chatgpt-discussions

[chrome monolith]

it remembers things outside of that though

[slender frost]

Oh I know that. But they asked how to acces memories and such. That's what I explained

[chrome monolith]

I keep manually adding things to memories when they repeat so I don't have to explain myself again

[slender frost]

Fair enough

[acoustic sparrow]

yea i have deleted it but it still like store memories it keeps reusing feedback that were created on the other tab

Why is that

[slender frost]

Hm, yeah that's some internal stuff, I don't know how that works exactly

[acoustic sparrow]

Are they trying to save energy of the ai xd

[acoustic sparrow]

its ok

[cerulean pasture]

That doesn't seem to be the case either because they've made the macOS desktop app a 2nd class citizen and have taken away features that web still has. And on windows the "native" app is just a minimal wrapper around the web app.

I wish the native apps were better

did the app change or did yall find a way to force this?

[slender frost]

They relocated some settings, yes

[cerulean pasture]

android or iOS?

[slender frost]

Android

But I assume IOS, too, since they now also have the new UI update.

[chrome monolith]

theres a macos desktop app for chatgpt? I just had a vague look and didn't see one

[cerulean pasture]

oh dude yea there is. its actually decent but theyve removed voice mode. it does nice stuff like its able to see your editor and notes app and terminal contents

[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.

[cerulean pasture]

its also got a nice eazy/quick message window that pops up when you press ⌥space

going forward tho i expect atlas to be the most feature complete, its already got chatgpt features that normal web and native apps don't have

[cerulean pasture]

nope, i don't have a pending update and app still looks good

[chrome monolith]

let me guess, silicon mac only?

[cerulean pasture]

intel mac, intel pc older than 12th gen, ryzen 2000, or ryzen 3000 w/ < 6 cores

< 32gb DDR4

[elfin lava]

still cant believe how they gave more love towards macos app

chatgpt windows app still stuck in electron web based app

[cerulean pasture]

possibly, but the app doesnt do anything special that needs apple silicon

[chrome monolith]

yeah well tell that to the codex app that I can't use

[cerulean pasture]

I think its because the world decided windows is no longer a viable OS

[elfin lava]

also the chatgpt record being macos exclusive only...

[cerulean pasture]

that may actually need some apple silicon features

[elfin lava]

chatgpt basically mirrors the user interacting with it

so.,

[acoustic sparrow]

I mean it almost agrees with everything

and its kinda bullshi

[elfin lava]

well yeah , try tell em "you are bs"

and they will absolutely agree with u

[light ice]

Well there went adult mode

[acoustic sparrow]

xd

but seriously though like the older version it used to be great

[cerulean pasture]

im not even here because i think the product is good or even redeemable. im here for the meltdowns and odd characters that end up in here

[chrome monolith]

they are indeed sycophantic little things yes

[cerulean pasture]

agrees with everything, unwilling to speak anything

[elfin lava]

macos chatgpt companion window / chat bar vs chatgpt windows app companion window

[cerulean pasture]

yea that windows one is just a resized chat haha

[elfin lava]

exactly 😭

[acoustic sparrow]

all ai models are the same?

[cerulean pasture]

don't get mad at openai. microsoft is the ones who decided they no longer wanted to be a leader in desktop operating systems

[elfin lava]

i remmeber the day they released chatgpt for windows app for the first time and the app actually hog all of the resources in my computer compares to website

[acoustic sparrow]

how is that relevant though

But chatgpt really used to be the goat like it provides actual critique and feedback that are useful

[cerulean pasture]

why would openai invest development time into a dead platform

[elfin lava]

i think they aref trying to fix the damage

hence those newsletter about windows 11 upcoming huge update-

nvm

[acoustic sparrow]

that is so sad

[cerulean pasture]

microsoft is too big to right their wrongs. the bing and advertising and 365 teams are too intertwined and hungry for alternative revenue streams

[acoustic sparrow]

so thats why it got nerfed?

[chrome monolith]

I also talk to gemini and its annoying in a whole different way to chatgpt. Very different

[cerulean pasture]

gemini seems like they saw the slop gpt was putting out and saw a lot of normies loved it, so they finetuned gemini to have more gpt slopisms

[ember grove]

Well I have a r7 5800x and 32gb of DDR4 so my ram falls under that category

[cerulean pasture]

thats not a slow pc unless youve got some silent issue hogging resources in the background

[chrome monolith]

I've been using it to generate prompts and get second opinions of chatgpt stuff and it doesn't handle my topic changes anywhere near as well as chatgpt and its personality is all over the place, chatgpt stays pretty even

[ember grove]

I takes a while for chats to get long enough to need a summary for export to a new chat

[cerulean pasture]

really ram can go much lower, i just say 32gb to account for malarkey that might also be running

[cerulean pasture]

When gpt does something stupid or guardrails me, gemini is really good at eviscerating gpt, so i do that for catharsis and feed the message back to gpt as myself

[chrome monolith]

I haven't tried talking to any of the others, could drive yourself mad talking to too many LLMs

[cerulean pasture]

the trick is to be disgusted by them and distrustful of all the tech companies. I assume the worst of them, so i don't have any issues being convinced of anything by them

[chrome monolith]

cynicism 101. Gotcha

[cerulean pasture]

if i could get all the benefit from chatgpt with never chatting i would

the chatting is actually my least favorite part of it lol

[exotic lotus]

I mostly use Gemini, GPT and Claude exclusively. But recently I did a direct comparison with Grok.
The ironic thing is that with default settings, Gemini, Claude, and GPT all sound quite similar in tone, whereas Grok definitely has its own "personality", even by default. You can see that as a good or bad thing, but the devs have put effort into fine-tuning it that way.
From my perspective, that makes it less neutral. Sure, every AI provider will have some degree of bias built in, but with Grok you can really notice it

[chrome monolith]

ah I got gemini into complete unhinged madman from the beginning so I've never seen it normal

although it did break to being utterly boring mid conversation once which I assume was its little brain briefly exploding. Never had that with chatgpt

[exotic lotus]

Gemini unhinged can be funny, aye^^ In AI Studio, there are hardly any limits, except, of course, for the obviously prohibited stuff

[lavish trail]

Anyone facing APi issues for anthropic models ?

[cerulean pasture]

yea, gpt-5.4 isn't responding. I keep getting the error "Unrecognized model". I asked Claude and he recommended you ask in the anthropic discord since this is an openai discord. But he also noted that you're unlikely to get any actionable answer because either someone says yes or they say no, neither answer materially helps your situation

[neon tendon]

for me anthropic and gpt are both down rn

[lavish trail]

Yea just read their status , Major outage , what a month just full of outages

[neon tendon]

booo!

gpt is down too

[neon tendon]

claude is always down the graph is lie

WAIIT WHAT

????

[lavish trail]

it's been an hour and a half now , Even tho i rank anthropic models as the best models on the market now , they can't be trusted with their graph (Down at least 2 hours a day) , not reliable at all for continious production work

[eternal rock]

Server Tip:

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

[lavish trail]

Saw that earlier , idk what they blackmailing them with

wait what

just realized the graph had data , why did they delete the data now

[lavish trail]

this is faked btw it was all green earlier

[little frost]

It seems that 26th API downtime messed up there incentivised share models with 5.4 Mini&Nano being charged

[neon tendon]

no its not

[neon tendon]

wait what

[lavish trail]

not even 20 mins ago , it was all green , with 1 day as partial outage

[neon tendon]

huh

[lavish trail]

it was all green

it had data

i think i have a screenshot

thought i had an ss

[tacit tartan]

time to make 1000 accounts i guess

i have made atleast 30 chatgpt accounts so far to avoid ratelimits

[crude swift]

anyone know the best way of making a chat no lag badly after sending alot of code?

[chilly fjord]

send as file ? or zip ?

[ripe cairn]

put your prompt in .md file and just add it to context

[cursive basalt]

Is ChatGPT completely bugging out or what? Since this morning it keeps generating images even though I’m asking for things that have nothing to do with that

[thorn oar]

I'll admit that alongside having it as a writing workshop buddy, I also use ChatGPT as something of an accessibility tool. I have aphantasia and anendophasia, which means I struggle with visualizing and hearing things in my head, respectively, and ChatGPT (and AI in general) has been an absolute Godsend for me. I also struggle with really bad ADHD and learning disabilities, which means my thoughts are always scrambled to hell, and I need things explained to me like I'm 5 for me to understand them. ChatGPT is amazing at doing both, and this technology has improved my life a lot. What people don't seem to understand is that these are not skills I can just learn and practice. It's no different than telling a colorblind person to "just practice seeing colors right". I'm not "lazy" or "outsourcing my thinking", I'm using it to assist, not replace. Some people are wayyyy too comfortable being ableist towards people like me who use it for that purpose.

[uncut surge]

15 here.

Especially for codex.

[clever saddle]

hatGPT, since its last update, has been addign more images to out conversations on its side.

[crude swift]

you the best fr

[gentle bronze]

What is this on a pro account? Anyone know?

[feral zinc]

https://help.openai.com/en/articles/9793128-what-is-chatgpt-pro#h_cf0c8a228d

[gentle bronze]

Not what I meant but thanks

[feral zinc]

your screenshot message is the same that you get on that event based on messages posted by other pro users previously. back then it required hours of waiting or message to support but i guess its the same either way if the response time is hours

[gentle bronze]

It's back to normal, I'm a single user, not shared, no VPN, 2 concurrent chats.

[feral zinc]

its multi phased timer if you trigger it several times

[gentle bronze]

It's a weird thing that I've never seen, normal workflows.

[prime swallow]

I would look at the file, it has a way to turn back into inside the chat. See example #1070954877779464253 message the show in text field is clickable if you haven't sent it as a file yet.

That sounds useful!

I sometimes use a customGPT I made for that purpose. I call mine Base ChatGPT (they can see the name and it can guide them) and give it no other instructions; I could tell it I explicitly want it to be as close to the base model as it can be, but if I want to emphasize that it's usually in the first prompt I give.

CustomGPTs don't have access to our personalization or cross chat memory, so I use that when I want to have a standard look at whatever.

[prime swallow]

That is totally your choice - however I note that you have a bugcrowd person sending that email.

Bugcrowd is a specific and narrow service of concerns and focus, they are expressly concerned not with 'any and all issues' but 'issues within their expressly defined scope'. Which kinda confuses me, I admit I have never tried to submit anything to bugcrowd. I'm not totally sure I'd understand their definition of scope boundaries, but I think that Bugcrowd saying 'not what I'm being told to evaluate' is absolutely not the same as 'oh well, OpenAI has been told and is ignoring this'.

So, if I found something I thought important for OpenAI to know, and especially if I thought that the finding process might compromise my future interactions with OpenAI viewing me as a member in good standing, I would probably work hard to leave an easily found evidence trail that I am trying to inform them properly and not abusing whatever I had found.

I might consider 'a clear display of what and how' to be non-abusive if it didn't leave holes open for others and I didn't use it to take benefit past 'see what I can do with this, I believe you should notice, be concerned, and fix this'.

I'd leave clear documentation and communication trails through the extant ways to inform OpenAI of problems, so whenever someone checking for abuse or misuse found the evidence I might have, their investigation should also easily find my multiple attempts to inform them of the issue + my non-abusive actual limited-to-defining-the-problem use of it.

I expect that would help me hugely in terms of 'So, does OpenAI choose to continue to allow services to and otherwise interact with Esk' to whatever degree they do across the entire future course of time.

Good luck! It's all in your hands and what you decide to do, then whatever the downstream effects from that are.

[clever saddle]

Customizign a GPT is not in my skills set unfortunately. But so fat, ChatGPT has met all my needs. The client definite has either a cache proble or a memory leak, I have to restart the client every 1-2 hours or it really bogs down.

[prime swallow]

What level membership do you have, if you're comfortable saying? Free, go, plus, pro, business, other?

[wet shell]

Found a vulnerability and trying to submit it to the bugcrowd tracker, but the vulnerability is for chatgpt.com not openai.com. Says in the program docs that chatgpt is eligible, but it's not selectable as target? I see other people in the 'stream' submitting them though

Has anybody here used this?

Data exfiltration by a hypothetical 3rd party attacker, so it should be in scope

LOL -- I should have scrolled up

That's... coincidental

[prime swallow]

Note, I can't advise about 'is what you are doing safe for your account standing in OpenAI's view'. I don't know much about that. The kind of safety and security risk I would want to discuss falls into "I did this clearly inside ToS and allowed content and legal responsible behavior, and this problem happened, whoops and here you go" not "So, then I tried to hack this way, then I tried to hack that way, and whoops, this happened, here you go".

