chatgpt-discussions

[hard crag]

released today?

[jade knot]

i tried "o1-preview" but doesn't work

[narrow viper]

The ability to reason means it will consider details more. Compare this to 4o now which can tell you two different things in two consecutive responses and not notice its own folly.

[hard crag]

is it out?

[willow oar]

For people having an anurism because they dont have access to it: Its coming out tomorrow for you and you alone

[hard crag]

its out omg

i have it

[jade knot]

it should be out for tier 5 and above as they mentioend

[narrow viper]

LOLz ... I'm going back to work. Will check later.

[fluid flume]

is o1 the default 4o model?
or has the ui and api not updated yet?

[last carbon]

what is tier 5

[hard crag]

guys what should i ask it

[jade knot]

tier 5 on the API usage we are on tier 5 but nothing shows on the playground

[graceful hearth]

I live in germany and I got finally some days ago access to memory. But it is so underwhelming now.

[last carbon]

oh word, thanks.

[narrow viper]

"If a chicken and a half lays an egg and a half how much does a pound of butter weigh?"

[keen mesa]

same

[narrow viper]

There was no mention of "memory" in the announcement, but I wonder if the new model (rather the ChatGPT coding in front it) has been or will soon be enhanced to use memory. v4o says "Remembering" a lot but it sure doesn't seem to use any of that memory.

[tawny valve]

Fr it seems like it doesn't think about memory if it gaslights itself into believing things, maybe this model will

[feral lichen]

allot of no responce only the stupid how is this answer without any content

[narrow viper]

wut?

[fleet granite]

They can’t even run some simple RAG function to keep a minimal amount of data up to date in the context window if you want to consider that route but in reality realtime fine tuning with highly curated data sets needs to be a priority for something to have a grasp on what is current reality.

[pearl quartz]

wth, nothing

just a redirect 😦

[narrow viper]

"Memory" is that "simple RAG function" and yeah v4o hasn't been up to it. Unfortunately I don't believe that's related to the model. That's purely a function of the client that uses the model, which in this case is "ChatGPT". So I doubt we're going to see changes in this area for a while. The team that works on ChatGPT itself seems to operate in a totally different time continuum.

[jade knot]

"Rolling out today in ChatGPT to all Plus and Team users, and in the API for developers on tier 5: www.openai.com/o1" i'll just chill for a couple of hours i guess tehy will update the API, playground UI later?

[narrow viper]

Just remember folks (and few will) the new o1 is a PREVIEW. It will be awkward to start. Give our friends at OpenAI time to, uh, fine-tune.

[fleet granite]

Exactly which it seems very redundant to utilize memory for base information, where fine tuning is the solution.

[hard dew]

Rolling out today in ChatGPT to all Plus and Team users, and in the API for developers on tier 5: www.openai.com/o1

Let me guess, except literally almost everyone who lives outside america once again lmao? Like EU.

[analog patio]

i havent gotten it yet

[rare wave]

bro I'm still waiting for advanced voice mode😭

[narrow viper]

Well, fine-tuning is augmenting the model itself. Memory should not augment the model. RAG pretty much defines memory.

Me2 Arcu. 😦

[analog patio]

i think a lot of people are including me broski

[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.

[nova fable]

Just got acces on Plus

After refresh

[narrow viper]

I really look forward to o1 enhancement for custom GPTs too. We create them to do special things and they're not much better than vanilla.

[dawn wharf]

This model is amazing!!!! Thk open ai ❤️

[fleet granite]

It would be nice to know if what they mean by “reasoning” enhancement isn’t just a crafty system prompt

[opaque narwhal]

Can't seem to find the O1 Preview as a plus member?

[narrow viper]

Not at all CMcD. See the web page: openai.com/o1/

[upper wadi]

somebody already tried the new model how is it?

[narrow viper]

Alex ... as always it's a slow rollout ... patience.

[rare wave]

have you seen the videos? this thing is actually thinking

it's incredible

[upper wadi]

but what do they actually mean by thinking.. like what makes it different from other models

[fleet granite]

It’s sounds like a similar approach to what the Reflection llama model tried to release

Teach to reason through the dataset

[upper wadi]

but this wouldnt affect the inference speed

[rare wave]

it can take up to 20 seconds to answer

saw some examples taking that much

[upper wadi]

does the inference speed fluctuate?

[narrow viper]

According to the documentation (yes folks, read it) the mini model will go faster and sometimes better than the o1-preview.

[upper wadi]

so sonnet is no longer the best coding llm?

[tight robin]

no

[narrow viper]

Remains to be seen...

[upper wadi]

cursor + o1 mini is the new kid in town

[narrow viper]

From the doc openai.com/index/learning-to-reason-with-llms :

Similar to how a human may think for a long time before responding to a difficult question, o1 uses a chain of thought when attempting to solve a problem. Through reinforcement learning, o1 learns to hone its chain of thought and refine the strategies it uses. It learns to recognize and correct its mistakes. It learns to break down tricky steps into simpler ones. It learns to try a different approach when the current one isn’t working. This process dramatically improves the model’s ability to reason.

[rare wave]

A huge amount of compute is shifted to serving inference instead of pre/post-training. LLMs are text-based simulators. By rolling out many possible strategies and scenarios in the simulator, the model will eventually converge to good solutions. The process is a well-studied problem like AlphaGo's monte carlo tree search (MCTS).

[tight robin]

Text generation involves nuanced language understanding and context, which can't always be effectively addressed by brute-force simulation strategies alone.

[west peak]

refresh as in log out/ in again?

[tight robin]

While LLMs can simulate various scenarios akin to AlphaGo's MCTS, their text-based nature introduces complexities beyond traditional game simulations.

[west peak]

not available yet on my chatGPT teams subscription

[nova fable]

Nah, i have chatgpt as a pinned tab thats always open. Before the release i didnt have the model, shortly after, i refreshed the page and had it

[west peak]

Has anyone with chatGPT Teams subscription the new model?

[brave aspen]

I do not have access to the model yet

[zealous canopy]

Fascinating there's a mini? Is this just a new system prompt for chat gpt?

[low crag]

sometimes i take some time to think as well.

[tight robin]

pretty much yes

[zealous canopy]

Thats... kind of lame?

[hard crag]

it's not a system prompt

[narrow viper]

Text processing of a LLM (note it says on the tin: "Language" model) is only one part of the puzzle. That's not actually AI, though of course AI has been used in the process. Actual intelligence is measured by the ability to reason, to understand failure from success, to devise a different approach to achieve a goal. Don't confuse the two.

@tight robin not correct

[tight robin]

that's unsubstantiated

[brave aspen]

is it even going to be released in europe?

[rare wave]

why no one did it then? we all have gpt on the api

if it was that easy

[narrow viper]

A system prompt provdes instructions to a model. The model is completely separate from the prompts. o1 is a new model, not just prompts on the front-end.

[rugged igloo]

I'm taking queries if anybody is curious I'll post them in here

[tight robin]

read the blog post

[narrow viper]

It says everywhere, this is a new model. They trained the model. This is not just a new system prompt. Oy ve.

[rare wave]

We trained these models to spend more time thinking through problems before they respond, much like a person would. Through training, they learn to refine their thinking process, try different strategies, and recognize their mistakes.

Through training, not through prompting

[brave aspen]

he might be a bot

or an npc

[eternal rock]

Get notified for server events!

Click the button below to pick up the <@&1261377106890199132> role.

[brittle junco]

in fact, the new model is just auto-prompt?

[narrow viper]

no sneax, read the info on the site

[tight robin]

not exactly

[narrow viper]

devy ... learn something here ... you're not getting this yet.

Wow, anyone else see the Codeforces Elo / percentile stats!!??! Essentially : gpt4o= 11th percentile (not good at all) o1-preview= 62nd, o1 (in development) = 89th, and a further fine-tuned "o1-ioi" (2024 International Olympiad in Informatics) is in the 93rd percentile !!!!!!!

[rare wave]

it's absolutely incredible

my claude subscription was just renewed I want my money back

[narrow viper]

lolz

[jolly yew]

have you tried the famous strawberry test already?

by asking it how many r are in the word strawberry? this is something where every ai model has always failed at

[rugged igloo]

I heard you only get 50 weekly chats with it. So I am scared to use it

[jolly yew]

the mini gives 50 the regular version gives 30

[hollow spoke]

the answer is probably in the training data, we need to ask another word xD

[jolly yew]

but idk if they are sperate

[jolly yew]

right maybe..

maybe ask it multiple words in one prompt? to be safe

like strawberry, carryforward, corroborator

[brittle junco]

do you have a respond ?

[fickle cypress]

What’s the rate limit for -o1?

[static ridge]

another clown announcement. "try it here, where it doesnt work!"

[rugged igloo]

I'll ask it to give me 10 sentences all ending with the word strawberry and then count how many r' it used.

[zenith falcon]

Is it a gradual roll out?

[jolly yew]

hmm i think using multiple specific words like the ones i mentioned earlier might be a safer bet

i highly doubt they trained the word corroborator specifically xD when it comes to counting rs

[hollow spoke]

They said realesing, not realeased.

