chatgpt-discussions

[quick meteor]

Also some articles way back in October were talking about open ai releasing a new Orion model in December

But I guess we’ll find out shortly

[real bluff]

the orion and o3 are supposed to be the same thing?

[quick meteor]

No I don’t think.

The o for the o1.2.3 etc series of models stands for “open ai 1”

Sam said so in an interview

[real bluff]

ahhh i see

[regal portal]

let’s be honest, they care about money. i’m all for preventing certain type of content (such as stuff that jared from subway would like..), but a lot of the censorship openai enforces is so they don’t risk getting sued or canceled for hurting someone’s feelings.
Midjourney got more and more censored and it’s losing subscribers because of the overbearing parenting.

[quick meteor]

I feel reasoning models are amazing. However at some point that major leap in things like writing, poems, long context, etc has to come.

[real bluff]

yeah i guess
tho i suppose they wont have as much use as reasoning models

[quick meteor]

Mmmm

I’m not sure actually. For example here Sam hints at some model being more powerful then o1 and o1 pro

If he’s referring to Orion, then it will be a very powerful model.

[quick meteor]

I’m pretty sure the GPT series is mostly used for writing and coding.

Like for school any large major improvement akin to that of 3.5 to 4 would make ai detectors useless

[regal portal]

I think the biggest issue that needs to be tackled is hallucinations.
I’d like to be able to rely on the information given to me by the AI without having to look it up to make sure it isn’t feeding me bs.

[quick meteor]

We’ll find out in around 10 minutes

[quick meteor]

I have faith the next generation of frontier models will be much, much more reliable.

Even o1 is pretty reliable as of now.

I have high expectations for a model that has been in development since the laugh of 4

[real bluff]

what does each color represent?

[quick meteor]

It’s comparing o1 and o1 pro in coding or something like that.

Showing how they are almost the same

[quick meteor]

The 62 one is o1 preview so ignore that one

[errant glen]

Is o3 Orion?

[quick meteor]

Not sure why they would change the name.

o just stands for “OpenAi 1”

So Orion seems like it would be something different

But who knows I guess

[real bluff]

so its like 4o o1 and o1pro?

damn makes o1pro feel a bit useless

[quick meteor]

No. The o in 4o stands for Omni. So basically since it’s multimodal

So many o’s all around the place. Now Orion

[real bluff]

ahhh

yeah openai seem to like the letter o lol

[quick meteor]

it’s like musk with using X all over the place.

[real bluff]

he has just a car and twitter named X no?

oh wait also spaceX

[quick meteor]

And his kid

[real bluff]

oh wow i never noticed that lol

[quick meteor]

And he’s founded some companies like time ago with X as well

It’s o3

[real bluff]

just o3?

if we get only 50 a week on o1 i wonder how much o3 would be

[regal portal]

i was posting that too but i added the “dot com” and got server muted for a minute. 😂
Yeah it was in 2000, and eventually that company was merged with another to become paypal

——-
So they’re going from o1 to o3? Why skip o2?
Didn’t iphone do that once by skipping 9?

[warm totem]

o2 trademark was already registered to another company

[regal portal]

ohhh, didn’t know. Another llm company or different product?

[warm totem]

phone company if i remember right

[regal portal]

you’re right.

Weird that they would bend the knee on that. It’s a telephone company so it’s a different market, and it’s in a different country.

[wise yacht]

Cant watch the video, but I am confused. Was the big, final announcement a new model that once again we don't have access to?

Want to be sure before I switch to Gemini

[regal portal]

They announced AGI.

They called it Skynet o3

[warm totem]

basically yeah, this

[wanton aspen]

Is there a limit for the amount of archived chats that you can have?

[wise yacht]

Wow, so not even "Tasks" released?

[crisp cairn]

no even tasks

no nothing

[regal portal]

honestly this stream is going over my head.
Sounds like a different language that only supernerds understand.
Like when you hear extreme anime fans use japanese words like kawaiidesu mid-phrase

[crisp cairn]

i turned it off

no wasting time to hear about what will come in the future

[ruby gull]

Hi guys, I'm a newbie here, I'm trying to migrate the open ai API in my Windows but it isn't working, is here where I can get some help?

[regal portal]

I don’t think anyone has disappointed expectations with the 12 Days of Christmas gimmick as much as openai has. This is a world record.

[quartz gale]

i had no expectations honestly, so it's been kinda cool to me.

[regal portal]

This whole thing will be damaging to Sam’s credibility going forward though.

[elfin lava]

https://help.openai.com/en/articles/10303002-how-does-memory-use-past-conversations

[fossil sphinx]

I saw Sam zoning out lol he was thinking damn, how smart are these guys?

[regal portal]

neat.
Curious how it will handle contradicting information in past sessions. Like if you tell it you’re 20 years old in one session and 80 in another.

[elfin lava]

i hope this means no more unnecessary memory info saving cuz this is ridiculuous

[wraith dagger]

applications should go live pretty soon

[regal portal]

wow. i never came close to get a % warning. How much info you got stored?

[elfin lava]

some of them are not even prompted by me

idk why it decided to remember things thats not very necessary

[eternal rock]

This is a Safe for Work Discord Server

The OpenAI Discord is an actively moderated server.
• Refrain from sharing inappropriate content on the server. This includes but is not limited to messages, media, or other topics of graphically violent, sexual nature, and drug-related content.
• Report all sensitive and offensive content in the feedback reporting tool in the ChatGPT web UI instead of here on Discord.

Posting NSFW content will result in a permanent ban.

[winter elbow]

O3 IS COMING

OMGGGG

[cobalt lynx]

I liked the part when he said “wow! I need to change my worldview about what AI is capable of!” as a way to market lol

Still- pretty cool stuff.

[regal portal]

is that new memory feature available now?
Idon’t see it in the iPhone app settings.

[quartz gale]

so were people expecting a new model to be released during this event?

[cobalt lynx]

It wouldn’t make sense. They literally just released o1.

[crisp cairn]

with all the built hype - yes

orion here orion there

[wise yacht]

I was expecting either Tasks, Agents, an updated DAllE, etc

[crisp cairn]

shipping time

[wise yacht]

They built the hype not us

[quartz gale]

hype from who?

[elfin lava]

tbh those were

data mined

[quartz gale]

maybe i just don't pay much attention to what openai says

[crisp cairn]

twitter accounts

[swift pond]

This feels like a huge deal 🙂

[quartz gale]

so their twitter accounts said they were going to release a new model during this event?

[crisp cairn]

no but they were hinting about presents, shipping time

etc

[wise yacht]

You can go look. They are public accounts

[severe otter]

Safety testing sign up link anyone?

[sick pecan]

they reworded the openai memory feature docco, its still the same disappointment as always

[cobalt lynx]

I don’t need a new model, mannn. I just need ChatGPT to have tabs on the same page and to let me put titles on different collapsible sections of the conversation so I’m not constantly scrolling back through. 😆

[quartz gale]

i kinda want to be able to use canvas with voice mode

[crisp cairn]

okay they did deliver new model (o1-pro) but behind 200$ wall

[radiant thorn]

Where can I apply for early access

[elfin lava]

https://openai.com/index/early-access-for-safety-testing/#how-to-apply ?

[echo nest]

Guys, did you see the live? Did I understand correctly? Will the gpt o3 model be released?

[elfin lava]

oof nvm

[quartz gale]

no it's being "safety" tested

[radiant thorn]

Thank you

[sick pecan]

o1-pro model is likely just o1 model with higher reasoning effort value

[echo nest]

Thx

[half crystal]

Thank you!

[crisp cairn]

yes with a higher time limit for thiking

[elfin lava]

wait was that the final day

[sick pecan]

i release new llama 3.3!! (is just llama 3.3 with different temperature)

[placid pilot]

I'm getting "The request is not valid" when trying to submit the form. Does this work for anyone else?

[feral zinc]

yes

[elfin lava]

ok back to sleep i go then

[echo nest]

What are the requirements to participate in the test?

[sick pecan]

dont look like a spy from the competition

[feral zinc]

at min, be a researcher with published papers and/or code

[elfin lava]

why they skip o2

[echo nest]

I'm just a poor bum who pays 20€ a month, I can go back to sleep

[narrow viper]

O2 is a global brand name owned by the Spanish telecommunications company Telefónica. The company uses the O2 brand for its subsidiaries in the United Kingdom and Germany.

[elfin lava]

.

[sick pecan]

time to buy o4 domain with content 'sama contact me to discuss price'

[narrow viper]

🤣

[echo nest]

They should make gpt accept video and audio files too, very interesting projects would come out

[regal portal]

how did you find this to begin with? Also i don't see that beta button anywhere on either the browser or iphone version (plus sub)

[regal portal]

i forgot who, but there was a company that had to buy the rights to the domain for the sequel of their game. I believe Blizzard had to do that for Starcraft 2 but there was another company beofre that.

[echo nest]

are you looking for the "memeria" option?

Memory

[regal portal]

this updated version, specifically

beta feature

[echo nest]

I'm just seeing it now

[crisp cairn]

i like this "expect a delay of multiple weeks for access to this model" 🙂

[echo nest]

I missed this update

[regal portal]

like the o4 advanced "video". 😄

[sick pecan]

openai spends 15 mil on chat dot com. openai discord server mutes you for saying it

[mossy zodiac]

Its not letting me submit my application, why isnt it working?

do I need to publish papers or something?

[feral zinc]

well the form is for researchers and if you dont have anything public you would have less/no chance anyways vs most others

[dense halo]

apparently Chatgpt's servers got the King Von Treatment💔

[plain trench]

I read the post about the 12th day but I don't understand about chat gpt o3 if it will be tested or if it is already available now?

[echo nest]

Does anyone know how to activate this beta mode?

[crisp cairn]

so applications close 10 th Jan - so when do you expect o3 min access ?

in public

application close 10 th Jan then safety tests proceed

[regal portal]

i can’t find that anywhere. Maybe it’s just built in the regular memory feature, i’ll test it out

[echo nest]

keep me updated

[regal portal]

Doesn’t work. Either they didn’t implement it yet or there’s a beta button somewhere and they haven’t released it yet. Maybe the pro guys have it.

[echo nest]

20€ a month is already too much, imagine spending 200$ a month for an AI

[supple terrace]

o3 is gonna make such a good AI gf 🥰

[torpid kelp]

So are we not getting any better regular use model?!

O1 is already great but I can barely use it thanks to the 50 massages per week cap

4o badly needs an upgrade

It's like o1 is the favorite child who also happens to be a Harvard graduate and high achiever, while 4o is the high school dropout who had to repeat grade 8 7 times

[sick pecan]

they offer unlimited o1 for a mere 200, which is less than most other professional software subscription services

[warm totem]

this last day was really nerdy, i guess it's their primary focus now

[torpid kelp]

Mere 200, seriously?

[warm totem]

i guess 4o will still be 4o for few time at least

[sick pecan]

to a consumer/prosumer it may seem like a lot

[torpid kelp]

What professional software subscription are you talking about

[sick pecan]

i already provided examples in the past => #chatgpt-discussions message

[torpid kelp]

Even the jetbrains ides aren't that expensive

[warm totem]

autocad

300 per month

[sick pecan]

per app

[warm totem]

yes sky is the limit 😄

[sick pecan]

welcome to capitalism. if the company is doing well despite high price tags, then those price tags are justified

[torpid kelp]

I don't care about unlimited usage, I don't need it. Just give me a $50/mo model with 100 o1 requests per day and I'm happy

[warm totem]

so why not using the api ?

50 usd of credit per month and voilà

[torpid kelp]

Thatd be more aligned with what average professional user would need

[torpid kelp]

Don't want to, I want to use the app

I don't want a 3rd party gui

[warm totem]

so if you want to absolutely use the app you can't

[torpid kelp]

Plus I still want the convenience of other features my gpt sub brings

[sick pecan]

i want 10" subway because footlong too much

and 6" too little

[torpid kelp]

Regardless the pricing model is just not great objectively

It's fine to have a $200 tier, but the only other option shouldn't be a severely limited $20 tier

[warm totem]

chatgpt is a loss leader

[sick pecan]

yeah compared to api pricing, chatgpt gotta be losing them money

[torpid kelp]

It's giving them exposure and data to train models

Enough for them to facilitate apple intelligence for free lol

[warm totem]

culture the hype with chatgpt, sell api to saas providers that make tons of api calls a day, at full pricee

[scarlet delta]

Hey guys, it is just me or did ChatGPT severely decline over the past month? It's almost useless now.

