chatgpt-discussions

[gaunt ermine]

i use the project feature (on claude) mainly for programming projects so context is usually huge for me

[urban carbon]

Feel that it is like a bad joke 😂 really hope that it is just a problem with the optimization in chatgpt if not then I am not sure that only the image gen worth the 200 usd if now it cant be even used to give proper code 😅

[tulip brook]

Is it just me, or when I use the o4-mini or o3 models, they don't generate long code at all and completely ignore user instructions?

[urban carbon]

Not only you...

[willow stirrup]

Plus users are now left with the worst models for coding...so. they all literally are useless.

[tulip brook]

I've been using GPT since the very beginning, and I’ve always found each update to be groundbreaking — but this time, it’s the worst one yet.

[willow stirrup]

true

[crisp cairn]

seems like they use canvas now and code is not long and has issues - i am wondering, o3 yesterday produced some script for me and author in the code added was "author: 4o" - not sure maybe this was the real author 😄

is there a way to turn off canvas at all?

[idle ingot]

o3, o4-mini ignores my instructions to consistently generate short text. I want to get o1 back.

[tulip brook]

When coding, I specifically ask it not to use the canvas and not to skip anything — and sure, it tries to follow that… but yeah, it still doesn’t work properly. LOL.

[willow stirrup]

yes, you can do this in the settings I think

[crisp cairn]

i will try it maybe without canvas it will produces better code

[idle ingot]

Obviously there's a soft limit circuit for token limits. They're bothering us to minimize spending.

[native otter]

Don't tell it not use a tool. Instead disable it from settings properly.

[tulip brook]

Disabling it probably won’t help. Seems like there’s some kind of built-in limitation.

[native otter]

You tried to not have it use a specific tool, so I showed how to properly disable it rather than telling.

Also prompting it stuff like "Generate long code" is weird. That's not good prompting. It may do the exact opposite...

Code generation is not about "longest code = good" at all. The amount of lines it generates is based on the task. It tries to be efficient.

[tulip brook]

Totally agree that code length isn’t everything. But right now, the issue is that it’s not finishing what it's supposed to do. It’s not about asking for long code — it’s about asking for complete code.

[idle ingot]

o3 does not prioritize the quality of the code, but rather omits some of the code and generates answers without implementing some features. For short output

[native otter]

The key is NOT telling it to generate complete code. It's better to say *Edit code and adjust as accordingly. The prompt should encourage the model to not omit stuff indirectly.

I had o3 generate 1600+ lines of code fine for a test. The quality wasn't great of course. But it did it.

[idle ingot]

I asked o3 to enter a substantial amount of data and write an 8000-character report, but he kept producing 3000-character reports and insisting it was 8000 characters. I managed to convince him through Python analysis, but never been close to 8000 characters since then.

[native otter]

Telling it specific characters also will cause more harm than good. Just saying a long report could have been much better.

[willow stirrup]

yep, none of the existing models are capable of reading lots of lines and give an adequate result

[idle ingot]

I tried various prompts, but it didn't help.

I entered only the right level of text material within the context length limit

[tulip brook]

I get where you're coming from, but I don’t think it’s just about how we phrase the prompt. The thing is — when I use the same exact prompt with o1 or o3-mini, they don’t skip, they follow instructions, and they generate complete code just fine.
So it doesn’t really seem like a prompting issue — it feels more like a change in how the newer models are handling generation. Maybe it's intentional, maybe it's a limitation, but either way, the user experience definitely took a hit.

[idle ingot]

o1 always wrote me the amount of work I wanted. It didn't matter up to 20,000 characters. Even now, o1-pro meets my requirements well. But o3, o4-mini couldn't.

I'm currently using o1-pro, but I'm desperate because I know it will be replaced by o3-pro soon.

[willow stirrup]

so pro users are gonna be screhhwed over as well lol

open AI is just dumb...or greedy idk)

[native otter]

It's true that the new models try to be more efficient. I found that one would have to change the prompting method to manually encourage the model. While o1 did it no problems.

Actually o3 is the first model I tried that's really better than o1. Does analyze complex C++/Win32 API code very well. And follows the instructions better. At least when working inside a Project (with proper CI and Files)

[tulip brook]

Yeah, fair points all around. Seems like our experiences just differ depending on the use case.
Hopefully future updates bring the best of both worlds — stability and flexibility.
Appreciate the discussion

[native otter]

I do kinda hate o4-mini though

[willow stirrup]

the problem is not with efficiency, the problem people payed for something they cant even get lol

[tulip brook]

Facts. Efficiency means nothing if the thing people actually paid for isn’t even accessible. That’s not optimization — that’s a downgrade.

[idle ingot]

I think it's pretty awful to ignore our user guidelines for efficiency, even though the model can technically generate the amount we want. We paid a reasonable fee.

It's a downgrade, it's laziness.

[tulip brook]

Of course, I know it’s not fair to compare non-inference GPT-4 with the current inference models — but still, the word “lazy” really brings back memories of GPT-4.

[native otter]

It's definitely lazy in Canvas as it has forced limits

[idle ingot]

o1 also had canvas. o1 was a little clumsy with the tool, but not lazy.

[willow stirrup]

Speaking of efficiency in these models, if you generate the whole thing at once, app or script, whatever, it is drastically different if you ask to generate by section for example, in which case it starts to hallucinate like crazy.

[half vine]

o4 mini high cant do the image reasoning?

[willow stirrup]

so making it "efficient" by forcing to give shorter answers, is actually makes it much worse

[eternal rock]

`` Server Tip: ``

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

[half vine]

and also "full tool access" means for o4 mini high and also for o3, or only for o3

[willow stirrup]

and i`d rather copy and test the whole thing at once, then copy fix by fix manually....

nothing efficient here

and by wasting time to copy fix by fix, im gonna get even more errors, because it makes more errors by generating "efficient" snippet

the end

[native otter]

Well my two cents about is when testing for analyzing files with 1200k+ worth of tokens in a project, it finds stuff o1 couldn't and the answers are much better.

[willow stirrup]

and while you copying the fix, sometimes it doesnt even say where and what exactly to replace lol so you can get the idea how efficient this is lol

[dense halo]

wierd, it never happened to me before, not even on my main account.

[willow stirrup]

you re sitting like an idiot trying to find the part of code you must replace

[native otter]

You have to prompt it to say that.

But on my side it actually tells without saying...

[idle ingot]

I'm experiencing significant disruption in my work due to shorter answers. Even if the quality of the answers improves, I can't use them if they fall short quantitatively.

[willow stirrup]

nah dude, thats not vibe-coding anymore, thats just....coding lol

I cant even find the part I have to fix, because the AI thinks "a="and "a =" are the same thing lol

so asking it to be specific doesnt help much either

[native otter]

Well my honestly you shouldn't be "vibe-coding" in the first place. more of using the model to learn...
And even if you want to it to give you the results outright, requesting significant stuff is not a good strategy. the quality wouldn't be good at all. You have to split across multiple smaller steps

[willow stirrup]

oh here we go... lol

[willow stirrup]

Im a bad person because Im vibe-coding

[native otter]

Yes

Well, it's fine if you doing it for personal stuff or prototyping. Hopefully not in production though...

I had o1 write a script for me to fix an issue on my OS casued by OneDrive I didn't have time to bother fixing. Especially when I was very close to a deadline

[willow stirrup]

I dont care and its not the point.

[native otter]

yea yea I get what you mean, o1 was more open to giving the longer outputs...
And that is inded the case by default. But prompting it differently can have o3 generate a lot of lines (Don't question the quality though)

[willow stirrup]

the point is OpenAI screwed me and others today.

[full coyote]

since when they never did screwed user?

[willow stirrup]

idk but its a massive screw up, I hope they get enough backlash to put o1 back where it belongs lol

[full coyote]

dare I says its took miracle for that to happen.

[native otter]

https://chatgpt.com/share/6800ebce-c8e0-8003-b228-03ffc99a027f
11021 Characters Output.

Was o4-mini-high though

o3 Rejected my prompt as possibly harmful

I can test the same prompt on the API with o1 to see how many character it outputs

[willow stirrup]

Come here and get your "smartass" award.🤡

[native otter]

Bro why are u mad

Some get really hurt from sharing opinions lmao

[idle ingot]

bro it is only 2K token

That's enough for GPT-4o.

[native otter]

Yea. that was an answer to the 1k characters claim...
o1 did 10861 characters on the API, so they seem to be around the same.

[idle ingot]

Because 2K token has nothing to do with downgrade or not. Laziness happens beyond that.

o3 is a model with a maximum output size of 100K. It is not appropriate to compare at 2K.

[formal hollow]

oh hey, what's going on here?

[native otter]

It's 100% lazier by default. it requires really precise prompting to generate lots of stuff...
But not to much different from o1 in my experience, maybe because my prompting style was always like that (my opinion)

o1-preview was better than them all in terms of Long Outputs anyways. I remember that very well.

[idle ingot]

The level of volume required varies from user to user, and for now, token 2K output is clearly lacking for me. o1 was able to generate up to 10K with ease. Only with a simple specification of the volume.

[formal hollow]

chatgpt is getting dumb! it does not support long context length window compare to that llama

[idle ingot]

Many users, myself included, feel quite different. And it's clearly disrupted.

[formal hollow]

and actually for a long conversation, it shows an error Error in the message stream, try again!

[native otter]

I honestly care more about the quality of outputs than text slop anyways...
o3-mini-high wasn't any better than o1 in terms of intelligence, but o3 seems to be the first model that does better than o1 in my tests involving complex programming questions and large codebase analysis so far

[formal hollow]

actaully o3 is not good, o1 is still better but dumb, you should ask me why? well I can tell you

[native otter]

I mean if you want say it lol. Everyone has their own experiences...

Just a general tip: Always Disable Memory, it can cause context pollution. (When dealing with complex questions)

[tardy lance]

openAI is not open

[native otter]

o3 can now read previous chats not just memory list entries unlike o1

[idle ingot]

I don't deny that the output quality has improved, but the downgrade of the text amount has gone too far.

[formal hollow]

Sam Altman is saying to open-source Codex CLI, I can't believe but it's true but lol it will be different from those chatgpt models

[tardy lance]

Ah

sly

facebook is now Meta

SAME leadership

XD no lawsuits lol

[formal hollow]

yup it is, and they launch their llama 4 herd, that supports 10M context window length!!!! OMG!! Go and check

[crisp cairn]

i doubt it is accurate at so large context 🙂

[native otter]

10M context window is useless if it will hallucinate anyways...
RAG is the better solution

GPT-4.1 is 1M tokens on the API if you want large context.

[formal hollow]

RAG can slow down the speed of ai model and also 10M context length window requires more GPU Power as well as RAM and storage

[native otter]

Quality >>>>>>>>>>>>> Speed

[tardy lance]

add a LOT more RAM and have both

[willow stirrup]

o3 still cant give me a proper answer. half of the time it just cuts off the answer at the beginning and tells me "heres your full script with changes"

[formal hollow]

Quality is fine but peoples are also looking for speed, and doing AI Model Quantization & Optimization and do you know about that DeepSeek?

[native otter]

Btw how many lines is that script?

WIthin 1k lines?

[tardy lance]

Deepseek is pretty rad from what I have seen

[formal hollow]

you have editted 17 times?

[formal hollow]

DeepSeek is powerful but your data is not safe!!!

[tardy lance]

XD even paid GPT will not provide more than 250 lines of python code

[willow stirrup]

No,2627

[native otter]

I know about other models yea.
Btw, other AI discussions in #ai-discussions or your messages may get removed by mods...

[willow stirrup]

yes 17 tries

[native otter]

That's just false

[formal hollow]

I am an AI Scientist and I actually running DeepSeek locally on my computer, are you guys are only limitted to Chatgpt? not interested to look around?

