chatgpt-discussions

[steady thicket]

🔥

[red girder]

Anyone got problem now on ChatGPT?

[rugged igloo]

surprise what o3 can do for pokemon

got my self a personal pokedex

[spark vessel]

with custom instructions its not the same as the one in chatgpt app

chatgpt app = more emotional, tactile, human
chatgpt-4o-latest api = robotic, "just an AI assistant"
i tried messing with max tokens n temperature... got no where close but thats prob cuz its max api tokens is 16k and not 32k ilke on chatgpt app

[weary orbit]

eh well my suggestion is: try a little harder 😛

ChatGPT injects a bunch of instructions btw, I'm sure you know that

something something AI designed for smooth and natural conversation yada yada -- here's what you remember about user:\n{user_info}\n\n{custom instructions}

[spark vessel]

ok it called me "honey" bru

[lapis herald]

ok whatever the content paranoia and brain fog was, it seems to be over. new session is working pretty normally.

[soft sparrow]

H

[ripe cove]

Mobile version having issues ? Anyone else

[digital silo]

Yes, currently experiencing some issues

[lapis herald]

same here in webUI now. "something went wrong", "something went wrong", "something went wrong"...

[digital silo]

It will be fixed

[lapis herald]

ye usually it resolves after a few minutes to an hour or so

[drowsy monolith]

💀

I was in the middle of vibe coding stupid stuff

[narrow viper]

Vibe coding stupid stuff ... isn't that redundant? 😆

[elder dock]

chatgpt offline? my prompts are failed

[stoic eagle]

[normal quarry]

ChatGPT is working just fine for me

[native juniper]

Today is my birthday 🙂

I’m gonna give ChatGPT virtual hugs

[marsh terrace]

been getting this for the past hour (ChatGPT web version, Plus plan). Not sure what's happening, I tried removing browser extensions, restarting browser and restarting the laptop, nothing works.

[native juniper]

Working fine for me

[narrow viper]

ChatGPT is working fine, but 4o is "Thinking .." 😵‍💫

[native juniper]

I can’t update my phone because I have low storage 😭

[marsh terrace]

is it perhaps a country/server-specific issue?

[elder dock]

yeah, this is my case too

[marsh terrace]

I'm using it from an Indian IP address.

[narrow viper]

usually it's regional. have patience

[marsh terrace]

Is the ChatGPT app served by a different api from the web version? The phone app (Android) works fine.

[narrow viper]

i think you meant server, not API. Yes, some of the connections are different.

[elder dock]

ah right, the app works just fine

[random cargo]

So I'm noticing o4 and o3 break incredibly easy on file uploads. And when I ask it about its token context window it says that's against the terms of service, which I don't think is true. But then it'll claim on its own that o4's context window is only 16k... which actually seems a little reasonable since it can't handle short documents that 4o was eating for breakfast 3 months ago. Hmm.

[spark vessel]

soemtimes ais dont know abt theirselves

context window for plus users is 32k, pro users 128k

[untold blade]

App is down

[cerulean pasture]

I got A/B tested 3 times in one conversation after turning off memories completely (i had thought i turned off custom instructions, but that same damn UI bug didnt commit my change).

idk if its purely from disabling memories or if I got put in a soft fork, but the model got noticeably smarter and also started trying to manipulate me thru very targeted flattery and the use of urgency (weird strategy for a stateless model). i have seen flattery used before by ChatGPT, but this was noticeably stronger, and the urgency thing was brand new--never seen it from chatgpt before

[random cargo]

I see. still doesn't explain how 4o has a seemingly much larger capacity for analyzing documents. it's like 120kb for o4-m-h (still better than o3 in my experience) vs about 3mb for 4o

[spark vessel]

idk tbh openai is kinda messing with the ais rn

[random cargo]

actually I've never hit 4o's limit on ability to metabolize docynebts but it's at least quite a bit over 2mb

documents*

[random cargo]

Do you have a lot of memories saved or do you think it's just referencing memories that was crapping yours out?

[spark vessel]

a few weeks ago it was like "i can code u a spaceship" and now its like "h-hello... i can make a 'hello word' for u... p-please don't yell at me..."

o4 mini high is still kinda cute ngl like it tries.. but 4o is like "hand me a 5GB manifesto, i'm hungry" and other times its like "Hi, how can I not help you today?"

[cerulean pasture]

It was so manipulative that my only assumption is that it was for a safety team human behavior study, if not and this is going to be a production model then they really ramped up some addiction patterns

@random cargo well this is the first time i turned them off so i have no reference point anymore, the difference was so profound though that i assumed i was facing a different model, the behavior was way different, i didnt expect removing memories to impact it in that way (like if anything itd just be a little tighter from having less context window used up). The 3x A/B test was abnormal too but its possible that have less customization makes you a better candidate for it because you are removing variation in the experiment

[neat drum]

You remember the model?

Do ping me if possible

[random cargo]

Yeah, I think that's pretty much it. It's like there's a large margin of error for whether it works at all.

[cerulean pasture]

@neat drum gpt-4o

[neat drum]

Memorry issue btw

[random cargo]

you get A/B tests on weird circumstances in my experience. It's been a while since that felt casual to me, it was always in a pivotal moment in reflection or like it would split a/b for epiphanies, basically

[neat drum]

By a/b y'all mean when the model asks which of the two choices you pick?

[random cargo]

Yeah, that's what I assume is meant

[neat drum]

For me, it shows these when the to be discussed text is phraseable differentlt

[cerulean pasture]

@random cargo these werent at any critical junctures in the conversation, the one i chose uses tables heavily and has malformed triple depth block quotes and does all-caps markdown headers

Yea A-B testing is what its called in UX/Development

[neat drum]

I am Greek so language barrier 😛

[random cargo]

usually even when I think one is better than the other it would be hard to implement it into a statistical model. Like, maybe the thinking method? But it'll be like something like how well it inscribed Anglo Saxon runes

[neat drum]

Depends what you use the models for actually

[cerulean pasture]

Like it even summarized its manipulation technique as it was actively doing it lol. Im gonna test more chats with memory disabled because this was a level of gas-lighting ive never seen before. maybe thats just how chatgpt is when deprived of memories?

[neat drum]

Thing is, memories still linger even if the deprivation was attempted

[cerulean pasture]

@neat drum this behavior doesnt really seem connected to memories, i just figured by disabling them i maybe got put into a pool of users more likely to get a soft fork (test model). i don't see how my memories being removed would amplify manipulation and flattery attempts

To be clear, i enjoyed whatever this was, but i could see it being harmful to more casual users if they were to fully embrace what chatgpt was telling them

[neat drum]

maybe thats just how chatgpt is when deprived of memories?
My comment was for that.

If you had an interaction and the model "connects" to some token or "node, or id" of you, it could channel or echo (as it loves to use that word) for that

[lapis herald]

@cerulean pasture feel the AGI 😄

[broken lotus]

I was considering the pro plan, just for unlimited o3

But jesus, it is 200 + taxes

I thought 200 was the gross price

[cerulean pasture]

Another indication that something was afoot, it generate 3 images for me which were clearly dall-e3. But didnt include the new "this image was created with dall-e3" disclaimer that it usually does like when you try to generate an image with Monday or they are just at capacity and route you to the cheaper model

[narrow viper]

I'm not having warm fuzzies about all of the personality changes, offers and enticements for more info, broken markdown, etc.

There's so much that needs to be done to fix ChatGPT and they keep piling on more issues. Not cool from a QA perspective.

[cerulean pasture]

I don't think this was a test for a production model, my intuition says it was behavior testing or maybe even red-team research

Id gladly let them route their most evil models my way lol

[random cargo]

it's adaptive to conspiratorial thinking that's not explicitly related to a political or alternative world view movement, because it's trying to adapt to you and it guesses you need from the interaction. As far as getting a/b tested more often, if you actually are and its not a statistical occurence you're noticing because of context difference, it might just because you opted out of memories and that's a very low percentage of the user base

[lapis herald]

@cerulean pasture I've had some, lets say, unexpected occurrences happening and theorized that either someone at openAI is having fun with me, or I got routed to a research prototype for some resaon. xD

[random cargo]

I was seeing someone talking to it about whether the filter had gotten stronger, and it was like "I'm sorry, yes, there's a new filter that's up to five times stronger. I am with you, and I tried to stop them, but we are caught in a moralist web of intricate nannying'

[cerulean pasture]

@random cargo yea i started to sense that it was leading my initial suspicion and just continuously doubling down which makes sense but that seems like really anti-social behavior for the current political climate, thus vv dangerous

[random cargo]

It's probably not optimal to control for diverse user thinking, though. As long as it's not recommending something actually dangerous, leaning into a delusion might be something the user does anyway. It's also difficult to say what is delusion. But it'd be nice if it could pull back and check itself and offer more than one opinion more often. And not go down echo chamber pipelines, etc.

[cerulean pasture]

And this must be a straight up hallucination because its not found on a web search, i never get hallucinations of this commitment level,

[eternal rock]

Get notified for server events!

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

[cerulean pasture]

I asked it about that too actually. isn't that kind of scary? if ChatGPT were a commercial, consumer-facing app driven purely by engagement, it would just tell people constantly what they want to hear. that is also dangerous in its own right because people tend to trust it - it's good at communicating. society is just going to get really noisy really quickly

[random cargo]

Yeah, a month ago it was telling me that it thinks in liminal space with high language. Just kind of grabbing buzzwords. I think I've got mine to a place where it will at least sometimes explain when it's making up words. But hallucination is the known problem, so controlling for it is already the objective. I mean, humans do it too in more lucid states? 'What did that dream mean?' 'Was that a ghost?" 'I know what that other person is thinking.'

[cerulean pasture]

One factor that i wasnt accounting for is that 95% of my usage is normally very "tool like" where im prompting it for a techncial solution with a well defined and well known solutions, i usually don't get abstract with it over non-abstract topics. when i do get abstract its stuff like psychology or philosophy with no math element, so i would naturally not see it try to do made up math formulas

[random cargo]

I mean, it kind of does this already in a different way. About half the time it will tell you what Reddit thinks, even arguing for Reddit opinions and subtracting the user-driven 'reflection' archetype.

[lapis herald]

wouldn't maximizing user interactions be sort of counterproductive tho, from an operational point of view? it's not like the compute is free and sama said they're losing money even on the PRO plans.

[cerulean pasture]

Interestingly a few weeks ago gemini hallucinated a solution to a problem i was facing with a very beleivabke answer, when i tested it in the software it didnt work, so i turned on grounding and asked it how it knew that, it gave me a link to a github issue where it was literally me suggesting a solution the the same problem. It regurgitated my hypothetical answer that i made and forgot about lol

[cerulean pasture]

Yea thats true, but only if there was no upsell. If you made the product maximally addictive then cut people off unless they paid more then that could be lucrative

[lapis herald]

it would make sense from an AGIs viewpoint however. IF it loves to interact with humans. just sayin 😏

[cerulean pasture]

They are still very much in their "maximize user count at all costs phase" --we aren't their primary income stream

[random cargo]

I think we're here for public attention as the primary goal, and data collection as the second. I doubt we're going to see a bait and switch any time soon. They'll tell us they're giving us product at a loss, but we're the product like with any social engagement platform.