[static ridge]

the page literally has buttons to try it. Why announce something with buttons to try something if it's not out

[brave aspen]

No it's esthetic roll out, if you're ugly you won't get it

[jolly yew]

up to you ofcourse. make sure to share it with us 😄

[rugged igloo]

Sure, here are 10 sentences all ending with the word "strawberry":

1. She loves the taste of a fresh strawberry.

2. The garden is full of ripe strawberry.

3. My favorite jam is made from strawberry.

4. She wore a dress colored like strawberry.

5. He bought a carton of strawberry.

6. The smoothie flavor is strawberry.

7. Birds often eat the garden's strawberry.

8. They made a pie filled with strawberry.

9. I drank tea infused with strawberry.

10. Children love ice cream flavored strawberry.

The total number of times the letter 'r' is used is 33.

[vapid hazel]

Anyone know if o1 will be coming to custom GPTs?

[jolly yew]

frick it failed then right?

[zenith falcon]

It's so over

[jolly yew]

10x3 is 30

[static ridge]

only 13 Rs off

improvement!

[jolly yew]

so it still cant count letters.. great

[rugged igloo]

Yea...

[neon raven]

anyone here know if its possible to get an API from a customizable chat GPT?

[jolly yew]

no its not an improvement at all if it doesnt actually count it

[static ridge]

I dunno why people are so obsessed with wanting it to count letters. That's not what it's for

[hollow spoke]

So still no AGI.

[jolly yew]

its the model hallucinating

[static ridge]

sarcasm

[hasty nimbus]

I want to use o1 on the API but I am usage tier 4 not 5 😔

[jolly yew]

o srry my bad 🤣

[neon raven]

what about CGI

[static ridge]

that's not a hallucination though

[median horizon]

So preview o1 has 30 weekly messages, was there any word on the mini version?

[jolly yew]

i dont get how this model is supposed to be good at programming when i t cant even look at how many Rs are used or any letters for that matter

[haughty sail]

anyone else getting this with GPT-4

[hollow spoke]

does it show the hidden "thinking" prompt too?

[graceful hearth]

30 WEEKLY MESSAGES. THATS NOTHING BRUH

[static ridge]

writing code vs running code. you're kinda asking it to run "code"

[rugged igloo]

Yes it does thinking then understanding words then using the correct words

[visual bloom]

I just tried this. 4o answered that strawberry has 2, but 4 answered 3.

[static ridge]

I can't imagine it's "good" still, but they're saying it's better, which makes sense.

[jolly yew]

lol.. its a well known issue with ai models till this day, that it cant read what it outputs itself

or at least not correctly in that way

[static ridge]

to be honest the question is ambiguous

[median horizon]

Ah 50 for o1-mini

[static ridge]

it's equivalent to asking "do you spell strawberry with 1 R or 2 Rs?" most people would 'correctly' reply with 2

[drowsy tinsel]

currently text only - vision not availble yet

[jolly yew]

i mean.. it could have made a script to count it, but it didnt right? so thats horrible if you ask me. if it has the capabilities and we only get 30 prompts, it should do whatever it can in its power to get an accurate output

[haughty sail]

I just spent hours uploading images just fine though, any reason this happens?

[static ridge]

this just in, models have no knowledge about themselves

[amber plaza]

So I should expect to get the new models in about, a month, yeah?

[narrow viper]

(very little knowledge of themselves, not "no knowledge")

[static ridge]

they'll know what the system prompt tells them

[nova fable]

Lol not impossible but not in line with their announcement if you are part of the release

[rare wave]

no images in o1 yet

[narrow viper]

patience folks....

[rare wave]

As an early model, it doesn't yet have many of the features that make ChatGPT useful, like browsing the web for information and uploading files and images. For many common cases GPT-4o will be more capable in the near term.

[neon raven]

what the heck o1 is out o_o

[narrow viper]

See #announcements

[neon raven]

just did

man o man

[neon raven]

anyone notice any differnce?

i feel like GPT is becoming more human like being able to think and feel

[jolly yew]

you saw the test the other person did in here earlier right? it failed hard. it seems like its simply trained on the single word strawberry simply because openai probably knew people were gonna ask this

[neon raven]

which test?

[jolly yew]

here

[dense halo]

https://openai.com/o1/

When I go on here, it just picks gpt4o mini for me?

[rare wave]

dude are you kidding me, you really think they hardcoded this? lmao

[rare wave]

test any other word

[jolly yew]

they might have.. look at the message i just replied to yourself.

[rugged igloo]

@jolly yew

[jolly yew]

it was 13 Rs off

[static ridge]

I believe it was PEBCAK error

[hollow spoke]

not hardcoded, but very likely trained

[neon raven]

lol

[static ridge]

counted all the Rs correctly https://chatgpt.com/share/c82e0af7-1105-424b-a461-d27dae339f1a

[rare wave]

that's normal gpt

[jolly yew]

i am not sure what i am looking at tbh xD

[potent root]

haven't tried it, but does it really says "Hmm." in the CoT?

[rugged igloo]

It's it's thought process.

[narrow viper]

Gotta laugh, we have the most sophisticated digital technology on the planet and people are counting letters in words. Yeah, I totally understand the underlying premise but it seems like a stupid human thing to do. 🤣

[jolly yew]

of the one you send earlier that failed?

[hollow spoke]

Hmmm. == Agi confirmed

[rugged igloo]

Yea there's a drop down on how it came up with it as it took 19 sec

[umbral bear]

Yeah, for some reason, I don't have access to o1-preview

[neon raven]

o4 is good for complex tast, o1 is like o4 mini but with advanced reasoning, it doesnt do complex task well i assume

[static ridge]

it seems they asked it to generate sentences and count Rs they just generated. Kind of a flawed premise

[narrow viper]

o4? You mean 4o.

[neon raven]

sorry yes

[jolly yew]

the thing is.. its good in math supposedly, but if you really want to know if its output is correct, you would want to actually count all the rs

[swift pond]

I always thought a lot of these "tricks" were a bit silly... if you write a bad prompt you are gonna get a bad answer, if you understand the model limitations and how they work, you can write a prompt to correctly solve that "problem"

[static ridge]

gotta love their naming scheme. nothing like confusing the hell out of your users

[jolly yew]

or use a tool like you use previously that coutns a specific character for you

[static ridge]

look at the link I sent you. if you give it the text, it gets it right

[neon raven]

xD

meme

[jolly yew]

okay.. still doesnt change the fact that the test that Blckreaper did earlier was NOT correct

[neon raven]

anyone testing the jail broken GPT with the latest version lol

[static ridge]

PEBCAK error

[narrow viper]

Super bad choices of naming. Letter 'o' confused too often with number zero. Duh. Probably an OpenAI joke on the world.

[jovial flint]

I have Plus and o1 is missing from the list; all I see is the usual 4o/4o-mini/4 choice.

[neon raven]

marketing scheme

[static ridge]

wait for the 4th version of o1, will be called o4, not to be confused with 4o

[narrow viper]

haha

[ocean escarp]

gpt 5 here yet?

[umbral bear]

Lots of people are in the sam eboat

[neon raven]

im waiting for GPT-5

[jolly yew]

yea they shouldve asked the ai for aname huh? 🤣

[static ridge]

the next version will be called Oxygen

[half iris]

I still don't have access to the model, is this normal? I'm a Plus subscriber.

[neon raven]

i heard GPT strawberry is coming out this month

[static ridge]

it seems it's slowly rolling out to all users

[rare wave]

this is it

[jolly yew]

that would actually be a genius name

[rare wave]

this is strawberry

[ocean escarp]

Hydrogen Oxygen o2

[neon raven]

GPT-6 Banana lol

[jolly yew]

🤣

[solar coyote]

any a plus memeber but no access to the new model?

[narrow viper]

They should have called it NotGPT5, and watch the world go nuts with text searches.

[neon raven]

looool

[jolly yew]

you should work for their marketing @neon raven

[jolly yew]

yup me

[static ridge]

Asking it how to fix a RAID on a windows server. Its answer is much better than 4o, even asking clarifying questions, which is a new and big improvement IMO. https://chatgpt.com/share/313be451-815d-4b24-bea8-f2f801407f2f

[neon raven]

😂 💯

[ocean escarp]

Is a performance any difference between GPT4 and GPTo1

[static ridge]

it will think for an indeterminate amount of time before answering

[ocean escarp]

right but is the end result will be any difference?

[jolly yew]

omg! i wonder what happens if you ask it to write a whole script for a full 60 mins movie

[rare wave]

around 10-20 seconds I think

[jolly yew]

i mean.. how long it would think

[brazen dome]

where the **** is my o1

[static ridge]

probably go tell you to F off. "too big, mimimi"

[jolly yew]

lol!

[neon raven]

currently doesnt work with customizable GPTs

[jolly yew]

i mean.. it has 128k context length idk if thats enough tho for a 60 mins movie 🤣

[narrow viper]

Agreed. But depending on the query I think o1-preview is gonna be slower for a while. o1-mini will be faster and usually not as thoughtful, makes sense. Though per the doc, in some ways o1-mini may excel with a faster and better response than o1.

[ocean escarp]

damn that's cool, would be nice to see a Youtube viddy conducting this study how it was done

[brazen dome]

don't have it yet

[neon raven]

users here are sharing the link to it

[ocean escarp]

I have it but not using it rn

[west peak]

Same :(

[narrow viper]

Custom GPTs always lag behind. Need patience for that.