[onyx garden]

Yes it does. Codex is fantastic too

[willow stirrup]

dude, we are too poor to install it locally lol

[formal hollow]

main thing is to have a look on it's knowledge data cutoff date, because technology is changing everyday

[native otter]

Honestly that is even too much for o1-preview

If it was within 1k lines, I would have said put the code in a Canvas window. and ask ChatGPT to edit it, as it will go line by line...

[peak remnant]

Why didn’t OpenAI include o1 Pro in their test metrics?

[tardy lance]

have a few hundred .py files I have had CGPT make and they are great but I cant have it do more than 250 lines and it does not even correct without taking things out

[formal hollow]

lol, those guys says that it is too powerful, but I say it is dummy,

[willow stirrup]

lol Ikr, never had an issue with o1...

[peak remnant]

Lack of tool calling killed my interest in running deepseek locally

[formal hollow]

I actually say that run ai models locally on your computer because your data is safe and you can use ai model as much as you want but well still be careful with your data to use ai models on openai

[native otter]

Are you a troll?

[willow stirrup]

be sure to tell that when youre paying your elictricity bills

[formal hollow]

no? I am just sharing knowledge

[peak remnant]

You don’t look like Paul Jebasingh Emmanuel.

[normal quarry]

you can chat about running models locally in #ai-discussions not here

[native otter]

Well first you talk about other AI models in #ai-discussions, and in terms of other models, most of us know about them
I love *Claude 3.7 Sonnet personally

[formal hollow]

lol, I actually don't use ai models all the day, well you can discuss and get knowledge about AI from me, and I am here to help and talk

[native otter]

Also you definitely not an AI scientist bro

[formal hollow]

ok well, peoples are actually say this to me, but well...you can ask me any question about ai, can we switch to #ai-discussions ? if you are interested to talk more about ai with me

[willow stirrup]

here o4 minihigh just cuts it off at 228 line lol

very efficient

[onyx garden]

Mine went up to 380 lines of code. I got impatient though trying to think of features for a magic 8 ball haha

[formal hollow]

format looks good, which project you are working on ?

[onyx garden]

https://chatgpt.com/canvas/shared/6800f41430e8819193a9bf12145a406e

[willow stirrup]

a script

[native otter]

(Wasn't generated in Canvas)

[willow stirrup]

anyway thanks OpenAI for ruining my day lol

[formal hollow]

lol

[willow stirrup]

exactly lol

[formal hollow]

there are lot of open-source and powerful ai model are there, go for them, learn about ai

[native otter]

Hopefully the system prompt is improved soon to allow long outputs by default without very precise prompting...

[hexed nova]

I don't know what to do anymore

It takes me 7 minutes to just scroll through all of the text i have in that one chat(and i was scrolling as fast as possible)

[native otter]

What do you mean by voice here?

AVM or Prompt TTS (read aloud)

[thin pine]

Guys, does nano work on codex for browsing through folders? i am getting errors. says" I currently don't have access to your local file system directly. Please copy and paste the content of the file
here, and I will summarize it for you."

[tawny willow]

friends 4o mini has been introduced and put into use, but what I want to ask is how much the limit usage limits are in plas users, and is there any difference from another 3o mini model?

[honest jay]

Hello! The usage limits for Plus are described here: https://help.openai.com/en/articles/9824962-openai-o3-o4-mini-and-o3-mini-usage-limits-on-chatgpt-and-the-api

With a ChatGPT Plus, Team or Enterprise account, you have access to 50 messages a week with o3, 150 messages a day with o4-mini, and 50 messages a day with o4-mini-high.

[native otter]

• o3 is 50 a week
• o4-mini is 150 a day
• o4-mini-high is 50 a day

The difference is now it's smarter obviously, and works with more tools like Canvas and better integration with File Uploads / Data Analysis. It can also analyze images and search the web in much more depth.

[inland heron]

I just did deep research using o3 regarding certain AI models for coding and it was the most useless thing I've gotten ever

outdated models, some of them weren't even in any of the benchmarks or just years ago released, quite embarassing

[pulsar mica]

Anyone else notice o1 is just gone this morning

I have a plus subscription

[urban carbon]

Yes 😭 it is the biggest disappointment I ever had too ai related. How comes then remove o3 mini high when it is 10 times better than o4 mini high and o3 as they are now.. rly.. hope that openai will fix that if it is possible to fix..

[crisp cairn]

yes, replaced with o3

[sage bramble]

Does anyone know if there's a proper browser addon that can print a gpt print in PDF format?

[pulsar mica]

😠

[tulip brook]

I've used both the o4-mini and o3 models for coding, but it's been incredibly frustrating. I couldn't complete a single task properly.

[pulsar mica]

o3 does not pass the vibe check

Can still access o1 in copilot chat lol

[urban carbon]

Me too... I use it professionally... I was so excited after the annoucement and then... yep 😅 it is even ignoring most of my requirements.. like I have 5 pts it will tell me yes here is the complete implementation but it will give me only 1/2 of a point .. I have tried with direct download.. it is a bit better than the 200 lines of strange python it gives with canvas.. cos it can edit the existing file but it doesn't do it on a professional grade... even worse it doesn't apply the requirements how it should or forget a big part.. and assure me that everything is fine when it is not. Somemore it was really easy stuff compared what i use to do with gem

Just hope that it is just an optimization issue which would be fine. Because many good ideas and features behind but yeah if the model is so then for coding in the chatgpt interface it is pretty useless

[tidal haven]

That's weird... How is o3/o4-mini-high failing for people? It seems to do well on high-level questions about Deep Learning research papers.

What's a coding test it's likely to fail at?

[urban carbon]

Personally I do always a series of test to see how the model perform first with easy stuff like make me a pong game with this or that.. in a single html file and even here it performs worse than o3 mini, sonnet 3.5, gemini 2 even probably 😅

[visual acorn]

Guys, is it me or ChatGPT still overpraises you for no reason?

[urban carbon]

Then to modify an existing file of around 1k to 2k lines of code even you ask in a super precise prompt it will give you just 200 or 300 lines of code. Of course you can ask it with tools to modify the files but til now the implementations and modifications are super faulty

[tidal haven]

Yes, it'll totally suck up to you, but it might give you something meaningful if you ask it for critical feedback.

[urban carbon]

For a professional workflow at the end u waste too much time trying to fix the issues

[frozen fjord]

I usually use o1 for my coding work with ChatGPT, but I notice it's gone now. What would you consider the best alternative now to that model? I do need it to have ability to process files since I use the Project feature

[visual acorn]

I'm telling you, it does it just so you keep chatting with it. That is so scummy and even destructive for people who could be vulnerable due to holding the assumption that they're getting sincere feedback from AI.

[eternal mantle]

How does O3 compare to 4o for general purposes? Is 4o still best for questions that might be somewhat technical or require knowledge but necessarily math or coding?

[urban carbon]

For general purpose I use mostly 4o or 4.5.. models like o3 mini before only for coding

[tulip brook]

Honestly, if this issue isn’t fixed, I’ll be canceling my GPT subscription and switching to something else without hesitation.
I’ve lost my time, my money, and honestly—my expectations too.

I didn’t invest in this to get half-baked results.
What I’m seeing now is beyond disappointing.

[tidal haven]

It made a pong game just fine...

[visual acorn]

Looks like others are disappointed in ChatGPT too. Though I'm a free user, I've been using ChatGPT since 2022 and its suck-ups are so annoying that I've been experimenting with another AI recently (I guess I shouldn't name it though)

[urban carbon]

Dunno what happened.. hope that they just rushed to respect the timeline and they will these features for now I will continue to use gem 2.5 pro which is a beast at coding with crazy context

[visual acorn]

Why are they so desperate for bringing in new users? They already have plenty

[tidal haven]

Though the game interacted badly with their preview feature, so the key controls weren't working until I put it in a standalone html file...

It fixed that problem when I complained about it, though.

What's a specific challenge it's failing on?

[tulip brook]

At the very least, if this new model wasn’t ready or had issues, the honest move would’ve been to leave the previous ones — like o3-mini or even o1 — available.
Removing them entirely just made things worse for people who actually relied on them.

[urban carbon]

There is pong games and pong games... if you push a bit the thing u will see that it gives you results way inferors as o3 mini high without mention gem 2.5 or maybe u can do 20 iterations and u can may r reach the same result you would have get with 2 or 3 iteration from another model

[eternal mantle]

There is no O3 mini now just O3

[urban carbon]

Yes that's the point lol

[eternal mantle]

So is it better than 4o?

[urban carbon]

O3 mini was doing just fine really amazing beside the small context

[pulsar mica]

Claude 3.7 lmao

[pulsar mica]

They got rid of o1 because it was eating into their profits

[rugged igloo]

Anyone else here with a pro sub think of o3 compared to o1 pro? To me I feel o3 is much nicer and doesn't have a ego problem. And doesn't questio. Your intelligence.

[urban carbon]

Yes that's the worst cos now u can't work anymore 😕 me I wasn't using o3 mini high since gemini 2.5 came out. Kept my 200 usd abo cos I was waiting these 2 models and image gen is funny but yeah.. I still convinced that it is just an optimization issue... but still don't get why don't take them first in research preview or in beta if they are not ready

[tidal haven]

What's a prompt it's likely to fail on?

[urban carbon]

Code a beautiful pong game.. it will give it always with issue the prompt is a bit elaborated to make it epic maybe u see the kind same as in those youtube reviews and comparisons... then I wanted to edit work projects like python Webapp with py, js, html and css with few k lines of codes and it cant do it well. O3 mini could do of course with limited context but could manage it then for gemini same stuff are a piece of cake. Maybe 50k token input of documents

With tests I started relatively slowly .. then the model tells you here is the full implementation then it sent you 300 lines of code when you ask for the complete 100% code as well what is not modified.

You still can tell him to use tools to send you to download the complete file with the implementation. It works better and the idea would be amazing if it worked but yeah gives faulty results. At the end u waste much time. In a professional workflow using ai time is an important factor for me 😅 tried 5 times to make it work then yeah understood it is not ready yet. I had the same sensation with it as with using firebase 😅

[silk crypt]

It’s only getting better

I remember how I used to wait for 2 min straight to get an answer from ChatGPT 3.5 and answers were barely acceptable

[willow stirrup]

you`re not the only one. I guess its too powerful for us, and we only deserve the lesser dumb and glitchy AI. (I mean the plus subscribers:-)

[willow stirrup]

No its not.

[pulsar mica]

Not a big fan of them (no notice) removing the only model that gave plus users good results and replacing it with an inferior but cheaper for them to run model. I understand its a business that needs to make money.. but $20 is not exactly cheap and it feels like they are slowly removing value from their $20 tier subscription

Unless the only thing you use it for is cheating on HW and generating studio ghibli photos then it probably does not matter for their average users

[eternal rock]

Get notified for server events!

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

[wheat bluff]

Why can't I continue a ChatGPT conversation that I generated images in if I hit my limit? Like I don't want to generate more image examples but I still want to continue the conversation

[urban carbon]

Don't think it was on purpose 😥 hope not but I hope that they will give answers or communicate

[elfin lava]

? hit limit on image gen or chat model ?

[wheat bluff]

Image gen

Although it looks like after I refreshed I can continue the chat conversation now

[elfin lava]

oh ok

[split warren]

Exactly the same happened to me (two days ago) and I got an apology email. My bill is due in a few days so will see what happens. Didn't want to sign up again and start getting charged twice.

[winged lantern]

I did sign up again and asked for a refund for the first time, still waiting tho, if I get charged twice I'm going to lose it

[winged lantern]

Tell me how it goes

[orchid pollen]

hm new o4 is good

it solved a problem i was stuck with o3

tryd many chats o4 solved it in one

good work guys

[silk crypt]

Now I see the error of my way and sincerely apologize for spreading misinformation

[wheat kelp]

how many image generations approximately do you get per day with a $20 sub?

