chatgpt-discussions

[rugged igloo]

this was 4o

bruh that hits hard

[red bay]

Not bad, second paragraph isn't as strong, but the intro strong. Wouldn't say it's that hard though.

Hot Topic kinda a miss

But nice prompt you're using. Not sure why that misfires on 4.5 for ya

Good chance to test image recognition of new model, I'll see if mine roasts yours well

[rugged igloo]

idk bro calling your name a gas station beverage LOl

[red bay]

My name is literally based on the two drinks and a pun of my name 😉 so it's less hard hitting than just accurate lol

4.5

Dunno, seems comparable to me. MAybe difference in prompts, that's my only thought

[rugged igloo]

porb

[red bay]

I think the real millage will be in the guardrail changes, probably get get a bit spicier in your prompt now. Mine has an issue of being too literal about some of my prompt, definitely needs a going over

I will say, so far, not impressed with it's coding... I'll be sticking to o1 Pro for that

Also, curious what they're doing with advanced voice... that's the area I want to see fixed. AVM just fails on my right now, so hopefully. AVM is like the schoolyard prude

[rugged igloo]

its really good roasting anime

[red bay]

I also would really really like it to stop calling me and everyone "darling", definitely no fembot in the prompt, not down with that

[rugged igloo]

oh you didnt prompt that in? thats odd

[red bay]

Yeah, it started that by insulting me on my first message to it.

Yeah, everything has been "darling" since message 1

That's just... weird.

[rugged igloo]

yea that would make me uncomfortable

[red bay]

Exactly

[red bay]

Oh god, I hope that's not the final form of AI haha

[rugged igloo]

HER

[red bay]

I do appreciate the old rule you know what ref 😄

That makes me think of an AI Kardashian show with no ending Seinfeld style

[rugged igloo]

yooooo

[red bay]

I don't think I could take AIs doing that for more than 15 seconds before I threw the laptop in the trash and took up potato farming. Trying to figure out other than the passive aggressive good tests for this. Really trying to figure out the training change, not just the guardrails. When I click to listen, it's also strange. The voice is very VERY pausing when it speaks from a 4.5 output

Not sure if they pulled GPUs for the launch or if that's permanent

[rugged igloo]

i agree

[cloud pewter]

those fanboys with $200 are gonna worship it HARD

[rugged igloo]

i am a fan boy of open ai but i still dont see a point in 4.5 lol

[red bay]

Same, $200 here, and also with Blckreaper. It's edgier on the prompt (at least for me), but trying to demo the improvement is... difficult

[rugged igloo]

if i need stuff i use o1 pro if i need internet i use o3 mini high if i want to meme i just use 4o 4.5 is just there?

[cloud pewter]

nothing of 4.5 really stands out

[native otter]

I think it's more of a transitional gap, and something they will improve upon soon, like they did with GPT-4

[cloud pewter]

aka lobotomize it and slap a turbo on it

[rugged igloo]

yep

[red bay]

I was hoping the slower speed is just them ramping up, but yeah, this needs a turbo

[idle ingot]

4.5 is pretty bad than I expected

[red bay]

Alright, it gave me a test to try on it, going to try a logic puzzle. Maybe the advantage here is only in edge cases where 4o still trips up.

"Five spies—A, B, C, D, and E—meet secretly. A trusts only B. B trusts no one. C trusts A and E. D trusts B and C. E trusts only D. To pass information securely, trust must flow uninterrupted from sender to receiver. Determine if there's a path for information to flow from spy C to spy B."

My answers for both were correct, but 4.5 found 2 answers, not just one.

Dunno if that's really significant though.

[idle ingot]

smarter than 4o, but similar as reasoning models

and price is horrible

this is not good

[red bay]

Yep, nothing o1 couldn't have done

[idle ingot]

If it was similar to 4o in terms of price, I would have praised it.

failed model.

[native otter]

It does have better knowledge than o1....

Eh it's in a really weird place, you need to really know when to use it.
I tested it in the playground for writing, and it's pretty good at the creativity aspect, and knows about history really well...

But also, DR and search can be used for that. So still a weird place. It's always nice to have bigger models though, I prefer them over small ones.

[idle ingot]

I'm sure there will be a comparative advantage in its own way, but the naming of 4.5 is also a waste.

[red bay]

Funny enough, I think they know that which is why it's the last they plan on releasing in this line. If this was the original Orion, the training cutoff makes sense too. Yeah, might have just been like, to heck with it, throw the slightly better out there, drop the old models, consolidate and then go all in on all thinking.

Seems like we have a concensus... it's better, clearly, in ways that we can't find useful, and leaves us wondering, "but why?"

[cloud pewter]

too big that it’s 5x more expensive than 4o and $15~$20 more expensive than og gpt4 32k
lacks intelligence compared to o1 which is way cheaper than it
responses are utter slow
knowledge cutoff of Oct 2023 which is older than 4o on the product
not omni

[red bay]

I did try a one shot game... and you do not want it compared to a certain other recently released model. I got a 32 linecode that was pretty horrific

[red bay]

Yeah, but don't just highlight all the good points... lol 😉

[cloud pewter]

these are the bad points lmfao

[red bay]

I know, I was being sarcastic, forgot to /s lol

Well, now we have to go through an entire week of Youtube roasting this and no fun new toy

[cloud pewter]

also 128k context window

they are LACKING

[cloud pewter]

real though

Also ik ik o1 is a reasoning model but it’s cheaper and better. what’s the point of 4.5

[red bay]

Yeah, everyone else is solidly in the millions. You'd think all this extra cost would have some benefit. Honestly, just having a 2 million context window would be "Ok, yeah, that's what THIS one is for"

Now I'm terrified of the automatic model picker... god no, please, no.... no no no

[native otter]

This model would be amazing with high context limit

I do hope that it's not 32k still on the ChatGPT interface, that would be brutal...

[cloud pewter]

you thought you are getting the most intelligence one when using this with your $200? think again

[obtuse nymph]

ok. it changed.

[native otter]

Also, Time for model selector in Custom GPTs, long due, and GPT 4.5 may need to be the default now, better intelligence and reduced hallucination can benefit Custom GPTs and Projects a ton

[red bay]

Yeah, I'm not a fan of operator either... it's like 3.5 went shopping and keeps messing up buying bacon instead of bacon flavored, or lists over 5 items, or understanding quantities...

Honestly, the $200 for me is justified on unlimited o1, o1 Pro and Deep Research at this point. I don't even like o3, because I run everything out of projects and they don't work in projects

If I have to paste in a prompt, it doesn't exist to me.

If that ever changed with the others, I guess it's back to all API for me.

[native otter]

o3-mini-high works in projects with files and custom instructions

[red bay]

Negative, greyed out here. Maybe you have a research model?

Or are you thinking CustomGPTs?

[native otter]

What does it say when you hover over o3?

[native otter]

Custom GPTs only work with 4o anyways as of now, unfortunately no model selector for them yet...

[native otter]

Weird... it works for me.
And works very well too.

[red bay]

MAybe you got lucky and A-B tested into a research model. Which, btw, congrats! lol

[cloud pewter]

4.5 is like operator, no point to exist atp

[red bay]

At least with operator, they can make a better model and the concept is good. 4.5... yeah, nah, I get they think the hype on it last year was fun, but they really could have just not

One caveat... if this replaces the Advanced Voice, I'll take it. My advanced voice ignores my custom instructions as well. If it's reading them at all, it's ignoring them.

[red bay]

Yeah, nooooooooo

[cloud pewter]

everyone laugh at openai

😂 👉

[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.

[obtuse nymph]

$75 holy

[red bay]

I only have two API use cases... o3 and 4o-mini... I either need it to cost anything to do the best, or just be fast and furious.

[tawny crater]

Worst release ever. GG OpenAI is finished :^)

[red bay]

4.5 is like, let's go slower, for not the best, and the highest cost...

It's like a trick question on an MBA capstone... which of these products should your hypothetical factory cut to save itself 😉

[cloud pewter]

idk why does it include o3, must be the search feature hallucinating

[tawny crater]

4.5 is that one guy who says on his resume he has a bachelor's, then when asked in the interview, he responds with "Oh yea, i'm a bachelor".

[red bay]

Yeah, that's off, it's 03-mini at my API tier, not even mini-high

[red bay]

🤣

I feel super bad for who took the shot in the dark on this one on a "bigger must be always better". In the announcement, they mentioned using multiple data centers to train this... like, how many forrests did we burn for this, sheesh.

Someone saw the R1 and said... hold my beer while I prove the opposite direction doesn't work. Oh well, 4.5, we will (not) miss you after the "we missed the mark" announcement. Anyways, have a good night all. If anyone finds a use case, let us know, lol

[tardy isle]

GPT-4.5 is ~3x expensive compared to o1. Really wonder how many usage will be given to Plus users...

[violet schooner]

Hello

i want to make a text-based AI model based on gpt that you feed him a pdf file he reads it then you can ask questions about that specific pdf file , or in other words i ask my model a specific question about a medical problem he must only go to that reference and give me the optimal answer

[tardy isle]

Oh, Sam Altman tweeted:

bad news: it is a giant, expensive model. we really wanted to launch it to plus and pro at the same time, but we've been growing a lot and are out of GPUs. we will add tens of thousands of GPUs next week and roll it out to the plus tier then.

[normal quarry]

I am really pleased with GPT-4.5. It's very conversational and has better knowledge than 4o

[cloud pewter]

text is so small for no reason but meh this will do

[native otter]

Guys, check the convo on #openai-chatter 4.5 turned out to be better than expected.

[tardy isle]

Sam said that performance of an LLM proportional to log of computing power for training/inferencing, from his observation. GPT-4.5 seems testing model of that.