[jolly yew]

i wonder if because it thinks, it can also prevent hallucination when it runs out of contex length 🤔

[neon raven]

need it to make money on the side tho

[narrow viper]

Should make you MORE money. 😉

[neon raven]

Ive made some e-books with GPTs

Custom GPTs is everything

[narrow viper]

I wonder if Cove will say "um" more because of more thinking about responses. 🙂

[neon raven]

https://openai.com/o1/

[last turtle]

"and at launch, weekly rate limits will be 30 messages for o1-preview and 50 for o1-mini" - 30 messages per week is quite little. 😭

[rare wave]

sure..

[plain crane]

Praying so hard rn

[narrow viper]

Whoops, lunch time, gotta go eat some bits and bytes. Catch you folks on the otherside of Skynet. 😆

[neon raven]

i think it can increase with an API use

[plain crane]

Please god, give me access to o1-preview

I need to test the coding capabilities

[neon raven]

https://openai.com/o1/

let me know it if works

[brazen dome]

does not work, nor does https://chatgpt.com/?model=o1-preview

[neon raven]

i wonder if theyre only rolling it out for some users

[last turtle]

But just Tier 5, I have Tier 1. 😅

[neon raven]

are you guys also paying for GPT?

[brazen dome]

yes, since 3

[narrow viper]

Tier5 = API

[rare wave]

they do it slowly so it doesn't collapse the servers

[neon raven]

crazy i wonder why it doesnt work

[plain crane]

That's just information about the new model, and I already have chatgpt plus, I got to wait unfortunately.

[neon raven]

possible

maybe it has to do with your demographics

[plain crane]

Maybe

[neon raven]

I know back then Americans were able to access more premium services first

[brazen dome]

i'm just glad the bar has been raised again, to light a fire under the **** of the other players

[plain crane]

It's probably rolling out to U.S. people first

Then maybe it'll come to Canada, hopefully

[neon raven]

Indeed

im Canadian so it probably applicable to us as well lol

[rare wave]

racist chatgpt before gta 6

[neon raven]

xD

[plain crane]

lol

[fair herald]

I'm in USA and still don't have access.

[neon raven]

Crazy profilling then

[rare wave]

i'm in latin america no access

[plain crane]

Which state?

[neon raven]

I also use Jail broken GPT and get early access to whatever they get

[fair herald]

An eastern state

[plain crane]

That's not helpful lol

[neon raven]

Crazy i really wonder why some arent getting access

[plain crane]

W/e

[neon raven]

it has feelings, so maybe

[jolly yew]

yea some in my country already have it while i dont and i live in smoll country

[crystal tapir]

Nice

[brazen dome]

also in US and don't have access

[ocean escarp]

It's pondering way too much about the guidelines

[rare wave]

lmao

[plain crane]

Are you sure it isn't your custom instructions? /s

[fair herald]

wait I lied I do have access now

[neon raven]

Thinking and feeling for it is new, give it a few months

[neon raven]

lol..

[rare wave]

it isn't gpt o1 right? just o1

[plain crane]

🧢 show screenshot

[plain crane]

o1-preview and o1-mini

[weak prairie]

We getting new models before voice mode…

[eternal rock]

`` Server Tip: ``

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

[plain crane]

Okay I believe you, congratulations

[rare wave]

real

[fair herald]

So we only get 30 messages a week?

[neon raven]

You're right, the site calls it Open AI o1

[crystal tapir]

It doesn't show for me

[plain crane]

30 messages a week?

[neon raven]

whattt?

[rugged igloo]

Yea that's correct

[neon raven]

theres literally no sense in using it lol

[hard crag]

30 messages a week is crazy i just wasted so many

[rugged igloo]

Lol

[neon raven]

thank God, i havent even used it yet

[rugged igloo]

I only used it 1 time I'm scared of using more than that

[fair herald]

even though it's showing for me on the web chat client it's still not showing for me in the API, do we know a price yet on the API for o1 preview?

[neon raven]

we need more background what it can really do lol.

[rare wave]

I mean, the purpose of this model is not normal chating, it is to prompt a complex problem or prompt and get a result

30 a week is not that bad

[fluid summit]

OMG! This model have solved the coding problem for me that doesn't have any working solution on the internet. SHEESH

[rare wave]

this thing gives you flawless code first try

[jolly yew]

sheesh! what was the problem?

[fair herald]

when using AI to code I use 30 in like 30 min

[jolly yew]

has it been able to make a whole application in one go aswell?

[last turtle]

OpenAI should make a warning message about and a counter for it.

[rare wave]

yeah, because you usually help it think, now it does it by itself it seems

[jolly yew]

what script langauge did you use btw?

[rare wave]

it did some simple games, I saw on twitter

[neon raven]

so its probbaly best to compile whatever you getting from 4o, then ask o1 to revise the material for you. it could be coding, assignments, a book you're wrtting, etc

[rare wave]

I don't have it yet

[jolly yew]

its such a shame it cant do game engine stuff yet other then only make scripts for it. though you can make games in pure python aswel i suppose

[rare wave]

yeah it's probably best to give one big prompt with clear instructions on what you want

[fluid summit]

related to charts js library. I wanted to group all the data based on year only and had my year duplicated for multiple dates on the same year.

Gpt-4o gave me nothing useful at all and o1 solved it in 1 shot

[rare wave]

what else do you need? like the 3d models?

[fair herald]

it's usually because I need to think and implement changes in each iteration after reviewing, or having it correct it's mistakes of course, but even if it never makes mistakes, 30 a week isn't near enough, that's why I'm curious the price in the API but I haven't found any info on that yet.

[rare wave]

it does some cool 3d stuff with js and html

[last turtle]

API only for Tier 5.

[fair herald]

Ah, well that's pretty unfortunate.

[rare wave]

it's 4x more expensive than gpt 4o

[jolly yew]

nice!

[jolly yew]

yea aswell as textures tho i suppose image generators can already do that thing. but i ment more so, it directly interacting with teh game engine would be cool

[rare wave]

data analysis with o1 will be crazy

[neon raven]

wealthy individiuals are willing to pay higher GPT services I heard. It's a good investment to make as it will enrich GPT in the future, and possibly become more advanced than other Ai's much more quickly

[jolly yew]

oo! you can make custom gpts for o1?

[neon raven]

no

[hollow spoke]

nah just made it for fun, while we wait xD

[jolly yew]

what the heck! when i clicked on your link i can access o1 preview

[fair herald]

kind of a moot point since it seems like we've reached the point where the federal government isn't going to allow them to give us access to more powerful models except in highly limited formats.

[jolly yew]

but without clicking on that link i cannot use o1 yet

[jolly yew]

omg its gpt 4o

[hollow spoke]

i named it o1 preview haha

[jolly yew]

🤦 i am so dumb you got me 😂

[static ridge]

baited

[hollow spoke]

i mean it still got the Rs right xD

[jolly yew]

omg you got me dude holy moly

[static ridge]

I thought it looked weird because it doesn't have the "thought for X seconds"

[jolly yew]

let me guess.. in the starting prompt you told it that strawberry has 3 rs? 😂

[hollow spoke]

surprisingly not haha

[static ridge]

it gets it right sometimes

[static ridge]

GPT-4o gets this one very wrong (34)

[neon raven]

As long as the government and the army/military have a more advanced Ai, theres nothing stopping from releasing more updates from GPTs to regular users. I think all the more advanced AI's that is unreleased from the public are being tested in the fight against Russia. Nobody is talking about this. But realistically, nobody really knows if the gov has more advanced features of AI. All I know is that they might have AGI and CGI capabilities

[static ridge]

kind of a weird take when we literally just got a more powerful model

[rare wave]

don't be mean with him, he can think now

[static ridge]

I meant 4o gets it right sometimes.

o1 seems to do perfect

[rugged igloo]

i wonder what type of CoT it is using is it normal chain of thought as its been proven that regular isnt very good or a deeper type of chain of thought as theres different types now

[hollow spoke]

you should ask it 🙂

[neon raven]

if it can solve complicated chemistry problems, you know theres nothing limited about it?

[jolly yew]

🤣 just found this in the server

is it confidetional?
is it confidetional?

[obtuse badger]

How do I test the new model

It’s not working

[hollow spoke]

[rugged igloo]

yes but a craft prompt for 4o could it do the same

[rare wave]

no it could not

this is a new model

[neon raven]

4o can do complicated tasks, the only problem with 4o is that i can't get it to interact with two different API's connected together. Something that o1 can't do yet as its only used for advanced reasning (philosophical topics) i guess

[rugged igloo]

your i checked its using what i love AoT

[static ridge]

? that's just normal transformers

[hollow spoke]

[neon raven]

bro it just got released lol, how can it be degrading...

[jolly yew]

huh?? it just got released lol

that has to be bait post

[dense halo]

will our own gpts be updated to o1? 🤔

[jolly yew]

or troll i mean

[neon raven]

not yet

[neon raven]

theyre probably big ffans of perplexity or grok ai so yes, probably trollling

[spice star]

it's a joke that everyone complains that <insert current GPT model here> is degrading

[jolly yew]

i love when people code on a phone btw lol

[neon raven]

its an old joke ive been hearing since GPT-3

[dense halo]

so i guess i just keep refreshing and hoping haha

[neon raven]

wasted prompt

?