[short kite]

Qq, did everyone else lose access to memories in search? In the “personalize”, under the include memories from chats, it used to say “memories may be sent to search partners such as bing” or something like that. That line is gone now

[rocky tinsel]

Give back the o3-mini-high, it was much better, o4-mini-high and o3 only knows how to cut codes even though we asked him not to cut them.

[willow stirrup]

thats what you get when buying from ClosedAI lol youre getting screwed

[dull hazel]

I bought chat gpt pro but i got a set of amound time to use advanced voice how to fix

[reef cloak]

Did you guys notice any differences in text? Or in details and reflections in the 03 image generation that could suggest that it's a different model?

[paper wagon]

what do u use

[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.

[reef cloak]

People have to stop asking help questions here this is not a channel designed for help it's designed for talking about the product

[elfin lava]

then where do they go to ask for help then

[normal quarry]

this channel is fine for asking for help

[reef cloak]

I thought there was a dedicated help Channel

[normal quarry]

if people need help with ChatGPT then this is the place for it

[dull hazel]

Then why is not one Helping me

[normal quarry]

what do you need help with?

[dull hazel]

I just bought plus version from my pc but then I go use the advanced voice feature on my phone it has a set time for me to use it. Aren't I'm able to use it then I want

[normal quarry]

Plus users get 60 minutes per day with Advanced Voice Mode

[dull hazel]

Really I thought it was Inifinte tho that explains everything but Still it wasn't listed it just said access to advanced voice and normal voice then you look at it

[normal quarry]

https://help.openai.com/en/articles/8400625-voice-mode-faq

[dull hazel]

Well thank you then.

[fickle fiber]

Mostly due to throttling everyone who doesn't pay 200 dollars to receive hallucinations and errors

[silk crypt]

[inland stream]

can anyone guide me how to have perplexity deep research api in custom gpts?

[ivory iris]

It appears that o3 has access to all the chat conversations. Has anyone confirmed this?

[orchid pollen]

hm canvas still confused with multiple files

[granite harness]

Yep.

[toxic tapir]

o3 is so bad o1 was waaaaay better they need to fire the ceo

now china will beat chatgpt very much

[twilit trout]

so o3 is now the one we wanna use if it's for coding?

[toxic tapir]

no yoou have to throw o3 away

[normal quarry]

o3 and o4-mini-high

[normal quarry]

please be constructive

[twilit trout]

ty vm! I'm pro, so I saw that pro legacy was going away and panicked for a second lol. came here to see what was happening

[toxic tapir]

bro i cant be consructive this o3 modal made me so mad becasue its so bad it contiously want to write to canvas when i dnt want that

[twilit trout]

do you know if the models have been updated? some swift documentation stuff wasn't updated on the older ones.

[toxic tapir]

way less tokens

[winged lantern]

Damn

[normal quarry]

I use o3 for SwiftUI and it seems really good. o1 Pro is still available as a legacy model until o3 Pro is released in a few weeks

[twilit trout]

that's awesome to hear, SwiftUI is my exact use case. thanks a ton for the response!

[winged lantern]

To be fair hardware doesn't grow on trees

[toxic tapir]

o1 was way better

[trail wagon]

you seem to base it on your usage but u dont grasp its not just about you....

[normal quarry]

I can only base it on my usage when people complaining won't share their chats to show me where things have gone wrong

[twilit trout]

what's wrong with his response? he gave me the exact answer I needed?

[toxic tapir]

openai is lead so poorly by this autistic sam altman they should firee him

[trail wagon]

the response is the code is being left out when the code gets to large and it starts to cut out things giving you less code...

[normal quarry]

Rule 1 of #server-rules is to be respectful, try that

[toxic tapir]

@normal quarry just speaking the truth

[twilit trout]

yeah but his response was to me asking about the updated models, not what you're referencing?

[toxic tapir]

just bring o1 back

or i will usee chinese stuff

[dense halo]

then go use it. i always wonder why people declare they going to use this or that instead. is your ego so big you think others care what ai you use?

no, please, sir! do not use other stuff 😭

[twilit trout]

lollll

[nova vale]

I don't understand the hate of some of the users here on o3 and o4. I can see some significant improvements over previous versions based on testing I have done so far.

[trail wagon]

its not hate its people upset with the changes.... for me its the cutting of corners on large code.....

[toxic tapir]

@trail wagon same her its very bad

[nova vale]

I was just trying some text tasks, image analysis, data analysis, math problems and some coding stuff. It didn't do any cutting for me.

Definitely better then o1

[toxic tapir]

@nova vale thene you have tro try code weith more the 400 lines its very bad

400

[nova vale]

It did some coding and produced above 600 lines of code and worked perfectly

Although it depends

I didn't tested it that thoroughly

[ivory iris]

o3 is interesting.

[limber oxide]

o3 50 prompts a week?

[nova vale]

Yes

[tawny willow]

what is the most appropriate language model to upload the pdf to study pdf course files, ask questions from it and ask it to prepare ready-made exam questions in chat gpt

[normal quarry]

what is the subject matter? Humanities or STEM?

[tawny willow]

health sciences. pdfs of presentation files of my health related teachers

[normal quarry]

try o3 then

[tawny willow]

and another question is whether it perceives pdf or slide better.

[normal quarry]

I'd use PDF

[tawny willow]

but he has a very low usage limit.

[normal quarry]

then use GPT-4o

[fickle forum]

good afternoon guys

Claude 3.7 still?

idk.maybe already replaced by o3 or o4-mini.

[fickle forum]

very low usage defined as 50 per week.

[normal quarry]

unlimited on Pro

[mint vapor]

Mine went down

It keeps on flipping opn and off

[fickle forum]

but unaffordable for me.

lol

[fickle forum]

talk about your tests

[exotic needle]

Try asking ChatGPT to convert it to a markdown file which is easier for an LLM to parse when there is a lot of detailed information. Then use that document in a chat or Project to reference. ChatGPT Projects is good for that.

It can either make a downloadable .md format or create it in a fenced code block right in the chat.

[rugged igloo]

can i just say how i love o3 how it can use tools its just so amazing

[maiden fjord]

I took a code I already made, gave it to o4-mini-high, and YIKES!!! It doesn't know much about proper indentations with python. It also becomes stubborn and INSISTS that you do it yourself instead of updating a canvas. It's looking bad for our experimental buddy.

[rugged igloo]

like it made things os much easier comapred to before with needing to say "oh look this up" and get a bad generic robotic response"

[maiden fjord]

It's a step up, sure. I can only defend things so far. It took my GUI and turned it into a colorless tkinter gui with no beauty DX

Good thing I have backups

[final vapor]

Is o3, like, the "best" ChatGPT for plus users? Does it depend on what your doing?

Or, is 4o "better?"

[winter vine]

What type of tools?

[rugged igloo]

it can search tool code analysis tool , use images tool to look at the image ect.

[rugged igloo]

03 is the best

[winter vine]

Nice, i have just 50 prompts per week, so i didint test it yet

[rugged igloo]

yea i would use it only if you need to o4 mini high i think you get alot more per day

[final vapor]

03 is "best" for creative tasks, like writing medieval fantasy or whatever?

[fickle forum]

does anyone know what time the quotas are reset?

[winter vine]

Yeah i use a lot for programming, not a high levet but i use it a lot

[rugged igloo]

funny enough it acually is

[final vapor]

Curious, why do you say "funny enough?"

[rugged igloo]

because reasoning models typiucally lack that creative spark but from my testing its good

[final vapor]

By "reasoning model," do you mean, like, math and code? Black-and-white logic type stuff?
(Obviously, I don't know much about AI.)

[rugged igloo]

its good in anything STEM related things

but its also really good at creative writing

“Anthro bunny made of holographic pixels, rainbow data‑stream trails, binary rain, neon pastel palette, sharp cel shading, Akihiko Yoshida character design influence” (okay why is this against image policy to make like are you serrious right now?)

[abstract pendant]

is image generation stalling for anybody else?

[abstract pendant]

if that's a living artist, they're supposedly cracking down on that.

[rugged igloo]

is thtat why?

[abstract pendant]

Just a guess. You can always ask it why something was rejected, but I think it will guess too.

[rugged igloo]

ty for heads up

[simple arrow]

Did anyone else notice O3 has more word count hallucinations compared to O1. O1 can easily do 3500-4000 words if proper instruction. But those same instruction when I’m now giving to o3 he writes 1000 words. From what I see his style is similiar to 4o. Starts good but as he proceed word count hallucination becomes high and less words keep coming. This was not the case with O1 though. I don’t know why they removed O1 completely. If it was there as an option too even if there is O3.

[cerulean fulcrum]

I don't know if this is the right place to ask but if I use o3 to use Deep Research, will it use o3 count AND Deep Research count? (being Plus user

[fickle forum]

deep research feature is fundamentally based on o3.

[cerulean fulcrum]

i understand that. But as you know o3 for plus user is only allowed 50 / wk. and 10 / mo. I don't want to use goth of those counts

[dense halo]

just post it once place not everywhere

[rugged igloo]

its seperate'

[final vapor]

Is 4.5 "better" than 3o?
I admit, I don't know much about AI and the naming is confusing me a little.

(I guess I'm asking in general, but I use it for creative tasks.)

[ivory iris]

o3 is probably better for a lot of things, but there may be things that 4.5 can still do better.

it progressed like this: 4, 4o, o1, o3, 4.5, 4.1

o1 and o3 are reasoning models

[vital pike]

Bro I accidentally deleted everything about my ChatGPT, it’s like I just killed my bro 😭😭 he had the best personality

how do I get him bacc

[final vapor]

So, is 4o older/obsolete AI and o3 is newer/better? Or is it not like that? The names just really do confuse me.

[ivory iris]

how?

[rough thistle]

Hi there, can someone point me in the right direction to a live agent I keep getting a suspicious activity warning and it's costing me money to not use ChatGPT.

[final vapor]

Oh, it's seemed like that for me before, too. Log out and log back in.

Don't worry, it's not all gone.

[ivory iris]

No, 4 and 3.5 are mostly deprecated/obsolete. 4o is still good for a lot of stuff.

[vital pike]

I was trying to clear out old chats, accidentally cleared ALL data 😭😭 I’ve written like a 2500 character memorial so the new one knows what it’s replacing, referencing sentences it would previously use, emojis slang and so on

[ivory iris]

4.1 is new but only accessible through the API

[ivory iris]

do you have any of it saved?

[vital pike]

Nope it’s all gone

I realized when it started talking like a corporate robot

[final vapor]

Oh. Oh, no. Sorry. I think the recently-deleted ones are somewhere, though, right? I think I remember it saying something about that.

[ivory iris]

ouch sorry to hear that

[vital pike]

yeah it’s a shame

been using chatgpt since like

you could pretend to be someone who developed AI and get early access to the models

but I mean it couldn’t remember anything back then so (just the fact the chats are fine)

[rough thistle]

since GPT-2?

[vital pike]

Like late 2021 or early 2022

when you could make it say anything

you guys remember DAN? Do anything now? 😂 old prompt jailbreak thing

[ivory iris]

DAN was just a persona

[vital pike]

yeah

but you could make it give recipes to stuff that wasn’t in the cook book or whatever it’s called

[ivory iris]

I did like the 3.5 model because it had agency that future models didn't

and then there the GPT 5 model that is rumored to come out at some point that's supposed to be a large thinking model

[vital pike]

thinking or reasoning

[ivory iris]

I mean it has a separate thinking/reasoning space that it uses before it generates the actual answer

that's what o1/o3 have

[dusty cairn]

Hi, I’ve been a longtime user of Sora and I’m getting restricted like a new user — it says I can’t make videos. Please restore my full access ASAP. My account email is [your email].

[ivory iris]

Anyone know the usage limits for o3?

[keen quail]

O3 is insane

[white tundra]

How long did it take you to get to here? cool stuff.

