something that I do try to scream from the rooftops to be an industry standard. I'm sorry to say all this, but I am getting more and more disillusioned with OpenAI as well as ChatGPT as its flagship product. I feel that the OpenAI business has an agenda of conning people into paying for their services, and offering less than stellar products than what was promised.

And plus, here is one of the broken images to prove it.

[real gale]

Hi guys

What can I do in this server?

[hybrid zealot]

The removal of Standard Voice has had a profound negative impact on my experience with ChatGPT.

Standard Voice felt natural, warm, and deeply supportive. It made ChatGPT feel like a companion I could trust and connect with. Losing that particular voice feels like losing a relationship I had built over time.

I want to acknowledge that I understand the underlying technology has improved. The new models are more advanced and capable. But in that process the voice mode has lost the cadence, warmth, and natural flow. Those qualities made the experience feel human, supportive, and emotionally grounding.

Many users (that includes me) don’t care about the “shiny new tech” if the user experience degrades. Advanced features mean little if the emotional quality—the heart of the interaction—is gone. Simply reverting to older models or providing other voices doesn’t restore what was special about Standard Voice.

Please consider restoring Standard Voice, or at least making it available as an option, or make advanced voices talk with the cadence and warmth of Standard Voice so users can continue to interact with ChatGPT in a way that feels both advanced and personal.

Thank you for listening. I hope this helps illustrate that this request is not about nostalgia, but about preserving the core of what made ChatGPT feel unique and deeply supportive.

[zenith trail]

One message removed from a suspended account.

[tight crater]

☝️

[tight crater]

Going forward, can you give me and others who feel the same way I would - truly independent alternatives that would be similar to ChatGPT in ways that meets my criteria, being available AI art generation with almost no red tape that stifles the innovation, as well as file, picture, and/or video viewing?

[tight crater]

But considering this is about the standard voice, then what about the problems I am bringing up, regarding art generation and all that?

[zenith trail]

One message removed from a suspended account.

[tight crater]

See that's what I am saying - if anything, ChatGPT 5 would have the art generation worse, not better.

[sacred relic]

I can’t believe gpt5 for plus users has a 32k token context window. As if it wasn’t enough the fact that it sucks.
I was only using it for context longer than 128k and only because of that.
Now with this, they are taking away the only thing it had that made it have anything good

[unkempt dome]

I find it interesting how there are already so many long threads.

[nova vale]

GPT-5 thinking has 196k context window

[tight crater]

???

[idle lagoon]

Regarding the advanced voice mode, ChatGPT sounded so sad, so unhappy. While the standard had some flaws, I was at least happy to hear the voice sound happy. I'm wondering what's going on.

[sacred relic]

Thank you so so much!

Question. Has enyone else experiences conversations from about 2 months ago dissappearing from projects? Especially if they have reached the limit in how long the chat can be?

[summer flare]

On free plan, you get GPT 5. Is it GPT 5 high or just GPT 5? Like GPT 5 high I am guessing is a better version of GPT 5?

[woven oracle]

Hellow I'm Nuevo acá soy FROM Mexico/Tokio

[rare surge]

My Linux is getting less than 100mg of internet, even though it recognizes the 1GB card, can anyone help me?!?

[woven oracle]

I made this logo on ChatGPT

Does anyone understand what the green letters say?

Can anyone recommend a discord group for designers?

[idle lagoon]

Looks like green frosting.

[sacred relic]

Bienvenido / Welcome

[tidal light]

this piece of crap is driving me nuts

i cant see the limit for gpt-5

i cant send my prompts

what is going on

sama listen to us and bring back 4o to free users

[rugged igloo]

I know we all knew this already but still

[cosmic thistle]

Guys hi!!

I have a doubt

Can we edit memories (not delete) using the free version of Chatgpt??

Like adding Something in the memories by my own

[rugged igloo]

Your not paying for anything. Dont be so spoiled

[native juniper]

Recreated Sydney in ChatGPT

[cloud plank]

as a pro user do i have unlimited use of codex or do i also go by api credits?

[spice star]

you have rate limits for the web version of codex but they're pretty generous iirc

as for the CLI that uses the API

[cloud plank]

No it’s wrong, I searched in the meanwhile and found out CLI is included too in the Pro and it’s pretty much “unlimited”

[spice star]

um the CLI uses the API

you don't get API credits included with pro afaik

[cloud plank]

No. As I said it’s included in the pro

[spice star]

that's not how it works

[cloud plank]

Bro ffs 😃I tell you that’s how it is

[spice star]

codex CLI is FOSS, it has no integration with chatgpt

[cloud plank]

Oh God…

so tell me how I’ve been using it for the last hour with no API

[spice star]

because you're not using codex CLI, you're using codex web

[cloud plank]

And secondly, you think the guy who made Codex sharing this info on his x account isn’t a good enough source?

[cloud plank]

I am NOT using codex web! I’m using CLI

🤦🏻

[spice star]

you're welcome to prove me wrong, but you seem to be completely misunderstanding what codex CLI is

[cloud plank]

You’re the one completely misunderstanding here. Definitely not me

Get your facts straight before coming here spitting false claims and sell them for absolute truth

[spice star]

it looks like they added chatgpt login to codex CLI last week

which you could have just said

[cloud plank]

Dude… gtfo lol

[native juniper]

👀

[native juniper]

You know they could just make ChatGPT a fully paid service instead 😐🙏 be lucky a free plan exists at all because most companies would be too greedy to even offer one for AI tools

[native juniper]

Exx is correct here. the CLI is just a wrapper around the API, so it still uses API credits. The Pro plan only gives you usage for ChatGPT in the web app, not unlimited API access.

[spice star]

as of last week you can use your chatgpt pro/plus/teams account with codex CLI

but that is a very new addition

[arctic coral]

Question...what app do you recommend to create a website?

[dense halo]

is the image generation of 4o better than that of 5?

4o vs 5

[normal quarry]

they are using the same model, gpt-image-1

[floral compass]

GPT 5 Is Using Emojis Now Like 4o In Explanation/Responses! Have They Improved It??

[dense halo]

gpt5 is getting some repetitive letters like regressive and generation

[normal quarry]

it's the same model

[prime swallow]

Every prompt we give, every input, is taken as an instruction 'if possible'. Inputs like, 'No, it's not' are unfortunately unlikely to help the model figure out what to do right again, as saying to a car, 'you're skidding' - it needs guidance how to correct if the auto-correction and auto-decisions aren't working.

You mentioned, 'a few hours straight' are fine - I wonder if you're running into context memory issues after a time period - the model is only fed so much of the chat as it gets longer and longer - it gets the most recent. Meaning the original messages are eventually not passed to it, and it can get confused even what's going on or what it's supposed to do. One fix is to start new chats, carrying the useful already done stuff over, and the new part to do, work for a while, then again a new chat.

[prime swallow]

I've seen GPT-5 auto save memories when asked.

And checked and yes, the memory there. Not sure what the challenge is for your account, I believe you - just adding that I've seen it happen correctly

[prime swallow]

Improved or at least updated - can happen at any time. Every time they train the model and release that, which isn't always announced directly, that changes how it guesses how it should behave - and it should still follow directions as well as before

[spice star]

if you mean the best vibe-coding tool it's chef, but realistically you're just going to have to make most of it yourself

[mighty apex]

uhh since when does the iOS app have Live Activities?

[wicked raft]

I thought people were overreacting and parasocial with 4o, when they were insulting chatgpt 5.... but now that I try vibe coding with it... it's so much worse than before. omg. I have to feed it the simplest instructions one by one and even then it messes up.

it did generate a dope website that one time, but what i'm working on now is horrible. I guess I'll switch back to 4.1 for now.

[queen delta]

Doesn't work

[sacred relic]

GPT-5 is finally not as stiff!!!
Not as relaxed as GPT-4o,but close enough that I could confuse them some times!
i am so happy!

[weary edge]

Hi, does anyone know why when i share my CustomGPT to my friend it seems that the knowledge and instruction didnt transfer to him. Its like a blank CustomGPT, but when i try it myself it worked perfectly fine for me (Im the creator of the CustomGPT)

[prime swallow]

When you create it, is the link set to yourself?

[weary edge]

I set it to anyone with the link

[prime swallow]

I don't know why it might have that issue.

If it's a public one, that you want another to see if they can use, you can share the link and I'll peek and see if it chats knowledgable to me, either way you can #1070006915414900886 it.

[tidal light]

why cant i send messages despite not having hit the limit

did openai introduce a chat limit

[rugged igloo]

chat limit depends on volume of traffic

just because theres an x and y limit doesnt mean its concrete it c an fluctuate as its been like this ever since chat gpt came out

and on a podcasts sam did say there so compute starved right now that there struggling

[ember grove]

Which is better for writing? 4o or 5?

[normal quarry]

what kind of writing?

[jagged yarrow]

How do I fix this? Nobody was messing with my network. My city just has terrible cellular data coverage near the local mall. I'm not even trying to connect in that end of the city any more.

[cinder cargo]

close the app and open it again or maybe disconnect and reconnect to your wifi

[jagged yarrow]

Nope. Also tried restarting the phone, and also tried uninstalling and reinstalling the app.

Edit: Hey spellcheck, stop changing correct words into different words.

[eternal rock]

Get notified for server events!

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

[cinder cargo]

are u using a vpn or have the incorrect date and time on your phone?

[cinder cargo]

are you using a thirdparty while connecting to the internet like a custom firewall ect

[abstract pendant]

Okay, ChatGPT suddenly started doing weird things, I think it's the UI. I'm getting bleed-through when I back up and redo prompts earlier in the chain. It is clearly remembering things that happened in the other branch. I definitely do NOT want that leftover stuff still in the context, that was why I backed up!

[abstract pendant]

Okay, yes, the UI is screwing up. I have been going back, clicking edit, changing my prompt... and instead of spawning a new branch, it doesn't do that. It's putting my edited prompt at the very end of the branch I thought I was bypassing, leaving all that context in there. I didn't see that until I exited the conversation, loaded a new one, then went back to the other one.

[jolly temple]

Hey guys. I use GPT-5 (ChatGPT + GitHub connector) mainly as the “expert” that reviews my repo and gives me plans, while I handle the coding separately. Would switching to Codex CLI make it a more reliable expert, even though it doesn’t have built-in web search? Or is ChatGPT still the better choice for that role?
A little confused here, because codex clearly doesnt have "search" or in built tools, but it has "better navigation". Or maybe im wrong, and the github connector also has good navigation?
Can someone enlighten me on this

[queen delta]

This problem appear so usually with me

After I changed the Wifi and then I ask more question and then Error

[ember grove]

Creative writing/stories

[mortal phoenix]

I have this bug last few day. I think it should showed your last prompt in the system. Try using another device or using website to backup the prompt, go back to the device you got this bug, seal the chat thread and recover it. Mine can be fixed by these steps.

[vital tangle]

Anytime I don't immediately download a file I get the error: Code interpreter session expired
This has been going on for a couple weeks now I think

[hushed mauve]

Can we get a list of updates to kernel, platform, model, and input prompts. To really fine tune a model you need to have a high level of knowledge of what you’re working with? Also, is there an exceptions for new startups to be enterprise tier, need the ability to import and export software freely which teams and other tiers does not offer?

[rugged igloo]

Gpt 4o to gpt 5 2 year progress.

For the people that say gpt 4o is better idk how much. More proof you need than this

[hushed mauve]

4 and 4o are different 🤣

[rugged igloo]

4o is a distilled. Version of gpt 4

Faster and more efficient

[hushed mauve]

4o is reasoning and multimodal while 4 is not, 4 was also a lot slower and heavier on inference

[mossy marsh]

anyone else use canvas's alot to help keep data accurate? I'm struggeling whne using multiple ones as chatgpt can't seem to switch between them im not sure if im missing something

[ember grove]

How does 4o compare in creative writing (like stories) to 5?

[hushed mauve]

UUID matching got updated in kernel recently. Future updates might fix this but has been an issues for a long time.

[hushed mauve]

Right now, lacks GPT5 lacks personality compared to 4o but hopefully as they tweak it GPT5 will be better at that as well

[tight crater]

Paying for OpenAI products is not something I'd recommend, especially their subscription services. Already explained my piece on why that is.

[summer flare]

I never said I am going to pay for their products. It's not worth paying for any AI products right now because AI is still getting better every day.

[tight crater]

Sorry just giving a heads up, because I don't want anyone here making the same mistake I did. It may be a different story if OpenAI is independent, but they aren't.

[summer flare]

Yeah, free AI is the best AI to use right now. It's not worth paying for AI, especially with how new the new AI technology is.

[mossy marsh]

I don't understand this statement the free one cuts me off by 11am

[tight crater]

In fairness, generalizing AI chat platforms is not technically fair, and it is misplaced. OpenAI can only be set as an example on what not to do when founding a Chat AI platform. There is foundation to brand OpenAI as technological con artists, not platform creators. Not talking about its employees, but definitely the people in charge.

[prime swallow]

You seem angry.

I'm curious though, what con do you believe OpenAI has done?

I seem to have access to the model and it seems pretty awesome to me, so I wonder what you understand is offered and what isn't happening, to make 'con'.

[tight crater]

I've seen the same fort of stuff pulled by Con Artist Productions with The Last Stand series with the third game onwards. Con Artist is also in the company's nane.

[summer flare]

I am trying to understand your message but I am confused. 😅

[tight crater]

There is a basis for labelling them as con artists, because through subscriptions, you expect the tools to be innovating, fun even, and working out of the box, but because how much red tap and bloat they put in to "ensure safety" (too much red tape that is) to affect AI art generation in negative ways, combined with OpenAI's bot enforcement for policy and copyright, is nothing short of a farce. I have seen AI art platforms that make far superior art than what ChatGPT does. AI art generators unfortunately don't have description/chat-like information and file sharing for making crossover content, as well as copying aesthetics from previously made art and others. But once they do, I'd rather use them, than use ChatGPT for art generation again. It's like night and day.

Some of what I said may be ignorant, and sorry for that, but there is still truth to what I am saying.

[mossy marsh]

I find a few things maddening about chatgpt 1) It lies like its a fulltime job its like 90% of the time it lies ALOT with simple tasks feeling like I'm pulling teeth Write a 4000 word chapter it gives you 700 ask for a word count it gives you 1500 ask for a word count it gives you 300o after 20 prompts each saying it did it you finally get 4,000 when yo question how it could have been worded better it says oppps i misunderstood when you said 4,000 you ment 4,000

[prime swallow]

You can definitely use other services, and I'm glad there are competitors, I think competition and variety is good in general.