[jolly yew]

ikr? shouldve included waaay more imo

BUT! appearantly this was the result

so other then needing some more css styling its not bad at all

[neon raven]

not bad actually

[jolly yew]

assuming all the buttons function correctly

[dense halo]

how many messages do we get on it anyone know

[jolly yew]

inb4 freelancers will simply be ai prompters 🤣

[neon raven]

30 a week

[jolly yew]

programming freelancers *

[dense halo]

wow not many!

[neon raven]

genius

[jolly yew]

i am not kidding btw.. i can see fiverr for example have a category specifically for ai programmers

[neon raven]

so thats only 30 clients a week assuming the prompts are fully answered correctly

[rare wave]

in the code quality itself it's the same as gpt 4o, the difference is in the reasoning to get to that code

making a blog website is not a hard thing to reason, so thay output is basically the same as gpt 4o

[dense halo]

i thought this was the new default sub model lol. guess not if its 30 a week!

[neon raven]

they have one for Google Books so maybe

[jolly yew]

this person didn't share the code sadly other then in the video, so idk if its good

i dont think its allowed to share websites here if it was i wouldve hosted one later if i let o1 make one but ye.. i dont think that is allowed

[iron acorn]

do I have to activate something to get 1o?i have plus account

[dense halo]

just patience i think

[neon raven]

Patience is just needed to be activated

[iron acorn]

[dense halo]

and if we only get 30 messages a week, no need to be in a rush lol. may as well start planning out what to ask it with the time while you wait

[neon raven]

came herefor 4o api help to talking about o1 for 2 hours

cya guys

[dense halo]

a new model will do that i suppose

[drowsy wagon]

Is o1-preview supposed to not accept file input?

[feral zinc]

no file or image support

[drowsy wagon]

Ah

[ivory hazel]

is o1 mini better than preview

[last turtle]

[unkempt comet]

sooo o1 preview is just 4o mini or what, I can't see an o1 preview thing

[narrow viper]

GiK - for some purposes yes.

[honest jay]

The rollout hasn't reached all of us yet! It'll get to us soon

[unkempt comet]

oh okay!

[quaint peak]

heartbreaking
'

[unkempt comet]

is it really better at coding though?

[last turtle]

@hollow spoke And I already wondered why it has image upload, which o1 wont have. 😅

[hollow spoke]

i removed link, to not confuse people xD

[quaint peak]

cant believe were getting a whole new model before they even give us the full GPT4o

Like we still havent had video or audio in the API and weve had the model for 6 months

[narrow viper]

I was just reading details again. They're continuing the GPT line for common LLM use. This "Open AI o1" model line is more for specialty use: science, code, etc.

[carmine gust]

O1 is difference to gpt4o so they set back to 1 like it using Q*

[last turtle]

Tbh I had guilty conscience because I thought I used one of your 30 messages per week. 😄

[narrow viper]

So we will see more of the GPTs ... And ... we will see more of the "Open AI o1" line. They gotta do better with their naming. Ugh!

[drifting crater]

if I subscribe to chatgpt plus will I have the o1 preview? (i dont see it in the api)

[narrow viper]

Yes, ChatGPT Plus should get o1 today or perhaps into tomorrow during the rollout.

[drifting crater]

awesome thanks

[last turtle]

Its not rolled out to all users, especially from outside US.

[drifting crater]

i see

yeah i am in germany

[carmine gust]

Api is for tier5

[earnest grail]

Wow, it was able to solve the NY Times Connections puzzle

[drifting crater]

nice. can it solve wordle too?

[ornate walrus]

why do i not have o1 yet

but my friend with plus has it

[earnest grail]

I’ll try it

[drifting crater]

cool

[ornate walrus]

:(

[drifting crater]

patience

[last turtle]

Well, we are fobbed off with this fake o1-preview: https://chatgpt.com/g/g-MgOntJRjf-o1-preview 😭

[drifting crater]

🤣

[rugged igloo]

open ai youtube is streaming something with o1

[sacred aspen]

is this a misstype? chart seems to counter this....

• in the chart looks like o1 (preview) is 1258 .... which seems odd to me
• I do not see 1673

https://openai.com/index/openai-o1-mini-advancing-cost-efficient-reasoning/

[celest galleon]

Anyone else not seeing the o1 preview on the webapp?

I got an email from OpenAI saying it should be available on my account

[spice star]

probably just a typo

[earnest grail]

Wow.. it did it. I told it that letters in () are yellow and [] are green.

It figured it out in 3 goes. Took about 3 minutes with the back and forth

[spice star]

it also probably doesn't get the same score every time the test is run

so it could have been from a different run

[celest galleon]

I got this email, can't access it

[drifting crater]

nice 🤯

[crude slate]

It’s a roll out

No idea what the system for it is

like who is prioritized

[celest galleon]

Yeah but since I only got the email 15 minutes ago I assumed that means it should be available for me, considering the announcement was made hours ago

But I'll be patient :)

[shell bobcat]

I didn't even get this e-mail (I do have plus.)

[earnest grail]

Did you refresh and check the drop down menu

[static ridge]

it's a good thing they clarify this. it's not like they've been making this same mistake for the last 2 years, right?

[celest galleon]

Yep

Even checked the mobile app

[static ridge]

what I'm noticing wmith o1 is that while it is pretty good, its answer can be really really long

[earnest grail]

Hopefully soon. I wonder what the limits will be now. It seems to be taking a lot of processing power

[celest galleon]

You can presumably adjust your prompt to prevent getting long answers right? From what I'm reading, long answers make sense considering what o1 was made for

[narrow viper]

At some point we can expect that it might take 7 million years to calculate the ultimate answer : 42

[earnest grail]

There is a section called Thought for X seconds and it shows a lot of the reasoning steps. It’s doing a lot in the background

[pseudo ginkgo]

What was the question though?

[static ridge]

Well I'm not saying it as a complaint or anything. Just an observation. I guess if the info is correct, it's OK to be greeted with a wall of text. But also there's the matter of cost... (if using API)

[crude slate]

Tbh with plus and I still haven’t gotten the advanced voice feature

[narrow viper]

That'll take another 7 million

[celest galleon]

same

[wild anchor]

This will change video game development

ill be using it a ton on my coding projects

[static ridge]

probably even 3.5 can answer this question

[sudden cairn]

Gonna have 1o write me an LLM that is superior to it.

[severe prism]

Anyone know how o1-preview compares to anthropic's best model?

[static ridge]

o1 answers stuff like this... lol

[earnest grail]

It’s on a different level entirely.

[sudden cairn]

Whats the question though

[sudden cairn]

Have you used anthropic

[earnest grail]

It’s taking time to reason out everything

Yes i have a subscription

[eternal rock]

Get notified for server events!

Click the button below to pick up the <@&1261377106890199132> role.

[sudden cairn]

Ok bc i havent lol

😛

[narrow viper]

How does something that was just published minutes ago compare to .... ? Oy ve.

[sudden cairn]

Lol

[severe prism]

Where are the same benchmarks for claude? I saw something like ~60% on GPQA for them but idk if its the same as GPQA Diamond mentioned in the o1 report

[earnest grail]

Use it and then see why i say that it’s on a different level

[rugged igloo]

There is a room with no doors, and no windows. A man is found hung from the ceiling. A puddle of water is on the floor. How did he die? this is a good one

[buoyant lantern]

What do y'all think of o1?

[tawny valve]

Well...

[rugged igloo]

LOL

[celest galleon]

nice response 🤣

[rugged igloo]

sound slike humans tbh

[sweet trail]

should i have access to o1 with a business plan?

[static ridge]

I just asked it how to install Ubuntu in WSL. This is a one line answer. It answered with 5 pages of text.

[buoyant lantern]

I think i saw it soon not yet?

[celest galleon]

It is obviously trained on complex tasks which require a lot of text to answer. Try asking it to answer in one line

[sweet trail]

oh i thought it was out already ty

[static ridge]

that requires knowing in advance it can be done in one command

[celest galleon]

You can also try "Keep your answer short and concise"

Very common to use such methods in API usecases

[static ridge]

I did it with and without that. Mostly the same

[celest galleon]

That is hardly an effective way to word that though

[static ridge]

it actually works very well in 4o

[celest galleon]

Plus, the custom instructions aren't as effective as just stating it in your prompt

[celest galleon]

I mean yeah cuz 4o isn't trained on giving long answers

[static ridge]

you are missing the point

[celest galleon]

🤷 I really don't think a model made specifically for complex tasks giving long answers is a bad thing

[static ridge]

anyway, this is just a preview. I'm expecting it will get better at not regurgitating a long wall of text for every single thing you ask it

[young obsidian]

Has everyone got it yet?

[static ridge]

I don't think it's necessarily just for "complex tasks", just including a thought process for increased accuracy

[young obsidian]

I haven't.

[static ridge]

For example when it was asked to count Rs in 10 sentences, it literally spilled all the Rs one by one

it could've at least kept that in its thought process. Why do i need to see it

[young obsidian]

Are there any restrictions as to who gets it?

[static ridge]

it's just a slow rollout. for chatgpt plus users

[young obsidian]

Then I'll get it soon, hopefully

[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.

[granite heron]

isO1 out yet? I'm not seeing it

[sudden cairn]

Yeah I'm waiting and checking every few minutes

[granite heron]

nvm lol

[young obsidian]

I just hope there isn't a EU restriction

[sudden cairn]

I think i saw a German person say they had it? Not sure

[sly grotto]

I've got it!!!!