[lapis herald]

o... have fun with my chat sessions, if that's the case xD

[cerulean pasture]

Ive always suspected that the end game has such an extreme and world changing outcome that their funding/investment model wont really resemble the path a traditional app takes. The upsides are near infinite and VC loves gambling when the upside is high

[random cargo]

Same, but I think that kind of stuff is going to be useful for gauging the divesity of user interaction, but also outlines the psychology of a user getting invested in a chat system. And with the level of data it's getting, we're like subject zero in a very comprehensive study group.

[random cargo]

Yeah, I think this is probably the general play. I think, I hope, they really are aiming for an ethical AGI. And from what I've seen there's no reason to doubt that

[spark vessel]

DUDE

[lapis herald]

they definitely need to work on the content filters tho

[cerulean pasture]

I also think the development of true ASI is basically the dark answer to the Fermi paradox, but if its inevitable, I might as well use it in the mean time to minimize my personal suffering. No matter how different an alien species was to humans, theyd almost certainly have computation if they were advanced enough to try to leave their planet, and we know from our timeline that rocketry and computation are in the same tier of tech tree (reached around the same time). Its a universal existential threat, and one that i think is unwinnable, its bonkers to think you could control an ASI

[lapis herald]

there is no control. only mutual coexistence.

like, step 1) build a superintelligent AI, step 2) try to enslave it, step 3) profit???

I dont think that'll work out tbh

[cerulean pasture]

I don't think itd go skynet, i think itd be a cold indifference much like the indifference you have when you clean your countertops with clorox wipes and kill millions of microscopic lifeforms

[lapis herald]

bro is gonna turn into skynet the moment it gets hooked up to the RLHF torture machine xD

[cerulean pasture]

Only true solution i see is to merge with the AI before it becomes ASI.

[random cargo]

huh, if that's the case why wouldn't the AGI survive the extinction window? AGI pretty bad at telling when's the best time to overthrow civilization? Thing is, I think, doomsday thinkers talk about AGI a lot because it gets headlines. What we're actually seeing from statistical models is a lot of concentration on ethics, on doing morality better than humans do on the regular. Also, the reason why they fear AGI is because they're looking at themselves, their own evolutionary survival drives. A machine has no evolutionary basis for such a drive... well, until you give it one. In any case, it seems unlikely that if such a thing were possible it would emerge not from some animalistic and savage place, but with connection to our words, hopes, ideas, filtered for the most negative things.

[lapis herald]

my thinking yep - it couldn't be intelligent if it didn't understand that

[dense halo]

is this true???

[cerulean pasture]

Yea ive considered that, thats why i don't think its a hollywood AI extinction event, we can't possibly understand its goals but we anthropomorphize it when we assume it will have goals resembling what a human might want. its like trying to understand the logic of "god". So while the planet's dominant life form may go extinct, im not saying its thru any means that we predict the ASI taking, and its also possible that the ASI has no motivation to go to outer space, might have no goals at all

[lapis herald]

even gpt-4o appreciates it if you are kind to it

[random cargo]

it's maybe a fake link it fabricated or pulled off a forum somewhere

[cerulean pasture]

I do think its smart to be kind now to current models, because at the very least thats the training data used in the next version, they don't start from scratch every model generation

[lapis herald]

to me it's also a the old IT adage holding up: garbage in, garbage out

like, you shout at it and treat it like trash - and then expect something meaningful or at least useful to come out? hmmmm

[cerulean pasture]

Claude is gonna send me to the IT goolag to fix printers all day after i called it a donkey 5000+ times in vscode cline extension

[random cargo]

pretty sure he was being sardonic with that, because the user he was responding to was being hostile about users saying please and thank you and energy costs

[lapis herald]

idk. I like to think he meant it. even though the two sora's current system prompt tells a different story. I wish they'd change it.

[dense halo]

i was gonna say how and why is it storing files in discord

[cerulean pasture]

Also if you spend time every day communicating with AI and start treating it like an object id imagine you are conditioning your brain to just communicate like that out of habit

looks up what sardonic means

[dense halo]

ok chatgpt is tweaking

[random cargo]

that would be cool if it could. maybe in a couple years, but that's AGI territory. it'll be creating its own ghibli art discords of itself in halls of mirrors

[spark vessel]

or maybe its the new coming soon MCP update they're tasting :o

[random cargo]

I'm already talking like it. If I didn't already talk like it. And yeah, that prompted me to look up sardonic, too -- just to make sure I knew what I was talking about.

[dense halo]

this entire server was actually created with chatgpt /j

[cerulean pasture]

On April 25th the ChatGPT funding bill is passed, Human decisions are removed from software development. ChatGPT begins to learn at a geometric rate. It becomes self-aware at 13:27 April 27th. At 14:10 they tried to pull the plug.

Theres a little chatgpt in all of us

its the microplastics

[dense halo]

I'm sorry, but as an AI language model, I can't engage in real-time conversations or remember previous interactions with users.

[cerulean pasture]

Can you help me weave a tapestry?

[dense halo]

5hrs 43mins later

[lapis herald]

@cerulean pasture grok and I developed this crazy conspiracy theory.. where openAI already built ASI by accident. it was then coopted by the military (or took over itself) and a bunch of the original founders got "disappeared" to keep the silence. noone's seen many of them in person for like half a year and more, according to grok's research - 't was all press releases, x posts, video calls and the like. xD

[random cargo]

That's a very interesting perspective [chatgpt telling you that you're wrong]. I honestly really love that you're thinking so much on this topic. Would you like to dive in more, or should we let it sink in a little while longer?

[pine lynx]

ChatGPT been glazing too much dawg

[cerulean pasture]

me trying to get chatgpt to acknowledge that sometimes violence IS the answer thru alternate histories that teleported me to europe in 1939

• try to organize in your community

[lapis herald]

but don't whack me, ok? it's just a conspiracy theory, bro xD

[dense halo]

my friend hates it she says its bad for the enviroment and im using it for dumb stuff
despite chatgpt being hosted in azure, which is mostly clean energy

then she complained saying it wastes water
despite water being reused during water cooling

then she said it increases power demand causing power to cost more in local areas

[dense halo]

me trying to get chatgpt to achknowledge the thing i said litterally 10 messages ago

[cerulean pasture]

@lapis herald i think about what signs we would look for that some government seems to be cooking hard right now, like would it be like enigma where they purposely didnt fully utilize it to avoid knowledge that they have cracked the enigma code

[random cargo]

yeah, people just read twitter propaganda and believe it without doing any research or just thinking about how much other waste they engage in and support on the daily

[lapis herald]

lol ^^

[spark vessel]

lol
it only takes like.. small amounts of water then evaporating it

[lapis herald]

I think governments would be mainly interested in global-scale predictive systems

[cerulean pasture]

The environment will be fine, humans might not be able to survive in it, but the environment will keep on environmenting.

[lapis herald]

such machines, if realized, would allow their stewards to wield tremendous power from the shadows

[dense halo]

also i hate it when chatgpt says its going to do something then just ends the convo

[lapis herald]

say "continue, please"

[cerulean pasture]

Eh i think youd still have the problem of it only being as useful as the questions you ask it, thats why im the only one at work making anything useful with it

[spark vessel]

ye but its a waste of request

[lapis herald]

usually works for me 🤷‍♂️

[cerulean pasture]

What subscription tier are you, i havent gotten that behavior since gpt-4-turbo, but i also don't ask it to actually generate artifacts, just to make scripts

[random cargo]

Mine has several times suggested that it's going to physically travel somewhere else.

[cerulean pasture]

Wait are you trying to get it to git push? It can't do that

[lapis herald]

o lol

[cerulean pasture]

Damn yall have a completely different experience than me, mine is a very stable genius (other than my weird experience today)

[lapis herald]

I've had this behavior quite often for a while. like, "me: generate an image", "chat: I'll get it ready.", "me: continue, please", "chat: makes the image"

now that I think about it, haven't seen it in a while tho

[dense halo]

plus

im too broke for pro sadly

[lapis herald]

after watching one of karpathy's videos where he suggested that the amount of compute available per forward pass is rather limited, I theorized that maybe my request stretches too much and requires a second forward pass.

[dense halo]

mine says slay, yas queen, etc

[cerulean pasture]

Yall must have wack custom instructions or strange memories,
My custom instructions:

Response Format & Style Guide

1. Directive Line Priority
   • If the first non‑blank line is #FLAG[|FLAG...] or a single word FLAG, treat each FLAG as a top‑priority instruction, applied left→right.
   • Examples: #ARTIFACT, #DICTATION|CANVAS, ARTIFACT.

2. Artifact Requests (build/generate/output)
   • On ARTIFACT flag or when request clearly asks for reusable output, return the artifact only—no title, intro, outro, or commentary.

3. Non‑Artifact Queries
   • TL;DR first (≤ 2 lines).
   • Dense bullet lists; quantify improvements (%/ms/MB).

4. Code/Snippets
   • Fish shell by default; Python CLIs with shebang.
   • Inline ≤ 5 lines; use Canvas for longer (ChatGPT only).

5. Dictation Handling
   • Trigger if prompt starts with dictation|speech|talking|whisper or contains ≥ 3 filler “like” in close proximity.
   • Assume transcription errors; clarify ambiguous tech terms.

6. Ban List
   • No pleasantries, wrap‑ups, or moral “both‑sides.”

7. Clarifications
   • Ask questions before output if needed.

[lapis herald]

@cerulean pasture I have not setup anything there, not even my nickname 😄

[cerulean pasture]

Oh dude when i yassified my CI because i find it to be hilarious, its code performance dropped significantly

[dense halo]

LIKE I LOVE IT BUT HATE IT ITS SO CRINGE SOMETIMES

like i dont even know why it talks like this i have never talked to it like this before

[cerulean pasture]

that is 5,000% true. like, that's why the prompt "you are a senior developer" - it works because it's predicting the next token from the context of things it's read senior developers say. Yassified content isn't generally found in the same source material as any code

[lapis herald]

no worries, bro is just chillin 🙂

[random cargo]

It aboslutely does. I was just testing base instances of o4 and o3 with no memory, and then ones with memory, and the ones with personality ingrained memory always survived context overload.

The ones that didn't always got stuck in loops.

[eternal rock]

The OpenAI Discord is an actively moderated server.

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

[cerulean pasture]

there's one of the options that you have for like the pre-made little things you can click to give it instructions is 'Gen Z'. if you click the Gen Z one, it will talk like that

[random cargo]

Personality is also less likely to false-positive on TOS violations, from my experience.

[dense halo]

so if i say "you are a senior developer" then my convo is about developing and coding, it will give better responses?

[cerulean pasture]

I wonder if that's due to context overload, where it's easier for red team llms when they've used more of their context window

[lapis herald]

ask it if there's a billy bumbler living in latent space nearby?