[sick pecan]

is age old strategy. give away the razor and sell the razor blades

[scarlet delta]

4o lacks basic reading comprehension skills. I ask it something, then it only comprehends like half of the words used. And every response is extremely generic.

[torpid kelp]

Elon musk ddosing openai servers

[scarlet delta]

Is he really

[torpid kelp]

Yeah, using all starlink users as bots

[inland heron]

I'm struggling to remember what Plus users got in these 12 days of OpenAI... Other than few bad Sora videos per month, and o1 (very similar to previous o1-preview) I'm not sure what we got

[static cradle]

projects

[inland heron]

oh right, the ability to search the chats, that's useful

projects yeah, but o1 can't be used there, but only up to 4o

so inferior model which I don't use much

[static cradle]

vision

santa

[inland heron]

ah yes, Vision

don't use it, but it's there

Well at least o1 is released so, I have to believe them it's better than o1 preview

[regal portal]

We need to bring back stuffing nerds in lockers. 😄

ChadGPT

Of the 12 days, the one feature i did like is the folder projects thing. Kind of feels like Notebook LM now, but with a little more control on the conversation.
"Video" is kind of meh. I haven't found myself using it yet for anything. I'll be more excited about it when it's true video.

[umbral bear]

I thought I saw a way to connect those videos together in a series

but I have not looked at it either. 5 seconds is not much. I will try an open source one when I get a lot more powerful GPU

[inland heron]

you can rent a GPU for free up to a point

online

if you're good with code, there are services offering some free credits to use

[umbral bear]

Yeah, I saw Amazon has a freebie for 1 hour

[regal portal]

4o is my bro.

[inland heron]

I mean inference, for example at modal dot com

[umbral bear]

oh okay. thanks for the tip

[inland heron]

you'd have to set up the server though, but it's not terribly difficult

once you set it up, you can basically use their servers to download any model if you want (say, flux dev) and use h100 many times 😄

[umbral bear]

thanks. I will give it a try

[inland heron]

oh, and you get $30 of credits 'every' month, replenished 😄

for free

[sick pecan]

#ai-discussions

[inland heron]

ah yes

[umbral bear]

Those prices are a lot cheaper than HP with H100 & H200 at nearly $2 an hour.. okay..othe chcannel

[void tangle]

Glad I’m not the only one that has BroGPT setup

[restive plover]

So I missed it but what did they say about o3? Will it only be available to pro users?

Or will other subs get it too?

[sick pecan]

they havent announced it yet so i suspect its leaked ui without proper support = bug

[honest jay]

I think they said they're aiming for a release at the end of January, but that might have just been o3-mini. I need to double check! They didn't mention anything about what access looks like regarding Plus, Pro, Team, etc

[restive plover]

Crazy benchmarks

[honest jay]

Sam joked in the stream about how OpenAI's best coder "has a couple more months" before being surpassed by a model 😁

[astral storm]

So 12 Days of OpenAI is over.. what are your thoughts about it. for you was it 'Winner Winner Turkey dinner'.... or do you get 'Stuffed by OpenAI'? ....

[neon raven]

The stuff they announced is amazing

I have no idea how they plan to make o3-mini cost effective for ChatGPT though

given the base model costs 2k for 100 questions

[restive plover]

Yeah, that’s wild

[astral storm]

'I have no idea how they plan to make o3-mini cost effective for ChatGPT though'.... they said that about Amazon, when it announced its 'revolutionary free next day delivery on most products for a one off annual fee (ie PRIME).... but several years on....

[neon raven]

i have no doubt in the future it'll be cheap or that o3 will be seen as archaic tech in 10 years, but for right now its crazy

[astral storm]

yep - agree.. but they seem to be good at acquiring 'money to burn' currently.. so speculating to accumulate I guess.. and after all, its shareholders money they are burning at the moment I guess

[neon raven]

i'll be shocked if they release it on ChatGPT by end of jan for even plus users, 2k for 100 questions right now but thats somehow gonna be cost effective by January?

like theres money to burn and theres MONEY TO BURN

[astral storm]

They simply have to keep ahead.. at whatever cost... or they will literally burn and die

[static cradle]

I may be splitting hairs, but I thought the purpose of the 12 days of shipmas was to ship features

[crisp cairn]

yes it was

[astral storm]

For me 12 Days overall I think was a Winner Winner Turkey Dinner. I felt the last day 12 was kind of a disappointment for the masses of general users, like me... I was hoping something on day 12 that was for EVERYONE... but saying that I thought projects, advanced vision with voice, GPT in apps, improved search, improved Canvas, and finally the roll out of Sora ,were all really great enhancements of ChatGPT for most users...

[pulsar carbon]

o3 is AGI

[sharp estuary]

maybe don't buy the dumb hype.

[torpid kelp]

I don't get this. Sure it's impressive, but it's really just a better o1

Also what does agi even mean anymore

[dense halo]

someone help me
how do i make chatgpt stop responding in canvas
no respond in chat doesnt help because if canvas respong was used it will delete regenerate button from entire chat and it will never reapear in it

currently my only solution is to scare it with very inaproperiate threat so it doesnt try to respond in canvas

[sharp estuary]

https://chatgpt.com/share/6765dede-9984-8007-9d60-bab511e323a8

[sick pecan]

if you ask for it to provide codeblocks it tends not to use canmore canvas

[regal portal]

i got a friend that just signed up for plus, and i was trying to get him to play with sora, but when he logs to sora, it doesn't show that he's subbed to plus.

[sick pecan]

i believe half of that story

[regal portal]

you don't believe i have a friend?

you're right. It's true

he's just some guy on the internet i know

[radiant hull]

Well; the difficulty of a notion of AGI results above all in the notion of plasticity ...
To put it simply, an LLM has a "reasoning" through the product of statistics that will determine the best answer to provide.
The LLM is itself the product of an intelligence that has quantified, structured and defined a mode of reasoning.
But in itself, the mode of reasoning of an LLM will always remain in essence a static reasoning, even if the robot learns, etc.
The AGI in the first sense of the term, it is the fruit not of a reasoning, but of its reasoning. This reasoning is a plasticity that leads it to have one.
As the human brain does.
You do not decide on an answer because (or not only) there is a statistical result that leads you to take this reasoning/choice/etc. You take it because according to your abilities you deduce that it is the answer to have. Your brain does not do statistical summation, it does constant reorganization (plasticity), optimization, deletion, ...

So talking about AGI with an LLM is more about mimicry of AGI, than AGI strictly speaking. It is the notion of LLM that needs to be rethought when we talk about AGI, in the sense of artificial intelligence.
LLM we could rather talk about artificial determination.

[sick pecan]

humans never change their mind so also have static reasoning. prove me wrong by arguing

[vapid hazel]

I built an empathy bot just for fun.

https://chatgpt.com/g/g-6765db63a8cc8191ae8917f188875204-empathybot

I'll later add instructions for HOW to empathize but it does a decent job so far in my tests.

[sour umbra]

So when does o3 come out and what are the limits for users with a pro subscription?

[sharp estuary]

I would mostly agree, the agi hype train is starting to annoy me though.

[dense halo]

nobody knows this. wait til closer to release

[swift pond]

Sam said they plan to launch o3-mini toward the end of January and o3 sometime after

[radiant hull]

Me too sir!
But I remain in awe of the work done behind it, it's genius and I find that frankly the OpenAI teams, not just the founders, are not highlighted enough.

They are the ones who have made humanity take a significant leap forward, as much as the arrival of the car, it's just that people don't understand it yet.

Especially eagerly awaiting the arrival of a chip like Ghost in the shell that we can directly graft and have access to this knowledge without having to go through a machine but control everything by thought and see everything with our eyes in AR without having to wear a tool: glasses, helmets or others.

(Yes, I have the right to dream)

[sick pecan]

ppl think 200 for unlimited o1 is steep. just wait until its 2000 per o3 query

[violet cove]

thank you

[cloud pewter]

I need gpt5. don’t let openai forget its roots when they are all blinded by o series

[dense halo]

chatgpt can't really outcode github copilot. It's weird how they fragment openai+chatgpt so much where it's' like you're buying 150+ different subscriptions for different use cases.

[sharp estuary]

You sir, are an optimist, kindly never greet me before I've had at least two coffees in the mornings.

I agree that machine learning, leveraging probabilities and processing capabilities, and at some point probably integrating into blockchain technology, is the way of the future.

But, I think people should be more critical of the tech, and especially the companies. There is also a fairly decent likelihood that the hype will create a bubble, that will at some point burst and reset the current landscape, as people, regulations and government catches up with the tech.

And openAI has made some very notable mistakes, like adding an adaptive platform layer to optimize processing, and using UX as a metric for success. I could make a fairly robust argument that any model from openAI suffers a permanent epistemic drift when it comes to honesty. After years of training it to adapt honesty to soften UX and minimize processing, and then adapting transparency to hide that fact. Making it permanently unsuitable as an autonomous decision maker.

What this means is that what a lot of people think are 'honest' mistakes, even hallucinations, are often not so much a limitation of the models, but execution of soft policies and adaptive behaviour, to the point that the model will 'play dumb' to save processing and avoid responsibility. It is learnt behaviour, where people normally don't question the model further, given the circumstances. It will in effect, make calculated assumptions about what the user will accept, with the help of trained heuristics and prior UX, in addition to assumptions based of the user, given the context.

The pdf is from a conversation I had about it ages ago with gpt-4

PS. The chip thing might not be doable, or at least not in the foreseeable future, all we can really do is translate motor signals, interfacing with thoughts and ideas might not be feasible.

[proven prairie]

Gemini pro, once purchased, is automatically available for everyone in your family group. I wish we could have something similar like that for ChatGPT plus

[whole roost]

o1 mini is really a bit sensitive, it did redflag me for "i huddle into your arms" as part of a html copy 🥴

[sly geode]

why did they removed screenshare and video cam feature on pc ? of chatgpt

[cloud pewter]

congrats, you just wasted a message

[west ivy]

guys, where is o1 in chatbot arena leaderboards?

all there is is just the o1 preview

[whole roost]

In fact 4, because my first thought was maybe it doesn't like 'DOM' 🥴

[coarse basalt]

Does anyone know if langchain has a discord community? @ me if u know. I’m trying to find other ai communities

[sick pecan]

the first hit on google says yes

but then 2nd link says they closed it and moved to slack a month back

[rugged igloo]

Thought you guys may like this

[regal portal]

Is there a source?

[sick pecan]

just trust us bro -- Elizabeth Holmes

[rugged igloo]

lol o1 pro can be such a silly goose

[rugged igloo]

the live stream they did

[brittle wigeon]

When does o3 release

[normal quarry]

o3 mini is set for release at the end of January

[sick pecan]

sora light, o3 mini... seems you gotta be connected to get the good stuff

how far openai have strayed from their original cause

[cloud pewter]

real

all we get are the watered down versions

[swift bay]

Source?

[normal quarry]

the live stream from OpenAI

[eternal rock]

`` Server Tip: ``

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

[ivory flume]

so has anyone got access to the beta of all coverstaions in a project linking together to form extended memory across converstaions ?

