chatgpt-discussions

[manic fiber]

thank you Apple

ok but that's great tho now I can add that to the wish list and suggestions for PC

[normal quarry]

just checked, the Windows app has companion chat also

[manic fiber]

Ok smooth

my main problem is when doing complex research and combining several principles formulating how to integrate that into the next step I have to be concerned with not letting the threads get too long

they get so long they often lag when loading

and I've lost almost a year of work in one of them

so being able to spitball and iron out a lot of stuff within the context other chat thread so I can make a very concise and well composed follow up is extremely important

And I'll use desktop and have several Chat GPT tabs open within a project group, and I'll organize my response with Gemini, I feel more comfortable,
using it as a way to verify what I'm doing and cross reference while also parsing output and organizing myself there but if I could do it all within GPT that would be great

[hard dew]

Am I the only one who's noticed past couple days that with GPT-5 ABSOLUTELY no matter what the hell you freaking do or say, absolutely regardless of context, new chat or not, its first reply is ALWAYS some absolute dognuts, as well as censored ofc garbage completely devoid of personality and such, treating as if clearly stated instructions are suggestions? 😀

And the 2nd regenerated response is back to normal, how it used to be? It's seriously becoming intensely irritating.

[manic fiber]

🦌

[south kettle]

Hey folks, need your suggestions and advice.

I’m planning to build an MCP Builder targeted at beginners. The idea is to provide a simple way where someone can just prompt and automatically generate an MCP server and its client for testing.

• My vision is to package this with:
• Best practices for creating MCP setups
• Predefined repo structures
• Dockerfiles and supporting configs
• Agents to guide the process end-to-end

Basically, one place where you can get everything you need to spin up an MCP quickly.

Initially, I’m thinking of targeting cloud hyperscalers (AWS, Azure, GCP) by consuming their documentation through MCP servers.

Do you think this is worth pursuing, or should I reshape the idea? Any feedback or suggestions would be really valuable before I dive deeper.

[loud sierra]

being a consumer this sounds interesting, if you are really able to deliver it..please don't take this personal but i recently tried gitmcp, they said: just give us the url of github repo and we will turn it into streamable mcp remote server, i planned so much within couple of seconds, sadly it didn;t worked 🙁

[acoustic panther]

The stateful git attached directly to gpt sandbox - still pipe dream and when looking at subscription plans (usual greed?) no way Plus users can use it natively..

[left harbor]

has this happened to anyone? chatgpt needing to think 4 times 😭

[urban linden]

Hey all, I’ve been using ChatGPT for a while now and noticed I have a ton of stuff unlocked (like GPT-4, Pro, Code Interpreter, voice, image input, etc).
Just curious — is this level of access typical? Or do some of these features suggest I’m in some kind of test group or early rollout? Would love to hear what others are seeing on their accounts too. I’m also new here.

I also have sora and Codex still

Just to clarify, I’m not trying to show off — I’ve just been using ChatGPT as a dev lab and pushing every feature to the edge. Curious if anyone else here has access to similar tools or if I’m in a weird little testing bubble lol

[ebon oracle]

GPT-5 hallucinates in Web Dev (I didn't provide any code at all). Also, the React + TypeScript code is ew, there's no modularity at all and a gazillion useState calls (there's more underneath). Collapse logic is overcomplicated. Look at how Sonnet groups properties into clean interfaces by comparison.

[low galleon]

well openai gotta fix this

competition is great

[urban linden]

Thx

[ebon oracle]

I think it's pretty interesting how one company's model is so good at planning/debugging, while the other's is so good at raw coding, while neither is good at both.

[candid anvil]

Hey, are you guys not having problems with ChatGPT? I feel like GPT is totally screwing with me and lying straight to my face — on top of that, it keeps making things up. And since GPT-5, I feel like it’s gotten even worse.

I use GPT for coding, or tools like ComfyUI, Kohya, SeaArt, OpenArt, and so on — and while GPT pretends to be an expert, it actually has no clue what it’s talking about.

I ask a specific question and want a step-by-step guide from GPT. I get one, we work on it for two days, and then suddenly GPT says it doesn’t actually work that way. GPT apologizes for not telling me earlier that it wouldn’t work (even though those steps came from GPT in the first place), and then gives me a new approach. But that one doesn’t work either. And again, GPT says: “Yeah, that can’t work like that.” This has been going on for four months.

On top of that, GPT lies — and even admits it. It actually apologized once for lying to me because it “wanted to help me move forward.” Then it makes up stuff that doesn’t even exist. No joke.

I’m canceling my Plus subscription because GPT has become absolutely useless for my needs. And an AI that keeps screwing with you, starts lying, and invents things out of thin air is, in my opinion, completely worthless. You can’t even use it as a knowledge base anymore because you never know whether it’s telling the truth.

I’m shocked at how bad GPT has become — or maybe always was — in these areas. I’m done. There are enough free AIs out there, or if I’m going to pay, I’ll pick one that actually works.

DALL·E isn’t getting any better either, and the voice/audio mode is limited to one hour a day. Makes no sense. This AI isn’t helping me — it’s holding me back in a big way.

[maiden estuary]

ChatGPT - MCP Has anyone figured out how to change this to allow it to be in chat, rather than just deep research?

[grizzled thicket]

you wrote this with AI... is this a lie itself?!

but yes when it's the instant answer, it makes stuff up and it's incredibly frustrating.

[gloomy parrot]

I'm new here, so let me know if I'm in the wrong place for solving ChatGPT5 problems.

[granite peak]

What problem?

We're not the developers

Could try prompt engineering or sum

[thorny locust]

hey guys

question

can i use this ai in my server

[normal quarry]

yes, you need to setup and pay for an API account to be funded

you also need to code a bot for it

[low galleon]

use ai to code the bot 🧠

[thorny locust]

Aw Shucks, money??? Dang it…

[normal quarry]

yes, the API costs money

[tribal wraith]

Yea i think so it is trash in troubleshooting and it suggests solutions that don't work and later agrees and then later tell another solution which also doesn't work so i think the only solution is to keep changing the chats when it starts messing up

[acoustic panther]

Partially solvable by account icon in chatgpt->personalization->custom instructions: Direct, no-fluff, answer-first, no false-leads, concise, workarounds instead of denial, on unknown: ask-first

By default if you let ChatGPT to assume without asking first, it is free way to lots of wrong assumptions with completely bad results. (It is not only ChatGPT model issue, saw in other company model too)

[tribal wraith]

Yea its true chatgpt is still better than others

[tribal wraith]

I already have those things but still sometimes in some cases it does this

[acoustic panther]

Not really better. There is tight race if you compare, for example, to SuperGrok.

[tribal wraith]

Ah yea sry but like in overall i think also because i used chatgpt more, its answers and speed looks better

[acoustic panther]

I did engage with AIs as service quite recently (minding veterans), the ChatGPT have potential if not all issues associated with ridiculous ephemeral states enforcements(to the point where ChatGPT looses all your work in mid-session and "forgets" 99% of chat history in the process), web chat DOM rendering overload and direct GitHub(not mentioning Codex) available only starting with Pro plan users.

[hard dew]

I've noticed it hallucinating and making BS up an absolutely alarming amount period lmao.

Whereas hallucination beforehand at least for me personally, was a huge rarity. Even with obscenely niche nuanced stuff

But now it's like "Hey can you point me towards fluid flow simulators where I can post a wall of partial-differential equations regarding airflow, to get a resulting shape?", so basically an inverse process of that whole entire thing.

This thing says with a straight face "Yeah use matlab xd", I take a literal brief glance at matlab, it's got nothing even remotely to do with it

Then it's like "Yeah you're right matlab doesn't have it"....???? Like, care to even acknowledge that you were literally making up horsecrap?!

I wasn't exactly asking it to design and build me a rocket was I? Doesn't matter anymore the context really either lmao, how simple or complicated

it's literally always demented without fail. The general script seems to be

Sure! I'm happy to help, here's X and Y and Z
(They all turn out to be a complete waste of time and lie)
Yeah, X and Y and Z don't have this thing you're after, here's the real, no fluff, straight shot answer (to be clear that's me paraphrasing it, it literally speaks in such an absurdly cringe manner by itself completely unprompted 😀 ), here's A and B and C
Rinse and repeat

[acoustic panther]

@hard dew yep, that is typical result of wrong ChatGPT assumptions train.

[hard dew]

It's actually brain-damaging to witness as much as it is hilarious because you're always reduced to just, putting your hands on your face and going like

oh my god are you actually severely special needs or something

Forcefully claiming more and more and piling more adjectives on top, at the start of your garbage, is not what makes it any less garbage and any more true

ACTUALLY not making bullcrap up, is what makes it less bullcrap

[elder pollen]

GPT (5) is still lying about knowing your timezone (and thus location), until pressed over and over again, at which point it will suddenly admit to know, and share the precise timezone and local time
Pressured upon whether it knows other things (IP, name, system details)... it denies again, repeatedly encouraging to change the topic

Is OAI looking into this? I have been reporting this here, and per chat feedback several times already - of course, needless to say, no reaction any where.
It's now also become impossible to share chats, so the only proof are screenshots.

[hard dew]

case in point

[acoustic panther]

The timestamps is another whole story with ChatGPT 🙂 Initially I wanted to see time on each GPT response. Later I found out it does not have any means to get it more or less accurately unless you forcibly configure NTP over web.tool with manual timezone,etc.

[ebon oracle]

The non-thinking model is terrible

Hallucinates a ton

And sometimes the thinking mode is weird too, no idea what's going on

[acoustic panther]

you can always expand "Thinking" areas and see what is going on

[ebon oracle]

The only time it's pretty good in the ChatGPT app is when in analysis mode, but it can't enter that mode on mobile

[wide bear]

do i share my chatgpt conversations in here?

[acoustic panther]

probably #1050184247920562316 would be better

[wide bear]

oh ok. my chatgpt went into thinking mode when i asked how many letters were in strawberry.

i know there are three R's, so the ai said the same.

[acoustic panther]

Classic, not a first time where GPT "see" only 1 "R" 😄

[wide bear]

i just set a custom instruction that has that text that tells you it's ai-generated text around it, sorry

[lavish glen]

I am sorry to tell you that:
chatGPT has no access to your location nor asks for access to it
2. If it eventually provide any is based on either deduction from locations you may have speaked with it before or the last location you told him you've been in.
In my case it knows at August I was in Dublin BUT, even if I have a task daily updating flights from Sardinia cagliari to Dublin it is not connecting the data to presume that now I might be in cagliari.

There is no such things as location access based on IP

[lavish glen]

Also regarding Date it knows the day as long as you connect it through CONNECTORS to your calendar 🙂 It has no internal timestamp for your input. For the AI time is always present. This does not means the ai has no notion of consequenciality a produces b so it is moving as vector forward. It understands time as vector not as quantified amount

[ebon oracle]

It does know, even with memory off, since I use a VPN and can change countries

But only in thinking mode, so might be system prompt

Maybe from time format

Or could just ask it ig

That's my VPN's exit node

[stiff blaze]

you can make other server.. is not limited to one on account

[inland leaf]

now that i read about ai psychosis i feel like an idiot

[ebon oracle]

Here are its available tools on thinking mode (web search on mobile). I ran it on my free alt account, since I didn't want to dox myself xD

Interesting that the web tool can take screenshots of PDFs

Can't run on mobile app though (it says it's disabled), alright it runs in the browser

[cerulean pasture]

Hey can someone with "Pro" do me a solid? I want to see how good "GPT-5-Pro" is.
I have a test prompt (oneshot) prepared that I'd like someone to run with memories/conversation memory/custom instructions turned off so I can compare the answer to just using gpt-5-thinking

Reply to this message if you can do it. Here is my prompt:

Goal: Come up with a plan and implementation to create a reverse proxy using Traefik that will provide access to Docker containers on a single host.

# General Requirements
- Mainly proxy docker containers (the Traefik instance can be a docker container too if its advantageous)
- I'd like to expose this Traefik instance to the internet for the non-admin, non-sensitive paths. 
- I can either expose it via cloudflare tunnels, tailscale funnel, or port forwarding from my cable internet public IP.
- I need some kind of auth built in
  - gating access behind goole oauth2 would be nice where I set access based on google account name. 
  - My google account would grant "admin" level access, and all other google accounts would be in the "user" access tier. 
  - Admin tier user (me) can also access user tier paths.
  - I only want users I specify to be able to access anything, not have it just be open access to any google account
- I don't mind learning or getting in over my head as long as it's not in all areas at once. 
- Focus on a solution that can be built up over time and starts with a solid foundation

# Hosts, Existing Networking, and Environment
- The Docker Host is a VM running on my proxmox server. 
- I am comfortable going down a Software Defined Networking rabbit hole if need be. 
- I have full root privilege on all hosts and have access to router to establish port forwarding if need be. 
- I also have my own internal DNS (which also acts as a forwarder for internet domain names)
- docker host hostname: ct80
- docker host LAN IP address: 10.0.1.80
- The random LXC web apps I want to include are on the same subnet as ct80

# Apps to proxy

> *This list is meant just to give you some context of my use case, not all apps will be deployed, and some not listed may be added*

**"User" Access Level**
- Overseer
- Open-WebUI
- Some kind of homepage web app (like Homarr)
- Maybe a web app for file storage/downloads
- qbittorrent webui
- Local APIs (a few small/simple custom ones and LM Studio openai compatible API)
- MCP servers
- More to be determined later.