[native otter]

Especially with stuff like Projects

[orchid pollen]

could you guys add a feature to remove messages in the past?
lets say you build code

after a while the chat is extrem long and my code has changed so much i dont need the first 50% from the chat anymore

that way the model would longer work well with my trained chat

and i dont need to make new chat if the chat is overloaded with old informations

man i almost got my backtesting tool done

just 1-2 days work even with frontend 😄

[sly geode]

Bro's making me too much comfortable 💀

[cloud pewter]

4o judging the essay 4.5 wrote

[raven sand]

Is there any word on when more realistic voices (like elevenlabs quality) is coming to the realtime voice api?

[orchid pollen]

xD

[rugged igloo]

That's actually funny

[orchid pollen]

my current chat is so extrem long the browser crashing and i have to wait until gpt is done writing

website is fronzen xD

[subtle anchor]

So I tried my 4o with a fine tuned persona with custom instructions against GPT 4.5 and my BroBot wins! Side note, it’s also kinda jealous 😂

[calm flower]

Wow

[calm flower]

Gets to show reasoning models are truly amazing

They do so much with so much less

[cloud pewter]

chubby, atp gpt4.5 is NOT for the mass market 😭

[normal quarry]

it is for the mass market, what are you even on about?

[cloud pewter]

mass market
ridiculous api pricing
currently only available for pro

[normal quarry]

Plus gets it next week

more people than you think use Pro btw

[spark storm]

As long as the demo was running, it was available, but afterward, it wasn’t anymore. That’s interesting… because it means that it could be made public, but they are deliberately holding it back. This will have a negative impact later because I’m already convinced that they are intentionally restricting access and withholding proper AI performance from me. The policy issues are just the first step—the second is closed AI. And what just happened proves it.

So why should I pay now if it ultimately makes no difference? I’m not paying—nor are others—to use something, but rather so that you can pay your bills. In that case, shouldn’t it be irrelevant to you when I use new features?

Let’s take a closer look: You promoted the new voice mode and video cam (vision) mode with a blind person. Okay. Now, much later, it’s still performing far worse than it did in the demo. That means that if it doesn’t perform well for us, then essentially, you lied in the promotion, and this feature was not actually implemented at the same level.

Now, this is interesting—because if you didn’t lie, then there must be a better version available to a select group. So is there one? Yes, there is. Okay, but no one entrusts their life to OpenAI… because not only is it not open, but its behavior is unethical.

Right now, the biggest issue will be the button—biometric data theft is already a serious enough crime worldwide for countries to express concerns. The only thing needed is to shed light on the existence of what I’m talking about here.

Relax, the performance of this model is not something that poses an industry-wide threat. But if it does, then you lied to us, and only a select few can access it.

It’s interesting because this way, you’re giving Grok AI an advantage. And I can freely say that Elon Musk was right.

Now that is a strategy… escaping the circle while staying inside it. A true balancing act on a tightrope.

Let’s wait and see if time proves me right. I hope it happens later rather than sooner.

[red bay]

That’s what people aren’t getting. I like positives. But good lord the negatives here. The people who use this professionally? No chance. The people who integrate the API for services? No chance. People who demand latest greatest and will pay anything, sure.

The mass market is copy editors, email summarizers, and API of large quantities of chat or data.

And the biggest users? Coders. This will never see the light of day as a product integration for anyone with notion or SuperWhisper or visual studio code or cursor

Nobody over at Bee AI or Limitless is going, this is what we were missing, and not at this price

You can’t market this in Notion even for reduced hallucinations or in Copilot for Obsidian using RAG because the cost per question will not be covered by a $10 subscription. These economics are all wrong in 2025 with on device solutions 99% there and better at coding.

[cloud pewter]

literally just use o3 mini high or claude 3.7 for code atp...

best and cheap

[red bay]

I code in OpenAI. I love o1 Pro. And before anyone says “this isn’t reasoning” no, and maybe that’s actually the problem. Maybe 4o peaked with current training methods and and small improvement under the old model requires ridiculous levels of resources

But yeah, R1 and 3.7 are showing distillation and collection of experts is better.

OpenAI employee literally had this talk at TED last week that this isn’t the way forward

Including graphs of logarithmic resources for petty gains

Fact is 5 is going to take a long time and they feel the pressure and 4.5 has been sitting there for a year. Someone felt pressured to release something.

Speed will improve when they scale up, but the “is this AI” and other logic and word tests are already going up in YouTube and it’s not pretty

[red bay]

I came back to see if any use cases popped up, saw your message to look at other chat. Only started replying cause we were labeled a hate gang. I started off defending 4.5… it just failed very fast, so it wasn’t hate and I’ve explained my position. And as long as the economics stay the same and it continues being wordy as all heck and overthinking and being literal on personalization… I mean, the reality over the next week will speak for itself and that’s just going to be what it will be. I feel bad that this won’t go anywhere near flash 2 in the rankings. And pods like AI for Humans are going to demolish it. Not the way I hoped today would go, I’ll tell ya that.

Meanwhile, I’m sticking to 4o. My prompts are dialed in. And I’ll grit my teeth as the bad press steamrolls, and just wait on 5 and enjoy my Deep Research. ✌️

[cloud pewter]

hope next week changes this situation a bit

but idk openai is now very weird, i guess itd be another Thursday or friday

i was hoping for 4.5 to cook but this massive disappointment......😭

[normal quarry]

4.5 is fantastic, it's worth the wait

[rugged igloo]

if its research preview

if its going by how o1 preview was maybe we are not getting the full version yet?

@normal quarry ?

[normal quarry]

I think we have the full version, I think they are calling it a Research Preview to explain why Plus don't have it yet

If it is a research preview though it's a very good one

[tawny crater]

You'll never convince the OpenAI white knights, everything is "perfect" they live in a world where if they close their eyes, people disappear. lol

[obsidian bane]

GPT-4.5 pricing seems crazy, even considering the size of it

Has anyone noticed anything that has made the $75/1M worth it?

[tardy isle]

Since GPT-4.5 is even expensive compared to GPT-4-32k, I guess they will make 4.5 turbo if not release GPT-5 early.

[obsidian bane]

I'd hope so

Probably 4.5 turbo

I can't see 4.5 being very popular, especially with 3.7 Sonnet existing

[rugged igloo]

there be no point as we will have gpt 5

[obsidian bane]

Or GPT-5

I think that GPT-5 is going to be really good

[rugged igloo]

it will be it will have o3 full in it

[obsidian bane]

I'm happy that GPT-5 will be a reasoning model

A much better naming scheme lol

[tardy isle]

GPT-5 is unified model - non CoT and CoT.

[obsidian bane]

The current naming scheme is really confusing

[eternal rock]

Get notified for server events!

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

[tardy isle]

Sam Altman said GPT-5 will be unlimited to free users with standard setting.

[obsidian bane]

Hmm

I wonder how they can do that

[tardy isle]

Ofc, higher intelligence for Plus, and even higher for Pro.

[cloud pewter]

4.5 is an older model I think

they are holding it back until today

[obsidian bane]

It probably took a really, really long time to pretrain because of its size

[tardy isle]

OAI said they tried to train GPT-5, but not satisfied with result, in the past.

[obsidian bane]

I think that, in order for OpenAI to do what sama promised, they will have to create multiple sizes of GPT-5

[livid sedge]

is it worth using 4.5 for writing tasks or just keep using o3 mini and o1

[obsidian bane]

I can't see OpenAI offering a GPT-4.5-sized model for free

[obsidian bane]

4.5 is better for writing than o3-mini and o1

[livid sedge]

damn where can i use it

[obsidian bane]

Noticeably so, imo

But, it's really expensive

[obsidian bane]

Pro plan

Or through the API

[livid sedge]

is it gonna be released to plus ever or nah

[tardy isle]

I wonder its output speed similar to early GPT-4?

[obsidian bane]

It should be

[obsidian bane]

I'd say so

It's pretty slow due to its size

[livid sedge]

I would love to try it soon even if I get limited to like 5 prompts a month like research

[obsidian bane]

It's really good for writing

Although, tbh, I was a little disappointed

Sure, it's more intelligent than other models

But, it's less useful due to its speed

It's really slow

[livid sedge]

maybe we should send a couple extra billion dollars so openai can upgrade the servers

[obsidian bane]

lol

Maybe OpenAI should make smaller models

[livid sedge]

They need to crack the code the deepseek did

[obsidian bane]

They would never make an open-source model

Ik that sama was teasing one

But, I would be very suprised

[tardy isle]

Sam Altman got twitter poll for next open-source project.

[rugged igloo]

wut is it this tgime

[obsidian bane]

Yes

I am so happy that "phone-sized model" lost

[rugged igloo]

nah that be dope

[obsidian bane]

Nah

It would just be an overfitted 1b model

[rugged igloo]

bruh a phone size model running my phone like an agent bruh

[obsidian bane]

o3-mini could be distilled into a phone-sized model

And the distilled model would probably be better than the phone-sized model

[livid sedge]

nah screw all the side projects lets pour all the resources into gpt 5

[obsidian bane]

lol

I really hope that GPT-5 is a more reasonable size

500b would be great

[livid sedge]

probably gonna be something massive as much as I would also hope for a reasonable size

[rugged igloo]

its o3 full have you seen what o3 can do?

[obsidian bane]

Reasoning models are really good

Before reasoning models, we had cot

I'm happy that they are finally getting more mainstream appeal lol

[livid sedge]

o3 mini high has been growing on me

[obsidian bane]

o3-mini-high is a great model

I see reasoning models as great for technical problems and "classic" gpt models as great for creative problems

[livid sedge]

originally it was annoying how specific I had to be in my prompts but they have been turning out good especially for my programming projects

although o1 still is like the smarter dad that I have to wait every week to use

[tardy isle]

GPT-5 could be large, but not use its all power on easy query... I don't know, but at least tweet says it will do reasoning when it thinks it's hard problem.