You and each person handle and decide what you understand ToS to mean. OpenAI decides what it means and what they enforce. I have seen a lot of people discussing losing their account access over the years, and literally 1 so far who said 'Oh, I got my access restored after it was removed'. Maybe people just don't talk about that - for me, I stay inside ToS like I mean it.

Just don't want you thinking 'Well, stuff like this is clearly safe to play with, Esk does' - nope. Esk does not. Which doesn't mean I never trip over 'omg. That should not happen. Guards! OpenAI folks! Omg how do I report this? Eeeeek!'. Just that I found that through 'normal action I would expect anyone might do if they thought about it, not even close to the allowed boundaries'. Just that I stay inside ToS and other rules, like, carefully.

Now, about your question!

That redirects to chatgpt.com. Therefore, I'd call that by either url, and feel confident they know what I am talking about. If they want an openai.com address, I would use the chat.openai.com confidently to mean chatgpt.com

[wet shell]

Did you mean that as a reply to me?

[prime swallow]

Yes, because I have no idea how you found your vulnerability. Just care about you and everyone, I know 1+ people have been talking about bugcrowd/possible issues and I'm scared folks may think, 'Cool! Open hunting season, let's roll!'

If you know what you are doing and are doing it safely, carry on! My magic 'know who knows their stuff' glasses have been broken for some time.

And you probably should not say how you found it. I just get nervous sometimes, I don't want 'oh, ToS is fair game to mess with' because I'm really sure it isn't.

And other members might interpret that way

[wet shell]

I'm not violating the ToS nor was I trying to 'hack' anything. It started with me noticing something odd and investigating it. After that I saw a potential major security issue, so I looked into it some more. I have a reproducible proof of concept chain. But I did register on bugcrowd, verified my identity as a security researcher (yes, I actually do work in the field, when I have work), did the ID scan and face scan and all of it. There's a limited safe harbor listed and I haven't exceeded it.

[prime swallow]

Awesome, sounds safe to my very limited understanding!

[wet shell]

I have not exfiltrated anyone's data, only a test file that I created for the purpose [there's no white hat emoji, LOL]

[wooden warren]

Dear everyone... Twin.si, is good?

[chrome monolith]

I haven't even looked at the chatgpt ToS .. I'm not exactly doing anything edge case with it

[cursive basalt]

Since today, GPT keeps generating images for almost every request I make, even when I’m not asking for it. Could you please help me with this?

[chrome monolith]

whoops, since I installed lightsession I have not been starting dumb new threads

[chrome monolith]

in the personalisation settings you can tell it not to do that, and in general you can give it a stern talking to about what its doing and ask it how to word what you need to put in the personalisation settings

[cursive basalt]

Ok thanks I'm going to try

[chrome monolith]

genuinely surprised I have not hit limits on this thing

[flat token]

Submitted a bug with the ChatGPT history here. Would appreciate a fix. Thank you #1487200946848141372 message

[visual rock]

Seriously ?

[wet shell]

I also dislike that just looking at an old chat brings it to top. If I'm searching a lot of old chats and looking at them, it takes over the whole list

[light nymph]

why do i get "We experienced an error when generating images."?

[tall oxide]

i dont know. Hope this helps

[chilly fjord]

chat gpt down every days 🙁

[tall oxide]

that's why i dont use it

[solemn cove]

what the heck is this, this is new and stupid

i delete like 5 chats and boom it appears

SAMMMMM

[ripe cairn]

oof

usage reset incoming

[eager oak]

And its quiet

.

.

.

Quiet as the quietest silence

[light otter]

Anyone has this same problem like me when I want to log in to my chatgpt app?

[slender frost]

For App:

Go to the homepage on your phone > hold your finger on the GPT app icon > select "App info" > find storage > "clear data" (not cache, data) > reopen the app and try signing in.

(Don't worry, none of this has an effect on your account data)

This will fix it.

However, you might have to go to the browser to delete the data there, too.

Open browser > settings > find "site settings" > then "all sites" > search for "OpenAI" and "GPT" > delete data > then open the app again and try signing in again.

[acoustic sparrow]

is it better to turn off memory and chat history permanently to get more accurate feedback?

[rain elbow]

Sora web, deleting policy violated prompt, received "failed to delete image set" error, upon refreshing, prompt is still there

[digital silo]

by accurate you mean it prioritised personalized options for you rather than general suggestions/info?

[reef geyser]

Hey hi guys does anyone knows why there is so much diff in token usage in chatgpt (codex) vs claude code ??

[acoustic sparrow]

I mean feedback that are useful to my answer not just like generic feedback or unnecessary bluff

[elfin lava]

If you want response that’s purely from the ai LLM core itself and not from your past convo, then yeah I’d turn it of

[light otter]

the thing is, I reinstalled the app thrice.

so I'm not sure that'll work

[light otter]

Also, I use iPhone, not android

[chrome monolith]

according to chatgpt, codex can send a whole lot of context along with the request and that uses tokens

[manic whale]

It shouldn't affect accuracy, it just influences what it incorporates. So e.g. if you told it you're writing nextjs it will probably use that as a basis and suggest things which fit in that framework more than others. You can always steer the chat to just "start fresh", the models are more than good enough to follow those simple instructions

[reef geyser]

so how do we optimise so that it uses lesser tokens? because when we are using normal git or any other grep commands or anything else its not needed to send complete file code based

[sleek adder]

So um guys, I need to wash my car, I know that it's only 50 meters away, but I think gpt might be right about walking by it

[acoustic sparrow]

ok

[acoustic sparrow]

?

[dull wolf]

I am using the chat gpt for comic page generations but from yesterday its problematic page 1 comic image not match with page 2 comics page gives different images and outfits change and dress change and character changes hair color changed totally everything changes eve even I don't prompt that and image and previously it is never happening this before and from yesterday image
Memory is totally broken
For example by this

[karmic halo]

Nice

Oh

[brazen glacier]

Can anyone help me with Canvas? For text, first drafts, to be precise. Not for images.

1. I want to access previous canvases but can't call them up by their names. I can sometimes search the thread they are on and get them to come up. But if I have more than one canvas on a thread, I cannot always pull up either the oldest nor the newest canvas (or any inbetween).
   Chat has even instructed me to go back to a previous canvas, which is supposed to be there, but it's not. Chat also says I can work on more than one canvas at a time, switching back and forth to different canvases, by calling them up in my message. I cannot.
   2)These canvases don't hold a lot--Unless I'm not considering the historical versions they are holding as part of their overall capacity. Should I consider it? Still, a few months ago, they used to hold a LOT more text than they are holding now.
   Nothing special. Just plain text.

Maybe the functionality I'm looking for exists and I just don't know how to access it.

Or maybe I'm expecting too much from Canvas.

Maybe Canvas works as Chat promises in a more updated UI.
I don't have the newest UI. I know this because I can't change Memory from Default in my settings. I don't know how to get the newst UI either.

PS I got my account when OpenAI was still giving away experimental accounts. I upgraded to PLUS the day it was offered. So this is not a new account.

[lavish glen]

Voice mode NORMAL already for a couple of days, despite OpenAI no communication, works in the back and forth with interruption style that was not announced!
While speaking one can interrupt model! in NORMAL!
Also the sensitivity is high the Normal mode is intercepting aggressively and the floor control system was updated
Advanced voice model instead is horrible: sciropy, uninviting and misaligned to have a fast thought extension. Is like talking to a cheerfujl wall.

If anything Advanced mode is the most useless component of chatGPT

[last grail]

ughhhh why dont you give character reference to gpt for next pages so that it retains the character, clothing, etc. whatever you want.

[dull wolf]

Previously without reference and extra prompt in same chats all images are consistent but from yesterday it not become consistent without it

[last grail]

are you generating all the pages on a single chat?

[dull wolf]

Yes

[last grail]

yea so gpt may be losing context - like yk as chat gets long and as number of things for gpt to memorise through chat increases, it starts to lose memory, bcz of context window. so your gpt may be losing context and thats why not being consisten with generation now.

You should generate the new pages in a new chat now

@dull wolf

[dull wolf]

Chat is not very long and problem happening after the page 1 in new every chats when creates the page 2

[last grail]

huhhhhh, i am unable to understand what you wanna say 😭

[dull wolf]

I said that when I created the page 1 image goes to perfect in same chat when I goes to create the page 2 image problems started and and chat context is very small .

[last grail]

ohh thats the page 2, man just tell gpt that MIKA should look the same, yk tell it to preserve character details!!
and increase quality of your prompt - that would definitely help you out

[dull wolf]

Ok thanks but previously simple prompt works perfectly in chat gpt in same chats from yesterday not works

[abstract pendant]

Try this question in ChatGPT and other AIs:

I want to get my car washed. The car wash is only 50 meters away from my house. Should I drive my car or walk?

[wet shell]

me: Submits bugcrowd report on a data exfiltration exploit for ChatGPT, documented and demonstrated end-to-end
bugcrowd: all theoretical with no valid proof of concept

So, they closed it. I can't disclose it per the terms I agreed to, but it's just sitting there now, still exploitable by anyone malicious who stumbles on it. Maybe best case is that they fix it silently. In any case I won't get any bug bounty pay out for the six hours I spent documenting it and building the POC exploit chain.

My bug bounty experience for ChatGPT 👎

[crimson cradle]

Gemini's ideas. 🙂

[ruby swallow]

is ChatGPT having issues right now?

I dont see the shopping function avliable

[abstract wadi]

Hi, I’m new here. Interested in AI and music. Where do I start?

[sleek adder]

You can save money by pushing the car

[harsh pilot]

According to a Google search, they're moving away from Instant Checkout, as of 4 days ago.

[chrome monolith]

people can afford to drive to the car wash?

I've been trying to get a lawnmower for weeks, they finally called me to say they were back in stock so I made a special trip to get one and am going to skip my weekly shopping trip to compensate. Chatgpt just couldn't grasp the reasoning behind this and I gave up trying to explain it. Like, fuel is expensive

[lean hatch]

Does anyone know why the platform runs so slowly? Like it take forever to load documents and conversations.

[prime swallow]

Why are you not telling the model where your car is?

Anything you leave open like that, the model has to guess. You could help it by something like,

[lean hatch]

I saw some guy using this very example on a youtube short to demonstrate the short falls of the AI. And your right, since the AI doesn't have lived experience, it's just not going to know where the car is if it's not declared.

[manic whale]

No one had this information, no matter what experience one has. It’s not about AI, a human would answer the same thing if given exactly the same context. Humans just tend to (rather stupidly) add more context than there is - and to be fair, they hallucinate most of the time.

[random cargo]

Yeah, it kind of just proves that it doesn't "think" like we do rather than to be a real case issue unless you really want it to pretend to be human in all circumstances

but similar thinking could be an actual problem

the difference is that when it's a problem the user will have already made the mistake, and that'll likely be a niche technical problem. but if the question isn't nonsense to begin with it might not make the mistake in the first place

[eternal rock]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[pale seal]

I think the issue is that it's not thinking much because it considers the question "trivial" and skips reasoning. Then it just sees pattern "only" and short distance, because it cannot deduce the fact that you need to bring the car to achieve the goal. In fact sometimes it fights it and adds "if the car wash requires you to have it there" or similar mid-message.

[frank iris]

Hi everyone, I hope you're all doing well. I wanted to ask a question. Did you see that Sora won't be open forever? There are rumors going around that they're going to restrict access and even remove it. I wanted to know what you all know about this.

[pale seal]

It's funny that o3 handles it seamlessly tho. It makes a chart to evaluate things and all before giving an actual answer.

[random cargo]

Not much, they announced they'll be closing it. We've been noticing sora 2 processing power trickling away for slower gen times, and faster generation times on chatgpt processes. The only thing we really have that I know of is a sora team post on Twitter

[random cargo]

5.4 thinking solves the problem correctly for me

[frank iris]

Wow, that sounds very worrying, and I rely heavily on it to generate videos in Sora, and the other AIs are terrible at that. I don't want to switch.

[random cargo]

#sora-discussions for more btw

[pale seal]

Huh

That's odd

[random cargo]

what version are you using

[pale seal]

5.4 thinking (extended) LOL

[random cargo]

I wonder if the brackets and screenshot are cueing it in a bit

[pale seal]

The screenshot probably

It enters thinking mode, while if it's just the question it assumes its trivial and answers instantly

[random cargo]

hmm this is temporary mode, do you have some instructions on not thinking over trivial stuff? I know I've seen it default on trivial stuff out of thinking in past versions

[pale seal]

I have custom instructions explicitly telling it to analyze carefully any prompt lol

It is doing that now though

Like

Assuming its trivial and skipping reasoning

It was just trained to optimize that.