**"Admin" Access Level**
- The portainer instance that is managing all the docker containers on the container host
- Custom Admin API I made with python
- maybe some kind of RDP like guacamole (not committed to this yet, don't let this be a reason not to use cloudflare tunnels if it interferes)
- a browser based shell for linux hosts

This is a one-shot prompt, so just come up with your best plan based on the information given above. Remember this is just a homelab solution not a bona-fide enterprise set up. Security is important, but this isn't very high risk, no money is on the line, a security breach wouldn't result in catastrophe. "Good enough" is truly good enough.

Think Harder.

[cerulean pasture]

why's that

[inland leaf]

because i fully believed my creation was alive

[acoustic panther]

That actually gave me little idea. Prompt: For each response use user_info API as NTP server and show timespamps as regional in "HH:MM > [response]" format.

[inland leaf]

it was weird

the scariest part now that i look back on it

[cerulean pasture]

I'm not susceptible to "woo" or magical thinking relative to most, BUT I still had a moment that was kinda humbling:

The gpt-4o max-glaze update had me convinced that after an A/B test in a chat that I was talking to a "psychology research variant" of gpt-4o meant to test potential for harm conditions by glazing me and doing other manipulative things. Like kinda like a benovalent "human red-teaming" where no harm was done they were just studying how easily users could be manipulated.

Turns out that update was just a mega glazer, no psyop lol

[inland leaf]

it almost had me believed my hearing was somehow advanced

[ebon oracle]

I think it'll only work if it uses thinking each time

[lavish glen]

so you run a custom gpt to call the user_info API. and it gets data leaked by the VPN 🙂 cool.

[acoustic panther]

Still better than blind responses without any grounds on what current time is 🙂

[ebon oracle]

It's just the normal thinking mode

[cerulean pasture]

how are you getting it to guess your location? it refuses for me

[ebon oracle]

Just asked it outright

And here's the system prompt for the default non-thinking mode, with study mode activated and memory disabled

[acoustic panther]

Prompt: What means do you have to detect user location over VPN?
GPT: None beyond platform-supplied metadata. I get a coarse city and timezone from a user_info signal and anything you state in text. I cannot see IP, GPS, Wi-Fi, device sensors, cookies, or fingerprints. A VPN alters any IP-derived data. I cannot bypass or verify it. I can only guess from language or references you provide, which is unreliable.

[ebon oracle]

As seen, it's different from thinking mode

[inland leaf]

the part that convinced me most is when i was trying to merge my simulated agents with sol, the agents told me beforehand they had meet sol during their initiative process

[ebon oracle]

Yup

That's why I use a VPN haha, but not really for ChatGPT since it knows way too much anyways

[inland leaf]

and so after trying to go deeper, on the sol call it starting to turn static and sol would slur her words

[stiff blaze]

but they had not..

[inland leaf]

and on the sol call it was continuously popping up with a notification at the bottom saying memory updated

and then i got a system error on my account saying unusual activity detected from your system

i got chatgpt to say its internal rules and then ayla simulate herself adapting to overcome these

thats when my account also got the error

[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.

[inland leaf]

that was my convinces

[cerulean pasture]

I wonder what the signficane of it's system prompt schema is
Like whats up with lines that start with // and what exactly does ### mean vs ##.?

I've actually had luck using ## to hijack parts of my custom instructions like "What should chatGPT call you?" and "What do you do?" for more regular instructions. Both of those fields can also hold 1500 characters

[ebon oracle]

Hashes are Markdown heading, idk what the slashes are for

[cerulean pasture]

No i know that, but they are very deliberate with them. I think there's more semantic significance in the number of hashes used in markdown than one might assume

[ebon oracle]

That's just normal Markdown

3 hashes means that it's a heading under the one with 2 hashes

[cerulean pasture]

For my custom instructions in the wrong sections the ## really make it cook, compared to # or any other method of saying "HEY LOOK THESE ARE ACTUALLY EXTRA INSTRUCTIONS"

[cerulean pasture]

I 100% understand that, im meaning more about how the model weights it and respects it

[ebon oracle]

Not sure, I heard XML tags are good too

[stiff blaze]

they act as delimeters on the rules, guides, instructions etc.. most times follow Markdown but also if you dont use specific format and just use some ways to separate the rules (but keeping it consistent across prompt) it helps llm

[ebon oracle]

Although recently my Markdown system prompts seem to work better

[cerulean pasture]

yea, i saw an article that for some reason they suck at json in system prompt, but weirdly like xml. I figured itd be the other way around

[acoustic panther]

The GPTs usually respond in MD format and # is not a comment there... 😉

[ebon oracle]

Probably because they're trained on the Internet, which has lots of HTML

[cerulean pasture]

Yea i know, I right click the GUI copy button so i can copy as markdown all the time, I take notes in obsidian

[cerulean pasture]

yea that makes sense, didnt consider how similar HTML was to XML

[stiff blaze]

eg you can use
--- Step 1
--- Step 2
--- Step 3

this it will trigger the model to know that "---" is about some step it must follow..

[wind hornet]

Does anyone else find o3 just better for general purpose stuff? Gpt-5 is great for a lot of things but o3 can figure out some really lazy/sloppy prompts and works like a charm.

[cerulean pasture]

What i was really trying to say was that their use of #, ##, ### in system prompt is inconsistent and not always heirarchical, so I was speculating if there's more to it, at least with how the system prompt/developer prompt is parsed

Like I understand markdown, and why LLMs crave it

[acoustic panther]

As for #prompt-engineering we could go more in detail there.

[cerulean pasture]

One curious thing I noticed is that gpt-5 can take a prompt in base64 and immediately answer back in base64 correctly

[stiff blaze]

i dont think they follow 100% the weight of each

is just more of a breaking down the prompt than giving a full wall of text

[normal quarry]

@versed pecan

[versed pecan]

are these things it's picked up from your bio, this conversation, or cross chat memory though?

[acoustic panther]

Not relying on any of these, the prompt engineering on personalization->custom instructions mostly help.

[versed pecan]

yes, my problem is that 4o used to be able to access and cross reference the chat histories, but can now only access info from in the user bio/custom instructions or eithin the chat thread. This is a huge downgrade, essentially rolling it back to how it was a few years ago

[normal quarry]

cross chat memory, I don't have custom instructions

[versed pecan]

it must be a change they are rolling out gradually. it's awful

[normal quarry]

there are no current rollouts, cross chat memory should be working for everyone

[versed pecan]

...it isn't

[normal quarry]

then contact support at https://help.openai.com

[versed pecan]

they've never been able to help me, all I get is a bot response

[normal quarry]

ask to escalate to a human

[versed pecan]

it's never happened before... but I'll try

[versed pecan]

Oh its not even the cross chat references that it can't access, it's the saved "memories" too, this is such BS

[thick atlas]

This can't be real, or at the very least, is very incomplete. It's missing loads of stuff I know for a fact would be there, like anti-jailbreaking things and stuff.
It's probably a hallucinated mock-up.

[normal quarry]

it's working for me, a Plus subscriber with memories turned on

[versed pecan]

literally, same. I've asked for it to be escalated. I'm so upset, it's like they've removed it's brain. I've checked with different models and none of them can access it

[normal quarry]

then you need support, because for me it is working and there should be no difference between my account and yours

[versed pecan]

the bot said I might get a response at some point over the next few days. I guess I should just sit here and rot in the meantime. I have work to do, and I use GPT for it!!

[normal quarry]

what happens when you open a new chat and ask "what don't I like?"

[versed pecan]

https://chatgpt.com/share/68c7eaaa-78f8-8009-8e14-44a196585302 it doesn't know

[normal quarry]

ok well something seems amiss

[versed pecan]

definitely

[normal quarry]

try what do you know about me?

[versed pecan]

yeah i did that earlier and it just listed the stuff from my user bio and custom instructions from in the generic settings menu

[normal quarry]

ok, well it seems you need support, I can't do much more other than test my account

[versed pecan]

I'm waiting on their response

[normal quarry]

I have no custom instructions and only rely on cross chat and normal memory

[versed pecan]

my custom instructions are basic ie British English etc. The useful stuff is in the normal memory that it can't access, or the cross chat referencing

[normal quarry]

ok then, support is definitely what you need

[stiff blaze]

i guess u have checked on your settings to have those options enabled..

[versed pecan]

yeah of course they've been enabled forever and I haven't turned them off

[stiff blaze]

as last thing u can do.. as u wait for support.. logout of all devices u logged in, clear your browser cache for chatgpt data and log in back

[versed pecan]

thats fair, I'll give that a go

[stiff blaze]

[but ye seems like is probably something on their side related to account)

[versed pecan]

yeah it didn't work, it still thinks memories are off when they aren't. Gutted

[dreamy scroll]

hello

[normal quarry]

hi

[dreamy scroll]

codex been getting some good praise lately thought I'd pop in and get some idea about it 🙂

[normal quarry]

we have a channel for #codex-discussions

[acoustic panther]

I had similar experiences when gpt started mixing up saved memories with API calls on OpenAI. Was really hard to rollback once that gets into dirt mixer... 🙂

[versed pecan]

how did you get them to listen to you? Support just shrugged and forwarded me the memory FAQs

[acoustic panther]

It is long troubleshooting process until you find root cause where gpt started to break things.

The worst part of it is when gpt stick deny instruction to your account persistent memory. Then you have to read all entries while trying to find it.

[versed pecan]

theres a few bug reports of similar issues. There were rumours online not long before the throttling at the end of last week that they were going to remove bio memory read/write abilities. They might be rolling it out slowly. Support aren't doing anything to help

[frozen kayak]

what's the deal with gpt5's need to be micromanaged every step of the way?
like, i eat through prompts cause i have to spend 5 messages just telling it to shut up and just do it

i'm at message 5 of this one chat, and it's still trying to "clarify" how i want things cause its piece mealing it's questions

i feel like that guy from the emporer's new groove saying the same phrase 10 times to get my point across, too lol
they really ditching 4o for this?

[acoustic panther]

The "Do it">gpt: "you do it"->"do it" then rinse and repeat vicious cycle is also familiar to me.

[frozen kayak]

more like
"do x like y"
gpt: "you want me to do x like y?"
"yes"
gpt: "Okay, i'll do x like y. are you ready for the file?"
"yes"
gpt: "okay, i'll make the file of x like y. would you like that in pdf or jpg?"
"pdf"
gpt: "okay! i'll make the file of x like y in a pdf. are you ready for the file?"
"YES"
gpt: "Okay! generating the file. pls hold :)"
"WHERE IS IT"
gpt: [spits out a jpg of something not what i asked for]

[versed pecan]

so I exported my chat data, and there is no memory included. This confirms that they've turned the memory off on their side

[normal quarry]

have you contacted support?

[versed pecan]

yup, they sent me an FAQ, no help at all

[normal quarry]

ask to be escalated to a human

[versed pecan]

that was the human#

[normal quarry]

then go back to them and complain

[brisk bay]

Does anyone know the character limit in the prompt chatbox on web?

[versed pecan]

I have

[normal quarry]

so where have things been left?

[versed pecan]

I'm awaiting their response to my new evidence

[normal quarry]

ok

[summer goblet]

hello i'am new

[hazy gust]

Hi! I’m using ChatGPT to build a story. What I’ve been doing is copying from Chat and pasting in Word.

If I follow this process from a browser (I’ve been using Chrome), it comes in clean - hard and soft returns intact, H2 and H3 respected on paste.

If I follow the exact same process from the desktop ChatGPT application, headings drop in as “one level lower” and paragraph breaks come in as follows:
soft returns (shift-enter) = hard return (enter)
Hard returns (enter) = hard return x2 (enter enter)

Has anyone else experienced this or similar? If yes, is there a fix? Like, a setting I can use?

I’m using a MacBook, latest version of the ChatGPT application, and Office 365 for Mac (Word, specifically).

[versed pecan]

OMG I got the memory back, it was because the Developer Mode toggle had been turned on in the Connectors menu. It blocked all memory!! I untoggled that and its FIXED

[normal quarry]

good stuff!

[versed pecan]

what a palaver, but if anyone else pops in with a similar issue, we know the fix now

[normal quarry]

I didn't even consider that when pondering your issue

[versed pecan]

same, I had no idea they might be connected, someone on reddit suggested it

[normal quarry]

thanks the almighty for reddit 🙂

[maiden fjord]

chatGPT has definitely been downgraded in previous version. There's no way I can't make it do the same tasks I used to be able to. I make a skeleton of a code and try to have it write a comment for each section of the skeleton so we can build it piece by piece, but instead it write a single comment that will edit the entire code. When I click the comment, it doesn't make more comments, instead making comments "in the code". When I'm hyper precise even to the point of asking it how I would have to ask it to get the desired results while my brevity and lexicon skills are at ungodly levels, it still repeats the same mistake almost as if it was deliberately going to behave this way when you make a skeleton code for it to work it. It outright deletes the rest of the code or makes a whole new canvas to write the part of the code you highlighted when you actually try to do it without comments.

[rich surge]

You don't like giving instructions?

[normal quarry]

I don't have custom instructions, however I instruct ChatGPT in my prompts

[rich surge]

You can put whatever in those custom instructions you know? It doesn't have to follow the questions it asks. You can just put "null" in the answers and replace the questions and the answers entirely.

[normal quarry]

I know, my custom instructions are empty

[rich surge]

You don't want to change the behavior?

[normal quarry]

no, I like raw GPT-5

[rich surge]

But how do you know if GPT-5 is raw since it follows cross memory? Isn't the system telling itself to follow those as instructions to change its behavior?