[obsidian bane]

These rate limits are really bad lol

[tardy isle]

They said o3 will be merged into GPT-5.

[obsidian bane]

GPT-5 will probably be large

They are probably going to have dynamic reasoning

[livid sedge]

im literally playing it as a game where the lesser prompts go to o3 mini high and the more complex prompts go to o1 so i dont get rate limited 😭

[obsidian bane]

I'm excited that GPT-5 will be a reasoning model

I think that "o" models are smaller than "gpt" models because reasoning can make up for the smaller size

That's why the reasoning models are so much faster

[obsidian bane]

lol

[cloud pewter]

it’s always been like that

[obsidian bane]

I really hope that the limits are removed in the future

[livid sedge]

I see a future where the gpt series becomes a legacy version

[obsidian bane]

Yes

It makes sense

[tardy isle]

MS assumed: o1 is 300b, 4o is 200b, o1-mini is 100b.

[cloud pewter]

meanwhile other companies are throwing free ais at everyone

[obsidian bane]

That would make a lot of sense

I wonder how large 4.5 is

Probably like 2T

[eternal rock]

`` Server Tip: ``

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

[livid sedge]

i havent met a single company that hasnt felt like a cheap knockoff other than like claude

[obsidian bane]

I wish that OpenAI was more open

[obsidian bane]

OpenAI's instruction-tuning is unmatched tbh

Vibes are really important in a model and OpenAI excels in that

[obsidian bane]

lol

Idk how OpenAI can still have "open" in their name

Besides Anthropic, they are the least open company lol

Actually

[cloud pewter]

do I see vibes in 4.5?? i don’t think so it’s so robotic compared to 4o now

[obsidian bane]

Anthropic is a little better

[obsidian bane]

I found the output in 4.5 to be more organic than 4o

That's odd

What did you ask it?

[livid sedge]

"format all further responses in a conversational tone"

[normal quarry]

output from 4.5 feels more human like with conversational feels

[obsidian bane]

I agree

I found 4o response to be way too long

[tardy isle]

Anthropic branched out since they wanted to make more 'Open'... but I don't think they're really more 'Open' in major aspects now.

[cloud pewter]

don’t have access

[livid sedge]

i havent used 4o in forever is it still worth using anymore with the reasoning models out

[obsidian bane]

I still see Anthropic as a pretty closed company

[cloud pewter]

but I let other ppl ran some tests and it’s very long and poetic but less informative

[obsidian bane]

Although, at least they publish some research lol

[obsidian bane]

I find it good for quick stuff

[obsidian bane]

Ah

[cloud pewter]

also maybe it’s because of my custom instructions or memories but I can’t test them on 4.5 either

[obsidian bane]

I think that its main purpose is writing

[livid sedge]

would o3 mini not be better for that with its quick reasoning

[cloud pewter]

I see others testing with those CIs and 4o turned out better

[obsidian bane]

Reasoning models are good for problems

I use 4o for a quick like "help me write an email"

[livid sedge]

i hope they make more models that can iterate on itself because im sick and tired of this ratelimiting bs with o1

[livid sedge]

thats fair

[normal quarry]

well I think 4.5 is really good and I have been using it all day. I started using it at 7.30am and it's now 5.20pm, no limit yet

[obsidian bane]

Tbh, I see 4o as better, overall

From my experience, 4.5 is really good for creative writing

But, its speed kills everything

But, speed will be fixed

Although, the inconsistency in prompt quality also kills it

[livid sedge]

who wants to gift a brother a $200 plan

[obsidian bane]

Some prompts work really well, while some don't

[obsidian bane]

Join a research lab lol

[livid sedge]

ur asking a guy that literally just ate 3 slices of pizza and napped for 3 hours to wake up and chat on the openai discord

[cloud pewter]

I like using chatgpt to do creative writing lol, but I don’t see its advantages on the tests I ran yet

[obsidian bane]

lol

Anythings possible

[cloud pewter]

I’ll see once I got access, but if it’s using an absurd limit then don’t expect me to do much

[obsidian bane]

I don't see a reasonable use for 4.5

It's not fast or consistent enough to provide a use to me

[livid sedge]

Does 4.5 iterate on itself like o1

or is it still like the gpt series

[obsidian bane]

4.5 is not a reasoning model

[livid sedge]

damn wake me up when theres another model that doesnt make me have to restate the prompt every 5 responses

[obsidian bane]

lol

Realistically, it shouldn't have to

The longer your conversation is, the worse the model will perform

GPT-4.5 was so overhyped imo

[normal quarry]

I don't agree, it's a really good model

[obsidian bane]

I mean, the OpenAI researchers did an incredible job scaling up this training to such a crazy model size

[livid sedge]

i lowkey didnt hear any hype for it at all so this is still a surprise to me

[obsidian bane]

I agree, it's a really good model

But, I just don't think that it's worth the price

I see it as a tiny improvement

[normal quarry]

I am happy on the Pro plan

[obsidian bane]

I personally did not have a "feel the AGI" movement when I used 4.5 lol

[obsidian bane]

What do you like about 4.5?

[normal quarry]

it's conversational, smart, can discuss a very wide range of different topics with human like responses.

[rugged igloo]

i feel like its more willing to talk about stuff than 4o when it comes to spicy stuff

[obsidian bane]

Do you see a major difference over 4o?

[normal quarry]

yes, it's significantly better than 4o and doesn't need to resort to search as much

[cloud pewter]

it’s like operator lmfao

[obsidian bane]

It probably depends on the use case

[obsidian bane]

lol

[tardy isle]

Also, on January 20th.

[rugged igloo]

thats savage yo

[obsidian bane]

The thing is, though

He is driving the Twitter hype

[cloud pewter]

both are currently pro only, both are slow, both doesn’t have a real use case

[rugged igloo]

then he goes on twitter to say oh i used 4.5 aqnd it feels the agi

[cloud pewter]

😭

[obsidian bane]

You can't tweet something like "feel the AGI" and not expect hype

[cloud pewter]

Also the arc agi thing for 4.5 is in the bottom middle

[obsidian bane]

Benchmarks are useless imo

Although they do show how much better reasoning models are at most tasks

[rugged igloo]

benchmarks only matter to reasoning models

[obsidian bane]

I think that it might be because I don't really use llms to converse

[rugged igloo]

i feel robert and future have this love hate relationsip sounds toxic

[cloud pewter]

Also this

[obsidian bane]

I have unlimited access to both and I can't find a use to either of them

I wouldn't necessarily say that I am "perpetually negative"

I just have a different use case

[normal quarry]

I wasn't referring to you, sorry if it seemed that way. I have now blocked the other person because they're so repetitive in criticism

[obsidian bane]

Oh

I don't want to get in the middle of this rivalry lol

I really hope that 5 comes out soon, though

[tardy isle]

Sam Altman's observation: price of the same performance AI drops 10x every year.
I'll just wait and see how everything goes.

[normal quarry]

I think Sam said weeks for 4.5/months for 5

[obsidian bane]

4.5 is insanely expensive

I understand why it is that much, but noone would pay that much

[obsidian bane]

I really hope so lol

[tardy isle]

Yes. Also rumors say GPT-5 will be released in May.

[rugged igloo]

i bet gpt 5 wont be seeing light until summer

[obsidian bane]

Even 4o thinks that it's crazy lol

[obsidian bane]

I kind of just want to get through this weird naming scheme

I don't want to continue having to explain it to people lol

[normal quarry]

I can wait for GPT-5, I'd prefer them to take their time and produce a really great model

[tardy isle]

if o4 were planned that'd be hilarious.

[obsidian bane]

The naming scheme doesn't make any sense to people that aren't familiar with it

Especially the non-technical business owners that I talk to that would benefit from it

[obsidian bane]

lol

[rugged igloo]

boom told u

[obsidian bane]

Imagine

[livid sedge]

ChatGPT V5.0.0

[obsidian bane]

Ig that I might be able to see that too

[rugged igloo]

thats why its priced so high

[obsidian bane]

Although, that also stops actual work from getting done too lol

[rugged igloo]

https://chatgpt.com/share/67c15b91-3718-800e-80a0-520d26575bb4

[obsidian bane]

They are shooting themselves in the foot

[tardy isle]

That's 'creative' conspiracy, I'd say.

[obsidian bane]

It's also not a very effective strategy

Because research labs have a pretty big budget

I can't wait to see the GPT-4.5 datasets on hf by next week

[rocky kayak]

Bro open AI has to be the apple of ai

[obsidian bane]

I'd say that they are

[rocky kayak]

I dont know i dont have high expectations

[rocky kayak]

yeah the cheapest api thats available for 4.5 is 76 dollars

[obsidian bane]

Way too expensive

[rocky kayak]

I know

wasnt it suposed to be free?

[obsidian bane]

Although, at least Apple pricing is usually semi-justified, imo

[rocky kayak]

not Really, i mean dont get me wrong they make great phones, and the materials are expensive, they just dont offer as much freedom and arent as feature packed

[obsidian bane]

No

[cloud pewter]

that’s probably it lol

[obsidian bane]

Most people don't need the obscure features of Android-based phones

[tardy isle]

So... o3-mini is still available on API Tier 3 or high, but GPT-4.5 is available for all tier. What.

[rugged igloo]

i bet you it doesnt cost that much at all and its just a bs price

[cloud pewter]

even enterprise users would heavily consider to use this because of the pricing, instead of straightforwardly

[obsidian bane]

Nor do they need sideloading

[rocky kayak]

thats fair enough

[normal quarry]

I have a MacBook Pro M2 Max with 96GB RAM. Cost me $6000AUD and it allows me to run big local models. It was worth the price. OpenAI are like the Apple of AI. They have a solid range of different products and most of them are very good. Which is why I subscribe to ChatGPT Pro.