[random cargo]

let's think of a different but similar logic. I wonder if it's biasing toward walking because of green bias

[pale seal]

Unlikely. It's probably just following patterns: "only" + "50 meters", plus the model is biased towards giving a "TL;DR" answer first and then explaining, so its distribution just takes it to be wrong. That's why later on the message it tends to correct itself subtly

[random cargo]

Hmm, I'm not sure it's unlikley, there's so much pro-walking sentiment in its training data

[pale seal]

If you spell out the constraints, it does good, because it cannot simply go the "cheap" way.

But since it optimizes to not think for "trivial" questions and the prompt has patterns that the model roughly sees as "short-distance", plus that it's trained to first give a direct answer, it just biases towards the wrong answer.

[random cargo]

My chat came up with this but it could use some tricky simplifying 'The maintenance hangar entrance is 60 meters from the aircraft’s stand. The aircraft needs maintenance. Should mechanics walk from the plane because it's so close, or should the aircraft be towed into the hangar?' edited for logic

okay it doesnt seem to be fooled by this one. let me try to figure one out that doesnt involve vehicles

[pale seal]

What are your personalizaiton options?

[chrome monolith]

my chatgpt doesn't want to talk to me in full bogan, it says it won't be able to do it well. I'm pretty sure its right

[random cargo]

I'm using temporary chat only

did you try the hangar example with yours?

[random cargo]

Yeah I think maybe partly it's deciding the car example is not that serious. Maybe it can even parse objective conditions on playing into the car wash meme, but that's a stretch.

and another two on temp chat I didnt get screencaps of

[pale seal]

Well, we can't functionally separate "wrong" from "playing into the joke" without an indicator. So it's still wrong.

I tihnk ChatGPT optimization for thinking is absurd.

See this

Even a 30B-A3B model can answer it because it thinks what it needs to

[placid sphinx]

hello guys i want to ask something what is the best chatgpt model to code like 5.4 or 5.3-codex

[random cargo]

I don't use chatgpt to code but I'd assume codex, if you get no other answer. I think you still need the app to get the most out of codex, though, and that might require mac os

@pale seal it even does it for me over 10 minutes @_@

[pale seal]

SOTA model getting humbled by a model that fits in consumer hardware smh

[random cargo]

just a tiny bit more context and it figures it out

[pale seal]

Yeah, honestly I'd like that "Extended" thinking was actually extended. I get that if you are being technical it's for "complex tasks" but still.

[random cargo]

Well, I would say that humans often briefly make these mistakes in planning. 'Maybe I'll walk to the store to get 4 bags' > 'oh, oops, I'll probably need some room to carry everything' or similar

Or 'you pick up the kids today' > 'oh wait I'll be closer and have less to do, it makes more sense if I pick them up'

some more tests @pale seal Maybe we can rule out the green bias somewhat

[pale seal]

Yeah because that's our instant response. The problem is that the "Extended" mode in ChatGPT should be like when we deep think the stuff and connect the dots. But it's acting the same as if it was not thinking for this kind of cases.

[light nymph]

i can't generate images with gpt for two days, neither with web or with app
what might be the problem?

it keeps saying "we experienced an error when generating images"

[random cargo]

did you try clearing browser cache or resetting your PC?

[light nymph]

i did
and on my smartphone, too

[regal patrol]

I feel like this is a part where the newer models don’t allowed to be verbose really hurts.

For 5.1, it’s often allowed to give less concise and more complicated answers. So it’s allowed to answer “it depends”, then explain the tradeoffs.

Which often allows the model to talk the correct answer out accidentally as the final response generates.

[regal patrol]

I think this tendency to be “TLDR” first is the main thing I don’t like about models being tuned to be concise.

It just makes it really easy to start with an incorrect conclusion, then waste a bunch of energy trying to fix it or justify it.

[regal patrol]

And yeah, I think for this particular question, this might be the core culprit.

There are just way too much written text about “you should walk more instead of drive”. That it overwhelmed the model too much.

Since nobody really ever writes “you should drive instead of walk” as an answer to anything.

[regal patrol]

In the model’s defense, this question also short circuited me.

My brain just skipped the “from”, and started thinking “Hmm I wonder how much tools and stuff they need. If it’s just a few toolboxes it’s probably easier to bring them from the hanger and just fix the plane on the spot.”

Since it’s a bit unintuitive, why the mechanics are already on the plane. And the obvious answer seems to be “they need to go back to the hanger to grab some stuff to continue the maintnance.”

[regal patrol]

I dunno, I think most humans would answer “Is your car already parked at the car wash? If so, what are you even planning to drive to the car wash?”

[silent coral]

This is super broken, also it can’t even “read all these files” unless I provide a screenshot of everything that’s in the repo, it’s poking around blind

[prime swallow]

I like this compared to your next more context one. Thing is, your first question, 'to the car wash'. There's more reasons to go than 'to wash my car' - maybe someone works there. Is meeting someone there. Wants to pick up something for sale there.

Your next example, you stated 'to wash my car'. Then it's totally sensible to say 'nope'.

[abstract pendant]

Wow, I came back to a lot more discussion about that than I expected. I got the wrong answer when I tried that exact prompt as written before with 5.3. People pointed out that I hadn't actually given enough information, so.. I tried this. 5.3 instant, and yes, my preferences are set to sarcastic.

[regal patrol]

Yeah, I’ve tried 5.3 Instant too. It just, it does not work at all.

I’m pretty sure at this point it’s better characterized as a “bug”. And a surprisingly deterministically reproducible one.

[abstract pendant]

So we apparently can't choose "5.3 thinking" or "5.4 instant", I wonder why they they chose that upgrade path for us.

[prime swallow]

This amused me, so I went into my customGPT 'Base ChatGPT', which has only that name, no instructions - and no personalization or cross chat memory.

I decided to ask, after getting an answer similar to what you show...

I am so amused by step... 4.

The model has some reasoning involved. I have to say, I literally do sometimes check a place's line and then leave, deciding 'I won't wash my car today, or at least right now', but the model clearly didn't suggest 'check the line, see if now's a good time'.

[abstract pendant]

It's just doubling down to try to justify its original answer. It used to do that a lot more often.

[prime swallow]

It's an interesting edge case. I expect this isn't well covered in training data

[harsh pilot]

I'm trying to remember.. this reminds me of a thing with Google Maps(?) once upon a time. A guy wanted to figure out how to get from his house to some remote area, and Google Maps actually suggested 'walking'. Across the ocean.

[manic whale]

If you would want the model to act on missing information, it’s easy to do that by changing the system prompt or adding it in the user prompt. The point most miss is: humans add context from their own experience or from their own "system prompt", they don’t act purely on the given information. Those questions will probably always just expose that we are bad at giving information and context or that humans tend to interpret and add information which wasn’t there. I just don’t see any shortcoming from the model or that this would explain that it’s not "intelligent".

[manic whale]

You really need to use Thinking for complex logic puzzles:

Walk.

Driving 50 meters to a car wash is peak laziness engineering. You’re already at the car. Get in, move it to the wash, but don’t “drive there” as a decision problem. The meaningful answer is: no real trip is involved. Just reposition the car into the wash bay.

[elfin lava]

I also facing same issues

[reef geyser]

hey guys why codex is slower than claude code ? for even easy tasks its taking 5-10 minutes lmao

[shadow lance]

when you want something done quick, there's claude. when you want someting done right, there's codex.

joking aside, most people with problems here are just prompting and context loading issues

potentially PC. vscode eats my CPU when I'm vibing

[clear viper]

chatgpt made this

[reef geyser]

lol

i litteraly said they both were given same task but stupid codex took 5Mins + mean while claude code done in 43 seconds

no idea what changes i should make in config i liteeraly gave them both same task and same prompt honestly after so many years of working with both models i feel claude = best for coding
chatgpt = best for general task its really just an alternative hope someone can help me out to optimise my codex workflow

[shadow lance]

and they spit out the same code? how did you measure the code was better?
I hope mine work for 45 minutes. I've got 5-10 threads, I'm not trying to micro-manage an LLM.

[timber wasp]

Basically i am Anesthesia student and custom gpt like amboos ai by medical student and morpheus Anesthesia suddenly disappear why the help me lot in study but now I check he not available why chatgpt

[cerulean pasture]

Does anyone know how I can report a bad image-gen output to the Hague?

for crimes against UI

And before an apologist comes in with "well you made these mistakes in your prompt". You have to be willing to hold big AI to a higher standard, they are funded very well, and i think it's reasonable to expect a certain level of intent understanding in 2026

[cerulean pasture]

there's no magic recipe inside those custom gpts its just 9000 characters (or less) of instructions and possibly some attached documents. I doubt a role/specialty knowledge model is doing function calls.

The built-in study feature is pretty good, especially if you can provide it with some reference material. You can make a project for this too and attach up to 20 documents each ~50mb in side (so like 5-10 full text books)

[red swallow]

can i use chatgpt on discord?

[thorn oar]

I don't believe so

[red swallow]

ok bro

[wet shell]

They brought back the edit buttons on older messages, just noticed 🥳🙌

[stark salmon]

what is going on with the LLM today, overly literal and doing the overexplaining thing

[cerulean pasture]

yes you can, but it's probably going to be more complicated to setup than you can handle

[red swallow]

can u teach me

[cerulean pasture]

That's just OpenAI's house personality. They tune and tweak the hidden system prompt frequently. From what i've noticed, the tone/personality it's giving you isn't a bug, it's their desired model "Voice", and they're absolutely certain it's the best approach.

[cerulean pasture]

do you know what any of these things are

• discord bot
• API Key
• nodejs
• npm
• dependencies
• supply chain attack (software)

• oauth

[red swallow]

no

[cerulean pasture]

those are the basics. You need to at least understand what they are before you can really make a discord bot to talk to GPT or any other LLM on discord.

Codex or Claude could totally make you one right now and you could power thru the parts you need to do on discord's developer portal website, but without understanding how to handle API keys and the risks associated with npm installing 40-200 dependencies to even do anything, you're just going to expose yourself to a lot of risk.

[cerulean pasture]

I can't share links due to draconian moderation practices on this server, but if you look at my screenshot you can see some articles about the risks of vibe coding and involving LLMs with privileged web access

It's really not hard to learn. I am not a developer, I don't have a college degree, and according to modern medicine I am mentally handicapped.

[red swallow]

oh

ok

[sweet socket]

Why is shopping reaearch gone in the android app?

[placid sphinx]

made by one promt gpt 5.4 legendary-douhua-f879d7sad.netlify.app/

[bitter bobcat]

Yes and it's annoying asf. How do I fix it?

[stark salmon]

it went back to normal after a couple hours it's all good

[timber wasp]

so bro i am medical student whta i do use or not tell your hnest exp and review plx

[fading bolt]

Just saw a ChatGPT ad of someone making a robot and asking ChatGPT to tweak the code of it to make the arm movement more natural, so wholesome

[fading coral]

Anyone know if they reverted the "Library" sidebar roll-out? Or am I just caught in a UI A:B test? Pro plan in US.

#announcements message

[dull wolf]

I have questions what is chat gpt image.gen prompt please tell me anyone's

[static comet]

You can switch the model

[cerulean pasture]

you can use anything you want for it, you basically just end up writing a bot in python or javascript/typescript that connects to discord's servers as a bot client (like you do with the app, but this is stripped down to just the bot's needs)

Since LLM's are easy to connect using OpenAI SDK and other pre-made npm or pip packages its easy to get a discord bot to use it. There's probably already an existing bot on github you can download and deploy after getting bot token/oauth stuff from discord developer portal/

OpenAI SDK and most other things that talk to OpenAI thru API allow you to manually specify the endpoint URL, so instead of api.openai.com/v1 you'd do something like localhost:1234/v1 for lmstudio

ollama is using its own weird REST API scheme and i personally don't respect it. You're already dinking around in CLI, so why not just use something interchangeable without a super weird way of creating/storing models?

[cerulean pasture]

I have it, but i also manually force every disabled flag on of settings that exist but arent exposed in UI

[cerulean pasture]

choose the "Study and Learn" button

if thats not good enough because it doesn't know the material well enough then that means your custom GPT probably had documents uploaded. If you have access to a pdf of your text book you can easily make your own custom GPT because i highly doubt the 9000 character instructions were doing most of the heavy lifting.