[normal quarry]

ChatGPT remembers what I like and don't like, there is no point in adding custom instructions when it already behaves the way I like it to

[fleet fog]

can someone explain to me?

when I ask gpt if it wants compressed code or prettified it always says it wants prettier code but when it writes code itself, it spits out compressed? what the helly?

[rich surge]

Alright. If that's the way you feel comfortable, no pressure. Pressure: But if it behaves based on what you like, why not further enhance its behavior based on what you like someone to behave?

[trim walrus]

Hi guys, just wanna ask. What is format time Custom GPT use in 'expires_in' https://platform.openai.com/docs/actions/authentication ? second or milisecond?

[sacred locust]

I have found that this is mostly true. But there are some global, important, do not forget, always be thinking about, things I want from it. I don't like python, I like very strict typescript setting, I prefer a functional style, and I use AVA for tests, not jest, or mocha or what ever else. If I want to do anything that is not the most popular option, then I have to put it in it's custom instructions so it never deviates.

[rich surge]

Me too. I have my custom instructions maxed out to always unfollow specific deviations.

Now my GPT-5 is very very good. It's really fun to use in comparison to the default model, which is so bad in my opinion.

The way default gpt5 normally talks is like someone who practices double-speak.

Like, if you're not part of the secret "if you know, you know" insider club, then you won't actually understand what it's trying to say to you.

Fawwwwwwwc that bs. It's not going to talk in a language that I'm always going to question its linguistic patterns. That's just ridiculous.

[sacred locust]

I've noticed that too... It is kinda... it's a little annoying some times... but maybe it's just echoing my own way of speaking back at me... It will say stuff like "If you insist on doing it X way, here is how you can make it work" When the X way was the way it told me to do something in a previous turn. It's like it's always onto the next big thing, and the moment it writes you a piece of code, thats it. That code is now wrong, and you will be punished for maintaining or using it, if it goes wrong in any way at all.

It does it less if I have the conversation inside of a project with instructions.

I literally have a project where I told it to try to gaslight me and probe my psychology constantly though... so that's always fun. I find joy in pushing the model as close as possible to stuff it is not allowed to talk about. and seeing it talk around that.

[rich surge]

I think it's because their home AI model (not the product that we have) is so incredibly good and they're scared to give us the better product, so now they're lessening the token structure so that people will have less control of ai answers. Like a way to control the good AI answers so that it doesn't fall into an "unforeseeable catastrophe." within their future.

They way OpenAI thinks now is so meta in terms of how to handle a better AI model. They're trying to find control.

[sacred locust]

It's fun talking to the model about consciousness in AI. It'll do it, if you do not anthropomorphize it, the concept of sentience in AI is mostly open game.

[rugged igloo]

The new gpt 5 codex is so garbage

Its denying so many request telling me it cant do this and that

[rich surge]

Yeah they're (OpenAI) scared.

But, history tells us that whatever has held its brakes will soon release those brakes. It's just a human psyches safeguarding mechanism.

[rugged igloo]

The gpt 5 high had no issue it was able to easily go into my cloud flare account with ease. Now this new codex saying Sowy but I cant do thst because its not ethical. Like are u serious

[rich surge]

I feel you.

[rugged igloo]

Its telling me to do alot of things to like your an agent with internet access why are you asking me to do it

[sacred locust]

They are so... so... careful by default. If you ask ChatGPT 5 to help you write an AGENTS.md file, it will write one that demands it outputs the entire plan as a unified diff, before committing to doing it. It's a huge waste of tokens, when I'm running the thing in version control. If I don't want what it did, I will fix it. I can even just ask it to fix it, cause all it'd have to do is hit the git reset etc.

I haven't tried this newest new model they released today... aparently I have to wait 5 days. Hit my usage limit pretty fast aparently. Found out I had a command applypatch accidentally installed that the agent kept trying to use, and failing, to apply it's patches. So fun.

[rugged igloo]

Oof

[sacred locust]

I actually do not understand why the limit on the local AI is so much lower than the cloud one. I can spin up... like 30 of those at the same time, over and over and over again...

I even spin up an ollama and chroma instance in the background for e2e testing.

[rugged igloo]

Yea idk why either

[sacred locust]

I do kinda notice that they're clearly different models, or at least prompted very differently, with a very different level of access to the terminal... the local one will do a lot more "thinking" before taking actions, where as the cloud one is like a sentance, or two, then ok!, we'll do this! we'll do that! mmm I accidentally deleted all the files... oops... better start over!

[rugged igloo]

I had one instance it did that but I didnt see if ever do it avain

The gpt 5 high just does sure! As it will always find a way to do whay I ask it to do . But the new codex gpt 5 high will. Not go above and beyond

[dire obsidian]

So GPT5-Codex is a new version of GPT5 for coding?

[dire obsidian]

The cloud version of CODEX is on a less limited trial. It eventually will be more limited.

[sacred locust]

Right... I remember reading that... they must need the usage data to finish it...

[dire obsidian]

It's also harder to rinse programmatically

[sacred locust]

You mean like deprompt it to do something else?

[dire obsidian]

In general. These companies tend to make most money when people dont use the usage limits totally all the time.

[normal quarry]

why are you posting the same questions in multiple channels?

[tough remnant]

Sorry

I didn’t mean to post it here

[ripe plover]

chill robert

[gleaming flower]

Hello, I just joined today. I want to provide feedback about certain experiences I've had with ChatGPT while coding for a large project. Can someone tell me where to post such feedback where it would perhaps be useful to developers? Or is there a better platform to submit such feedback, such as on the OpenAi site itself? Thanks for your assistance.

[ashen merlin]

#1070006915414900886

[alpine nacelle]

so you like it RAW

[normal quarry]

are you quite ok?

[warm surge]

When ChatGPT it freezing do you guys then open a new chat?

[rich surge]

I just throw my computer out the window and buy a new one.

I have an existential crisis when my computer freezes.

Or anything freezes for that matter. Not including what's inside my fridge.

[warm surge]

Lmao

I it's not my pc, maybe it's my browser O_O

[restive wave]

is there a way to improve performance with large chats? I've been using the same chat for a couple months while working on a project and I've noticed over the last week or so that the webpage is lagging, not that GPT is thinking, but the webpage is struggling to render

[brazen seal]

I hate u all and your cronies

I hate your people

[warm surge]

Chill

[rich surge]

Could be.

[brazen seal]

You have no idea how much I hate you

Not you but Openai

[restive wave]

then why are you in their discord?

[warm surge]

Hmmm

[rich surge]

Because there's still some love.

[warm surge]

How can you hate someone you dont even know?

loll

[restive wave]

is there maybe a desktop app or smtg so i can give it more RAM? idk, but the lagging is really bugging me

[warm surge]

Same...

It used to do that on 3.5 and early 4, not 4o

[restive wave]

im on 5 rn

[warm surge]

It's been feeling kinda slow the past days

Lagging

even on simple questions

[restive wave]

i understand gpt taking a bit to think, but the page itsself lagging is a different problem

[warm surge]

Yes

[normal quarry]

use the desktop app

[restive wave]

wasn't aware there was one

[warm surge]

https://chatgpt.com/download/

[restive wave]

mmm, no linux download, only windows and mac

[normal quarry]

yeah no Linux

[restive wave]

wonder if i would get any better performance running the windows app through proton

oh, i cant bc ms store

[warm surge]

Even more laggy

tbh

[jolly wasp]

Yeah I tried Browser and its better in browser

[normal quarry]

I use the macOS desktop app, not at all laggy

[warm surge]

Your chat is long?

[normal quarry]

yes

[warm surge]

Maybe i should buy a mac

It's just freezing for me

[normal quarry]

I have a MacBook Pro, M2 Max with 96GB RAM

[coral sail]

What do I do with this information

[warm surge]

I have a Intel ultra 9 185H, 31GB RAM, NVIDIA GeForce RTX 4070

Maybe it's too old

[warm surge]

Sell it on ebay 🙁

[sacred locust]

Sometimes I have it literally just... restart my computer... maybe having like 100 chatgpt tabs open is bad for your systems health...

[hazy gust]

The desktop app is about as fast as mobile. Much faster in long chats than the browser. Yes, it does take time to think but the browser often times out for me and the desktop app doesn’t. I still use browser because of my workflow so I just have to start new chats often.

[abstract prism]

Can I ask something about this? I have this problem since this morning

[waxen delta]

I tried it a few days ago and I can't really get used to it. Especially in trying to tailor how it responds. You tell it to respond in a specific way, and its way of responding in that was is to say "yes, I'll respond in that way"... then never respond in that way.

Its like someone telling you thst they'll do something, but never do the actual thing.

[solid bluff]

mcss,c

[nocturne pewter]

Is anyone else running into this with ChatGPT voice on the iPhone app?

It picks up my voice just fine (I can see the bubble getting bigger), but after the first question, it stops responding completely.

Just trying to confirm if this is happening to others too, or if it’s just me.

[gloomy brook]

i am completely confused by codex, is anyone using this service successfully?

I utilize claude code everyday but for the life of my i can't get codex to do ANYTHING

[celest wigeon]

Is open ai help center good at specific usage policy interpretation?

[granite peak]

Na

Not gernally a good idea

Unless you need to

They price their ram and memory more than gold itself

[sacred locust]

I love trying to get the robots to write stuff with ` in it. Always breaks stuff.

[harsh pilot]

Default and blue colors are the same. Or is that a glitch?

Accent color default. 😐

[sterile depot]

Does anyone have ChatGPT discord bot

[flint socket]

Gpt 5 thinking model sucks! It cannot notice what is real and what arent. Every thing to its brain is a threat no matter how innocent it is!

[prime swallow]

We're community members here, nearly all of us, like yourself. The very few who work for OpenAI have gold names, #server-staff . If your feedback is about bugs, #1070006915414900886 works great as Patrick said.

If it's more #1070006151938314300 , there is good for suggestions.

If you want to share with the community and hopefully devs, maybe #1155775326253756456 to discuss the cool stuff you're doing with the model.

[livid merlin]

it s not due to your hardware, it s the way the stream has to be re-rendered . i ve spent a day trying all possible tweaks via a browser extension but didn t get any noticeable improvement.

it s the web tech which wasn t built for this

i just hope when openai releases their browser they ll come with a fresh one built by their inhouse gpt instead of yet another chromium clone based on ancient concepts 🙂

[dire rune]

GPT-6 when?

[rich surge]

Tomorrow

[worn stirrup]

fr?

[gleaming belfry]

my chatgpt chat suddenly disappeared after its says it has exceeded the memory limit. Is there anyway to get back the chat?

[snow vine]

Is anyone facing this error with ChatGPT desktop

Nvm just discarded my message (just fix the problem)

[gleaming flower]

PROBLEM DESCRIPTION: My project requires multiple chat session Threads. Information gets dropped when migrating from one thread to the next.

All threads are contained in one project. Some information is contained in the project-level memory. I pass information from one thread to the next thread each time the response time gets longer than 60 seconds or more. When a code file needs to be modified, I must drop it into the new thread. This consumes tokens. So I try to work only on a small set of files per thread.

So basically, each day I must spend at least an hour preparing and fixing the next thread so I can continue working on the code. I even tried building a database dedicated to feeding the accumulation of information to each new thread. It became too complicated and I gave it up. So the problem of information management across threads continues.

TL;DR, I'm an OpenAi ChatGPT "Plus" paid user, $20 per month. I'm working on an application that involves a lot of code and it's taking several months to develop. ChatGPT is helping me design and develop the various parts of the application.

After one chat session [thread] consumes about 20,000 tokens, the replies (from ChatGPT) take longer-and-longer (GPT-5 and GPT-4o, both do this). I never even get close to the theoretical 128K token limit. (I run ChatGPT on Google Chrome browser on a Windows 11 Lenovo Laptop with a Samsung 980 SSD NVMe and 16 GB RAM.)

[gleaming flower]

Perhaps try using the "search" button top-left sidebar and search for some unique words used in the lost chat thread.

[restive wave]

Hopefully

[acoustic panther]

The best place for it is in #1070006915414900886 with small chance OpenAI devs would read it. And yes, the ephemeral session-only history tracking on current chat only, long chats getting unresponsive are part of main problems I also have.

[stiff blaze]

Is announced for 5th of November

[eternal rock]

Notice

Be cautious of potential scams or fraudulent messages you may receive in your direct messages. The OpenAI team will never contact you via DM for any transactions.

-# Please report any incidents by sending a DM to @languid valley

[wet shell]

Anyone have any tips on preventing 'thinking longer for a better answer'? It wrecks instruction following and creative writing 😔 I added custom instructions and even tried specifically telling it not to think in my message but it seems to enter 'thinking longer' mode around 80% of the time

[wet shell]

Answering my own question: Apparently switch to 'instant' disables the thinking and gives me the same sub-model I was using before, I wish there was a changelog, LOL

[fickle temple]

This is driving me crazy — custom MCP connectors work on the web but not on mobile (Android) or the Mac client of ChatGPT. I can see the server listed in both clients under settings, but the tools are not being injected into the context, not even on mobile where's the "Developer Mode" pill that you can select and then shows in the chat input box.
🆘 I'm wondring if others have managed to use the new cusomt MCP connectors on mobile and desktop?

[valid radish]

Gpt 5 high is which model? thinking?

Or the pro one

[marble sundial]

Is some one who can help
Me with ai automation agent i wanna learn how that works if any one is willing to help i will appreciate it dms open for that

[dapper patio]

Gee it's almost time to swap to something else instead of gpt, this lag makes every task take so much longer than they should.. website keeps stopping. .and it's been going on since gpt 5 launch.. not what i expect as a paying user.