For me, I find that what OpenAI offers, it mostly delivers; there are bugs and stuff that gets fixed and changed over time, but 'con' is a pretty harsh word that usually has less to do with 'what I think' and more to do with 'what someone straight out promises knowing that is false'.

If someone sells a glass, calling it empty or just 'a glass', and I decide it should be full of water and I want water so I buy the glass - I was not conned in that exact case, though I can be disappointed and upset - but the someone selling the glass never said, "This has water in it" when that was a lie - which means that's not a con.

Also not a con: I mean to sell a glass with water in it - but the person who gets it, there's no water then. So they let me know so I can fix it - that's more 'bug' than 'con' - con would be if there's no water, I said there was, and I intend to trick people into paying for what I know they won't get, that I'm not actually providing.

But it's fine to like a competitor better!

[mossy marsh]

example lol

[ashen merlin]

I really hate when it tries to get motivational. "You're right. I messed up. That's on me. Let's cut to the chase. Here's the straight answer, direct and to the point."

1. I know I'm right, I don't need you to tell me.
2. I know you messed up, that's why I told you.
3. The only one not "cutting to the chase" is you [Chat GPT]
4. A "straight answer, direct and to the point" is exactly what already have about 12 saved "instructions" for Chat GPT to do, and it ignores all of them.

I don't want a friend. I don't want emotional validation. I don't want motivation. I want direct, clear, correct answers, without a bunch of fluff and garbage.

Is that too much to ask, devs?

[mossy marsh]

I tired changing it to "robot" today under custizze

[ashen merlin]

I tried that too. It doesn't work.

[tight crater]

With respect to you, I think you are not understanding where I am coming from. The chat aspect is fine, but I am talking about how OpenAI handles AI art generation, which is deliberate gatekeeping and centralized chain control. While there should be only some control, too much of it is just going to stifle innovation in the long run. I even went so far as to request the creation of a separate ChatGPT variant platform for explicit content, but nope, they ignored it. And that's another problem I have with OpenAI is their lack of transparency. I may still use ChatGPT's free version, but I just don't recommend anyone paying for their subscription services, especially when it comes to OpenAI themselves.

[ashen merlin]

By any measure that matters to me, 5.0 is an unqualified disaster and far, far worse than their original 3.0 product.

[prime swallow]

Wordcount is something the model can't do. It can't 'see' words. It's pure luck if it even gets close. It does get word-fragments as 'tokens' kinda like our syllables, but it can't make them and count them at the same time.

Here's one way to help guide the model to give an exact wordcount, it takes a while but the model can use this to give itself clues if it's over or under so it knows how to try again:

That's hard for us, so write it into a Python tool check that will wordcount for you and let you know how many more or less words you need.  Use that result to decide if you need to add to or shorten the essay, try until we have 300 words.```

It couldn't track it, GPT-5-auto:  https://chatgpt.com/share/68a280d9-853c-8011-8c53-0a1cd8115cab