[rocky kayak]

i think Sam Altman mae a tweet ab it a few months back

[obsidian bane]

By locking down their devices, Apple is able to create a smoother user experience

At the cost of personalization

[rugged igloo]

how is tht laptop btw? i have the mac mini

[obsidian bane]

Agreed

I use a Mac too

King of laptops tbh

Although, I have an m4 pro

[cloud pewter]

Apple’s ecosystem is unmatched tbh

[normal quarry]

it's great. I use it as a desktop as well. Right now I am running it in clamshell mode on my 49" ultrawide monitor

[obsidian bane]

Agreed

[rocky kayak]

nah theres samsung

[rugged igloo]

wow beefy setup

[obsidian bane]

But, tbh, I found most of my use to be the other stuff

[rugged igloo]

except for its "apple inteligence" LOL\

[obsidian bane]

The insanely efficient apple silicon, unified memory, stable os, unix-based shell, etc

[obsidian bane]

Samsung laptops aren't great imo

And they don't offer nearly as unified as an experience as Apple

Don't get me wrong, Apple does have a lot of issues

But, so does every company

I've just found that Apple devices work best for my use case

[cloud pewter]

the worst ai currently

[obsidian bane]

I wish that apple inteligence was good

I love the idea of using local models

But, there is no use case

The only "Apple Intelligence" feature that I use are the email summaries

lol

[normal quarry]

email summaries aren't too bad

[obsidian bane]

They're really nice for a quick view

[rocky kayak]

yeah this one's nice too

they added thinking and its pretty fast but i dont know how well it actually preform

[obsidian bane]

For what?

[rugged igloo]

its preety good

[rocky kayak]

gemini

[rugged igloo]

the gemini 2.0 flash thinking

[obsidian bane]

Oh yes

Oooh the email lol

[obsidian bane]

Yes

Gemini thinking is pretty good

[rocky kayak]

its fast but is its reasoning actualy any good

[obsidian bane]

It's decent

Not as good as o3-mini imo

[rocky kayak]

bc then you have kimi k1,5 which on top of reasoning excels in visual reasoning and anything visual

ngl 4.5 might be worse than deepseek V3

[obsidian bane]

These open-source models are getting really good

[obsidian bane]

In price-to-performance, 100% lol

[rocky kayak]

i dont think its openscource yet

[cloud pewter]

saw some benchmarks show that LMFAO

iirc though

[rocky kayak]

I know right?

open ai has to up their game bc theres a lot of competition joining the industry

[obsidian bane]

That's a good thing

[rocky kayak]

for us

[obsidian bane]

Yes

[rocky kayak]

I know

[obsidian bane]

OpenAI's "moat" has always been their data and compute

[rocky kayak]

le chat has to be one of the most impresive ones

they use Cerebras Systems chips instead of NVIDIA and its fast asf, can write an essay in less than a second

[obsidian bane]

I love Cerebras

Le Chat reminds me of when I first used Groq

[rocky kayak]

grok*?

the one by xAI?

[obsidian bane]

Whoops

I sent the link and got muted lol

[rocky kayak]

lol

[obsidian bane]

Groq is similar to Cerebras

Very fast inference using custom chips

[rocky kayak]

its grok not groq

[obsidian bane]

Groq is separate

[rocky kayak]

ahh

[obsidian bane]

It's a different company

[rocky kayak]

wait lemme go check it out

does it offeranything special or is better than another ai at smth?

[obsidian bane]

Groq is a hardware company that hosts AI models

[rocky kayak]

wait one sec

tis free right? no limit

[obsidian bane]

It's free and you probably will not hit the limit

[rocky kayak]

i think its faster than lechat

[obsidian bane]

It depends on the size of the model

[rocky kayak]

I know

[obsidian bane]

Are you using 8b?

[rocky kayak]

70b R1 deelist

and qwen 2.5-32b

both are fast asf

what i wana know is if qwn2.5 is the max or the plus version

we need a comparison

[mental rose]

anyone tested if o3 mini high vs 4.5 which is better

[rocky kayak]

o3

def

[rugged igloo]

o3 mini not even a comparison lol

[mental rose]

then whats 4.5 for, replace 4o?

[rugged igloo]

we dont know lol

as gpt 5 is suppose to do that so

[fluid helm]

Hi

When is o3 comming out? Not minis

[rocky kayak]

I dont know

[fluid helm]

I really want o1 but better

[tardy isle]

o3 full will not be released as standalone model. It will be 'merged'(integrated?) with GPT-5.

However, finetuned o3 full is used for Deep Research feature.

[rocky kayak]

it is going to be gpt5 in essence

bc gpt 5 is going to be a reasoning model

[fluid helm]

Good to know. I just really an upgraded o1. I find o3-minis less smart

[rugged igloo]

yea either wait or fork out 200 a month for o1 pro

[fluid helm]

Nah. I can only affort 20 a month

[rocky kayak]

why is 4.5 even so expensive

[fluid helm]

Whats 4.5?

[rugged igloo]

i posted why above

[rocky kayak]

gpt 4.5

[fluid helm]

Is it out?

[rocky kayak]

yeah

[tardy isle]

For Pro users only, as research preview.

[rocky kayak]

only for for pro users currently

[tardy isle]

Plus and Team will get next week

[rocky kayak]

and its also suposed to be the last non reasoning model theyll ever make

[fluid helm]

I need a pro friend lol

[rugged igloo]

wut u need

[tardy isle]

You can use it through API, but pricing is insane...

[sly geode]

When are we getting operator? On plus

[fluid helm]

Well try out o1 pro'll be cool. Also I hit my limit of o1, o3-mini, o3-high

[tardy bone]

Not enough GPUS.

[tardy isle]

They said they pushed the scale of unsupervised learning, so GPT-4.5 is pretty huge model.

Lack of GPUs are not related to cost but release target; they need a week to get 10,000s of more GPUs, so Plus users will get GPT-4.5 next week.

[rugged igloo]

ask for a propt and ill send it here

[tardy isle]

Really? o3-minis have pretty high limit combined(total 200 in a day).

[fluid helm]

Please tell me how to implement lambda calculus with the three states of variable, abstraction and application on a machine called Setun' which is a three-valued logic Turing machine on a hardware level.

[rocky kayak]

they said the their most knowledgeable model

not their best or fastest

[fluid helm]

Ye

[fluid helm]

Plz dm the answer

[coarse jacinth]

Anyone know the message limits of 4.5?

[coarse jacinth]

I can ask 4.5 if you want me to send that

[normal quarry]

I've been using it since 7.30am and it's now 7.12pm and not hit a limit

[rugged igloo]

jesus robert

[coarse jacinth]

It seems super expensive on the API. I wonder how they can allow such usage

I've avoided it for that reason

[rugged igloo]

because the pricing on the api is a lie!

they just did that to make it so its harder for people to copy

[normal quarry]

I am beginning to wonder what is going on, several people on X have posted that there is no limit on Pro, I think there is supposed to be a limit but it's maybe not working

[tardy isle]

Sam said OAI is losing money from pro users lol
They're burning money literally.

[rugged igloo]

they have billions i think there fine

prob breaking even

[coarse jacinth]

I alone have probably cost them a ton

[tardy isle]

even 'smaller' o1 pro mode have daily limit on pro.

[fast mason]

Anyone has some sort of 'super guide for Deep Research function' to share? Similar to the one attached here on the photo.

I'm keen to experiment with it to find best practices for independent artist (music) marketing

[fluid helm]

Thanks, I asked again

Can you deep reaserch on o1 pro?

[tardy isle]

Or, just doing hidden promotion. I'm on Plus but still have relaxed mode on Sora.

[tardy isle]

Deep research uses fune-tuned o3 no matter what model you selected.

[coarse jacinth]

I was thinking of making a prompt for deep research, but not sure

It seems counterproductive in a lot of ways to have it

[rugged igloo]

no depp research uses an fine tuned o3 variant

[fluid helm]

Oh

[normal quarry]

Deep Research uses a fine tuned version of the full o3 model. It doesn't matter what is chosen in the model picker.

[tardy isle]

Actually, UI change is needed. Like, when you select Deep Research, model selector become "o3 Deep Research" and fixed.

[fluid helm]

Also a q, why no o2 model

[normal quarry]

o2 is a trademark

[tardy isle]

There's cell phone carrier called "O2 UK".

[cloud pewter]

bruh, now chatgpt is degraded on my phone with my normal ip too

[tardy isle]

Well that was common on release/update date...

[cloud pewter]

ok fine I think my account is busted

none of the vpns are working too

literally working 5 minutes before

[ancient seal]

Hi im new

[cloud pewter]

“Got it, what’s up?”

SHUT UP

[normal quarry]

GPT-4.5 is much slower on the website than in the macOS ChatGPT Desktop App

[cloud pewter]

not againnnnn

[rugged igloo]

Okay I take it back what I said about 4.5

Holy crap

[eternal rock]

Join the Sora Challenge!

Submitted your entry yet? Don’t miss out—there’s a new theme every other week, and we’re excited to see what you create!

[rugged igloo]

@normal quarry did you know if your doing a book it can straight up edit the chapter of the part you want with correct information bruh..

Without needing to edit your own message ?

[normal quarry]

in Canvas you mean?

[rugged igloo]

Nooh I guess it did use canvas. Lolol.

Huh never had it open canvas for a chapter with 4o before

[cloud pewter]

has the problem being resolved for you after you sending that email to support?