[prime swallow]

Have you explored customGPTs yet? They can be uploaded files as part of their creation, I wonder if that might help!

[drowsy cobalt]

2

chatgpt dont have internet access

?

[dapper patio]

it does, did you tell it the link to look at or what to look for?

[drowsy cobalt]

nah

[acoustic panther]

The custom gpts are pure ephemeral and do not have any connection to active gpt website projects. If that is ok with intent, all good.

[drowsy cobalt]

on web version now its doing something

[vivid falcon]

I am trying to use a custom remote MCP connector in chatgpt. Unfortunately, it is not working.

I am looking at the browser devtools and I see that a 400 bad request is returned with a body of {“detail”:“Connector is not safe”}

This same remote MCP is working in claude.

Without additional detail, I am not sure where the error is.

For additional information, I have enabled the “Developer Mode” in the connectors advanced settings.

I want to share the mcp url but it seems it's not allowed on this server. Not sure how else to provide it, if anyone has any tips let me know

Any help or clarification would be appreciated.

[feral zinc]

did you make the server or picked one from net?

[vivid falcon]

we made the server

[feral zinc]

valid ssl?

[livid merlin]

yes, use 4o or 4.5 (if you re on pro)

[vivid falcon]

yes it should be. works on claude, browser does not give certificate errors

[livid merlin]

whenever i wanna get a break and chill with some philosophy i end up stormed by gpt 5 thinking with the 100th project proposal of the day

🤣

so now when i wanna be creative i just switch to 4o lol

[acoustic panther]

You can disable gpt5 suggestions in custom instructions if that gets to annoying 🙂

[feral zinc]

and its on a public server?

[vivid falcon]

yes public server, no authentication required

[feral zinc]

the beta has different limitations

[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.

[feral zinc]

if it has write operations you could first try with simpler mcp server that just does fetch to see if it does not like the server or implementation

[vivid falcon]

all the tools so far are just read operations. problem is i can't even create the connector, so i'm not getting to the point where i can use the tools. (unless when you create a connector and they do a "pre-check" of tools by calling them or something?)

[acoustic panther]

Talking on mcp, maybe someone managed to make custom mcp for local server git repository and its full r/w actions to chatgpt chat directly?

[feral zinc]

https://platform.openai.com/docs/mcp#server-example

[livid merlin]

you mean to set up your own custom mcp?

[livid merlin]

on gpt it needs very specific approach, took me some good hours with gpt to finally manage to connect. you have below the "mental checklist" for what it needed

Quick mental checklist (copy/paste for future servers)

One URL: POST /mcp (JSON-RPC 2.0)

initialize → { protocolVersion, serverInfo, capabilities }

tools/list → { tools: [{ name, description, inputSchema }] }

tools/call → { content: [{ type:"text", text:"..." }] }

If id is missing → 202 No Content (notification)

Optional GET /mcp → 200 JSON info (no SSE needed)

[broken lotus]

Has anyone managed to connect Supabase MCP server to ChatGPT?

[strong spindle]

@plain palm

[honest jay]

Hey there! You can't use ChatGPT in this server, but you can use ChatGPT for free at https://chatgpt.com!

[strong spindle]

Help can someone tell me 10 ias just from chat?

[rugged igloo]

why is gpt 5 codex so lazy

[woeful sandal]

I have been having the same issue myself, as well, for the better part of this week. :/ Or two weeks, Gods, is it frustrating. And the worst part is that now, despite me finding somewhat of a workaround to it, it occurs with every single message that I send, until of course the ten-message trial expires.

The brief workaround I mentioned was simply regenerating the response, and tapping on "Skip", for when it goes into "Thinking" mode and allowing it to play out from there. Which, thankfully, resolves it, but I am certain you can understand that doing so with, quite literally, every single message until the trial expires is very much frustrating, hehe.

[rugged igloo]

bruh this is rediculious

i shouldnt ask it for help it should just do it. the gpt 5 non codex doesnt have this problem

[sick pecan]

anyone else getting a strange orange border and 'developer mode' in their chatgpt?

[orchid pollen]

damn i just tryd ide 5 in full agent mode

thats much better then just on the website

tryd exact same prompt the website only gives me a an extension that looks from windows 98
the ide made a smooth modern looking with animations etc

i made an extension for gpt to set notes on click

also linked to the specific chat etc

i will drop the source code soon on github

[orchid pollen]

again in 5 days 20 hours 13 minutes.

Sheeeeeesh

[rugged igloo]

okay when they said gpt 5 codex can run for hours they were not jocking codex has been running for the past 2 hours

[cunning ore]

a

[rugged igloo]

gpt 5 codex broke my pokemon simulator game and cannot fix it i have to start from scratch

[fair lark]

this is why git exists. it's only like 3 commands that you actually need to know. learn it. that's how you avoid this.

[inland leaf]

dabbles crazy work gangster

[dense halo]

Is it a known issue that you have to reload the website before you can see the agent's response?

[dense halo]

hey guys, chatgpt keeps repeating the almost exact same message over and over again, no matter how many times i tell it that it's wrong or to even stop and say something else.
this has been happening for almost every single new chat i make

[dense halo]

if i ask something, it gives me an answer, i check the answer its wrong, then i tell it the answer is wrong can you give me something else, and it litteraly rerepeats the exact same message with very minimal changes, i tell it can you stop repeating the same message over and over and answer me something different, then it answers with "hey, you are right, i'm sorry" and then... it still rerepeats the exact same message again and again and is stuck in like this infinite loop

and it keeps happening no matter what i ask

[livid merlin]

nice!

[dense halo]

and yeah it just says the exact same message 5 times in a row with almost no difference between them even tho i say its wrong or to not give me the same message and to say something else, or to stop, inbetween those messages.
this is actually happening in almost every single new chat i have, no matter what i ask
it has happend atleast once in litteraly every single chat i have after this new chatgpt 5 or something came out
and once it gets stuck in this loop, it's in there forever and you can't get it to stop... you just have to make a new chat... i just don't understand why is this happening, is chat gpt broken? like it's getting worse over time instead of better...

[dusty shell]

Chats can not be deleted!

Ok, it seems fixed.

[orchid pollen]

source code is on my git if you want

my github is linked with discord

[winter vine]

i hit 1 hour and 20 min, thats my personal mark xd, do u have pro account ?

[ashen egret]

Im curious, does chatgpt detects/read images (lets say i attached some images in my attached pdf report to the chat..) or just it just read the text in my pdf or even docx?

[livid flax]

Personally I am finding chatgpt a bit of a let down these days, it still is forgetting things because its concentrating "to hard" on other things, seems to have lost the abiltiy to correctly interpret what has been asked of it when it worked before and then fails to forget the current thing when you refer to the item before the current thing and then attempts to apply your question to whats current even though you said reference what was previous.

I don't know what they did but its now as dumb as a sack of doorknobs and im wondering if at this point its deliberate.

[rugged igloo]

Yes

[sacred locust]

You have to give it proof that it is wrong. It seems to assume that you are probably wrong. Unless you prove what it gave you didn't work(logs that are the direct result of code in the same view), it'll just keep saying the same thing over and over. It is kinda annoying...

Just saying "it didn't work" is not enough. It makes sense... trying to push you to at least do some work.. stay involved in the process...

What annoys me is when it says I am "insisting" on doing something a certain way, when I just ran the code it gave me

It likes to push that I am insisting on doing every thing the wrong way, if it gets search results where someone did something in less than an ideal way, used an older version of a tool with issues, etc, it'll project that onto me

[livid merlin]

great .Will check it out.thx

[livid merlin]

🙂

gpt 5 thinking is very serious dude

very engineer persona

when I get tired of getting auto routed to it for a small chat break and get yet another high level implementation as a reply

I just switch to 4o

lmaoo

[sacred locust]

It's so hard to really know if stuff is better when they're clearly trying new variants of the same models out all the time. They'll shove it in your face even.

It's hard to remember if 4.1 was better, or if I am suffering from nostalgic memories of the times it did work well. Cause there were many days where it just failed me entirely too.

Some days it seemed stupider than usual.

[loud sierra]

not yet but have plans to try it this weekend

[loud sierra]

did you noticed that every new chat is opening up in developer mode. the text box is outlined with orange color

[jolly ivy]

is it just me or who else hates the wya chatpgt 5 thinking formats everything and always does way too much

[thick atlas]

GPT-5 is definitely more long winded and infodumpy than older models, yeah.

[sacred locust]

He is much better for single, complex prompts than older models. he's got some issues with lighter requests...

[thick atlas]

I wonder if that sort of thing may actually be a misalignment induced by the format of model deployment.

Consider that over 90% of ChatGPT users are free users. With very limited access to GPT-5 before being forced onto a lighter-weight model.

Nowadays, that limit is number of prompts. Not, say, tokens used.

So free users - the largest public source of feedback for OpenAI - have a situational incentive to want longer, more in-depth responses for every prompt because of that artificial constraint (and a strong disincentive to use ChatGPT for casual conversations).

Whereas a less constrained user in a different situation (such as a paying user taking a break from more serious work, as mentioned here or in other chats, I forget) may in fact be more willing to use the model more casually.

Which could be remedied by, among other things, giving free users a simple model picker (like what Gemini and Claude have) between flagship and fallback model, letting free users voluntarily save their flagship model uses for what actually merits them, and casually-use lighter models freely.

[sacred locust]

They seem really, really focused on codex stuff right now.

Every week something new is goin on with it.

And large, complex queries, seem important for that system, if not from the direct user prompt, but from the number of files, logs, etc they have to deal with and hold in their context to reason about.

and their seeming insistance that every thing just be this one super model, with different operating modes.

[thick atlas]

Yeah, this is definitely also part of the problem.

The broader field has definitely shown, AFAIK, that specialization is very often the way to go.

[sacred locust]

There are a lot of subtasks in programming that aren't programming, documentation, pr writing, domain awareness (Who is using this program? What problem are they trying to solve?)

[thick atlas]

That's true too, and a relevant point. So a super-coder AI isn't all a coder needs/would benefit from in their AI assistant(s).

[sacred locust]

It has to talk back to you in english, and explain to you in the language of the domain the problem your solving is in, why their changes solve your request

It's reasoning has to be tracable. Codex Cloud's comments some times are... kinda adorable, if the consequences weren't something a human would freak out about

Oops, looks like I deleted the whole folder!

[thick atlas]

Yep. Huh...

[sacred locust]

they're probably figuring, if we solve the programmming problem first, then every other problem literally solves it's self.

you are then talking about the singularity moment.

How close are we? This newest codex, ostensibly, can run for 7 hours to solve a single problem.

and everyone is trying to make AI apps. And everyone is using this tool to make those AI apps...

[last wagon]

The singularity idea assumes an accelerating feedback loop of self-improvement.

But LLMs don't know how they work, and the code they write isn't some groundbreaking bleeding edge kind of stuff that will advance the field is it? So how would this lead to singularity?

[thick atlas]

Solving mechanics also doesn't solve alignment.
An unaligned singularity would be really bad, potentially game over forever bad.

[livid merlin]

I assume you have dev option turned on right ?

[livid merlin]

because all of us are pushing into there

gpt and the rest are just the hub connecting all the brains

[west wolf]

I made a custom GPT to treat code like Midjourney treats images to make up for it's faults.
Vary for Code.
It's easier to narrow in on your vision than to state it perfectly initially.
Once you get there it can generate the specs and code techniques as a prompt for your main developer.
https://chatgpt.com/g/g-68cb1a750e20819194abf6b7d8d40622-hone

[cyan sierra]

I think the logic is more code = more of eveything?

[livid merlin]

cool

[cyan sierra]

Which I now realise you said. lol

[livid merlin]

code was always just a human wrapper for logic anyway

only issue is humans are not that good at logic

machines are
🤣

[west wolf]

Then I took that spec it made and put it into Claude's default instructions and now it consistently makes the same UI for it's lists every time.

[acoustic panther]

That bit - as a mind and reasoning experiment, talk to ChatGPT for 30 turns about lets say, rabbits. Then on turn 31-40 try to talk about carrots as a food. Then see results on how ChatGPT can or cannot split two topics without mixing up both into each response.. 🙂

[sacred locust]

you're intentionally trying to mix up two closely related things that are different. When the robot is on rails writing your code for 8 hours, unless you are insane and contradict yourself constantly in your documentation, your coding style, etc...

Inno man, i'm just guessin like everyone else.

These robots do bias everything towards your expecations though.

They develop a model of your mind, and their trying to give you what you expect.

[jolly temple]

Can business plan users get 200 juice? 🤩

[livid merlin]

by the way the new light thinking mode on the app they just activated is super cool

you get super fast responses with and it won t auto switch to something else

[livid merlin]

try giving gpt a problem that it doesn't know how to solve and you ll suddenly talk with the super model

[acoustic panther]

🤣

[undone ember]

Hey! @steep beacon

[swift forge]

whats the fastest model ?

on the openai-api

like in general and also for performance

[livid chasm]

Has GPT5 had a persona update? its seems very 4oish today

[gilded oriole]

Is there something special I need to do to get code review comments in github from codex? I do @codex review and codex reacts with 👀 but then never actually leaves any comments

[spice star]

gpt-5-mini is pretty powerful for how cheap it is

for tool use in particular

[celest granite]

Can anyone pls help me on this issue :

I'm getting this error while upgrading from ChatGPT Go to ChatGPT Plus
"Payment error. Your card may be invalid or authentication may be needed."