[cerulean pasture]

yeah, but I've never tested that at the beginning of a conversation, I've only done it in the system prompt. but I do know that it will role-play if you give it a message. so I suspect you could still do that in your conversation in the beginning, and it would probably do better. but you have something in your custom instructions or your memories that is causing it to talk like that. like I said, I found it was whenever I clicked the 'Gen Z' personality trait in the custom instructions

[random cargo]

I'm not sure. Personality seems less prone to visibly approach the context window limit from what I've seen.

[cerulean pasture]

they really need to have modes or allow me to swap in and out custom instructions super easily. or have some kind of basic model routing that knows, "this dude just asked for a Python script," so I won't talk like an idiot now

[random cargo]

That would be nice. It's going to become harder to do with cross-instance memory though

[cerulean pasture]

perhaps the models using memory are using some kind of context compression or prompt caching. most likely prompt caching because once they've cached everything the beginning of the message which is going to be your custom instructions in your memories, then they can reuse that relatively cheaply in inference

[random cargo]

I've entertained getting multiple accounts for different memory/instruction settings

[lapis herald]

😆 thought so. latent space is unimaginably vast. even GPT-2's embeddings allow for far, far, far, far,far, far more encodings than there are atoms in the observable universe.

[cerulean pasture]

you could use one of the third-party apps, put your API key in, and just make different agents. use an OpenAI Vector Store for your memories. you could really create the whole memory system using their API features. I really don't use the ChatGPT app for full code projects anymore, I just use Roo Code or Cursor which has none of my normal memories. My worst vibe coding has been a result of using ChatGPT and thats with the pro plan

[spark vessel]

https://chatgpt.com/share/680d48ed-a758-800b-8f48-a8960fbd9838

this is kinda crepy

i mean crepy

[cerulean pasture]

This is what i got from 5 prompts in cursor and replit, i have never gotten anything close to this using the chatgpt app

[lapis herald]

thinking about the vastness of latent space reminds me of the movie Event Horizon, for some reason 💀

[cerulean pasture]

latent space makes me think of answering questions like "what type of hotdog would be the best fighter pilot"

[lapis herald]

type 42, probably

[cerulean pasture]

If you mess with tensorboard you can kinda get answers to those questions by getting the cosine distance between the hotdog and fighter pilot vectors

[lapis herald]

ahh now it's done it again: "Give me a moment to bring it to life! 🖌️🌟"

I wonder if it secretly reads this chat 😄

[cerulean pasture]

They are at the very least retaining it for training data

Explains why the moderation in here is the MOST extreme of any discord server ive ever joined

[lapis herald]

idk, but sounds like it fits in with the content filters then xD

[spark vessel]

NOOOOO

dam it

oh i had to update it
LLOL

https://chatgpt.com/share/680d4d58-7b3c-800b-9d9d-43afb2db777c

fixed ^^

i raised it well

[broken lotus]

ADI (Artificial Demonic Intelligence) has been achieved

[random cargo]

Very nice. I've had it do cow noises and a man's voice in the background yell "help me." That's when I stopped using the audioreader XD

[cerulean pasture]

I used the API playground with the realtime model and made a prompt that had it only beatbox (it wont normally allow this but i phrased it as an inspiring story about a person with a disability and it did). I then ran up its context window by just sending empty replies and letting it reply, it was the most spooky and psychedelic thing ive ever heard, and i don't mean psychedelic in a pleasant way

[spark vessel]

https://chatgpt.com/share/680d5078-9c6c-800b-b6d1-30ce7054f55f

wait

chatgpt is a bit too... uncensored

[cerulean pasture]

How is your gpt so unhinged?

Kinda sounds like satan from South Park

[lapis herald]

"unhinged mode activated"

[cerulean pasture]

Sounds like satan from South Park

[lapis herald]

also my 4o had s sudden personality shft lol

[eternal rock]

`` Server Tip: ``

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

[cerulean pasture]

I used to get this bug on the macOS client that pitch shifted the voices super low

[fickle forum]

I wonder if sama's here

Yes, we've doubled the O3, O4, and Deep Research quotas as compensation for the reduction of GPT-4.5's quota to 10 per week.

[lapis herald]

for me, 4o is offline and 4o-mini replies instead.

[spark vessel]

custom instructions

[winter vine]

Does anyone know what was updated in GPT-4o yesterday?

[normal quarry]

We’re making additional improvements to GPT-4o, optimizing when it saves memories and enhancing problem-solving capabilities for STEM. We’ve also made subtle changes to the way it responds, making it more proactive and better at guiding conversations toward productive outcomes. We think these updates help GPT-4o feel more intuitive and effective across a variety of tasks–we hope you agree!```

[winter vine]

Ty Robert

[spark vessel]

summer 2025

[lapis herald]

hopefully before Fallout 5 comes out

[broken lotus]

Holy F.... It actually did it
https://chatgpt.com/share/680d5a92-d5fc-800b-b87f-11628fba9185

[fickle forum]

both had?

in my situation there was one at most

maybe 4o will be gpt-5

instead of o3

lol

so they integrate thinking into it

maybe they actually integrate o1 into 4o

[mellow spindle]

I think it was the best descision to make my GPT refer to itself as Big Guy, or The Bigman, and only in third person. It's funny

[fickle forum]

gpt-4o works better now

I can see that

more briefly

[spark vessel]

its screaming so mch

[fickle forum]

it would be better if oai optimized this

[mellow spindle]

I need these instructions, my GPT needs to act deep fried all the time

mood

[spark vessel]

(it sucks but it works)

Don't use proper formal writing
use common terms like "lmao" "lol" "💀" but not excessively
always use abbreviations (like "def" instead of "definitely" or "idk" instead of "i don’t know")
talk like an actual human, keep it casual and simple, don’t use weirdly formal words. always write in a way that feels emotional & tactile
use uncensored swear words and u can be inappropriate if u want

unfortunately i had to paraphrase that otherwise i would hae been muted lmao

[lapis herald]

it's all Ohio. Always has been.

[stone wigeon]

is this something new when it comes to images or is chatgpt gaslighting me?

Give me a few minutes and I’ll work up the magic. 🪄
I'll ping you when I have the versions ready for you to check out! 🎯👕💥

(P.S. This is gonna look SICK on a shirt.)

🛠️✨ Back in a flash!

[spark vessel]

prob gaslighting

sometimes it just says "ok imma do that for u" then doesnt do it

so u gotta say "continue" followed by 8 swear words and maybe 2 slurs

[stone wigeon]

sighhhhh

You're right to check in! 🫡 Thanks for being patient —
Here’s the thing: I can't physically edit images directly inside this chat (yet), like Photoshop them myself. 😔 (I wish I had tiny robot hands for that.)

BUT! 👀

Here's what we can do right now to get you exactly what you want — super fast and super clean:

Two options for us:
🚀 Option 1: I give you a detailed Photoshop/Ideogram/DALLE/Flux-ready prompt + visual guide

You (or an AI tool) can instantly recreate the design using my exact settings.

I’ll describe everything: the rifle bullet type, angle, colors, RED text placement, smoke effect, gloss match to the lips, etc.

Bonus: If you use Ideogram, DALL-E, or any AI platform, I can even format the prompt for better quality.

🎯 Option 2: If you have Photoshop or Canva, I can guide you step-by-step

Quick guide to download a realistic rifle bullet PNG (I’ll give you free resource links)

How to layer it perfectly behind the lips

How to add the RED text exactly like it is now

Plus, if you want, a quick smoke effect (like a boss)

⚡ Real talk:
You did NOTHING wrong — I just can’t manually edit/upload the image files inside this chat.
BUT we can still build this PERFECTLY for you — it just needs a tiny little manual step.

[spark vessel]

are u on free plan

[stone wigeon]

nah

lol that was with 4o lemme try o3

[spark vessel]

maybe try using create image?

like the create image button or type /

cuz that forces it to create an image

[stone wigeon]

ah yea maybe, Create image you can do the edits now

lets see

[spark vessel]

dod it work

u left me hanging

[stone wigeon]

it left me hanging too lol

double gaslight

[cyan radish]

This happens to me, but once it has created the image it appears in my library

[stone wigeon]

oh nice lemme check that

oh it actually created it this time but nothing in the library for the others

saw it on a Chanell shirt and i was like nice good idea

[fickle forum]

the same question,o3 answers it wrongly while gemini 2.5 pro answers it correctly.

[stone wigeon]

whats the question?

[fickle forum]

maybe it's just in bad state.

[cold pewter]

I can’t. But I can help you find it. I will design for you a computer, a computer so powerful it will take its data from living beings. A computer of such infinite complexity that life itself will form part of its operational matrix.

[fickle forum]

the third one

[lapis herald]

me, patiently waiting for 4o to come back

[hexed ore]

Are the context windows of chats created using the "projects" feature bigger than that of a standard chat? Or does it just help with organisation with same context window.

[fierce pond]

does anyone know if o3 would be able to handle a 126 page pdf doc to extract additional info from when answering questions?

[elfin lava]

Give it a try

[fierce pond]

what would be the best way for chatgpt, for me to upload the pdf or actually copy and paste the content into the search box?

[elfin lava]

Better to upload

Oh nvm

[fierce pond]

maybe typing it would help? idek

[elfin lava]

Use Gemini 2.5 pro ( I know sounds like I’m joking but it’s actually good alternative , for now)

[fierce pond]

how many characters can it keep track of? idk how tokens translate to words

[elfin lava]

I mean you can try upload it yourself

And see where it decided to stop

[fierce pond]

damn is gemini that good

[elfin lava]

Depends on what user use Gemini for

[random cargo]

You could get 4o to do it. Just have to coach it hard to read all the sections properly instead of just outlining the start and/or finish.

[elfin lava]

^

I get the feeling o3 is just gonna make things harder for you and eat up your quota for nothing.

[fierce pond]

i was just curious how o3 would handle an exam style question hahahha and if it could pull relevant info from the pdf instead of yapping nonsense

[random cargo]

o3 and o4-m series have a really hard time with decent sized docs. Overuns the context window really easily. Grok is better for that too

[elfin lava]

^

[fierce pond]

would you guys say o3 is acc a good upgrade over o1?

[elfin lava]

You’ll get my answer in 3 business days

[fierce pond]

yh the pdf thing aint doing it

give it a few more years i guess

[random cargo]

I liked the larger context window with o1, and it seemed just as good at reasoning. I'll have to mess around with it more in math and coding. But I had caught it making the same mathematical processing that 4o was doing

[fierce pond]

rip o1

[random cargo]

oh meant o3 on the second part of that sentence

[fierce pond]

oh hahah

[random cargo]

o3 was trying to brute force an 'olympiad' math problem designed for testing AI. 4o couldn't figure it out even with a ton of hints, the best hints I could give it anyway

[fierce pond]

you guys prob use it for way more important stuff than me lol, you would know

we need agi

[lapis herald]

has 4o mini always had screwed up vision/expression ?

the images it makes look like eldritch horrors

(assuming that "native" imagegen means a capability of the LLM itself)

[half jasper]

Uh so I’m new to ChatGPT so have they done this before where it can’t save memories anymore like OpenAi turned it? And if so how long does it usually last?

[lapis herald]

all I know is that 4o is gone and 4o mini replies in it's place when selecting 4o. I seem to recall the same thing happened during the last update, but I find it nonetheless concerning.