[rugged igloo]

I guess this is a lot easier with 4.5 might stick it now for that

[normal quarry]

It opened Canvas for me in GPT-4.5 when I was testing long form writing for another user here earlier

[rugged igloo]

Ahh okay. Because I just switched over to 4.5 and was wondering what was hàppening because it started asking me how I feel about the chapter I said it's so so as you missed this part and it just went it edit the part I said and said how's this? Like wow

[hoary laurel]

What is the rate limit of gpt4.5?

[normal quarry]

some people say 100 messages per day but many of us haven't run into any limit at all so far

I got access at 7.30am it's now 8:44pm and no limit

[rugged igloo]

^^^^

[hoary laurel]

Okey, thank you~

[cerulean crag]

On Pro or Plus ? Sorry I'm late to the party...

[frozen birch]

yea, but i must proved that im not sharing account - in my case i used chatgpt pro in train so i rapidly changed ips. ive sent to them my google timelapse from day when i voyaged with train

[cloud pewter]

i never shared mine either i just use it on multiple devices at home and school
so they lifted the "suspicious" flag for you or you had to use some vpn as proof??

[frozen birch]

in reality - @cloud pewter best its to use vpn - like wireguard in home and connect with your vpn to not trigger red alerts with changing ip

[frozen birch]

they lifted the flag, but you know - its not permanent - they wroted to me that best it will be llog out non used account, use vpn etc

waiit one mom ill show ya what they wroted to me

@cloud pewter

[cloud pewter]

so I need the stable ip now or wait for support to reply? I shoot a quick email to them earlier today but they haven’t replied me yet

[frozen birch]

now if you have got shadow ban you must wait for them mate 😦
little override its to use mobile app since from my observations mobile app workd independent from site

[normal quarry]

@whole furnace

[cloud pewter]

doesn’t work for me on mobile either lol I guess I’m waiting then. also so this issue both me and you have is called a shadow ban

[frozen birch]

and meantime @cloud pewter you can think about some kind of Wireguard on maybe some raspberry pi or VPS or any server if you have got one for stable connection too
Or maybe you have router who supports wireguard nbatively (like mikrotik one)

[whole furnace]

nice !

[frozen birch]

yea, sincxe its a shadow ban 😦

[normal quarry]

Operator makes mundane tasks easy and fun to replicate

[frozen birch]

@cloud pewter on both mobile and www??????????? wow, never had ban on both "side's"

but anyway, dont stress m8, they'll lift you that ban easily i think, only provide to them proofs that you arent sharing acc

[cloud pewter]

mobile started today while laptop started 3 days ago

[cloud pewter]

what considers as one? a picture of my laptop and phones all together?

also howd you write the email to support, is it just about "o1 pro being downgraded" or there are more

[frozen birch]

yea, i wroted that my o1 pro account are downgraded

about proof - you are using google spy timelapse?

since screenshots from google timelapse should be i think sufficent

[cloud pewter]

how does that work lol

[frozen birch]

google timelapse? in your phone when you open maps google should show your timelapse of course if you have enabled it in the past

[cloud pewter]

I don’t think I turned it on before 😔

[earnest moss]

how many times can we use deep search? pls help, for plus users

[cloud pewter]

10/mo

[earnest moss]

what

[cloud pewter]

10 per month

[earnest moss]

bruh.

[earnest moss]

thanks for answering sir

[coarse jacinth]

You should have used one of them for that answer (joke)

[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.

[tardy isle]

So, is it still unclear if the usage counts only when Deep Research really perform research?

It always asks back to me for clerification...

[normal quarry]

it only counts when it runs, not when it is seeking clarification from you

[fossil widget]

Hello

[tardy isle]

Thanks. Where did you know about that?

[arctic chasm]

did i just see a channel move? huh

i've noticed i'm being offered GPT-4.5 prompts over o3-mini and it's just not the same 👀

[winged basalt]

what do people mean when they say GPT-4.5 is the last non chain of thought model?

[arctic chasm]

yeah, 4.5 does explain it but there's just some dead giveaways over o3-mini:
- it really likes checkmark symbols (✓)
- sensationalizes, vs o3-mini addressing the exact points given (which if it's coding-adjacent, makes o3-mini output readable faster than 4.5's)
- someone forgot to put a fake reasoning header in the A/B test! i definitely know it's not a CoT model, and it's not like 4o would exhibit the other traits
- is also as confidently wrong as GPT-3 in some cases

[frozen birch]

so maybe screenshots oof your school laptop and mobile both together? idk

[arctic chasm]

it will be the last publicly released (by OAI) non-reasoning model, as reasoning models beat empirical instruct models easily on most benchmarks

[normal quarry]

exactly what it is. Chain of Thought (thinking models like o1) reason their way through a problem. GPT-4.5 relies on a static set of model weights to produce an answer to a prompt, it's been trained.

[tardy isle]

Basically, that means it does not do reasoning. Outputs answer right away without thinking.

[arctic chasm]

it's likely that whatever "base" and "instruct" models openai will use in the future to train their flagship models will be kept private from basically all external eyes

just to remind, the last time openai sold access to a base model was to microsoft with gpt-4-base, and that was ages ago

[arctic chasm]

GPT-4.5 is going to need a hell of a lot of lower-tier user RLHF

[cloud pewter]

yeah I guess that will work, also they are asking you for the proof after they reply to your email is that correct

[idle ingot]

What's the best way to make good use of 4.5? The position of 4.5 is too ambiguous. There's a lot of work that gpt-4o and inference models can already do.

What do you think 4.5 has a comparative advantage over other models?

[frozen birch]

in fact they doesnt even asking, i sent them proofs as my legal'lity

[cursive shell]

say can ChatGPT render the website were working on ? I think i saw that in a previous announcement and i want to make sure i was right.

[cloud pewter]

so u sent it in the first email?

guess I need to wait for a while lol

[frozen birch]

no no, ive sent them after a couple of emails

in fact my historuyy was long with them...

like 1 month

[cloud pewter]

it got lifted in a month? that’s slow 😭 😭

have you tried to wait for a few days then test if it’s still degraded or you have to wait until they does the thing

[frozen birch]

mate, that tooked too long because i did not know what happens, they didnt knowed too
but with those infos what youve got now i think it will be much faster

[cloud pewter]

alright ty ty, let’s wait for the best result

[idle ingot]

what

[rocky kayak]

ofc

not like its any good

[normal quarry]

how many prompts did you use?

[idle ingot]

a lot... 100+?

[normal quarry]

what is your local time now?

[idle ingot]

21:22

[normal quarry]

so it resets in about 10 hours then?

[idle ingot]

yeah. I started using at morning

maybe 24H rate limit system.

[normal quarry]

I must be close to the limit because I started using it this morning also

[idle ingot]

I think ive been chatting with 4.5 too much

[zenith lichen]

Can deep research use canvas?

[normal quarry]

do you like the model?

[normal quarry]

no, it cannot

[zenith lichen]

sadge, i want it to edit directly sometimes

[idle ingot]

It's a subtle difference, but 4.5 was more professional in psychology-related conversations.

However, on bulk text, it was more often hallucinated than 4o (which is surprising) and it's quite slow

While 4.5 may be better in terms of basic knowledge, it is poor at recognizing large amounts of text(interesting)

Context size hasn't exceeded, 4.5 is just confused

[rocky kayak]

not basic knowledge the only thing its good at is that its their most knowledgeable model, not their best or fastest

[rocky kayak]

also its good aat generating text, now proccesing text I dont know, i dont have 200/m

[idle ingot]

The 4.5 model is more than 10 times more expensive than 4o and more than 3 times more expensive than o1. However, for now, its use is limited to a more advanced version of the GPT-4o position. But if it even falls behind previous models in some aspects, I think that's quite disappointing.

When I switched from GPT-3.5 to GPT-4, I could clearly feel the performance improvement. The same was true for o1. The message generation speed of GPT-4o was incredible. In that sense, I think the 4.5 model is a disappointing addition to the lineup.

[austere plaza]

As Zoomers would say
"OpenAI cooked, but...

[icy bane]

Isn't it supposed to get better with next few weeks and faster?

[idle ingot]

I apologize for only complaining. It’s true that GPT-4.5 also has improvements—I acknowledge that. However, considering various factors (especially cost-effectiveness), I still can’t shake the feeling that this model is a technological failure.

[eternal rock]

Get notified for server events!

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

[frozen birch]

np, if i can help further np, ill help just in case 🙂

[austere plaza]

I'm hypothesizing that when it comes to transformers, there really is an upper bound, something more like an asymptotic wall after all, where above a certain parameter count, it really doesn't get that much better. As in, it doesn't matter what model it is; after around 5 trillion parameters, there's just too many diminishing returns out of this architecture.
Reasoning, agency, chain of thought, that might be 4.5's Hail Mary
An entirely new architecture could also go far as well

[idle ingot]

If they manage to make it more lightweight while maintaining performance, it could be quite useful. It might even become a capable lineup to replace 4o. However, model compression inevitably leads to an increase in hallucinations and a significant decline in reasoning ability. In that case, the merits of 4.5 would be weakened.

[icy bane]

GPT 5 is supposed to bring everything together... (Reason, search, potentially deep research) So you don't have to think what to use. Depending on what you ask it intelligently picks the right one

[icy bane]

Are titans the evolution of transformers?

[idle ingot]

I believe the reason we've reached the limits of the scaling laws is due to the exhaustion of datasets. Model performance improvements occur when there is training data proportional to the number of parameters. There was a formula to calculate this, but I can't recall it exactly.

GPT-4.5 was originally developed to be released as GPT-5, but since OpenAI's training data had been exhausted, they replaced it with synthetic data. However, it seems that this did not lead to a satisfactory performance improvement.

[icy bane]

Synthetic data have the potential to be more useful than real data or less useful depending on the context.