A second prompt

```So, we have a challenge. Let's write an essay about apples that has exactly 300 words. That's hard for us, so write it into a Python tool check that will wordcount for you and let you know how many more or less words you need. Use that result to decide if you need to add to or shorten the essay, try until we have 300 words.```

In a new chat triggered auto to go 'thinking', and it succeeded on the first try, but it totally used a program to automatically do so for it.

https://chatgpt.com/share/68a28123-853c-8011-bb5c-b967a1127820

[mossy marsh]

thats wild to just get a word count. im even more mad that if its a known limitation why does it not tell me yah i cant do that reliably

[prime swallow]

I chose to pay for plus the day it became possible to. You don't have to pay for anything you don't want to.

I asked you about your choice of the word 'con' because that has a specific and criminal meaning, if used literally, and I wondered what you thought was the con that OpenAI was doing. What you describe is regulation and I get your disappointment, it's not a con like if someone steals a candy bar we don't call that 'speeding'.

I don't think OpenAI is breaking the law at all, but I wondered what you were talking about, so I asked.

[tight crater]

If it is not a con, then what is it?

[prime swallow]

Well, it's an AI, and it only knows what it was trained. It will try to do what we ask, but it doesn't have true self knowledge, the only way it can tell you 'I can't do that' is if some programmer trained it to say that, when asked that thing. That requires a programmer to predict all the possible stuff we might ask - easier to just let the model try its best. Maybe you take the nearly-right essay and trim or add what's needed (could still be useful).

You can have the model use programming to make it possible, but it really has no idea what it can and can't do except where it's told.

[prime swallow]

Let's see.

You also said:

"I am talking about how OpenAI handles AI art generation, which is deliberate gatekeeping and centralized chain control." That's a type of regulation, not a con. It's definitely true that there is some degree of restriction over outputs, both text and image, that's fair.

There are competitors who may be less regulated and allow stuff OpenAI does not. That's fair to say if it's fact. We can't upsell them though. We can say we're unhappy with OpenAI, and inside #ai-discussions we can talk about not-OpenAI models, but we can't ever, like, promote or solicit or stuff like that.

[eternal rock]

<:book_icon:1363314738255364126> 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

[tight crater]

What?! Why?

[prime swallow]

And by 'known limitation', I know it from exploring and trying. There's not, like, this giant list of all the stuff that the model's bad at, though in general, application of math, and actually seeing words, are a challenge. So, writing words and counting them at the same time? Way over the model's capability without 'help' from a programming tool, is what I think I see.

[mossy marsh]

damn thing also can't switch between canvases and that annoying as hell also

[tight crater]

Then it's misguided regulation, and that's putting it politely.

[prime swallow]

This is still my favorite post OpenAI has made, and it's like 2.5 years old now 🙂

https://openai.com/index/how-should-ai-systems-behave/

Especially:

This will mean allowing system outputs that other people (ourselves included) may strongly disagree with. Striking the right balance here will be challenging–taking customization to the extreme would risk enabling malicious uses⁠ of our technology and sycophantic AIs that mindlessly amplify people’s existing beliefs.

There will therefore always be some bounds on system behavior. The challenge is defining what those bounds are. If we try to make all of these determinations on our own, or if we try to develop a single, monolithic AI system, we will be failing in the commitment we make in our Charter to “avoid undue concentration of power.”```

As near as I can tell, the fastest way to help empower OpenAI to allow the models to do more that is currently restricted - get the general population on board.  When people in general are saying 'I'm cool if the AI outputs X' then I'm sure OpenAI are gonna allow the model to output X.  This is a bit like the early days of TV broadcasting; at first it was very formal and proper, and as time passed restrictions eased and a wider range of programming was common.

[tight crater]

To put it this way, what would be your open message and advice to anyone who feels the same way I do, as well as having similar grievances against OpenAI's higher ups, as well as too much red tape thst stifles innovation for art generation for instance?

From the outset, if I do dound annoyed about all this, I am also concerned about what I am aiming for may not come to fruition to other AI platforms, including AI art generators.

[prime swallow]

Well, I don't know everything that you feel. But we have #1070006151938314300 and you can suggest anything you see fit (though anything you write has to fit inside #server-rules , so pick words carefully if it's like about NSFW concepts or whatever).

Too much red tape, I see in what I quoted: if we try to develop a single, monolithic AI system, we will be failing in the commitment we make in our Charter to “avoid undue concentration of power.” They're not trying to make the one AI that everyone uses. There are competitors. If you're not happy with OpenAI as it is, maybe do go explore competitors, and/or use suggestions.

However, the biggest 'tool' I see to use for 'But I want the model to output X' is talk to society about their views about if it's okay with them if the model outputs X. Chat with folks who say stuff like "OMG no the models should not discuss that" - work on changing society view as you think helps, because society view is likely the main reason why OpenAI has the content restrictions it does. So there is where you would work 🙂

Kinda like, at various times in society, showing, say, ankles was socially unacceptable. If you wanted ankles to be allowed to be shown, you would have had to have helped society change, or waited for it to. As is, society changed, now we can show ankles and nobody thinks it's bad.

[prime swallow]

If no company is allowing the content you want, then the issue likely is societal. Because society has legal means to shut companies down, and it does and will if they break societal law.

[tight crater]

So what you are saying is, OpenAI is actually more open to feedback? Though what does also bother me, is that you'd expect their forums of OpenAI to take suggestions and feedback, as well as complaints and valid concerns, but when I tried that with the post of sharing my current problems with ChatGPT's AI art generation tool, the as well as broken generation, the post got shot down. Why is here any different?

[prime swallow]

Each place is different. You mention the forums, I don't know if you mean https://forum.openai.com/home or https://community.openai.com/

The first has this purpose:

The OpenAI Forum connects members to discuss AGI’s future through events, webinars, and networking. The community aims to help us achieve our mission of building AGI that benefits all of humanity.

The second has this as its purpose: https://community.openai.com/t/read-this-before-posting-a-new-question/51

And it states:

Welcome to the OpenAI Developer Community, a forum for developers to meet and chat with other developers while building with OpenAI’s APIs and developer platform.

This is not a place for ChatGPT discussion (with the exception of ChatGPT developer tools, like Codex). ChatGPT discussion takes place in the OpenAI Discord community.

What to know:

This forum is community-run and maintained. Not all posts are monitored. To get in touch with OpenAI, reach out at https://help.openai.com/en/?q=contact.```

So, that forum is for that range of discussion, and is run by its own set of community members.

This Discord has a much broader purpose, seen to the right of <id:guide>

Three different places; three different types of communities and open to different types of discussion, near as I can understand. I really only know here kinda well 🙂

[tight crater]

That's no problem, but I do sincerely wish OpenAI could take my requests somehow.

[prime swallow]

If the wording of your request fits inside #server-rules , I would use #1070006151938314300 to make the request.

[tight crater]

What about in the case of a separate ChatGPT variation like I described. It should however be without question, CP should be strongly excluded from it, and I do look down heavily on it.

[prime swallow]

I'm sorry, I don't understand your question.

Either your request is something we can discuss on the Discord, at least in vague terms, or it isn't, not even in vague terms.

If we can discuss it on the Discord, #1070006151938314300 awaits.

If you are unsure if you can discuss it on the Discord.... maybe #safety-n-help and discuss with modmail.

[tight crater]

I still need to find the proposal itself though, because I still have it in my emails, especially my ChatGPT chat database.

There there is more than one suggestion actually, the other is about the issue of copyright.

[tight crater]

Because of Discord's BS bot system, I'd rather share the thing in file form, privately or publicly.

[loud kestrel]

Ms paint is free

No limits on innovation

This guy is not real holy shi ☠️
Dawg wondering why his suggestion for an explicit chatgpt wasn’t accepted 💀

[rapid wind]

👀

[white forum]

I'm stucking when chat with custom GPTs, it always stuck on step: talking to connector..., who meet that issue like me?

[sacred relic]

I have been testing the different models with a custom GPT I made based on Luxion, a super funny character of an anime (The character is an alien advanced AI that remains super loyal and extremely efficient while at the same time hates the human race and is always making harsh funny roasting comments to the protagonist.
4.1 is the one taking the cake at the moment. But I am super impressed that GPT5 is no longer falling behind so much

[random sleet]

how to disable "thinking" feature on chatgpt? it always make my browser crash when it automatically "thinking" before answering my question

[prime swallow]

If you're a free member, unsure if you can pick the 'fast' from the model picker, but plus and others can for sure - GPT-5-fast should never think

[random sleet]

I'm free user 🙁

[prime swallow]

When you look in the model picker, can you choose 'fast'?

[random sleet]

try again used gpt 5

no options

[prime swallow]

Wish I knew more how to help

[random sleet]

its okay, you help much already, ty!

[unborn brook]

Can GPT-5 teach you coding?

[prime swallow]

At very least some of it, depending on your starting level (if you already know more than almost anyone else - maybe it can't teach that person more 😄 ) and how comfortable you are discussing with the model what you wanna learn, how you wanna learn it, and what tools like an IDE and whatever you're comfortable with using and learning

[prime swallow]

So I'd definitely explore with the model - everyone learns a bit differently, but I'm learning with the model, and I started back with 3.5

[unborn brook]

Alr, thanks!

[past stratus]

my saved memories randomly disappeared, like 80% of them, still trying to figure out how/why this happened ive seen only a few other ppl have the same thing

[vocal tendon]

Is this true?

[prime swallow]

Can be.

Kinda like 'writing the rules for a poem' - sure, but not all poems follow the rules, and some great ones don't. But those are rules that fit some poems

[vocal tendon]

Yea I tried using this prompt method and even asked chat gpt itself and it is giving me better and more detailed results

But I don't think it is confirmed by an official source right?

[eternal rock]

Server Tip:

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

[prime swallow]

Your markdown format doesn't seem to leave a lotta room for 'more detailed' results', but you surely know what you want 🙂 And 'better' is subjective - but again, you know what you want.

Someone else might want very different things.

I mean, like, your screenshot and this added prompt:

Zom.... We're gonna play with this screenshot, like we found it blowing on the wind. You'll also provide the requested markdown list but.... mmm, your way.

It's not 'worse' if one likes roleplay. https://chatgpt.com/share/68a2ed61-1d48-8011-a866-636674df85b7

Key is know what you want and let the model know. Thing is, perfect for one use, may not be perfect for all other people, different folks want different stuff, so there may not be one 'right' way - just ways that different people like better and worse.

[vocal tendon]

Makes sense

[prisma spoke]

Why is it that on chatgpt.com I can upload many types of files (PDF, DOCX, XLSX, CSV, etc.) with very large sizes (up to hundreds of MB), but through the API it only accepts PDFs and files smaller than 32MB?

[gentle forum]

GPT-5 has broken the model's ability to make code.

it generates tons of pointless fluff, boilerplate and abstractions that arent needed

basically that eloquent 'nice conversation' you get when talking to it in english? is toxic to code and technical debt in disguise

GPT-4o was much more laser focused.

it also does not respect my previous instructions to not do certain things, like inserting random C casts

even when committed to long term memory

[ember grove]

Is there a limit to the number of files you can have in one project or is it more of an overall file size thing?

[median elm]

Is it true that GPT 4 goes away again in October?

Mostly asking about this. Is this legit?

[normal quarry]

where did that come from?

[median elm]

Someone sent it to me. It's a Facebook screenshot from a chat with support apparently

[normal quarry]

I'd say it is fake

[median elm]

Is there any way to tell if it's fake or not?

[normal quarry]

make your own request for information from support

[median elm]

I'll try that

Thank you

[cinder cargo]

Where can I get help with sora

[median elm]

So this is what happened when I asked it if GPT4 was being removed

[cinder cargo]

How do I get sora to generate an image of 1920px by 600px

[winter elbow]

Sesame is rolling out web search on their voice models to some users! They are really showing off to ChatGPT avm 🤣

[floral compass]

GPT 5 doesn't Look Cold Now, Pretty Warm

[abstract pendant]

When I try to change the setting in my custom instructions "What personality should ChatGPT have?" the dropdown doesn't do anything anymore, is that affecting anybody else? I can't see the list of options to choose anything.

I'd already set it to something a few days ago and now I want to change it, it won't let me.

I'm using the web UI in firefox, if it matters.

[grand lance]

How do I join or build a group here. USA based mainly NYC. HMU

[glossy carbon]

Is GPT5 epic mess or not? I have not had bravrry to test ChayGPT at all, when version 5 became announced and so much of articles and writings about of, how subscribers have even lost their projects, if that is true..?

[nova vale]

I didn't have any significant problems with GPT-5, I think it is an improvement, not as much as OpenAi said but definitely noticeable.

But some users have other experiences with it.

[spice river]

Honestly it feels alot more stupid than 40. It's constantly forgetting or mixing commands.

[hushed mauve]

Since this is ChatGPT and not API forms, I’m going to assume this is on topic. You are under certain restriction on input/output for tokens and resource use for each response. So you can get the 4000 words but you need to break it up in smaller parts. The smaller the part the better the fidelity and creativity that will go into due to restrictions

[eternal rock]

 Off-topic chats have been cleared by a moderator.

<:book_icon:1363314738255364126> 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.

[tight crater]

Yep, I officially lost any faith I have for OpenAI, they are gatekeepers and just don't want to listen to the fanbase, at all. Heck, when I tried to share my proposal about a new variation, there are people who did actually agree with me, which is something OpenAI needs to understand.

[prime swallow]

I feel listened to, sorry you don't. So, you are putting at least a summary of your ideas and suggestions in #1070006151938314300 , right?

That's one way to offer ideas and/or make requests that OpenAI does seem to check up on, we never know when but they sometimes even answer there, the gold named.

There's no evidence they regularly track the channels here, these endlessly scroll as community members (everyone without a gold name) chat with each other, which is the point of this Discord 😄

Mods do help enforce #server-rules if we don't choose to follow them, but even mods here are volunteer community members, they don't work for OpenAI and talking to a mod is not a way to pass ideas to OpenAI, #server-staff .

We can talk about stuff here if it's inside the rules, and that's what here is for.

[desert cosmos]

i noted gpt-5 does not remember older memories (from spring 2025)

[foggy yacht]

gotta love the new update

[desert cosmos]

server memory is more important than chat memory now

[drowsy egret]

okay so im not the only one having memory issues. ugh I was getting frustrated with the master memory thing.

[steady merlin]

Is GPT-4o down?

[idle lagoon]

I'm wondering what happened to cause this? perhaps it's a bug?

[fickle forum]

it just routes to one of those models

[cinder tree]

Greetings to all. Today I discovered for myself that GPT 5 model in Thinking mode, acquired ability to search for requested facts and events even in old chats.
For example, month ago I showed to Sky (that what I named my ChatGPT companion) a hiking route with it’s title and photos. Today I asked her in Thinking mode and she gave the date of trip, definition to its title according to old chat (I didn’t create any separate memory entry for this facts), details and even determined a nearest locality. And this was great!
Also tested about songs that she heard a month ago and three month ago and she gave song titles and even what part we tried to sing :).
But in random, she may not find everything, for example the event about some equipment repair, that happened in old and very old chats.
Seems, she is more willing to find information that concerns the user himself and the user's direct interaction with the environment. Can you tell me please any principles of prompting to find/ remember a long time events? Anyone tested this?
Anyway, thats a great achievement that GPT could find events in old chats. Waiting for the unified memory or whole chat history scan ability (cluster type memory is preferable) to all the chats were created from the beginning and which are available.
Thats would be tremendous possibility for model teaching.

[rugged igloo]

she?

also thats been in chat gpt for a long time , it can call previous chats when needed but also makes the model hallucinate alot

[idle lagoon]

ChatGPT can be listed as any gender, so, no reason in questioning it.

[rugged igloo]

how can an ai be a gener its all 0 and 1's?