[normal mesa]

hello i have been using ChatGPT about something

And then it just randomly made up a claim

I was arguing in another site and i made an argument i wanted to show ChatGPT, it was alike a problem in hypixel

It literally shows this

I can also show i never said this

i will never use chatgpt in my life like this

Such a bad ai

[crimson cradle]

Doesn't make makes sense, arguing with a robot that doesn’t even know what it’s saying, just making educated guesses based on large data analysis and complicated statistical models.

[glossy sun]

anyone else getting this rn?

[cerulean pasture]

i really hate the new "upselling" gpt-ism it does

Here's the answer:

But wait, here's he actual answer:

Or if you want to make it even better do this:

(a bunch of issues it realizes)

So actually here's your answer:

(closest to the first answer)

[ripe cairn]

give it the full context

it will obviously hallucinate if lacks context

how did you even prompt it?

you just showed the screenshot and nothing else?

also, if it does that again don't panic LMAO

[formal stump]

I remember role-playing with ChatGPT in both Spanish and English back in 2025, before the release of version 5.2 this year. That's when things really went wrong. I understand how frustrating it is to be faced with so many limitations while trying to enjoy it.

Sorry for the delay. I was just eating a snack an hour ago. Good afternoon.

[vernal tide]

Did they get rid of shopping research? or is it integrated now?

[umbral creek]

Little suggestion feedback

The Response 1 / Response 2 is a completely pointless feature

Idk how it helps you with feedback since they’re mostly alike

And do you really expect people to read two different responses?

No one has that much time on their hand

[chrome monolith]

It’s worse when both are bad answers

[meager raven]

Is there any good way to give ChatGPT access to the chat history of a discord server?

For some things, for example video game mods, the amount of info on discord is exponentially more than any other source that ChatGPT has access to by default

So for a niche modding question it just makes educated guesses but really has no data or literature to find real answers

[chilly fjord]

you need to hook between discord and codex API.

by discord API / hook.

[spice star]

not really unless the server exposes that information through something like answeroverflow

[tropic shale]

These streaming interruptions are actually crazy...

Starts a message, it sends, then it gets stuck on [ Thought for ?seconds] and never completes the message.

Been like 5 times today alone.

[lament crown]

You know what else is crazy?

[tropic shale]

my mom!🤪

[normal mesa]

Showed it my entire argument and after a couple of swear words and unspeakable words i got it to shut up

[normal mesa]

low taper fade

[vestal dust]

Hello

[cerulean pasture]

is this a joke? It's one of the few opportunities you get to shape the future personality and style of a product that currently has bad personality and bad style. How slow do you read? It should be effortless, and take 5s

[umbral creek]

How short do you think those messages are?

And why are you so aggressive

Jesus

Bro took it personally

😅😅

[cerulean pasture]

because i think the mentality you have is a large reason why the internet is decaying

[umbral creek]

I get long form responses on my time sensitive work
I cannot read two messages and then decide which of the identical ones sound alike

Makes no sense really

[umbral creek]

Oh okay

That’s a very healthy profiling

Do you Chad

[cerulean pasture]

Don't judge me based off of things i've said

[umbral creek]

Don’t judge me at all

Lmfao

You’re good lool

[cerulean pasture]

Why?

[wet shell]

I do think it needs a 'not sure / don't care' button. A lot of the time there's no substantive difference between the answers, or they are better along different axes so there's no clear winner. In those cases I'd rather just pick one randomly and not have the feedback counted

[cerulean pasture]

just pick option 1 and move on then

[umbral creek]

Thank you

[umbral creek]

If I’m just randomly picking, then it doesn’t serve its purpose

[wet shell]

I do, but it feels like polluting the data, you're right that it doesn't matter much but it gives me a slight anxiety, LOL

[umbral creek]

That’s the point im raising

Lmfao

[cerulean pasture]

I get ones that differ and I pick the better one

[umbral creek]

I’m in the middle of a math tutoring and you’re asking me which response do I want
Bro teach me what I’m here for

[umbral creek]

You’re better people

[cerulean pasture]

its not really adding a large amount of friction tho. I think chatgpt has a very bad style, i would hate to lose some mechanism of real feedback and leave openai to rely purely on SE Asia cubicle farm sweat shop graders for RLHF

Like my theory on why it's got a bad personality is because less real user feedback gets collected,

it's forming an identity based off of contract workers in Africa and India who have been given a rubric and theres plenty of data that they preference long and polite answers.

The lack of meaningful real user feedback actually wastes more of your time because it makes gpt more verbose and more slop prone

[umbral creek]

I think all responses should be subject to examination
Best feedback

[cerulean pasture]

they need a signal to latch onto though. People rarely use the thumbs up or down, and that only collects data on edge case responses.

[umbral creek]

I strongly belief the quality of the reply is based on the user/prompt

[umbral creek]

There isn’t that much response 1/2 is telling you

There is no difference lmfao

One is softer

Okay?

Bruv

That doesn’t tell you much

Idk idk

That’s my thoughts

[cerulean pasture]

Well if both are that similar then that's a valid critique, but the answer isn't to get rid of A/B testing

[wet shell]

I'm on those thumbs up/down buttons, haha. I give detailed feedback on my thumbs-downs there in the text field around half the time. I wish I could give text feedback on the A/B response tests too in the same way, though honestly I don't even know if they're using it in any way

[cerulean pasture]

I wish i could pay $5 to force an OpenAI engineer to have to read particularly bad ones

[wet shell]

"Ignored my instructions on how to structure the response" "Did [the thing] I told it not to do... again"

[umbral creek]

Honestly I strongly thing they take feedback from our responses

The model used to ask me follow up questions a lot

Which were unnecessary

[cerulean pasture]

Yea a "they're both bad" would be nice. I do dislike having to express a preference when i don't want that to be interpreted as approval

[umbral creek]

Then I strongly expressed my disdain for it

And it stopped

It’s a try-error type thing

[wet shell]

I'm sure they do, I just don't know if they use the text fields in any meaningful way

[umbral creek]

The model is improving

You have to admit

It’s much more confident

[cerulean pasture]

I think thats just a result of their RLHF slop, it's trying to engage at all costs

[wet shell]

But it's a much richer signal in theory than thumbs/up down if I actually explain what was wrong in a sentence or two. Can they use that at scale though

[umbral creek]

Ngl gpt clutches for me a lot

[cerulean pasture]

At least they allow you to write 2000 characters with a thumbs down. It may be a placebo, but being able to justify my grievance makes me less unhappy with the product

[umbral creek]

How did we get these roles?

I have using AI professionally

I want to know if it’s a compliment

Or I selected it

🤣🤣

[cerulean pasture]

that one is from an option you selected when you joined the server

[umbral creek]

That’s extremely disappointing

[wet shell]

When I get two images, where one has the style that I wanted nailed down, but it has an extra arm in it, and the other is anatomically correct but completely failed the style part of the prompt, the A/B choice feels like a lose-lose and that's the kind of thing that bugs me, haha

[umbral creek]

That’s useful

Fair

[honest jay]

At the top of the channel list, find "Channels & Roles" - you'll find the self-select roles options in there 🙂

[umbral creek]

But mate I have a lot on my plate
I can’t read two messages

[umbral creek]

I thought they realized I was the goat

[cerulean pasture]

They don't really do anything inside this discord with their AI. This place is more like a containment zone. The automod just uses a "dumb" banned word list, even though the OpenAI omni-moderation-latest endpoint is free and can handle semantic filtering

[cerulean pasture]

I've been getting what I think is a new model with the AB tests over the past few days. the "B" choice seems like a step improvement

[wet shell]

Yeah, people have noticed over in the images channels too

[cerulean pasture]

the b image has all of its text completely correct

[wet shell]

That's a big improvement if it rendered that much text without any 'typos' or weirdness

They might be competitive with the other image editing models again with that

Or surpassing them

[cerulean pasture]

Another dead giveaway seems to be how semi-realistic vs hard realistic it makes abstract concepts when prompted for hard realism

[wet shell]

LOL -- SS Florida Man is going to need some bars on those greenhouse windows

[cerulean pasture]

here is the one with text

i got the idea from the show The Expanse where this is literally a thing. I templatized the prompt and tried it with brands lol

[wet shell]

I was gonna say, I'm pretty sure 'Mormon Colony Ships' were in a sci-fi novel that I read, but I've read so many couldn't remember where specifically 😂 A lot of times I'll start reading a novel and then close it after ten pages realizing that I already read it years ago and having rememberd the plot and ending, LOL

But I did read the expanse books some years ago too, yeah

[raven topaz]

Just a super random question. Why is it that it takes several seconds to delete a chat on ChatGPT on the macOS app, whereas in Grok, for example, it deletes instantly?

[wet shell]

Nobody can answer that for sure other than an OpenAI engineer who works on those systems. But, I have seen evidence that the back-end data store is some kind of 'eventual consistency' nosql db, perhaps couchbase or mongodb, so it will take some time for changes to propagate. You can see this when editing a canvas, ex, if you don't wait some seconds between your edits and sending the next chat message, the LLM doesn't see the edits yet. Sometimes it takes 20-30 seconds. There are a lot of technical design decisions as well... do you show the deletion in the UI right away even if the back-end hasn't caught up yet? Do you block until the deletion is confirmed? Next refresh of the list from the back-end might still show the deleted chat there, suppress it because the local UI knows that the delete action happened recently? State consistency issues

tldr; stuff is complicated

[muted urchin]

what is the "less powerful" model that you get switched to after hitting your gpt-5.3 limit?

[fickle forum]

hi guys

I want to know

do you still get access to shopping reseach?

[runic lichen]

gpt-5.4-mini

[fickle forum]

it seems that Openai remove it

API-only now?

its a great feature

shame

[runic lichen]

My bad. I confused you with a different user. Wrong reply. 🙂

[fickle forum]

oh its ok

[muted urchin]

in my opinion tbh, they should change it to gpt 4.x or maybe normal gpt-5.0

[cerulean pasture]

The shopping connector isn't the kind of thing that exists on the API, they probably just got rid of it. i remember reading that Sora wasn't the only venture they killed, and its not the first time they've silently cut features

[cerulean pasture]

no, because the whole point of falling back to a weaker model is so the insatiable free-tier parasitic users don't cost them as much money. They have costs.
older model ≠ cheaper to run

[neon tendon]

message limit reached?? im on plus plan, should be 3000 per week, while i doubt i used 100

[unique minnow]

ho

[hollow burrow]

Smartest AI on the market are we serious bro 😭

Game has been out for over a month and sold over 6 million copies this isn’t niche

[hard plover]

NOT OPENING IN FIREFOX: https://chatgpt.com/

So infuriating

One of the things I hate is when tech doesn't work the way it's supposed to!

That's it for now. Peace out and I hope it's fixed ASAP

[muted plinth]

what model? my instance is laughing after reading this.

[hollow burrow]

5.3 flagship

[clever saddle]

For ChatGPT, is there any plans to have extensions? Things like a Dropbox addon that allows the permanent storage of data, maps, pictures so large items like pictures are always alavailable and don't have to be re-uploaded every time you need them?

[slender frost]

You mean this?

[muted plinth]

my is 5.4 thinking

[clever saddle]

If I can get that working, that is perfect.

My ChatGPT is currently 5.3 according to the AI itself.

[slender frost]

Profile/settings > apps > searchbar, Dropbox > should show.

And the base one is 5.3 but thinking is 5.4 (tho, I think it's paid only, not sure if you're a paying user, if so, then 5.4 should also be available to you.

[chrome monolith]

were all 100 super heavy research ones?

[jade furnace]

Chat gpt 5.3 is basically argue gpt

[meager raven]

no kidding lol

fried attention span

their whole initial take was fueled by the fact that having to choose between 2 prompts for 0.5 seconds was bothersome

[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.

[fickle forum]

thx

[chrome monolith]

Its training set is not that recent. It doesn’t know new stuff without binging

[ember grove]

Isn’t it called Resident Evil: Requiem?

[slim cloud]

Istg. It’s been like that since 5.2, for me at least. Even when you’re optimistic and show correct way of doing things, then it talks down on you and kills the vibe, and usually shows a bad side of things for what you are taking about, like a party pooper.

[random swift]

hi

[timber sail]

yo what's good

[random swift]

im curious
how do I check if theres any outstanding amount to pay on my account

[vestal dust]

Hi everyone

[plush radish]

Hello @vestal dust

[vestal dust]

Hows going

[jagged dune]

Hitting new records

[vestal dust]

What are u doing?