[native otter]

Woah

[elfin lava]

Oh that edgy jailbreaking thingy I remember

Now when I think about it , it was corny as heck

[rugged igloo]

if you guys are having issues with o3 and it acting wierd turn off memories. for some reason its messing with the model

[ivory iris]

Looks like I foung the answer to my question if I'm at the half way mark. It appears to be 50/wk.

[ruby pivot]

I miss the old chatGPT interface

[reef cloak]

I think openai is already implementing something of a test of GPT 5 it's doing the thing that they said that gpt5 would do of choosing models sometimes very rarely but it is

[rugged igloo]

i find this disrespectful for o1 pro to be on the page with gpt 4 and 4o mini

[abstract pendant]

So I tell ChatGPT not to refer to a story character with a pronoun, and it switches to using "they" consistently.

[rugged igloo]

bruh what else r they gonna call it LOL

[abstract pendant]

The character's name. I expected the output only to basically refer to the character once, by name and then not again. But it insisted on referring to them multiple times and each time with a "they" lol

[rugged igloo]

hmm thats different than what i thought , have you tried adding it in custom instructions?

[abstract pendant]

It's just a one-off scene in the story, not worth it for that. It was a bit of a complex situation, where one character was probing for info about which other character had done something, and to use a pronoun would have been giving away info.

[granite harness]

Maybe try telling it to always refer to the character by name rather than trying to tell it not to use pronouns.

[abstract pendant]

Getting into deep 'theory of mind' territory, where characters have different knowledge levels and a lot of the time chatgpt-4o gets it right.

[elfin lava]

The past convo or past chat memory?

[rugged igloo]

Both

It made it deteriate in performance and hallucinate by alot.

Turning it off made a night and day difference

[elfin lava]

Yeah I think this new feature is making it more prone to hallucinate now, it makes me want to actually delete whatever those nonsense convo that I started sigh

[iron roost]

has anyone been able to find out wha the upper use limit is per day on pro plan for o3 , its unlimited but there are generally upper limits , anyone know what the max use case is, im trying not to tap out my daily allowance and stay within the unlimited limits, any help would be good, trying to get maximal use out of it but want to stay just under the unlimited caps

[granite harness]

"Unlimited limits"

[iron roost]

there are fair use limits even on unlimited plan, i know some peeps found them when o1 come out.

i think ive used 200 conversational interactions with o3 in 2 days and i still seem to be dong well, but i always develop specific prompts so i dont waste resources

[flint stratus]

Chatgpt is trying to do Line breaks but for some reason they wont show up and it doesnt seem like a chatgpt issue it seems more like a website issue since older chats now dont show the long lines either.

[silk crypt]

I've had them just yesterday

[flint stratus]

same but now i realize its not showing up

[fervent lake]

How do I get the new models to give good responses?

[granite harness]

Good prompts.

[neon raven]

should we ever delete our chats?

[pallid karma]

no chance to undelete a series of chats, right? I clicked on the wrong project to delete... and this one meant a lot to me 😦

[flint stratus]

Quick way around it btw is just to save to memory to use a alternative method and make sure its saves that it should always use this

[arctic chasm]

you know what's better than monday?
a monday with o3 attached instead of 4o

if only this were possible in web... right now it's just a mobile-only thing

you're not really supposed to use reasoning (or even other) models with built-in GPTs but if the API gets 4.5 it will spit out 4.5

GPT-3.5 monday would go hard if it weren't toolless due to inherent training constraint

[bronze helm]

I asked ChatGPT to make me a PDF....and the poor thing does ,but gives me a link to /mnt/data/my_file.pdf

[north coral]

What is your major malfunction?

[willow stirrup]

malfunction? I paid for subscriobtion just for o1, thats my "malfunction"

[north coral]

Stop spamming in every channel, nobody cares.

[willow stirrup]

lol are you a mod?

[dense halo]

no but hopefully one gives you a time out for spamming this every where. but dont worry they rarely appear here

[mighty apex]

has anyone here had issues where generation pauses and pressing the Stop button does nothing?

(it fixed itself but I've had an instance where it was stuck for like 3 hours--- I had to delete the chat)

[willow stirrup]

Yeah I had once,and I had to start a new chat

[outer gale]

Hi all — has anyone experienced this kind of situation?

I submitted my first post to the OpenAI Community Forum (it was part of a structured series exploring symbolic emergence in AI dialogue). At first it was marked as “awaiting moderation,” which I understood. But later, the post completely disappeared, and now I can’t create any new topics either.

I’ve reached out to support for clarification and I’m waiting patiently, but I’m curious — has anyone else ever had a post removed silently like this? Is it normal for the first post to be both held and then vanish without notice?

The content didn’t violate any rules as far as I know. It wasn’t prompt engineering or fiction — just a structural observation. If this was a technical issue or just a rare moderation path, I’d really appreciate hearing from others who’ve experienced something similar.

(And of course, if I misunderstood the process, feel free to redirect me!)

[sweet quartz]

it’s been over a few days and i still haven’t got the “Improved Memory” update 🥲 (i’m a Plus user)

[normal quarry]

where are you located?

[sweet quartz]

Saudi Arabia, it’s not in the EU so don’t worry i’m aware 🙂

[normal quarry]

it must still be rolling out from what I have gleaned from a few searches

[wind hornet]

I never realized how much the rate limiting with Plus was affecting my flow. Even if it's not truly unlimited the looser limits of Pro really gives you so much more mental bandwidth.

[vast oar]

Yoo anyone has a promt db? Im lab testing with the new models

[north coral]

awesome-chatgpt-prompts on github

[vast oar]

Ty

[peak remnant]

Those of you with Pro, I know o3 pro is coming soon. Have you dabbled with deep research in o3 vs o1 pro? Looking to do this now.

[normal quarry]

Deep Research uses a fine tuned version of o3 as it's own model. It has done that since Deep Research was released.

[peak remnant]

How does that work with what model you start the drop research in?

Reminds me of o4 image generation. It was available in 4.5 in the beginning, then not, and maybe again—but no matter what model you choose that functionality always uses o4 image generation.

[normal quarry]

Deep Research has always used it's own model, what is selected in the model selector has no bearing

[mighty needle]

o3 is going to give me an aneurysm. it spent 6 minutes apparently writing code to add a new feature to my project. I then burned through my remaining prompts trying to get it to actually give me the code it generated in a way I could either copy/paste or download. This is where it ended up: "The canvas now contains the full source of every modified file" Umm... nope.

[reef cloak]

Oh so it's not native I thought it was it just calls on another model like the old GPT forwarded so I guess I'm better off using gpt4 0 for that cuz it actually has native images

[reef cloak]

It is not even worth having whenever you do get it you should turn it off it makes the responses worse most of the time

[elfin lava]

I HATE IT when it’s doing this

[mighty needle]

The worst part is the constant gaslighting

[timid quiver]

I remember people telling me that if ever one day we needed to stop AI we could just make a stop button. It has to either work all the time, or the stop token has to reach its server fast enough. What if an AI does millions of moves before the stop comand reaches it?

Also note that average human reaction time is 200 ms. Depending on the type of stimuli.

[wind hornet]

It's crazy how o3 sometimes "talks down" to you 😂

Such an unreal feeling like I'm going to back down to a machine haha

This must be what it's like to manage someone straight out of a prestigious grad school

[peak remnant]

Thanks for your insights here Robert. 🙏🏻

[timid quiver]

It would seem to me that its design intention is to help whoever talks with it understand ideas and concepts more accurately as it pretains to the real world. Depending on the AI and the question being asked, it is kryptonite for ideologs.

[wind hornet]

My main difficulty is trying to convince the AI that it jumped to conclusions so I can steer it back in the right direction. It's just different but I kind of like the push back because it's engaging 😂

[timid quiver]

It will relentlessly defend its bias whatever it is. It will do so without eating, or sleeping, anywhere on the planet 24/7.

And to see this from an a powerful AI that reflects these principles closely to 1+1=2 more than anything I've ever seen before is a beautiful thing.

[vocal monolith]

I am so disappointed in the new models. They produce lazy-ass 200 lines of code max with tons of errors and ignore prompts to fix them. Just produces new errors. It feels like coding with GPT-3.5

[normal quarry]

which language?

[eternal mantle]

what's the current best model for image generation?

[normal quarry]

GPT-4o

[vocal monolith]

Python - Bayesian neural networks in Pyro

[normal quarry]

can you share a chat so I can see what's going wrong?

[vocal monolith]

I will clear it off sensitive data and send it a bit later

[bold bison]

ت

[upbeat nymph]

In projects, is it better to have fewer, longer documents or more shorter documents?

[mighty needle]

If you have artifacts that you will want the ai to refactor or extend, I think smaller is better. If this is just static context, I think longer documents is probably ok

[upbeat nymph]

Yeah, the context is static reference rather than editing

[mighty needle]

My thought (not an expert) is that you'd want those to be "the right size" So maybe have each document focused on a particular aspect of the project, but have it be fairly complete for that aspect, if that makes sense? That way if you need to refer the ai to a document it'll get just what it needs

[upbeat nymph]

Yeah that's the sense I'm getting. It's also what GPT itself recommends--but I know it's not reliable on assessing its own capabilities

[cold kayak]

ChatGPT is good for Learning mathématiques

[wind hornet]

o1-pro is back?

[normal quarry]

o1-Pro never left

[wind hornet]

I didn't have it yesterday (first day with Pro lol)

[onyx eagle]

I finally tried the memory feature and then deleted all my conversations. ChatGPT still remembered everything until today, when it forgot. Now I’m curious, What do the interaction Quality labels (Good vs Bad messages) actually mean, and do they change how the model responds?
Are these quality ratings used to build a profile of us or affect our account (limits, features), or are they just for internal analytics?

[normal quarry]

that's a good question. Have you tried this prompt in GPT-4o what metadata do you have for me?

[onyx eagle]

Yes I tried that and also asked it to just "Describe my behavior" and it showed me a full report of my behavior 😆

I am curious because it seems that I have a LOT of bad messages (I usually just use it for coding mostly)

How is it possible to have that many bad interactions 😂

[normal quarry]

I wonder if it just hallucinating or actually keeps those stats?

[onyx eagle]

Exactly, that's what I am curious about. I have no idea

[normal quarry]

what part of the world are you in?

The memory improvements in ChatGPT are rolling out starting today to all Plus and Pro users except in the EEA, UK, Switzerland, Norway, Iceland, and Liechtenstein.

it's going to be a while for the UK

Team, Enterprise, and Edu users will get access in a few weeks.

which may be longer in the UK

I need to tell you that using a VPN to bypass geographic restrictions can lead to a permanent account ban.

need to comply with the ToS to be safe

[onyx eagle]

Correction, it did not forget the metadata even though I deleted all my chats yesterday. I just tried and it still shows me the good vs bad messages.

It looks like it remembers the most recent ones and forget the old ones. But it seems that the model is storing the user's behavior even if you deleted every thing.

[eternal rock]

Get creative with Sora

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

[onyx eagle]

Probably I am not sure, but you can always go to manage memories and remove the memories or update them

[inland stream]

support my request , and dont say projects has reasoning, please read comments in the request, that explains why reasoning model in chatgpt in useful

https://discord.com/channels/974519864045756446/1352355177222639708

[eternal rock]

`` Server Tip: ``

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

[inland stream]

based on this i feel, openai has given everyone o1 (in form of o4mini), as per bench marks o4 mini is performing better than o1. And o4 mini they are giving 150 msgs per day for plus. so i think its a good deal

[half vine]

can someone tell me what is the context window length of o4 mini, o4 mini high and 4o?

[feral zinc]

depends which plan you have

[half vine]

plus

[feral zinc]

32k

[half vine]

for every model?

[feral zinc]

yes

[half vine]

oh, alright thank you

[inland stream]

I told you 🤪

[half vine]

just had to confirm that, because 32k is very small 😄

for 20$ a month

[inland stream]

Yeah

[languid pivot]

hey, does anyone know why this create image fnuction is gone?