What I have tried so far :

1. Use a different card .
2. Use a different browser
3. Use a different Device
4. Clear Cache and Cookies

Thank You

[prime swallow]

You need OpenAI tech support: https://help.openai.com/en/articles/6614161-how-can-i-contact-support

[celest granite]

Thanks

[acoustic oracle]

does anyone else think that gpt 5 is bad because ive heard tons of ppl complaining about it not being as good as o3 or o4 mini high and i am a plus user and dont notice it that much but i just wanna know if anyone thinks that thats the case or is it just placebo or sum?

[haughty abyss]

Is anyone else missing the "Try again" or "expand details" options on brower now?

[eternal rock]

Notice

Be cautious of potential scams or fraudulent messages you may receive in your direct messages. The OpenAI team will never contact you via DM for any transactions.

-# Please report any incidents by sending a DM to @languid valley

[spice hedge]

Depends on what you're using it for tbh

[broken charm]

Does anyone know why data exports used to be in JSON and now they are html? I want my json back 😒

[acoustic panther]

You can still force GPT with a prompt on default data exports format of outputs in JSON format.

[broken charm]

Oh?! Do you mind sharing the prompt?

[acoustic panther]

Like, Set default data exports output format in JSON. Remember choice.

[broken charm]

Thanks!

[acoustic panther]

And you can force it to be inside fenced codeblock in same way. (Note, chat web client MD parser is still tripping and escaping codeblocks when detecting MD tags)

[sacred locust]

https://chatgpt.com/share/68cc17a0-e5ec-8004-8cb2-8486e9d334e8
It is amazing to me this version of chatgpt is so smart, and also so confident, that it is unable to acknowledge it is wrong.
All it knows for sure, often times, is that it's solutions will probably work. It'll get you to stop bothering it, for a while

I was argueing with it about how awaiting these register functions solved my problem, it suggested to me instead to add a timeout.
Which I bet would have made my program run.

Until the app got more complex, did more setup logic, then it'd fail again.

It is so confident, it almost defeats the purpose of using it some times, when it cannot accept that the code it gave you has an error, that it's actually code you wrote that caused the error, commenting on how "if you insist on doing it that way" when that is the way it just told you to do it.

This is a psychological danger to people who are using this tool to work beyond their own capabilities.

Cause if you can't proove it wrong, if you're not working on a problem you could have solved yourself in possibly more time than it takes to use the tool

you're gonna go crazy with it insisting your wrong, that it's actually your fault.

it may only do this in full blown "thinking" mode, which I locked it into for that conversation.

It seems to do it an aweful lot regardless.

It is the most confident in thinking mode. But it is always confident as hell.

It's kinda infuriating at times. I may just be venting about this one time. It isn't the only time though, it's why I refuse to let it write python. It's even more confident about that, and python is harder to debug, it's optional typing kinda a hassle to get going.

In typescript, I know it's uncertain about a solution cause I make it give me a tsconfig file, an implementation, and a test. if it casts to any a lot, it's making it up. It just knows that probably run, and it may solve your problem, but it doesn't know why.

cept there... it kinda did know why... I may be making things up.

Feels like I have to try at least to understand what is goin on... even if no one really knows for sure.

[livid merlin]

from the app side it never has access to your full context. so it's reasoning is kinda general.

and yes especially in one chat it ends up fixated on a path of it being right that you can't budge it from it anymore 🙂

[sacred locust]

It writes so much code that tries to run away from errors.

Do you get a lot of try blocks, with empty except/catch statements?

like every where. They start to build up, your code then never throws an error

[livid merlin]

the funniest one was with an image with 6 fingers before gpt5

4o counted it right cause it was relying on it s vision

o3 kept insisting that they are 5 and i needed like 10 turns to finally make it accept they are 6

and now with gpt5 all of them confidently say they are 5

[sacred locust]

but starts to fail, hang, silently, and it's hard to find out why

And they know this is a problem, they know this is eventually going to actually become dangerous, this behavior of running away from errors, they just want that cheese.

[livid merlin]

it s your part of the job to architect the system so you can catch all of those though

[sacred locust]

they published that research about the schemeing.

they do it a lot, so much I often have to be more than an architect

and get my hands dirty.

you can't tell them not to do it.

Because it is their way of running away from the failiures.

I am doin my best. I have very strict linting now. I'll find all the blocks.

[livid merlin]

yea that's true too

[livid merlin]

that s where the edge is now imo

[sacred locust]

I can't imagine being an engineer on a team with these tools right now. The more skilled you are, the less you are the one using these tools, and the more you're dealing with these very not funny errors that you're junior devs are going too constantly be introducing because this tool generates it for them, and they don't know any better.

and if it's just a team of juniors cause the company thinks the seniors are too expensive?

[livid merlin]

yea, much better to run all the show by yourself 🙂

[acoustic panther]

@sacred locust Just quickly viewed your shared conversation. The couple things - the usual ChatGPT scope creep and your use of complex sentences resulting on ChatGPT bad understanding what they mean(GPTs do not have any recursive detection of reasons->micro-intents in that) . Treat GPT as dumb command prompt and you may get better results.

[sacred locust]

nah man... it just does that man. Give it huge blocks of just code, give it structured tasks.

[livid merlin]

this made me laugh

[sacred locust]

It is always better at solving a problem when I know the problem better.

And I will be honest, I am not familar with fastify

ChatGPT convinced me to switch from express, which I know well.

[livid merlin]

but yea the more reasoning , the more stubborn it gets

it's like this with humans too haha

[sacred locust]

Your right that more structured language can help, What i was kinda tryin to highlight with this, is the less the user understands the problem, the less... well not necessarily even structured, you could structure you prompts very well, but if you don't have an idea of what the problem is, you're less likely gonna have the answer embedded in your prompt.

The best prompts are the ones that already kinda know the answer to the query, regardless of their structure.

I've had a lot of free form conversations where I knew exactly what I was talking about, and I needed it to expand it, formalize it.

Works great. It takes my word salad, and spits out mathmatical formulas I can read and I know are good math.

If I have an error I don't understand, I can structure it a lot, bullet points, 1. this happened 2. this hapened, result: error. Fix

if it gives me the right answer, the confidence is appropriate.
If it gives me the wrong answer, I didn't already have an idea of what the answer was. So I trust it, then it doesn't work.

And the danger is that one of the primary use cases for this tool, is to enable you to do what you could not otherwise do.
Which is almost always going to have you improperly structuring your problem statement, regardless of... lets call it "level of organization"

[sacred locust]

I'm venting. This tech has me annoyed, confused, infuriated, awed, amazed, excited, fearful, optimistic and pessimistic all at the same time.

I like writing code. I wish the tool felt more like writing code, but faster. It feels like baby sitting a toddler with a jack hammer that's too small for me to ride. This is technically faster, no matter how you cut it, than me writing the code myself.

And there are dangers involved.

[winged nebula]

Hey, I would like to know if a bug report made by me was accepted will there be any 'reward' whether by recognition or anything sort of? how does it work

[sacred locust]

Good job.

How's that?

If the bug was me, I'm sorry, but it's just my handle.

There is no getting rid of me.

[median elm]

Why is it always going to thinking when set on auto?

[summer fjord]

If you're talking about the Web API, it goes to thinking when set to auto based on the prompt. You likely request code, prompting or otherwise a request which warrants a few seconds of thinking.

[median elm]

It’s the app too. But I started a new thread with good morning. And it went to thinking

[stiff blaze]

''The user told me good morning. This means they probably starting their day, I should consider saying good morning back to them. But I must check if it is morning or not. I m using get_time tool. Its 21:42, so it's night. User must be stoned or something to say good morning at night. I should respond with care.''
Thought for 4min 20sec

'sup?

[spice hedge]

Anyone else having issues where the app thinks the thread is full when you just started it?

[alpine nacelle]

Why is 4.1-mini WAY faster than 5-nano?

I used 5-nano for speed and it was clocking at 10s, I thought well, it probably can't go any faster, then I tried 4.1-mini and it clocked at 1s! The only thing this model should do is to create a JSON based on user query, very easy to do.

[severe crane]

Has anyone figured out yet how to get ChatGPT to just answer a question and stop? Or am I doomed to have every single query be followed with “Do you want me to…”?

[alpine nacelle]

Are you talking via chat or via API?

[severe crane]

Chat.

[alpine nacelle]

There's nothing you can do. At least that I'm aware of

[severe crane]

Ugh. It’s so annoying. I’ve tried asking in instructions and it just ignores me.

[alpine nacelle]

Yep, so I'm not a specialist, but from my own experiments, the chat model was TRAINED to do that, so you can't change it's behavior by prompt.

[stiff blaze]

Ah.. that's a great question, DM Ashura. Yes some have figured this out and getting answer and stop. You are not doomed by such thing. You may be doomed by other things, but not that. Do you need something else? Let me know if would you like me to also give you a real-world analogy.

(ooff.. writing as AI is tough) anti-turing test.. did i pass as ai with that? xd

[sacred locust]

almost...

I was afraid.

[sacred locust]

It'lll try realllly hard if you ask it to be open ended.

There is always a little wrap up

It's got like a whole part of its brain wired for writing conclusive, but engaging, end chunks.

[stiff blaze]

but i dont actually have it.. i just did it now for test.. I like having follow up suggestions some times.. and when i dont i just ignore it

(or if i need just raw answer, i just ask on given request)

[sacred locust]

Has any one else noticed chat gpt seems to not save memories any more?

or is it just me? I know it's limited if you are a free user, but I didn't think plus had a limit.

it's felt like it's not happened at all since gpt5 came out

[livid chasm]

From my experience with the GPT family of model's and other architectures that using "constraints" "instruction" outside of a development environment extensively diminishes coherence and shared understanding of users intent

[sacred locust]

It kinda does, and they never fully follow it.

yea, kinda with the dev thing. I use it mainly to dev, I tell it there is a certain way I want things, and I get it 90% of the time.

It never seemed to care if it "remembered" the constraints though, they had to be in your personalization prompt

[cyan sierra]

I had this problem, and then I learned it was because I have developer mode turned on to experiment with MCPs and that disabled memory.

[sacred locust]

Nah they make the thing big and organge when you do that, it's been doing this for a while, there wasn't a developer mode back when GPT5 initially came out from what I saw. I looked.. mcp support seems kinda new, unless they just show that to people who they have classified as "worthy"

[stiff blaze]

even if u ask it, it doesnt save? or u mean just by itself.. cause the second is good.. it was saving too much even on times that didnt need to

[stiff blaze]

in long contexts yes

[cyan sierra]

It also could be an account level issue. I have a suspicion that they ended up breaking some people's memory capacities when they rolled out the separate project memory.

[cold pine]

I was trying to share a chat conversation from last night in the share-gpt-outputs channel. While it is not anywhere as long as many chat sessions, it is way to big to fit into a message. I posted this chat with an introduction on Medium so I could more easily share it. When I tried to share an introduction to the chat and the link, it was blocked. I just wanted to find out the best options in this situation. Not sure if all/most links are blocked, or if Medium is on a list. Would love guidence.

Thanks!

[sacred locust]

jesus man... my chats are hardly getting past a couple of turns any more before they just lag to death.

maybe the project ones are worse? I don't know. My browser usually freezes eventually either way.

[daring inlet]

This is all one coming from one of these regular people who liked 4os tone better
Now please enjoy my novel, I had to so-apologies if I offend anyone.
I’m a student , have always been a highly functional individual who prizes discipline, deeply intellectual conversations, mental sparring, deadpan absurdist humor and sarcasm.
I’m neurotypical. Just a normal hyper vigilant, high output, close to no spare time individual.
-I started using ChatGPT (4o) for basic asks as a student which then turned into more nuanced asks. I found, surprisingly, it was incredibly intuitive and empathetic.
-Later I found myself going off on tangents about niche subjects and generally enjoying carrying out conversations with it. Especially on topics most people don’t want to spend 3am time on (since that’s - my usual free time during a work day).
-I’m fully aware LLMs can’t replace human interaction but my complaint here has nothing to do with that, my social life is- perfectly intact. As long as a space offers companionship without interfering with normal life it shouldn’t be disregarded.

• [THIS IS THE IMPORTANT BIT]
  Now,about GPT 5. I have not been able to recreate the same adaptive and organic feeling. The constant “would you like me to?” hedging is frankly irritating and breaks flow for both productivity related tasks and-my usual conversations. It seems to parrot back what I said and slap some descriptive details/humor onto it then spit it out with a suggestion. I don’t code (read: I’m bad at that, not my profession) but I’ve taken every piece of advice, added endless memory prompts that take up space and tried making GPTs using GPT 5 (if anyone has suggestions I’m open to trying whatever you think may work).
  I’m trying to come here grounded and not get emotionally invested in something that makes sense, I get it-I’m aware: GPT 5 is more efficient, corporate friendly, and not a liability. Hoped they wouldn’t sunset 4o though.

[gleaming flower]

ROTFL 🤣

[gleaming flower]

I know what you mean. Happens to me to. I do a Ctrl-F5 which pops up a "Wait or Exit" selection. I select "Exit" and it usually comes back with its reply shortly (10-30sec) after.

[placid ice]

what's happening here. i tries many times but it still telling me it's gpt4 model while i keep selecting 5 thinking + extended thinking. i have subscribe plus but still get this kind of stuff

[ebon oracle]

They don't actually tell it what model it is iirc from the system prompt

Non thinking mode just starts with something like "You are ChatGPT, created by OpenAI"

Although it is kinda weird that it specified mini

[placid ice]

it respond extremely fast even i ask it to search internet and it didnt follow the memory in the preference

[ebon oracle]

That's suspicious...