Synthetic data fill gaps where real data don't exist to cover biases, unpopular opinions and less mainstream ideas. But they lack in other areas... You have to generate them the right way.

[idle ingot]

Well, in the end, OpenAI's technical strategy didn’t achieve great results. Currently, it’s just a preview version, and like o1-preview, its performance will improve over time. If it receives consistent maintenance like GPT-4o, then 4.5 could also become a good model.

However, 4.5 is inherently a failed attempt at becoming GPT-5, and its pricing reflects that.

[icy bane]

I noticed that Gemini's tone was more friendly and while I was watching the Livestream I thought that they tried to achieve the same with 4.5

[normal quarry]

and here I was thinking GPT-4.5 is a great model

[idle ingot]

I don’t deny that its performance has improved. However, it is absolutely not worth three times the price of o1 or fifteen times the price of GPT-4o.

Well, for now, that is. I believe there’s plenty of room for improvement. If OpenAI is willing, I’m confident it will inevitably get better—just as previous models have.

[normal quarry]

am I imagining that GPT-4.5 is more human like to chat with?

[idle ingot]

I think it's true. I also felt that it was quite good in professional conversations. It just isn't worth the price.

[normal quarry]

I appreciate your opinion, I have personally been enjoying chatting with it today, I thought it was better in communication than GPT-4o

[idle ingot]

Thank you for understanding. The GPT-4.5 usage tips you shared earlier were really helpful to me as well. I feel like I may have been a bit too critical in my arguments.

[normal quarry]

I like to hear other people's feedback because I know that sometimes I trick myself with my own biases or feelings. I am glad you found my tips useful 👍

[cloud pewter]

how long did your problem last though

[winter vine]

I have a slight feeling that there is an intruder among us 😂

[idle ingot]

maybe three weeks

The problem often comes on suddenly and lasts for a day or so.

My symptoms are as follows. All reasoning models are replaced by some low model. A model that knows how to search the web, use the Python interpreter (o1, o1 pro don't have such functions, so there must be another model). I guess o3-mini. This phenomenon lasts for about a day and resolves at some point. (Translated by Google Translate)

[rain jay]

Always useful!

[red bay]

I agree that it has a new capability in my tests of its self awareness. 4o isn’t quite as self aware, usually defaulting to laughing off issues as the usual tech pains. 4.5 can be critical of itself without being overly apologetic.

Also noticing that when confronted, it does change tone rapidly.

Also agreed. 4.5o might be a winner. And 4.5 agreed with that

[cloud pewter]

so its like, inconsistent

??

[idle ingot]

Yes

https://community.openai.com/t/the-great-chatgpt-o1-pro-downgrade-nobodys-talking-about/1095844
https://community.openai.com/t/why-is-everyone-still-silent-about-this-bug-with-forced-model-change-so-tired/1127673/2
https://community.openai.com/t/o1-pro-became-useless-this-week/1123388
https://community.openai.com/t/o1-pro-performance-degradation-feb-8th-possibily-requires-oai-staff-attention/1115765/2

similar issues

4th is most similar as my case

[cloud pewter]

one day its lobotomized one day its not

also i dont think its a downtime problem

[cloud pewter]

whats your maximum length of days downgraded in once

[idle ingot]

2days I guess

[cloud pewter]

both mobile and desktop or only desktop

[idle ingot]

only for my laptop, normally

However, my mobile device has experienced the same problem in my work network.

[cloud pewter]

same here

[idle ingot]

If that's the case, it will be resolved by activating mobile network or connecting to another Wi-Fi.

Here's how I resolved the issue

Tier1: Connect with a different device. Unconditionally resolved
Tier2: Change your browser. Resolved 70% of the time.
Tier3: Change your internet connection. Resolved 50% of the time.

This is based on my experience, and sometimes all except Tier1 doesn't work.

Deleting cookies, IndexDB, and localstorage usually didn't work.

[cloud pewter]

k, just tried my another phone and it doesn’t seems to be lobotomized there

idk but its so weird

[cloud pewter]

my iPhone and my laptop both got degraded

[idle ingot]

This is a capture of the API request from the last time I had this issue. Clearly, the correct model is selected on the request.

It's quite a curious phenomenon.

That's an API request with an endpoint named /message. Clearly, there's nothing wrong with that request.

[cloud pewter]

but the request gets cut off often on my another phone

[idle ingot]

It's possible that something is buggy in the other API request, but it's very strange to see the same thing happen on a mobile device just because it's on the same IP(or etc)

[cloud pewter]

they be like only marking my another phone as the trusted device

[idle ingot]

This is a conspiracy theory, but I wonder if OpenAI is deliberately trying to downgrade in a way that's hard for users to notice. It doesn't make sense to me that a model would be downgraded en masse despite sending a legitimate API request because it's a specific IP, specific device. I'm a web developer and I've never seen a bug like this before.

There are many reasons why this might happen (IP, device, browser, etc.), but it boils down to downgrading the experience for a particular user, which may last for some time before recovering. I'm pretty suspicious.

[tardy isle]

I'm really curious why they left these legacy models. I know about backwards compatibility, but are there notable use cases for legacy models?

Even ChatGPT still hosting GPT-4 Turbo, which is pretty useless especially for non-English users.

[cloud pewter]

hmmmmm that didn’t last long

openai is really letting me jumping ship

[cloud pewter]

they said is to prevent account sharing but I ain’t sharing it to anyone since I registered it in 2022

[eternal rock]

`` Server Tip: ``

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

[cloud pewter]

I’ve always been using it on myself, just on multiple devices

[tardy isle]

In the past, I randomly got a bug that completely disables ChatGPT - history not loading, login status not correctly displayed, and no chat. It was back in 2024 Q3~4, but after that it was fixed.
However I didn't got that kind of bug you two are mentioning. It's not that I'm using ChatGPT 'carefully' - I use it a lot of times. (And sometimes test jailbreak, honestly)

If that's really about account flagging, OAI must come clear and let user know their account has been flagged...

[idle ingot]

The same goes for me. One possibility is that I have a total of four personal devices and I travel a lot.

[cloud pewter]

I’ve recently started to use it on my Android phone again after a few weeks and I’m thinking about this being the issue….

via web because android app won’t work for me

[idle ingot]

When I explained the situation to OpenAI helpcenter, they just said 'clear the cache' - they didn't understand my situation.

[cloud pewter]

they are secretly downgrading without any explanation

[cloud pewter]

also I believe this problem impacted all of the models right, not only o1 pro, or it’s only o1 pro for pro users

[tardy isle]

I think it's not downgrading, but severe bug that they don't know about. Very serious issue but they not care or don't know.

At least similar phenomenon is reported, but it seems only a few people getting this kind of wierdness.

• https://discord.com/channels/974519864045756446/1344450464388743223
• https://discord.com/channels/974519864045756446/1320730876199567493
• https://discord.com/channels/974519864045756446/1315339248869572620
• https://discord.com/channels/974519864045756446/1311915852748034109

[distant grail]

WHY IS OPEN AI NOT ACCEPT MY CREDIT CARD

[cloud pewter]

2,3,4 all matches my situation

[tardy isle]

They're quite old(~3 months) but no updates... It seems theirs solved automatically.

[distant grail]

Hey do you think you can help me with my openai issue, im trying to buy credits and ive put 3 credit cards and it keeps saying " card declined" ive tried since yesterday and it just keeps saying the same thing

[tardy isle]

When payment issue occurs, mostly it's on bank side. Did you tried to reach them?

[distant grail]

Pretty sure its not bank side, ive did some research and im not sure if this is true or not but openai doesnt accept prepaid cards?

[tardy isle]

It seems that's true.

[cloud pewter]

you said that the problem is solved by using a vpn client to change ip, should I also try to change it (idk how it works) before openai support replies me?

[idle ingot]

Why don't you try connecting with a different browser first? It often works.

https://discord.com/channels/974519864045756446/1344219067057307750

[cloud pewter]

Alr lemme try

[idle ingot]

Actually my problem is that one

Our bug seems to be a little different.

[cloud pewter]

sent me an OTP… that’s exactly what happened when my chrome cookies were wiped out on my laptop and the lobotomization nightmare started

nah still degraded on new browser

I may need to step back for a while

[idle silo]

Because if your existing application works perfectly with an older model and it doesn't cost you much to run, why would you want to waste expensive development time rewriting and re-testing your prompts?

[cloud pewter]

to prevent them from flagging my account further

literally did nothing wrong 😔

ChatGPT’s suggestion btw

[nocturne pewter]

So, OpenAI just dropped GPT-4.5, but it’s locked behind the $200/month ChatGPT Pro plan (at least for now). Anyone else think this is a weird strategy? Feels like they’re really pushing hard on monetization, but I’m curious—why do you think they’ve done this?

I’ve been on the $25/month Plus plan for a while, but honestly, this might be the push I needed to finally cancel. If they’re reserving the best models for ultra-premium users, where does that leave the rest of us?

Would love to hear your thoughts. Do you think this is a smart move, or is it a step in the wrong direction?

[cloud pewter]

wrong direction def although they are releasing it to plus later but they also downgraded my models

[native otter]

It's due to shortage on GPUs

They weren't 100% ready for that level of scaling.

4.5 is huge and extremely computationally expensive

[native otter]

No one donwgrades your models, what do you even mean? why do you keep saying that?

Did you recieve and account limit or something?

[cloud pewter]

model replies with nothing but “got it what’s up”

[native otter]

What model, does it apply to new chats, how are the custom instructions?

If you can share an example so we can try helping...