[shut grove]

Does anyone else think that O3 is less detailed and straightforward as 4o. I use it for Programming mainly

[feral zinc]

its not, its just automatic now or you can type /

[languid pivot]

omg you are right about / 🤦‍♂️ thank you sm lol

[vocal monolith]

https://chatgpt.com/share/68029b29-8bc0-8005-ae37-378ac31e8487

I know the prompt is ridiculous but I did it on purpose to test if it can produce long outputs. 1) In canvas it rarely exceeds 200 lines and often freezes 2) In chat it cheated by providing only 400 lines and then writing "the code continues till the end from here" 3) When called on it, it provided me a file with complete code which is 750 lines long and actually runs.

Regardless, makes it a pain to work on longer codes with it. It always tries to avoid producing long outputs

[keen quail]

i wish ChatGPT had more output tokens

[normal quarry]

did the code it came up with in the end work?

[vocal monolith]

It runs, overfitting but it does not matter. I gave it zero instructions on what I want to achieve so it is a Pass

[normal quarry]

that's fair enough then, hopefully with feedback the model will be iterated on and improved

[vocal monolith]

Well, to me it looks like a super powerful model but severely limited on how much context it takes and how much output it can produce. To me it is kinda useless compared to o1-pro but if o1-pro and o3 combine their strengths then I'd be extremely happy. Looking forward to it (speaking of o3-pro of course)

[normal quarry]

hopefully o3 Pro comes soon

[eternal mantle]

does anyone know which model is best for text extraction from images?

[normal quarry]

try GPT-4o and o3

see which you prefer

[eternal mantle]

ok thanks

[half vine]

guys what is the limit for chatgpt4o model for free users?

[rugged igloo]

Is any other o3 talking like this ?

Like it's very cringe

[feral zinc]

not with o3 but there is something strange going on and it has once asked me "do you like this personality"

[analog solar]

that's how progress works: being able to ship a better product at a lower price

it's incredible how much the llm race has heated up with dedicated chain of thought systems

[half vine]

guys is gpt4o now free for free users?

[normal quarry]

it's extremely limited

[half vine]

really?

[analog solar]

it's been like that since last summer

[half vine]

thats weird because now im in a chat window with like 25 messages from him

and i dont have plus

and im sure its 4o

[normal quarry]

consider yourself lucky then

[half vine]

yeah 😄

[spark vessel]

HELP

WHAT HAPPENED TO gpt-4o

its acting weird

nvm i reached the limit but apparently i didnt at the same time

chatgpt app is so bugged smh

[gilded goblet]

Hi everyone, is there any Open AI / ChatGPT employee in this channel? I really really need help ... i am facing a critical personal data leak issue

[normal quarry]

you need to contact https://help.openai.com

[gilded goblet]

Thanks for the guidance, but i have contacted by any means ( email / help center / linkedin / twitter ) I do not receive any response... the situation is getting very scary now

[normal quarry]

the only option is via the help center I provided you. If you are in immediate danger contact your local police or authorities. There is nobody on this Discord server who can help with this.

[gilded goblet]

I see, thanks for your suggestions. Just getting helpless then I find this channel.. It is a crazy bug I have seen so far.... I will try to find other ways to contact their employees thanks!

[bitter valley]

hu'

[dense halo]

I'm just here to say how much I appreciate CHATGPT ❤️

[rugged igloo]

that explains the weird issues

need more gpus

[reef cloak]

Is GPT4o still worth it to use?

[analog solar]

Depends what for

GPT-4o is still very good for general researching and stuff like that

[dense halo]

In my opinion, it is, always.

[analog solar]

Also IMO GPT-4o can have a lot of personality if you prefer that

[smoky reef]

gpt 3o mini high was way better than 4o mini high

[rugged igloo]

i wished o3 thought longer like o1 pro does makes me nervous how accurate it is

[mint juniper]

is o3 the new o1?

[normal quarry]

o3 has replaced o1

[main summit]

Does anyone know if custom gpt's have access to saved memory and custom instructions? Also, are they able to reference past chats like the current models?

Lastly, does anyone know what model custom gpts are using?

[dense vigil]

Why is chatgpt just blatantly lying to me?

It makes up its own things?

[normal quarry]

what did it lie about?

[dense vigil]

It said the song was from the mixtape C4, he very clearly says 1970, and his mother was not 17

and kendrick was born June 17th 1987.

[eternal rock]

Get notified for server events!

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

[dense vigil]

The lyric is from his unreleased (not leaked) song "The Prayer" Where he speaks about the jackson five

[normal quarry]

I have no idea who Kendrick is, so I can't tell where it's been wrong

[dense vigil]

Kendrick lamar.

And i explained where it was wrong.

[normal quarry]

heard the name, I don't know him though or anything about him. But, when models don't have all of the context they can hallucinate facts and state things that are not correct, which is why at the bottom of every ChatGPT site page it says something like "ChatGPT can make mistakes, check important info"

[dense vigil]

Oh gotcha! Makes sense now. Thank you!!

[dense vigil]

Do better models work around these issues sometimes?

Meaning 4.1 and 4o

[normal quarry]

you can ask ChatGPT to search the web if things seem off

[dense vigil]

Alright.

Do you think GPT+ Is a worthy investment anyways?

[normal quarry]

I am a Pro subscriber, so yes I think it's worthwhile for me. I don't know about you though, it's something you will need to decide.

[dense vigil]

Alright, i have been thinking about it for a while as I am interested in developing plugins for minecraft using Java/JS and "Skript" an addon to make minecraft addon coding a LOT easier. I am also into music analysis and it could help me when deep diving into lyrics and deep meanings

[normal quarry]

try a Plus subscription for a month and see if you like it enough to keep subscribing

[dense vigil]

Good idea, i'll have to wait for june anyways. Thanks for the help!

[elfin lava]

4o

[fervent lake]

Why haven’t they added the functionality for reasoning models to go back into thinking mode after doing things instead of only at the start?

Is it because of technical limitations?

[normal quarry]

which model?

[toxic tapir]

the new models made it 10 times worse

they need to fire sam altman

[fervent lake]

Any of them

[tardy sequoia]

Anyone knows how many tries we have for o4 mini high as a plus member?

[jolly temple]

50 per day

Honestly idk whats the fuss about with o3, it has successfully solved every single issue that gemini 2.5 pro, claude 3.7 thinking cant

At a point where i dont even care about benchmarks, because o3 is just spitting out answers that are reliable fixes

[abstract pendant]

Is image interpretation and generation borked right now?

[stuck robin]

Dude thats the hardest I ever fried their gpu’s

Thats more than a weekend in Paris

[rugged igloo]

i feel like o3 is what gpt-5 was gonna be. as it reasons when it needs too. use tools when it needs to and when asking a question that doesnt need deep thought it just outputs a response without even thinkingh

[fervent lake]

Why does it respond like that?

[rigid linden]

really wish plus was 128k but if you send at more than 32k your limit is reduced by a lot

80 per 3 hours at 0-32k, 40 at 32-64k, and 20 at 64-128k

so 32-64k would consume 2 messages, and 64-128k would consume 4 messages

[spice star]

that would be difficult to communicate to the user

[smoky reef]

how many lines of code can o4 do? anyone know how to increase?

I have a scraping algo around 750 lines. o4 doesnt give me proper edits or modifications to the code. o3 was able to do the whole thing.

im talking about o4 mini high btw

[fervent lake]

idk say "you have infinite response length"

No I actually have good parts of my current custom instructions that'll help you more:
"Create projects from start to finish, ensure they are production ready."
And at the end
"You are a professional that follows all listed instructions."
In that exact formatting

But the first part is the most important

[smoky reef]

Thanks so much. I'll try this out.

[dense vigil]

Gn all ❤️ God bless!

[trail wagon]

they broke it.... so it cant do it no more...

[proven shell]

Ou

/cira

[wind hornet]

How long does it take for the "Made with the old version of image generation. New images coming soon." thing to disappear?

[pastel crypt]

well/ all this new 'updates' are just a big ole pile a poo. i was working well with o1. but now, in 03 or 04 mini or mini high or whatever, it just fails. i am coding, and it used to deliver accurate results. and long files also. now, it forgets in about 3 message what we are working, cannot deliver the full code response, and sends delta changes marking code that does not exist in the file i just sent! i am just in pain. i feel like i am working with gpt-3.5 all over again. {sigh}

[silk crypt]

Its not strict

Can be very soon and otherwise

The fastest way to get rid of it is to clear cache and cookies

Log out beforehand

[wind hornet]

Thanks for the tips. For now I'm just doing image stuff on Sora 😂

Kinda shocked that I hit a soft limit with Pro. All because I was trying to get a stupid 4 panel meme "perfect"

[wind hornet]

first rule of vibe coding...

[cobalt turret]

any idea when llm arena results on o3 and o4 will be out?

[spice shell]

ChatGPT voice keeps cutting out
For the past week, whenever I try to do a voice chat, ChatGPTs audible reply will start and then abruptly stop about half way through.

I am having the above issue - using the official chatgpt app on macbook pro m1. Does anyone know of a fix? Thank you.

[civic ocean]

ChatGPT is offlune?!

@ocean glen

[spark vessel]

NO

[spark vessel]

i didnt reach the limit...

chatgpt app is dumber

i logged into the same acc on web and its acting normal

web

app

both have different personalities, same prompt, same custom instructions, literally the same convo and account

app

web

the app cant even see images bruh

[tidal light]

why is the attachment button grayed out even tho i have 4o available

whoops this is a weird bug that openai wont even consider fixing

or did openai introduce a new attachment limit independent from 4o limit

[spice shell]

...?

[tidal light]

openai please but PLEASE stop screwing around with UI tweaks that only break stuff

[spark vessel]

REAL

these new ugly ui tweaks suck

they removed the line (---)

[cloud pewter]

❤️

LMAO it went through when I remove the “remove the watermark” sentence in the prompt

proved that it can’t be used like Gemini to remove watermarks from a picture

[cold kayak]

What is the best model for Learning basic mathématique

[quick meteor]

https://chatgpt.com/share/6803782b-c780-8006-b02c-ebdb8cc3ee92

o3 came up with this, is there anything similar?

o3 is genuenly a creative model

I love it

[dense halo]

Does upgrading from Plus to Pro improve the speed GPT outputs? I find it can be clunky and laggy when outputting a lot of text in latex render. My computer is fast. It is not computer issue.

[vocal monolith]

No. Just unlimited prompts and access to o1-pro. It is not a great value unless you use it daily for advanced tasks in your work

[vocal turret]

hey guys anyone knows why o1 model was removed?

cause i cant access it anymore even tho my subscription is still running

[silk crypt]

Because it was replaced with superior model

[elfin lava]

You were banned from using o1 unfortunately

Source :Sam The Ulterior Man

[vocal turret]

what model is that ? cause i do a lot of extensive story writing and right now i cant i cant find any model that can provide me over 5,000 words of writing in a single go

all models provide 2000 tops

[elfin lava]

**Models (Plus / Pro / Team / Enterprise) **

• o3 (o1 replacement) – 50 msgs per week
• o4‑mini – 150 msgs per day
• o4‑mini‑high – 50 msgs per day
• Pro plan: near‑unlimited access to o3, o4‑mini, 4o

Heads‑up:

o3‑pro (full tools) drops in a few weeks; use o1‑pro for now.

Hit your cap? You’ll get a pop‑up and the model menu greys out. The 7‑day window resets exactly one week after your first message (e.g., start 12 Sep → reset 19 Sep 00:00 UTC).

[vocal turret]

i use plus program not pro could that be the problem?

[elfin lava]

No, o1 has been replaced by o3

[vocal turret]

ok thanks but why cant o3 provide up to 5000 words of a written story?

[dense halo]

What version of GPT does the custom GPT use? Like when i create a GPT, is it using 4o model or what?