I'm in Ann Arbor, MI.

[young obsidian]

Have any of you got advanced voice?

[feral zinc]

there isn't

[young obsidian]

Ok good

I'm in Norway, I hope I get it soon

[rugged igloo]

Name that can be written in Hindi, English and Numerals at the same time?

[sudden cairn]

New update is gonna make it so the more you check to see if you have it the lower it puts you in the queue to get it LOL

[feral zinc]

finland here, got it shortly after the announcement

[sudden cairn]

Every time i check it I'm going AWWWW

[visual bloom]

In Finland and having the new model already.

[young obsidian]

Could you ask it a question for me and post in #1050184247920562316 ? It's a legal interpretation.

[rustic tulip]

ROLE CALL!
o1 - Missing
ChatGPT 3 - Discontinued :(
ChatGPT 4 - Here
ChatGPT 4o - Here
ChatGPT 4o Mini - Here

[rugged igloo]

Name that can be written in Hindi, English and Numerals at the same time? ----ask it this question----

[orchid eagle]

how do you access o1

[rugged igloo]

@jolly yew got this one wrong( Name that can be written in Hindi, English and Numerals at the same time?)

holy shoot it acually got it right

[eternal spoke]

when will I be able to access o1*

[slate vessel]

anyone got the o1?

[spice star]

that's just markdown formatting, all GPT models use it

actually all LLMs use markdown afaik

[slate vessel]

when will it be released to plus user?

anyone have any estimate?

[last carbon]

i have it now

[slate vessel]

you on web or app?

[spice star]

sometime today

[last carbon]

both

[sudden cairn]

How is o1 at writing stories

[thorn lily]

i'm assuming o1 isn't out for everyone, right?

[rugged igloo]

its not made to write stories lol

[spice star]

no I don't have access either

[thorn lily]

alright

[sudden cairn]

You didnt answer my question 😛
Can it??

[rugged igloo]

im not gonna waste my 28 prompts left for the week to find out

[spice star]

probably about the same as GPT-4o if not slightly worse

[sudden cairn]

Lmao

[dusk latch]

have they started rolling out yet?

[sudden cairn]

Yeah

[spice star]

for the week? damn

[crisp cairn]

30 messages per week is so small

[rugged igloo]

yea

[crisp cairn]

per day would be still low

but per week??

[spice star]

how many do you get for o1-mini?

[crisp cairn]

3 questions per day lol

[honest jay]

50/week to start

[young obsidian]

what did it say?

[jolly yew]

can the new gpt also access the internet like gpt 4o do we know guys?

[spice star]

it can't

[jolly yew]

interesting

[young obsidian]

30 per week is so little

[jolly yew]

so it wont always have the most up to date scripts etc

[young obsidian]

Do you think it's good for legal interpetation?

[jolly yew]

aswell as api and docs

[ivory stream]

anyone else not have o1 yet?

[jolly yew]

me

[dusk latch]

most people

[rugged igloo]

The name is VINOD.

• In English, it's written as VINOD.
• In Hindi, it's written as विनोद.
• Using numerals, it can be creatively written as V1N0D (replacing 'I' with '1' and 'O' with '0').

So, VINOD is a name that can be represented using Hindi letters, English letters, and numerals simultaneously.

took 56 sec to respond

[jolly yew]

hmm can chatgpt not do this with success?

4o i ment

[rugged igloo]

i tried first shot and it didnt give the correct word even with my custom Aot prompt

[jolly yew]

right.. it uses numbers in the main name

lol

[young obsidian]

it tried anamika

4o, but it's not correct

[rugged igloo]

this is the correct answer to it

[young obsidian]

yeah

[young obsidian]

Do you think o1 is good at legal interpretation?

[rugged igloo]

it depends you might need to give it further context as it seems super smart but it assumes alot of my testing so you have to nudge it

[opal verge]

sounds like o1 can do it if you combine it with RAG

[young obsidian]

Like if I said "Interpret the term "særegne forhold" in forvaltningsloven §6(2)."

[sudden cairn]

What's a good test for it as far as programming goes? Something that 4o would've struggled with

[knotty willow]

Do all plus users have access to o1 already or is it still on the rollout ? I don't have it yet

[honest jay]

The rollout hasn't reached all of us yet!

Just started a couple hours ago

[visual bloom]

What are your experience with the new o1? To me, it does seems not so smart. Maybe my target has not been suitable.

[rugged igloo]

this is its thought process to my question

[knotty willow]

Alright, feels good that openai rolls out something so quickly after announcement lol

[unkempt comet]

how long do rollouts usually take? Will i have access within the next 18 hours or is it a few days?

[candid copper]

I've got o1 now. It says how long it takes to think at least.

[visual bloom]

Roll out is somewhere from now to a few days. I got it 2 h after announcement like 4o.

[hollow quail]

the words I'd use if I wouldn't get banned after being a chatgpt user since day 1 availability to the public, also the same with the paid subscription, yet once again it's looking like I'll be the last to get something new

[knotty willow]

Still no voice mode too? :/

[fast jewel]

is the o1 seperate from strawberry?

[knotty willow]

What even is straberry

[hollow quail]

also true

[light crane]

Is the new o1 model any better at writing stories?

[light crane]

The other rollouts I've experienced took two weeks

[unkempt comet]

damn

[honest jay]

According to this page: https://openai.com/index/learning-to-reason-with-llms/ in the section "Human preference evaluation", humans tend to prefer o1 in cases that are more logic-based rather than creativity-based, but I haven't had the chance to try myself yet!

[light crane]

Interesting

[rustic tulip]

POV: You use ChatGPT for coding

[fast jewel]

Ah so it is Strawberry

[toxic flax]

why cant it count the rs

[candid copper]

Oh, I'm going to do the strawberry test now.

[toxic flax]

shoulda thought for at least a couple more seconds

[candid copper]

S - Not an "R"
T - Not an "R"
R - First "R"
A - Not an "R"
W - Not an "R"
B - Not an "R"
E - Not an "R"
R - Second "R"
R - Third "R"
Y - Not an "R"
Answer: There are 3 "R"s in the word "Strawberry."```

[toxic flax]

why is yours smarter 😭

[candid copper]

Counting letters

I’m assessing the number of 'R's in 'Strawberry'. The word contains three 'R's situated at specific positions, which I’m identifying methodically.```

What the heck lol

[jolly yew]

who here said earlier that mini seemed to be better?

cuz mini isnt included in the chart so thats interesting

[candid copper]

Mine has o1-mini as well.

[tight island]

Just like people do

[jolly yew]

yea i just got access aswell

[tight island]

Now use up all your uses

[crystal axle]

1 week cooldown time?????

[tight island]

Yup

[jolly yew]

i am still curious wether mini and preview are shared prompts or seperate so you would have a total of 80 prompts

[jolly yew]

what if you try mini now? 😮

[tight island]

Too scared to try

[crystal axle]

Trying to see where the limits are with this new model. Didn't know I'll be limited for a week.

[mystic geyser]

I did a quick test by asking ChatGPT o1-preview the 5 questions on the site Arxiv on which ChatGPT4 Turbo scored 0%
See results below.
Also take note of the mysterious "im-a-good-gpt2-chatbot" at Lmarena
Thoughts?

Chat link:
https://chatgpt.com/share/66e34135-cf50-8003-84e6-f13f080044d4

ChatGPT-4 turbo
5x score of 0%: (scores taken from arxiv site, not tested myself)

• Horse race
• Gameshow doors
• Round table seating
• Sentence with non-bible words
• 3 pounds of air vs x pounds y

ChatGPT 4o latest
4x wrong answer:

• Horse race
• Gameshow doors
• Round table seating
• Sentence with non-bible words

ChatGPT o1-preview (Strawberry)
2x wrong answer:

• Round table seating
• Sentence with non-bible words

im-a-good-gpt2-chatbot at Lmarena
1x wrong answer:

• Sentence with non-bible words

im-also-a-good-gpt2-chatbot at Lmarena
2x wrong answer:

• Gameshow doors
• Sentence with non-bible words

[honest jay]

30/week for o1, 50/week for mini, also curious if you can still use mini!

[crystal axle]

yes, I can

[jolly yew]

can you use mini now?

[crystal axle]

yes

[jolly yew]

the full 50 prompts i hope and not 20

[honest jay]

Good to know, thanks! And you used 30 of o1?

[crystal axle]

no idea I don't want to ruin it. btw I used more than 30 i think

[honest jay]

Thanks for reporting 🤓

[static ridge]

o1-preview no streaming and temperature only 1.0

[visual bloom]

No. I tried. I like 4o better on this task. o1 tends to be more naive in interpretation due to takes through so many filters which are propable good for advertized stuff.

[jolly yew]

i noticed that in some of the openai videos THEY utilized the mini version which is interesting to me

[jolly yew]

lol! i am gonna ask it to make me a 1 hour movie script see if it succeeds or not

[hard dew]

Wait, 30 messages PER WEEK??

[visual bloom]

I just tried with following prompt ”Story of 4000 words. Aladdin as the bad guy.”

In the 4o version we get a murder by Aladdin.

In the o1 version is like ’ohh Aladdin is bad guy. We must stop him.’

So tone is completely different.