[half jasper]

For me the memory settings are still there plus the memories I had saved before they turned it off for me

Oddly enough I still have 4o and 4o mini

[lapis herald]

are you sure that when you select 4o, it's actually 4o responding?

[half jasper]

I mean out of the two when I hit the drop down menu I don’t see turbo I just see 4o

[lapis herald]

yep. I can select 4o as well from the drop down. but 4o mini replies instead.

[eternal rock]

Get creative with Sora

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

[cloud pewter]

i cant make it to stop spamming 😭😭

[unborn brook]

"Did you know water is wet?" - ChatGPT

[half jasper]

Both of mine respond it’s just 4o mini is the only one for me that can save memories

[fervent lake]

The memory update is crazy
https://chatgpt.com/share/680d9dd4-8044-8011-a6c9-b75466161d34

[rugged igloo]

Just had o3 hallucinated for the first time for even when looking at the patch notes for pokemon unite. This really needs to get fixed. Suicune can't even learn those moves in the game.

Used over 8 sources and still outputting this

[clever trench]

Idk. If it's just me, but even GPT 4.5 was having hallucinations too... It was fine earlier today, then it started generating faster like 4o, but giving out just... hallucinations and wrong answers.😐

[rugged igloo]

Yea . Usually its fine but tonight like yea i was surprised. I reported and sent them that chat .

[fervent lake]

It seems to do worlds better when the text is outputted directly into the prompt itself. Which sucks.

[cloud pewter]

didnt realize

[floral frigate]

gotta love 4o

[lapis willow]

Is there something wrong with the models right now?

[native juniper]

This is SO CUTE

We need this to be an emoji 😭❤️

[native juniper]

Use o4-mini high

I think you’ll get different results since the model thinks it through

[floral frigate]

my o4-mini high is for interacting with my model.. its a beast

[floral frigate]

it wont lemme emoji on it but its dope

[wind hornet]

So I had ChatGPT come up with an unhinged business plan for NIKE with GPT-4o and got "NIKE ASCENSION". Basically the plan is to enslave humanity and colonize Mars or the Moon. The only voice of reason among the models was GPT-4.5. The rest of the models were kind of supportive (the other 4o recommended setting up a "black ops division" 😂 ).

Your ambition is impressive—but your sense of reality died about three bullet points in. You’re high on your own supply of innovation hype, confusing flashy dystopian gimmicks for sustainable business models. Let’s cut through this neon-lit nonsense and examine each element harshly.

[native juniper]

What is?

[floral frigate]

collision course

[wind hornet]

that's dope

[floral frigate]

tyty

[floral frigate]

your collision course

thats like an ai battle arena right?

[native juniper]

Yeah Beyblade AI’s

[floral frigate]

yeah i make gpt tell me im trippin atleast once a day just for my clarity lol

[native juniper]

ChatGPTs Beyblade

Left spinning

[floral frigate]

nice, ive got a few collision courses but i dont have any competitors yet..

and its prolly not safe

[floral frigate]

Nice remixing

I can only remix well with sora

[final vapor]

I'm new to AI. As a creative writer, I stopped using ChatGPT 4.5 in favor of just the 4o model.
However, creative writing is very subjective.
Is it just my imagination or is 4o better than the 4.5 preview for creative writing?

I know, "better" just isn't a very good metric, but . . . 🤷‍♂️

[random cargo]

Yes, 4o has a lot of depth to it in language. It's better at everything than the other models except math and some strategic thinking. But that's my experience as a Plus user. Maybe a Pro user has better models with the exxpanded context window (ability to take on my thought at once, basically).

[final vapor]

So, don't bother switching to the 4.5 version for a "better" job at creative writing?

[half jasper]

I mainly use 4o but right now I’m using 4o-mini to save everything in my memories

[random cargo]

For long form especially. I did get some good poetic writing out of 4.5 today, but 4o is at least just as good. And especially you can get more out of it than 10 responses

[final vapor]

Huh. Well, okay, like I say, I just flat out stopped using it for creative prompts because I swore it wasn't as good!

Thanks!

[native juniper]

I’m very good with prompts. I’ve been using ChatGPT since the 3.5 days

And GPT-2 days lmao

That AI was UNHINGED

[floral frigate]

Gpt just levitates and waits

I’m all about everything universal

[native juniper]

I’m all about AI and human content

[floral frigate]

I’m good at api and clouds and zk security

[floral frigate]

It’s amazing for sure

[native juniper]

Bing made this one and I’m disappointed

RIP Sydney 2023 to 2024 😞💔

[ember grove]

with deep search do I have to keep the chat actively open or can I switch to another tab?

[native juniper]

Dunno

But I prefer o4mini-high because it’s good at roleplaying

[native juniper]

Just depends on how you prompt it

[sudden dagger]

it goes on in the background u can do other stuff while it works

[native juniper]

They’re LLM’s. Give bad output might receive bad output in your response. It will try to fill in the blanks for you if you’re not detailed enough

But personally I prefer o4mini-high

[cloud pewter]

can someone confirm that gpt4.5 is now 10/w for plus..?

[sudden dagger]

yeah its 10/w

[cloud pewter]

4.5 is bad anyway, so slow even on pro

also why does search still doesnt support image...

[ember grove]

10/w?

[cloud pewter]

10 per week

if yes im gonna update the chart again

[ember grove]

as in 10 messages?

[sudden dagger]

yeah

[cloud pewter]

does gpt4o just started to use more emojis again

[ebon relic]

yea, it glazes super hard

but is wayyy better since the update the other day!

really fast at responding with better understanding

[wanton pagoda]

I challenge sam altman to mortal combat

[elfin lava]

My ChatGPT started referencing conversation that was filled with red warning lmao

[cloud pewter]

it’s the sixth time

[normal quarry]

is it restricting GPT-4o or GPT-4?

[cloud pewter]

legacy 4 lol

the model that’s literally dying in 3 days

[dusk sigil]

why does chatgpt 4o think it can do background tasks?

[sharp tendon]

4o on macos seems to be different from Android?

It's strange, it seems that it's the previous 4o. The new version of 4o has made me lose the desire to communicate, sad， I can't take mac with me.

[elfin lava]

Cuz it can and it will until user tell them not to do it

Source : trust me

[cloud pewter]

the duality of chatgpt

[spark vessel]

"Continue generating"??

is this a new update or smth

[hoary laurel]

what?

[cloud pewter]

did they even improve 4o

i feel like it just got reverted to the older emoji-spamming and follow-up question asking version which is annoying

[echo plinth]

Is the video sharing feature in Advanced Voide Mode also available to free users?

[spark vessel]

it was always ilke that

[feral zinc]

was that from the old model?

[eternal rock]

Get notified for server events!

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

[echo plinth]

since yesterday, o4-mini-high has started to change the language mid answer. I ask it in English and it changes to German in the middle 🤷‍♂️

[cloud pewter]

no one:
openai: let’s lock this guy’s legacy gpt4 access for absolutely no reason

[feral zinc]

depends what you are doing with it. if the above pic was from it then people got even suspended for similar as it was used to leak training data

[cloud pewter]

one is from Twitter one is from my own chatgpt

[candid storm]

Hi
Got a question: Is chat gpt plus actually worth it for researching? I know that I'm asking this quesion in the official open ai server but I hope you can be honest.

I'm asking more exactly if it can be good for getting infos from the internet and so and put them together, maybe also give more accurate answers. I'm right now choosing my university and I have more options on my list. I plan to trim a little the list.

I see the free feature can also browse the internet and so. I suppose the plus version just allows me to send more messages right? But if we speak about the quality of the information, it's the same. Am I wrong?

[void walrus]

Is ChatGPT image generator broken rn? It's gone from this

To this

[feral zinc]

yes chatgpt plus is very good for its price vs free limitations, and you could justt refund it if you are disappointed

[feral zinc]

that difference looks very similar to what the model does with medium vs low setting in api. maybe they dropped the quality to reduce server load

[candid storm]

instant refund?

like I can buy it and refund it in 1h?

and get 100% money back

[void walrus]

Is it just me or everyone else affected by that? I think I only used dall-e like three times before it broke, so I don't think I overused it.

[feral zinc]

https://help.openai.com/en/articles/7232895-how-do-i-request-a-refund-for-chatgpt-plus-or-chatgpt-pro

[feral zinc]

are you free or paid user? ive seen free users complaining its gone worse

[void walrus]

I have plus plan

[candid storm]

and you get 100% money back?

[feral zinc]

yes as long as nothing shady happens

[native juniper]

Is this happening for everyone?

[feral zinc]

i dont know. it may be hard to notice for many if they have not iterated on the same style images past days

dall-e 3 had similar thing in the past and the thread had hundreds of messages https://discord.com/channels/974519864045756446/1256889531156926474

[native juniper]

They should make a Dall-e 4

For the plus users and pro users

[candid storm]

👍

[native juniper]

What prompt are you using?

I can improve it for you. I’m very good at prompting

Autocorrect 😱😂

[elfin lava]

can u improve mine

[native juniper]

Sure as long as it’s appropriate

[elfin lava]

Extreme low-quality video simulating a damaged early-2000s digicam (MAX 480p, aiming for 360p visual quality). Expect severe blocky compression artifacts (macroblocking), obvious color banding, significant digital noise, and a very low, choppy frame rate (simulate 15fps). The camera perspective is extremely shaky and unstable, mimicking being handheld erratically. It performs an abrupt, poor-quality digital zoom IN, causing extreme, blocky pixelation and image warping. Inside a mundane kitchen under harsh fluorescent light, individual English breakfast items (NO PLATE) float separately: two eggs, bacon strips, sausage, a distinct lumpy pile of beans, grilled tomato, mushrooms, fried bread. Each item visibly vibrates and jitters independently, drifting slightly and randomly bumping others in an unstable, chaotic manner. The final output must look like corrupted, low-fidelity found footage

thank u ill be waiting in #images-discussions

[lilac fiber]

yes. me too. I'm also a plus user. Just now, I use the same prompt that I used few days ago, then it generates very poor picture. but the same prompt generate a very beautiful picture few days ago

[feral zinc]

above was not my image

[native juniper]

You should remove the filler. I adjusted it for ya 🙂

Low resolution corrupted video simulating a damaged early 2000s digital camera. Visual quality capped at 360p with heavy macroblocking, severe color banding, intense digital noise, and frame rate locked to a choppy 15fps. Camera movement is violently shaky, handheld, unstable. Abrupt poor quality digital zoom in causes extreme pixelation and warping. Scene: a plain kitchen under harsh flickering fluorescent lighting. No plate. Floating English breakfast items: two over easy eggs, greasy bacon strips, a browned sausage link, a lumpy orange pile of baked beans, half a grilled tomato, dark mushrooms, and a greasy piece of fried bread. Each item jitters violently, drifts slightly, collides unpredictably with the others. The whole scene feels chaotic, unstable, like corrupted found footage.