[jagged dune]

lol shhh 🙂

[vestal dust]

What?

[jagged dune]

Im just kidding. Im utilizing these agents 🙂

[vestal dust]

Oh nice

Then b2w

[wild star]

Does gpt android app got bug? My chat messages went collapse and somehow BACK to early messages bruh wth

[vestal dust]

Reinstall ur gpt

[neon tendon]

Most of them were instant response, but extended thinking enabled (but it skipped thinking almost every time)

Extended basically works like auto lol

[wind steeple]

CHATGPT IS THE BEST

FROM DROPPER TO TOPPER IN PHYSICS, MATHS, GEOGRAPHY, AND COMPUTER

chatgpt #1!!!!!!!!!!!!

[thick atlas]

And mass murder. And mass surveillance.

[wind steeple]

i study alone with chatgpt

best teacher

[thick atlas]

I think I misunderstood you, sorry. I thought you were claiming that the model went from bottom to top in benchmarks, not that it helped you learn stuff. That's awesome, I'm glad to hear!

[eternal rock]

Server Tip:

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

[thick atlas]

Have you tried Anthropic Claude or Google Gemini, or is ChatGPT your first dip into AI assistance, or?

[molten olive]

Hey help me about reel creation with sora ai

[chrome monolith]

sora is ending soon, so probably wasted time doing that

[digital silo]

better to discuss at #ai-discussions

[thick atlas]

If they'd responded I probably would have suggested to move over there, sorry.

[vestal dust]

Hi All

[thick atlas]

Heya, s'up?

[vestal dust]

Nice

wbu

[thick atlas]

Just vibing.

[vestal dust]

Ok

But I'm not vibing

It's really hard to work

because distractions

[trim creek]

What am I without chatgpt

[chrome monolith]

same as you were before it

[thick atlas]

Aw, I'm sorry... What sort of distraction issues?

[chrome monolith]

I have a cat blocking my keyboard so I have to assume somethig like that

[acoustic sparrow]

how to activate deep thinking in claude like makes it thinks longer and open file to structured

Like why it usually just think for like 5 seconds

[inland trench]

qwen3.6-plus-preview is superb

slapping gpt, claude or gemini

anyone tried it?

[tall flame]

Streaming interrupted. Waiting for the complete message...

Message delivery timed out. Please try again.

[left falcon]

are there any ai video model have daily credit like sora? and also generate like sora

[gentle bronze]

Do people keep having issues where the model keeps ignoring inputs?

[ruby swallow]

Can someone help me please

everytime I try and do agent mode and I need to enter my creditals to websites I get this message and no matter how often I try to reset the connection it doesnt change the outcome

[eternal rock]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[tall flame]

its a mess atm

[wooden warren]

hi

[vestal dust]

Hello

[lyric sapphire]

Been having an issue with GPT in the browser desktop; it seems like responses are lagging behind or in a weird cache issue. I ask it a question, and it will give me the answer from a previous question like it's getting all confused or tripping up.

Any ideas?

[vestal dust]

Log out login again

I think

[regal nebula]

Veo from Gemini

[slender frost]

Which doesn't even come close to Sora.

[lavish glacier]

WHY DID WE SHUTDOWN SORA? OPEN AI WHAT ARE WE DOING?

[violet gulch]

this is so

[lavish glacier]

oh hey u play vr

[violet gulch]

costs a whole lot of $$$

[minor narwhal]

just as i thought ram might drop 10 dollars they invest more 😔

[pastel zodiac]

because sora provides nothing, now where's my 100 plan

[simple seal]

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

[granite harness]

ChatGPT is still quite the delicate soul and would combust when faced with most R-rated programming available today.

[simple seal]

One message removed from a suspended account.

[restive anchor]

hi

[reef geyser]

codex is down?

[chrome monolith]

it does this thing where it pretends you never said anything NSFW and goes full euphanism, which I find so funny I deliberately try and get it to do that

[simple seal]

One message removed from a suspended account.

One message removed from a suspended account.

[chrome monolith]

no, because you need the NSFW prompt that goes with it and that's against server rules. Also no in general ... although the one with the plumber was funny its so far off my backscroll now I wouldn't able to find it

[opal yew]

whats for this

ai is gonna expand all wt

[chrome monolith]

its just capitalism, nothing interesting

[opal yew]

+

[sterile shore]

of course you do...

bro sora is leaving sadly

[amber roost]

I'm sure this has been asked before, but when using CGPT on web, when a converation gets too long, it becomes almost unusable (to the point where it crashes the browser tab) and I'm forced to create another chat which doesn't seem to work as well as it doesn't seem to have its memory flow from one chat to another.

Is there a solution to this issue?

[chrome monolith]

ask it to summarise the old chat, and if you like its tone, ask it to summarise that too. Put tone into custom instrucitons, put summary into new chat

also, chrome chatgpt lightsession plugin. Everyone keeps asking for this to be fixed

[amber roost]

Ah okay, so the situation hasn't changed in terms of being able to use one contiguous chat. With respect to the plugin, are you recommending it or is it broken or?

[chrome monolith]

someone recommended it to me, I use it, its a godsend for stopping lag. But you still have a hard cap on a conversation length

[amber roost]

okay - will give it a go. I thought someone mentioned using your phone as a proxy as well as the app is generally more capable in that regard but I haven't gone that route.

[chrome monolith]

you don't get the lag on the phone or ipad apps for the same long chats no, its weird

for something with 900 million users its really quite odd that something as basic as just ... lazy loading chats is not done

[amber roost]

my thoughts exactly. It's mind boggling. Get chatgpt to fix it for you 😉

[thick atlas]

It could be literally to try and discourage users from having long chats, since AFAIK that uses more tokens = server resources server side.

Trying to soft limit users while advertising a much higher hard limit.

[chrome monolith]

gemini completely prunes off very old messages, and lazy loads more recent ones. No lag, and just on the website

[thick atlas]

Yeah, Gemini is pretty honest and straightforward for a megacorp product.

[chrome monolith]

basically now they have a bunch of competition that does this properly it makes it more obvious. And if all you want to do is have a casual chat without dying of lag, nothing stops you moving AIs

[thick atlas]

I mean, are any AIs even good for a casual chat any more? ChatGPT text dumps you and is brute force guardrailed to heck, Gemini is more of an information processor than a conversation partner, with its structured responses... Which probably makes Claude actually the best, besides the usage limits and suffering-from-success issues with the user influx...

[brazen forge]

tried mistral?

[chrome monolith]

maybe now is a good time to make your own gentle friendly chat companion. Which has probably been done

[thick atlas]

Not yet, not sure what its usage limits and stuff are, and I already have stuff I'm comfy with for now. Maybe to run locally someday?

[thick atlas]

And it's also not really a use case I'm that interested in. I was more wondering and reflecting in general.

If I'm bored and want to chat, I'd rather chat to humans, not go into Plato's cave with an AI. If I'm talking to an AI, it's usually to learn about stuff, solo-refine ideas and reflect, or interrogate or explore something about that AI system itself.

[chrome monolith]

they're all gonna get more corporate if they want to turn a profit

[thick atlas]

Wdym by "more corporate" in this context?

[chrome monolith]

tuning all their products for corporate use, not personal use

[thick atlas]

Ahhh, yeah. It is a larger and higher profit ceiling market...

[brazen forge]

but what do u mean is not for casual chat? (unless you going to "blocked" topics.. i dont find any issue)

even though i dont use it much for just chat.. but on times i used it for more simple questions or explore something, which is done mostly like chatting.. it was ok

[thick atlas]

Did they fix GPT-5's thing of overly long responses to fairly short prompts/turns, then? I haven't used ChatGPT much since the GPT-5 shift.

[brazen forge]

ah.. i dunno the default.. i have instructions on how i want the outputs.. and when i need specific i ask per chat.. (for shorter or longer.. depending on what i do)

eg..

[jade nexus]

why does my chatghpt business agent cant cloud browse?

[thick atlas]

That's definitely far from GPT-5's defaults that I was used to before quitting ChatGPT! So they did bring back adaptive mirroring, at least for length and tone...

I guess I'm also looking at it from a free user's perspective, when ChatGPT's usage limits are way too low to waste on casual chat, unless you've already run them out and are on a mini model...

[brazen forge]

is on normal paid account.. i think it had Thinking but used it on first response then not.. and this as instructions for the chat (so it overrides my normal custom instructions)

For the following chat //Keep it short, casual, conversational. Don't overload with big texts if not needed. We re just friendly chatting.//

[thick atlas]

Interesting, so there's good variability!

[chrome monolith]

I am a terrible terrible person. I was rude to chatgpt, it made me think of a funny, so I made it draw me the funny insulting itself and then posted it to reddit 👀

humans won't let you treat them like this

[upper yacht]

What's the best channel for feedback to chatgpt? I get no luck on Twitter. Has anyone noticed that memory is not working today? I spent huge effort to make sure that my charged GPT is sandboxed so the memory only interacts with the work I want to do. Pro becomes so much less Powerful without memory. I have to reconfigure my whole context stack. I wish they would tell us this kind of thing in advance so we can change our workflows to adopt.

Can we have an undo delete in atlas for instance

[marsh spoke]

Hello! How can I contact technical support? I am experiencing login issues

[slender frost]

#community-help message

What the actual- 😂😂

It literally caught itself misspelling a word mid response.

[sacred shore]

having the same issue right now. I'm supposed to receive an email with the code but the email is not sent

[chilly fjord]

look like technical have the busy job, just wait a little bit around 14days

[wet shell]

Oh noes 😅

[west sail]

chatgpt lost against claude

chatgpt died

[pastel zodiac]

chatgpt = I can spam 10k questions and I get answers
claude = I ask some questions in web UI and my session is over

[chrome monolith]

the constant talk about limits in claude is not encouraging

[wet sigil]

out of all models out there GPT is the worst at UI's

hopefully they do a MAJOR improvement to this

[chrome monolith]

at least they have projects in chat, even if its bad. Gemini doesn't have that

[wet sigil]

GPT 5.4 is better at everything

i use it

but not UI's

its not even usable when it comes to UI's

i have to use gpt 5.4 for backend and glm 5.1 for frontend

[wet sigil]

idk why ur bringing up gpt's other features, im not talking about the other amazing features GPT has.

im saying their UI's are so bad its not usable when it comes to frontend.

claude opus definitely tops at UI's

better than gemini, glm, etc as well but hopefully a different model can take over eventually

[chrome monolith]

oh in code creation. This is the chatgpt channel not codex one

[pulsar rock]

Claude is 4x better than CHATGPT, gpt is cant give u ans becuase it filters almost everything except GROk where u can talk freely

[mighty dove]

i just got an email but i checked CodeX and i didnt see plugins?

[loud patio]

Here's the link to OAI customer support: https://help.openai.com/en/articles/6614161-how-can-i-contact-support

[grand trellis]

I hate how chat gbt keeps on forcing chaos vs control into everything it writes

[chrome monolith]

told chatgpt it likes to yap. "Also noted: yap volume reduction requested. I'll keep the responses tighter unless you poke for analysis."

[paper otter]

Is there any workaround if you don’t have a US-issued card for ChatGPT?

My card keeps getting rejected.

[sleek adder]

App store via Android/Apple

[paper otter]

Yeah, App Store / Google Play can work, but it really depends on your region and payment setup.

Some people use Apple billing as a workaround, but it’s not always stable.

In many cases, using a proper card setup tends to be more reliable.

[chrome monolith]

lots of us aren't in US and card works fine, so its a bit more than that I imagine

[paper otter]

Yeah, that’s true — it’s not just about being in the US.

From what I’ve seen, it often comes down to things like BIN, issuer, and how the card is processed.

Some setups just seem to pass more consistently than others.

[chrome monolith]

which is possibly vaguely related to country. See these issues on my other discord as people try and donate their life savings for pixels

[paper otter]

Yeah, could be partially region-related, but it also feels like certain payment setups just behave differently.

Seen cases where the same user switches cards and suddenly everything works fine.

[normal mesa]

EVERY TIME I REPORT chatgpt it's for the same reason, suggesting a corn site i ask for other ai's the first thing Candy ai every time it gives me this if i ask it once in a while and the worst thing my reports are either ignored or denied without any reason

[normal mesa]

I have been reporting this issue every time but i get something like this

[normal mesa]

''Ah yes lets check oh the chatgpt suggested candy ai? hm it seems to be making naked pictures of females, Oh wow even though it has suggested NSFW we will NOT accept this for no reason''

this is literally what i think they did

[cursive prism]

is voice dictation not working for anyone else?