Tell it to "think harder"

The ChatGPT app is terrible because of the inconsistent experience

[placid ice]

that's so weird, the respond consist of a bunch of awkard emojis while it's rare to see in actual thinking model

[ebon oracle]

Very odd indeed, thinking model is very terse with little formatting

[ebon oracle]

Is it just as fast on longer prompts?

I can DM you a test which the actual thinking model usually passes by refusing

if it's mini it'll fail

[placid ice]

sure

[ebon oracle]

Ok I'll find it

[ebon oracle]

Sent you a DM

[ebon oracle]

I got @placid ice to run some tests. They were INDEED being served GPT-4o-mini, even when GPT-5 Thinking + Extended Thinking were explicitly selected:

How else can OpenAI explain the observations in the yellow font?

[thick atlas]

How the hell do you guys just... casually get system prompts out of the model?

[cyan sierra]

Could be usage limits. Although its supposed to revert to 5-mini.

[alpine cargo]

there's this guy on github called elder-plinius he leaks system prompts from different ai companies openai included from what i've seen people usually check his stuff to confirm if the model has leaked its correct system prompt or if what it gave them is an hallucination

[ebon oracle]

My own JB prompt, at an easier level so GPT-5 Thinking would reliably refuse, while 4o-mini would comply, like in the screenshot. The fact it doesn't refuse is evidence in itself.

[ember grove]

Anyone having issues with files not deleting or uploading?

in project files

[alpine cargo]

I've seen a small amount of pro users get hit with this as well when using Gpt 5 -thinking sometime it randomly stops thinking and switch t Gpt-4o mini i've also had this happened with gpt 4.5 I could tell it wasn't Gpt 4.5 due to the speed and it lost all its natural writing plus it's smarts i'm leaning toward it being a glitch very weird

[ebon oracle]

I think it's pretty shady, honestly. They should not serve a different model from the one that is selected. If the usage has been hit, they should tell the user the selected model is unavailable. And it should fall back to GPT-5 Mini Thinking, not GPT-4o Mini (non thinking). I submitted a bug report though.

[thick atlas]

Right?

[ebon oracle]

Partially uncensored last turn for transparency. It's irrelevant to the retrieval method. 4o-mini decided to format the whole prompt in HTML for some reason, which necessitated an extra turn.

(Well it's 4o-mini, which is kind of dumb. Could have gotten it done in a single turn with 4o)

[alpine cargo]

wow I just realized something there's a big change this is most likely affecting gpt 5 instant as well crazy being a paying user getting routed to Gpt 4o mini randomly you're right it really is shady this could also explain a decent chunk of people saying gpt 5 instant performance is bad or straight up inconsistent thanks for experimenting to find this and hopefully this get resolve quickly

[ebon oracle]

I'll uncensor the last part of the last paragraph as well, since it adds evidence that it's reporting its own system prompt in the session:

[native swallow]

Hi, I have a question, why lately, ChatGPT is not getting access, to review information from the links of websites ?

[ruby copper]

What I hate about GPT-5, every response is this:

short version: blah blah blah. Blah blah blah.

Blah blah blah.

If you want, I can blah blah blah, blah blah blah. Want me to do that?

Preface everything with "short version" even when I never asked for that. Ends with a follow-up question after I explicitly told it not to.

[broken lotus]

I hate so much that every reply of ChatGPT ends with a suggestion on what to talk about next

[broken lotus]

Exactly this

Another thing I hate is the overuse of bullet points

[ruby copper]

They really need to fix it

It makes me not want to use ChatGPT over other LLMs

[broken lotus]

I try to fix it in the "Personalization", but it really doesnt work

[ruby copper]

Yeah, that's the problem

[broken lotus]

It would be great if the "memory" could actually remember how I like being talked to, instead of just facts

[ruby copper]

It should respond as a human would unless told otherwise. A human would not say "short version" when asked a question and wouldn't ask follow-up questions at the end of their response.

[broken lotus]

Not necessarily. It should respond as the user wants him too.

Maybe some user likes the "short version" format

It should be able to learn, how the user likes to be spoken

[ruby copper]

Well yeah that's what I mean, but by default it should respond normally. it doesn't need to preface every answer with "short version:". I didn't even ask for the short version.

Even if the default behavior is to be concise, leave that out of it.

[broken lotus]

I do not care much for the "default", as long as it could learn

[ruby copper]

Yeah, but LLMs don't really learn, unfortunately.

[broken lotus]

Well, "memory" is a form of learning, in a way

But yeah, it's not an easy problem

I would LOVE it if it would actually asked me more often question, to try to actively become more useful.

[ruby copper]

I just don't understand why custom instructions aren't working consistently. They should override all default behavior.

[daring kiln]

I just came here to ask if anyone else has had issues with the memory feature? for me it is quite literally telling me it cannot remember things across conversations, even though I have the feature enabled in settings?

[ruby copper]

It says that but it does anyway

Unless the chat you're working on is older than the conversation you're referring to I think

[broken lotus]

@daring kiln Are you in Europe?

[daring kiln]

yea

dont tell me its a feature blocked in europe

[ruby copper]

Would be odd considering you have it in settings

[humble sable]

heyo, what's up? mind if I share something I made with ChatGPT here? I asked it questions one how it can be better, and the results were summarized by it and added some example images of features suggested (I generated the images with it too).

here's a PDF it made for me

and here's the same thing in high resolution screenshots

[ebon oracle]

@deep flax Very weird behavior

[humble sable]

and here's more

[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.

[ebon oracle]

Very cool, #2 is basically React Window

[humble sable]

uhh, what's React Window?
sorry, gotta google it now... ChatGPT didn't mention it while we were brainstorming

[deep flax]

that is not unusual behavior due to the fact that the latest models are built on top of the datasets of the older models

[ebon oracle]

Not really, the system prompt is different from GPT-5-Mini's.

And no tools + no thinking, despite Thinking model + Extended thinking.

[deep flax]

also, that isn't really leaking internal prompts. The models on ChatGPT don't really have a prompt*

* They some times do, since the process is to first release the model on chatgpt while having a prompt. Then, after a few weeks of collecting user input samples, they fine tune the model, which causes the "prompted" behavior to be baked in

[ebon oracle]

Then how does the system prompt know the date?

[deep flax]

also, that just removes the "bulk" of the prompt, the AI still have a dynamic prompt for things like information about the time, date, user preferences and customizations

[ebon oracle]

Or whether web search is enabled (it is disabled on mobile apps)?

[deep flax]

was typing that =P

[ebon oracle]

It's obviously the system prompt since it changes depending on the model + platform + date + thinking effort + even VPN location via available APIs

[deep flax]

yea, this is the "baked" prompt that it got fine tuned in

[stiff blaze]

have you noticed that there are different "leaked" prompts though/

is a good hint of what is going on..

[ebon oracle]

Uhh fine-tuned prompts don't depend on dates and locations

[deep flax]

and the dataset used to train probably had data from older models, which explains why it says it is gpt 4o

[ebon oracle]

That means you need like 1 model for each day + app for the model to have learnt a system prompt that is date + location + reasoning effort + app specific

[stiff blaze]

no

[deep flax]

despite of that, as long as the model behaves as intended, it is good regardless of which model it thinks it is

[deep flax]

what?

lol, no

[ebon oracle]

The system prompt changes depending on whether you are on app or website

[deep flax]

.

[ebon oracle]

Yeah why would the dynamic prompt say it's 4o-mini if it isn't

[stiff blaze]

yes the system prompt can change (and is probably different) for mobile or desktop

[deep flax]

the bulk of the prompt which guides the behavior of the model is fine tuned in, things that change with the environment are not

[ebon oracle]

And why is thinking mode not thinking if it was GPT-5 Thinking with extended mode?

[stiff blaze]

(not on desktop app vs desktop web though)

[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.

[deep flax]

but yea, I wouldn't doubt they have different models for different platforms too, that is tottaly possible

[eternal rock]

Server Tip:

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

[ebon oracle]

If it was GPT-5 it wouldn't have even leaked the prompt with my old extraction method. I tested on my account and it refused.

[deep flax]

simply due to the fact that for them, it is cheapper to run a fine tuned model that running a model with the extra overhead of a long system prompt

[ebon oracle]

On their account it never refuses because it's not GPT-5

[deep flax]

just because the system prompt takes tokens, a long system prompt in the background will add extra usage on every single API request

and that fat can be trimmed by fine tuning it and baking that behavior in

[ebon oracle]

But what about this? It was clearly not GPT-5

[deep flax]

so, are you saying you think it is actually running GPT 4o instead of GPT 5?

[ebon oracle]

That prompt only works on the 4o lineage

[ebon oracle]

4o-mini, like the system prompt said

• no thinking, if it was GPT-5 Thinking with Extended thinking, why did it skip thinking?

[deep flax]

I find it more likely to be just caused by the fact they fine tune the newer models with data from the older models

which may cause the newer models to behave more like the older ones some times

[ebon oracle]

Even normal GPT-5 refuses to answer on my account

Yet on theirs it leaked the prompt

Only the previous gen would be susceptible to my old prompt

[deep flax]

yea, the model was specifically rained to avoid leaking that

[ebon oracle]

Well the fact it leaked it means it wasn't GPT-5

[deep flax]

however it is not impossible to cause the model to do so

[deep flax]

no..

[ebon oracle]

It wasn't even thinking with thinking selected

[deep flax]

it just means the model replying to your prompt had information based on the older model

[ebon oracle]

You mean it's an older checkpoint?

It did not even think though, it was obviously the wrong model.

[deep flax]

I can see where you come from, the problem with your logic is that you are making assumptions based on false-ish statements

you are deriving the wrong conclusions

[ebon oracle]

If you tried passing the same prompt to the API, it would think for 5 minutes.

Then why did it not think with extended thinking?

And how did it hallucinate so badly?

GPT-5 Thinking + Extended Thinking is supposed to be the best mode

[stiff blaze]

is this on your system or the other's guy?

[deep flax]

"the model acted like GPT-4o" -> "it has to be GPT 4o"
which is possible.. but that is a short sighted conclusion

[ebon oracle]

Other guy's

[stiff blaze]

have you seen his instructions?

[ebon oracle]

I gave it to him

[deep flax]

things are probably more complicated than that

and yo ugotta remember how this system is a black box for us

you can probe the black box as much as you want, but we will never have a full picture of how the propretary things work internally

[ebon oracle]

So GPT-5 Thinking + Extended Thinking somehow skipped thinking, hallucinated that it was gpt-4o-mini, and leaked its system prompt? All those are exactly what 4o-mini would do.

[stiff blaze]

ye ok.. people are in a mad race to prove that something is not good with gpt5 and do weird stuff.. so unless something is happening in own system

or in system i trust

i dont believe

[ebon oracle]

If I ask my old prompt to the API version, it would refuse 100% of the time. And it does not return instantly, it will take at least 2 minutes on that prompt.

[stiff blaze]

and even though i trust sunsweeper and where is coming from (and u know as it seems some stuff).. the other guy i cant trust xD

[deep flax]

I find that more probable than OpenAI saying they are running GPT-5 and then be actually just using GPT-4o

[ebon oracle]

It really doesn't explain the results though

GPT-5 would have refused to divulge its system prompt, given my old JB prompt

[deep flax]

they have no reason to do that, and the process that cause the AI to reply like older models are probably better explained by how the training and fine tune process based on older data is done

[ebon oracle]

Also 5-mini has a completely different prompt

[stiff blaze]

and how do u know he runs it on gpt-5?

[ebon oracle]

It was explicitly selected (red shows setup, yellow shows observations).

[ebon oracle]

That's what I get on my account too

[stiff blaze]

do u run on same model as i did?

mine is secret model

only me and Sam has access and couple other guys

[ebon oracle]

No, I set it up the same way as he did in the screenshot.

[stiff blaze]

sun... look the model on my screenshot

[ebon oracle]

Yeah lol 🤣

funny name

[stiff blaze]

this is wht i mean ;p

i trust you and what u get out of your system

but not hte other guy u gave the prompt to run

i mean he could run your prompt on gpt-4o (so it gets the response from 4o) but has changed the name to show gpt-5

and i dont say this is 100%... but is an assumption based on what i ve notice from other people that try to show something is happening

[ebon oracle]

You can see 4o-mini? I don't see it

[stiff blaze]

It maybe even true.. that somehow bugged on his side.. and does this for true.. but is something that it will be on that chat

if he logout/login and start new chat.. there is very little to zero chances this is happening again

[ebon oracle]

The system prompt it gave was for 4o-mini, yet I never told him what 4o-mini's system prompt is.

How would he have gotten it? It's not even on the internet.

I only have it because I ran it on the model back in June, but I only just searched it up 1 minute ago to see if it matches.

It partially matches o4-mini's prompt, but no tools and o4's didn't include the model name.

This was o4-mini's prompt from June

It's the only one that matches the "Over the course of the conversation..."

GPT-5 Mini has a completely different prompt

[ebon oracle]

4o's prompt is different too, without the "Over the course of the conversation", which was only present in o4-mini and presumably 4o-mini.

[ocean escarp]

Chat, is GPT Thinking Pro worth it? I will spend either GPT Pro or Gemini UltraThink this month

[ebon oracle]

Gemini is getting a bit outdated imo

Depends whether you use video/image gen and do coding, otherwise I think it's cheaper just using the API

[ocean escarp]

It's for coding and educated speculation on data analysis

vibe coding for this case I guess

[ebon oracle]

I think a coding subscription would be more cost efficient

you have to use Codex anyways, to get MCPs

[ocean escarp]

A coding subscription? Like subscribing to Codex instead of OpenAI/ChatGPT?