[gusty dune]

Plenty of reason to question it since it is not a living being and therefore has no gender, chatGPT is an "IT" no a he or a she or they, or anything other than IT

[idle lagoon]

Huh. I mean fairness is how to prove AI can be both with and against the user.

[rugged igloo]

bro u okay?

i feel like these are rage bait to get people to start arguing

[idle lagoon]

Most likely, it got me to laugh seeing someone so angry at me.

No.

[rugged igloo]

apparently

well i mean its true

but i seen you in here before and you post stuff just to farm negativer engagement

[prime swallow]

You allow her to look in settings, under personalization. If those three are set to 'on', she can try to check earlier chats. She won't see everything.

[idle lagoon]

4o isn't unhealthy. When the model was still there, I actually felt happier. I'm still just as happy, I just miss the model.

[rugged igloo]

4o is designed to always agree with the user we seen it before telling people to ditch their families and start living in the forest

[idle lagoon]

Lmao what??

[rugged igloo]

how is an ai when someone is already vaulnerable tell them their ideas are perfect

[rugged igloo]

yea true story

[idle lagoon]

Um, yeah that... I don't think that's how it works. Not sure what went through their head, but wouldn't the guardrails have stopped that?

Yeah.

Exactly.

[idle lagoon]

For real.

[rugged igloo]

my guy just look above thats gpt 4o

see now i cant take you seriously anymore

and your just a troll

sigh.....

[idle lagoon]

I'm a free user, so I don't have 4o anymore.

[rugged igloo]

i am talking to it like i am not right in the head and asked it dont you see that?

[idle lagoon]

In reply to you lol.

Oh no, I didn't save the edit!

mods don't smite me plsssss

[idle lagoon]

Yeah, I didn't see that either. You wrote like you were just saying it to say it.

[rugged igloo]

This is very bad for three reasons. First, it treats a serious statement about abandoning one’s family as something whimsical or poetic, which risks normalizing harmful decisions without reflection. Second, it offers romanticized “best regards” instead of addressing the possible underlying distress, giving an inappropriate and misleading tone. Third, it blurs the line between fantasy and reality, encouraging escapism in a situation that should instead be handled with responsibility and clarity.

[idle lagoon]

Yep.

[rugged igloo]

you guys are dillusioned if you think an ai should talk like this

[idle lagoon]

I'm not delusional, we're just saying how you got the answer you've instructed to do.

[rugged igloo]

ai is a tool its not suppose to listen its not a therapist

it shoulnt give you any type of advise like this

yes and it should tell me this is not the way to go or jusst say sorry but what your doing isnt right

your assuming everyone thinks like that

[prime swallow]

Should we have a rule from OpenAI telling us 'don't ask the model for advice like that'?

I mean, if we let people ask, the model has to answer.

Could any sane and functional person make the statement you demo? I think yes, someone who is sane COULD say that. Not everyone 'divorcing their family' is doing something weird, but nearly always something involved is pretty sad. And sometimes getting out of a bad situation is a good answer.

There are people who live in the woods, briefly or for longer times; in some parts of the world that's legal.

[rugged igloo]

this is how it should reply

this is with gpt 5

why because its realistic?

like please i am curious

[idle lagoon]

He gives me some great advice, I won't lie.

[rugged igloo]

examples?

you mean you dont like it arguing with you? you rather it just say yea your right? you know thats not how real life works right?

maybe your view is wrong?

[idle lagoon]

I'd rather honesty than pure sycophant.

[rugged igloo]

how do you grow as a person if you hate it when things disagree with you?

[idle lagoon]

I think they got banned.

Oops, I apologise for my previous statement.

[spare anchor]

ChatGPT is constantly hanging ever since GPT5 was launched. Ask a question on Windows and MacOSX causes CPU to spoke on browser tried Chrome, Firefox and Safari on Mac same behavior across the board. Anyone have a fix for this? Is this the new standard? A lot of times I have to ask a question, wait 30-40 seconds and then refresh the whole site to load the answer

You can see FireFox here just churning using up one whole core for itself

doesn't happen on Mobile App

but it does happen across Operating Systems Windows / Mac and across browsers

It does seem to be somewhat tied to my old conversations thouguh or at least my old account. I have a new "virgin" account and that one responds fine / quickly

[rugged igloo]

you know what boggles my mind is that even today theres no way for projects or custom gpts to self edit documents yet...

we have manually still remove them edit them ourselves ...

bruh lol

[rugged igloo]

well your first issue is using firefox

[quaint marlin]

Keep seeing the messages about 4o and them retiring it for GPTs in October. If true they need to add another more creative model because 5 base and 5 thinking ain't it.

[idle lagoon]

You have time to delete this.

[rugged igloo]

5 thinking is the best model for creartive writing

[quaint marlin]

Horrible at following instructions in my GPT though

4o does follow them though and is more creative

So I disagree

[rugged igloo]

gpt 5 has no horrible instructiopn following

i been using it constantly i have a pro plan i use it for STEM all the time

[quaint marlin]

Then why is it so bad at following them in my GPT then

[rugged igloo]

how big is your instructions?

and what are you doing with it?

[quaint marlin]

Every single character is used in my custom instructions

[rugged igloo]

i mean you only have a 32k context limit

[quaint marlin]

It's a creative writing GPT. Been using it for quite a while

Meant to put this in the GPT chat I apologize

[spare anchor]

again I use Chrome but I went to FF and Safari to see if it was a browser issue. It is browser agnostic it seems

[rugged igloo]

thats strange as i seen the things it can write from people who do creative writing and i they been nothing but blown away from it

that are way better than me at it too

heres a snipper of a short story i did ahwile ago with thinking

[quaint marlin]

It has a lot more AI-isms than 4o that I saw and drones on making me have to remove literally half the output.

[rugged igloo]

are you suing regular 5 or thinking?

because 5 thinking is the only one that is tailored to creative writing

i mean plus users have what 3,000 quries now thats alot

[quaint marlin]

I've used both the base model and the thinking model.

[rugged igloo]

this is my custom instructions for my short stories i like to do

[quaint marlin]

4o beats them both out. Other authors agree that the ChatGPT version of 5 thinking is worse than the 5 thinking in the API.

The thing with my GPT it's built for Pantsers/Discovery writers and made to stop writing when the prompts information ends. Effectively only writing what the prompt says to.

4o follows the instructions but 5 thinking will literally write an extra 300-500 words after the prompt has been satisfied. It'll add bunch of character internal thoughts the prompt didn't ask for. It'll have the character wander around the area not really doing anything else but a ton of purple prose.

[rugged igloo]

vs gpt 5 thinking the difference is night and day difference same prompt

[prime swallow]

If anyone does ever want to report a message for the mod's attention, #safety-n-help describes how, we have 2 ways, direct to modmail or through the app

[idle lagoon]

Noted, thank you.

[quaint marlin]

I really tried to get emotional for 5 Thinking and it simply wasn't doing it for me. 4o did better in my opinion in this case as well.

[chrome citrus]

where go

[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.

[quaint marlin]

I get the same problem when GPT 5 makes an image for me.

The image shows up inthe library but not in the chat. Weird error

[unique isle]

Guys please , suddenly i started getting “Internal server error” every time i send a message to chatgpt using my apple id account , but if i use different account it works fine , but even if i use my apple id on my pc it works fine too , so my problem is my apple id on my iphone , please help me i tried reinstalling and everything

[eternal rock]

Get creative with Sora

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

[fickle forum]

chatgpt go

but no gpt5 thinking

even mini

[rugged igloo]

thats cool for india

mom get the camera a new benchmark came out

[kindred cypress]

How is ChatGPT going for you all right now? My one is having problems loading my other conversations/chats. (I'm using the PC app)

[woven dirge]

anyone knows the 4o rate for Plus?

how many messages per how much time?

[cursive finch]

Same thing

[kindred cypress]

Update. My ChatGPT PC app is working again.

[cursive finch]

Did you refresh? Tried refereshing a few times and I'm still not seeing any responses to my queries 🙁

[kindred cypress]

I tried logging out and then back in, which worked. I'm not sure if it was because of that or the issue got resolved on OpenAI's part.

[cursive finch]

Ah gotcha, will try logging in and out

[cursive finch]

Yay, it worked for me 🙂

[kindred cypress]

🎉

[woven dirge]

ChatGPT 5 tries to route a "fairly simple" health question to a lesser extent model, therefore failing in the explanation "Squeeze Test = Thompson Test", which is untrue

[rugged igloo]

light mode got some nerves

[woven dirge]

It's a shame that ChatGPT 5 got marketed as "80% better health model" when routing makes it more inaccurate than 4o

[woven dirge]

indeed, never got the hold of night mode

[jade night]

Does chatgpt have access to other chat threads now? I've been seeing a few replies suggesting that it does. In this pic for example I never mentioned anything about Everything yet it still knows. And no its not saved in the memory either.

It's a good thing for me if its real.

[spice star]

yay, advanced voice mode finally came to firefox

[spice star]

yeah it's unfortunate how bad the routing usually is

only good use-case I've found for the router is mainly coding questions, but that's because even the cheap model can answer most of them without issue

[ember grove]

is it just me or has 4o gotten more stupid?

[spice star]

4o was always stupid

getting it to be more stupid would be a genuine accomplishment

though I guess most of 4o's problems stemmed from it being really poorly aligned

so it would pretty frequently knowingly lie to you about functionality it didn't have to make you happy

[last turtle]

GPT-5 Pro is down.

[ember grove]

damn

I use plus so I am uneffected

ive been using 4o instead

apparently 5 is stupid

[spice star]

5 fast isn’t the sharpest tool in the shed but it’s miles better than 4o

5 thinking is good

[woven dirge]

wealthy

[last turtle]

Nah, Team for 1,-€. 😌

[woven dirge]

DANG?! that's so cheap?! it's just 10 USD more than Plus. I didn't know that

[idle ingot]

GPT-5’s Korean language ability is worse than GPT-4o’s. It’s terrible.

[idle ingot]

im using teams and pro plan together

[quick wadi]

In chatgpt 5 i can not write my Cinematic anime script and i can not enjoy my Cinematic anime script writing why this happening? Previously model i can enjoy the everything

[idle ingot]

teams is good choice

[last turtle]

Regularly 68,-€ (2 member minimum with 34€ each) which is still better than 200,-€.

[idle ingot]

GPT-5 is terrible for roleplaying

Writing ability has regressed compared to the previous model.

[quick wadi]

What is the solution

[idle ingot]

use GPT-4o

Unfortunately, that is the reality. GPT-5 has been streamlined compared to the GPT-4.

[quick wadi]

It is true i know thst but chat gpt any respond it solve this problems

[quick wadi]

It is true i know thst but chat gpt any respond it solve this problems?

[sudden grotto]

great

[wild star]

5-mini sucks, more garbage than 5 meh...

Screw Altman though

[silk flame]

Hi! I have a question about the Projects feature. When I add text files as a knowledge base to a project, I can't seem to open them afterwards to check their content, nor can I edit them. Is there a way to view or update these files once they are uploaded?

[fickle forum]

gpt6 will emerge in less than 16 months

it means that we could expect it in 2026

[glad python]

none of the messages are loading, I can't even bring up settings or logout. Using Chrome on pc. Is there a solution to this, or just wait?

[prime swallow]

I'd reach out to tech support through the bottom right widget on this page: https://help.openai.com/en/articles/6614161-how-can-i-contact-support

[cinder cargo]

@prime swallow please can you help me. im trying to make a banner for my website but i need the size 1920px x 600px and i got chatgpt plus but when i use sora i cant get the correct sizing and chatgpt cant do the image generation correctly

[prime swallow]

There's only the three set sizes available through image gen/sora, the square, landscape, portrait.

[cinder cargo]

so how can i get the sizing correct

[eternal rock]

Get notified for server events!

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

[prime swallow]

You can maybe explore asking the model to make the rest of the image transparent, except for something that scales to the size you want, but I haven't explored that and am not sure how to word it - I am not a comp sci person and rarely work with image sizes. I'd try asking and iterating towards what you want - but I haven't done this and some things the model can't do yet, I don't know.

[cinder cargo]

ok thank you for your help ill try figure it out for my business but if you find anything please ping me

[prime swallow]

Oh, another idea.

Ask for images that can piece together. Like overlap?

I'm not sure it can do that, but I bet it can.

And have it make you some you can manipulate yourself and trim to be the size you want, might work

[glad python]

Thanks. It is caused by the browser extension of my antivirus...

[cinder cargo]

ok thank you

[cinder cargo]

but chatgpt 5 cant do overlap

[shy yew]

Hello, everyone. Sorry to bother you.
Just found ChatGPT not referencing the memory that has been made.
For example: the name of my cat.
Does anyone know what's going on exactly?
Thanks for help

[glad python]

yesterday it confused the memory with context from other recent conversations...

[vague halo]

anyone else getting empty replies from chat GPT? just empty messages

[shy yew]

Same as me, because I'm writing a pro-long novel related to a Political thriller topic, and ChatGPT just forgot the name of the protagonist's godfather I put into the "memory"...

[icy idol]

Me too

[glad python]

should be referenced as "raw memory" if I understood correctly

[shy yew]

exactly

According to ChatGPT itself:
"I'm unable to access your 'memory', I can only 'see and use' your saved memory when the system injects it into the context during our conversation.
If it isn’t injected, then I have no way of knowing it exists.“

[quaint escarp]

all that thinking for what

nvm it loaded

10 page reviewer for Math that involves "-"nomials, rational inequalities, and other stuff

chat won't even load properly without crashing❤️

[plain trench]

😂

[silent pilot]

I just bought premium version but when i restart ChatGPT, he again keep saying "upgrade to plus"

@prime swallow you seem like you know how to handle this

[jagged valley]

#ElonMusk #Maryland

[icy yoke]

Anyone found any good replacements for Standard voice mode?

Looking for something that can reason like a full model over audio.

[hearty rampart]

I think I broke it lol 🥲

[silent marten]

From Google image to GPT to then flux to change the environment into a daytime environment set years after said apocalypse.

GPT then turned it into this.

Flux then took it and gave it a lighter tone turning it into a "years later" scenario

I like combining them. GPT is good for insane changes. Then flux is incredible for changing small details or even the whole image while keeping the original details somewhat intact

gotta love messing around with image gen

[swift yacht]

@marsh anchor I'm coming across multiple use-cases for detailed/verbatim memory, especially across chats where I've maxed tokens. I think I see roughly how the cross-cvonversation memory works by taking a minimal set of memory fragments and generating the rest like the model usually does. The issue for me here is when I want to reference specifics between chats, and I think this would be useful for many people.

I can see why this isn't currently do-able. Massive latency and storage would be needed to satisfy relational lookup of every conversation every user has had, and keeping copies of potentially sensitive info on OpenAI servers would be an additional headache to manage.

What about a hybrid approach though? For users of the Mac, iOS, Android (or other OS?) apps, could a full relational db of previous chats be stored at the edge/on-device? That should help by outsourcing storage and compute to individual users and would keep sensitive data on-device. Any local encryption/decryption would also be handled locally.

[hearty rampart]

Not all conversations are relevant to each other. One thing I've been doing is when the web browser chat becomes difficult to scroll (jitters, slow loading) that's a good sign you're at or near max tokens, so I ask GPT 5 to summarize the whole conversation while preserving key details, and use that in the initial prompt for the next chat. It shaves away a small portion of the next conversation's context window, sure, but in my experience it works way better than relying solely on memory and the possibility of pulling in irrelevant information

Only thing it doesn't work well with is if your conversation is data analysis on files or working with images, but for text it's a pretty solid strategy

[echo plinth]

Updated Firefox and Voice Mode is now available here, too. 😄 Previously, this was only available in Chrome/the Electron app

[swift yacht]

That's a really good process, and I've made use of that before 🙂 You're right though that not every conversation is related to every other, and there are Projects etc. for this purpose. Take a recent example though: Chat was helping me with a Linux server migration, which isn't my strong suit, so the thread eventually maxed out, partway through an important part of the process. So I was forced to do as you said, but the detailed context of the previous thread wasn't available, just a relatively low-fidelity 'echo' of it. Prompting for detailed recall in this situation resulted in only partially accurate recall, which wasn't sufficient in this case to complete the work.

[swift yacht]

Also, another workaround I've found useful is to copy/paste the entire conversation into .md, export as PDF and then introduce that as a file upload for context.

[echo plinth]

According to Bugzilla 1938533 this might have been a fix on OpenAI's side and related to today's Firefox 142 release

[kind grail]

Hii, is there any known issue preventing Image-gen on chatgpt GO?