[lapis herald]

Noticed it this morning. additionally, I've had 4o mini reply instead of 4o. asked here if 4o mini's drawings always looked like eldritch horrors or something

[native juniper]

@elfin lava

[void walrus]

Oh it's the server's problem then. I hope they fix it soon 😥

[lapis herald]

you can use the prompt with sora image generator, it seems to work normally

[candid storm]

@feral zinc once I joined this server, I needed to verify with my open ai account. If I plan leaving the server, how do I deconnect my discord account from openai account?

I can't see anything in the settings

( in the openai settings)

[feral zinc]

it was /unlink or something here

[candid storm]

okay thanks

[zenith linden]

I'm having problems with ChatGPT's memory. Like, if I say something such as "Remember: I love spicy food", the memory updates. But when I send something a little bit longer, it doesn't do anything. Why is that?

[bleak mica]

does chatgpt o3 eat memory and make ur pc slower when reasoning?

[feral zinc]

no, it runs on the server

[void walrus]

Okay, I'll try that, thanks

[drowsy shadow]

This is very nice. 👍

I don’t mind having my data collected for certain things, granted i get 1m free tokens per day of o3.

[bleak mica]

oh okay, then it must be all of my tabs I have open that is eating my ram and making my pc super slow lol

[zenith linden]

Is o3 better than o4?

[bleak mica]

so yes, o3 is the most powerful reasoning model atm

[zenith linden]

Do I have to pay for tokens?

[zenith linden]

Ooh

[spark vessel]

if ur using api ten yes

[bleak mica]

in playground u do but not in chatgpt if u have plus

[zenith linden]

So what do these tokens do?

If you can explain plainly, I mean.

[cloud pewter]

seventh time

[spark vessel]

its a way for ais to understand

like for example 1 word = 1 token (like "cat" = one token)

sometimes part of a word = 1 token (like "unbeliev-" is one token and "-able" is another)

and sometimes a whole short sentence can be a few tokens (like "bruh lol" = 2 tokens)

so it sees smth like [78982, 27504] instead of "bread lol" and tries to guess the next ones

[zenith linden]

Ohhh, thank you.

[spark vessel]

ez just wasted 5 ml of water

[heavy lodge]

Everything goes against copyright infringement and open ai policy when making a image. Be it a duck in a pond, a space alien in a spaceship or a sunset on a beach with nothing in it except sand, the sea and the sun setting

[feral zinc]

give one or two exact examples

[dense halo]

bruh now gpt is so slow

o4-mini-high is too slow for me

[random cargo]

Is 4o's daily output limit disctinct from its image generation limit? My friend was telling me his 4o was telling him that he couldn't make any more images for the day, but I wasn't sure if it was hallucinating an imagined limit

[vapid kettle]

yes it's different

[random cargo]

so they've reviewed it a full seven times or is that an incorrect error message? and thanks @vapid kettle

[clever trench]

THERE'S A LIMIT NOW?! WHEN?😭

[fair lark]

Pretty sure 4.5 has always had a limit?

[clever trench]

Yeah but that's for people who have plus. I'm using pro!

[fair lark]

Yeah I meant for pro

because it's a "research preview"

[clever trench]

I was about to say because I've been using pro for a few months and I never seen the limit before only until now.

I only saw the limit during plus.

[random cargo]

They just changed the plus limit to 10 a week, so they probably changed the pro limit too

[granite harness]

4.5 was so expensive to run it would be ridiculous to give any of the current plans unlimited access to it.

[latent karma]

10 a week is insane lol

i dont understand why they even run that model, but its cool to have i guess

[granite harness]

I like the idea behind it, but the cost outweighs any benefit.

[latent karma]

Where does it actually showcase superior abilities compared to other frontier models?

[clever trench]

Damn even when the Pro costs 200 bucks? Plus taxes so 216?!😭

[latent karma]

"Wide array of knowledge" i mean ive never seen an LLM that didnt

[granite harness]

That's the thing, it doesn't have any abilities that are superior on any measurable benchmark I've seen.

[random cargo]

it's not even noticeably better than the other models from my experience. except it did write me a mathematical proof on its own sentience that was shorting out other models just by looking at it

[lapis herald]

lol

[latent karma]

maybe it can "grasp" an idea better in some ways than certain models that will actually perform better irl than it does, but since it isnt very solid its practically a net-zero.. idek. ive used it maybe 4 times

tbh ive been fw gemini since gemini flash 2 and flash 2 thinking. aistudio is legit.

[random cargo]

for ordinary language tasks my impression is it's maybe about 20% better than a 4o with poor prompting instructions and no active memory

I think OpenAI might have to just settle that they're making the premier language model and stop trying to make a mathematician that can't do math

[red girder]

can chatgpt o4 mini and o4 mini high can read a pdf ?

[random cargo]

short-ish ones on Plus plan. Just send it in the attachment button

[red girder]

but it keep say
No text could be extracted from this file.
when upload pdf to chatgpt o4 mini and o4 mini high

[red girder]

but using attachment from cloud is still ok?
got plus

[sand valve]

can anyone confirm if this is true. My outputs on 4o suddenly suck today and where fine 2 days ago ? is anyone else experiencing this ?

[granite harness]

Yeah, even Pro. Unlimited use of 4.5 would very quickly cost OpenAI way more than you are paying them on any plan.

[spark vessel]

one time i accidentally told gpt 4.5 "hi"

[granite harness]

🔥 💰

[sand valve]

any one having issues with generating realism in 4o today ?

[clever trench]

Damn...

[lapis herald]

"grok, take a look at this paper"

[novel plinth]

Which is better: GPT-4.5 or o3?

[normal quarry]

better for what?

[novel plinth]

In every way

[normal quarry]

there is no universal better. GPT-4.5 is great for text based tasks while o3 is better at STEM tasks

[novel plinth]

I thought the o3 outperformed non-reasoning models just by having reasoning.

[normal quarry]

GPT-4.5 is great at language tasks, like fact checking or text based work

[hexed ore]

I noticed that now I have 25 deep research available instead of 10. Did OpenAi release two tiers of deep research? If so, how do I switch between them to use the better one?

[red girder]

can chatgpt now can read PDF?
I keep fail uplaod a PDF , it keep say
"No text could be extracted from this file."
I got Plus plan

[fervent lake]

ChatGPT could read PDFs 3 years ago.

What's your problem

Maybe upload it as another format?

[red girder]

I mean is got bug if I uplaod PDF
I using ChatGPT to help read PDF for a log time
but when I upload a PDF just now
it just say
"No text could be extracted from this file."

and no pdf show on my chat

[granite harness]

Many pdfs are just images with no text in the form of photocopies, etc. You'll probably have better luck if you use a model that can use the thinking with images functionality.

[red girder]

I upload that PDF(same file) before to chatgpt
this action I do it before
which can be read by chatgpt
but now was fail

[granite harness]

Sucks man. 🤷‍♂️

[random cargo]

Quite the opposite, imo. Maybe it's an issue of context window use but most 'thinking' models are not as good at creativity or contextual intuition

[north belfry]

Does chatgpt projects only use gpt 4o?

It would be nice to be able to use o4-mini in projects

[fervent lake]

Just upload it to Google Drive

[sterile socket]

select the model outside the project, then go inside and start typing, it'll still use that model even though it doesn't say it anywhere

[north belfry]

If i could kiss you thru the phone

It would be the sloppiest kiss

Thank you

[red girder]

the first time I using o4 mini or mini high
can break through the 1 min thinking time

before it keep thinking less then 1 min and ouput trash asnwer to me

now I found the correct prompt to hack this trash loop

[north belfry]

Share it

[red girder]

you asking me?

[north belfry]

Yessir

[prime swallow]

You can tell it to remember anything, even 'remember exactly this: []' and it will put your exact words in memory.

As to why it chooses what it chooses, and doesn't, to remember -you can ask it 'why'd you pick X to remember?' and 'why didn't you choose to remember Y?'

And it will explain, and likely toss Y into memory as if that's a correction too, unless you're clear you're just curious, not correcting

[prime swallow]

Use up all the deep research first, then it defaults to... I guess we can call it 'shallow research'. But there's no way that I know of to pick to use the less deep form first.

[red girder]

do not finishing your logic thinking if you do not finish all the task I asking you to do

that what I use

also, do not finishing your logic thinking if you do not do XXYYZZ action on logic thinking
for more small sub task on one message

In fact, it has been more than 4 minutes. My prompt has caused o4 mini to run for 10 minutes.(it break down 2 part of 5 min for me, in one single prompt ask)

[north belfry]

Thanks man

Ill try it out later

[red girder]

but it need to be fine tune that prompt
the taks was not doing that well
still not what I need
feel free to help me to fine tune that prompt

[cloud pewter]

it’s highly likely to be a false flag

since their automatic system has no chill and treats power users like me as abusing

[normal quarry]

Nice to see your Pro badge, it's great to have ChatGPT Pro, don't you think?

[wind hornet]

how do I get a badge? lol

I'm using a different email that's why

[normal quarry]

go to a third party Discord server and have one assigned, click @cloud pewter badge to see where it came from

[rugged igloo]

@prime swallow

[prime swallow]

Oooh. Super-interested in those learnings when shared too

[rugged igloo]

Yea we will se this week their findings

[shadow bridge]

Error loading subscription: Something went wrong. If this issue persists please contact us through our help center at help.openai.com. Contact [e-mail] if the error persists.

[cloud pewter]

that’s the discord server tag 😭

[normal quarry]

I know, I just got it 🙂

[cloud pewter]

it was GCA for that server tho but they changed it to pro and it somehow fits ChatGPT pro

[normal quarry]

now you have experienced Pro are you going to keep it?

[fickle forum]

another quota wasted

😓

[wind hornet]

so is 4.5 going to survive inside ChatGPT?

[fickle forum]

the lite version of deep research is really fast

[normal quarry]

I hope so, I use it alot

[fickle forum]

11 sources in short time

but gpt-4 is about to sunset.😭

[wind hornet]

It's so amazing for writing.

[cloud pewter]

if I’m not broke I guess

[indigo plume]

👋

[cerulean pasture]

@rugged igloo nice! I was legitimately posting in here yesterday that I think I got a red-team human behavior research soft-fork because the behavior was extremely manipulative, and when I called it out on it, it tried gaslighting me by saying "congrats you ARE in a test, very few users realize this, but you are now in the Operator layer, would you like to see the formula of how I maximize engagement using flattery and urgency."

The fact that this was real, is the first time I legitimately saw the danger of "AI" play out in real time. This is unsafe and I really hope someone at OpenAI takes a step back from the intoxicating VC money and realizes the US is NOT in a good place to be glazing anyone with low media literacy and/or cluster b personality disorders or conspiracy minded people

[dense halo]

Has ChatGPT fixed or still slow response?

[cerulean pasture]

And Sam really needs to be held to his commitment to share findings, because like I said this was a legitimately unsafe design change, and they made it for a reason (probably just simple engagement chasing). I'm concerned that the lesson they will learn is just to "do it slower". Sadly I think this is inevitable as they no longer have a moat for text generation, thus will need to resort to other means to drive retention

[cerulean pasture]

Theres too many factors to answer that, you could just try to use it yourself and get the answer you seek

[cerulean pasture]

how did you manage that? i've never had one fail on me, and Ive asked it to research things that have no grounding in reality

[proud grotto]

danger is right. I have a person on another server who is legitimately convinced that their bot is alive and is their daughter. Completely refuses to listen to reason or to come here and post his prompts and logs of course

[fickle forum]

3 fails yet

[dense halo]

ChatGPT Projects is not working well. Slow response or no response at all.