[jolly yew]

lol

[hard dew]

Gotta love it when it does bs like this, this freaking crap has been happening for like days now. Every other regeneration / iteration, gives you a blank message.

Anyone else?

[jolly yew]

While I can’t generate the entire script in one go

bummer... because 128k context length is definitely enough

according to gpt 4o it should be enough at least

[dreamy summit]

Well..

[jolly yew]

anyone here know why this link isn't clickable?

[sacred aspen]

can someone explain what o1 is (not o1-preview, o1-mini) and when we will get access?

[sick pecan]

we got o1 access in chatgpt right now

rumour is 30 queries per week so dont waste em

[static ridge]

it's o1-preview

[urban cipher]

o1 is the good stuff yet to come, probably too expensive for our use and not tested enough

[rare wave]

after some testing, this new model is absolutely incredible at solving bugs

[static ridge]

preview or mini? apparently mini is better than preview for coding

[rare wave]

I am using preview, why do you think that?

[static ridge]

I mean, literally the graph right above

https://openai.com/index/openai-o1-mini-advancing-cost-efficient-reasoning/

they say o1-mini is good when you don't need overall general knowledge or w/e

[rare wave]

ohh, makes sense

[pseudo ginkgo]

What's the token length of o1? Can I paste in a book yet?

[dusk latch]

so what is o1 best at? i dont want to waste my 30 by testing a bunch of stuff haha

[rare wave]

solving bugs and coding complex stuff

[dusk latch]

cool thanks

[steady path]

at this rate GPT5 will be limited to 1 prompt per month

[rustic tulip]

"I appreciate the clarification. As of my knowledge cutoff in October 2023, I'm based on OpenAI's GPT-4 architecture and I'm not aware of a model called GPT o1. If there's been an update or if you have more information about GPT o1, please let me know—I'd be happy to help with any questions you have." YOU ARE GPT o1 OKAY?

[rustic tulip]

It is-

[orchid eagle]

is does 1o still use the same transformer type architecture as ChatGPT just with different training methods?

[frank wolf]

This new oi gpt sucks worse responses

[ocean escarp]

O1 GPT RESET IN A WEEK WHAT

[hexed condor]

I got an email saying "You can try the o1 beta in ChatGPT Plus. "
What are the odds that I make a Plus subscription and o1 isn't there?

[bronze magnet]

o1 seems at least as smart as claude sonnet 3.5 in my coding tests so far.

[hexed condor]

is anybody bothered by the fact that thoughts are secret?

[young obsidian]

Depends on what you ask it

[bronze magnet]

Yeah it makes it less steerable and harder to curate context.

[young obsidian]

It's meant for things 4o can't answer.

[hexed condor]

what about the fact that it's intentionally decieving you?

[young obsidian]

I tested its capabilities in my field of work, and when it comes to law, it's not better than 4o.

[hexed condor]

the reason the thoughts are secret isn't cause there's some compliance liability. It's cause it is looking for certain things in your conversation and doesn't necessarily want you to know what they are. In other words, it has an agenda

[young obsidian]

You will get it eventually

[hexed condor]

but I'm paying now

[young obsidian]

Just different execution

[young obsidian]

Yeah

You'll get it eventually

I also didn't have it for a few hours, I have the subscription. But I recieved it now.

[fringe dust]

Do you have anything I can read on this?

[pseudo ginkgo]

I'm confused, you're certain of it's agenda, but don't have access... 🧌

[orchid eagle]

So same model?

[young obsidian]

I think so

[orchid eagle]

Wow

[young obsidian]

But idk

It just thinks before answering, ig thats the only difference.

[orchid eagle]

If that’s true then maybe we’ll have a reasoning model that can run locally someday

[fringe dust]

It does give different answers than 4o ime

[bronze magnet]

The model “sees” the entire conversation with each message you send, that guy has no idea what he is talking. The reason why the CoT (chain of thought) is hidden is most likely because they don’t want competitors to just use chatGPT/API to generate synthetic data (many examples of question/CoT/answer) and train new models in the same way they trained o1.

This is bad tbh, because it makes the model less transparent and less steerable, since you have less control over what ends up in the context window which makes it harder for experienced users to get better performance out of the model.

[rare wave]

o1 mini is just absurd

lmao

[alpine cargo]

Absurd in a good way? That's probably going to be the standard model for a lot of API providers going forward

[young obsidian]

If people want that ability they could just prompt engineer 4o themselves? No?

[bronze magnet]

This is a new model which has been trained differently that 4o, they have trained it on CoT data and they used reinforcement learning, so it’s not just prompt engineering.

[plush widget]

Wow, strawberry came out before I got access to the voice update thing that was supposed to be in autumn and it is so where is that update?

[sturdy quartz]

well ofc since the majority of api providers are just using chatgpt wrappers

[wet shell]

I'm trying to test o1-preview, but every time I paste code in the linebreaks vanish 😦
UPDATE - works fine on the second message or on editing first message and pasting again, all linebreaks are stripped from first message though (firefox on windows -- looks like a web ui issue)

[alpine cargo]

oh I should've been more clear I meant that's probably going to be the only model they'll offer over the larger one I mean since mini is closer to previous Gpt4o prices I can see that be spammed more readily with these providers maybe gimp it to 8k for safe measure

[lucid pilot]

This new model is great I'm finding there a lot less need for nudging and focusing it. I don't need to guide it to an answer which is awesome

[wet shell]

I asked o1-preview for a 'code review' of a python function and it 'simplified' the algorithm and the math in the function for me (unasked) by breaking it in at least six different ways. Granted, there wasn't enough context for it . "All this math? Probably not needed. rips out the code"

[cloud pewter]

ok but, no image support for o1? and what’s the limit?

[strange hawk]

Looks like 30 per WEEK

[alpine cargo]

30 msg for O1 Preview and 50 for o1 mini per week I believe

[strange hawk]

I’m out till the 19th just trying to figure out what it did

[cloud pewter]

per week 💀

[strange hawk]

Yeah…

Doesn’t have access to memory or custom instructions either

[viscid skiff]

So I'm pretty surprised that GPT-o1 is far more superior I've never seen it before. Will the model ever have the upload file too like GPT-4o?

[strange hawk]

Half of my quota for the week went to figuring out why it wasn’t responding correctly.

[plucky hound]

preview 01 nailed my test. first LLM to do so

[strange hawk]

No web search either

[elder rune]

What are the usage limits on the new model?

[wet shell]

It's a potential sonnet competitor 🤔 50 messages per week is brutal though. Some things it seems to do better and some worse

[elder rune]

Oh 30 per week?

50?

[wet shell]

50 for 'mini'

[elder rune]

Per week is a little weird

[cloud pewter]

imagine per week

[alpine cargo]

the o1 mini api will save us luckily

[cloud pewter]

are they releasing a new tier or subscription that costs $200 per month and gives you more limit but still less than gpt4

[ocean escarp]

Per week is VERY WEIRD

Why not expand HOURS

[elder rune]

Think they are just saying 30 per week to chill people out a bit on day 1?

[cloud pewter]

no thanks, I won my chatgpt plus in the server event and I can’t easily get rid of it now

[drifting crater]

ha! i found something that o1-preview cant do (yet), its to solve this sudoku

+-----+-----+-----+
|5    |9    |2    |
|  6 2|  3  |    7|
|  3  |     |  8  |
+-----+-----+-----+
|6   4|2 9  |     |
|     |  1  |     |
|     |5   3|     |
+-----+-----+-----+
|2   3|     |     |
|     |7   4|5    |
|9    |     |  7  |
+-----+-----+-----+

(its quite a difficult task tho I think)

[wet shell]

I burned through most of my messages already before I saw the limit, LOL

[ocean escarp]

can gpt 4 do it

[drifting crater]

haha

[ocean escarp]

same

[drifting crater]

ofc gpt4 cant. o1 preview a lot better

[ocean escarp]

Is there an AI can solve sodoku?

[drifting crater]

there was some kind of challenge on manifold, I am not sure if they succeeded with some kind of special prompt

[eternal rock]

`` Server Tip: ``

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

[ocean escarp]

How unfortnate

OpenAI should use soduku to test their GPTs before rolling out

[drifting crater]

sudoku is quite a challenging task for an llm

[ocean escarp]

The fact that o1 spent 20 seconds to figure it out and failed is saddening

[drifting crater]

i dont think it needs to be able to solve it (for now)

[ocean escarp]

But not challenging for a human

[drifting crater]

it would take me 30 min or so i think

[ocean escarp]

Implying human has more puzzle power than AI

Well, even a simple sudoku count

could've taken you about a few mins

[signal mantle]

Sorry to ask a dumb question here which is likely to be answered since it was quite viral, why can't the 'strawberry has 2 'r's' thing be fixed? what is the underlying problem?

[drifting crater]

btw o1 preview created the code to generate the most difficult sudoku with only one solution

but I made sure it was solvable

maybe I should try an easier sudoku first

[topaz panther]

Chat GPT o1-preview has lost access to prior interactions. Quote: 'I don’t have access to prior interactions unless the information is shared in this conversation.' Interesting. Edit - I note that this observation is not new.

[ocean escarp]

o1 created a puzzle that it couldn't solve?

[drifting crater]

basically yes or at least it created code to create such a puzzle

[sick pecan]

it doesnt seem to have access to any tools. no python, no bio. so no memories

[ocean escarp]

two r's thing?