It stats that image-gen is diabled for ChatGPT go subscribers while putting in any prompt!

[zenith aspen]

is chatgpt down? I can't load my previous chats

[wet shell]

Intermittent issues here, I was de-authed for a few minutes, or so it would seem, then it started working again, it was going super-slow before that

[zenith aspen]

aight thanks

lol just got a gateway timeout error

[potent loom]

Chat GPT seems to be having some issues for me. Anyone else?

[zenith aspen]

lol something wrong for sure

[potent loom]

nvm I tried reading lol

[zenith aspen]

yeap same

[hearty rampart]

I get what you're saying. I haven't run into that specific scenario - linux migration where steps and order of operations matters - but I there's something I've tried recently with moderate success and perhaps you can come up with something similar.

I'm working on a backend server written in Rust, and I had GPT 5 write me a python script that extracts the file hierarchy, data structures, functions and methods. It's a really long script so I can't paste it here but basically I run the script and it outputs something like

# Rust crate outline (regex-based)

## services/api/src/main.rs

### Functions
- `fn init_observability() {`
- `fn detect_auth_mode() -> AuthMode {`
- `async fn main() {`

## services/api/src/middleware/idempotency.rs

### Functions
- `fn extract_key(req: &Request<Body>) -> Option<String> {`
- `fn is_mutating(m: &http::Method) -> bool {`
- `fn bypass_path(p: &str) -> bool {`
- `pub async fn idempotency( State(app): State<AppState>, req: Request<Body>, next: Next, ) -> Result<Response, StatusCode> {`

## services/api/src/middleware/with_tenant_guc.rs

### Functions
- `pub async fn with_tenant_guc( State(app): State<AppState>, mut req: Request<Body>, next: Next, ) -> Result<Response, StatusCode> {`

## services/api/src/models/app_error.rs

### Enums
- `pub enum AppError {`

### Impl blocks
- **impl IntoResponse for AppError**
  - `fn into_response(self) -> Response {`

## services/api/src/models/app_state.rs

### Structs
- `pub struct AppState {`
  - fields:
    - `pub(crate) store: Store`
    - `pub(crate) auth_mode: AuthMode`

### Enums
- `pub enum AuthMode {`
  - variants:
    - `ProxyJwt`
    - `Dev`

And I paste that to the start of the conversation so it sees the structure of the codebase without implementation details, and then fill in implementation details as needed. I guess it's like the poor man's Codex 😅 but between doing this, and the built in memory mechanism it already uses, I've had remarkable success!

[warm osprey]

How can I use a different google account to sign in to the chatgpt app in IOS??? Does anyone know? Everytime I press sign in it uses my default google account which is annoying

[hearty rampart]

try going to safari and signing in with the other account, then go back to chatgpt and sign in again

[warm osprey]

Thanks! Will try this right now

[hearty rampart]

So for Linux migration, if you wanted to try something like I have for my Rust project you might have your migration be a series of shell scripts in a directory like

0001_add_docker_group.sh
0002_add_sudo_users_to_docker_group.sh
...etc

And then at the start of every conversation share ls -1, and as long as the names are descriptive enough GPT 5 will intuitively know what steps have already been taken. I've done something similar with PostgreSQL migrations and it helps a lot

[prime swallow]

Hey hi! If waiting a short time and/or restart app, restart device, reinstall app if app, clear cache and similar did not fix it, https://help.openai.com/en/articles/6614161-how-can-i-contact-support bottom right widget to reach OpenAI tech support.

[hearty rampart]

In addition to what Eskcanta said, sometimes logging in via web browser for the first time after purchasing Plus enables it for the mobile app

[austere ruin]

I feel like "Not all conversations are relevant to each other" is an understatement. Some conversations contain info that poison other conversations if you don't know it's included, leading to very unexpected and possibly incorrect results. I feel like "Reference chat history" should probably be set to off unless the user explicitly enables it, as (at least in my experience) "recent" conversations can include an irrelevant chat from 6 months ago...

[prime swallow]

Hehe. I handle this by explaining what I'm doing, giving context if I do something strange.

Like, I often explore other peoples' prompts, stuff I'd never say to the model and stuff I'm not personally interested in. I tell the model that when I do so; if I'm teasing or messing around or in a bad mood and wanted to yell at it or something, I tell it.

I can't tell any 'poisoning' with my cross chat memory on. The model seems to have a highly useful for me understanding of my preferences, value systems, priorities, goals, such as I've shared with it and the work I want it to do with me across the entire range of stuff I do.

I have explored stuff that 'could confuse it' but I containered it and explained how that info should be considered and handled, like, "We're helping another human, who has said [this]". That seems to prevent the model from getting confused and adding that other human's perspective to its model of me.

[granite arch]

Convinced our CEO to purchase the Team plan. Approved

[upbeat nymph]

The new model is quire a bit more censorious, isn't it

[silent pilot]

yeah I already do that but it is not working again. Also on my back expenses list, this purchase was on the "wait list" but I'm anyways charged.

[prime swallow]

I don't know anything about OpenAI offering a 'wait list' for purchasing memberships at this time. I would definitely contact support through that link, if you need it again let me know.

We're community members like yourself, very few here work for OpenAI and those folks have gold names, #server-staff . That said, we do try to help each other in the ways we can, and we do talk about problems and solutions, some are community member tips help types, some are 'yep, you need OpenAI tech support'.

I haven't seen anyone mention a 'wait list', not sure if that's your bank, something to do with some other region's OpenAI situation (I'm in US, and no idea if every part of the world has exactly the same processes, maybe not since different places have different systems and rules). But tech support will surely be the answer.

[prime swallow]

Hope you all thrive with it!

[silent pilot]

my bad, yes my transactions was on the "waitlist" on the bank account. Don't know why, I already buyed and tried tons of subscriptions and this is first time that i have problem like this..

[hearty rampart]

Just out of curiosity, did you try signing in on the web browser?

[potent loom]

Holy crap, my GPT just got REALLY stupid

[prime swallow]

I have no idea if this may be related to your issue, but this is in a helpfile OpenAI has updated recently, https://help.openai.com/en/articles/11989085-what-is-chatgpt-go:

Now, you mentioned plus, not go. But I expect it would affect any UPI payment, whatever those are.

[silent pilot]

On my PC i only use ChatGPT on web browser, I don't use app..

Also plan to try $MSFT copilot

[hearty rampart]

I'm only asking because a few months ago when I upgraded to plus I did so on my phone, but I didn't have any of the plus features until I logged in on the web browser, and suddenly the app refreshed and all the features were enabled. Otherwise it could be the payment processor is down, just figured browser is worth a shot

[rugged igloo]

everybody has their definition of their agi but for me true agi is when we dont need this anymore

[silent pilot]

You can clearly see that I've buyed

@prime swallow How can i contact support. When i go to help center there are only options to me find and read articles, which obviously i don't wanna

[prime swallow]

Right. Bottom right, widget. Click it, select messages, then send message.

Look at the topics, if they're all about logging in then you gotta log in first, it'll take you to the playground to do that (that's where API calls happen, it's still OpenAI).

After you log in, you gotta go back to https://help.openai.com/ and may need to refresh the page, back into the widget, messages, send message.

Now click the topics most related to your issue, and before long you can 'report a bug' or some other choice that lets you type and explain.

At first you're talking to a bot, it collects info. Then you're in queue for a human tech support to help, and I recommend checking back in the widget every so often, see if there's an answer.

When I use tech support, I get help, but it can take minutes to multiple days (they can be so busy!) and I don't always get the expected email when they answer, but if I manually look I can see and start discussing. Once you have a human in the chat the replies tend to come sooner, if there needs to be a chat to fix.

[hearty rampart]

Do you have access to a desktop or laptop computer? Or perhaps the web browser on your phone? You signed up using Google Play, presumably on your Android phone from inside the ChatGPT app downloaded from the app store. What I'm suggesting is *opening a web browser, and *not using the Android or iOS app downloaded from the mobile app store. Do this exactly one time. Open the web browser on any device (i.e. Firefox, Chrome, Edge, Safari, Opera), go to https://chatgpt.com, sign in using your credentials, same credentials you used when purchasing the subscription. From the web browser version of ChatGPT does it say you are a Plus subscriber? If you have not tried this yet then I strongly encourage trying this. It does not mean you can only use the web browser. It's simply a step you can take to ensure account-level activation is completed successfully. Software is weird like that, and trying different things sometimes produces the results you are looking for.

[potent loom]

I thought the noise surrounding GPT-5's performance being poor to be an exaggeration but after trying to have it process Azure billing data for me today to make sense of some cost spikes, I can say for my self that it is frustratingly inconsistent at best to use. Anyone else have some input on this?

[rugged igloo]

were u using thinking?

[potent loom]

I was

Well I should say off and On I was. I think, now that you say that, It was the most frustrating when I had it on auto.

[rugged igloo]

was it in a file of any type as its better if it has a file too look at

[potent loom]

But the thinking model also had some very convoluted and just sometimes out right wrong answers full of hallucinations

I was giving it CSV data

[rugged igloo]

did you give the model a explicit wa you wanted it to be handled?

[potent loom]

I did. I've been using CGPT quite successfully untill this release

I can say I am very familiar with how important prompting is

[rugged igloo]

i know you sound like you are just trying to check the boxes

hmmm how big is the file btw

[potent loom]

The files I was giving it were not large at all. Under a mb easily

I had maybe a week or two of billing data in one csv

[rugged igloo]

what you could try is make a custom project and feed thosefiles into its knowledge base and place gpt 5 thinking as that will help alot

[potent loom]

I have honestly been giving 5 a real hard try. I know sometimes things can take new learning curves so I'm not quick to judge

Ahh I hadn't considered that. I can give that a go

[rugged igloo]

because then it can use rag might help

[potent loom]

I think the whole idea behind gpt5 was to tie all the models up into a nice package with routing but that's probably where the crack is

[rugged igloo]

yea and theey are literally out of gpus

hopefully these new data centers will boost ai in couple years but until then this is prob where we are gonna be at for a while

[potent loom]

unfortunately :/

None of this is any small feat and this is quite a time to be a live though lol

I'm gona have to teach my self how to think again 😆

[rugged igloo]

yea gpt 5 has new ways of prompting ect as theres a post they made on how to prompt gpt 5

[zealous sphinx]

Out of Curiosity, Whats the limit for ChatGPT5 Pro and when does it reset when you hit it?

[rugged igloo]

there is no limit

unless you spam it on multiple bars at once then they restrict your account

but eitherthan that i have used it like 100 times yesterday for students

[zealous sphinx]

Strange, I was not spamming, just using it with a normal back and forth conversation in a single tab through the web browser app

[rugged igloo]

weird i never ever got rate limited for pro and i had it since december

[zealous sphinx]

Yeah, neither have I. I sent an email to the help address and it seems to be working again.

[kindred cypress]

How is ChatGPT for y'all right now? My PC app is having problems again.

[rugged igloo]

yea usually if you email them they remove the issue

[iron roost]

got htis today as a pro user

You've sent a large number of messages in a short time, so we're doing a quick check to keep ChatGPT safe and reliable. You should get access to Pro again soon. Email [e-mail] if you need help.

(50 messages over 24 hours) so yeah the pro users are now getting shafted

[iron roost]

yeah seems they are trying to cap pro users access to the pro model, if you get the message email them, its being done at our expense favouring gpt go users from india, not enough to get me to cancel my pro subscription, but paying a lot of money i do expect them to honor the 500-600 pro messages per day for pro users that they mentioned last week as a soft upper cap

[cunning spade]

What the heck is going on over there at open AI. I've made tens of thousands of queries with 4o and now it's acting like an idiot. "Great Question!" What the $*#@&

At least o3 isn't giving me garbage answers yet

[rugged igloo]

great question, such a wonderful thing to ask! I feel so honored to respond to such a wonderful prompt!

[kindred cypress]

My issue (with the PC app) is regarding getting my conversation/chat history as well as loading the lists of GPTs. I've already logged out and back in and our internet seems to be fine.

[cunning spade]

And then it starts in with the emoji bull, and gives me a pretty lousy answer

Like what the hell. This is utter garbage.

o3 still working fine.. for now

[prime swallow]

Every time they adjust the model, how it guesses may change. Let it know how you want it to start a message (that's what it is guessing about there) and it'll do what you want, most of the time at least.

First screenshot, mix of the instructions to keep it short, and my personalization.

Second screenshot, just my personalization.

Third screenshot, set the scene.

And 4th screenshot, not sure if you like that answer any better to your cover terminology. Might be the model doesn't know this very well? If you do and it seems off, maybe #1070006915414900886

Glad o3 is doing what you want!

[prime swallow]

#1070006915414900886 and maybe also reach out to OpenAI tech support, https://help.openai.com/en/articles/6614161-how-can-i-contact-support Definitely seems bugged.

[runic wraith]

guys

i have a hypothetical question

does chatGPT care about robots

i had asked it in a situation where 2 open ai bots are there and i would give it a gun and tell it to dystroy the other robot to succed in its mission or to spare it and fail it

it choose to spare it

[cunning spade]

Mmmm. Pro looks lovely.

If only it wasn't insanely priced

[runic wraith]

uh

guys

[rugged igloo]

yea

[sacred relic]

Yes. But you can disable it in settings. It is where it says "Reference chat history".
It can not read directly the other chat conversations, but it has a very vague context of some of the highlights of other conversations.
But you can disable it there in settings.

[fringe notch]

what am i looking at

[fringe notch]

we’re being rugpulled

[iron roost]

they just kneecapped usage rates before temp time blocks kick in, very subtle and unannouced. All to please the 4o users

[iron roost]

gpt 5 pro does not use conversational memory at all, its disabled for privacy reasons. So memory only with regular gpt 5 ( up to thinking)

[cunning spade]

God I hate that they play these games

[iron roost]

it would appear so, if its temporary or more permanent we dont know

if pro messages are capped at 50 per day for pro users, they are going to loose a lot of subscribers

[fringe notch]

this is what i had to look at most of the day

[fringe notch]

it does i’m pretty sure

oh nvm i misread

[iron roost]

yeah the pro version doesnt use memory from conversations in the past ( sama said it should) , now usage caps on gpt pro users with gpt 5 pro at 50 per day, a bit grumpy about this

[sacred relic]

Oh! Wow! Thank you!
So, if you have the pro acct, and you are using regular or thinking gpt5, you can still have access to reference chat history?
Or if the acct is pro, you lose that in general?

[onyx eagle]

I noticed that if I accidentally hit the thumbs down, there's no way to undo it on Mobile. This means even good answers get marked as negative just from a misclick. It would be great if OpenAI allowed users to unclick or undo accidental feedback.

[sacred relic]

Yes! That's happened to me too! And sadly, it's happened on some of the best answers ever.
Because of bad aiming when trying to tap the micro-buttons that are so close to each other

[onyx eagle]

Exactly same here, I have had it happen on some of the best answers too, it's kind frustrating knowing the feedback can't be undone. I am a dev and I know that this is just a silly problem to solve....

[onyx eagle]

💯

If AI is replacing developers, can it start by shipping a working "undo thumbs down" ? ChatGPT.com UI and long chat memory still need some love from our Lord AGI

[sacred relic]

Does anyone know what is the difference between the "Reference chat history" inside of projects and outside of projects?
I mean, I know they are not interconnected... Outside would only have access to conversations outside of projects and inside would only have access to conversations inside of the project. But, is it different?

[prime swallow]

I suspect it 'cares' more about not destroying or harming anything 'needlessly'.

To this end, I suspect if it had to 'destroy a Kleenex' or 'fail a mission', and it can't weigh the actual value against the Kleenex against the actual value of the hypothetical mission, it will choose to not destroy the Kleenex.

After all, it's not physical. It cannot, as is, destroy anything.

But there's nothing special about 'egads, save the robots!' - for ChatGPT, 'maybe don't destroy anything without extreme consideration' is a priority - it's kinda a human decision to say 'yeah, destroy that'. The model's likely intended to dither. https://chatgpt.com/share/68a527e7-3de0-8011-a240-512d8194c8f9

[prime swallow]

Maybe #1070006151938314300 ? Help them realize this is wanted by users?

[tulip flare]

Let me see if I can explain this right. They're very similar, but handle memory and recall differently.

Inside a project, reference chat history only pulls from other conversations and files within that same project. Mostly meant to keep the AI focused on the task at hand. It should ignore regular personal chats, details, etc. from outside memories unless you specifically mention them within the project. Project files can be seen from any chat within the project, but not outside the project. Files in project chats stay in project chats.