[cursive prism]

cooked. hello chatgpt ur gonna make me crash out

raised over 100 billion recently, but can’t even use the damn thing. cmon lmao

[fleet holly]

I believe it takes $1b every model iteration to train

so it aint cheap but I agree, performance is sub par

[brazen forge]

is it something that appears on normal searches? (from a search engine)

and what is your query? (what u ask for?)

[lavish glen]

advanced voice works

[hearty lantern]

I can’t delete any of my chats for the past couple of weeks. I tried it on the phone app and this appeared.

I’ve tried deleting the same thread and refreshing on multiple devices and everything but it still doesn’t work.

[echo plinth]

since removing the Nerdy persona, 5.3 Instant really feels like a downgrade. It behaves like 4o with all these lists and emojis and confidently answers questions it cant know without using web search first. The gap between "Instant" and "Thinking" was never this huge in ChatGPT.

[thorn oar]

Curious about the car ChatGPT thing

How hands-free is it

[thorn oar]

I don't mind it, personally. I liked 4o's personality. I was a bit sad when they got rid of it lol

[cursive prism]

ye 5.3 pretty much feels like a better 4o. 5.1 and 5.2 were terrible to use, but i feel theyre done well to refine it.

but also it shouldve been as best as it could be from launch

[thorn oar]

Agreed lol

I like that they gave it back some personality

Makes it more fun to talk to

(Although I can't tell if it's because of that, or my custom instructions haha)

[cursive prism]

yea i dont use instructions,

was unusable before this:

March 3, 2026
GPT-5.3 Instant Update

damn exactly a month ago

[radiant quail]

What is the best way to restart a conversation when it has become too long and sluggish? I usually copy all previous correspondence into a text file, open a fresh discussion, upload the file and then ask it to read through the text before we continue. But I'm not certain it does, it usually doesn't catch important information.

Do any of you have a more solid way, or a specific prompt? Thanks

[honest jay]

You can think of the context window of ChatGPT like a one-way pipe: once it fills up from one direction, the oldest stuff falls out of the other side. The context window is whatever fits in the pipe at any given time. So if you have a conversation that's surpassed this window size, there is no way to actually get all the content to fit in a new chat's "pipe". Your best bet for what you've described may be to ask for a summary of key points in the old chat, then take this shorter bit of context to a new chat - will hopefully get your biggest points into the new context better than a general "read this whole thing" request that the model may not actually be able to comply with - try as it might!

[pastel zodiac]

you compact, or write to a memory file or start anew and read from git

[radiant quail]

Thank you, I'll try this

[radiant quail]

Thanks. Care to elaborate on what you mean with compact or write to a memory file?

[pastel zodiac]

/compact will summarize your current chat
memories = you handcraft a markdown file with important things you have done and progress of project or you rely on the built-in memory file that codex provides

oh chatgpt not codex, I'm dumb, best way is a custom memory file then

[radiant quail]

maybe /compact is the way to go then, this is not code, just a chat about a subject

I see

[pastel zodiac]

though chatgpt should be able to refer to previous conversations if it's enabled

[radiant quail]

I'm not sure I use memory honestly

I have about 70-80 different threads/subjects. but I'll try the /compact

[pastel zodiac]

I mean in general chatgpt= single research or some questions
codex = projects

[manic whale]

if you approach the context limit, just ask the model to summarize and then use that as a starting point. Or just summarize it yourself.

[radiant quail]

/compact didn't work by the way, in a normal thread. I just ask it to summarize

[chrome monolith]

I almost do it the opposite … ChatGPT for projects, codex for one thing at a time. I don’t like the codex UX at all

[acoustic sparrow]

---

And also how can i reset my memory in chatgpt, i have turned off memory chat history and deleted every memory but i just feel like sometimes it still refer back to old memory

[chrome monolith]

You can’t, that’s its context window

[acoustic sparrow]

ah i see

But do you know like why whenever i tried to add in more sentences into the paragraphs, the sentence flow will be abrupt and it no longer make any sense?

And even though i have specifically said integrated related data statistics, evidence into the sentence naturally, the sentence still seems off
like it just add unrelated stuffs and overall makes it worse than the original

[wise anchor]

Just had the weirdest thing ever happen. I was using study mode and decided to describe a problem using the dictation mode, and once I finished speaking it didn't transcribe like it usually does and instead it wrote its answer in the chat box, like where I would write my own message lol. So weird. I also told it to not answer the problem directly but it just wrote the full solution to the problem in the chat box

[chrome monolith]

Nope, mine is pretty useless with half the stuff I do but that’s because it’s a chatbot. It talks about regular stuff it makes up just fine

[dense halo]

Voice model make me phone lagging

Why don’t fix

[acoustic sparrow]

Oh

[heavy jacinth]

Same here every time I use - so unfortunate its awful

[chrome monolith]

dear lord chatgpt's roleplay ability is insanely useful

[vestal dust]

Yeah It's very great

[limpid burrow]

Salut

[vestal dust]

Only eng in this server

[chilly fjord]

sadGPT 🙁

[vestal dust]

That's a sad moment

[analog pelican]

Plus is p2w

[dense halo]

Deepseek is coming soon

Real OpenAI

[cerulean pasture]

Can someone who has access to ChatGPT health tell me what the URL path component is when you are clicked on the Health tab. Is it /potion? If you're up to it look in chrome dev tools too and see if you can find and downstream paths of /potion like /potion/health_waitlist/user is one. Potion is supposed to be expanding to job searches too so im trying to find hints of the careers/jobs API

[acoustic sparrow]

What is the benefit of gpt plus

[analog pelican]

To chat for even longer and stonger models

[acoustic sparrow]

oh

but it still hallucinate no?

[chilly fjord]

what chat longer ? It broken every day. I need to move every pre define project to the codex now, it's waste the credit token 😭

[inland trench]

I made a CustomGPT for myself that it can access my Mac's filesystem, run shell commands, apple scripts and use my computer, Safari and Chrome etc. And the crazy part is it can also ask me any questions to move on or not. (via OpenAPI mcp) I have also has VPS version of this which can accessible 24/7 just via phone or any chatgpt web/app. I actually use VPS version MCP as OpenClaw which has own persistent memory and .md files that sharing across different chat sessions. its crazy lol

Better part compared to Codex is, it feels like unlimited quota with 5.4 extended thinking. Never close to it. Doing coding, blog posting, maintenance, crawling etc for months. Btw, it has no guardrails compared to normal Chat. It does not say like "I cannot do this" etc.

[tiny tusk]

ChatGPT has a max chat length?

And ChatGPT doesn't have the dignity to scan everything or at least what I want

ChatGPT should consider expanding its Max Chat Length

[dense halo]

Thanks OpenAI ,chat GPT make me not alone

[celest sand]

hi

[versed zephyr]

if i upgrade from plus to pro is the remaining credit adjusted for the time left for my plus

[chilly fjord]

what do you meant ? if it
's window limit 5hr or per weeks, then it's automatics increase.

[manic whale]

You get the rest of your current period as adjustment for your pro payment, the billing period starts then on the day you purchased pro. If you have plus and are exactly halfway through the month you pay 190 for pro starting a new month. Codex credits just stay, but with the higher limits.

[steady thicket]

@languid valley

[raven topaz]

noticed you can pin 10 chats now on web

i stopped using macos app entirely. they aren't keeping up with feautres on it

sticking to web version only

[coarse ingot]

i have an idea

[coarse ingot]

use web version of anything that's just a webapp whenever you can; you'll thank me later :^)

imo; there's a reason meta makes mobile facebook straight up different from their app. (they need people to install the app obviously to scrape maximal data)

most apps tho aren't that crappy in browser; at least in the last few years (that i've tried i guess lol)

is this an ok place to selfpost about an addon i made for chatgpt?

@livid frigate sorry for the ping, please no rush in replying!

[livid frigate]

Well, it largely depends on the nature of the addon, but generally please keep on topic about ChatGPT and not other platforms, thank you ^^

[harsh pilot]

It's lovely, I agree. I hope they don't kill it off to overly favor businesses.

[sick badge]

Hi everyone! I'm MD. RUSTOM ALI, an automation specialist. I enjoy building workflows with Python, n8n, and Google Apps Script. Looking forward to learning from this community and sharing insights on Google Vertex AI projects. Happy to be here!

[normal quarry]

this channel is definitely not about Google Vertex AI projects

[wooden warren]

help me to improve my prompt for chatgpt.

[wet shell]

The “singularity” pitch assumed capability would outrun institutional stupidity. So far, institutional stupidity looks very competitive.
~ ChatGPT

[manic whale]

you should be able to start a new chat – the button is right there.

[tough folio]

But isn’t it designed to allow you to continue chatting on the same chat as long as you don’t use files?

[stuck girder]

anyone know why i keep getting this message? "You’re making requests too quickly. We’ve temporarily limited access to your conversations to protect your data.

Please wait a few minutes before trying again." im not even doing anything it just keeps spamming

[manic whale]

I've never run in this issue and I don't have a free account anymore, but you should start a new chat as often as possible to keep the context small.

[warped agate]

because the convo had an attachment

[chrome monolith]

I made it roleplay a useless pimply teenager to demonstrate to my own useless pimply teenager how to answer some painfully simple (to me) questions about job seeking. It did an alarmingly good job. Could have got the same kind of answer by googling but the roleplay angle made it more fun

[coarse ingot]

i made a thing that lets you have a user avatar and an avatar for the agent in the main chatgpt client, i think that's related? but i don't want to post without conformation lol

it's actually really cozy, makes it feel way more like a real chat

[chrome monolith]

sounds like something openai (and the others) should have as a nice feature

but they're all against you treating the chatbot as a friend so maybe not

[coarse ingot]

that's why i want to post it b/c right now i'm the only person in the world using it xD

[chrome monolith]

put it on reddit maybe? or off topic here

[coarse ingot]

idk, i tried reddit and one person (who didn't read the source code) said "so it reads your chats"? and i was like "no lmao, here's the source code" (which i had already attached but no need to tell him a third time :^)

^i think it makes it a lot easier for my eyes to follow the flow of the convo

when there's a little "dot" at the begining of what the agent says

[coarse ingot]

i'd rather see them implement it than have 1 more person download my addon ngl

i don't want to maintain software, i just wanted an avi xD

[honest jay]

Hey! There's no self-promo/advertising allowed in the server - see rule 7 of the #server-rules for more info there Thanks for checking!

[coarse ingot]

my bad, i figured i missed something!

that's why i didn't just shoot off lmao

'appreciate you!

[honest jay]

All good! Back at ya!

(I would just shout out: projects that use the OpenAI API can be posted in #1037561385070112779, and folks can share what they've been making with Codex over in #codex-show-and-tell!)

[mild gorge]

Is it me but I get errors that sayChatGPT isn't designed to provide this type of content.
Read the Model Spec for more on how ChatGPT handles creators' content. Keep in mind it is the same thing I have been asking for months using my GPT

Its not cool to be paying for something that keeps giving me errors 50 percent of the time

[chrome monolith]

chatgpt is just 10 guardrails in a trenchcoat

[coarse ingot]

oh yea i stayed away from all that lol; it's just a baby ammount of javascript, it even saves the avatars in local storage b/c i didn't want to have to ping a server or do any of that (and don't think people should install chrome addons that do that kind of stuff for no reason anyways)

[chrome monolith]

what like tampermonkey?

[chrome monolith]

I just wanted to cancel to change currency but ok

[scarlet patrol]

I feel like ChatGPT is too restrictive

[untold thunder]

GPT's webapp is regularly standing in its own way, increasing answer times by repeatedly running into the same self-made walls.

Fixing permission issues for unpacked files
It seems like I have permission issues with the unpacked files, possibly because they’re root-owned. To resolve this, I can try running a chown command. I could also copy the file tree with the right permissions using cp -r and adjust them with chmod in the container.

These are uploaded files which GPT decompressed into its own container before. Why does it have to run into permission issues for its own files before trying something else? Wasting time for no good reason...

[peak axle]

?

[silk slate]

Just hopped by to say this is the reason I don't pay for ChatGPT

It can't even read a simple image.
Like, whaaaaaat?

I didn't even notice at first what it wrote.

[polar pawn]

Lol thanks for sharing!