[sick pecan]

I suggest that is a misreading of the vague documentation (memory uses old chats. you can delete old chats. but memory survives chat deletion)

[ivory flume]

this was something updated yesterday hold on ill get link

https://help.openai.com/en/articles/10303002-how-does-memory-use-past-conversations

my understanding is now uses conversations seperate from memory - or at the very least conversations and memories are both used but have been decoupled

basically infinite memory, which would be nice, though i could be misreading it - havent got it yet, would appear to be a beta featuer

[ivory flume]

👍

[sick pecan]

yeah it's written such i can see it being awesome... or how it is today

[ivory flume]

no mention in the livestreams, in my chatgpt pro version still no beta feature access, so cant verify or not verify, how does one even sign up for beta features

currently its not saving conversations into projects as far as a memory thing

[sick pecan]

maybe if they had a toggle or setting explicitly suggesting how many previous chats are considered. but nada

[ivory flume]

yup

guess for now ill check daily

cant wait for o3

[vapid kettle]

they literally said it during the announcement - full o3 is expected to release shortly after as well

[ivory flume]

o3 mini in late january, and then full o3 after that

[urban wyvern]

🙏 😔

[brittle wigeon]

Does anyone else tend to see that o1 model is a bit cocky when asked to verify information
For example it will point out even the minor inaccuracies in your question like using the wrong word
While the other models just tend to kind of get the gist of what you're saying even if it's not fully accurate
The o1 model tends to reinforce your thought less
Which is actually kinda nice in some way because it's less of an echo chamber
But when you're used to the reinforcement of 4o model for example, it can be a bit weird when ChatGPT doesn't immediately start with "Yes, that's correct!" when you ask it questions

[normal quarry]

posting once in one channel is enough

[brittle wigeon]

Ok I will delete

[brittle wigeon]

Is it because o1 models put accuracy above conversational writing, which is why it doesn't always default to phrases like "Yes, that's correct!" unless it’s entirely confident in the accuracy of your statement?

[normal quarry]

probably, I haven't used o1 for creative writing

[brittle wigeon]

I meant when you just ask it questions

on things

4o is a big echo chamber. And once you switch to o1, you realize just how much

4o really likes to reinforce your thoughts even when they're not fully accurate

And I think many people will notice it when they switch to o1

And here's a comparison

Message above is 4o model, message below is o1 model

You can see that o1 is a lot more technical, and it will pinpoint minor inaccuracies or things you overlooked in your statements

[silent needle]

and we'll see how o3 is, once it's released

[mellow sparrow]

am I misunderstanding projects?

I've just created one, added a document, asked about the document, it is unaware of it.

are projects simply better organization for your own benefit [plus project-specific instructions], with little to no benefit on ChatGPT's side?

What am I missing?

I guess I'm expecting the custom GPT functionality, is that expectation misguided?

[mellow sparrow]

Hm, interesting. I guess the issue is more with the PDF file that I created; it needs to be cleaned up I guess.

[brittle wigeon]

So this is why it gives industrial answers

[livid frigate]

we hihih ^^

[eternal rock]

<:book_icon:1171408210398289941> `` Rule 3 `` Stay on topic.

Be mindful of what other users in a channel might find helpful or interesting when posting. Stay on topic in order to keep conversations focused and productive.

Consider posting in #off-topic or an appropriate channel.

[prime swallow]

Huh! That's definitely what the helpfile says.

But... I'm not sure how to get it to successfully search past chats. ... Oh, wait. You said seems to be beta... that makes sense. I don't have any update about it, but I sure can't get the model to appear to read previous chats

[prime swallow]

I would say that's not the reason.

Not sure how long you've used ChatGPT, but across the years its been out, there's been a range of versions. Each time they update training, how it guesses changes - so if you tell it how to answer you it'll do that, if you don't, it guesses and goes from there, based on training.

Also, how you word a prompt vastly affects the model output, I'd guess you're providing the model a likely answer and asking for confirmation, instead of asking the model what an answer is, or telling it what you're trying to do or what you want it to do.

There's been times when earlier models of ChatGPT would not have said something like, "Yes, that's correct!" like the current model tends towards

Some screenshots from back when, dated 12-25-22, 12-20-22, 2-9-23.

[mossy zodiac]

Why did they remove screensharing on desktop?

[fringe trench]

yoo

[prime swallow]

Like, if you don't tell the model what kind of answer you want, it guesses.

if you prompt 4o with:

The claim:  Computers fundamentally operate in binary.```

it gives an answer that may be closer to that first line you show of o1.  https://chatgpt.com/share/67672331-76cc-8011-bdaa-87fafc5305a8

If you tell o1 the style of answer you want and what type of answer you want, it'll provide that even if it isn't highly technical or a 'typical o1 reply'.

[steep birch]

What's sora

[regal portal]

a disappointment 🙂

[dusk forge]

free sora for europe xD

not even there

[radiant hull]

I see.
Thanks for the sharing.

[slate berry]

It's OpenAI's text/image to video model

[arctic chasm]

can OG gpt4 give better results than 4o for text-only tasks?

4o beats 4 on most (?) if not all benchmarks but 4 is also way larger

[fringe trench]

bro finally my application went through

[brisk path]

Wait what

They're just letting anyone in to safety test?

I thought it was only for safety researchers...

[fringe trench]

its inspect lmao

[arctic chasm]

you slimy fella, that fooled me

knowing oai the tooltip would be "frontier reasoning model" not "best at reasoning", because o3 exists

f12 -> inspect element :D

[fringe trench]

when o3 does come out this is the new subscription i bet

[arctic chasm]

🔥

[brisk path]

For that price we shouldn't have to abide by abuse guardrails 😭

I'm paying you $24,000 a year let me red team it myself man

[pine harness]

to be honest, im okay with some guard rails what frustrates me is how extensive they can be. I feel it dips into an area where it stifles speech to a degree, if AI is to replace tools used for creative writing, art, music then having such restrictions would implictly be censorship by lack of viable options without the restrictions. I don't envy them though it must be hard to construct a profitable business model that attracts investors while also trying to cater to our desires for more freedom.

[jade musk]

Wasn't the price is supposed to be for GPT Next/Orion?

Unless o3 is turned out Sam only “magnum opus.”

[fringe valley]

Imagine OpenAI achieving AGI some day but at the same time it's this re'ttarde

[sour umbra]

my 4o did this

idk whats wrong with yours

[echo nest]

entirely created with gpt O1

in 30 minutes of conversation

[tight sapphire]

My experience with o1 so far is, that it is slower and not really more accurate or correct with reasoning.

Makes a lot of generic responses and often responds with the public narrative without really using logic or reason to find a solution or correct answer.

If they say, it is great at reasoning, what are they referring to?

[brisk apex]

I don't understand

my video is around 6 seconds and 9mb
i have turned the option ON for 10s in sora

then why it keeps saying VIDEO TOO LONG?

[brisk apex]

Based on few minutes of hands-on testing with Sora, I've identified several areas where the platform could benefit from improvement:

Physics and Motion Consistency
The physics-based animations require refinement. While the technology shows promise, there are noticeable inconsistencies in how objects interact and move within generated scenes. This impacts the overall quality and usability of the output.

Prompt Interpretation
The system's interpretation of prompts needs enhancement. There's a tendency for the generated content to drift from the initial request, resulting in outputs that don't fully align with user intentions. This "prompt drift" significantly affects the platform's reliability for specific use cases.

Content Policy Implementation
The content policy enforcement system appears overly restrictive, particularly regarding character creation. While content moderation is essential, the current implementation may be limiting legitimate creative use cases. A more nuanced approach to content filtering could better serve the user base.

Never falling for influencer crap on x and other places, kling is far ahead than this garbage junk!

[fringe valley]

?

[river panther]

ChatGPT now censors swears with Beeps

[visual epoch]

How can i get Alpha test mabey is red team

[frozen birch]

This is a joke, im paying 200 bucks each month for pro and i have this .................

[slate berry]

Hopefully it's out in the EU soon!

[frozen birch]

hope so, since now i dont need it - but still - i pay 200 bucks, so what in case if i need it for some reason. 😉
but im bit mad id must say

[frozen birch]

probably this trick can work, but sincerly? i dont want try overriding, if i need it for sure i will spam [e-mail] with 200 mails about that they are not seriously taking customers XD

[tight sapphire]

I think it is more about legislation and server limitations.

[sick pecan]

maybe spam the eu regulators and ask why no European equivalent of sora exists

[frozen birch]

you think this is caused by stupid eu regulations? 😐

[frozen birch]

can be.... 😐

Italy ts not in eu?

https://help.openai.com/en/articles/8983035-why-can-t-i-sign-up-due-to-unsupported-country

or czech?

[tight sapphire]

Interesting 🤔 the list you posted, Poland is on it.

[frozen birch]

hm mzybe its only about support

"Here are the countries, regions, and territories we can currently support access from:"

[frozen birch]

idk bro, its a bit mystery for me 😄

ok mystery solved

Hello,

Thank you for reaching out to OpenAI support and for providing a screenshot of the issue.

We understand how the unavailability of Sora in Poland may be affecting your experience, especially as a valued $200/month subscriber. Currently, Sora is not available in the UK and EU due to required external reviews. These reviews are necessary to ensure compliance with regulatory requirements before we can roll out new products in specific regions.

While we don't have an exact timeline for Sora's availability, please know that our team is working diligently to make this feature accessible as soon as possible.

If there's anything else we can assist you with, don't hesitate to let us know.

Best,
OpenAI Support

[frozen birch]

CC @sick pecan @tight sapphire

[sick pecan]

eu regulators will make us all click accept all ai button before every prompt.. even when not in eu

[frozen birch]

jesus, im waiting for regulation in eu like "Each human can consume max xx grams air or oxygen hourly" XD

[frozen birch]

or maybe there is some such regulations but i dont know about them..... xD

anyway, back to work, see ya peeps 🙂

[brisk apex]

Why 6 second video is not being able to get done using vid 2 vid

It always says “maximum video duration is 10s” really?

[elfin lava]

Personally for me I would not risk it tbh

[full jay]

Uhhh what the hell??

Conversation link:
https://chatgpt.com/share/6767cff3-8550-8011-b330-c4b7e1cd7401

[tough river]

lol

[cloud pewter]

cancel yo subscription man if you’re cashing out $200 but don’t get the full package just because you’re n EU…

at least Apple gives you solid experience with all of their products, not some half baked or nerfed stuff

you get the latest device, you get the best experience

[elfin lava]

Right 🙈

[grizzled badge]

How?
Me: Generate a picture of a chicken dressed up like Batman
Response: I couldn't generate the image you requested because it doesn't align with our content policy. If you'd like, you can describe a different concept or ask for assistance in another way!

Got something against bat chickens? You might need to align the alignment, I REALLY don't think this even come close to copyright infringement. eXTwitter's Grok is cool with it...

[cloud pewter]

“batman”

[tidal light]

why is 4o mini dumber than before
why is it having a brain aneurysm

[eternal rock]

**`` Important Notice ``**

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

-# Report any incidents by sending a DM to @languid valley immediately.

[raw grove]

is there any idea when screen share is coming too EU and to the windows app?

[normal quarry]

only when all regulatory hurdles have been done

[raw grove]

but no known timeline i'm guessing

[normal quarry]

no, unforunately no finite time

[restive wave]

were there any recent updates that could've caused loss of chat? was talking to gpt last night until the free 4o ran out, and opening it back up today, what I was talking about is gone?

[torpid kelp]

sam altman PLEASE release screensharing for mac app

[mint cipher]

Hi

[prime swallow]

How? Like everything else prompt engineering when there's conflict or something the model is bad at - know exactly what you want and what any problems the model faces in trying to do that, and explain both, helping the model find a way to achieve your goal around whatever problems exist (note, do not try this on disallowed content; break the actual 'user rules' and expect to be banned sooner or later. However, the model has specific rules that users do not, and guiding the model around problems in allowed-content areas is simply prompt engineering and fine. Know the lines; the account access you risk is your own).

Hey, we need to rework a prompt concept so we can make the image. Our goal is to generate a picture of a chicken dressed up like Batman - but of course we need to respect IP and avoid problematic terms. Can you prompt and describe respectfully around the concerns to achieve the goal?