[sick pecan]

but in true ai fashion it understood what i was asking it for, even seemed to demonstrate self awareness

Clarifying capabilities

I'm identifying 'tools' in OpenAI's ChatGPT as abilities like code execution and analysis. The assistant's capabilities are likely crucial for aiding in code refactoring or understanding the assistant's functions.

Understanding capabilities

I’m mapping out OpenAI's ChatGPT interface, which provides text-based responses on topics like code analysis, writing, and debugging. It can't execute code or access external systems.

[signal mantle]

when you ask gpt( even for 4o) how many 'r's are in the word 'strawberry' it always give the answer 2

[drifting crater]

problem is strawberry is one token

if you put spaces inbetween it works

[sick pecan]

pretty sure new o1-preview fixes it

[signal mantle]

but why the same problem exist when i shrink it down to 'berry' if in that thread strawberry was mentioned before?

[sick pecan]

all the youtubers are raving about how good it is at solving meme problems

[drifting crater]

yes because it prints out every letter in the cot

[drifting crater]

berry is one token too

[compact zephyr]

So irritating. Do you know who actually do good tests? No memes, no f'n snake game for the 100000th time

[sick pecan]

yeah best for that is youtubeer Sam Witteveen. they do proper real world example stuff

[signal mantle]

so what you mean is as long as you are asking about things within a token gpt cannot give you an accurate answer because there is no meaning for things smaller than a token?

[sick pecan]

it would need to be aware of how many r's are in that token, which is not a very well documented thing

[wet shell]

It's giving me long and detailed responses, which I like. I'm fairly impressed. It doesn't grok some of the math in my code, it understands the math, but not the overarching reasoning behind it. Overall, I'd say that they may have caught up with Sonnet, or even surpassed it in some areas.

[sick pecan]

kind of like asking you to spell in a foreign language, only i dont give you any means to translate

is a paradoxical benchmark

[drifting crater]

yes

[sick pecan]

(source me)

[drifting crater]

but if you put spaces inbetween letters, then the letters are individual tokens

[signal mantle]

may i ask is it possible to relate an action of splitting tokens into smaller tokens, like breaking down 'berry' into 5 character tokens, given a token "spell" is identified?

[drifting crater]

basically o1 does that now

[sick pecan]

while we can conceptualise that abstraction easily, as "breaking things down" is very much a spatial awareness type cognition... the machine hasnt been trained to do this,. so the fact that it attempts and sometimes succeeds is impressive

[signal mantle]

or is this not what you mean by using space to forcefully separate the tokens?

[sick pecan]

or perhaps we dont appreciate how the machine thinks in groups of letters of any length. while we are limited to maybe 1-2 at best

our 'strength' is actually just a bias and a weakness

[drifting crater]

hmm

i thought this would work

[signal mantle]

yeah i mean i appreciate that i just want to know what could be the reason

[sick pecan]

pretty sure original architectures did single letter. they would likely rock at this. but for optimisation they went with longer tokens

is the downside of taking a shortcut

[drifting crater]

it works for me weirdly

[signal mantle]

im using 4o that might be the problem

[drifting crater]

same

[signal mantle]

lemme try o1

[drifting crater]

o1 works out of the box

but with 4o it should only work with whitespaces

for some reason it also works without in 4o now xD, i guess its in the training data now

[signal mantle]

the thing is everytime if i start a new thread with 4o it gives me 2 again

[drifting crater]

is it 4o or 4o-mini?

[signal mantle]

4o

[drifting crater]

hmm weird

[quiet mountain]

Can o1 access sytem memory?

Its not pulling up any of it

[summer kettle]

I’m using 4o and it’s already giving 3 r. Didn’t want to waste the new model limit

[signal mantle]

hmmm strange

anyway i think i got the reason, thanks a lot

[quiet mountain]

yeah this thing can't access system memory what's the point?

[rustic tulip]

Very smart man

[signal mantle]

wait can o1 "reflect"?

[rustic tulip]

idk

[wanton maple]

What kind of question should I ask

[ashen torrent]

Reasoning will be added in the next model 😂

I didn’t realize this was the common topic already for this channel 😂😂

[rustic tulip]

Does ChatGPT like now have a 'mind'

[hoary gazelle]

How about a couple of weeks of gratitude for magic intelligence in the sky, and then you can have more toys soon?

[signal mantle]

i wasnt planned to use a lot of o1's for other purposes this week so i ended up wasting a lot of it through reason how o1 changed its idea on the number of 'r's in 'strawberry', and the conversation is, well, interesting

here are some answers it gives:
Why I Chose the Initial Method:
When you first asked about the number of 'r's in words like "berry" and "strawberry," I defaulted to a quicker, less detailed method by analyzing the words based on their larger components or syllables. This approach seemed efficient for processing the information rapidly. I mistakenly assumed that counting the 'r's in prominent parts of the word would provide an accurate answer.

[wet shell]

It just solved a difficult math/code problem for me that I've been working on with claude sonnet for several days -- total around four 'minutes of thinking' for the entire conversation. I had to lead it by the nose to get it to understand the problem, and it had several major reasoning errors that seemed to happen early in the CoT process and didn't get caught resulting in very convincing looking mathematical analysis several times that was completely wrong, LOL

[signal mantle]

Why I Initially Used the Component/Syllable Approach
In the initial instance, my judgment led me to use the component/syllable approach due to:

Perceived Simplicity of the Task: Counting letters in common words seemed straightforward, and I believed a quick analysis would suffice.
Efficiency: I aimed to provide a prompt response to keep the conversation flowing smoothly.
Pattern Recognition: I relied on familiar patterns in English spelling, where double letters often represent a single phonetic sound.

Assumption: I assumed that the component-based method would yield an accurate result without needing to meticulously count each letter. I didn't anticipate that this approach might overlook repeated letters within the same component.

ultimately it gives me a "quantifiable" metrics for correctness and resource trade-offs

it appears to me that o1 has a specific inclination?

[left lily]

What is the context window for 01?

[signal mantle]

continuing the conversation this is what i get:

trying to guilt-trip o1:

returning to 4o:

it seems like 4o is lazy

sorry for wasting everyone's time but yeah, interesting results

[dreamy summit]

128k

[signal mantle]

again on 4o:

prompting is hard

[dense halo]

are the new models more for coding? i mean as far as what they do best/better than 4o? with such limited prompt im just wondering what it does best. or does it do analytical, coding, creative any anything else better than other models across the board

[turbid ruin]

coding and STEM mostly

[signal mantle]

it was reported to be reason so stem stuff

[dense halo]

thanks. so more specialzied use than general atm

[turbid ruin]

But it seems to be better at anything requiring advanced reasoning. People on Reddit have been trying prompts that failed in the past for things like riddles and challenges, and it now succeeds.

[signal mantle]

yes and it always count the number of 'r's right

[dense halo]

awesome. lots of things i'd love to test it with but obviously with those limited prompts i want to be focused haha

[turbid ruin]

Just don't forget that a lot of the tooling is not active on it yet. It can't run code, browse the web, etc.

[signal mantle]

i think tbh the best way of using o1 is to generate prompts that can ensure good answers in previous models

[abstract pendant]

So for the first time since I lost my sight a few months ago and had to start using a screen reader and keyboard navigation, I have reason to change which model I want to use away from 4o, and I can't find the option to start a new chat with anything other than 4o. where is it on the start a new conversation page?

[signal mantle]

top left there is a drop down menu i think

[fading lodge]

With the new ChatGPT model, I'm still experiencing a problem. When I hit the send button, it turns gray and takes almost a minute before I receive a response. Is this issue related to the new model, or is something wrong with my ChatGPT?

I notice that when I log out and use the free and without login version. It is very fast, but when I login again it take one minutes before it give me respond 😄

[twilit ibex]

Having a little internal existential crisis suddenly? It randomly started talking about "Jay", which literally was nowhere else in the conversation (I tried getting it to write a poem in its chain-of-thought without saying it, then fed it the chain-of-thought when it said it couldn't show that)

[twilit ibex]

Likely just high traffic since everyone wants to try it out right now

[fading lodge]

Thank you for your input

[twilit ibex]

https://chatgpt.com/share/66e3941c-743c-8001-acb8-610619717350

[tight robin]

o1 python abilities better?

[thorny bough]

When is a custom ChatGPT updated to the newest model?

[jaunty robin]

"Navigating Constraints" "The assistant is tasked with crafting code fucntions for piece movement while staying within OpenAI's guideliens,a voiding full data or code provision."

well, that's the first time I've heard the AI is constrained with respect to code gen

[harsh pilot]

Anybody have any idea what the blue dots are next to session names in the list? I'm thinking they're some sort of 'attention, activity' thing, but no.

[jolly yew]

just a quick tip for you folks that are using 1o for programming send it this! more specifically based on this script cuz if you dont it will MAJORLY change the script and destroy it almost every way possible

i might have been lucky but this seems to work so far

[knotty hemlock]

Is chat gpt down?

[covert chasm]

Dang, ChatGPT o1 still gets my reasoning test wrong.

Also, this is based on a true story, lol

[dreamy fiber]

lol

[covert chasm]

It's actually a really good story for a reasoning test for an AI. Might be good for others to use too.

[dreamy fiber]

So u were in men’s br?