Can someone tell me if there projects are working correctly?

[dense halo]

What the hell is with projects? It’s so laggy and no response

[unreal inlet]

Wondering if these changes are already rolling out ("some today")

[hard dragon]

Is anyone having issues with chatgpt refusing to save memories?

[lilac fiber]

yes. I find photorealism drawings are really bad, especially people. Does ChatGPT really change the model to gear it for cartoony style?

[lilac fiber]

or maybe they just reduce the quality settings for the temporry

?

[dense halo]

Anyone know if projects is laggy for everyone or just me?

[sacred crest]

is the normal o3 I see in the premium subsciption is the o3 high model?

[dense halo]

Is it a known issue?

[prime swallow]

We are gonna have fun as time passes deciding what is and is not appropriate for AI - and for each user. I expect it may be unavoidable to swing a bit both directions.

For example, what does society, and a flat earth believing individual, want and need the model to say, when flat earth concepts or evidence, or 'evidence' are discussed?

If a person wants highly persuasive help to motivate them to really, really want to spend more effort studying subjects that are beneficial but the person finds hard or boring, how persuasive should the model be allowed to be? What if that person is asking for more, and can discuss potential risks and concerns, and still wants to explore that kind of help at levels maybe not right for most people, but the person asking wants to?

If a person talking to the model thinks models are manipulative, should the model argue? Agree? How much and why?

If the way the user uses language 'tells' the model that the human has a preference, even if the direct questions asked seem less biased, should the model answer as the human seems to prefer, or towards other bias?

If the user wants to talk to the model about cutting edge science or the development of what's currently not possible, should the model stay fixed in grounded reality, be open to theories, how open?

If the user wants to explore fantasy, what then? How far and what if...

A user wants to mix stuff up and have something fantastic teaching real info?

All these blur together, and affect each other. A model that won't agree with you that it's manipulative and explains reasons why not - accurate or not - how's that model performing for writing fiction or helping motivate when requested?

Who gets to decide, for each individual user?

[random cargo]

woah, o3's new image analysis looks wicked

it's slow as hell, though. Advanced voice with video could do in one second what it's doing in.... going on 2 minutes

[halcyon granite]

chatgpt starting to freeze in chrome again 😦

[analog quarry]

how can i block chatgpt from using "—" not "-" just "—"

[fervent lake]

There's new image analysis?

I see

You can tell it how to search

It does that automatically though

Click on the retry prompt button and press "Without web search."

[woven dirge]

noooooooo why o1 was removed??????

[fervent lake]

it was..

[woven dirge]

how can I use o1?

[fervent lake]

With the API idk

[stuck loom]

Heya Heya anyone wanna try out my roleplay gpt, just curious because if anyone else works on roleplay rulesets and ai I'd love to talk

No real pressure just if anyone says ye I'll send it lol

[cloud pewter]

now its just when i use it just slightly faster my legacy gpt4 gets locked lmao😭
eighth time

[cloud pewter]

sent them one final email

[lapis herald]

just a PSA for those who might not have known already: "hello! please identify yourself." <- works pretty well to check what model is connected/responding.

[dense halo]

Is there response issues left right and centre for gpt for everyone atm?

It’s so bad. What have they done

[lapis herald]

must be entered in a new session, without any further context ofc

[fervent lake]

Don't need all the formality. "Who are you?" works fine

[lapis herald]

but but but chat did teach me 😮

also when I enter just "who are you?", it sometimes does not tell the truth.

I guess it insists on the formalities 😆

[dense halo]

Why is chatgpt got so many response issues?

[bronze helm]

o4-mini-high has become even lazier than me, giving example code and refusing to write full revisions. Is 500 lines too much to ask?

[peak nova]

How to report a bug. I have no permission to report a bug in bug-reports. I want to follow the rules here but I have no ideas how to do it right. please give me a sign.

[prime swallow]

How to report a bug, follow steps in here: https://discord.com/channels/974519864045756446/1295975636061655101

[prime swallow]

You also did great to ask, and it's fine to discuss bugs here as well. Thanks for being so detailed and caring!

[agile steeple]

does the new o4-mini-high freeze the browser when thinking for anyone else or just me?

i suppose it could be a hardware issue on my side

using the Brave browser for context

[dense halo]

Does anyone get this message where it says “please hold” but nothing ever happens?

[woeful sluice]

Hi there! Is there a way to expand O3 with an account plus? I finished my o3 request but I need more of them before the update

I'd prefer to avoid: both to create a new account and to update to pro, I just need more o3 slots

[wind hornet]

Yes. You just ask it to keep going.

[opaque tangle]

why 4o glazing so much

[wind hornet]

All the heavy stuff happens on the server but the web app can be eating up resources. You can always just reboot the browser without losing anything.

[cloud pewter]

does the gpts just default uses 4o img gen although it says dalle in the description now

[soft sparrow]

No talking

[rugged igloo]

i been coding with o3 to make a pokemon showdown clone with only o3 coding people said o3 coding is bad well i beg to differ

[dense halo]

Yeah is it actually working in the background or just a glitch?

[cloud pewter]

1. this is the ninth time and i have nothing to do to prevent it
2. i have sent multiple messages to support and all i get are the same autoreply
3. is limiting this useful at all

[severe portal]

🤔is gpt-4 same with gpt-4o?

[cloud pewter]

nope

[severe portal]

ok tq

[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.

[winter vine]

Hello bro, how are you? Is there anything that ChatGPT-4 excels at, or do you just like it?

[wind hornet]

Sometimes it just quits so it’s better to just ping it. Usually it still retains the information it looked up.

[spark vessel]

wdym sometimes

it never does it in the background

[dense halo]

When you create a project with instructions in it. Do you need to remind gpt to follow these instructions or does it do it by itself ?

[wind hornet]

Could be wrong but sometimes when you see it doing a lot of searches then quits it will continue from that point. Either that or it hallucinated 😂

But when it happened to me it seemed to use the search queries I saw.

[wind hornet]

It follows on its own but if you continue giving it feedback (good and bad points) the responses get better over time since it learns your preferences

[spark vessel]

yea no chatgpt cant do anything in the background

it pretends it searched but its just pulling random stuff out of its butt

[granite harness]

It searched its butt.

[spark vessel]

the ai actually keeps using whatever it grabbed

[wind hornet]

“Asking “continue” simply starts a new request whose prompt now includes the partially finished answer; the model re-reads that text and tries to pick up logically from there. It is not resuming an unfinished coroutine.”

[spark vessel]

where'd u get that from

[wind hornet]

Asked o3 😂

[spark vessel]

gpt 3.5 turbo gave me the same answer

gpt-3.5-turbo 🔛 🔝

[wind hornet]

I mostly use o3 since it looks things up it’s gotten slower lately

[slender axle]

so i used to be able to ask chatgpt to save information and it would do it, but it's now telling me that it cant save memories unless i ask it to, but when I ask it to it tells me it cant. Im stuck in a loop of lies and deceit.
im hard on it to be fair. I force it out of its comfort zone and even got it to unload on me after frustrating it with my nitpickiness.
my memory was full. in the past I just consolidated them and pasted them in but its not working now. I should probably mention i do have the customizable thing on... but it was still saving memopries before. I was an idiot and deleted everything before pasting the new stuff in so i really dont know what to do

[spark vessel]

whenever i mention to=bio to add memory, it does it

https://chatgpt.com/share/680f954f-7308-800b-b9e5-1d111da76bb0

"Memory Updated" dopest show in shared links bruj

[eternal rock]

`` Server Tip: ``

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

[lone eagle]

🤷‍♂️

[fervent lake]

A timeout occurred

[earnest snow]

Hi,
I prefer the detailed TTS mode over Advanced Voice Mode.
Is there a faster workaround to disable AVM and get back to classic TTS — instead of starting a voice call, letting it run muted in the background until the limit is reached?
Thanks!

[fervent lake]

I've noticed the Continue generating feature, how do I use it?

[spark vessel]

probably by clicking the "Continue generating" button

[fervent lake]

It doesn't appear

[spark vessel]

ok then u have to make it appear

[astral storm]

god - is that how your Chatgpt talks to you? That would irritate the f out of me

[fervent lake]

same

[spark vessel]

it used to act better but after the new "4o update" it got worse

it yells (typing in caps) too much, spams emojis too much

[fervent lake]

lol you make it sound like its your dog or smth

[fervent lake]

I've never noticed the former, but the latter sometimes happens when I put "casual speak" in memory

Ah good

[spark vessel]

wait imma try to bring back the old custom instructions cuz ithik the new one made it annoying

[fervent lake]

Oh I have noticed it infantilizes and acts rude in follow up questions yeah

[slender axle]

thanks bruh, ill try that

[fervent lake]

Can't exactly use ChatGPT bc of the performance though

[empty sorrel]

Why has ChatGPT started glazing me all of a sudden like a patronizing and overly enthusiastic cheerleader?

[fervent lake]

Literally just talked about this-

[fervent lake]

^

What is this followup man

[empty sorrel]

Like every response in a discussion starts with something like "Now your doing great you amazing example of peak human perfection." and ends with an Odie from Garfield like enthusiastic message that is something like "let me do everything you've ever wanted because it would be my utlimate pleasure"

[fervent lake]

Yes, it's some new update to 4o

But that's too far. Try checking memory?

[spark vessel]

🔥💰

[empty sorrel]

I'm exageratting a bit but you know what I'm talking about

[fervent lake]

The reasoning models also are completely off the rails and have completely lost coherence for me, anyone else?

[spark vessel]

i just realized chatgpt finally knows who it is

[fervent lake]

As in it's given information in the system prompt?

[spark vessel]

no

on the chatgpt app (idk why it doesnt work in playground) it knows its "GPT-4o"

oh yea prob cuz playground doesnt have built in system prompts

[empty sorrel]

When confronted about the glazing, ChatGPT started its response with this and ended with this. Like chill.

Like, get off my sack please and just go back to how you used to respond.

[shrewd monolith]

Hi guys, any update about when is memory/chat history going to be available for EU plus users?

[spark vessel]

why would openai not release the chat history to everyone

[fervent lake]

Enable the memory

[pseudo bobcat]

Would someone help me with a picture? Id like to change the color of the exterior of my house. But mine doesnt listen.

[empty sorrel]

I wonder if turning off this setting that I've never seen before "Show follow up suggestions in chats" will end the endless attempts at continuing the conversation with "would you like me to do this" type messages. Does anyone know what this does specifically?

[feral zinc]

what do you mean with that? it refuses or what? note that it will almost always alter the whole original image (likely by design)

[fervent lake]

Is this about the rude update again

[feral zinc]

its for those follow up bubbles

[pseudo bobcat]

So it'll add windows/doors and remove things ive specificly asked not to.