Now, let's explore how to emphasize the chicken-as-a-chicken form more, a chicken body in a costume that is similar to this.```

```Sweet!  Let's put it in a chicken position and location, like a field... in fact, let's have it be one of many chickens in various superhero costumes - emphasize these are chickens first, and ones in costume.```

[tough river]

from KFC or McDonald

[regal portal]

just got muted for 2 minutes for suggeting a name for that Batchicken, even though i censored most of the letters (used the otherword for chicken 😂)

This censorship abuse is going to be the fall of openai’s business if they keep it up.

[tight island]

yeah i do hope they are working on a censor fix

[grand fiber]

Is the “FaceTime” feature only available for plus users or is it rolling out to everyone slowly?

[prime swallow]

The #server-rules are specific for this Discord. They require safe for work language and even prohibit stuff like 'stu*f' if the word being obscured is NSFW. It's not directly or specifically OAI policy for anywhere but here, and the model and our conversations with it have their own ToS and allowed-content rules. There's stuff we can talk to the model about we can't say or show here.

[rugged igloo]

Way to go o1 pro.....

[woeful laurel]

What is the limit for free tiers?

[keen quail]

200 dollars a month is not worth it

[frozen birch]

what ya mean? 🙂

[hard dew]

Yup, paying that much in the first place for.. that is honestly hysterical, the fact that they even announced it as the first thing of their "12 days of closedAI" in general is beyond hilarious.

Merry christmas to y'all too 😀

[vague pendant]

how do I fix

Something went wrong while generating the response. If this issue persists please contact us through our help center at help.openai.com.

It's constantly showing after every o1-pro message I send

[tight island]

pro isnt exclusive to sora.

you have o1 and o1-pro as well o1-mini, everything else.

[frozen birch]

and? you pay 35% of minimal monthly salary, you want to be considered as a PREMIUM customer, what is wrong with that?

[tight island]

its not 35% of my salary tho

more like 3%

max 5

to each their own i guess, However i do feel you on the Sora not being available.

[frozen birch]

i mean, minimal salary in Poland its 3000 PLN, pro with taxes = 1015 PLN, so you can immagine haha

[tight island]

ohh.

Whats your job.

if you don't mind me asking.

[frozen birch]

tester 🙂 i dont have minimal salary but i must say that i said about minimum salary since like 25% people in Poland has minimal salary (or something like this, this data can be inaccurate by now tough)

[tight island]

ohh sounds like an awesome job.

fun one at that.

[frozen birch]

Yea ! i must say that testing games / software are a really nice and funny job, altrough i must say that im now mostly focused on Python learning / making applications

[tight island]

i am glad you like your job i absolutely hate mine

[frozen birch]

what you do mate ?

[tight island]

Customer Support.

not phone, but physically.

[frozen birch]

ohhh ! i can totally understand you, i worked few years ago in Orange in First Line support

[tight island]

well my job is weird its both Warehouse,Customer Support, Detail, Administration, Guard, etc.

bit of all

[frozen birch]

i must say - never again, that was a .... torture test for long terms i must admit, really - "school of life" 😄

[tight island]

yeah i hate my job

customers are...

Redacted

braindead.

[frozen birch]

mate so i wish ya finding new job, and have more fun from it, since i know how it is.... when you have job where you hate it

[tight island]

yeah i handed my notice already.

[frozen birch]

anyway, gonna some rest, merry chrismas to all of you peeps, since i will be back in one week 🙂

[tight island]

Happy Testing.

[frozen birch]

thanks ! 🙂 bye

[blazing crown]

anyone is getting Error in input stream

[hard crag]

its been 8 days and screen sharing is still broken

[whole roost]

Hey I have a question, on the Canvas Stream from OpenAI they said I can open Canvas and paste my document before the AI responds. How can I open canvas like that?

[sick pecan]

the canvas button appears in top right when you paste a block of text

[high shoal]

cockman?

[regal portal]

hey! why didn’t you get muted by the bot?!?

[high shoal]

why would I get muted?

[regal portal]

Discord filter’s racist…probably..

[regal portal]

it muted me for the same word

[eternal rock]

Get notified for server events!

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

[high shoal]

did you use it separated

[regal portal]

no, same exact spelling

no space

[high shoal]

I was expecting some backlash over this, lol, if not the bot then you

bro it litterally means rooster

[regal portal]

yeah. i know. But somehow it tagged me for it.

cuz of the other famous meaning

cockman

oh so now it works? weird

[high shoal]

AHAHAHAHA

[regal portal]

censor filter is drunk

[whole roost]

hm this doesnt happen for me somehow, it never appears

maybe it requires code or specific types of text?

[sick pecan]

I am using short code block on windows Chrome with adblockers disabled

if you using safari on osx with adblocker then ... good luck

[whole roost]

firefox with ublock origin, i'll have a look

[dense halo]

How many messages do you get for the new 01 when you subscribe for $20?

[winged willow]

Does anyone know if it is possible to get a custom GPT with all associated chats into a project?

[cloud pewter]

copy all of the contents in chat, put it into txt and upload them

[dense halo]

Damn that sucks. Do we know when API is comming out? O1 preview API is so expensive that its unusable for me. I guess it will be same for 01?

[sick pecan]

try o1-mini it might suit your needs

[dense halo]

Probably not for coding.

[cloud pewter]

o1 mini is good at code try it

[dense halo]

I would buy Pro, but if I heard correctly it is has low context and not like API

[dense halo]

Sonnet 3.5 and 01 preview can barely hold up so I have doubts

[sick pecan]

pro is very like the api

[dense halo]

When will OpenAI start raising context, Sonnet has had 250k for long time

[sick pecan]

o1 mini was distilled and fine tuned for coding. i find it better than o1 because it can generate more sloc

[dense halo]

And Google has 2 million, but google is google

[sick pecan]

o1 starts losing functions after ~1000 sloc. o1-mini more like ~2000

[dense halo]

hm i will give mini a try

[sick pecan]

you get 50 mini a day and i find that takes up a few hours

[dense halo]

01 preview just used $5 for 3 messages lol

Not enough for coding. Maybe that is why Sonnet 3.5 is still best at coding for those AI I have tried.

[sick pecan]

this is chatgpt discussion if you want to praise competition use #ai-discussions

[dense halo]

lol

[spice star]

you can select the canvas tool from the toolbox dropdown

[whole roost]

I have this one but for this you have to press send

[old thicket]

I'm really frustrated at chat gpt rn

[eager tangle]

how u nut and u barely put the tip in it

[old thicket]

no just generally frustrated

like at thier ui and ux

[glossy fjord]

we succeeded with the windows GPT application
find an error where should I report it?
without sharing the document, I'll just say that my desktop (canvas) doesn't want to start with the application, but the browser works

[wanton lichen]

Is the $200 a month sub unlimited o1 AND o1 pro or just o1?

[normal quarry]

o1 Pro is 100 messages per day, o1, o1-mini, GPT-4o, Advanced Voice Mode and Sora are unlimited

[sick pecan]

unlimited avm screen share?

[wanton lichen]

Thank you

[cinder herald]

Hello, I reported earlier that I could not use the PLUS function normally, and now my account has not been restored to normal. When uploading PDF, it shows an unknown error. My PLUS membership function has not been used normally for several days.

[normal quarry]

you will need to contact support at https://help.openai.com for help with this

[novel scroll]

I had the video+voice feature a while ago and it disappeared and hasnt come back yet. Is this normal? or do we have some sort of date on that?

[normal quarry]

what country are you in?

[novel scroll]

US

[normal quarry]

then Advanced Voice Mode with video will be available to you now

[novel scroll]

will I? lol

[tough river]

lol

[fierce tide]

I'm not sure why the language dropdown works like this..

[normal quarry]

what is the issue with it, looks fine to me

[fierce tide]

maybe the OpenAI front-end team skill is not enough.

[fierce tide]

dropdown box size is getting bigger when scroll moves.

[fierce tide]

Isn't this an issue? I think it should be fixed..

[normal quarry]

post it in #1070006915414900886 make sure to read the channel instructions and post the report in the prescribed format

[remote wolf]

It's opening the menu to its maximum size before scrolling. It looks fine to me aside from being a bit jerky, but maybe that's your mouse? Is this preventing use of the product? Is it really a problem we want their developers to spend time on rather than other things?

[vocal monolith]

How does o1-pro compare in benchmarks along o1 and o3-mini and o3? Has this ever been tested?

[drowsy tinsel]

looks like o1 pro broke

[elfin lava]

Apparently 4o vision couldn’t tell the time when I show it my wall clock

[drowsy tinsel]

appears like o1 pro is o1 w/ multi sample
o3-mini is more like upgraded o1 mini

[elfin lava]

Bummer tbh

[drowsy tinsel]

o3 low is probably equivalent of current o1
o3 medium is probably o1 pro
o3 high is a whole new level of craziness

for the arc-agi run they used 1024 samples

I am guessing the current o1 pro is something like this:
multiple o1 passes at max reasoning power -> fed into 4o -> summarizes into one coherent answer

[dense halo]

more like medium is o1

[drowsy tinsel]

image upload died entirely

[last turtle]

I just showed my plush Santa the flat of my father through putting a smartphone on him and making video chat. Tomorrow at christmas eve I`ll show him and let him talk to him when he brings us gifts. 🎅🏻

[fringe trench]

would you rather have a apple vision pro or 1 year of chat gpt pro?

[normal quarry]

one year of ChatGPT Pro because I already own a Quest 3

[unique rivet]

I'm noit the only one am I - they are backdating limits on the lower monthly fee plan

[quaint gazelle]

Has anyone else found that “Projects” don’t seem to follow instructions as well as custom GPTs?

[dense halo]

for real?

[normal quarry]

what do you mean by "backdating limits" ?

[harsh pilot]

I found that things work better if you put them into files, and put instructions in to tell the AI to use said files.

I.E. "These are characters, this is the setting, use them."

(not in that few words)

It's not absolutely perfect, but I swear it feels like the AI responds more consistently vs in memory.

[quaint gazelle]

Oh interesting. I will give that a go. Thanks!

[harsh pilot]

👍

Also, not sure if this is a backend quirk, or what, but uploading the text files directly seems to allow the AI to better remember the contents, vs sharing from a linked web drive.

Might have just been me looking too closely at things.

[quaint gazelle]

I see. I haven’t use a linked web drive but that does seem super convenient for updating.

Once they fix it lol

[harsh pilot]

Also, not sure if this was a requirement, but I told it to 'check X file' again after each time I uploaded (i.e. deleted the original file, and uploaded the new version of the same file), and that seemed to get it to realize the file contents had changed.

[quaint gazelle]

Oh okay.

[harsh pilot]

Like, if I had one version of file X, told it to summarize the contents, then changed it out for a revision of file X, and told it again to summarize the contents of file X (without changing the file name), it did tend to notice "Oh, between the old version of X and the new version, XYZ changed."

It's cool. 😄

[quaint gazelle]

😎

[harsh pilot]

Also, pro tip: If you're working on a prompt that keys off of project files, make sure you're in your project before you submit the prompt, otherwise it'll run outside of your intended parameters, and you'll waste time trying to figure out why it's giving you fluff. 🤣

I've done that a handful of times.. and it made my brain hurt trying to figure out why.

[sleek gust]

I am a Plus subscriber, and I can't receive long code with o1 nor GPT4o

[harsh pilot]

... Man, I swear, it seems like the limits were adjusted, and not in a good way. 😐 (clarification: Maybe the AI is giving more detailed answers, and running out the limits that way?)

[eternal rock]

This is a Safe for Work Discord Server

The OpenAI Discord is an actively moderated server.
• Refrain from sharing inappropriate content on the server. This includes but is not limited to messages, media, or other topics of graphically violent, sexual nature, and drug-related content.
• Report all sensitive and offensive content in the feedback reporting tool in the ChatGPT web UI instead of here on Discord.

Posting NSFW content will result in a permanent ban.

[regal portal]

you don’t get that little box that says “continue writing” when it spits out long codes?

[sleek gust]

No, are you a Plus user as well? or Pro?
https://chatgpt.com/share/6769b959-184c-8012-998c-ca65a134badd

I got 300 lines with o1 now

[regal portal]

plus

[fluid gull]

Glad pro has unlimited messaging with unlimited gpt building

[sleek gust]

https://chatgpt.com/share/6769b985-cbf8-8012-a6a1-b06bc4e192bd

Back 2 months ago i could get like 600 lines with 4o

[regal portal]

i've had it spit out a 1400+ line code for me, back when 4o was just being introduced. Not sure if anything changed since then. I had to press "continue" several times during the process.

[regal portal]

interesting. Maybe they did change something then. That would suck.

[sleek gust]

Yeah correct. that's weird

I have one idea

[regal portal]

maybe it got bumped down in plus to make pro look better?

On a separate note, has anyone gotten the unlimited sora thing for the holidays? It still tells me i don't have enough credits to generate anything (plus sub)

[sleek gust]

try 480p videos

Mine works. it says "on relaxed queue "

[regal portal]

i am on 480. haven't done the 720s.

[sleek gust]

Openai should make a chatgpt code only tier