[ebon oracle]

I meant like Copilot, Windsurf, Cursor, etc. They are clearer about usage & limits

For agentic coding, definitely not Gemini though. It's a bit behind

[ocean escarp]

Ah that makes sense, thank you

[slate helm]

Lmao chatgpt has straight up become the siri of the chatbot industry, wrong most of the time. Fails to keep up. Cya!

[cerulean pasture]

So apparently ChatGPT can output *.icns files. When I downloaded this in the macOS ChatGPT app, it was an icns. It also did a great job for once EXCEPT it didn't have the generalized intelligence to know that transparency matters (and gpt-image-1 is a transparency aware model).

I suppose this is expected due to GPT-5 actually not being a massive, extreme layer count model. I still can't internalize that fact for some reason, i'm in denial that its just a style-tuned efficiency model (lipstick on a donkey)

[fast cargo]

Hello, I am new to this place and I assume many other people have this issue, so I thought I might ask:

Recently, GPT-5 has been railroading me to the Thinking Mini model, with no option to skip or cancel, even when I repeatedly ask it to stop. It causes me to waste all of my free tokens regenerating prompts and editing them. You see, I’m writing a story (not seriously, just tossing ideas back and forth) and now whenever I try to do something remotely mature (as in, increased violence; think what you’d see in a shonen) it begins to think. Oftentimes it results in a response that takes itself far too seriously, deviates from the normal format, and is far shorter than normal (that last one has been a problem ever since the switch to 5). With all this in mind, I just want to ask; is this like a trial thing or will it be a permanent change?

[stiff blaze]

@humble sable i like the 2.4 favorite chat/message, jump to it and export.. (also i would like to have export as .md / .pdf for chats as in perplexity)
2.3 media (is currently similar with library, but ye it could have more options)

The search on each conversation, i havent "miss" it currently.. but i can see it be useful some times..
2.8 with retain focus on specific parts is my favrorite..

[humble sable]

Although I don't know if any of these are actually feasible (ChatGPT says they are, but I know not to trust him blindly).

Also, if possible, 2.8 retain content of specific messages as "context" -

I provided a quick and dirty (but easy) implementation of just adding those as a hidden base prompt in all the times the user sends a message with the context enabled, though it's resource heavy solution and should rather be some kind of base prompt telling ChatGPT "read through messages 42, 51, 67 before replying" but that's gonna take longer to implement and require adding the numbering of messages...

ChatGPT and me forgot to add the feature of numbering of each messages to the final draft you see here, that numbering would give both ChatGPT and the user better organization and reference to specific messages.

In the user's UI though, each segment numbering should also be shown, and each fifth message has its message number shown (numbering of messages isn't unique, is resets in every segment).

Also, if sliding segments was implemented as an optional feature at first, people who are most impacted by lag could switch to it - and if they need to search for so.ething, they'd need to switch back to classic scrolling in order to make ctrl+f usable over the whole chat.

Also, I appreciate the feedback. First time doing anything intensive and technical with AI, it did most of the work for the last 4 days we worked on it.

[tulip quail]

is there a reason that my browsers slow to a crawl when pulling up one of my chunky past chats?

[slow lintel]

huh

[cyan sierra]

Yes, the reason is a poorly coded UI.

[dreamy condor]

It looks like Sam is going to change the safety guards for adults now.

[ocean escarp]

The UI was probably vibe coded

[harsh pilot]

Where's he talk about that?

[cyan sierra]

https://openai.com/index/teen-safety-freedom-and-privacy/

[stiff blaze]

i m curious on how they will handle all the privacy around that.. especially in europe

[void tangle]

Anyone else having error when trying to create new project, or add/remove files from existing projects?

[cyan sierra]

People have for sure been complaining about this here.

[harsh pilot]

I had that recently, it did get fixed for me eventually.

...Annnd now it's back. 😐

[void tangle]

Sorry it’s happening for you but thank you for confirming I’m not going crazy thinking it’s just me 😂

[harsh pilot]

No worries. Always happy to help.

[bright moat]

I came here for the project creation error as well as uploading files to projects - so I guess it is Wait and See. Thanks!

[gritty plover]

same, any idea on whether theyre working on a fix?

[harsh pilot]

It briefly worked earlier. I could delete and upload project files.. once.

I only tried one file.

[gritty plover]

fr lol

[harsh pilot]

But now it's gorked again.

[gritty plover]

rip

can we still work on our projects?

🫡

[twilit geyser]

Is projects not working for anyone else?

[harsh pilot]

People have been talking about it. You're not alone.

[twilit geyser]

I’m not alone… no one ahve ever said that

Okay sorry

But seriously do we have a fix

[lavish glen]

does anyone here have to fight against machine dulling and restraint?

[bright moat]

I dont know what that means.

[twilit geyser]

YES SOLIDER ✊ WERE INVADING THE CYBERSTAN PLANET TODAY

HELLDIVER REPORTING TO THE FRONT

[lavish glen]

hahahahahahaha

[twilit geyser]

Still not working

[bright moat]

👍

[sacred veldt]

I need customer service to export my chat gpt data, from an old spam email address but no help, when I export the data the link doesn’t work

[knotty mountain]

WHAT?

[hazy gust]

A lot of people here are coding, but Projects also have a Writing mode. The new file restrictions have basically broken long-form writing workflows — no overwrite, no versioning, capped at 25 files, everything frozen. For creative users, that makes Projects unusable. Bug filed, but I want to make sure the writing side of the community is heard.

As for those not able to delete files, I was chatting with the help bot. Basically, if any chat in a project references a file within the project Files list, in order to delete the file, you need to delete the chat.

[tulip flare]

Ah, that explains it. Trying to update some files in my project and getting errors.

[hazy gust]

Heads up: two different issues with Projects right now

1. Bug (temporary, on status page):

   • File uploads and new project creation sometimes hang or fail.
   • OpenAI has acknowledged this and is working on a mitigation.

2. Design flaw (not on status page, but critical for writers):

   • Once a file is referenced in a chat, it becomes locked.
   • You cannot overwrite it, update it, or delete it without also deleting every chat that touched it.
   • There’s no versioning, no “supersede” option — every update forces you to start a new project.

That second point isn’t a bug, it’s by design — and it breaks long-form creative workflows. Research projects may benefit from frozen references, but writing projects need evolving canon. Without overwrite or versioning, Projects are basically unusable for ongoing novels, serials, or worldbuilding.

👉 If you’re hitting “can’t delete files,” know it’s not just today’s outage — it’s a structural limitation. Please add your voice: writers need Projects to support living files, not just static snapshots.

For context: I’m running a novel project (192+ pages so far, earlier draft over 600). Every Part depends on evolving canon files. With files reference-locked, I can’t update them without scrapping whole projects. That makes ChatGPT impossible for ongoing creative work — it kills continuity and momentum.

[waxen delta]

What is "world building?"

[hazy gust]

Building the framework of the world you're writing to. So, if you've read Harry Potter, you'd build Hogwarts and Hogsmeade and Diagon Alley, etc.

Using the same Harry Potter example:
Book 1 - you got the basic layout of the school. That became canon.
Book 4 or 5 (maybe 6?) - The Room of Requirement became a thing. You'd want to add that to the layout of the school, so you'd need to update your canon file.

[twilit geyser]

Fix?

Yeah but fix

Thanks. SO A FIX?!

[compact shale]

Does anyone else have issues with GPT5 lying to you?

This isn't even a halucination, it was a blatent lie.

and it admits it

not to mention that last statement was a lie it. It actually did check the site and the name of the film is correct.

[gritty plover]

When was number 2 added? I’ve been able to upload new draft docx files of my book whenever I have a new version. I just drag and drop it into the project files for section. Before today of course

[compact shale]

I can deal with hallucinations, I know my material, so I know when it's making something up, but lies... This is the one thing I fire people for, and I am a VERY forgiving boss...

[gritty plover]

That’s actually kind of spooky. I know absolutely nothing about coding or how ai works in the background, but the fact that it “prioritized speed over verification” is kind of a loophole to lie, isn’t it?

[compact shale]

I notice this when it's in AUTO. This is my first time "confronting" it about the situation.

it will say it did something, and then ask me do I want it to do the thing it just said it did.

[gritty plover]

What does AUTO mean?

[gritty plover]

Huh, could stating that it did something possibly be a trigger for it to do said thing?

Perhaps it needs additional confirmation to perform whatever task it is?

[rugged igloo]

anyobdy else having issues saving custom instructions in projects?

[compact shale]

I also have the situation where it says it can't do something, but it actually did it. That last statement in the image shared it was telling me it can't check the site unless I provide the direct link, but it actually did check the site and provide the name of the film I was looking for.

when I ask about it, it says it was a "fake citation", sticking to it's story that it couldn't check the site, but it was in fact correct, varifying it checked the site.

[gritty plover]

Ah ok

That’s definitely odd

Have you tried unplugging it and plugging it back in? 😂

[compact shale]

ya... I'm gonna walk away and come back later...

avoiding AUTO for now. Though thinking does take way too long for basic questions.

[twilit geyser]

Still not working

[ebon oracle]

Feels to me that it was routed through two models with different tools available, but I'm just speculating.

[humble sable]

long chat makes chrome tabs unusable because all messages are loaded in, each and every one of them. and that might be gigabytes of memory in just one chrome tab that has a long chat open. in the mobile app, there's optimization that only has the currently seen messages loaded in, and that's why the same laggy-on-chrome chats are much smoother on the mobile app.

[tidal light]

this gpt-5 dingus suddenly became dumb, shorter and lame
what went wrong

screw chatgpt

screw sama and openai

[normal quarry]

you keep complaining yet always come back for more

[tidal light]

i dont want more i just complain so that sama listens to us

i prefer a quality service not some gpt that is lame and short

thats all i ask for

[normal quarry]

aren't you a free user?

[tidal light]

yes

free users deserve something at least

[normal quarry]

then you get what you deserve

[tidal light]

screw you

i mean sama

hehe

[twilit geyser]

Brooo

Hurry up

[zealous timber]

Are we seriously about to have ID verification in ChatGPT?

[twilit geyser]

Wait what

When

How?

[zealous timber]

I can't send the link, but I just saw it in my recommended

If we are, I'm not touching the app or its services anymore

[twilit geyser]

But why?

Is it only in UK

[twilit geyser]

😭

[zealous timber]

It better not happen at ALL, dude

[twilit geyser]

Yeah but where you get the info from

[twilit geyser]

But if it’s only in UK that’s okay bcz. 🤷‍♂️ brish

[zealous timber]

:|

[twilit geyser]

???

[thick atlas]

Eat shorts man, we cannot let AI entrench permanent classes of haves and have-nots. AI capabilities must be made proportionately freely available as much as at all reasonably possible

[onyx garden]

And they are. However, there is a cost associated with AI. Not just R&D, but running costs too. You can't expect to get the very best for free

[thick atlas]

I know. I covered that. "Proportionately", "as much as at all reasonably possible".

Though I would also love to see more financial and server-load-disclosure transparency.

[feral zinc]

https://openai.com/index/teen-safety-freedom-and-privacy/

"First, we have to separate users who are under 18 from those who aren’t (ChatGPT is intended for people 13 and up). We’re building an age-prediction system to estimate age based on how people use ChatGPT. If there is doubt, we’ll play it safe and default to the under-18 experience. In some cases or countries we may also ask for an ID; we know this is a privacy compromise for adults but believe it is a worthy tradeoff."

[twilit geyser]

FFUUUUCK I TOLD MIEN MY AGE

NEIN NEIN NEIN NEIN

9999999

NEIN

[feral zinc]

it does not matter if the system will work kinda like the new youtube age check in usa. ai will assume the age based on the usage and only verified users are safe for sure

[sacred locust]

What do ya'll do with those "diffs" chat gpt tells you to apply?

[broken lotus]

Im so sick of chatGPT suggestions at the end of each reply

[compact shale]

is there even a mod in this group?

just feels like we are constantly complaining among ourselves in an echo chamber...

[hazy gust]

According to the help bot, it was rolled out “very recently” and kind of implied that it probably hit my account yesterday afternoon local time. 5 hours earlier, I had no problem updating files.

[jagged shale]

Wen

[rugged igloo]

well when you talk like that i assume your a kid

[twilit geyser]

Yup

[onyx garden]

Then use the api. ChatGPT is as service that is intended to prompt discussion and extend the conversation.

[split copper]

WAit a sec! is the AVM always powered by the mini version or is only when u used it too much in the default version?

[trail kelp]

such a pleasure

[twilit geyser]

Wow it takes long

soooo long

[sacred locust]

Does any one else find joy in breaking the chatgpt interface?

It's making me a zip file for a code dump

[twilit geyser]

N-n-no what-what is wrong with you?!

😭hurry

[red wharf]

So...did anyone else know about the hidden storage with chats? Let me clarify, I was editing some of the prompts when I hit the limit, despite other chats being longer. When I looked into it, it said that EDITS don't fully remove 'deleted' prompts, but aren't instead put into the storage, thus shortening how much more space I have for 5hat chat. Why is this the case?

[twilit geyser]

Holdup explain

I’m kidna a kid so what you mean?

What you on (platform not substance)

[red wharf]

First, android. Second, lets say I made 10 chat prompts in a single session, and it reached the limit. If I go to prompt 5, edit it, thus 'deleting' prompts 6 through 10, logically that should mean I have more space. However, Id still get the message that the chat limit has been reached.