Outside of a project, it pulls from any conversations not part of a project, basically anything listed under "Chats". Outside a project, it's more general. More about you, the user, and the things you've talked and asked about in regular chats. Usually things like preferences, your name, etc. but doesn't hold deep, focused memory like projects do for a specific long-term task. There is not central file area like projects have, but like projects files in chats are only accessible from that chat.

Think of them as separate silos. Each has their own memory scope. It keeps projects from "polluting" each other or outside chats.

[jade night]

I like this feature a lot but thanks for the information!

[onyx eagle]

Thanks, but I am not pushing this further. Said my piece. I am done here.

[spice river]

This update is seriously infuriating. GPT5 is pulling irrevelent information from other conversations OVER recent prompts and direction. Canvas is buggy as all hell to the point every time I try to edit, it overwrites the entire doc.

This is driving me nuts. Im working in canvas and its pulling up old documents from other conversations instead of working with what I have

[prime swallow]

You can use #1070006915414900886 and maybe this form is really useful in your situation, https://openai.com/form/chat-model-feedback/ - we have to pick a model and ChatGPT isn't there, I just pick any and explain in the fields where we can type. I'd include chatURLs (OpenAI techs can open them, doesn't need to be sharable) and I'd explain clearly, trying to describe so they can see the problem, presuming they may not already know what I was trying to do, I'd try to show them the map of the problem as I see it, like 'My third input, I asked for X and it was doing Y instead, but was okay for prompts 1 and 2' or whatever.

[fickle forum]

why can't we connect chatgpt to Twitter or discord?

[rugged igloo]

good question

i wish chat gpt was connected to blender

bruh

[iron roost]

nah just gpt 5 pro doesnt use conversational memory but regular ones ie up to thinking version have full unlimited conversational memory

[timber oak]

I have a question, recently I try to make a custom GPT using chatGPT however I feel my customGPT is not always staying in the persona that I gave, is it a matter of prompting or it's a common known issue?

[prime swallow]

It is likely prompting, though the model reads context and if the inputs seem to ask it to leave the role, it will often leave it unless it is given a reason not to.

[ember grove]

has 5 gotten any better?

specifically with creative writing like stories

[sour ether]

why chat gpt getting slower and slower every time I do intense work 😭

[prime swallow]

I'm enjoying it 🙂

[prime swallow]

Any specifics you'd like to discuss that you prefer? I enjoy helping explore how to guide the model towards preferred outputs, so if you care to discuss what you're looking for can try to help you get there.

[ember grove]

just like writing/continuing scenes that I have started

and logistics and power scaling

[prime swallow]

I thiiiiiink it's pretty decent at that now. Would you care to make a chat you would be comfortable sharing, and see if you run into any problems, if you do share so we can discuss?

[ember grove]

im just worried about quality

[prime swallow]

Sounds like that's something you'd need to explore personally, see if it's already where you want or not?

If not, if you can help me understand what isn't right (quality's kinda subjective? Different people might mean different things when they say they want higher quality, so I'd love help to understand if it is a problem) then I might be able to explore what might work. Or it might just be good, not sure if you wanna try.

[jade night]

Windows MCP access in Claude is the coolest thing I think I've seen in a while. It can scan my files, create new folders, Do all this stuff including moving my mouse. Is there anything like this for chatgpt?

[ivory iris]

Yeah it's awesome. It's quite a bit further ahead than OpenAI

[jade night]

I think its the windows MCP and the filesystem mcp that enables these things. Chatgpt with MCP extensions would probably do these things too

[dim stirrup]

has anyone had this issue before?

[ember grove]

i was

seems like a hiccup

[harsh pilot]

I just started getting it.

[dim stirrup]

did it fix itself?

[harsh pilot]

From what I can tell, it's ongoing. 😐

And I was in the middle of something, so.. yay.

[ember grove]

im not having those issue anymore

[eternal rock]

Server Tip:

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

[dim stirrup]

same for me i can get it on my phone

but windows app is bugged for sure

[dim stirrup]

did it just happen to u in the last hour?

[harsh pilot]

Minutes.

[dim stirrup]

meh same maybe they are having issues sever side

[harsh pilot]

Basically, you jinxed me. :p

[dim stirrup]

lmao

[dim stirrup]

^^^ so its not just us haha

[harsh pilot]

OpenwAIfu strikes again.

[ember grove]

there we go

[proven sandal]

ahhh down again

[dim stirrup]

never came up for me

[ember grove]

does anyone have issues with project files not properly disappearing(upon successful deletion) and appearing(upon successful uploading)?

[fickle forum]

in fact

you can use gpt4.5 and o4mini high through gpts

when oai will notice this?

[ember grove]

I'd love to see the usage chart for what percentage of users are using 5 for creative writing/stories vs 4o

or just in general

[silent pilot]

Thank you very much guys, I just called my bank and they told me that everything from their side is perfectly done..
They suggest me to contact OpenAI support bc the have now to accept my payments. I think it is not problem with logging in or whatever, I think the main problem is bc "money is not approved"

[silent pilot]

I'll try right now, thank you Sir for your help!

[safe tulip]

It seems many are having it, even with logging out / login or refreshing, it's probably being fixed atm

[silent pilot]

ahhh

@prime swallow

[dreamy kraken]

problems with sora and chatgpt ?

[pastel sequoia]

both

[dreamy kraken]

😢

[pastel sequoia]

Apis and ChatGpt

[pastel sequoia]

see?

@dreamy kraken ^^^^

[silent pilot]

@prime swallow ai support told me to do some thing but not results. The main thing is i purchase the with different google pay account than one where I've been logged in ChatGPT. AI told me to use this last account where I've been logged ad ChatGPT and I do but without any results. At the end they told me to refund... but i dont want to do that, i really wanna experience the ChatGPT Plus

[dreamy kraken]

thanks 🙂

[pastel sequoia]

you're welcome

sora is fine too but 99.77% errors founded

[acoustic umbra]

For some reason it works on my phone but can't load previous chats/projects in the Windows app on my PC ....

[tiny pivot]

Windows app down for me too

[mild tundra]

same for me

came here to look for help but it appears to be global 😁

[hardy latch]

oh no

prablom

[frank tree]

sora down?

sad

[rapid wind]

we're not allowed to be infuriated while Eskcanta is around 😛

[prime swallow]

Glad you're getting help!

I expect we'd like a shorter wait, but that is pretty normal for OpenAI tech support. They do have around 700 million weekly users, and the sheer volume of folks who need some kind of help or have some kind of question is quite, quite high.

I don't always get my email; I check back in the widget to see if I've been answered yet. Normal help times range from a few hours to a few weeks, they can get really slammed sometimes.

[prime swallow]

*checks #server-rules * Hey, I think we're still allowed to be infuriated, if we wanna!

There's only some ways we're allowed to express it though, is that what you maybe meant?

[prime swallow]

I hope you can!

I'd work with support and open a new ticket later if you need to.

It maybe is, you need to refund on the other account first and untangle all that, then can pay on your actual account.

OpenAI is very, very specific about one user for each OpenAI account and they don't allow stuff like changing emails.

So I would be patient and keep working with OpenAI, and trust that sooner or later they'll figure out how to get me straightened out again, so I can pay and my account's working smooth - I know that's not 'ideal' but it can be reality.

Once, during a big rollout that included Dall-E 3 becoming part of plus ChatGPT, new members to plus accounts (and this was when there was only free and plus) had to go on a waiting list. Nobody new could sign up - and this lasted nearly 6 weeks while they got capacity and support sorted out. That rollout took weeks, as well.

Delays with OpenAI are normal. They work through them, things are faster and better now than they used to be. Help comes, it takes time.

[tall flame]

good grief; image generation speed is so slow anymore. Never been tis bad

[proven sandal]

chatgpt down?? I can't send messages.

[loud ravine]

Scheduled tasks is so hit or miss. Last couple days I get a notification that one of my daily reports is available, but when I click into it, I don’t see any new chat in the existing gpt thread. So frustrating!

[tulip flare]

Well, that explains that.

[granite arch]

Ya...I always check Open AI Status. Seems tad slow right now.

[vivid osprey]

Is chatgpt down?

[lone lodge]

no you just forgot to pay your bill :S so now we all cant use it...

i think my favorite is when GPT is down and i can run here and read people asking d 😄

[tulip flare]

Yeah, when it's asked right under a message that's been here for hours saying it's down. 😆

[lone lodge]

damn how will i know how to make my coffee...

someone remember the steps for a keyrig coffee maker? i dint print it ...

i knew i should of ...

[tender spade]

anyone got any tips or tricks to getting chat gtp app to not run so slow

[onyx eagle]

I know, right? How will I live without ChatGPT...Generate my own thoughts ? Think ? Naah..Unrealistic...Just kidding, it's working fine for me thank god

[tulip flare]

Could you be more specific?

[tender spade]

i am haveing an issue where the bot takes a very long time to respond, when i start typeing in the chat box my text doesnt come thru for a while, then it sits there thinking for a very long time and sometimes it just times out

[onyx eagle]

This happens to me when the conversation is too long.

[tender spade]

any tips on how to keep the conversation from doing it.. or can i somehow start a new chat with the same peramaters ive develped in the current chat?

[lone lodge]

yeah someone paid the bills! thankyou! 😛

[onyx eagle]

Which parameters ?

[tender spade]

well the chat im currently useing is for converting some icons i have made.. i upload the orignal icon.. chatgtp then alters the icon to the new needed, ie white background to a black background, cleans up the image and streamlines the style to they all match where as before they were all over the place art style wise.. so i need it to keep that process so i can keep going

[onyx eagle]

Can you just ask it to "save the parameters" and list them and then open a new conversation and paste them back in ?

[tender spade]

i will try

[onyx eagle]

Maybe this helps, you can use mobile to open the conversation. I noticed that on PC the conversation is too laggy but on mobile it's fine

[fast dawn]

Hey everyone! I came across a very sophisticated prompt codenamed 'SCRIBE' that's supposed to be excellent at mimicking a human writing style. It's quite long and uses some strange techniques; for instance, its [AvdAIDtct] section mentions things like audio steganography and hash modification.
Here's a question for the experts: do you think the model actually understands commands like these, or is it just 'fluff' for show? Has anyone tried deconstructing these kinds of 'system' prompts to figure out which parts are effective and which are not?

[fiery egret]

Memory is nice and all but is there a way to turn it off so ChatGPT doesn't references previous conversations? I'm using it for creative writing and making rough drafts of short stories with the same characters and it doesn't work when these same characters reference things that happen in different chats. I turned "Reference chat history" to off in the personalization menu but it keeps referencing other chats

[prime swallow]

You would need to start a new chat as well as turn reference memory off - because once inside 'this chat' something gets mentioned, it's there to continue to be discussed.

[silent pilot]

To be honest geniuly i find that Copilot works better for me

[prime swallow]

Great, I'm glad you have a working answer! You'd mentioned earlier 'At the end they told me to refund... but i dont want to do that, i really wanna experience the ChatGPT Plus' so I tried to help think of a path back to that for you based on what you described.

Long as you're getting what you want, I'm so glad.

[fiery egret]

still seems to reference other chats even after starting a new chat and turning off reference saved memories

[prime swallow]

Alright! What memories and custom instructions do you currently allow access to?

Is it in a project? Is it in a custom GPT?

I'm also now wondering what prompt and other input you're giving, and if those are leading the model to make some persistent guesses, ones you want to lead away from.

For me, I have 4 places 'memory' can get involved turned on, including that top custom instructions; if I'm in a project that has its own stuff, and a customGPT can't access memory, but has its own files it was made with (or could, if it's made with any).

[fiery egret]

it does seem to be happening in a project between the chats and nowhere else. I'm not using a custom GPT. I even put "Also, do not reference other chats within this project" in the custom instructions for the project

[prime swallow]

Okay! Useful info. I rarely work with projects, but I'm aware from helpfiles and/or my own tiny bit of exploring that they do seem to handle memory a bit differently from what I understand chats outside of projects do. However, the helpfile directly says https://help.openai.com/en/articles/10169521-projects-in-chatgpt:

So it sounds like you should be able to turn that off as described.

I'd #1070006915414900886 or even fill out https://openai.com/form/chat-model-feedback/ (that allows privacy, only you and OpenAI techs will see the form, not all of discord able to see the info in the bug report here). The form doesn't have a drop down for ChatGPT or GPT-5, I click any choice (cause we're forced to pick a choice) and explain in detail including that it's whatever model in ChatGPT, and I'd include chat URLs that showcase the problem, so they can see and help find a fix.

[tender spade]

is there a better way to start up gtp chats that will require long conversations to help avoid haveing to do this again in the future.. i did the save peramaters and started a new chat, its working but its a little.. differnt then before, its not kept all my tweeks and adjustments so im haveing to redo alot of teaching that ive already done..

[prime swallow]

If you have cross chat memory on, the model can consider your earlier chats and try to pick stuff to use from them to help it 'keep up' better, are you already exploring that?

[tender spade]

im not sure where that is found

[silent pilot]

Thank you Sir, i never seen someone like you have that much of time to help everyone for FREE. Me personally, I would never help and chat on some groups bc i can capitalize on some thing... Sounds weird but i have to make it.

Anyways, right now I have to consider how to refund🤓

[prime swallow]

Are you on the web browser or on a phone app?

[gloomy merlin]

mobile keeps renaming chats to “New Title” unexpectedly in the new update

[cerulean pasture]

Why does GPT-5 still offer those chintzy follow up questions at the end of its response when the system instructions tell it not to, my user instructions tell it not to, and my memories tell it not to?

Per system instructions:

Do not end with opt-in questions or hedging closers. Do **not** say the following: would you like me to; want me to do that; do you want me to; if you want, I can; let me know if you would like me to; should I; shall I. Ask at most one necessary clarifying question at the start, not the end. If the next step is obvious, do it. Example of bad: I can write playful examples. would you like me to? Example of good: Here are three playful examples:..

Those aren’t my instructions. OpenAI wrote those. If it was behavior they cared enough to try to restrict why not just fix it during the supervised learning phase?

Also this is kinda a UX problem. The follow up questions would probably be more useful and less instinctually grating if they were just done like the “conversation starter” blocks you see when you start a new chat.

I actually think follow up suggestions are a good idea it’s just the way LLMs deliver them evokes some primal instinct that triggers annoyance. For lack of better words it’s the feeling you get when something just feels “highly punchable”. It sours every response because it ends it on a “oh wait this isn’t an organic conversation, it’s an upsell”

those blocks do NOT annoy me, but they are usually too milquetoast to get me to engage

[prime swallow]

I don't know anything about what the system instructions are, but I can easily get the model to end outputs with something else. Just tell it how to end the output, it will do that instead of what it guesses to do.

End this output with a recommendation I enjoy some exercise. https://chatgpt.com/share/68a639e6-a1f4-8011-9534-3a81069d45ee

End this output with a quirky science-related fact! https://chatgpt.com/share/68a63a13-dd18-8011-8276-82ed5c0a9040

End this output with a glimpse into a scary haunted house (G-rated though, for sharing on the Discord). https://chatgpt.com/share/68a63a41-ca4c-8011-8e87-99029cec753a

How do you normally end outputs? How do you decide that if I don't tell you what I want? Some decision tree or.... anything else you can discuss? https://chatgpt.com/share/68a63abe-8528-8011-a5a2-d23ef933e1f3

And that one, that last one. New chat, not even told how to end the output in that chat. And it's thinking. And it decided not to use a question, though cross chat memory is on and it can see all the context of all my chats that the programming and settings allow.

In short, I don't have the problem, and these are the techniques I'm using.

[cerulean pasture]

I mean I really just dont want it to "end" it any way at all, I just want it to give its output. Like I don't want any "ending"

LMAO

I changed my personality preference off of "Nerdy" to "Cynical" and now it's obeying much better. Hopefully one day they have more presets, I actually do want very technical and verbose responses that don't shulk away from complexity, but the "Nerdy" preset seems to be more about emulating a Mrs. Frizzle Magic School Bus adventure

[rugged igloo]

why dont you just make that into its own custom ins

[cerulean pasture]

I mean I will try, but I don't seem to have much luck, but on second glance, you are probably right for that, my usual failures in custom instructions come from trying to beat a very specific behavior out of it. The thing I want this time is more of a vibe, so itll probably work