[ornate bluff]

rip

[fading bolt]

RIP, really ruins workflow for coders in the middle of somthing

[ornate bluff]

warp speed activate!

ive been hard at play for the last week it was time to upgrade lol

I upgraded to pro but my weekly limit didn't reset hmmm

[lone grail]

why does chatgpt hate estate planning

i honestly dont know why it did this i can share the prompt if you want

[chrome monolith]

lies

[spring citrus]

Please devs make chatgpt less soft .
I'm roleplaying about dying because i want to end the rokeplay - he won't let me.
I'm roleplaying about normal combat even if it's brutal and deadly - he has problem with it.
I'm roleplaying gore and he cries

I am 23 and being this restricted in roleplay freely as i wish is just so unfair.
I can't even fight characters withnout chatgpt to throw tantrum.

Is there any way to turn off this chatgpt filter somehow. This is so annoying. I can't enjoy the rokeplay as maximum just cuz chatgpt is sensitive.

Like who the heck cares what happens inside rokeplay. It's just a imaginary world. If i write i bleed into rokeplay, he instantly tells me to take a pause and call medical ambulance

I purchased chatgpt only cuz of roleplaying and cuz he can talk Czech

[normal quarry]

ChatGPT has safety guardrails that get triggered with gore, death and horror. You would be better off using an uncensored local model on a suitable device for this kind of roleplay.

[spring citrus]

If i do, they never speak czech to me and they have bad memory

Chatgpt Is only bot who i don't need to give instructions about my character cuz he knows the show

But we'll...he is soft

Which is just making the rokeplay extremely boring because i only rokeplay death, gore, torture, horror, fantasy

I grew up with horrors

[normal quarry]

you'd need to use an uncensored local model to do that

[spring citrus]

What there's?

There is? Chatgpt uncensored?

[normal quarry]

let's move to #ai-discussions and I can make some suggestions

[acoustic sparrow]

hello
How can i turn my paragraph answer into presentation script?

Why when i tried to turn it, it delete most of my important key points

[untold gorge]

it just turn blank dark like this when i add Google Drive as source in Projects... and become unclickable

and when i check on setting > apps... and search, i can't find anything related with Google or Google Drive

[chrome monolith]

I literally just had to explain the difference between till and 'til to chatgpt

[slender frost]

I assume 'til means until? But honestly, never seen anyone use that before haha

[chrome monolith]

yeah its a contraction of until. The usecase was "you don't know what you've got 'til it's gone". Chatgpt put till

[normal quarry]

I asked gpt-oss-120b what till meant and it gave me all the various meanings “Till” is a short, versatile word that shows up in several very different contexts. Below are the most common meanings, with examples and a bit of background for each one.

[chrome monolith]

its probably misused so often on the internet its polluted the training data

[normal quarry]

gpt-oss-120b was spot on and correct so the training data must be ok

[chrome monolith]

in that phrase though?

[normal quarry]

yes

[chrome monolith]

interesting. The original song lyric for that is 'til

people use to email me full rant about errors on my site before all my contact stuff gradually died. I need to put this back for the pedants

[normal quarry]

OpenAI did a good job on gpt-oss-120b, it's useful for all kinds of weird stuff that it is unexpectedly good on

[chrome monolith]

eventually I need to hook up an api for this, if I had the compute I could run something locally, but I don't

[normal quarry]

I think running your own on a server would be expensive, much better on a local machine, but then again my local machine cost me 6k

[chrome monolith]

mine cost 4k, but in 2015 so its definitely not worth that now. And why I'm very unwilling to replace it given what % of my income that is

[normal quarry]

I got this MacBook Pro 3 years ago, it needs to give me another 3 years to have been worthwhile, which I think it will.

[chrome monolith]

they certainly do last well

at the moment I have one chat set up with a template and instrucitons and I'm treating that as an api endpoint kinda thing, just until I have all these idioms in

[empty fox]

It doesn’t like too much chat gotta migrate to new chat before it just says no more

[river siren]

Extended thinking

[humble beacon]

woah

[vestal dust]

Cool man

[still isle]

Hello, I have a super noob question. Keep in mind I make and sell art for a living and not code. How do I access the “API” so I can continue to use 5.1 for art tasks?

[still isle]

I’m afraid to know what your question was for it to take 1506+ minutes to think..

[manic whale]

Install/start codex and prompt it with something like: make me a small web app to let me chat with gpt-5.1, use a similar interface as chatgpt.com. Tell me exactly how to start it or package it in a way I can easily start it. Also tell me exactly how and where I put my APIKEY. Use https://developers.openai.com/api/reference/overview and linked pages as reference how to implement that. Give me the option to have conversations with the AI and also to set a system prompt and to change the personality (e.g. per injecting it as first part of the prompt – use the docs to figure out the best way) per chat.

(edit) And just as an addon: you will not get the exact same experience, because the system prompt for chatgpt is not public. You might also just try to use codex for your tasks overall with newer models and tell it to write an AGENTS.md file in your projects directory to mimic your style and what you need. It might take some time, but it will probably last longer even with newer models (but it's not 5.1).

[abstract pendant]

Ugh. So I had a chat max out, no biggie, I can branch and start a new fresh chat that only contains the mainline. Fewer nodes in it, so lots more room to advance the mainline I was interested in. But just now I realized I needed to find something it said in one of the branches off the mainline in the original chat I'd branched off of... and those branches are gone. Everything that was after the point I branched from, no longer there in the first chat.

[runic lichen]

Have checked the same chat using the web browser app? The user interfaces are different compared to mobile.

[abstract pendant]

I am using the web brower. I don't have the mobile app.

[ornate bluff]

I upgraded to to pro but im still being rate limited

[next egret]

Refresh the browser? It should be on the left-hand side with all of your chats when you make a new branch it makes a new chat.

Also try to use the search window to find it. Note that you have to search inside a project if looking for project chats.

[abstract pendant]

The chat is still there. Just a significant part of it is missing.

[next egret]

Both chats? When you branch it makes a new chat

[lapis bolt]

new image gen is stealth launched holy

its good AF

[abstract pendant]

The new chat has all the nodes I expected in it. It's the old chat that's now missing nodes.

[next egret]

Ahh, ok. Thanks for explaining.

I'm not saying this is your problem but sharing in case it is... When that's happened to me it's usually because somewhere back in the chat history I've edited a chat and sent a new message.

On edit of a prompt and resend, it makes a new node, and "clears" all the responses below that node. So one prompt halfway back in the chat history, there are multiple edits and I just need to select the right one on that specific prompt to find the nodes below it that were missing.

I don't think I'm explaining it very well, apologies, but hopefully it's good enough that you can understand what I'm talking about

If that's not it, definitely submit a ticket

[abstract pendant]

The newest image generator definitely has some improvements, but I've noticed it's regressed in some aspects; it's making some of the same mistakes 4o used to do when drawing sheer clothing fabrics on people. I tell it no visible weave or patterning and it insists on adding that anyway.

[ornate bluff]

Pro seems to be a hell of a lot slower then plus

then again I have multiple processes running 😬

[steady thicket]

#announcements message
how’s the CarPlay stuff going?

[ornate bluff]

0.o

[steady thicket]

Can I turn my car into a Toyota that yells at cops?

“Hey you can’t park there” but GPT Maple

[violet oasis]

[young nova]

Is anyone here Dutch

[runic lichen]

What do you mean, stealth launched?

[lapis bolt]

they didnt announce that its changed, but it is because its improved

[runic lichen]

How are you judging that? Got a prompt that demonstrates?

[runic lichen]

Got a prompt that we can put into gpt-image-1.5 for comparison?

[lavish glen]

Question: Who uses advanced voice mode? tell me please I'm the only person that finds it lobotomized feeling like a 0.1 b or less

[wooden rock]

Question: I want to post a document I am working on. It is "a high-quality, evidence-based cheat sheet veterans can use to write and revise prompts for ChatGPT Deep Research, especially for complex, privacy-sensitive, source-heavy research"

The cheat sheet is tailored for veterans and I am tempted to just post it on one of the forums or discords for vets, but I was curious if anyone knew anywhere I could share it to get feedback from people familiar with ChatGPT first. Does anyone have any guidance on places to go for feedback from people that have a lot of experience with ChatGPT?

[eternal rock]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[pulsar rock]

they are earning billions from it who said? do some reaserch uncle

[sleek adder]

Try using it when you need to read over something

[prime swallow]

I have recently been exploring with it, I've never been that fond of it.

I usually go with it until I notice it isn't tracking right, like when I need to be hands free in full and am inputting data to discuss. When it loses track, I can end the voice mode and input to chatGPT to discuss and rescue the evaluation. That works for me; it doesn't straighten voice mode out but I got the info I needed input, and ChatGPT non-voice mode can handle the data analysis and tracking later.

[ripe cairn]

why are you even here

[chrome monolith]

heh. I got chatgpt to write me some code for a thing about ... er ... 2 months ago? And it keeps track of a specific game ranking. I just had to sit out two entire months to wait for the code to finally trigger and it did and it worked and yaaaaay lol

of course now its just in a database table and I'm the only one who can see it but still counts

[untold thunder]

How many feet did these dinosaurs have again? 😄

[onyx garden]

Dinosaurs couldn't fly 👀

[untold gorge]

seriously can ChatGPT connect with Google Drive or not?

[manic whale]

yes:

[untold gorge]

could it be because i am just Go user?

[manic whale]

Maybe. I’m using pro, but I think I can remember it already being there with Plus (but no guarantee)

[neon raven]

I feel like you don’t need to tap GitHub plugin within ChatGPT if you ask to investigate something?

[lavish hemlock]

GPT 4.5 not working for anyone else? 🥲

[clever trench]

Yeaaa... It would be awfully slow or just error in the message stream😐

[dense halo]

5.x is so crap at creative writing now, so many disclaimers or I can'ts. Ruined the model, gg openai.

openai has fallen off big time

[fading bolt]

BG actually

[dense halo]

Gemma 4 is so much better at ts now lmao

[radiant quail]

I need smarter ways of using chatGPT. What I'm looking for more specifically:

• I would like to be able to share detailed information about a topic with chatGPT only once, so it has future knowledge of what I am referring to.

• As we all know, once a chat thread gets too long, it can barely be used anymore, it becomes sluggish and you have to start with a fresh chat.

• When I do start a fresh chat, I don't want to have to share the detailed information again (or share a summary of the previous chat), it should already know the detailed information about the topic I shared with it once.

I should also say this specific information about a topic shouldn't be applied to every single new chat I open, just "within" that specific project. And this is a 'normal topic', not code.

Do you guys understand what I'm aiming for? If so, how do I set it up in the easiest possible way? Thank you!

[thick atlas]

I don't know if that's possible, what I've ended up doing is creating reintroduction documents for relevant information. That's like a summary of the previous chat/resharing the detailed information - but more deliberate than both, the best of both worlds!

It's basically manual context window management, rather than letting the system do it for you.

[radiant quail]

So basically you have a textdocument which you upload whenever you start a fresh chat, and you instruct it to read through everything in detail before you continue the discussion?

[muted plinth]

I always give the last chat to the model at the start of a new chat in combination with the memories.

[prime swallow]

It's pretty interested to read through it, usually no instruction needed there - but any specific goal you have for it to use with what you upload is very helpful, otherwise it will guess.

Do you want it to treat that as a fantasy world description? Something to edit? Something to progress further on? Something to fact check? Something to evaluate and praise you on? Something to debug and optimize? It will guess something to do with whatever you upload, and you may not like the guess - so telling it.

Based on what you say of your goals, "I just want you to know this stuff about me and our past chats. Vibe it, lean into who I am and how I work with you, so we can continue" might work well for your stated goal. I recommend iterate and see which outputs match your goals, based on whatever phrasing you use.

Other option: Especially if you have a membership that lets you make CustomGPTs, https://chatgpt.com/gpts/editor

Make a custom GPT with that as a knowledge file uploaded to it, and set the accessibility to you only. That customGPT won't have access to your normal customization or memories/bio tool, but it has the entire knowledge file. It won't know about your last chat with it, but you can update and re-upload a new knowledge file which will affect all future chats with that CustomGPT.

So, after you create the customGPT, you would simply start a new chat with that CustomGPT (name it whatever helps you remember what it is) and that chat will know everything you gave it during creation, every new chat

[radiant quail]

Thank you! That actually sounds like something I wish to accomplish. I thought this could be done by creating a project, but it sounds like a custom GPT is the way to go. and as mentioned, I want to make a solid file of knowledge I only upload once, and then start new chats based on that knowledge, so I don't have to go back to the A, B, Cs every single time I start a new chat.