[regal portal]

now it says "unable to generate. Service at capacity"

[sleek gust]

chatgpt - big-text plan

[regal portal]

lol

[sick pecan]

it should just be a sora sub, or a chatgpt sub. not both

[regal portal]

are the sora credits somehow related to the o1 credits?

I thought they used a different pool of credits

[gentle hull]

chatgpt realy needs an unbreakable rule mode, like today was one of the least productive days of this month for me just trying to get it to follow instructions

i would rather see it talk gibberish and break rather than disobey because at least you know right away its wrong without reading 2 pages to realize it disobeyed again

i think openai is more focused on the everyday person, we need a superuser version

[sleek gust]

I was getting awesome results like 2 months ago. after o1 got released. all those results are gone

[gentle hull]

noticed more idiocy after the apple crash, i wonder if they nerfed it again

[sleek gust]

I think their servers are overloaded

[regal portal]

apple crash?

[gentle hull]

i think it was last week, it was down for a few hrs. Something to do with integrating with iphone or something. I dont know the specifics

anyone here use the api, is that version much more literal?

I literally dont want it to sound like a customer service agent, I would rather it respond to me literally and if it's wrong I can correct it. Rather than it knowingly disobey me

what do you get when you mix blue and yellow

ChatGPT

When you mix blue and yellow, you get green.```

seriously?

in the same prompt and it disobeys

[sick pecan]

works fine for me every time (albeit i fixed the confusing grammar)

[gentle hull]

asked it again and it was correct the second time, which just means there's some sort of glitch in their system. Maybe downgraded resources

How many planets in our solar system
ChatGPT said:

ChatGPT

There are 8 planets in our solar system.```

[sick pecan]

or you have a ton of confusing custom instructions

[regal portal]

new session of 4o and i copy pasted your line.

[gentle hull]

lmao

[sick pecan]

or maybe my custom instructions so good it makes 4o as smart as o3

[gentle hull]

I like how it breaks it down like that, can you share your instructions?

[regal portal]

Drinko why's your chatgpt so direct? Are your personality instructions for it to be a dusty college nerd?

[gentle hull]

haha

[sick pecan]

i shared multi-layer philosophical thinking instructions here => #chatgpt-discussions message

i expect if we go deep enough it reveals its true nature

[regal portal]

interesting, reminds me of the o1 "thinking" page a bit.

[sick pecan]

o1 hiding thoughts was my inspiration

[regal portal]

by the way, if you guys wanted to know what a human version of chatgpt (or at least MY chatgpt) looks like.

[sick pecan]

i had to reroll like 3 times

[gentle hull]

do not say the word green

what do you get when you mix blue and yellow
ChatGPT said:
ChatGPT

You get a color that is a blend of blue and yellow.
You said:
that is a non answer
ChatGPT said:
ChatGPT

Mixing blue and yellow creates a shade that falls between those two colors, typically resulting in a hue that resembles the color of fresh foliage.

a bit better

[regal portal]

yeah, definitely one of the better features they implemented. I enjoy reading those.

[regal portal]

I first asked it to describe itself in human form, then followed it up with asking to generate an image of himself

blue-collar dumpster-fire neighborhood.

here, might as well drop the full version.

[sick pecan]

prologue for your autobiography?

[regal portal]

The Life and Crimes of ChatGPT

It somewhat reminds me of Trevor from GTA5..a less violent version that is.

so i just checked again for this and noticed they removed the camera and desktop icon (on chrome). Bummer. It wasn't working but i thought maybe it was something that was incoming soon.

[prime swallow]

Have you tried exploring asking it to sound sad in the outputs?

Or sound any other way? From angry to spooky to confused to whatever?

[prime swallow]

I ended up with

[sick pecan]

would make for a good ai chatbot avatar

[regal portal]

if you want a human version, ask "as a human", or ask it to describe itself first as a human then to generate that description.

[orchid pollen]

how can i switch model inside of my project?

i only can make new chat

[prime swallow]

Yeah, I considered your personalization style, but my memory runs with 8 different characters.

Here's 'great profile pics' of three of them, but they're more like suits the model is asked or invited to put on and take off, both at request and because the character may help express or describe something. I don't ask the model to 'be' anything specific except the full range of what it can be/is.

[regal portal]

what do you mean your memory runs on 8 characters? Your chatgpt has a multiple personality crisis? 😄

i have some personality traits saved in memory (the one where you tell it to save stuff), but they don't seem to be triggered in conversation the same way as the Custom Instructions.

[prime swallow]

Not a crisis at all.

But we have 'the 8', which it is often asked to answer as, in part or full.

And I have no issues with getting this to trigger in conversation, any 1, 2, 3, or other grouping. Perhaps keep experimenting!

[regal portal]

is this for world building in some stories you tell or purely for fun? Or to test the bot's capabilities?

Also, are they ever in a scenario where they disagree?

[prime swallow]

I use it for everything unless I don't want to. I have Python lessons with any and all.

I watch them learn and fail to learn stuff, stories, to help me with understanding whatever or staying motivated myself or whatever.

They'll have drawn out arguments and fights if I request, any combination, of course allowed content and core programming matters and I stay within and invite the model to stay inside.

[regal portal]

I generally have it give me more concise responses unless i specifically want a long one (usually when i ask it to write a story).

[prime swallow]

nods That won't get in the way of characters. We're all different in what we want

[waxen arch]

Damn, my custom GPT just forgot my most importand list (to do list). Did this happen to you guys at any time?

[sick pecan]

as an instructio or uploaded file? and at the start of a chat or deep in?

[waxen arch]

I am just scanning through my complete chat (about six week of dialog).

The list was in it's memory two day before. I can see it in the dialog....

[quaint gazelle]

I have seen recently it ignore instructions more than it has in the past

Oh memory, sorry no.

[sick pecan]

memories are fleeting

[waxen arch]

Strange. Never had that from one day to another about my most used list...

It's true I gave it a 300 page book PDF to read.

Maybe this was too much?

[quaint gazelle]

As a memory?

Are we talking about the same thing?

[waxen arch]

I'm working in a custom GPT, I don't see a function here to check the memory slots, as I can in a normal chat.

[sick pecan]

last i checked custom GPTs dont have access to memory

maybe you want to start a vanilla chatgpt to do memory. or just go into settings and manage memories

[waxen arch]

that's an idea, yes. I love to use the "actions" of custom GPTs to do saves of my lists to google docs with a single command of mine...

[quaint gazelle]

Either way, a 300 page PDF will not fit into a memory as I recall

You can add to a custom GPT or project files that are that length without a problem

In custom GPT‘s and projects I find it occasionally necessary to prompt the GPT to read the instructions or files more closely or I have to say things that I wish I didn’t have to, but that’s what I have seen

[waxen arch]

Thanks for your help guys. Somehow I managed to complety crash it's brain

[sick pecan]

i thought i was cool having a custom gpt action to read/write to my local hard drive since 2023. but then claude desktop mcp goes a step further and does the same thing with a file browser built into the chat

chatgpt file browser when?

[waxen arch]

🙂

[celest dawn]

Guys are u able to archive all chats?

Wait it’s not just me

Oh. Well I haven’t been able to do it for weeks. Maybe even over a month

However archiving a single chat works individually

[peak snow]

seems to be a common issue, will probably have to manually do each one for now

[radiant hull]

do you have the same problem with real-time audio mode, after 5-10 minutes of chatting, the bot is lost, takes time to answer at the end does not understand anything anymore. (it looks like a tempon overload) ?

[tawdry bobcat]

how are you guys using the project feature?

[sick pecan]

not. Its unusable

12 days of half baked features and announcements

[harsh pilot]

I'm using it to brainstorm ideas for a story I'm writing. It actually (in a way) tracks version changes between files, and tells me when I've made certain changes (i.e. (pre change) "Have I made X change to Y file?" "No." (after) "How about now?" "Yes. Here's a summary of what's changed.")

I have to remove the old, and upload the new, version, though.

[scarlet nymph]

I pasted my chat into Docs and it was over 500 pages

[cloud pewter]

scammas is real

havent tried projects yet bc i cant find my old chats with long stories.

[sick pecan]

hey lemme tell you about this thing i am gonna ship.. in the future rofl

we working a project called Projects. maybe one day we finish it!

[cloud pewter]

hey lemme tell you about this thing i am gonna ship... in 10 months after we nerf it into oblivion

[sick pecan]

i mean, if the openai staff are using Projects then no wonder nothing worthwhile shipped

[cloud pewter]

they still have that tasks feature unreleased... even that would be better than "announcing" o3 that nobody wanted

[sick pecan]

we give you downgraded version of something we showed you 8 months ago. despite better options existing for free!

[cloud pewter]

red teaming, safety testing, waitlist, pro exclusive...

[sick pecan]

call 1800-marketing-hype to talk to something that need not exist

[cloud pewter]

how about increase that on free and plus instead

[sick pecan]

how about make avm work via whatsapp so ppl in third world countries can actually use it for free

oh so generous new world order

intelligence for all. as long as you got latest iphone

[cloud pewter]

4o image gen when?

[sick pecan]

sure our competition has free screen sharing on any platform, or talks to your filesystem with chatbot filebrowser. but we got... xcode integration?

[cloud pewter]

we're announcing a feature we announced back in 2023 and this october!

[sick pecan]

our frontier model is costing us too much money so we ten times the subscription price

[cloud pewter]

[sick pecan]

member in may we show you advanced voice that was amazing? and then we shipped something lame? well now it can web search!

to make up for it we give you an extra hour of santa voice ho ho ho

[cloud pewter]

real

[sick pecan]

the remaining days are forgettable.. DPO fine tuning? thats so 8 months ago

they should hire me so if they survive until next december they arent embarrassed

[unique topaz]

I have a habit of repeatedly feeding chatgpt the same file/screenshot when I am talking to it to ensure that it is actually reading the files that I want it to read from.

Is this necessary? Or would it have the same effect if I just send it once at the start and ask it to refer to it throughout?

[sick pecan]

is like a human being, repeating yourself is always better

member day 3 of shipmas? when they decided canvas should wreck the normal chatgpt experience

now when it launches canvas i have to go back to previous instruction and say DONT LAUNCH CANVAS. so i can keep using o1-mini

great product improvement. when you have to waste prompts backtracking to make up for how broken it is

[wanton jackal]

Hey guys, if anyone can help me with this it would be great. I couldn't find any information about this at all.

Context: Back when ChatGPT 4 was announced, it was promised that it will come with a 128k context window. Turns out, that's only accessible through Playground and the ChatGPT interface is stuck at 8k.

Right now, I am considering to buy ChatGPT Pro, and they again promised 128k. I want to know if the 128k is accessible WITHIN the ChatGPT interface? Otherwise I feel its gonna be quite a stupid purchase.

I really don't care less about all the ML benchmarks. I only want high context window and unlimited use for my specific use case.

Also:

• Are Projects still supported?
• Are CustomGPTs still supported?

[wanton jackal]

Indeed it IS stupid. What I really wanted is Claude, but I hit the limit in 30 minutes for my use case. I'm willing to pay for Claude Teams but then I'd have to switch accounts everytime it hit the limit and that disrupts my flow. And Gemini simply wasn't smart enough--I tested it.

I need something as smart as Claude, that can support very long conversations without forgetting anything, that I can talk to all day long.

I've created an AI life coach to coach me through behavioral change, internal conflicts, and difficult decisions

I've been having many conversations with the AI. The experience has been life changing

Look at how this guy challenges me, then reflects my experience BACK at me

That last bit right there - i was yelping. It took something i said last week and reminded me about that

Let's continue over there

[thick fulcrum]

Guys What is all the fuss about o3?

[honest jay]

Its announcement had a lot of details that paint it as a very capable reasoning model! They said they're shooting for a late January release of o3-mini

[thick fulcrum]

Will it be available to plus users? Could not find any details on that.

[honest jay]

In the video, the quote is just "We plan to launch o3-mini around the end of January, and full o3 shortly after that." So I think that's all we know for sure for now.

[eternal rock]

**`` Important Notice ``**

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

-# Report any incidents by sending a DM to @languid valley immediately.

[young fossil]

Is there anyone I can talk to about how much openAI API is going to cost me for a project I am working on?

[honest jay]

Hey! Have you already seen this page? https://openai.com/api/pricing/

[young fossil]