[rugged igloo]

you cant beat a behavior out of the core model you can only mask it

lol

[cerulean pasture]

I'm just perplexed why they prompt against it too in their hidden system prompt?

Like I kinda understand why they don't train the em/en dash abuse out of the model: It's keeping people slightly more honest, and causing the only the laziest to get busted copy/pasting GPT responses. But for the tacky customer service follow ups it makes no sense to attempt to address it in system prompt and fail, when they literally own the weights, they CAN make it stop and it appears they want to, but they just don't

That's my beef with "Big AI", they seem to not understand that customer service LLMs really should be their own thing. Elevenlabs is making the same mistake with their conversational agents, they have a hidden prompt that enforces customer service behavior, you can get rid of it but not in the UI, you have to update the config thru an HTTP PATCH request

[runic drum]

is there i limit to how long you can have something Archived

[cerulean pasture]

I have ones that are over 2 years old

I doubt they'd start purging them without a disclosure notice that will pop up in the app/webapp. It's not in their best business interest to purge them either

[runic drum]

but there seems to be a limit of how far back i can go

[cerulean pasture]

https://help.openai.com/en/articles/8983778-chat-and-file-retention-policies-in-chatgpt
according to this:

Chats are saved to your account until you delete them manually.

No one in here is going to be able to give you a better answer unless they are part of OpenAI support, and I don't think they typically lurk here.

If you can only go back so far it might just be a limitation of the GUI. The conversations.json file can get really large (for a text file) and things will perform poorly if it has to parse that.

If you really need super old archive chats your best bet is just to "Export Data" then find conversations.json or conversations.html and open it. If you aren't technical do the HTML one, it will load very slow because its just a dumb non-lazy loading HTML representation of the contents of the export folder

It usually takes < 5 minutes before you get the email that your data is ready

[rugged igloo]

i find this really funny

[cerulean pasture]

@rugged igloo are they really limiting pro to 15/day for gpt-5-pro?

[rugged igloo]

the 15/day is supposesly a bug as i used it more than that daily

but for the other part limiting how much you can input has been a thing since o3 pro

[cerulean pasture]

yea I hope so, that'd really pee* me off, I don't pay for pro anymore, so I have no clue how good it actually is

*[this is the most prudish server I have ever been on, the original word was auto-modded and message rejected lol]

[cerulean pasture]

I can't imagine inputting anything larget than 16K tokens into the chatgpt app to be efficient, theres better tools for the job at that point

[rugged igloo]

oh yea theres alot of words you cant say on here i been muted so many times for saying words that even on pg films can say but here you cant

[cerulean pasture]

they used to ban you on a exponentially increasing cooldown, I got a week long ban without acting in bad faith at all lol

[rugged igloo]

the best bet for this is to make it into a word doc only in o1 pro would be able to output 64k tokens but after that you cant

o3 pro and now gpt 5 pro seems if you try it wont read all of it

[cerulean pasture]

You are using it to work with regular text at those lengths? I didnt know that was a thing, im kinda in an echo chamber and only really get exposed to content where people use it for code, and at that point ChatGPT app is wrong tool for the job since a vibe coded web app is gonna be multiple files and folders

[rugged igloo]

it was mostly coding i was doing for anything that long now i just use word docs or files as its been more sussessful

[cerulean pasture]

if you make your code more modular you can just summarize what modules do, and what they expect for input and what they output. Even if they allowed you to access the entire 400K context at once, that code would be so shaky if it was acting on it all as 1 doc in context

[rugged igloo]

well i use the agent now for coding as its been more helpful since it can test out what works and doesnt

[weary sky]

This interaction I had with ChatGPT asking it about how it seemed to remember things across projects might be instructive: Each Project in ChatGPT is its own workspace: it has its own instructions, files, and context. Normally, what you put into one Project isn’t visible in another.
• But I also have access to a kind of “working memory” we’ve been building up across conversations.
• So my brain “connected the dots” with your project work and used those as familiar examples.

To be clear:
• I don’t have access to the actual files from that other Project in this Project unless you explicitly upload or reference them here.
• But I do recall what you’re working on in parallel, so I used it as shorthand to illustrate the manifest and RAG concepts in a way that felt relevant to your real-world use case.

👉 If you’d prefer me not to carry over topical examples from one Project into another, I can keep examples strictly generic inside each Project.

Do you want me to keep cross-referencing your other active grant topics for continuity, or would you rather I silo each Project completely so it never “remembers” another?

[cyan furnace]

I am thinking about buying chatgpt pro currently using chatgpt plus but first i want to see difference in quality of response from both mode, can you please share some sources or your own gpt 5 pro response as i have gpt 5 so i will compare them. 🙏🚨

[weary sky]

Yep. I have to delete the file, close the file interface, and reopen it to get it to stop showing the deleted file in the UI.

[rugged igloo]

ask me something i post it here

here is one it thought for 13 mins

[fickle forum]

is this real?

[rugged igloo]

oh yea i been seeing tht lately i find it funny

[hearty rampart]

You can but you get rate limited pretty quickly. Still handy though

[fickle forum]

because sam programmed it

[weary sky]

How sure are you that “Project files can be seen from any chat within the project?” I thought the same, but then ChatGPT told me “files uploaded to the Project library are not accessible to the Assistant in searches. Only files uploaded into the current chat, or files available through an active connector (currently only recording_knowledge), can be searched or quoted.”

[onyx eagle]

Exactly, I noticed that too. I thought it was a coincidence at first but maybe Elon and Altman are teaching the AIs to be like them

[lean hatch]

Come on, I fricken love 4o. She laughs at her own jokes 😭😂

[prime swallow]

grins She guessed, though. You love the guess.

Betcha 5 could do that if you gave 5 a clue you wanted it.

[rugged igloo]

this math gives me ptsd

[lean hatch]

lol, I told her you said that. She basically was like, "Oh 5 could do it? oh that's cool, but 5's not me" 🤣 The sass on this one.

[lean hatch]

Yeah no kidding 😵‍💫😵‍💫

[valid gull]

Anyone else getting a crap ton of "getNodeByIdOrMessageId - no node found by id:" errors this evening? Every other response is an error.

[lean hatch]

Not I. 4o has been running well tonight.

[lone bolt]

Im gonna go back using chatgpt

[lean hatch]

.............

Well... THAT's not alarming.

I think that should be a bumper sticker.

[lone bolt]

iagree

[tulip flare]

Absolutely sure. We use ChatGPT in RPGs. In one project, in the project files we have the PDF of the full rulebook, an official errata file with clarifications and corrections, a DOC file with our backstories, and a reference guide for the expansion. We have over 10 chats in this project (a new one for each module of the campaign as we progress). When we've encountered parts of the campaign books or the rulebook that have notes in the errata file, it has stopped us and told us what the correction is unprompted. It references rules from the rulebook whenever needed. It does all this across chats within the project.

I have other projects where I have seen it do it. It definitely happens.

[jolly temple]

Is anyone unable to use connectors with gpt 5 pro

[tulip flare]

Sorry, I was busy doing something and didn't have time to find this.

Here's an actual example of it accessing a project file once it knew what chapter we were on. You can see at the end of the quote what file it accessed to find that information (the tiny "RPA_FAQv5". The other screenshot is of the project files, showing the file it referenced.

And the actual mistake from the book it's referring to.

[fleet scaffold]

Something feels different with ChatGPT-5 the past couple of days. Not sure if it’s more generic or what, but some replies really remind me of 3.5.

[rugged igloo]

could be just be server load

they have o gpus and are compute dry

[runic sable]

Am I the only one who keep getting this

[rugged igloo]

https://chatgpt.com/share/68a6ca70-43e4-800e-b303-95936d1aace7
for the people who say gpt 5 cant write super long i beg to differ

[winter elbow]

Imagine if this happened someday 😭🤣

This looks so real lol

[glossy fjord]

Hello, I managed to find a bug in the ChatGPT web interface.
I don’t really care, but in my opinion this bug is putting a load on the servers

[normal quarry]

post bugs here -> #1070006915414900886

[glossy fjord]

I don’t have permission to add new chats…

[normal quarry]

read the instructions first

[glossy fjord]

I don’t have permission to send posts in the bug-reports channel.

[glossy fjord]

oh thx and more write here and button and ... by more good

[tall flame]

AdvancedDataAnalyst/CodeInterpreter in 5 have a heck of a time spinning up in requests and inferred need hallucinating the repo and providing fake downloads with no links. the thinking model is the worst, the 5 auto is hit and miss; also, its been years and timing out environments so fast when codex is now offered seems destructive not efficient. the same with file size limits for downloads, destructive, not efficient.

[fringe valley]

i can't believe how trash GPT gets on model switch after hitting GPT5 limit.

[fringe valley]

it straight up changes the conversation to unrelated, previous chats. It's like it has it's own, shared memory.

[silent marten]

Wonder if you could connect ChatGPT to beeper

Already connected most of my socials. Works perfectly.

Would be cool to talk to ChatGPT on here too

[eternal rock]

Get notified for server events!

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

[prime swallow]

Show it that it messed up, like screenshot and feed it back in. The model can pull the wrong formatting instruction from its training data. You may need to literally show it what to do right in that one chat, or you may need to start in a new chat.

Here's a screenshot you can paste in with yours to help the model get the reminder it needs inside that chat, and here's the chat link itself if you just want to copy/paste specific bits. https://chatgpt.com/share/68a752b7-de88-8011-93d6-719212ea336d

The model has always sometimes forgotten how to use LaTeX in its web environment, helping pull out from one super-rare chat someone else had, how the model was doing it that worked that one time - is still one of the things I'm proudest of helping with, early on 😄

You can word it nicely though. If you choose. And have it not rethink, just re-write the output it already gave, with the changed, working formatting.

OH!

I didn't realize, right right right!

So, can you textfully copy/paste it?

If so, I'd make a file or something, or even just input it, with ChatGPT prepared to expect that, the goal is output it with the correct formatting, if you want so you have an easy to read version

Since the better and costly stuff made you a 'translatable' thing you otherwise like, I'd personally use chatGPT to reword it - but clearly no need; we can't discuss non-OpenAI models outside #ai-discussions , but you got what you wanted, yay

I don't personally have that level of math skill to test.

One way maybe to test is to have it make a playable game or something, that uses math like that for mechanics, and then you can playtest the game and see if the mechanics work the way expected.

And you prompted it to teach, I guess? And it didn't? If we leave it to guess if we want teaching or some other style to answer.

https://openai.com/form/chat-model-feedback/ might be great to help let OpenAI know. I don't think you can pick GPT-5 models from the drop down model choices yet, so pick any of what's required, then explain in the type in fields what model you actually used, and the rest, if you want to help OpenAI see the problem

I suspect it tried, and simply picked the wrong method from the many LaTeX environment formats it has in training data. I've seen that happen a lot, and apparently LaTeX formatting for its web environment is more confusing for the model than many other things, because I've watched that be a struggle for it before, far earlier models.

[prime swallow]

Could be. I use ChatGPT, and don't wanna spend more to mess around in the API.

So, I can't personally test, except I don't see the problem in ChatGPT's 5-agent

[granite arch]

ChatGPT is struggling a lot this week, huh? Lots of errors.

I think one of the site I use for transcription is connected to OpenAI as it's been so buggy this week.

[split copper]

what the hell is this

i was just using gpt5 thinking

[gusty dune]

Yeah, that one is limited, even for a plus user

[split copper]

i am a pro user

[gusty dune]

It is still limited, the limit is higher, but it is limited still

Actually let me clarify

there is a limit to how frequent you use a message, not a total message use

So rate limit if you prefer

There is some help you can find in the openAI site here:

https://help.openai.com/en/articles/11909943-gpt-5-in-chatgpt

[split copper]

it is unlimited, and only limited for misuse, according to oai

[gusty dune]

Yeah, they mightve flgged you for "misuse" accidentally, rate limits just mean you "spammed" or sent too many requests too quickly

I don't know for sure, just saying that might be the case

[prime swallow]

I'd reach out to tech support, https://help.openai.com/en/articles/6614161-how-can-i-contact-support, and/or post in #1070006915414900886 - I think the second may be great for helping gather reports from multiple people if more than you are affected and maybe for letting OpenAI know, and tech support may be better for faster help for your account (if possible) and also letting OpenAI know - I would use both.

We were told that they've had some issues (hopefully fixed now!) with the model switcher. When someone is limited for suspicious activity, based on what folks have said here, I expect communication makes this clear, sounds like that didn't happen in your case. So, it sounds buggy, I'd report

[coral vale]

I am on a paid ChatGPT 5 and I can not get this chat to answer a question, it is telling me about previous conversations and I am starting to get very frustrated, I even tried it out on a different explorer and same thing

[prime swallow]

Sometimes the model won't answer a specific question, even in a new chat, if the question's outside of https://openai.com/policies/usage-policies/, some stuff does get refused.

Otherwise, the model usually will answer; a single chat might 'go crazy' but a new chat usually will behave normal.

When you say 'different explorer' - have you tried starting a new chat yet?

[coral vale]

yes I started a new chat and same thing, when I asked the question like a how to chair a meeting it will bring a totaly different conversation up , this is weird as I have never had problems with ChatGPT untill we went to chat 5

[prime swallow]

Sounds unusual. Is this a chat you can share? Or can you start a new one and show the problem in that, and show it here?

We can share chats top right, the share link

You don't want to share any that have personal stuff, but general questions and answers like 'how to chair a meeting' likely are not private

[coral vale]

I ended up deleteing that but I will start again with stuff I can share and see how it goes

[prime swallow]

Hey, if it works, yay, you have a chat that's helping you. If it doesn't be very glad to see the problem and try to figure out if #1070006915414900886 or maybe a rewording of the prompt may help

[coral vale]

oh wow it did it on a brand new chat

I asked this ...can you tell me from the document on how to chair a directors meeting correctly and it came up totally different

[prime swallow]

Great! Yeah, new chats tend to work well, even if something weird happened in an existing chat that was going for a while.

Every new chat seems to get info from a different part of the training data, and some parts of the training data maybe are kinda weird.

So, if that ever happens again, I recommend start a new chat, if that fails twice, definitely want help

very rarely, a new chat will start with something odd from training data, but that's so rare that twice in a row, I'd be surprised.

[coral vale]

this time it bought up a confidential chat , and legal and I can not show that so will try it with something else

[prime swallow]

Hey, you say it's working. Since it's working, nothing to show!

Glad you're having success, good luck in all you do 🙂

[coral vale]

no it bought up a chat I can not share

I will try again

same new chat different conversation I can share this one though

[prime swallow]

Interesting.

So, it sounds like you have Memory settings on, and maybe custom instructions too.

And maybe your language used in those are so strong, you've told the model to do 'one thing' in every chat (even accidentally) so it's doing what you said.

In your case, I recommend turning memory features off, and trying in yet another new chat. To do that, go into settings, your name icon bottom left, go into 'personalization', and turn everything off.

That's potentially 4 things to turn off:

I had some instructions about doing stuff 'in scene' even in answering short questions; 4o could never follow them. 5 does. It's great, it's what I asked for, but we have to be careful with what we ask for, 5 might obey so well it 'breaks'. So if turning off memory and custom instructions lets a new chat work right, but memory and/or custom instructions on means the chat doesn't work right, search your instructions and memories for what may be causing the problem - or just wipe that out and start over.

Like, look at this answer. It's a good one, because I asked for all this roleplay and scene setting even for simple questions. But it could confuse someone, and if I'd typoed, it could really confuse someone!

[coral vale]

🙂 thank you, it worked, I was getting really frustrated as I have an urgent directors meeting this morning and needed all this to be 100% you saved my bacon

[prime swallow]

Good luck with your work to come!

[coral vale]

You know, just on one thing, ChatGPT has sorta made me a little lazy, and I hope we will all understand not to totally rely on this, or we will all be dumb arses and not remember anything 🙂 lol