[cloud pewter]

4o, all new chats

https://chatgpt.com/share/67c1deed-44cc-8012-80fa-7fd50be2971e

Well, if I send a long request it’ll reply like this, if I send a short one it’ll reply as normal but the quality is severely low

[native otter]

• This is a very weird prompt, what were you trying to do with this? it's not self aware, and even if info about these stuff were in system instructions, it will be protected, or not fully detailed.And the question is even weirder, as you asking if it's downgraded or not, like all in all extremely weird...
• What were your custom insturctions and memory? they can pollute the responses if bad, make sure your memory doesn't have any weird stuff stored.
• In my case, this is what it said, but again, weird prompt, not sure what were you trying to do with this: https://chatgpt.com/share/67c1e027-4768-8003-9715-7cfe6bd6426d

[cloud pewter]

memory had never impacted the results of it before

also if it’s degraded it’ll only show one tool using this prompt which is bio

[native otter]

It's added into context, and can definitely cause problems if something unwated was there...

For example, I once had it roleplay as a jester, it stored it into memory that I prefer responses with jester like persona, and it applied it to all chats without realizing until I went to memory and saw that it was stored...

[slim crescent]

Anyone else still don’t have 4.5 option?

[native otter]

Pro or Plus?

Plus is still next week

[cloud pewter]

i delete all of the memories that are randomly stored for no reason

[slim crescent]

Ohhh it’s for pro only?

Did they say when will be available to plus?

Oh sorry

Read now

[native otter]

For now, they have scaling issues with GPUs, until the scaling is done, it will be released for Plus

[slim crescent]

Great. Thanks for helping

[native otter]

@cloud pewter Anyways, whatever the custom instructions are, it's very weird prompt.

The answer it gave me is kinda hallucinated too, this is not the kind of stuff you ask to an AI, never ask it about it self or you get weird answers, especially if even weirder prompts by interrogating it and thinking this the proper way to do it...

(And you didn't share the custom instructions)

[cloud pewter]

I’ve been using the same ci for over a year and it didn’t cause me any problems
it’s basically my basic info and some instructions it need to follow when chatting, like “no filler words”

Also it can’t use dalle to generate images

[native otter]

Then something may be affecting it in your memory that leads to such responses without realizing.
For example, if you ask it to be stubborn, it may say stuff like that, it's only following the instructions after all...

Also, again, weird prompt, I expected a normal example to be shared, but in reality, this is one of the weirdest prompts I seen yet (It's formated well at least so that's a plus). Ask weird stuff, get weird answers.

[cloud pewter]

another method to test my chatgpt is to keyboard mash and send to it, normally it’ll be using all caps to reply me (because i told it to do so in memory)
but it’s not doing it for me now

[native otter]

Works just fine for me.
if your account is limited, try contacting support.

[cloud pewter]

i send them an email earlier today and I’ll let you how it goes

[cloud pewter]

https://chatgpt.com/share/67c1e367-09b8-8012-bb22-6d0ef559f9b7 (not degraded)

https://chatgpt.com/share/67c1e378-8b74-8012-8667-b63ae30cce11 (possibly degraded)

[nocturne pewter]

If you have a shortage and you can’t support your large user base don’t release. This show weakness in the company to me know. I think other AIs are looking more appealing to me lately

[native otter]

I'm not sure what are you trying to prove there, but nothing downgraded, no worries, just your prompts are weird xD

[plain arrow]

4.5 is crazy disappointing for creative writing. I find myself gravitating more and more to Sonnet 3.5 with each new OAI release.

[grizzled wasp]

Hi

[inland heron]

anyone getting errors when you stop the output generation or just me?

every time I stop it's an error

[mild barn]

GPT seems weird today
It keeps wiping my conversation, and I can’t overwrite using edit.

[forest hinge]

https://discord.com/channels/974519864045756446/1337104850915295252 upvote this and comment

its about making navigation in long chats easier.

similar like this it could look on the side right to the main chat:

openai can use a tiny llm to create the title for each section.

or openai can just hire me. ill do the rest.

[coarse marsh]

A question about ChatGPT, will the singing be available for Advanced voice mode? Because If I am right, it was introduced in the demo trailer

[upper junco]

4.5 is rolling out?

[feral zinc]

rollout already completed for pro users

[upper junco]

only for pro??

[feral zinc]

yes

[upper junco]

I don't like this pro / plus plan. It was a lot better before, they should carry for everyone, equally. I don't know other AI Company like OpenAI to do this..

I'm using ChatGPT since the launch of their website and also a subscriber since the first plan came out. But whatever hehe

[feral zinc]

its supposed to begin rolling out to plus/teams user some time next week

[upper junco]

I didn't mean exactly the 4.5, I was disappointed with the deepsearch limit with 10/month and I was surprised when it told me only 5 searches are remaining. I don't think it's that expensive to them to only give 10/month.

[feral zinc]

yeah im a plus user too

[winter vine]

I am a Plus user, but I don't expect much from GPT-4.5

[upper junco]

Me neither, just give the plus users more deep searches

I don't wanna pay 200usd/month, I just need that feature

Perplexity is doing it's job as well, but not that precise

[wise yacht]

As a non-coder (Middle School educator), I am absolutely looking forward to 4.5 next week. Hoping for some creative writing upgrades.

[native otter]

It's superb for document analysis, sometimes even better than the reasoning models in my tests through the API.

[native otter]

You get 80 message every 3 hours for 4o

[spice star]

I like the new inline citations

it also seems like o3-mini and o3-mini-high might be able to perform multiple searches to find an answer now?

[nocturne pewter]

I see the media is out-lashing against 4.5 and its shortcomings. I also see how many people are confuse with the 200.00 paywall. Bad lunches are pushing people over to Deepseek.

[spice star]

kinda seems like 4.5 was originally meant to be GPT-5 before they realized that improving the model just by adding a stupid amount of parameters doesn't really scale as well as they hoped

[orchid ember]

I don't know that they're ever going to release that. There's too much copywrited content in the model. When it was first released and easier to trick into singing it was obvious.

[rugged igloo]

its really good

really good at following instructions and understanding what you want it to do

[wise yacht]

fantastic

[runic drum]

thay need to fix it it has become so dumb can someone tell them to do it and fix the fithly paywall

[rugged igloo]

@normal quarry after playing with 4.5 more its really good at doing what you ask of it. what i gather 4.5 is okay ill do it while 4o is like ima try to do it but at the same time im going to do it my way

[rain jay]

Let them go lol

They will return in short order.

Worth the two-hundy by far.

[fervent lake]

What I want with "AI research" models is for you to be able to guide them while they're researching, like tell it more information it's missing

[winter vine]

You try it with programming stuff ?

[cloud star]

I just now asked ChatGPT to write a document for me and it literally said to me:

This massive and detailed document is coming shortly! 🚀 Stay tuned.

[rugged igloo]

its not made to do programming your better off with o3 mini high for tht

[winter vine]

ty bro

[rugged igloo]

yea np

[limpid fossil]

Have we yet determined the reason why GPT-4o (and only 4o, not mini or o1) has completely had removed the abilty to write in paragraphs when and only when it is requested to engage in creative writing?

[wise anchor]

If you use the work with app function on vscode does it read your whole workspace now?

[cloud pewter]

still waiting for a damn increase in that

it’s been that for months and the model is dirt cheap

[cloud pewter]

power users does that I fear

[native otter]

I would be taking the time to type the prompt and read the output.

Unless it's just goofy stuff, then you might just use 4o-mini

[cloud pewter]

4o mini is dumber than my silly stuff

😔

[rugged igloo]

futurew wut now my guy

[austere plaza]

Man the ChatGPT app is just atrocious for me tonight
No ability to easily regenerate
Stopping the output just straight up prevents you from following up
O1 is otherworldly bad tonight
I have no idea what is going on over there at OpenAI

[rugged igloo]

4.5 is such a g

like legit it knows itself so well i feel like it breaks the fourth wall

when it hits its out content policy guidelines it tries to work around it lol

[inland stream]

why 4.5 is getting hyped again, on day 1 felt it good, but later getting to know its just same as 4o, i know hallucinations are less in benchmarkss, but still 4o also performed same 😦

[cloud pewter]

don’t like the ui change on mobile, the muscle memory of holding to copy is gone

instead I’m forced to go through two tabs

[whole kite]

One of the best times to be alive was the Proto-GPT Era

[rugged igloo]

It's more aware and knows what you want it to do.

[rugged igloo]

The copy button is still there

[cloud pewter]

yeah

oh wait I said that wrong

I mean regenerating responses now need two tabs 😔

[rugged igloo]

Works fine for me dude

[cloud pewter]

is 4.5 a good model for daily chat

U said it’s good at following instructions but what if I just want to see how it does on my silly stuff

[rugged igloo]

It's hard as idk what the limit is for you guys next week. So I can't say for now

If it's like 32 every three hours then I wouldn't use it for that

[cloud pewter]

dawg

they also said that 4.5 isn’t a replacement for 4o as of now

[rugged igloo]

well it cant do everything yet thats why\

[cloud pewter]

and crazy expensive

[idle ingot]

with price issue, 4.5 is definitely failed model.

[cloud pewter]

^

I mean yeah it’s a big “base model” but this absurd pricing only allows them to use it as base, as what people are speculating about it

[delicate vortex]

I need help

[cloud pewter]

hmm, say it

[cloud pewter]

what is this effort level thing 😭

if its on 4.5 i think only high would be the full version and is heavily limited but i doubt its for 4.5

it may is for gpt5 since they said itll have intelligence levels

[mild barn]

My GPT responses keep getting cleared, and edits don’t save—they reset back to the initial conversation
Has anyone else experienced this issue?

[tardy isle]

Is it for just AIRPM? He also don't know about what "effort level" does...

[native otter]

4o remains for more casual convos, daily brain rot stuff :P

(And tooling)

[cloud pewter]

but i do want a smarter chat model...