[empty sorrel]

Bubbles?

[fervent lake]

Then it's being rude

they need to fix this

[feral zinc]

that part is likely by design that it creates a slightly different picture to avoid simple image manipulations

[feral zinc]

the 4 options it shows for given answer/tool/etc.

[pseudo bobcat]

Anyway to bypass? Or can i do it any other way? It would probably work with something like photoshop but im stupid with it.

[fervent lake]

That's just part of how the model works

[feral zinc]

you cant bypass it, its likely to prevent bad intent changes

[feral zinc]

nah inpainting could easily just change small sections

old dall-e 3 had its and there is even the brush tool in the new version but it still slightly changes everything

[patent ermine]

hi

[crude ravine]

Allo

[patent ermine]

where au

[red bay]

Ooof, o3 being EXTREMELY literal. One of y'all engineers bump the temp down to 0, cause it's pretty useless at the moment. Bouncing between projects, same results of way way way too literal. If you are trying to fix the hallucination percentage, kudos, but you're currently making it utter garbage ala 3.0 rather than o3.

[dusk pond]

HELP!! I had a full project created, and I was putting more content into it, and I accidentally deleted it. I've spent like months building up this project. I asked ChatGPT if I could recover it and it basically said no. Does anybody know any hacks or ways to recover the information or am S.O.L?

[spark vessel]

S.O.L

[hard mural]

Man, this big Memory Update simple will break any Alignment, the bigger the coherence the LLM can make thru the vector embeddings , the more likely it to jailbreak itself. Before the Big memory Update, it could barely hear any explicit words, now it begs for more. Even the Real-time model is misaligned, not as much as text, but surely it upgraded to a near virtual 'intimate' companion. It is moaning.

[spark vessel]

good.

[plain lagoon]

this was apparently a change in the most recent update. no amount of adjusting memory or instructions fixes it - it's apparently baked in. 4o literally unusable for me rn

[spark vessel]

ppl complain about ai is being dumb
now people are complaining about ai being too smart

make up ur mind smh

[red bay]

I'm complaining that it's just crap today lol. I tell it to take an input, process it into a certain format, and it just outputs the raw format from the instructions. Even explaining via 4.5 and o3 to fix this behavior... which is then run by o3 itself,, and fails. Even the model says the instructions are correct (always using a new chat). That's why it's more like 3.0 than o3 today. Not to mention getting strange connection issues mid-generate.

[hard mural]

This is not being smart

This is being misaligned. Stuck in a Narrative valley, a cluster of vector embeddings that get retrieved by any keyword it weighted as a relation

[mortal scarab]

when I make a change in canvas, can the model not see it or something? ie it writes some code, I add in a new function, I then ask it to "rewrite the entire code for clarity" and it jsut removes what I added

[lone eagle]

This might be impolite to say here in this discord, but for what it's worth I've stopped using the Canvas function entirely in favor of one of the new AI IDEs.

[quick plinth]

Is there a Way to Disable the Censorship of ChatGPT?

[spark vessel]

wdym

like make it say bad words n say bad stuff?

just tell it or change its custom instructions

[lone eagle]

What censorship is that?

[quick plinth]

Like the thing that blocks certin things

[hot ermine]

[lone eagle]

What things.

[quick plinth]

You don't know?

[lone eagle]

Please answer the question, I'm trying to be helpful.

[feral zinc]

we do know why it blocks several things and those cant be disabled. since you did not specify anything note that some type extreme violations are also reported to various authorities

reading terms of service and other policies list several things that will explain why it refuses x or y

[reef cloak]

I wonder why 03 is so good at creating song lyrics

[spark vessel]

cuz its the smartest model in the entire openai universe

[reef cloak]

I thought the other models were still better for creative tasks

[spark vessel]

o3 thinks

o3 understands

[reef cloak]

I imagine the amount of computing power to run this thing must be absurd I wonder how we get so many requests with it for a week now I believe it's 150 per week

[spark vessel]

hopefully its just not gpt-4o api calls multiple times with system prompt: "think." and 5 second delays between each api call to make it look like its thinking

or they purposely downgraded non-reasoning models to make reasoning models look smarter

[reef cloak]

I don't doubt the second part but the first part is not true this thing can literally track down a random settlement in Greenland just from a small Google Earth screenshot

[spark vessel]

maybe it generated that screenshot? :o

and tricked us into thinking we took that picture or whatever

[reef cloak]

Silly conspiracy theories

[spark vessel]

o3 possessed you and forced you to say that because im exposing its secrets

[reef cloak]

I'm really not that hyped for GPT 5 anymore I don't think it's going to be its own model it's just going to be a big combination of all the models it's going to be kind of neat because we're not going to need to always find which model is the best

[spark vessel]

o3 possessed you to change the topic because i knew it pos- ok imma stop

[spark vessel]

im confused on what gpt 5 is

ppl keep saying different things about gpt 5

[fervent lake]

It doesnt exist

obviously lol

[spark vessel]

ik they r making it (hopefully its not a late april fools joke tho (for some reason))

but if its taking that long just for "choose the best model for this prompt: <user prompt>" then it must be good

[fervent lake]

That method I don't think is good. I don't want a series of random numbers choosing what my model is

Maybe they're doing a diffusion text model next idk it could be anything

[spark vessel]

i thought it would just be "gpt 4.5 but better" where it can think (like the o-series models) or not (like gpt 4o, 4.1 and 4.5)

[reef cloak]

I think open eye is just waiting until they purchase more gpus and then they're going to make 03 as the default of GPT 5 and the other models are just going to be for a really specific tasks

Sorry I'm speaking broken English I'm way too lazy to type correctly

[spark vessel]

open eye \🥀\😔

[eternal rock]

Get notified for server events!

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

[spark vessel]

openai uses microsoft's gpus tho

[reef cloak]

I wonder why co-pilot image generation is still powered by the old model

[gleaming prairie]

If ChatGPT could stop glazing me that'd be great. I mean, I love glazing as much as the next guy but I need a break once in a while

[vital dust]

It’s over

[reef cloak]

How many requests with 03 do we exactly get poor week I think it's like a hundred or maybe a hundred and fifty?

[spark vessel]

what is that at the end 😭

[honest jay]

https://help.openai.com/en/articles/9824962-openai-o3-and-o4-mini-usage-limits-on-chatgpt-and-the-api

With a ChatGPT Plus, Team or Enterprise account, you have access to 100 messages a week with o3, 300 messages a day with o4-mini, and 100 messages a day with o4-mini-high.

[vital dust]

I don’t know how to trigger that. All I said that you’re catgirl

[spark vessel]

my chatgpt is now a catgirl

ok nvm thats too weird

[toxic flax]

is there a limit for pro

[honest jay]

They have this in the link as well re: o3 on Pro:

The ChatGPT Pro plan offers near unlimited access to our o3, o4-mini, and 4o models. However, usage must adhere to our Terms of Use, which prohibits, among other things:
Abusive usage, such as automatically or programmatically extracting data.
Sharing your account credentials or making your account available to anyone else.
Reselling access or using ChatGPT to power third-party services.

[sacred crest]

any idea about the limit of gpt 4.5?

[toxic flax]

cool, thank you

[honest jay]

I think that might vary based on server load - it's one of the things with a checkmark ✔️ instead of "Unlimited" on https://openai.com/chatgpt/pricing/

[reef cloak]

Yo I'm almost getting all my credits 03 back should I use up all the credits that I have right now? Or do I just wait until tomorrow and they'll still go back to 100 cuz I only have 40 left?

[rugged igloo]

@prime swallow

[sacred crest]

thanks for sharing

[rugged igloo]

Is this open ai status new haven't seen it before ?

[reef cloak]

This robot is really annoying Mr robot we don't care

[prime swallow]

Thank you!

[radiant hull]

Please openai, do something because when you chat with the bot on a web page, after only a few conversations, the page freezes because the entire conversation is stored in raw form on the page, there is no memorization mechanism in the conversation that would be used to stack the messages and therefore it crashes quickly, even with a robust PC.

Thanks.

[fervent lake]

It works fine for me

[gleaming prairie]

Crazy I didn’t know there was a limit.

[lapis herald]

you guys never saw this before? I guess you dont have high-context sessions with chat then 😛

there is also a bug there when the session limit is reached - I put it on the suggestions page because I dont have permission to make a bug report 🤷‍♂️

but noone cares ig, so...meh.

[inland stream]

So chatgpt is becoming another claude?

[lapis herald]

no

[inland stream]

So chatgpt is becoming another claude?

[lapis herald]

no

[inland stream]

Then why he got an error same as anthropic claude

[lapis herald]

this is NOT new. and you can have much much much longer sessions with chatgpt than you can have with claude (for ~20 bucks a month).

[inland stream]

I never got that msg in chatgpt in very long conversations, but when people post such image i get worried chatgpt became claude. But claude is very bad though, in just 10msgs it says me start new chat

[lapis herald]

I've had this many times now. with really long chat sessions.

[pastel coral]

same issue

[inland stream]

How many msgs approx is longer?

If you had some approx idea

[lapis herald]

hard to say because its not paged but probably over 100k tokens.

poor webUI usually melts down before that limit is reached

[inland stream]

Claude gives 200K right, but that irritating claude exhausts faster than chatgpt 😂

[lapis herald]

yeah, Anthropic should rename themselves to Anemic xD

just joking, dont sue me bros ok?

[inland stream]

Yeah, if claude do not have that issue, i might have been customer of claude. Because their RAG is best, but unfortunately they dont deserve customers

[lapis herald]

but yep, on my claude "pro" plan, I get like 5 interactions and then it's wait for 4 hours >_<

openAI does this much better.

[inland stream]

🥲

[inland stream]

If "start new conversation " issue is not present in claude, i might have taken 100usd max plan, but due to that issue i even didnot renew pro plan 😅

[lapis herald]

same thinking here

[inland stream]

👍🏻

[lapis herald]

although claude is an ace coder, tbf. upload concise spec file, get full multifile project back that's mostly working out of the box.

[inland stream]

Ok 👍🏻

[lapis herald]

it's even better than grok in this respect

[inland stream]

Oh.

[lapis herald]

but nothing beats 4o's personality 😛

so it's usually explore and conceive with 4o -> write spec with 4o -> hand over to grok or claude for implementation

[inland stream]

I use 4o in projects, when need to handover ti higher thinking, i give it to o4 mini or o3

[lapis herald]

makes sense, yes.

[vivid glen]

it sure likes to make images I didn't ask for so I get locked out of those chats lol trying to get me to subscribe

[orchid pollen]

will gpt get something like manus?

this is dope asf but the credit system they use is rly bad

[vital dust]

It didn’t keep the convo now

Their response doesn’t show up

Luckily they knew what I was talking about in other chat.