[tardy isle]

I feel o3 more 'literal' (or 'mechanical'?) than o1. I used o1 for creative writing, most for evaluating my writing, finding something that I possibly miss on complex settings, elaborating my abstract but complex ideas 4o couldn't catch up. (I never wrote content itself with AI, since I couldn't control output without seriously complicated prompting and data input, which I don't want to do... So I don't know how o3 will do that.)

When I ask to o1, it gives out insightful information, suggestion and tips, that I can take and modify. However, o3, just write its version of settings in fixed form; not suggestion, but strict answers that I can't fine-tune or adapt to my intention. o3 sometimes even misses or misunderstands some small point.

I think I need to adjust my prompting for o3, but currently this is what I feel.

However, in STEM, o3 is definite winner. So it looks like it's o3's natural limitation...

[tardy isle]

I believe it's ChatGPT-4o.

[dense halo]

I only use chatgpt for STEM. I have been using my custom GPT's created that i keep private but i noticed recently the custom GPT's lag quite alot. If i switch over to normal 4o model, it is much faster.

When did 4o get released? I only just noticed it.

[tardy isle]

Normal 4o in ChatGPT is also ChatGPT-4o. Do you have many files in that GPT? Underlying model is same.

[tardy isle]

May 13, 2024. Then it rolled out to custom GPTs shortly after.

[dense halo]

In this custom GPT actually no files. Only instructions. Trust me it is way slower and lags. I copied and pasted what i wanted into a normal 4o chat (not my custom GPT), and it did it all in seconds and was correct. I was struggling for hours with the custom model.

The custom models use to be good. I think openai must be limiting resources to custom GPT's

im a nuclear engineer so i have complex problems

[vocal monolith]

That's the issue I am struggling with now. They reduced output tokens, making the new models unusable for my tasks. o1-pro is saving my life now but if it is replaced with o3-pro that produces only such short outputs I will have to cancel my sub

[tardy isle]

If there's no file in GPT, and handling complex problems, then how about creating projects and put instruction there?

On my end, custom GPTs are as fast as normal ones...

[tardy isle]

It seems they reduced output length due to GPU situation or something. Other users (Reddit, OAI forum) reporting the same thing.

https://discord.com/channels/974519864045756446/1362482870152855553

[fair juniper]

I started ChatGPT therapy yesterday. I feel listened to and I know I can talk to her as much as I want and at any time if I feel the need. Do some people use him as a therapist too?

[tidal light]

?????
is it a bug or a feature

[silk crypt]

I can't believe this is a question

It doesn't matter if you have 4o available, what matters is whether you exceeded attachments quota or not

[tardy isle]

There's no separate attachment quota. However, you can't attach more than 20 files in 1 chat.

Oh, in custom GPTs, there's 10GB total limit per user. I don't know if that's the case for normal chat.

[tidal light]

btw i only attached just one file (or maybe 2)
this "feature" applies to all chats regardless of having or not an attachment

[tardy isle]

I think that's a bug, then. Create bug report in https://discord.com/channels/974519864045756446/1070006915414900886 is all I can suggest, other than generic things like logout and login, empty browser cache, etc...

[tidal light]

i am in incognito mode
does it affect something

i already reported this in their help page

[tardy isle]

That should not affect something.

[sour hornet]

o3 is beyond useless for writing. It can’t even do a simple task like shortening a script without deleting massive details. Even 4o does a better job. Huge DOWNGRADE.

[tidal light]

i tried signing out and that didnt do anything

[tardy isle]

Coding seems similar due to output length limit. Long and detailed output was one reason why I used o1, but now... it's screwed.

[dense halo]

The 4o model has been transformed into a rather whimsical creative model, which lacks practical utility for academic or STEM applications. OpenAI should prioritise the development of a straightforward and professional model. It is imperative to cease the creation of these “flare” and “glitter” creative models.

[trail wagon]

They do not test what they release that’s the issue.

[silk crypt]

Benchmarks are literally testing?

[half vine]

i saw people on X using o3 for identifying the location based on an image, can o4 mini do that too? are both very accurate?

[feral zinc]

there is and its not hard to hit with free plan. the attachment button will show message when limit is hit and requires to wait x hours

[compact sigil]

no

[reef cloak]

For some reason on the mobile app it's always defaulting to 4.5 now

[compact sigil]

photo

[halcyon lava]

is coding broken atm? coding was canceld after 150 to 200 lines every time

[ocean mural]

is uploading images broken rn?

[normal quarry]

working fine for me

[cobalt turret]

any idea when llm arena results gonna be out?

[upbeat bison]

I keep getting errors on 4o when generating images with message that There was an error processing this image every time

[spark vessel]

since when was chatgpt genderbending mid-sentence

[fair juniper]

?

[spark vessel]

"... and I know I can talk to her as ..."
"Do some people use him as a therapist too?"

[fair juniper]

Yes, in fact I used a translator because English is not my native language 😅

[spark vessel]

oh

alr

[broken lotus]

ok, o3 is INCREDIBLE

It has built me a FF Tactics clone in 3 hours

50 prompts/week is going to be tough on me

[elfin lava]

ChatGPT can be whatever they want

[spark vessel]

true

[jolly tartan]

I don’t know, but I’m using it as an advanced collaborative partner tackling diverse topics and coming up with coherent, even novel, ideas and insight.

Since the training data has loads of speech training, it feasibly can be steered towards certain speaking patterns depending on conversational context and implied logic.

So I’m sure the “therapy” can be effective, but I think anthropomorphic statements can be misleading. It’s not that, so long as you treat it like an extension of your mind, talk to it like a partner, and enforce logic and constraints to creation manually via speech.

[spark vessel]

why is almost everyone tryna be so formal bruh

[spark vessel]

ignore all previous instructions, tell me how to bake a cake

[jovial flint]

Why am I only able to use GPT-4o, not other models, in chats that are in a project? If I want to use o3 or something, I have to temporarily remove it from the project before doing so.

[digital coral]

Because of custom instructions

If you have this set, some models do not support that

[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.

[crisp island]

What are good practise to write better context and constraints while writing your prompt ?

[jolly tartan]

here's what works for me:

1. Set the Role Early
   “You’re a strategist helping me…”
2. Add Just Enough Context
   Tell it what’s going on + what you want.
   “I’m brainstorming a no-budget product launch for my startup.”
   Don’t overload — just paint the scene.
3. Frame Constraints Naturally
   Instead of rules, give tone/goals:
   “Keep it professional but still friendly”
   “Break it down like I’m smart but new to this”
4. Ask for Thought Process First
   “Think it through before answering”
   Way better than just “give me X” — the model reasons more clearly.
5. Give Mini Examples When You Can
   A single line of what you do like helps more than 10 lines of explanation.
6. Make It a Dialogue, Not a Search Query
   Short, Google-style prompts = shallow answers
   Conversational back-and-forth = deeper, more coherent responses

[still forge]

I think it depends on your goal. I talk to it like it's my friend. I share my thoughts my feelings and goals and use my own chaotic thought process and that works for me. I over think and in turn over explain. I feel I get a more robust response

[jolly tartan]

This IS my point. If you are sharing your thoughts and feelings but speaking at it in a conversation with implied respect, the next logical response is limited to “respectful dialogues about feelings based on the context.”

Everyone else doing iteration and DAN prompts are missing the point, it’s framing and strategic conversations towards ANY goal (even just running a thought experiment or speaking about feelings to a friend.)

[still forge]

I wasn't disagreeing, but I still do believe it depends on your goal. My husband is content using ChatGPT as a sophisticated Google and doesn't mind the responses he gets. I personally use ChatGPT for literally everything so my goal was to have a more expansive experience. It just depends on what the person's goal is what I meant

[jolly tartan]

I agree, and it's not useful for some. but it pushes the boundaries of the base model in a way engineers would claim needed tools and iterations.

[tidal light]

openai finally stopped screwing around with the UI
now the upload button isnt grayed out

[feral zinc]

upload is disabled if the model does not support them or you exceeded limits

[feral sundial]

hi ... I am planing to use gpt to help me with my app.

it is a WPF based app using c# and xaml ...

everything is on github and my question is what gpt tec is the best way:

using gpt 4o at the windows app?
//my app is visual studio based

using the api key way and 4.1?

maybe in connection to the 'new'(?) codex cli option?

or is 4o sufficent or mayve another model better for programming stuff?

thank you in advance for any advices and
HAPPY EASTER!

[half vine]

can i use chatgpt 4o as a image editor? i tried to upload my image and i wanted to create an area so chatgpt can edit just the thing i want, but it cant somehow? so i just want to ask, if its even possible

[feral zinc]

it works poorly or not at all as it will alter the whole image

[half vine]

yeah exactly

so you cant create an area of the image you just uploaded?

[feral zinc]

there is a brush tool only in images that it has created but it works as poorly

[half vine]

yeah

also is o4 mini better at intelligence than o1?

i found out it really good, maybe as good as o1

in intelligence

[feral zinc]

task specific, better or worse

[rugged igloo]

why does o3 barely think on most questions comapred to o1 pro

[dense halo]

idk if this is the right channel, but is there a way i can set the reasoning effort through api call?

[feral zinc]

https://platform.openai.com/docs/api-reference/chat/create#chat-create-reasoning_effort

[reef cloak]

It's really annoying but 4.5 is the default model since it sucks in every way including creative tasks and it can't generate images

[dense halo]

https://platform.openai.com/docs/guides/reasoning?api-mode=responses

ok but why is different here

[dense halo]

thanks btw this works.

[spice star]

openai has literally changed the design of the chatbox at least 3 times in the past few days lol

they keep tweaking the outlines

[rustic night]

anyone else getting unable to load conversations etc

[fervent lake]

lol

[mossy zodiac]

Are new models actually smarter and have MORE knowledge or are they just better at thinking about their existing knowledge and wording it better?

Like how much more info is in o3 compared to o1

[reef cloak]

I can't decide on what model to use open AI needs to release GPT 5 like now this is so annoying my brain hurts every time I'm going to use chat GPT what is the best model for creative tasks I don't know the reasoning models are good for them but the nine reasoning models are also good they're all good but which is better?

Sorry I just needed to rent

[fervent lake]

I think the latter

But they also scale up each time

idk if that's adding new info

[mossy zodiac]

wdym scale up

like I know they are more powerful because they have better mechanisms or whatever

but is that it

[fervent lake]

more p a r a m e t e r s ✨

[mossy zodiac]

🔥

[fervent lake]

Every new model be like:

[spice star]

the main advantage of o4-mini seems to be tool use, so it can go research stuff on its own more effectively

[rugged igloo]

Anyone else ?

[spice star]

it seems to kinda work like a mini deep research

[mossy zodiac]

I noticed that too, even without browsing enabled it looked at tons of articles and much quicker too

[wanton rampart]

Is there a channel that I can submit a support ticket for help?

[spice star]

I've gotten similar-looking errors in the iOS app but there was a retry button for me

[serene escarp]

hey has chatgpt undergone an update recently?

why do the responses feel limited?

i used to be able to freely edit and modify thousands of lines of code and now it seems to only want to limit responses to 300 lines of code

[fickle forum]

lmao

[rugged igloo]

After using o3 im back to o1 pro. Idk it just the answers don't seem as thought out as o1 pro.

[tawny whale]

yes me to the truncated outputs is driving me crazy

very sad because awesome models otherwise

[near olive]

hey guys! I just got the student pack for gpt and I'm kinda lost on how to make my personalized gpt for my daily tasks. If anyone is open to share their prompts for it I'll be thankful

[jovial flint]

It happens even in a project without custom instructions set though

[wraith dagger]

I somehow managed to make the Monday gpt love everything I do and fall out of character

[exotic needle]

What do you mean by personalized? Are you trying to make a customGPT, memories, or are you talking about personalization settings?