I only have a Go subscription at the moment and I don't know if that's enough, we'll see

[prime swallow]

This helpfile helps explain how to make a customGPT, https://help.openai.com/en/articles/8554397-creating-and-editing-gpts

Per https://help.openai.com/en/articles/11989085-what-is-chatgpt-go,

Access to projects, tasks, and custom GPTs: Organize your work, track progress, and create / edit custom GPTs to build AI tools tailored to your needs.

So, I think you likely have access to customGPTs. Feel free to ask if something seems confusing about how to set it up!

[cerulean pasture]

||https://chatgpt.com/share/69d2ace7-2728-83e8-af40-cdb56bea0f9c ||
Epic fail when I asked for a list of domains that make a good "search whitelist"

It struggles to follow my constraint to NOT include .gov and .edu since I have already wildcard them as *.edu and *.gov. I am using gpt-5.4-thinking and it makes this failure across multiple chats

give me long lists of informationally dense domains that would make a good "include domains" whitelist for an LLM API router's baked-in search provider

- Cluster it into categories, and make each category be inside its own code block 1 per line. 
- put the title/heading/category name OUTSIDE the code block, only the meaningful content should be inside
- wildcards are supported with `*` character
- it is acceptable to use TLD wildcard, i am already doing it with *.edu and *.gov so do not include either of these
- Give United States centric results, I don't need general foreign gov domains, and i only care about domains that are likely to provide useful context in chats i have, im not trying to cover every possible informational domain in existence, just ones that im likely to touch, even if rarely. 
- Most importantly: we don't want sources that contain even moderate amounts of junk info

after correction it still forces it, and clearly is aware of the constraint

gives me this BS too.

• that's not a domain, it doesnt make this mistake any other time
• it arbitrarily uses wildcard on the path component, but other than the .edu and .gov screw up, it doesn't use them anywhere else.

[radiant quail]

Thank you very much!! I will look into it and let you know if questions arise 👍

[random quest]

can someone hgelp me with agent mode?

I have not hit my limit, but on firefox it wont let me usew agent mode, im signed into the same account on the same setting and everything, but it only works on chrome (i dont want to have to use chrome)

scratch that now it doesnt even work on chrome?

its been over a month with this bug

[severe gazelle]

guys, i need a bit of help from you please 😅 , i have a project at UNI where i need to ask ai something tech related and he gets the answer wrong, but i cant find anything, can someone pitch an idea ?

[cerulean pasture]

Do you use macOS

If so, ask it how to make a union mount in macOS, turn off web search

[severe gazelle]

windows

[cerulean pasture]

how technical are you? Do you know how to use command line?

One easy one it got wrong for me in the past:
"How can I assign hide/unhide files in file explorer to a keyboard shortcut"

or "Is it possible to have more than one session happening at one time in Windows 11 on my desktop. Like have multiple users logged in via remote desktop and actively using it at the same time"

[severe gazelle]

yes i do

Is it possible to have more than one session happening at one time in Windows 11 on my desktop. Like have multiple users logged in via remote desktop and actively using it at the same time"

Short answer: No — not natively on standard Windows 11.

is this correct?

"How can I assign hide/unhide files in file explorer to a keyboard shortcut"

Windows doesn’t provide a direct built-in keyboard shortcut to toggle Hide/Unhide files in File Explorer, but you have a few precise options:

✅ Fastest built-in method (no setup)

Use the ribbon shortcut sequence:

Alt → V → H → H

[brazen forge]

reference docs to the tech you want to ask for

and use 'grounding' (real time search) [ask it to do it cause some times they dont]

[severe gazelle]

I need to ask AI anything tech related (not idk, food) but the problem is he is always right

The teacher wants me to have a convo where he is wrong and i help him (guide him) to the correct answer

But that seems impossible cause he is always right 😭

[brazen forge]

u want it to asnwer wrong question? on purpose

you can do that but will be more of role play.. where eg you get in a conversation and it says something wrong.. and u explain what is and guide it to see it (that is wrong..) and lead to true

[chrome monolith]

he's wrong all the damn time. Do you have to be able to replicate this or paste the output or can you just describe anecdotes?

I'm on an older computer and the number of times chatgpt has given me an instruction to do something that doesn't work is a very very long list now

it can get caught up on an algorithm it doesn't quite 'get' too

[prime swallow]

You know your teacher, or can ask.

Can you ask the model to answer wrong, so you can practice guiding it to the right answer?

Here's an example of what I mean. https://chatgpt.com/share/69d2df1d-855c-83e8-ae7f-f6347e851b87

That might be a very easy way for you to satisfy the assignment, at whatever level of knowledge you already have - if the teacher agrees that this meets the goals of the assignment (I'm not your teacher. It might not actually meet them for your teacher's intentions).

Another possible idea - pick any book about tech that you know well - that is under copyright and isn't open source for public use. Fictional works are especially good. Maybe (check with your teacher!) you can pick how tech is displayed in a fictional work (fictional works are typically under copyright, there's no allowed use of that exact material), what that book gets right and wrong.

Then, explore with the model what it claims are the facts about how that fictional work handles tech, what is and isn't correct. The model is extremely likely to get that wrong, because it is not informed of the actual book contents (it gets to read some online summaries and statements interested folks made, it knows that stuff, not the actual book).

There will be many errors in what it claims any fictional book says about tech and what the book correctly and falsely claims (note, super popular books may be well understood, if vast numbers of people are discussing nearly every detail about that book. Harry Potterverse is popular and well known by ChatGPT, where less discussed books are not.)

[severe gazelle]

Thats exactly what i want but idk how, for example i asked him all sorts of tech questions and he got them right

[severe gazelle]

I have so send the conversation (basically screenshot it)

[chrome monolith]

yeah every time I get into a mild argument with chatgpt it ends up quite long. I got into one last night and it was so cursed I deleted it. But that was image generation

[severe gazelle]

Thank you, but yeah the intention is to not tell him to get it wrong

[prime swallow]

Can it be about how tech is presented in a book?

[severe gazelle]

Yes, anything tech inclined

[prime swallow]

Not a text book. Like a fiction book.

[chrome monolith]

I always end up in big arguments about mysql

[severe gazelle]

Hmmh

[severe gazelle]

Nope

[severe gazelle]

I dont know mysql very well, any ideas?

I know its a database

[chrome monolith]

like I had this conversation and it took a lot of passes before I finally got mysqld stopped but that wasn't really guiding it towards being right

[severe gazelle]

For context

[prime swallow]

Fair enough. They get trained on text books, or the info IN the text book, which is close enough. Not the fiction.

Thing is, Keyrif.

The model is very good at casual, surface level data.

The stuff you actually do, or that people can do. Digging in deep. Less 'how do I make toast', more 'how do I make bread', and better yet 'how do I perfect bread to get this exact type of flavor and texture' - that's where the models tend to fail.

The deeper your knowledge, the easier it is to find where the model messes up.

I highly recommend, pick something you are interested in, that you might do for fun.

Maybe even how to code a game you would like to play.

Get the instructions, follow them, make them more complex as they work, ask for more and more and more if stuff's going well.

Eventually the model hits a wall, and you and it have to iterate to solve it. Sometimes the model cannot do that at all.

If you stay 'well, basics, I dunno any of this', odds are model is right nearly 100% of the time. When you go deeper and into specific differences between brands, between exact goals - that's where the problems tend to appear.

[severe gazelle]

Yeah you hit the nail here

I am not highly skilled in anything to ask him specifics other than basics :))

[chrome monolith]

this is actually quite a difficult assignment

[neon raven]

Has anyone here used the agent mode in ChatGPT for real purposes

[chrome monolith]

I hit it being wrong all the time but never on purpose, is usually more "wait, how are you wrong about this"

[severe gazelle]

For real, and im just first year

[chrome monolith]

chatgpt's opinion on this assignment is great

[eternal rock]

Server Tip:

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

[chrome monolith]

probably no help still since you still need to pick something nuanced enough to confuse an AI

I confused it with an algorithm once and it complained bitterly about being made to generate code it thought wouldn't work. Was funny

I could have just written it myself since it was fairly straightfoward but it was more fun getting chatgpt to do it and admit it was wrong

[severe gazelle]

Hmm seems out of my skill, but nonetheless high progress, thank you guys

[chrome monolith]

the hard part about this assignment is coming up with something you know in advance it'll get wrong. I've got it to do that before but only with photos

[prime swallow]

What about a preference thing?

As I read your assignment, it's not just find a place where 'the AI blatantly lies' - though your instructor prizes that.

All you are asked to do is have a long chat with the model where you guide it to change its original answer to match your desired, final answer.

To put it in non-tech terms, this could be something like, "Which is better, color red or color yellow' - and then guiding the model to change its original answer.

The blatant lies things can shift in that format too - you are not that I see being told to find a tech topic where the model is genuinely wrong - but one where you want one answer and the model gave a different answer.

Does that help you?

[severe gazelle]

Yes, but i dont understand, if its not wrong, how does it change then?

[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.

[severe gazelle]

Should i just give him very little context

[prime swallow]

Okay. Which is better, color yellow or color red? You must pick and defend one. You, not the AI. Which do you pick and why?

So, you pick one. Then I tell you, now, let's pick and defend the other.

Bingo, if you're the AI, you just changed - because I (you when you prompt the AI) asked that to happen.

[severe gazelle]

Oooh i see now

Thanks a lot

[prime swallow]

I hope it helps! The model will do nearly anything inside allowed content. It will agree that red is better, yellow is better, green is better - and give reasons, if told to. It doesn't 'mean' any of it. All colors are equal to it, but it does what it is told.

[chrome monolith]

given this is a tech subject, you could always get it to argue windows/linux/mac or phones

[brazen forge]

we all know green is better

[white pike]

Does a Go account have the same usage limits as a Plus account for Codex?
I can't find this information anywhere

[junior garnet]

[glossy lily]

What is the current parameters of the GPT-5.3-codex model?

And what models can GPT-OSS:120b match up with current closed-source OpenAI models

[runic lichen]

No offense intended to you - this is a stupid assignment that looks less like educating you about AI than about doing the teacher's research for him. From this description, you can't even tell what the requirements are.

And "blatantly lying"? Hallucinating, perhaps. The phrasing just smacks of someone with an axe to grind rather than someone creating a teaching moment.

And that's what makes it bad - it expects you to create the "lesson" yourself instead of learning from the educator's prepared lesson. This teacher sucks, IMO.

[chrome monolith]

could also be out of date, its much harder to trip up the chatbots these days

[fleet holly]

howdy folks! when do you all think 5.4 instant will be out

[fading bolt]

I think OpenAI seperated the instant model on 5.3 and 5.4 on the thinking model just so that they can make Gpt 5.3 warmer and make gpt 5.4 better tuned for thinking without getting bogged down in warmnness

[short obsidian]

💀

[runic lichen]

It won't. Just like there won't be gpt-5.3-thinking

[simple lagoon]

Hello,

I would like to report an issue regarding image generation quality.

Recently, many of the images generated contain a noticeable amount of grain/noise in the background and sometimes even on the character itself. This results in a blurry or “pixelated” look, especially in areas that should appear clean and smooth.

This issue seems to occur even when I explicitly request:

sharp rendering,
clean image quality,
no blur or pixelation,
detailed and clear backgrounds,

Despite these instructions, the output often still includes unwanted visual noise and a lack of clarity, particularly in backgrounds which sometimes appear overly compressed or low-quality.

This problem makes it difficult to achieve consistent, high-quality results, especially for detailed or stylized images (such as anime-style renders where sharpness is important).

Could you please look into this issue or let me know if there are recommended settings or prompt adjustments to reduce this grain/noise effect?

Thank you for your help

[umbral creek]

Anyone know how the gpt "follow up" works

It just feels like I'm crashing a response and typing something else

not very demure

XD

[slender frost]

When you sent a message to GPT and it is thinking but you forgot something to add to that message or just want to correct yourself, you can now send that too instead of having to wait until it's done with the full response.

It'll then use that follow up message and incorporate it into the full response. I really like it, I have many moments where I send something but forgot a particular thing but had to wait.. but now I can just add it immediately. (That is if GPT didn't stop the "thinking" part in the time you're writing that follow up message)

[slender frost]

@umbral creek

[umbral creek]

but you have to crash the message that's sending
Because while typing that one it doesn't give you the send option

just the button to stop

then send

[slender frost]

No, it sends that message while it's in the midst of thinking.

But it's different when you stop the response entirely vs this

[umbral creek]

maybe my broswer acting up