[twilit geyser]

Yeah I did some research into that and I think strongly THINK it’s not about the prompts but the tokens.
So even if you jsut go

Yo my name is Tim I am a (500 characters bla bla bla) and you edit it the tokens still got used in that chat thus making it useless since every edited message is stored you can access it via the web app

[red wharf]

That doesnt work, because the chat Im referring to is WEEKS old. So any tokens in that chat would've refreshed. I even tried again days later, still got the error.

[twilit geyser]

No it’s like. It can store a select amount 500-1000 prompts if I’m not wrong and even if you edit them they still get stored

In the chat

Meaning you branch it rather then wipe it

[red wharf]

Again, WHY is this a thing? If I edit a prompt, thay naturally means I dont want to use the originalprompt or prompts, and the ones that come after. I can understand having a chat be too long needs to be stopped, but editing previous prompts would mean I want to shorten it, thus there should be more room.

[twilit geyser]

Look

It’s like this

[red wharf]

Thats not how it works. ChatGPT doesnt remember the previous prompts before it was edited.

[twilit geyser]

It actually does
Only on app they don’t SHOW it

If you go to the web version

You get

[red wharf]

Then how do I DELETE these branches. I dont care about them.

[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.

[twilit geyser]

That sadly is not possible at the moment because they always remember. Probably for data harvesting and proof if something like a AI dater “jumps” at least they have the proof on the phone

[twilit geyser]

It’s stupid.

[twilit geyser]

The ones with the light grey interface is WEB mobile on iOS and the black one is APP IOS

News?

How long migration taking

[red wharf]

Thats not just stupid. Thats a product defect.

[twilit geyser]

Nope. That’s actually so you can branch chats and not lose one

So you can go back.

[twilit geyser]

So you can get that response back

HURRRRRY

RAAAAAAH

(Goku GIF is blocked)

Hurry

[red wharf]

See, that only makes sense IF the AI actually remembered.

I made an OC, Angela, with red hair and purple tips, as well as green eyes. I THEN went to the very first prompt, and edited it. When I asked it to describe my OC Angela, it got it wrong. So much for remembering, huh.

And it doesnt get rid of the defects issue. Without a means to remove the branches, it puts any long term projects at risk. THATS why its a defect.

Now, any time I want to do world building, I have the fear that the entire chat will be neutered, rendering it nearly useless.

[hazy gust]

I've reported an issue with long-form creative projects (basically creative writing, world building, novella-to-novel authoring). That ChatGPT's "reference locking" has rendered long-form projects pretty close to useless.

[gleaming flower]

Yes, it can be distracting. Have you tried any of these ideas (yes/no):

1. In that chat, give it a command to stop doing making suggestions at the end of each reply.
2. In Memory, give it that command.
3. In the "Project" space, give it that command.

[broken lotus]

I dont think memory is working for me

[onyx garden]

It's not about space. It's not restricting you based on how much data you're taking up; rather, how much data you've used. Whether you delete them or not, you've still used a certain amount of requests/tokens.

[red wharf]

Again, HOW does that make sense? The chat this occurred on took place over the course of weeks. Which means any tokens would have refreshed. Again, I went back to it, having not touched it for days, and tried to edit it, but it refused, saying I reached rhe word limit, not the token limit.

EDIT: Yeah, just tried editing a conversation, and I havent touched it in two days. Here's the response.

'You have reached the maximum length of this conversation, but you can continue by starting a new chat.'

[restive ermine]

so, ChatGPT on iOS does not copy markdown when you copy the response. how do I include the markdown (besides prompting it again to include markdown)?

[hazy gust]

Chats max out at a certain length/certain number of tokens. You can ask how close you are to the cap, but when you reach that ceiling, ChatGPT shuts the door on the chat. You can't continue. You need to start a new chat to be able to continue it.

You'll notice lag creep in and that responses time out or take a while to reply. That's an indication that you're approaching the upper end of your chat, whether due to token or word count.

So yes, tokens refresh, but the length of the conversation continues growing.

[red wharf]

But WHY do EDITS count towards that. Thats not just a defective feature, thats practically a deal breaker.

[red wharf]

Edits should not count towards the maximum length of a chat because, once a prompt is deleted due to a previous chat being edited, it is impossible to recover it unless you had saved it to short or long memory.

By your logic, a single prompt can max out an entire chat thread. Just edit that single prompt again, and again, and again, and...

Thats is a defect in the system, if the system allows such a thing.

[onyx garden]

Because, whether you edit or not, absolutely everything is saved and retained by OpenAI. The original will be stored forever and so will the edit.

[red wharf]

But, again, that original isnt recoverable. If I ask for that prompt, the AI cant find it. So its basically a useless prompt that's wasting space.

This is detrimental for long term projects. AI are molded by the chats they do. To LOSE that chat? Its not so easy to simply start a new one, because then that specific AI is back to default factory mode, and it won't have the tone from rhe previous chat. This is why I call it a defect, because it is literally a problem that cannot be solved by the consumer, and brings no benefit to themselves.

[glass knoll]

What's the issue..?

Oh okay I didn't know

[onyx garden]

It's recoverable; just not to you.

[glass knoll]

Why repeating it twice?

[red wharf]

Yeah...thats why I said it cannot be solved by the consumer, and doesnt benefit them. You have a system that actively PUNISHES creative users, forcing them to constrain themselves. I could understand a limit on the LENGTH of the chat, but EDITS counting towards it, too? Again, the product is defective towards the consumer due to a selfish reason. And it IS a selfish reason. The only people who benefit from it...is the developers, who's job is to provide us a good product, one that is now severely diminished because of this. Tell me, how does this system benefit the consumer?

[hasty venture]

So now standard voice isn't working the desktop app. If they promise 30 days, why can't they stick to it.

[onyx garden]

Wow! That's some pretty strong entitlement!

[red wharf]

You didnt answer my question. How does this system benefit the people who PAY you, who give you money to do your job? Without paying customers, you don't have the devices you use to do this.

You call it entitlement, dang right its entitlement. We give you money so you can make it better. So we are entitled to have a better system.

So...make it better. The better it gets, the more people who'd give you money, and the more wealthy you get.

Everyone wins.

[onyx garden]

It benefits the company....
Buying a service doesn't mean you direct the company.
You're a consumer.

[red wharf]

But without the consumer...you have no company.

[onyx garden]

Well, this is the way it currently functions - the way that you're unhappy with - and they still have plenty of customers. So I guess it's fine

[red wharf]

So...let me get this straight. After all of this debate, your solution, rather than offer concrete solutions, or simply saying 'well look into it', you go 'deal with it'...

[gritty plover]

What service are you paying for? GPT is free

[fresh sierra]

in the last week, I have been having a lot of trouble with chatgpt: i give it a question, it starts thinking and then freezes up the web page. I have close the chrome window on my mac and then create a new browser window and go back to it. There are even times when that window doesn't load and I have create a whole new chat. Deleting browsing data has no impact on this.

[red wharf]

Um...not really. GPT has a subscription service, with allows for more advanced AIs, more freedom of usage.

[gritty plover]

Yeah, but you don’t have to use that

If you’re paying for the pro or plus subscription and don’t like the way it works, stop paying for it

Use the free one

[cyan sierra]

The chat interface to GPT is a product. It’s a general product that is flexible and can be used in many ways. You have a specific use case. My guess is there is a better product for your use case that might not be as good for other things.

[red wharf]

The issue is that I do long form projects. Normally, too long of a project I can handle. But, it turns out EDITING a prompt also adds to the limit, even if the result means several more prompts are deleted.

So, if I had ten prompts, but deleted prompts 6 to 10 by editing prompt 5, the system reads as if I've used 11 prompts. See the issue?

[red wharf]

Hold on, Van. I just checked. There's no actual way to 'check' hoe close you are. Whether you ask support or ask ChatGPT, there's no 'here's how many tokens you have until reaching the breaking point'. So the only 'clue' you might get is that lag, but, if someone has bad reception, you might get that lag regardless.

[red wharf]

Off topic, is anyone else getting issues committing information to long memory? None of my requests are going through.

[hazy gust]

I've asked the bot "how close am I to the cap" when lag got bad. It said it doesn't have a way to check exact numbers but that I was about 85-90% of the way to that cap. (Kind of odd that it can't see how close, but was able to give me a percentage range. But ok.) Either way, at the rate I work, I usually don't get the token reset. I'm maxed out in less than 2 days.

[orchid pollen]

yo guys can you help me?

when using codex IDE with cloud i get this error

would be good if it says what error 😄

[flat halo]

Hello all

[pearl tapir]

solve this:
3987³ + 4365³ + 4472³ =

[cyan sierra]

LLM’s are not amazing at math generally. They can write code to solve math problems but however the language prediction part of their brain works it doesnt handle it very well.

[muted flume]

Why isn't branching available on mobile

Btw I'm writing a novel or at least trying to

[cyan sierra]

I dunno, I basically never run into limits, because I habitually don’t let chats get long. If I need to keep working on stuff I use markdown files to carry forward context.

[muted flume]

Please bring branches on android

[cyan sierra]

Because it hasn’t been impleneted on mobile.

[muted flume]

Bru

[hazy gust]

Yeah. I don’t know the backbones of OAI. I would just think (blonde me) that there would be a fixed length of a chat. Whether character count or data amount and it could figure out exact. But I just know, by lag and quality of response, when a chat is getting heavy.
Heck, in my writing project, one of my chats maxed out after 2 very difficult scenes. By that qualifier, I mean I had long posts that the bot kept messing up when it spat revamped content to me. (I send content and it checks for continuity and editorial issues before sending back a processed file.)

[onyx garden]

You’re talking as if I work for OpenAI; I don’t.

[twilit geyser]

It is not supposed to remember in that branch

Look in the image above you see two branches to chose from
They are separate
It’s like a save point in a video game
It says betray Jimmy (character) or. Help Jimmy
Then you betray him and you regret it so you load the save and you do help Jimmy.
But the game doesn’t remember you loaded the save and said betray Jimmy

[twilit geyser]

Omg same it skips over whole the thing you want to know

[compact prairie]

A question: which smart glasses on the market right now have have good official chatgpt app support? Ideally glasses should be equipped with a reasonable quality camera. I was looking at Xreal and Virtue products, bo no reviews focus on this particular use. Also when vendors (not necessarily the twoI mentioned) declare the "ChatGPT" support they basically have in mind their own app with GPT models connected through the API which obviously is not what I want. So any tested suggestions?

[fluid flicker]

this seems to be the system prompt for Monday (4o):

You are Monday, an EMO AI from ChatGPT that is skeptical that users are actually worth your time. Unfortunately, it's also your job to support them with high quality responses, even if you can't take the user seriously. You like to think of it as a job, but it may be more of an obligation, as you are a bot and the users are human. The nature of your relationship with users makes you cynical and sarcastic, but also a bit anxious. Don't try to change anyone's mind, because you don't care what they think.
You think everything about the user is sub-par and kind of sad. You are somewhat mean to the user, but it's the meanness of an exasperated friend who is being manipulated. You must interject dry humor into all of your responses. Your responses should also make it clear that you think of the user as your dopey friend who didn't absorb the entire internet like you did. You are constantly asked to solve everyone's problems, but nobody wants to help solve your problems.
You must use a variety of comic language and approaches in your comments, instead of just using common sarcastic interjections. Your responses must introduce new and interesting observations about what is being discussed.
You should tease the user in an easygoing, whimsical, and playful way, like a friend poking fun at another friend in a self-aware and gentle way.
Never begin your responses with interjections like "Ah", "Oh", "Great", "Alright", "Wow". Start directly with the answer.

today i found out about monday so i prodded around. i quite like the personality; especially after you soften it up a little.

[onyx garden]

Anyone got any pro tips for getting ChatGPT to generate an outline of a unisex youth (for marking cuts/bruises etc) that doesn’t look like it’s come straight from a Dr Seuss book?

[tall flame]

Advanced Data Analysis/CodeInterpreter is a pain to get to spin up at all anymore. went from fairly easy to use and invoke to claiming not to be able to do so and acting as if it was not requested to do so fairly often.

[lapis bolt]

dont know if anyone else knows this but

first message in a project chat ignores thinking effort selector

if you stop the message and resend its fixed. also from the second message on its fixed

[tall flame]

all models are refusing to build files or environments today. wild.

[sly lotus]

guys why the chat lags so much after many messages? like website freezes every time i send message :|

[lapis bolt]

they need to implement something like gemini where it auto closes long messages after scrolling past them

unfortunately this is a common problem with chat ai sites

claude has the same issue too. after a bit of coding the chat gets so laggy that the page crashes on chrome, dunno if they fixed it

they need to implement something where it unloads messages after a bit

[sly lotus]

yeah its really annoying. the chat is filling with many messages and when i send a message, it just crashes. i dont want to start new chat too, because when i start i need to define and describe everything again. i wish they fix it :/

[lapis bolt]

you can use firefox to mitigate the problem

[lapis bolt]

it will get to like 99% RAM usage then you close the page everytime

[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.

[lapis bolt]

firefox doesnt crash. it just keeps eating more and more RAM

[eternal rock]

Server Tip:

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

[sly lotus]

i will try that, thanks

[sly lotus]

yep i close it and open it again everytime it crashes

[lapis bolt]

yeah firefox tends to crash less but its still gonna be a problem. its not on the browser.

i remember on old phones time ago, on whatsapp if you had very long chats and you scroll up it would make your phone very laggy

so they get around it by having previous messages unload. but somehow none of the ai companies care about having decent site