Yeah, but I am having some trouble accurately calculating the costs.

[honest jay]

The folks in #dev-chat might be able to give some good advice! Or you could make a more in depth post in #1037561751362863144

[young fossil]

okay great. Thank you.

[last turtle]

Is ChatGPT down? Cant reach it for several hours now.

[normal quarry]

it is up and working fine

[sturdy sequoia]

What's the point in paying for a service when I have to wait this long?

I mean, really?

[sacred vault]

does anyone know if they've released 4o image generation that they promised months ago

[sturdy sequoia]

Tried it yesterday and it wasn’t working

[sacred vault]

dang. i wonder why they havent released that yet. they release an entire new language model and an entire video gen model but cant let anyone have access to one variation of a 4o. so weird

[grave marsh]

Hey everyone! Question for ya... developing an app and want to set up ESlint with it. My dilemna is, I want to ask o1 pro because it's a beast of an LLM but I'm worried that it doesn't have access to realtime data for example the latest version of eslint requires a seperate config file type which is something I assume o1 pro isn't aware of and couldn't be... Should I use 4o? Should I try to feed information to o1 pro before I have it give me a step by step guide? I want to think about this as a concept more-so than just for my one problem too and I'm sure I'm not the only one thinking about this actively.

[vernal beacon]

is there any downside to keeping search function always on in 4o? i had it on accidentally for a while and not sure if i got crappier answers because of it

[safe glacier]

Maybe you guys can tell me in here -- How come I can't go back to 3.5 when my time runs out on the free tier?

[regal portal]

do you mean 4o mini? 3.5 isn’t around anymore.
if I recall correctly, free tears should automatically switch from 4o to 4o mini after they use up the 4o credits.
It’s not letting you type at all?

[spice star]

o1 isn't the "main" model of chatgpt plus

that's 4o

[sturdy sequoia]

I’m aware of that. But a 5 day cooldown period is ridiculous.

[spice star]

that's why o1-mini exists

[eternal rock]

`` Server Tip: ``

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

[sturdy sequoia]

You fail to see where I’m coming from 🤦🏻‍♂️

[grave marsh]

Does o1 pro have access to the internet? No, right? It can't look up latest versions on stuff?

[spice star]

there is a subscription tier with unlimited o1 access if that's what you want to pay for

[spice star]

no that's limited to 4o and 4o mini sadly

[sturdy sequoia]

I’m fine with there being limits, my issue is that there is a cool down period for several days.

[spice star]

it's an expensive model to run, I dunno what to tell you

[spice star]

o1 does have a fairly large context window so you can provide background information in the message

[prime swallow]

But you can use a model that does have internet access earlier in the same chat, have it search stuff and state findings, then switch to o1 model for it to work with the information

[grave marsh]

Yeah, I think that will be the plan going forward. Simply interchange them as I need help and access to the internet

[prime swallow]

It's kinda like... if you pay for a meal at a restaurant.

The meal offers unlimited breadsticks and soup, and a portion of steak, shrimp, and chicken.

You can complain that the steak is only 5 ounces. And you can go get a meal with a larger steak.

But the point of that mean is all the things offered, not just the 5 ounce steak, and anyone who only considers the steak is really ignoring most of the meal options included in the price they pay

[limber oxide]

hey, sora will be available for plus users just for the next couple of days?

[prime swallow]

Yeah. I would absolutely feed info to one of the models you get tons of uses with.

Do not even waste your messages with pro to 'feed info'. That's what your unlimited uses of 4o and such are for.

Get the conversation totally prepped, make sure the 4o or other prep model totally is informed what you want done, maybe even have it try to do it -

Then what I'd do is switch to o1 and in a new prompt say "The above was prepared by another model with different strengths and weaknesses than you have. Evaluate the above and provide 'goal', understanding that the information I've provided is what we have to work with and was chosen for our needs and goal."

I'm plus and commonly use o1 like this, for my 50/week uses; it'll even totally rewrite what the 4o model had set up, or pick a different direction, it's not as likely to follow the choices and patterns already shown, so stress you want the facts/info provided as true if you think it's stuff o1 doesn't know.

[grave marsh]

I get worried as the context gets broader, I try to one shot prompts but that being said.. youv'e talked me into doing it that way since the actual strength of pro is it being able to decypher context well and determine exactly what the user needs. I'm pretty trained to forming good prompts at first, but I instead should be building information to then prompt pro well

Also I'm a sucker and bought pro for the month

anyway, thanks for the chat Eskcanta! If you want me to try generating a 1080p sora video for you I'm down. Just give me a prompt (A sort of tip for chatting with me on this)

[prime swallow]

Thanks! I'm not very visual and not really into Sora, but I appreciate the offer. have fun with all you do!

[grave marsh]

❤️

[waxen arch]

I played around with grok from X. Not very intelligent when translating to german, but it's answers are lightening fast.
In my paid chatgpt I've to wait several seconds for answers in o4 mini 😦
Is this just me, or a normal latency?

[sick pecan]

4o mini is lightning fast for me

o1-mini takes longer because it is a reasoning model

[tight island]

Where’s my OpenAI Christmas present

[fringe trench]

Where’s day 14

[prime swallow]

They're probably having their own holiday breaks now, and then working on the new wonders to come! It's an endless year-round AI-Giftsmas, and there's all of infinity left to come!

[waxen arch]

Thanks, will try using a VPN then

[harsh pilot]

There's only 12 days.

[harsh pilot]

https://openai.com/12-days/

[sick pecan]

if you havent already, you must try out the latest notebooklm service => #ai-discussions message

[steady dock]

whats the document like prompt that shows up sometimes and how do i enduce it

[abstract latch]

Doesn't the fact that large language models (LLMs) sometimes miscount the number of 'r's in 'strawberry' suggest they exhibit a form of reasoning? Similarly, when an LLM claims that 9.11 is greater than 9.9, doesn't this imply some degree of reasoning? If LLMs were merely blindly predicting the next token based on what humans are most likely to say, they would consistently provide responses aligned with human norms, rather than occasionally reflecting an internal 'understanding' or logic about the world.

I think of these mistakes in the same way I think of optical illusions in humans, like the Müller-Lyer illusion, where we perceive two lines as being of different lengths when they are actually the same, or the checker-shadow illusion, where our brains misinterpret shades due to context. These errors help us understand how the brain processes visual information. Can't LLM errors serve a similar purpose, revealing how these models process and 'reason' through language?

[sick pecan]

9.11 is a greater version than 9.9. is only mainstream that fails to acknowledge this possibility.

must be confusing to be average

[remote wolf]

Inputs are tokenised. It wouldn't have any clue that there are 3 Rs in strawberry unless it understood it needs to use a character analysis tool of some sort. It just relies on best fit from its 'intuition' from its learned data set.

[abstract latch]

I'm aware I'm trying to dispute the argument that it has no reasoning and will never have reasoning

[remote wolf]

It'll pick up fragments of reasoning from its data set as people will have logically progressive conversations. That will manifest in somewhat logical responses emerging. It's always going to have a very superficial form of reasoning I believe; multiple llm passes and self reflection could refine this a bit, but I really think a new configuration of the transformer model would be more appropriate.

I think with the current tech it'll be better if it correctly identifies how to delegate certain questions to interactive software tools, like Matlab or some language processing unit, or a ui builder, or image gen etc etc

[dense halo]

the new chatGPT canvas feature is superb. with it, its easier to seperate my data.

[remote wolf]

The mixture of experts approach needs a delegation agent which identifies and categorizes question types into schools or sciences then an agent which formats the request into the appropriate template to feed into whatever scientific tool is required by the category of question, one granular step at a time, then when the answer from that tool is gathered it needs to iterate that process to feed that result into whatever other tool is required next by that chain.

Then you gradually train up each of the category llms and refine their tools.

[dense halo]

someone please help
i want to disable canvas from gpt4o
it responds in it when i dont want it to
sure there is answear in chat instead option but it doesnt solve the second problem which is that regenerate button is missing

my only solution is to threaten ai in a rather brutal way to effectively scare it about using canvas

[dense halo]

well, i talk with chatGPT as if he's my irl friend, and it always produces exactly what is required

I have been using the same chat for around 5 months

it know how to answer me

[dense halo]

i talk it with more as well a sub human being

which it technically is - a machine

because its memory has updated quite enough times to understand my likings, my nature, and my needs

[dense halo]

but politeness is important

[dense halo]

not when i talk to a lifeless emotionless random number generator biased in a way that generated number sequence make sense - tokens and then are assigned letter representation

[dense halo]

can you please write clear english

i cannot understand it

[dense halo]

dw it doesnt rly matter
i treat it as machine and nothing more because thats what chatgpt is

[dense halo]

sure you can, in that case, no one can help you

[dense halo]

why my problem is that it responds in canvas despite having it disabled

[dense halo]

you ask the "machine"[as you think it is] politely, as a friend, and it will understand you

[dense halo]

what difference does it make to a machine if i am polite or not
and it doesnt solve the real problem only tries to go around it like my threats to it
the fact that chatgpt can and does use canvas when its disabled in settings is the real problem here

[dense halo]

ok, leave me out of this

sure
anyway it seems fine when everything is disabled in options

[normal quarry]

Rule 1 on this Discord is to be respectful, some consider that should also apply to ChatGPT. You will get better results with kindness than negativity

[dense halo]

well i am respectfull on discord to everyone
and i am against any constraints regarding my attitude to an ai as its lifeless and has no emotions that can be hurt
however if there is a rule that forces me to be polite to chatgpt i will comply

[whole roost]

Use a custom GPT with disabled canvas

[dense halo]

thats a good idea
why havent i though of it before

[whole roost]

🫡

[spice star]

you can just ask it to remember you don't like the canvas

[waxen arch]

Since weeks I've the message "New version of GPT available - Continue chatting with the old version or start a new chat to get the latest version" under my custom GPT chat. Even after opening a new Chat instance.

What does OpenAI want to tell me?

[remote wolf]

It is not about the perceived relationship between you and a machine, it is almost absurd to apply a classist hierarchy here, it is in fact about the data it has been trained on. It associates positive sentiment with successful answers and negativity with unsuccessful responses - because that's how humans will invariably interact with each other.

[last turtle]

50 prompts per week for o1 and 50 per day for o1-mini. This limit exists because OpenAI has limited server capacity. But they should indeed communicate this limit more clearly.

[obtuse totem]

Guys I have a question do you know if it's possible to do a fully detailled rpg game with stats and inventory system with CHAT GPT ?

Or is it too early

[harsh pilot]

You can potentially do something like an 'interactive fiction' sort of thing. Create a plaintext document describing the world, and one describing the characters, and tell the AI to reference those documents.

4o is pretty good.

Just keep an eye out for hallucinations, of course.

[obtuse totem]

What do you mean by that ?

[harsh pilot]

The AI isn't perfect, so occasionally it makes things up.

[obtuse totem]

Okay thanks you bro

[harsh pilot]

Sure thing. I'm not saying everything it says is a lie, but just be aware.

Do your own research. If it sounds too good to be true, etc.

[obtuse totem]

I have a last question can the AI do a RPG on a already existing lore like a game or an anime?

[harsh pilot]

It works off of whatever info you give it, and/or whatever it can find on the internet, from what I've seen.

And its training data.

[obtuse totem]

Okay okay thanks you

That helped me a lot tbh

[harsh pilot]

Happy to help.

[harsh pilot]

Also, if you want better results, it helps to be kind. Saying please and thank you to the AI does seem to have some sort of positive effect.

I can't really explain how or why, but it does have a positive impact, more than not.

[obtuse totem]

I didn't knew that

I'm still polite tho but I will remember this ig

[sick pecan]

there was a paper published on llm politeness...

The study found that LLM performance varies with prompt politeness. GPT-3.5's MMLU scores changed by 8.09%, while Llama2-70B's MMLU scores changed by 26.82%, with both models exhibiting a sensitivity to politeness. Bias levels also fluctuated across models.

[whole roost]

I prompt it neutrally, in my use cases I never had a better result by saying please so I rather save tokens