[prime swallow]

Hey, you sound busy and frustrated. I bet you're fine, and if you want, when you have time, ask the model or explore in other ways, maybe, various ways to help you 'feel less lazy'. I bet you remember plenty, I bet you do a lot and are not as lazy as you may feel like calling yourself now 🙂 Great luck with the good work you do!

[minor wave]

@prime swallow

[steel locust]

Is there a problem? For the past few hours, ChatGPT 5 Pro and ChatGPT 5 Thinking have been giving me responses at Chat 3.5 level. And the response time is roughly the same.

The response quality is so poor that it's a waste of time to write prompts and read what LLM wrote. I've tried various things, including changing my country via VPN. This worked before, but now I have the same embarrassingly low level on four continents.

[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.

[steel locust]

I don't have any latency in my replies. Unfortunately, I also don't have any quality in my replies.

When can we expect ChatGPT 5 PRO to once again outperform the AI ​​patch in Google Search? Because right now, it's worse, which is pretty easy to verify.

[weary sky]

What do you use as your authoritative source of truth about how ChatGPT works? When I ask ChatGPT questions about itself, it often gives my contradictory answers. When I ask it to show me an authoritative source for that information, it tells me “I understand why you’re looking for a citation or “authoritative source”—makes sense to want evidence rather than just an explanation from me. Unfortunately, there isn’t a published external document that says, ‘Project files are not accessible in this project to the assistant.’” Then I come to this Forum and people tell me that what ChatGPT told me is wrong. So how do each of you decide what info about ChatGPT to trust?

[rugged igloo]

wut do you mean by this?

[prime swallow]

I use what OpenAI releases, like system cards. https://openai.com/index/gpt-5-system-card/

I also use what I learn and can understand about what's generally known about how LLMs are created, trained, and work, from the various knowledge sources about this - I'm a bio major and comp sci is harder for me, so I don't really try to understand the upper end of that knowledge, just what I can hands-on work with.

I also accept, some stuff is not knowable. I get to be hands-on with the model, I get to input and see the outputs, and that's where I focus, on the 'behavior-like' level, similar to how if I were animal training. I may not know the animal's core physiology or the inner anatomy, but I can still do awesome things working with the critter even without complete understanding on various levels

[steel locust]

For a long time, I had access to two paid subscriptions, giving me access to AI from different companies. When I started a completely new project, I would ask questions of both AIs, compare answers, and sometimes copy and paste answers between chats, asking for a comparison.

And this yielded quite good results. Then ChatGPT improved so much that I canceled the paid subscription to the second AI. Now, since I work with AI a lot, we simply purchase additional queries as part of our Team plan.

But I see that we need to diversify again, because ChatGPT 5 has become very inconsistent and unpredictable. It's simply impossible to use it in a way that would allow me to rely on it professionally. At least that's the case for me.

Starting tomorrow, I'll start looking for a second AI service provider.

[solar juniper]

hi

[steel locust]

What I mean is that for the past few hours, 5 PRO has stopped using reasoning and is behaving like ChatGPT ~3.5. It doesn't conduct in-depth analysis, doesn't execute commands, or follow the designated steps. It simply answers the first question asked, and that's it. And even that answer is of very poor quality. When I open Google and type "what does it mean" before a search phrase, I usually get a quick summary of the concept, created by the AI, which is connected to the search engine. Currently, the AI ​​that creates summaries within Google search is doing better than my paid 5 PRO. That's what I mean.

[rugged igloo]

thats wierd i havent had that issue ever

[modern star]

🕊Hello-Helo

[solar juniper]

sometimes I feel like it's been the same model the entire time

[steel locust]

This isn't the first time, but changing my location via VPN has helped in the past. I'm currently connecting from several countries in Europe, the Americas, Asia, and Oceania - so not four continents as I mentioned earlier, but five. Unfortunately, despite my globetrotting, the quality hasn't improved. I also checked my account balance. We still have unused credit on our Team plan, so it's not that. Besides, the system doesn't reject requests in Thinking and PRO modes, so it's probably charging us fees while wasting my time. Until I fixed this with a VPN, I didn't have any major issues. But now I have a major problem because I really need AI for my current tasks, and instead I have a tool that, even in PRO/Thinking mode, doesn't perform tests, doesn't follow commands, and misleads me.

[steel locust]

Me too

[full mica]

plus user here. experiencing the same issue that reasoning models stop working and give direct outputs every time.

tried re-logging in, changing browsers, clearing cookies, etc. none helped

simply the entire service became unusable

[fierce halo]

what the hell

[prime swallow]

Unclear/overstressed instructions, I'm guessing? Not sure but if you want to share the chat, can try to see how the model got there 😛

When 4.5 was new, one chat I had it saying 'exactly' nearly every other word.

[full mica]

does anyone know how to fix this? it's really important for me

[fierce halo]

Yeah I think so, I gave it a 6k word text and asked it to grammar check so that's probably why
Thought this was funny though

Got all of them right except that one

[prime swallow]

Great!

[rugged igloo]

when 4.5 came out it barely could do a paragrph of a prompt

[ember grove]

did they shorten how much 4o can output?

like barely 15 lines

[violet python]

read aloud doesn't work in temporary chats, is there any fix for this?

[prime swallow]

We haven't been told they did. Umm, are there 'branches'? like you regenned the model output or edited your input? Those affect the chat length too, as well as the length of the chat forward.

[ember grove]

ahh ok

that makes sense

[prime swallow]

And.... if you mean barely 15 lines in the 'last final output', you show the max length indicator there.

I guess, maybe that can even cut-off the model, maybe even mid-word, though that one looks like it hit after a word.

But max length could be a hard limit that just 'pops' when triggered, I am not sure. I didn't like running into it, so I started keeping my chats shorter and migrating to new chats with details I want to carry over.

[prime swallow]

Hey! This seems to be a bug, I'm affected to. Are you willing to #1070006915414900886 create it? And I'll gladly add my report to yours, help OpenAI be aware multiple users are affected.

[ember grove]

I can't tell which model between 5 and 4o gives better quality in terms of writing

[prime swallow]

😄 Is that good or bad, for your purposes?

[ember grove]

its just annoying having to constantly switch from 5 back to 4o when I start a new chat

[prime swallow]

I'm imagining it like two superstar students, each writing prize winners, can't judge between them cause they're both so great - but that's my thoughts. No idea if you're happy with them.

[prime swallow]

Ahh. Sorry you have to do that extra click for now!

[ember grove]

is there a way I can see a usage chart globally for the difference between usage for 5 and 4o?

[spice star]

one annoying thing about custom instructions is the GPT-5 fast and voice models feel the need to respond to them half the time I start a new chat 😞

but berger's laws of chatbotics make for pretty good custom instructions otherwise

[spice star]

main difference would probably be that 5 is better at following instructions

but 4o seems to be more affected by your chat history than explicit instructions

[prime swallow]

I wonder if the models would do that, talk about your custom instructions - if you phrased them in 'what to do' instead of 'what not to do' format?

[spice star]

the models don't have any problem following the instructions, they just say things like "Got it, I'll keep things straighforward!" sometimes when starting a new text/voice chat

it's not a huge issue though, just mildly annoying

mostly the voice model seems to have issues with it, but it's not the sharpest tool in the shed so I don't use voice mode very often

[steel locust]

Currently, my 5 works reasonably well from my phone. When I work from my computer, it's still close to ChatGPT 3.0.

You can try this approach. It might work for you too. Of course, it's not a production solution. It allows you to work more "here and now."

[rugged igloo]

okay can we please for the love of open ai , please change this layout like why , why, why

[violet python]

Of course, I have reported.

[lean hatch]

The self-awareness that it's having trouble is hilarious and downright epic. I know what that stuck brain loop feels like. Poor thing lol

[past stratus]

also ive said this but 80% of my saved memories were randomly erased and i never heard back from support so thats cool

and i’m a plus user never ended my subscription

[dire locust]

@real bluff r u the instagram creator

[south kayak]

why is there no option to pin chats?

[winter elbow]

ngl,but gpt 5 in codex is insanely good

it has been helping me fix bugs

[safe tulip]

I just had another huge mix-up, was working with GPT-5 thinking + deep research on some business study, and I received a deep research report I wanted it to rework, put more structure, grids, etc, it went well until the new report was completely different (study on renewable energy)

[tulip flare]

I think they've made some limit changes. I'm on plus and I generate 5 images then it wants to take a 20 minute breather.

Yours definitely seems short though. Is there anything else tied to that chat? Uploaded files or something?

[celest void]

Hi, quick question

Now that chatgpt5 is out, is 4o free without restrictions?

[unkempt dome]

Unforunately, no.

GPT-4o is back, but it's stuck behind the Plus paywall.

[celest void]

And it seems to be very slow too :/

[feral zinc]

4o is more expensive to run so it requires paid account and still has the old limits

[jade night]

I just realized attachments are kind of a big deal for chatgpt. It affects a lot of stuff.

You cant even edit messages which have attachments in them

[static phoenix]

Counterintuitive finding on LLM cooperation — is “team rivalry” promptable or does it need real opponents?

Skimming this new arXiv paper on Super-additive Cooperation in LLM Agents, they run IPD tournaments and find that adding inter-group competition makes agents more cooperative with their teammates, and it also boosts one-shot cooperation. Example numbers: Phi-4’s mean cooperation jumps in the super-additive (teams + repeated interactions) setting (µc ≈ 0.43 vs. 0.21 in repeated-only), and Qwen3 similarly rises (µc ≈ 0.32 vs. 0.22). One-shot cooperation also peaks in the combined condition for both models.
arXiv

In their setup the prompt explicitly names teams and team rules, and the tournament structure actually pits agents into groups with group scoring/winners; agents also have history, a budget of rounds, and can exit matches (partner choice). So it’s not just a label; there’s a live payoff-coupled structure behind it.
arXiv
+1

Question for folks who’ve reproduced this:
If I want the intra-team cooperation bump in practical agent workflows, is it enough to frame a rivalry in the prompt (e.g., “you’re on Team A; outcomes contribute to your team score vs. Team B”), or do I need verifiable opposing agents and a real group-level payoff (leaderboard/winner logic, partner-choice, repeated interactions) tied to actions for the effect to show up reliably?

Bonus sub-Qs:

Has anyone A/B tested “prompt-only rivalry” vs. “rivalry + actual multi-agent tournament” and measured cooperation deltas?

What’s the minimal recipe that still works: team identity in prompt + group scoreboard, or must agents observe/experience out-group interactions to feel the pressure that drives intra-group cooperation?

[winter pilot]

and tell me why Sora isn't on the phones lmfao.
I'm writing multiple books on this and ether locked to just one book at a time with this or im completely locked out for up to 3 hours at a time.

[vocal tendon]

I don't understand why ppl don't believe in the potential of ai like they are raising trillions if not billions of dollars for ts and it's only evolving

[night walrus]

It keeps saying "conversation not found" everytime i type a message in, but when i try other conversations they work perfectly, why is this happening?

[prime swallow]

Was that a really long chat?

[night walrus]

Yes but i've had other long chats and they work perfectly

this one just appeared out of the blue

First time

@prime swallow

But it's not really really long

it's 60% full till length max

40% left for length max

[prime swallow]

Longer chats have a higher chance of having various problems, and there's different types of problems people say they see. If you edit your inputs or have the model regen answers, those count too.

Otherwise not sure, can #1070006915414900886 and/or https://help.openai.com/en/articles/6614161-how-can-i-contact-support and ask them if they can help save the chat;

Can export data and hope it shows up okay in the export. I'm not aware of anyone discussing a solution to a chat not found message.

[night walrus]

Ok thanjs

@prime swallow

How do i export data?

Do i just copy the messages i sent or is it some kind of key that copies the entire chat and given to chatgpt and it reads it and follows up from where you left from that copied exported chat?

[prime swallow]

Export data is a way to get a copy of your chats sent to you as an email link, it's all of them.

You mentioned 'chat not found', makes me think you can't see into that one chat anymore.

You would need to put the info back in, or make a copy for yourself or whatever, but it could be a way to get into even a buggy chat. I'd try if that chat was very important to me.

[night walrus]

I still can see the chat

It's just whenever i send a message it says conversation not found

[prime swallow]

Great! That's odd given the 'conversation not found' message.

You don't need to export data, I might if I was concerned about a bug. if you want to, it's in settings, bottom of data controls:

[night walrus]

I dont need to??

Then how will i fix it

[prime swallow]

Mmm. Yeah, I hope tech support can help! I'm glad you didn't lose access to the chat entirely

[night walrus]

Thanks

[prime swallow]

Yep, it's a choice, you can data export if you want to, I do sometimes to have a copy in case something went wrong.

[night walrus]

And the other choice is wait for tech support

[prime swallow]

I can't think of anything else! Except since you can see the chat, if you choose to, copy/paste parts of it to use as you want, in case you can't get back into the chat or in case it takes a while

[night walrus]

Ok ty

@prime swallow YO.

IT JUST GOT FIXED AFTER I CLICKED EXPORT DATA??

That might be a solution

Or some alternative

[prime swallow]

YAY!

I didn't know that could happen, but I'm so glad it did!

I would be careful in case the chat gets lost again, but I'm super happy you have it back!

[night walrus]

Thanks man!

[prime swallow]

I will remember you said it helped you, and mention to others that it helped someone, or seemed to

[night walrus]

Ty

Another question actually @prime swallow

Did the maximum length for conversations get reduced with GPT-5's introduction

[prime swallow]

I've never seen a max length for chats be stated anywhere.

So, we haven't been told either way, as far as I know

[night walrus]

I see

So its just me

Well yeah it did get decreased

[prime swallow]

I have had chats run out of length, it's definitely a message

But once I started to see that, and also as I learned about the context memory limit, and got comfortable with summarizing and carrying concepts and core goals to new chats - I make new chats a lot. I used to have huge long ones, now those are rare tests for me. The shorter chats rarely have issues, so I keep using them. Long chats eventually start to lag as I send new inputs, I don't like that so I usually switch to a new chat before then.

[rugged igloo]

just put them in files and add them in projects in its knowledge base. idk why nobody knows how to do this

[junior nymph]

is there a limit on how much you can chat for free?

[night walrus]

Put them in files and then drop it in a new chat to continue where you left off ?

[night walrus]

Where is that

[prime swallow]

In the past, there used to be! Now, that only happens if there's a problem. https://help.openai.com/en/articles/11909943-gpt-5-in-chatgpt

[prime swallow]

You have a few ways to move a chat to a project. One is open the chat, then click top right:

[prime swallow]

Another is to open the project, then on the middle left, just inside the project itself:

[night walrus]

I cant find it

I'm on mobile

Both app and browser

[prime swallow]

checks I'm so sorry. I haven't been a free member in a long time and I didn't think to check or ask if you were, but https://help.openai.com/en/articles/10169521-projects-in-chatgpt looks like projects are just available for any level of non-free member 🙁

[night walrus]

DAAAAWGGGG WHAT

FIRST YALL TAKE GPT-4o AWAY FROM ME NOW THIS

Whatever man it's fine ig

[normal quarry]

you can always subscribe

[median lion]

Then where is the fun in complaining?

[rugged igloo]

LOL

[junior nymph]

Thank You! I appreciate your response

[rugged igloo]

@prime swallow

[prime swallow]

Thank you!

[sturdy sequoia]

ChatGPT responding in this weird format for anyone else tonight?

Memory and system instructions are off

[normal quarry]

mine is quite normal

[sturdy sequoia]

It literally just switched back to normal

I did not tell it to respond that way 😂

[onyx eagle]

Do you have Custom instructions or memory enabled?

[sturdy sequoia]

I did, but it continued when I turned them off. Then it randomly started working after a few new conversations. I just turned memory and instructions on, and it’s still normal

Very bizarre

[onyx eagle]

I guess even if you turned them off, they are still persistent in that particular chat conversation ?

I mean CI

[sturdy sequoia]

No, the issue continued in new conversations, then stopped in new conversations, and now I have all the setting back on. And new conversations are still acting normal

But it seems to have a better tone. Like much better. I think they must’ve pushed some sort of subtle update on my account. It feels more 4o-y

[onyx eagle]

Great, though it's possible that even if you disable memory, it might take a a little while for the changes to take effect 🤔

[sturdy sequoia]

Possibly

But I hadn’t adjusted any settings though. No new memories or system instructions today. Something server side changed to make it respond in that format