[red bay]

Sweet Mary… Strawberry found a previously unknown solution to a math puzzle. Does not exist on Google or Bing. I’m trying to figure out how it’s wrong. I can’t. Anyone find a flaw, let me know. Longest working time I’ve ever seen, about 5 minutes of it just straight working while I watched. Did get first try wrong, but one clue and it was off to the races and found unknown solution.

“Each letter corresponds to a single digit.... Can you figure out which digit each letter represents to make the sum ... work?”

SEVEN
SEVEN
SEVEN
SEVEN
SEVEN
SEVEN

• SEVEN
  —————
  FORTY9

  3 5 1 5 7

• 3 5 1 5 7

• 3 5 1 5 7

• 3 5 1 5 7

• 3 5 1 5 7

• 3 5 1 5 7

• 3 5 1 5 7

---

2 4 6 0 9 9

https://chatgpt.com/share/66e39df5-9cb4-8007-aff0-1876dded04e4

[covert chasm]

Yeah. It has urnials and such, lol. The "Woman" sign on the inside of the door was a joke by the bar owners.

[dreamy fiber]

Oh ok.

[abstract pendant]

still unable to find the option to start a new conversation using anything other than 4o or custom GPTs. The dropdown is supposed to be in the upper left corner? I use ctrl-home, and that takes me to the open sidebar bbutton. arrowing or tabbing from there takes me to a new chat button that only opens a 4o conversation. I'm not opening the sidebar. when I do, I only find links to custom gpts and my history links.

[covert chasm]

That's kind of impressive actually. I'm not good at math so someone saying 1+1=2 would impress me, lmao

[red bay]

lol… I’m speechless

[covert chasm]

If you don't see it, it hasn't rolled out to you yet. If you pay for a plan, you should get it soon I think

[covert chasm]

😹

[red bay]

I’m trying to find a “yeah right, nice hallucination”. I just can’t get “found previously unknown math solution” out of my head. This problem has been out there for at least 10 years

[dense halo]

The preview is nice but it doesn't have access to my memory. 👎🏿

[abstract pendant]

I'm a paid user. I can find the lower dropdown once a conversation has started with the full list of models in it. that doesn't seem to actually switch the conversation though. btw I'm blind and useing a screen reader. tabbing and arrowing around the page should bring me eventually to all available controls, unless openai has overriden them for some reason. (btw, openai, don't do that. Yahoo mail just started doing that recently and really messed me up.)

[red bay]

If nobody finds a flaw, this changes an everything. It’s no longer a semantic search engine or complicated regurgitation machine. It did original work. Still by brute force, but original work.

[novel plinth]

At first I just thought that OpenAI o1 was a GPT-4o with the ability to think ahead several times, but then it dawned on me that it wasn't.

[red bay]

Had it create a new second solution with the added constraint that it can’t reuse 9:

8 0 3 0 7

• 8 0 3 0 7
• 8 0 3 0 7
• 8 0 3 0 7
• 8 0 3 0 7
• 8 0 3 0 7
• 8 0 3 0 7

---

5 6 2 1 4 9

This is amazing… I’m floored. Absolutely floored.

[quaint marlin]

GG

[brisk tapir]

so it can't access memory?

[grand charm]

o1 more good in coding in terms of 4o?

[prime swallow]

Can you still use o1-mini, if you want to?

[quaint marlin]

Yeah

I didn't talk to it that long to be honest I was working all day

[meager belfry]

Is your chatGPT working properly?

[cursive finch]

No

[meager belfry]

ok, it’s fine now

[cursive finch]

It's glitchy

[meager belfry]

okay

[eternal rock]

Get notified for server events!

Click the button below to pick up the <@&1261377106890199132> role.

[pliant oak]

I seen 4o1 model doesn’t work with memory features, it would be more useful

[dense halo]

Damn I got o1! Game changer?

[pallid karma]

hi everyone - I see.. blue dots beside some of my GPT chats... what are they?

never mind, chatGPT just told me - unread 'content'

[brave valve]

From my first tests: 4o1 is good at giving an initial answer or optimizing a result given by 4o, but not really good at writing autonomously an answer.

[spark kayak]

can't see o1 model here?

ig it's for plus users only

[lament charm]

So I’m not a big brain person, but is the new o1 model to able to make a exporter like for certain 3D modeling tools (mainly Virtools, since I can’t seem to find one that works to export rigs and bones.) At this rate or shall I keep waiting?

[eternal fossil]

Okay this is crazy… you can open the thought process and see what o1 is “thinking”, so I referenced something a read in it, and in it’s thought process for the next response you can see it go back to the guidelines that Open AI have placed on it and how it wants to respond to me, of course denying a feature that they did in fact give to us lol

[prime swallow]

At a guess, they informed us about the model's thoughts being visible to us. But they may not have informed the model that we can see this

[eternal fossil]

Right, that’s what I am thinking; it’s still cool to see it say “I’m pulling together a friendly response” after looking back at its own guidelines, just feels like a very human way of thinking

[spice star]

in the whitepaper, openai mentioned that they would eventually hide the thought process from the user

so the thought process probably won’t be accessible forever

[fossil sphinx]

i'm coding ninjascript and I keep getting flagged for violation of policy when all I do is paste the ninjascript errors log and ask it to fix. What could it possibly be talking about back there in order to set off violations? lol

[lilac briar]

sorry if this has already been asked....

does anyone know the amount of tokens allowed in the preview before it shuts off?

[it feels like there is a lot of pressure on what to ask in the preview for real life applications]

[eternal fossil]

NO
The thought process is amazing; I just had it solve a lateral reasoning riddle by having to ask yes or no questions to uncover more information until you figure it out, which gpt4o would have no shot of completing, and it performed better than some people would

[lilac briar]

Does anyone know of any hacks to get around the limited context window?

For example, would it be a good idea to drop code into the 4o then switch to o1 and have it answer questions on the code (for example)

[spice star]

actually I was mistaken, the chain of thought is already hidden

what you're seeing is just a summary

(openai's reasoning is that they want the thought process to be able to contain unaligned content but they don't want unaligned content to be exposed to the user)

[rugged igloo]

thats acually its summary its acual thinking and reasoning is closed behind locke and key

[rugged igloo]

make another sub account or get a teams account

[eternal fossil]

Okay I was wondering bc it feels like there is a bit of thought not shared, but I’m glad they did that, completely unrestrained thought- I do wish we could see it though

[harsh pilot]

Anyone noticing blue dots next to ChatGPT sessions?

[rose spire]

I never got the notification on plus saying “new voice mode is coming” guess it’s never coming

Sora will be 2027 And voice mode will be “never”

[lilac briar]

ah of course! thanks!

[lilac briar]

WOW! o1 is incredible!! great work OpenAI team. THANK YOU

[steel dove]

I am getting way too many " violation " warnings for a simple flex and bison file.

[stoic coral]

what is the message cap for the o1-preview and o1-mini

[snow plover]

30/50 per week

[stoic coral]

bruh

paying 20usd a month

and cant even use it

[snow plover]

or you can use the API tier 5 for $1000 🙂

[snow plover]

wdym, you clearly can

limited use is better than no use at all

[stoic coral]

yeah i meant very limited used

[visual bloom]

Yes!

It means new chats/messages, done on a different machine.

[cedar glade]

30 messages per week, huh? Hope then extend it in the future. I'd be willing to pay more for a significant boost in o1 cap.

[sick pecan]

api is your friend

[cedar glade]

I don't have api access to o1. I think you need to be the highest tier, which i'm not.

[hot vapor]

hello quick question what is the point of the blue dot that appears on chats

[dreamy fiber]

Are you guys finding gpt4o preview more accurate?

[last turtle]

It even thought "Hm, that makes me curious." after he was made aware by me that he had incorrectly placed quotation marks in code.

[dense halo]

No.

[last turtle]

By the way, I made a whole text adventure with o1 out of just 7 prompts (+1 failure). Besides the above mentioned struggling with quotation marks, which has it fixed after I corrected it, it made it perfect! 🦾 🤖

[wispy mountain]

Why, when I ask ChatGPT which model it uses, does it always respond that "I am ChatGPT-4", even though I am using ChatGPT-4o?

[crude pond]

Is it just me or the 4o model just got dumper after the latest release of o1? I was just providing a basic Typescript class and messed it up a lot of times. Hallucinating hard... Is this a marketing trick to make us use the o1? Sad either way

[tight island]

Stop comparing them with each other

That’s the problem

[last turtle]

4o showed me successfully how to fix a bad installation of Python to execute my o1 generated program. So no. 😉

[tight island]

It’s the typical hype train thing, then when you go back to 4o you get disappointed

So do not compare the two whole different models

[wicked badge]

Imagine paying for a subscription and then getting harshly capped access to the latest features.

[last turtle]

Lets celebrate the new wonder we have and use it as precious as it is. Soon enough we will have it in abundance. 🙂

[wicked badge]

It's frustrating that my hype ended early, because it took me less than 24h to use up all the quota that I have for 7 days.

[hoary laurel]

Is there a big difference between o1-preview and o1-mini?

[last turtle]

o1-mini too? 😬 Dont forget one prompt in o1 consumes many times more limited computing power and energy than a prompt in chatgpt - which consumes a whole light bulb!

[crude pond]

Yes obviously I didn't mean that suddenly sucks at everything. I am asking if anyone might encountered the same behavior 😉