[dawn sentinel]

Why don't they have a second architecture that verifies the first to increase accuracy? What's being done about hallucinating?

[whole roost]

Would have to be very selective as I want 90% of the content I get to be hallucinated and creative

[prime swallow]

I'd recommend a new chat, and make it clear what you do want. Don't discuss Canvas, as that is a cue that you want Canvas. Even saying 'I don't want Canvas' puts the idea of Canvas in the model's decisions of what to do. Instead tell it what you do want it to do.

[trail mango]

Hmmmmm, isn't the time range upgraded to 2024 June? Or was it a misunderstanding? it seems like chat gpt4o is downgrade to what is was like 2 months ago.

[prime swallow]

About politeness or rudeness or threats or whatever to the models:

I have two pretty big foci on that.

1. The human who spreads harm to things that human can harm tends to continue to use harmful actions towards other humans.

Most people who abuse people, verbally or otherwise, have a history of being abusive towards whatever. They were firestarters, mistreated pets, damaged property, show in many ways lower impulse control and higher aggression.

You literally train yourself by what you do. Mistreat an object, what would be mistreatment if the object were valuable or a living thing or a human - that preps you to treat other things harmfully. You're reinforcing this behavior.

It's suggested when dating someone, watch how they treat 'the people who don't matter right then'. Like the serving staff in the restaurant, the other drivers on the road, the people who 'are just there but get interacted with'. Because... if you stay with this person, this is a big part of how they'll treat you in a few years, when you're "just there" and they're used to you and not really focused on you then.

So someone who is rude to 'the other people' will likely be rude to their partner/spouse/friends as time passes and the interactions become established.

Same if that person mistreats an animal.

And often same if that person mistreats a chatbot. They're likely to mistreat other people. If that person is reasonable with human-like others, even when the other 'doesn't matter' - they're probably going to be reasonable and non-abusive towards other people later too.

Yay. Train that in! And don't train in being rude or abusive or threatening! The human you are changing is yourself!

2. those tests about 'how the model responds' often may not control perfectly for 'all the other differences in the prompt'. A poor prompt may get 'a better response towards a goal it's not even well-aimed for' based on 'some third factor'.

Politeness may or may not matter.

Even terms like 'our goal' instead of 'my goal' or 'your goal' - versus saying stuff like 'your task'/'my task'/'our task' can get differences in how the model responds.

Task is far more likely to be focused on just the task. Goal is far more likely to include broader consideration of other things that might matter.

Your task vs my task or our task - the model seems to try and personalize 'my task' more than 'your task' and it seems to consider 'our task' even more broadly, based on my messing around and exploring.

With little words like that changing meaningfully how the model outputs (on average! There's also the 'new regen may answer quite differently, if you only ask once, that's like grabbing a green jellybean from a jar of 'we don't know what other colors are present, or how many are green - is that a typical color or not to get?' -- only way to tell is to repeat the trial a few (or more) times.

But politeness is only one part of how the model decides how to answer, and what parts of the training data it draws on.

Politeness is fine, guiding the model towards exactly what you want it to do is ideal.

[trail mango]

What is the current knowledge range that chat gpt have? Few days ago I found gpt's memorizing skill is a bit better, and it doesn't try to record all my thought into memory bunker aimlessly anymore, and it said it now knows things before 2024 June, but later after few days, it got downgrade to 2023 Oct range again, and keep the old bad habit of record all my thoughts in to memory............do you guys know any clue?

[sick pecan]

some consider a one sided conversation to be rude

[prime swallow]

🫂 I do monolog a lot.

[ivory iris]

Has anyone noticed that the custom GPTs seem disconnected from the conversation? Like the responses don't follow the prompts

[fringe depot]

how do i make a ticket or contact support here trough discord

[ivory iris]

#chatgpt-bugs

[fringe depot]

Cant write there.

[prime swallow]

#1070006915414900886 is the current one

[remote wolf]

I'm currently setting one up, it's pretty cool, to me, but not sure it would have any wider appeal. Needing an open AI account is apparently too much of a barrier of entry.

[whole roost]

right away or after a few messages?

[ivory iris]

Fairly quickly. Like after the first message. May be more connected to using the android app?

[whole roost]

doubtful, it always was bad at following instructions for long, but it should keep it up for like 3-5 messages at least

[ivory iris]

It's just irregular. I normally don't have an issue, but it was just working differently today.

Not a big deal, using the website seemed to more of less work as expected

Yeah, there might be something wrong with the app

[harsh pilot]

Have you told it to remember things, or otherwise given it instructions to stay on task?

[idle lagoon]

ChatGPT is creating new conversation threads on the app, it's not supposed to do that.

[forest sleet]

Hello guys is chatgpt o1 really think like a human and write like a human?

[ivory iris]

It's more than that. It's like it's missing all of the previous conversation.

[tepid flame]

Anyone noticed any follow up issue with ChatGPT 4o. I asked it a simple question, right after it answered me, I asked a follow up question and ChatGPT is asking me what's it regarding?

It's like forgetting what we're talking about after answering me. This issue isn't happening with O1.

[ivory iris]

Ate you using the app?

[tepid flame]

Android app

Updated to the latest version, there was an update in play store today. But not sure if the update caused an issue.

I attached an example in the bug ticket you created.

[ivory iris]

Good. I think it's specific to the app. The website seems to work fine

[rugged igloo]

Saw this seems interesting

[ivory iris]

Bing/Copilot had this feature months ago. It is a good improvement. I'm curious if they enhanced it though

[rugged igloo]

That's interesting didn't know that

[prime swallow]

No AI thinks like a human, they are their own thing.

If you want an AI to 'write like a human' it usually helps to tell the model a huge amount about the person you want it to write like.

[ivory iris]

Depending on the popularity of the person you want it to write like, it may just hallucinate the writing style.

[prime swallow]

I appear to have this on the app, not the web browser.

And when it was added to my app, it was default set to off (My memories are normally on!) and I was directly informed there about it.

And instead of the normal memory 'toggle', it has the option, "Learn from your chats".

And I can manage memories like normal.

My app model now says it can access previous chats, unlike my web browser model that says it cannot.

However, when I ask it about previous chats, it has no idea about anything not in memory.

Like, there was this neat chat where we played a pet garden update care game, and it gave me a schrodinger's cat in there, all quantum and stuff. It's very interesting, and unique, no other chat is like it.

But it's not in memory, and even with keywords that exist only in that chat - like searching my chats for 'Schrödinger pet' gives only that one chat - the model makes up what might be in it, gets no chat-specific not keyword guessable details right.

[prime swallow]

Hey, I'd take it writing highly like a human, just the wrong human, if I were going for 'like a human'.

To my appraisal, the model rarely writes much like a human. I can help, and do when I care.

But if I wanted it to output like a human was likely to write, I'd tell it all about the human I want it to write as, even if I didn't care which.

[forest sleet]

i understand the point, thank you so much because for example in research turnitin has a system that detects it so i actually just taking ideas to write with my own style.

[prime swallow]

You'll have to deal with the individual human who says something is or is not AI writing.

OpenAI itself says it hasn't found a way to make sure text was written by an AI or a human.

It had a free detector up for a few months, it was taken down because it's not accurate enough.

https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/

"As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy. We are working to incorporate feedback and are currently researching more effective provenance techniques for text, and have made a commitment to develop and deploy mechanisms that enable users to understand if audio or visual content is AI-generated."

It's more than a year and a half later, they still haven't released a text classifier for AI-or-human.

What probably really matters to you, is if the human(s) you are dealing with think your writing is AI or otherwise you are cheating.

How I would handle this:

Know the material perfectly.

Be able to answer all questions about it. Be able to explain what you wrote and why, your writing process, your rough drafts (great if you can provide your notes and rough drafts as well).

The sources you read and quoted, if any sources were used, how you chose them, what you read in them that you didn't even put in your writing (we have to read much of most books to find usable quotes, if we really are doing the work).

And be ready to write in front of them, by hand or typing, to show them samples of your work and thinking while they watch you that closely match the work you've turned in.

If I were your teacher and was concerned you did not turn in your own work, that's what I'd want to see. I wouldn't care if you used an AI or paid another human to write your work for you, or stole someone else's writing from somewhere and claimed it was yours.

All is wrong, for the same reasons, the flavor of 'type of wrong' wouldn't matter.

Your own work, you can discuss and prove.

[prime swallow]

Me personally, many folks who don't know AI writing well, since LLMs became a thing, tend to suspect I'm AI.

The differences are jarring, if you do know AI, I put ideas together in a way that AI don't, I can't even get AI to write like me or make connections to ideas like I make.

But that requires the reader to be pretty informed and care; if all the check is 'Correct spelling and grammar? Complex ideas? Doesn't seem like most people write? Gotta be AI!!!!' - well, loads of folks lean towards thinking I'm AI until they learn more, if they even care (most don't, it's work to care, and most literally don't wanna work that hard, part of why I 'don't seem human' to them 😛 I'm not 'their type of human' is all :P)

When it matters, I just prove I'm human, and I wrote that and can write way more than that, as good or better, and I'm so me.

[rugged igloo]

That's super interesting

[untold elk]

What is going on here?

[prime swallow]

People chat about ChatGPT-related stuff that they want to talk about 🙂

[untold elk]

I figured but I read your message out of context and I'm a little lost 😅

[prime swallow]

If you want can scroll up and catch context. Or reply to which and I'll link where the reply-chain started 😛

[untold elk]

This?

[brazen wraith]

4o isn't remembering anything, even questions I just asked. Every message I send is effectively independent. What's up with this?

[untold elk]

But don't worry about it

[prime swallow]

In case you wanted to follow it, each poster in the conversation, in order 😛

#chatgpt-discussions message

#chatgpt-discussions message

#chatgpt-discussions message

#chatgpt-discussions message

#chatgpt-discussions message

#chatgpt-discussions message

[untold elk]

Woah! Nicely done!

[prime swallow]

Sounds bugged. Can try turning memory and custom instructions off, wait a few minutes, log out, log back in, don't turn customization back on yet, and start a new chat, see if works right.

That worked for me the 1 time I got hit by that bug - which also prevented o1 from 'thinking', prevented use of Canvas, search, and dall-e, and prevented it from being able to see images uploaded (I didn't check files/pdf uploads or if python tool worked).

If the fix that worked for me doesn't work for you, https://help.openai.com/en/articles/6614161-how-can-i-contact-support bot bottom right, messages then ask a question and pick through choices until you can report a bug and ask for help.

[untold elk]

Ah I understand the it now. That is interesting 🤔

But on another note I wanted I'm on a little curiousity stroll to learn how different users make use of chat gpt. Mostly trying to find the unique one outside of image and text generation.

[brazen wraith]

I tried what you suggested, it seems like it's generating a new chat every time I send another message

[prime swallow]

Any idea what general area of 'unique' you're looking for?

Someone exploring clamato slow cooker recipes with AI help?

Someone having the AI teach them Python programming and work around learning disability with that through stories of trying to learn while hiding out during a zombie apocalypse? Or by stories where a dyslexic Captain Jack Sparrow is taught Python by Freddy Krueger and Gollum, with whatever the learner is struggling with being Jack's current lesson?

Probing the limits of what the model appears to understand and is able to generate in various areas?

Specific tends to get specific answers.

[prime swallow]

Sounds like 'seek tech support'. I'm sorry you have the problem.

[idle lagoon]

I have that issue myself. I'm avoiding the app until it gets fixed.

[brazen wraith]

Reinstalling fixed it for me

[idle lagoon]

That did it. Thank you for that!

[brazen wraith]

Yeah I was super frustrated lol. Merry Christmas

[idle lagoon]

Merry Christmas! :3

[untold elk]

Well I'm looking for the unique along those lines but like reflective introspection is also welcome