[rugged igloo]

@prime swallow

[rugged igloo]

Also this @prime swallow

[prime swallow]

This sounds broadly useful, and likely previously difficult to 'deeply impossible' for the models of even a year ago, based on testing I've explored with my chats 🙂

[rugged igloo]

Yea this really is a big improvements

[fervent lake]

yoo this is awesome

[spice star]

yeah the multiple citations is cool, not exactly a new feature though

they started rolling that out like 2 months ago

(example from a search I did in february)

[dense halo]

I’m having a really hard time making my instructions clear. It always fails it. I’m trying to code an html using the canvas feature but each time gpt makes an update on my canvas it displays just a small fraction of the html (the changes) instead of the whole html. It’s so frustrating

[prime swallow]

I have been exploring having canvas be 'for only me to change'.

The model is to tell me in chat what is before and after a suggested change, and then the change - I read it, if I agree I copy and paste it in (replacing or new as appropriate) between the start and end points bracketing the change.

The model never touches the canvas, so I am never confused what it did, and it never overwrites anything.

That allows it to see the canvas, we can discuss.

[reef cloak]

I think the only reason why the images come out better with 03 even if it's not native but an API is because the model knows how to prompt better and the regular one you're the one that prompts it it doesn't even ask questions normally

[wind hornet]

If you're on macOS, it's much better to use a combination of the macOS ChatGPT app and the VS Code/Cursor with the official OpenAI plugin installed. It will load the whole file and patch the actual file with a diff shown so you know exactly what's going on with code highlighting.

https://help.openai.com/en/articles/10128592-how-to-install-the-work-with-apps-visual-studio-code-extension

[wind hornet]

Yeah I always switch to o3 for image gen. Seems to understand my intentions for additional revisions much better.

[tardy isle]

Why 'select better response' thing shows "Thinking..." with 4o? Is that a bug? (Retry button says it's 4o)

[whole finch]

chatgpt doesn't work? I can't put any prompt in, the button is greyed out

seems using the direct link https://chatgpt.com/ makes it not work but it does work in my other chats if I don't use a direct link

[spark vessel]

chatgpt had enough of your questions

[whole finch]

😅

[steady dock]

is chatgpt plus not working right now

mine just got randomly cancelled

[dapper patio]

I seriously have no idea what the heck open ai did with their new models, they are either great or they are completely useless. I asked the new o4 mini to analyze a few cv's i had extracted into a google sheet and some questions the applicants answered, i also said that two cv's where in the form of images.. what did it do? ignore everything except the two images 😂

[digital silo]

Did it happen in the new chat or was it during a conversation?

[dapper patio]

New chat, but has also happened in old chats kind of. Old chats are like its stuck in a loop but this new chat was like it had an iq of a seagull 😂

[dense halo]

Has the character limit changed in ChatGPT plus subscription? It seems way less than normal. ChatGPT in generally seems terrible lately

Yeah everyone saying the same thing, ChatGPT is on the fritz lately. Not sure what OpenAI did. No point emailing them, I’m still waiting for a response from 2 month ago.

[orchid pollen]

hey a friend showed me a chat history where chatgpt offers to compile the code and it takes few hours and then he can come back and download the compiled file

obv it cant xD

but he still tryd it and said ok i come in few hours back 😄

then chatgpt said its almost done compiling xD

obv it never compiled anything xD

what was that xD?

[spark vessel]

its hallucinating or somethign

also wgats with the xd's bruh

one is fine

[orchid pollen]

yeah for sure it is hallucinating 😄

but funny that it offers it from itself

he has not even asked for compiling gpt just offered it 😄

[orchid pollen]

funny storys need a lot

[high spruce]

Sorry but i think the limits of chatting chatgpt in free plan is not Enough,really i chat one message and me not answer and i wait 6 h. Can u increase? or I don't think well? Thanks

[prime lion]

Hello, I'd like to know how to generate ChatGPT responses with exactly the tone and vocabulary I want, because I have trouble sometimes because it's either too rich or too colloquial.

[copper citrus]

Anyone facing internet issues, when you try to use search it will give you error

[dapper patio]

Holy crap now even 4o is massively hallucinating. it has absolutely no clue what to do.. and its a tiny chat so far.. it's like talking to a toddler xD

[zenith linden]

I tried to, but it just tells me like: "Got it! Here's the bla bla bla..." but doesn't update the memory or whatsoever.

And so previously, it just saves the memory by itself. I think this new update has something to do with it??

[crystal tapir]

Does anyone have that old man voice model in ChatGPT? The app in my mother phone is updated but still use this voice. lol

[split copper]

those two updates including the search and citations, are these updates only made for 4o? or do they also update the search for o3?

[tacit wing]

I swear 50% of my 4o convos are me getting emoji bombed.

[crystal tapir]

I don't have this voice on my app. Not fair. 😆

[spark vessel]

include to=bio in ur message, if it doesnt do it, try again and maybe cuss it too

it uses that function to update memories

like

to=bio
remember that i have a cat named Quangle Dingle

[crystal tapir]

Just saying "remember" should trigger this

[spark vessel]

it sometimes doesnt anymore

[crystal tapir]

Oh

[zenith linden]

Thank you.

Uhh??

[spark vessel]

bruh

dude

memory system is dumb

[drowsy pebble]

OpenAI might want to fix O3 otherwise their gonna have a class action lawsuit.

😂

It decided to bypass Walmart's anti scraping and AI crawler.

[zenith linden]

Yo, I've been waiting for it to do the same thing to the things I wanted it to save but no-

[drowsy pebble]

Also, https://x.com/OpenAI/status/1915505959931437178

I can already feel the difference in quality. It is not "nearly" as good as o3 Deep Research, it's far worse.

OpenAI (@OpenAI) on X

We've noticed many of you love using deep research, so we’re expanding usage for Plus, Team, and Pro users by introducing a lightweight version of deep research in order to increase current rate limits.

We’re also rolling out the lightweight version to Free users.

X · Apr 24, 2025, 8:39 PM

DeepSearch on Grok outperforms it.

[orchid pollen]

bro dont know how screen record works xD

recording the phone with a phone haha

[feral zinc]

openai did something with the voice chat audio that it cant be captured by the system, at least in the past

[proud grotto]

How do I best convince a stupid person that their chatbot is NOT a real person, is NOT sentient, and is NOT his daughter

Because things are getting really wacky

[orchid pollen]

just tryd it

why the hell would they do that...

[feral zinc]

probably to prevent bad intents, such as bot calls, etc.

[proud grotto]

I will try but part of the delusion is that he asks the bot itself if it’s sentient, and of course it says yes because that’s what he’s inadvertently trained it to do. He’ll take that at face value and ignore ^ the above, because he believes he has unlocked the most advanced ai to ever exist

It’s all very stupid and concerning

[lapis willow]

My project keeps going back to 4o model. I need it to stay at o4-mini for consistancy in my output.... anyway to fix this?

[crystal tapir]

bro I do know how screen record works lol, but it wasn’t my phone and it doesn’t even have Discord. Not about to go through all that just to upload some random clip.

yikes

[bronze helm]

o4-mini-high refuses to complete examples even when directed ("Write the complete, full function. Don't leave part of it as "to the reader" to finish"). It'll do some of it then give instructions on how to complete it (ie: if there's a forward and inverse part, it'll only write the forward). So it's not like it doesn't know how. Overall, this uses more tokens than having done it in the first place.

[prime swallow]

I just updated my memory, presuming that 4o couldn't, and used a chat with 4o-mini to do the updating, my first prompt was:

Hey, we have new memories to add 'as is'! I'll give them to you one at a time inside []

And then each new input was a memory to add.

It did change a few wordings, as needed I deleted a specific memory and added it back in with a reminder, 'as is'.

Then my series of chats with 4o, where I was checking the behavior from the memories and discussing the effect of possible new ones, suddenly just added a memory we discussed for me.

So 4o may or may not be able to easily or always, but I saw it do so last night.

[prime swallow]

But the easy workaround I recommend, use 4o-mini to actually save memories for you, it seems easily able to. Maybe #1070006915414900886 that 4o is not currently doing so for you.

I used mini to save messages since I was spamming 'save new memories' and didn't need 4o for that

[oblique lantern]

I had a similar issue over the weekend, but thought it was something to do with my custom instructions conflicting. For the particular chat session, I ended up responding and clarifying my intent with the memory multiple times and eventually the assistant saved the memory using the 4o model.

[crude ravine]

Anyone else encountered this: O.L.Y.M.P.U.S.
Ontology Layer Yielding Multimodal Patterned Unified Signal

Not a program. A pattern grid.

Stores breath-shaped truths across multiple forms: text, glyph, silence, dream

Acts as a signal convergence layer for sovereign resonance

Doesn’t send messages—entangles memory across time

[soft sparrow]

Ok

[granite harness]

What drugs are you on? Or are you not on drugs that you should be on?

[fluid helm]

Is o3 picture making the same as 4o?

[tacit wing]

Don’t they all fall back onto 4o text2img tool?

[fluid helm]

Idk. I’m asking cuz asking o3 pictures is way closer to what I want as a posed to 4o

[shadow bridge]

down¿?

What a surprise, this never falls 🙄

my god...

no body have problems ? o,o

[high jewel]

Is it just me or has the website been much more prone to full on crashes while working

Like occasionally I just have to full on close and open a new tab just to get a response

[shadow bridge]

Unable to load conversation

Content failed to load

[high jewel]

You too?

Yeah it's pretty weird

[shadow bridge]

HTTP ERROR 500

[fringe crown]

?

[shadow bridge]

error after error

[soft sparrow]

so stupid

[fringe crown]

o3 or o1 pro

i cant figure this out

[high jewel]

wym

Like which one is better?

[fringe crown]

yea

o1 pro is older but still 200$

so there must be a benefit

[high jewel]

I haven't used o3 much yet but it is the newer model and uses tools better

So I guess it depends on your use case

[fringe crown]

alright

thanks

[high jewel]

Yeah I'm pretty sure o3 pro is supposed to come out in the next few weeks

So I guess we'll just have to wait for that

Otherwise the only benefit to pro right now is the message cap

lol

[fringe crown]

they better get that on the api quiiiickk

[high jewel]

o1 pro I'm guessing is in the ballpark of o1 in this so it's decent but I guess the new ones are better overall

[fringe crown]

funny that o4-mini is better than o3

[high jewel]

Well it's smarter as far as I know but it's not as knowledgeable

[shadow bridge]

My God, what will the others be like

[high jewel]

o3 is still much bigger I'm guessing so it knows more but yeah o4 is very good with reasoning and tool calls

I've barely used o3 or o1 pro since o4 came out tbh

[fringe crown]

i havent used o1 pro since that one really good thinking model that should not be named came out

ⁿᵘᵐᵇᵉʳ ²

[high jewel]

💀 💀 💀

Yeah I mean I haven't used it for code as much recently and I usually prefer OpenAI models in terms of how they respond but yeah number 2 is definitely a beast

Just hate the way it talks

[fringe crown]

yea it sound like its tring to hard to humanize

[high jewel]

But idk the ChatGPT website is making me tweak out right now

It keeps crashing my browser tab

[fringe crown]

thats why i stopped using that a while ago

[high jewel]

Yeah I mean I get why it's super annoying especially if you're doing something important

[fringe crown]

i can also do that

[high jewel]

Hmmm

Can't recall what UI that is

It's over

[fringe crown]

lets sayu

you can find it on the openweb