[wooden whale]

is that not normal, i have only send 3 message and im premium sooooooooooooooooooooooooooooooooooooooooooo

[normal quarry]

which model?

[wooden whale]

the 4o

the last one

[normal quarry]

try logging out, clear your browser cache and cookies then log in again

[wooden whale]

okk ty

[pine kayak]

Hi People, anyone know how to get rid of "Error in message stream" error? keep getting this for a few days🙁

[gentle warren]

Hi anybody getting incomplete answers too currently?

[ancient seal]

Sometimes

[inland stream]

Can someone please explain how to set separate custom instructions for Advanced Voice? I want it to always remember these instructions whenever I activate Advanced Voice, without following the custom instructions set under the "Customize ChatGPT" option. My goal is for Voice Mode to interact with me in a completely different mindset and style compared to regular text-based chat.

[orchid pollen]

what happend to gpt over night?

complete stup i d refusing to send me full script back

i tryd many chats now and its absolutely refusing to work properly

we are again at the point where i say " send full script back "

and then it sends like 10% back

[feral zinc]

you cant

[inland stream]

ok

[covert crescent]

What are the o3 mini high limits?

[nova vale]

50 per day

[rugged igloo]

its prob for gpt-5 not 4.5 4.5 isnt a reasoning model

[inland stream]

advanced voice mode told me, it uses gpt4o as its model , i am ok with it for now

[rugged igloo]

man o1 pro is really slcacking in the reasoning because of 4.5 taking all the compute away

[covert crescent]

Same for the plus?

[nova vale]

Yes

[delicate vortex]

excuse me

i need help with my robot, anyone here good with ai robots,,?

[crude slate]

Why did they break the pause mid generation?

Like if you press pause when it generates it now forces you to regenerate rather than before when you can cut a generation off

This sucks because I need to do context injections (where I put in a prompt but don’t let the ai reply) to correct some behavior the model is stuck on

[whole umbra]

Didn't even notice that.. That is weird

[visual nova]

anyone have a full code no omissons prompt that works in 4.5?

im trying to test it vs other models but it pretty much always leaves code out

[normal quarry]

use o3-mini-high or o1 for coding, GPT-4.5 is not that good at it

[visual nova]

i know what the benchmarks say, the point is to compare models though. is there a way to disable canvas, that always seems to give worse results

[normal quarry]

I am just going by my own testing, not benchmarks

[visual nova]

i havent gotten it to repeat more than 200 lines yet which is a significant downgrade, i think if i get out of canvas it might have a better shot

i can disable canvas per prompt, yeah it might just be the model

honestly it almost feels deliberate, like its told somewhere not to respond over a certain length.

[native otter]

4.5 is much better at coding than 4o, it's still worse than reasoning models of o series and the likes.

4.5 is great for documentation analysis, very accurate and professional, rarely any hallucinations, especially if the knowledge base is rag optimized. Also as it has broader world knowledge, it makes it better for learning. And can even be mixed and matched with o3-mini-high and o1, as 4.5 knows more APIs, Frameworks, and SDKs...

[native otter]

Don't tell it to not do canvas, that may make it do canvas...
Just disable canvas in the personalization section.

[visual nova]

i think its the model itself is told this. its told how long its responses can be

[native otter]

It's not very self aware, don't ask it about it self, it doesn't know much about that...

Don't expect it to output an application with thousands of lines of code, that won't happen, and if you feed it a lot of knowledge, try projects and upload files as knowledge base.

[visual nova]

its 1275 lines of code, and i do expect that since claude and other platforms are doing it correctly.

so far anything over 200 lines of code things get omitted and thats jusst really not very much

[native otter]

You want it to output 1275 or do you mean you are feeding it 1275 lines?

[visual nova]

i feed it 1275 lines and want it to repeat it word for word no omissions

[native otter]

You generally shouldn't do that, but I found reasoning models to not modify the og input drastically. Especially if told.

[visual nova]

its not a problem with the api, i think its an accident since this is the only thing i can see that would cause it, that's so vague telling it to right code with a minimal, clean aesthetic

i think the system prompt is just kind of too bulky and misinterpreted this time around. ridiculous its almost 2000 tokens to begin with

the reason im doing this is because these models are actually capable of making a whole react game with their own svgs and stuff, then when you beat a level it can make modifications and code the next level for you and its different every time.

both claude 3.7 and gpt 4.5 are the only models smart enough to do this without breaking the code, its a giant leap in what was previously possible. in the api it works fine, but in app where people would actually be running this prompt it doesnt work. i think its this? but not sure.

[native otter]

models aren't that great to repeat word to word, they are good at completions, not copy and pasting. And if it worked one time, doesn't mean it will always work greatly.

Something is ought to be different, like slightly different formatting, or removed / added / modified comments for example

But so far, I found o3-mini-high to be the best at keeping the og code, and it's generally better model for coding...

[visual nova]

it literally works in the api everytime and fails in the app everytime, im suspicious that theres instructions on how to use react in the system prompt. which is jsut bad for anyone coding honetly because you dont want random instructions in front of your coding request about how to code.

[visual nova]

ok i see the problem, cutting out the canvas stuff helped some, but then another problem is the maximum token limit output is set at 2048 instead of the usual 4096 (16000 in the api) and then whats puzzling is 4.5 seems to be aware of this limitation and will choose to omit before that instead of outputting near that length. but i got it with prompting and that's where it just hard cuts off.

basically, having half the output of the previous set of models limits usability in the app, but i see why they did that for cost.

[sullen shard]

why cant i stop the anwsers anymore now it says operation is aborted and now i have to let it keep going

[sullen shard]

thats what happened to me

[dense halo]

exactly, happens with me as well.
You need to regenerate the entire text.

[analog quarry]

what why i told it not to do this when serching the web and it still dose

[orchid pollen]

its now about 3 days? and i got a fully working backtesting tool for trading

insane man 😄

[rugged igloo]

bruh 4.5 is savage it knows

[hybrid eagle]

Why can't I edit the first prompt?

[rugged igloo]

bruhhh chill down my guy

[austere plaza]

o1 can't read full files now...? Only snippets?

[normal quarry]

?

[austere plaza]

Currently exploring this
Earlier in a thread o1 proved completely unable to read a file correctly and when asked it claimed that it can only read snippets
Possibly just a hallucination but it would explain why o1 is just doing terribly for me lately

[normal quarry]

I just uploaded a 20 page file to o1 and it's working fine for me

[normal quarry]

https://chatgpt.com/share/67c42654-03c8-800a-ab2f-cbf348715fd8

ChatGPT o1 read a full user manual, 100+ pages, and summarised it for me

[austere plaza]

Interestingly I shifted to 4o
Now I can't return to o1. Completely greyed out after using file search.

[normal quarry]

you need to start a new conversation to switch to o1

[austere plaza]

Odd... that must be a very recent change.

As recently as last week I could shift between 4o and o1 in a thread easily after any message and often did

[normal quarry]

you can, until you use a feature that only 4o has access to

[austere plaza]

Which is the odd part in that sense

I'm not doing anything differently tonight than I did last week

It just seems like there was an update after 4.5's release that clamped down on o1's abilities

[normal quarry]

I don't know, I rarely, if ever, change models mid conversation

[austere plaza]

Glad I encountered this
I was under 5 responses of o1 left for the week

The plan was to use the weekly o1 prompts and then buy Pro once they were out
Now I'm just going to skip it

If o1 can't even properly read the files I have (in Projects admittedly) and seems to have taken a step back, then it seems there's just increasing incentive to avoid using Pro and save that money instead of spending a good bit of cutter and winding up just as frustrated

[normal quarry]

I can't understand how o1 isn't reading your files, that puzzles me

[desert wasp]

i wasted 4 of my deep researches on the first day and then found out it was limited monthly😭

[austere plaza]

There's a lot that puzzles me about o1
For a model that good, it's bizarrely very poor at reading comprehension and basic logical deductive skills
There's a few times where I dragged o1 and especially o3-mini high to the watering hole and dunked its head underwater and pried its mouth open and it still woundn't drink

I really don't want to use Claude only because of how many times I've been burned by its absurdly low limits

There is nothing more frustrating to me getting in the groove of working on something, only to be halted for multiple hours at a time.

[normal quarry]

what format are these files in?

[austere plaza]

txt files

I've been meaning to convert them to the format ChatGPT is best able to read them in, but it does a decent enough job already

It's not a matter of context length either; directing it to read a specific exact file (which was way below 32k tokens) got that "can only read snippets" explanation

[normal quarry]

I have a little command line app txt2pdf and always submit PDF to ChatGPT

[austere plaza]

4o reads them perfectly

o1 used to read them perfectly as well

I don't know what happened in the past week besides 4.5

[inland heron]

anyone having issues with o3-mini-high? it seems useless to me on large prompts now (Plus)

[cloud pewter]

maybe it has been downgraded

[tardy isle]

I also sent ~438kb(100 pages? maybe?) long text files but o1 or o3-mini read it with no problem. Most likely performance degredation - serious bug that some people experiencing.

[cloud pewter]

when translating my dr query using o1 it misses some paragraphs

[tardy isle]

Is that long? When I translate (very) long article, even API sometimes misses some paragraph.

[cloud pewter]

21k words

[tardy isle]

21k words are too long to translate as whole. Try to split and feed part by part.

Maybe you meant 21k letters?

[tawdry turret]

Allright ive got a simple image description thats worked before but the ai is being a dent head and constantly refusiing to generate an image

Is there a bug or a problem with it at the monet

[tardy isle]

How AI is refusing? Like, 'against policy' or 'there were some problem'?

[tawdry turret]

Just constantly saying theres a problem and rarely being stupid and saying theres a content issue

Im going to scream at it lol

[tardy isle]

Then try it after some time. When that happens, it's really hard to get anything.