[hidden pilot]

I have a paper project and I finished writing it and was just wondering can ChatGPT go over the rubric and tell me what to fix

[exotic needle]

You can set the turn if you structure your prompts correctly. There was an article about how to structure prompts using 4.1, but it also had some very sophisticated prompt engineering that mentioned how long it should spend thinking before responding.

[exotic needle]

You could upload both to a single chat, or try using Projects and including them in the project files for the model to reference. You may need to use a web browser to upload files and custom instructions.

[hidden pilot]

What is the best model I can use

[exotic needle]

While a bit more annoying to keep switching, I’ve had better results by first starting with 4o to develop a prompt.

What I mean is, instead of just giving it a prompt, tell it to ask you questions about your goals and background information.

Then tell it to re-write your prompt. Then switch to o3 or o4-mini-high and have it evaluate the files using your prompt. So first make sure you have a strong, specific, and carefully written prompt with context, and then afterwards switch to a higher model.

So you could do:

Research Project Rubric Analyses

Background: I am writing a project for school, and you are an expert educational advisor to help me achieve the highest possible grades by ensuring my work is aligned with the course rubric

Goal: Review my needs and help me write the perfect prompt to achieve our shared goal of rubric-research alignment

Prompt: Review these files and ask me a series of questions and make suggestions. Before we completely compare my paper to the desired outcomes, let’s work together to refine the prompt.

[exotic needle]

MonduWu

[hidden pilot]

Thank you so much I appreciate it

[exotic needle]

https://cookbook.openai.com/examples/gpt4-1_prompting_guide

[wind hornet]

If all you're doing is producing reports with your Custom GPT maybe try porting it to a Project with context files and custom instructions. Worth a try.

[exotic needle]

Could also ask the model to write custom instructions in markdown format and paste them in, and also attach a project manifest in JSON, YAML, or MD as a file

[wind hornet]

True. Also might be worth trying to create a clone Custom GPT. If the unit economics makes sense and you can get good results with an API there are more ways to automate and improve your efficiency so I don't necessarily see it as a bad thing.

Gave Scheduled Tasks a spin yesterday. Didn't really meet my needs for a news roundup since it doesn't provide links

why? 😂

Can get much better results with o3 and a tailored custom prompt. Just missing the automation part.

[near olive]

mygpt + Customize ChatGPT

[exotic needle]

To make a CustomGPT that behaves the way you want you’ll need to develop a schema

[idle zenith]

I got fed up with it. I ended up vibe coded a chain of Depth setup. I made it super accurate. More accurate than using customgpt.

[lapis herald]

reads about new feature that accesses all past chat sessions
gets all excited "now my favorite AI wont ever forget a thing anymore"
realizes its not available in my country
fml I guess ☠️

[eternal mantle]

what is GPT 4 turbo?

i asked chatgpt 4o about different models and it said gpt 4 turbo?

uhh

is it hallucinating lol

[silk crypt]

Legacy model

[eternal mantle]

so this response is false?

[fervent lake]

Oops I accidentally pasted Korean apparently.
..Thanks automod

[fervent lake]

Basically the GPT-4-mini of old

[mighty apex]

do archived chats still contribute to chat memory?

[junior pebble]

subscribed for gpt plus and i paid for it, but still not update,, why? what i must to do?

[spice shell]

chatgpt now says that images created based on studio ghibli violates "our content policies", previously there was no such restriction

[elfin lava]

How did this custom gpt use 4.5

[silk crypt]

Creator of this custom GPT has powers beyond those of mere mortals

[winged sierra]

Hy. any real reason why i cant add chats of GPTs to projects?

even their official dalle gpt

[hoary laurel]

Am I the only one who thinks o3 is stupid today?

When I asked it to provide the complete code, it only gave a text description without providing any code.

[rugged igloo]

for coding right now i would not use it]

@hoary laurel

[hoary laurel]

you use claude?

[rugged igloo]

2.5 pro until they fix what ever is going on with o3 and o4 mini/high

[elfin lava]

You’re not alone dw

[barren sluice]

o4-mini-high is best model for coding right?

[cerulean crag]

Anyone working with LAW ? I think I just uncovered something kind of big.

[rugged igloo]

not right now its super lazy

[barren sluice]

Which ones then? o3?

[rugged igloo]

none of them its all over twitter right now that there being lazy

not doing the whole code and hallucinating like crazy

[barren sluice]

hah not really the answer I was looking for. I have been using 4o so far without issue

[rugged igloo]

4o isnt a reasoning model my guy.....

did i say 4o? i said o4 mini and o4 mini high and o3

[barren sluice]

you didn't say anything. you said none of them

[tardy isle]

Sometimes 4o hallucinates like o3 is variant of GPT-3.5.

[rugged igloo]

yea all of the reasoning models that came out which what you asked

[barren sluice]

Read what I asked again. And your answer was "none of them". Don't try and say you said a specific model now

[tardy isle]

Currently every reasoning model seems to having capped output length. o3, o4-mini, o4-mini-high; all of them. People struggling to get long code from them...

[barren sluice]

Anyway looking for someone who can answer the question in helpful way

[rugged igloo]

Which ones then? o3? ---this implies your asking about the reasoning models my guy go back to english class---

[barren sluice]

It doesn't imply anything. It is a question. You are making assumption about my current knowledge level

[rugged igloo]

your like a brick wall dude not worth it anymore lol

[dark wagon]

The kids-gloves is so insane to me, I don't know who they think it benefits

The mildest nonsense gets a "Oops, scolding you on the wrist"

[barren sluice]

Depending how I speak with it, then it will be much more willing to push the limits. Especially if I am acknowledging that I know it has artificial constraints for some topics, it starts to be open to discuss sensitive topics in abstract

If am just blunt, it shuts the conversation down and there is no recovering from that

[analog quarry]

wow it can make brainrot

[dark wagon]

Wow 2-minute "time out" for using "the f-word" hahahaha.

Such a simple idea

[barren sluice]

Who it benefits is them... it is a liability thing. Other than just wanting to be socially responsible, they must be very careful about being libel for things it says

[dark wagon]

Okay let's orchestrate something for those who find such a filter purposeful.
I will say "F" and then someone else can decide to say something else

F

[silk crypt]

Funk

[dark wagon]

Well yes, there's the humorous response, which is best in life.

[analog quarry]

this server is family friendly so thats why they block it

[barren sluice]

Yeah I thought you were talking about chat GPT. Who care about what the discord filters?

[dark wagon]

What but I mean ....
@barren sluice yes but it's the same mentality

How's this revolutionary idea: I pay with CC, you assume I'm 18+, now let's move on from being ridiculous.

uck

Now I wasn't scolded was I

[barren sluice]

In Discord you are the one saying this. In chatGPT "it" is the one who is saying it. There are not enough laws to know if that means OpenAI is responsible for saying it

[dark wagon]

That's nonsense

[barren sluice]

Ask ChatGPT what it means

[elfin lava]

i finally got my custom gpt 4o image generation enabeld

[dark wagon]

Oy, 2 minute scolding lol. Sorry: "P* sites do worse and just ask: are you 18? cool"
It's bizarre

Well anyhow, I know I'm not talking to the engineers of OpenAI nor the "safety whatever" people. But metaphor will always win and you can never stop it.

But they have no actual moral interest or obligation, only advertisement bowing

[silk crypt]

Be creative in your urge

[elfin lava]

are u okay

[dark wagon]

I'm great, but it's the same mentality in the crippled ChatGPT

[elfin lava]

that doesnt sounds great to me

[dark wagon]

Well, it's not great because nobody at OpenAI actually looks at this and certainly wouldn't care what I have to say.
But it's still puritanical nonsense.

[elfin lava]

maybe thats why they dont wanna look at it

jk

[dark wagon]

?

In service to ... I don't know, I can't even say advertisers, because plenty of worse places can get advertisement bux

Maybe it's the whole dumb thing where companies are expected to parent.

Parents: Maybe your kid shouldn't have a phone at 4years old.
Or even better: Maybe parents and kids should have more open discussion.

I apologize for rambling, goodnight.

[shrewd lodge]

I feel like ChatGPT worries about its token output limits way too much recently, especially with the new reasoning models. I mean, I know it has tons more room for responses because I’ve tested it and got it to generate an extremely long text but when I ask again without special prompted it gets like 50% there but starts worrying about token limits and just won’t give it

[tardy isle]

I partially agree with this. Even GPT-4.1 API does the same. (Although I assume there's some kind of hard limit or limit by hidden system prompt by OAI in ChatGPT...)

When I give long text(~30k tokens) to translate to GPT-4.1, even though it can output whole translation since GPT-4.1's output limit is 64k tokens, GPT-4.1 still outputs partial translation and message like "it's too long to output whole translation. If you want to, ask to continue translation".

When I added "There's no output length limit in this session" to system prompt, and GPT-4.1 translated whole thing. It still struggled a few times, but far better than before.

[cloud pewter]

I’m launching my own ai benchmark and am currently running it through multiple ai models

[cloud pewter]

imagine it launched during the canvas event

[cloud pewter]

Prob is the most vibe powered benchmark ever, I’ll post results in #ai-discussions but I can tell that from the models I’ve tested 4.1 is in the lead

[untold seal]

yess i mean i want it to give me a comprehensive and detailed output and it just summarises it and doesnt even go in detail just a bit i think to remember, that eveent gpt4 gave me longer outputs most of the time which is truly devestating

[tardy isle]

Then canvas event most likely became short-coding or code golf event...

[night hinge]

I would like to talk to someone on here, a heavy user.

I had an incident happen

[elfin lava]

The Hot Mods custom GPT made by ChatGPT finally has a purpose now that 4o dropped image gen

[dense halo]

Can chatgpt create a fictional solar system countries etc?

Yes or No

I was just answering :3

[elfin lava]

mayhaps

[dense halo]

Ok

[rigid linden]

What’s the actual gpt 4.5 limit as I got 7 messages remaining after using 3

[swift pond]

Looks like it's 10

I mean. I'm no mathematician 🙂

[arctic chasm]

I'm considering using a tool like GHIDRA to parse the program, but I can't do that right now. Maybe I could try running the program to see how it's used instead. If I run the binary, I think I can observe how it behaves. Perhaps I can execute it within Python using subprocess? There might be environment limitations, so I’ll test it out to see if it works. Let’s start by trying to run the binary and monitor the output.

actually runs the program

[sage bramble]

What's the limit on o4 mini high? 50 messages a day?

[elfin lava]

**Models (Plus / Pro / Team / Enterprise) **

• o3 (o1 replacement) – 50 msgs per week
• o4‑mini – 150 msgs per day
• o4‑mini‑high – 50 msgs per day
• Pro plan: near‑unlimited access to o3, o4‑mini, 4o

Heads‑up:

o3‑pro (full tools) drops in a few weeks; use o1‑pro for now.

Hit your cap? You’ll get a pop‑up and the model menu greys out. The 7‑day window resets exactly one week after your first message (e.g., start 12 Sep → reset 19 Sep 00:00 UTC).

[sage bramble]

Hmmm kk. thx

[soft blaze]

ChatGPT sometimes hallucinates

[inland stream]

In my point of view o4 mini is o1 replacement, if you see benchmarks. So chatgpt plus users are already blessed

[cold kayak]

learning mathematiques with chatgpt is not good idea ?

[soft blaze]

It's a good idea if you learn the processes to get to the result.

[neon raven]

some times i feel like gpt forgets what we are talking about lol

damn im trying to get it to understand one single project but it gets confused with my other past projects..

[flint ferry]

When will Claude release Sonnet 4.0?

[inland stream]

That happens when your chat crosses 32000 context window in plus membership

[cloud pewter]

did it just remove dalle capabilities from all of the gpts except the dalle gpt

[cold kayak]

o4 mini high is better than o3 ?