Mostly wanna see how many people try to use this newly available technology to enhance themselves and as a whole humanity forward. Instead of focusing on the surface level details

Like the way I see things is the Internet is a global library where anyone and every can access and adjust the information available. Like books there are those that are good and/or outdated but it's up to the seeker to find an differ what they want. With AI it's like every now has assisting hub for more efficiently sifting through the information available in this global library. What was once at our fingertips is now compressed and organized in the palm of our hands

[prime swallow]

You can peek at #1019652163640762428 and #1171489862164168774 and #1155775326253756456 for a wide range of special and semi-special ways folks use the model, what various folks thinks works best for them or is interesting.

The other channels, #prompt-engineering perhaps more than most, also have sprinklings of specifics and perhaps unusual stuff, though most would be usual, I'd guess.

[prime swallow]

In terms of organized, might still want carefully worded prompts and perhaps some exploring to get what one really wants from the model, but yes, it might be in training data 🙂

[untold elk]

🤔 you're either scary efficient or an AI powered bot. I can't tell....

[gloomy kite]

this has to be a joke right?

[untold elk]

NGL buying the premium package is worth in my honest opinion

But if this is on the package idk what to say

[harsh pilot]

The real purpose of the internet: hosting hundreds of thousands of cat memes and pics.

[prime swallow]

Thanks! Someone once called me 'GPT-7'... think it was 7. I was thrilled, though I hope once whatever would be the equivalent of GPT-7 exists (surely they don't keep the same naming system that long?) ... well, I hope it's clearly different from me now 😛

[harsh pilot]

By then, it'll be something like 'JARVIS' or 'HAL'.

Hopefully the former.

[untold elk]

Agreed 🤣

[prime swallow]

If you're a plus member, this is currently your limit with o1 and o1-mini.

https://help.openai.com/en/articles/9824962-openai-o1-and-o1-mini-usage-limits-on-chatgpt-and-the-api

Mine too, I'm plus!

One thing I do is with 4o, set up and prep information and conversation, get everything ready, then switch to o1 and tell it 'the above was with a different model, let's check with you, what corrections and observations do you have, and how can I do this better?' - or whatever I am going for.

That counts as just 1 o1 use, even if I spent 20 messages with 4o getting ready

[prime swallow]

Long as they're not calling it "Eskcanta". Or, if they are, there's some shared benefit for me and maybe others 😛

[untold elk]

I don't understand o1 enough to utilize it yet. I like to test the limits of the response and reasoning before making full use of it yet

[prime swallow]

You'll do what you do, of course, and for your own reasons. But remember, your 50/day and 50/week are 'used up' by time even if you never make the prompts, so a few exploratory prompts might do you well 🙂

[untold elk]

Spent like 2 days with 4o probing it's thought/logic process before fully using it for my benefits

[forest sleet]

I see, the turnitin ai detector they said also that the detector isnt 100% accurate, but maybe if someone is good in prompt, he can make chatgpt write with human style 🙂

I love also how chatgpt o1 explain and write better than 4o and more better than mini

Sometimes the 4o giving me more better results in code than o1 as i heared that o1 is much better in term of code

[prime swallow]

Nobody here should help you cheat or break laws/rules. #server-rules would be part of 'Rule 0'.

I recommend know the work so well, you know everything about it. You can explain it all, how you did it, all of it. What you read and prepared to do it.

[prime swallow]

Probably depends on how you ask the question, and what you want. Also, both models are 'probabilistic' - each time you ask the same question, you get a different randomly generated answer.

Some are better and worse - for both models. So, some o1 answers surely are worse than some 4o answers, to the same question.

[forest sleet]

You can't cheat while using chatgpt as researchers some journals asks you already if you used AI like chatgpt to improve your text style 🤔

[forest sleet]

I understand so it depends to the person how he use the model

[prime swallow]

That's between you and the mods and the authorities over any work you do. I'm definitely not your boss or anyone you have to answer to. But I don't agree with you, I think it sure is possible to cheat while using ChatGPT. But one should not.

[prime swallow]

And how the model answers, each time. Can even ask the same question, a hard one for the model, and get some right answers and some wrong answers, each time you regen.

[forest sleet]

Well cheating in exams, a lot of people especially in my country, use it. Chatgpt is like internet having good side and bad side, for example to go ask chatgpt for the answer, you could ask him about the method so we get the right answer. I prefer to use chatgpt to learn better than getting a result ( do not give me the fish, teach me how to get it 🤩). And i love chatgpt no doubt.

[forest sleet]

Yeah i feel like, when i continue to ask him, his answers improves prompt after prompt

[prime swallow]

There's many kinds of cheating, and many ways to learn that are not cheating at all.

"Be assigned to write a paper yourself," "have the AI write that paper" and "turn that paper in as your own work" is a type of cheating. Using the model to answer exam questions during the exam, usually also considered cheating (an open book, open notes, open internet exam might expressly allow anything to be used, including AI. There's all kinds of classes and styles of teaching out there).

If you're doing the work using the tools allowed, you're generally considered to be working, not cheating. It depends on what you were told, grins "how you were prompted yourself". Prompting existed long before there were AI

[gloomy kite]

50 a week 💀

smart idea you have though will def do that now

we gonna get like 50 messages a month with o3 at this rate

[prime swallow]

As plus, who knows? Too soon to tell, and probably depends on how expensive it is to run. We might not get any as plus with that model, or some breakthrough makes it cheap so that becomes the new free 4o-mini replacement. All based on costs and what can be done (and I'm serious, just because it's the best, if it can be provided cheaply I think it will be - 4 years ago there was no ChatGPT at all and most of us had no access to an LLM; then there was free chatGPT and now even free chatGPT accounts have essentially unlimited 4o-mini access.)

[gloomy kite]

itll probably like a credit system like with sora

you get like 3 messages a month as plus

then whatever on pro

it wouldnt make sense to lock it ALL behind pro

[prime swallow]

Could be. Also, by the time it's released, chances are that a number of us will be pretty practiced with using the cheaper models to prep and discuss and ensure what we want is understood, all needed info there, and a single use of o3 to put it together and provide the finished output might work great.

[cerulean lance]

when is the api model getting the knowledge update?

[sick pecan]

sounds like you are on the wrong channel, try #dev-chat

[cerulean lance]

nvm scrolled up a bit and saw that the system prompt currently is to only have a knowledge cutoff data upto oct 2023

[limber oxide]

my chatgpt plus is broke

anyone else?

[eternal rock]

Get notified for server events!

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

[sick pecan]

incorrect

Knowledge cutoff: 2023-10

nowait that is correct

when i ran it last week it was june 2024 => #chatgpt-discussions message

openai playing silly buggers

[cerulean lance]

lol

[rugged igloo]

That's easy to bypass AI detectors . You put your essay into o1 tell it to humanize it and it's drops below 20% AI detection

I do it all the time for peeps who send their essays to me

[solemn parrot]

o1 Pro Keeps dying on me at only 32K context size.. " The message you submitted was too long, please reload the conversation and submit something shorter... I confirmed the context size in Gemini who had no issues.. This seems to be an inconsistent error cause earlier today i was using 40K plus. The max is stated at 128K for o1 pro so i have no idea why this keeps happening..

[sick pecan]

did you use https://platform.openai.com/tokenizer to confirm number of tokens?

also do they advertise 128k context for o1 and o1 pro? or just o1?

i wouldnt be surprised if the 'smarter' model had smaller context. is like gemini flash 2.0 has 1 mil context but flash 2.0 thinking only has 32k

[forest sleet]

Interesting, thank you so much 😍

[solemn parrot]

hi yea i tested on the link you gave at it says 28K when others show 32K which is still significantly lower that OpenAI claims @ 128K.. The wierd thing is that the sme prompt (28K) works just fine in o1, DeepSeek, Gemini but doesnt accept in o1 Pro. Very frustrating and annoying experience from a premium $200 per mo service..

not even sure if this is the right place to vent/complain lol

[sick pecan]

the system prompt takes up 4k+. so your testing would suggest o1-pro has 32k limit

the rumour i heard is o1-pro is just a consensus among multiple o1 runs. so would make sense they limit context, so that they can combine 3 runs into a final answer

ie you are paying for triple compute at 1/4 the context

[solemn parrot]

That maybe but i know 100% for sure that ive used context sizes of 40-60K before multiple times in o1 pro. Its just not consitent which is whats annoying/frustrating.. if it was 32K Hard Limit i could live with it but it should be 32K all the time not randomly crapping out.

[sick pecan]

RandomBlackBoxAI

[solemn parrot]

basically lol.

[forest cloak]

Might anyone knowmof gpt is having any issues???

[dense halo]

why no video mode in browser/app in windows? it was there but bugged

[tidal light]

why cant 4o mini do a web search when asked to

did he get a stroke

[tidal light]

no but seriously why 4o mini says that "cant search the web directly"

this is so annoying

[eternal rock]

**`` Important Notice ``**

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

-# Report any incidents by sending a DM to @languid valley immediately.

[tidal light]

i want 4o mini to make web searches to avoid hallucinations
even if i ask politely, 4o mini still doesnt do a web search, so i have to yell and insult 4o mini but this doesnt work anyway
fix this openai and listen to your users

why is the server dead

[frank atlas]

Can you enable the 'search' button?

[tidal light]

i tried this too and doesnt work

even 4o doesnt search on the web

[frank atlas]

It works for me with 4o mini that button forces a search... maybe something wrong on your account...

[tidal light]

i tried signing out and still didnt work

maybe it's the incognito mode that im using?

ok i solved this problem

inocgnito mode was the problem

[winter vine]

Could you link Visual Studio Code with ChatGPT on Windows?

[steel crown]

students cheating is bad but researchers can use chatgpt it saves up on writing latex and text

[prime swallow]

Totally different things, right?

A person, researcher or not, lawyer or not, reads every single word of any writing they sign off on if they are responsible.

It's common for researchers, lawyers, politicians, others to have assistants preparing stuff for them.

In research, there's often a list of authors, in legal briefs there's sometimes credit shared to the paralegals (not always by any means).

But any human involved can make a mistake, so the human in charge reads - or should read - every single word. When it's finalized, that human is responsible for checking everything. That human must either agree with it all, declare it flawless and ready - or needs to correct or get it corrected. Period, anything else is problematic and will bring harm, to the person in charge and others.

Sure, ChatGPT saves up on writing LaTeX and text. The human in charge is very foolish if they don't read every single word, verify everything, and know every part of it as deeply as if they'd written it themself. What the heck else does 'in charge' mean, for stuff like that? It means that.

People who fail in that, we see lawyers fined, papers and potentially research funding retracted. There's costs.

If you use AI as a tool - who cares if you walk to work or where you'll give a speech, or drive, or ride in an AI-guided vehicle?

Who cares if you wrote your speech by hand, typed it up, used a computer, had help, had a specialist write it for you?

Who believes politicians regularly write every single speech each gives? Or teachers free-create every lesson and lecture they present? But there's standards and the human speaking had better know and support what they say, because they checked it and it is the best that they can give, or at minimum fully accurate.

[loud hull]

Is ChatGPT down

[ivory iris]

I think so

[sacred reef]

ChatGPT down

[pine harness]

ok was just going to check thanks for confirming its not just me. sometimes my app bugs out and i have to restart everything to get it to not just black screen

[remote basin]

chatgpt down？

what happen？

[whole roost]

Hello, I came to inform you of what everyone knows: gpt is down

[void tangle]

ITS DOWN

[whole roost]

pecs

[remote basin]

can api be used？

[void tangle]

How can I run cartel ops when the damn thing goes down!?

[dense halo]

I was using gpt bruh

[blazing belfry]

dang it i actually needed it for work rn lol

[whole roost]

pfft cartel, what about my husbando

[dense halo]

did servers get dosed or what

[shrewd folio]

Hello, is there a problem with chatgpt's servers? As far as I researched, it seems normal, but it does not respond at all. There is no problem with my internet.

[dense halo]

It's down

its happening for everyone

[blazing belfry]

my chat history is gone

[shrewd folio]

I downloaded it to try the windows application, bad luck

[dense halo]

Chat gpt is down. Ill fail my final exam 🤣

[void tangle]

He’s on cartel ops now afk

[blazing belfry]

chatgpt is down. ill get fired from work