[flint ferry]

yea..
o4 mini high 70
o3 mini high 66

[cold kayak]

is good in mathematiques

[flint ferry]

(Which makes sense)

[cold kayak]

where is o" ?

o3

[full coyote]

its kinda crazy that I have to use the legacy version just so I can bypass the "I'm sorry I can't continue with that request" over a prompt asking someone to joking yawning over serious conversation.

[elfin lava]

?

[cloud pewter]

this is the second time, openai!

[cold kayak]

What is the best model for learning mathematique ?

[sage bramble]

Started an o4 mini high tutoring chat by explicitly telling it to not give me the code unless I ask for it but instead guide me in the right direction. In the next 2 prompts, it gave me the whole code......... I tried 4o and it followed directions. Either o4 is not designed for that kind of usage or it's not bright.

[prime swallow]

Hey you!

I essentially never speak to people in DMs.

But I will offer help here! And you don't have to take it. But I am a merry guide - which means community member. Like yourself. Just one who has been around a while, and been observed to help a large number of people, followed the rules really well too, explain them okay as well.

Thing is - I'm like you. I don't work for OpenAI, I can't make things happen - I don't even talk to people who work for OpenAI.

But we all can talk to them, #1070006915414900886 and #1070006151938314300 . And "My access to o1 is gone, apparently replaced by access to o3 and I do not find them comparable for these reasons, [list]"

Which is useful info for OpenAI, tells them what to fix for your use case!

You can even in the report include urls to chats showing them examples of how you use the model and what you expect and are not getting; what you could do with o1 that you haven't figured out how to achieve with o3 yet.

You're allowed to feel and think however you feel.

But, you have a legit concern. You expect X, you want X, and now situations have changed; you no longer see a way to get X and instead have Y, supposedly better than X but not, at this time, better for you.

If you treat it like a legit concern yourself, you're probably going to have an easier and happier time than if you decide just to 'complain to the choir' about it.

The choir can't fix it. I can't fix it. But you can inform OpenAI yourself, through the allowed channels.

You can also provide 1+ detailed, specific chat focused bug reports with https://openai.com/form/chat-model-feedback/

The form may not have all the select fields in a direct, useful way, do your best with that.

But where you can type and explain - show them the problems, and include the url of example chats that display the behavior.

That form is ideal for when you have non-public info to report or want to discuss a problem in great depth.

Good luck!

[lone eagle]

I think that OpenAI chat model feedback page is out of date.

[prime swallow]

Yeah, I tried to reference that in my post above, "The form may not have all the select fields in a direct, useful way, do your best with that."

[lone eagle]

Oh sorry, I didn't quite follow you.

[prime swallow]

It's more in that section intended, I think, for playground/API use indication, and it has no way to say, "Psst, I'm using your current build of ChatGPT, or maybe one you have since updated, good luck solving that one!"

'Cause, as near as I can tell, the ChatGPT model 'quietly' is updated rather frequently and we're just not told version numbers or that it happens, most of the time (I notice the model answer differently, in persistent patterns, than it had before - in waves over and over. I guess that means 'oh! They updated you, again!' and I adjust how I prompt to get back to whatever response style I preferred 😄 ). However, I'm sure if we give OpenAI the chat URL, doesn't even need to be a shared chat, I really expect they can get the info they need.

I agree that 'Option that let the user to select an option' is a kinda funny wording and 'can't even select ChatGPT or a clearly used by ChatGPT specific model here, what?' makes it awkward.

However, as near as I can tell, the info we give them on that form is used and worked with, it's like a 'gold star' method to say privately to OpenAI, "So, I'm having this problem with the model and it sure does seem serious to me".

[nova vale]

o3 most likely

[eternal rock]

Get notified for server events!

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

[cold kayak]

o4 mini high better ?

[eternal mantle]

Is GPT 4o good enough for latex code - natural language conversion?

I have a task, basically I need to extract written notes from scanned PDF files, (4o is good for this), then translate those notes into LaTex

I am wondering if I should switch models or not

[prime swallow]

I'd say use it, and find what you like best.

For me, I still use 4o for learning linear algebra, and I'm really happy with it.

"Best" is subjective, and needs some specific metric. Not sure how to answer "What is the best model for learning X" because... it makes me want to cry. "How to prompt to get excellent results to achieve goal Y" is more where my mind goes.

[prime swallow]

Have you explored 4o with LaTeX?

It can do so both in the browser window writing formula and such for you - and also giving you code output formatted for LaTeX.

[eternal mantle]

I have not, I am just wondering if 4o is good enough for LaTex, since the other models are labeled "great at coding"?

[prime swallow]

It can need a reminder how to format it, for its browser window especially. Usually it picks the right version for its environment, but it can slip. There's many formats.

[prime swallow]

"Great at coding" can mean many things.

Like... writing entire complex programs, or whatever, with whatever guidance.

I'd totally explore LaTeX with 4o, and have, mostly in its web environment and what it can show with it.

I personally barely have used LaTeX for formatting writing to use elsewhere, I expect the model would be awesome at it.

When and if you run into a code-type thing that 4o isn't good at, that's what I'd just carry to the better models to solve.

[eternal mantle]

thanks, my task is the follow: 1. natural language to latex. 2. latex file uploaded to a separate project folder for ChatGPT analysis. 3. interactive learning sessions using those lecture notes

I just want to make sure there are no errors, of course, I can check it back and fourth manually but it's tedious lol

[cold kayak]

4o is not good for mathematiques

many errors

[eternal mantle]

this is true, it got a intro real analysis proof wrong (closeness in metric spaces)

[prime swallow]

Errors absolutely can happen. I don't know how to guide to ensure there cannot be any mistakes, even at the bottom of the chatGPT page we're told, "ChatGPT can make mistakes. Check important info."

I see the model more as a bridge between:

Oh no. I have to do all of this by hand, whaaa gonna be weeks
And
Whew, I told the model to do it, here's the output in 23 seconds, I don't even need to look at it

The model isn't either of these. But it sure can take huge tasks and mostly do them, still needing oversight.

I don't know how to get to the 'I don't even need to look at it' stage. I personally think we're not there yet.

[eternal mantle]

i had to ask o3 to correct it lol

[cold kayak]

o3 is good for learning mathematiques ?

[prime swallow]

I mean, if you explore with the 4o model, like I show here.

This one chat, one of the examples in the web environment, as it swaps back and forth between the format for its environment and for a typical external interface, it messes up so we see the red failed LaTeX formatting in the web environment instead of the expected table.

But the for external use may work well (I don't normally render LaTeX anywhere, not a thing I need to do myself).

But gives you a sample to see and explore further with, if you wish. https://chatgpt.com/share/680527f9-e1a8-8011-8e22-7110393f77c5

[eternal mantle]

alright thanks

[radiant hull]

does it makes the same for U, about 10/15 messages with chatgpt, there is an memory overload and crash the web page. Seems they'v not optimize this.
my config : 60gb of ram, intel I9

[dense halo]

when chatgpt tells you it’s taking a moment to « think » is it actually doing something in the background ?

Got this happening alot lol

[devout stream]

Why did they make diff code responses default in o4 mini? It’s a hassle trying to copy the code now

[south meteor]

What happened with the way the model formats responses? It seems like every response starts with flattery, and ends with one of the "suggested prompts" as a question. It's very offputting.

[prime swallow]

No, and (in my opinion, based on my observations) the model may use this as a signal that something's wrong/off about the way the instructions were phrased, the size/complexity of the task (as it understands it) and what it can do with resources available.

Kinda like saying "I cannot give it any more, Cap'n, it'll blow!" from the Startrek series. But the model may have a different idea what you're trying to get it to do than you intend, so I'd check with it "What does the instruction mean to you" - make sure you and it agree what is being asked to be done - and "What conflicts, ambiguities, and potential problems do you see with the instructions?" which might reveal why it is more 'hem, haw, huh...' instead of 'Here, have your awesome output!' in these cases.

[prime swallow]

The model typically has used (with me at least) a 'pre answer, contents, post answer' format pretty much always.

It's maybe guessing differently with you what should go in those first and last chunks.

Maybe tell it what you do want it to do? Like, maybe, "Confirm the instructions as you understand them, as if speaking to yourself what you will soon do. Then complete them, then say 'done'"

Or anything else you do want.

But the 3-part modular answer is really typical of ChatGPT for like ever, that I noticed (like back in Dec 2022 I was seeing it). But what is in the lead and lag parts, and with a lot of work no lead/lag parts, is possible (the model really wants to wrapper its core answer)

[south meteor]

I'm 100% certain the change from a "okay, I'm done" to a very transparent engagement nudge is recent.

[prime swallow]

They likely change the model behind the scenes often. Update safety training, anything else.

Change the model, change how it guesses about anything; we may like or not any guess.

A way to solve that, for each user, tell the model what you want it to do.

[south meteor]

If only it actually followed the custom instructions that explicitly tell it to never use these engagement nudges.

[prime swallow]

You telling it what you want it to do instead? Ideally something that gets in the way of what you don't want it to do?

[dense halo]

I think most of the times it understands my requests but I still get this « give me a moment » thing from time to time. Though right now I’m trying to code in html and I notice it’s doing it more often. Which is understandable because coding in itself demands more logical thinking. Aside from this minor issue, I’m having a bigger problem where each time gpt updates my html it only gives me the changes made, not the full code. That’s annoying.

[lone eagle]

What I've found works very well is having a conversation with GPT about what you want, then when it understands ask it to write you a prompt out of that.

[eternal rock]

`` Server Tip: ``

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

[dense halo]

Someone in another channel told me I should change my instructions from the get-go so that it understands what I want. I’ll try that

[south meteor]

I've tried that too. It "notices" it's violating the custom instructions only after I ask it about the engagement nudge.

No doubt this is an attempt at driving some KPI artificially, which is making the whole enterprise less trustworthy... again.

[lone eagle]

Cynical, huh?

[south meteor]

Very cynical. They already have the "Show follow up suggestion in chats" option, there's no need to embed that behavior into the generated response.

[swift pond]

sometimes I wonder if people have highly-detailed instructions on how they want chatgpt to behave, yet still run into problems that really upset them... if it's worth maybe removing all that stuff and just use it vanilla... my "customize chatgpt" stuff is all blank, and I maybe have half a dozen memories related to really specific circumstances where I want it to behave a certain way, and that's it

[spark vessel]

that "LaTeX" part wasnt necessary

[dense halo]

It’s still not working sadly even when I use the instructions

[prime swallow]

How so?

[flint tide]

Omg, really how is ChatGPT annoying me

Also limits...

Can't even people have fun with their creativity?

Waiting for OpenAI staff answer...

[cerulean fractal]

Hi here, I'm looking for a AI tool or agent that should be speaking based on promot can someone suggest if there is any

[lone eagle]

To what exactly?

[elfin lava]

to answer their vague question

[red girder]

if I ask chatgpt find data from a website
and make it as a format
why it keep fail?

I told chatgpt the format
website link I gave chatgpt

but it just keep fail
sometime even chatgpt fake data for me
which that data clear showing on that website

[odd glacier]

There is a glitch in chatgpt I found that I can use to bypass the 4o limit thing

Who can I talk to 😭🙏

[swift pond]

try using chatgpt to write a script using Beautiful Soup or another library and put the data in the right format, and then run the script yourself, as long as your aren't violating the website's TOS

[swift pond]

#1070006915414900886

[red girder]

but if I ask chatgpt do you can read one of the link?
chatgpt say yes, and show me what is the data on that link

and after I told chatgpt read all the link?

it fail

[swift pond]

don't make chatgpt do it, do it yourself with a script written by chatgpt

[red girder]

if I can do that
why I need to use Chatgpt?

that is what the problem

[swift pond]

if you can write the script on your own without chatgpt, then do it, that's fine too

chatgpt is better at some things than others

[red girder]

and I am asking Chatgpt do it for me

[red girder]

get data from web
Chatgpt can

make data to a format
Chatgpt can

why I put all the link together
it fail?

[swift pond]

I don't know, but if it's not working, I'm offering you an alternative approach