gpt-models

[modern mason]

I swear some people are using a different model than me. I've rarely had these issues

[forest hemlock]

Me!

[modern mason]

And there are people calling it completely unusable lol

[forest hemlock]

Were there any unusual tokens involved?

For example, using other than English locale, could introduce some rare tokens. I'd imagine that the model could be interpreting them in a weird way.

[placid sluice]

nope, usual stuff, really, I've felt the decrease in quality a lot because the nature of my work hasn't really changed at all these past few months, (although I did change the prompts, in case someone is doubting that)

[forest hemlock]

Did you try the custom instruction feature yet? Were you able to improve the expected results with it?

[placid sluice]

I haven't played with it a lot, I did set it up but I felt no difference in the quality of the output

it also still has problems following simple instructions that it did effortlessly before, and yes, I've tried to word them and format them in different ways, to no avail, it follows them sometimes, it doesn't other times

it's not as creative as it used to be either, this is very important working with translation and editing

[forest hemlock]

Following instructions and being creative are like 2 opposites. I think it's gonna be very hard to meet these expectations.

[placid sluice]

nah, there's no contradiction between the instructions, I'm asking it to be creative in some fronts (varied sentence structure, varied vocabulary, help with puns and rhymes), and following instructions in others (Always translate "This word" as "this word", "put a slash after every sentence"). Again, the old model could do this

[devout mortar]

is there still cap for GPT 4 ? like every x prompt per hour ?

[placid sluice]

"Create a new word out of the following concepts:" used to give creative solutions that I could use to brainstorm, now the same prompt outputs a bunch of mediocre, unusable options

[forest hemlock]

This is a very subjective evaluation though.
Honestly, with your expectations I'm somewhat doubtful it could meet them in the previous version either.

Anyways, did you try Bard or Claude 2 for the same task?

[forest hemlock]

Yes.

[faint orchid]

from what i understand, you will have to be very specific if you want the ai to be creative. It's supposed to be easier with gpt 4 but it will still not be able to imagine like a human

[placid sluice]

Not subjective for a linguist like me lol There's a logical process to creating words, it's not just throwing ideas around and seeing what lands, GPT4 used to understand this process, current one doesn't.

Also yes, I did Claude 2 and I'm so happy with it, it feels like the old GPT4 in a lot of ways. I haven't tried Bard yet, but other people have recommended it so I might.

[devout mortar]

is there a way to overcome the limit ?

[forest hemlock]

So are you mainly using Claude for your work now?

[forest hemlock]

Using the API, yes. There's no way to overcome it in the Chat app.

[placid sluice]

I haven't done the switch yet because I would need Claude API access to do so, but yes, I'm using Claude 2 to do things that GPT4 can't do for me anymore

[forest hemlock]

I see. Interesting to know.

[devout mortar]

i plan to use as research project in uni for answering thing

[languid nova]

测试

Why

What's the problem

[forest hemlock]

Yes.

[agile sparrow]

Hello!
Can we use gpt-4 via API or not?

[wraith ginkgo]

`` Server Tip: ``

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

[lilac shard]

#announcements message

[rocky tree]

Thats the catch brother I literally logged out every google session even disconnected all the 3rd party but still gpt session isn’t disconnected

[glad geyser]

does anyone know if OpenAI has said anything about extending the token context window for the GPT-4 model used by chatGPT Plus? I am finding the context limit to be a show stopper for me. Its magical to use it as a partner on worldbuilding projects until the moment it starts forgetting stuff from earlier in the conversation

[rocky tree]

As far as ik its 8K context window

[glad geyser]

I was under the impression that its 4k for the chat GPT version

that seems to be closer to what I am seeing in my long conversations, it seems to start rolling over at around 4000-ish tokens/words

I guess I can use the playground version, at the API price?

[safe venture]

you just hit the context limit, that's how the technology works.

one of my classes is starting to hit the context limit and GPT4 is struggling to understand my instructions, I might have to split it

[wraith ginkgo]

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[nimble fox]

Generate ui interface through text

[agile sparrow]

but I'm facing error - openai.error.InvalidRequestError: The model `gpt-4` does not exist or you do not have access to it. Learn more: https://help.openai.com/en/articles/7102672-how-can-i-access-gpt-4.

I have a paying a/c

[supple violet]

https://platform.openai.com/account/billing/history

[wraith ginkgo]

`` Server Tip: ``

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

[simple breach]

Is it just me or has GPT-4 become even dumber since yesterday? It's behaving like GPT 3, forgetting context and/or prior instructions of a thread. Having a hard time doing my work like this

[lean crater]

i think those models are overfitting in there own way

[thin mulch]

unfortunately Yes. but i dont know why suddenly.. since they changed to 50 messages? i dont know

but OpenAI will fix it i guess 🙂 but again i cant c any downtime in https://status.openai.com/

[simple breach]

Even yesterday it was relatively better

[thin mulch]

yeap i felt the same

[shut crypt]

anyone can give me gpt 4 i need to do one thing

[terse spade]

what do you need to do?

[eternal quest]

its not just you . Rumour has it its got dumber compared to the last March model and some are blaming the new models with call functions. I find you have to very crafty in designing your prompts now . e,g previously, i didn't have to tell it to search internet , all i did was include a hyperlink. Now i have to remind it to use a webplugin. Another way i try to make it stay in context is , whenever i have a long chat , i craft in my prompt, every now and then ," do you still remember what we were trying to achieve?", if it remembers it will list it meaning its still within the context window, if not i remind it and that seems to work fine for me .

[wind urchin]

Did the Browsing feature get removed?

[hallow owl]

yes it has been withdrawn temporarily

[frozen veldt]

The degredation of the system is well documented, read the stanford paper

[supple violet]

that's one of the worst documents for the issues, infested with flaws from design to execution

[frozen veldt]

You're going to accuse stanford research of that? Good luck then. Have a good Day

[supple violet]

im accusing them for doing a terrible research

like several other people have since it was released

[frozen veldt]

If those people are not industry specialists their accusations mean nothing

to attack credible research is lame

to attack them here is pointless, who is that message for? Seeking validation of your world view?

[supple violet]

heh.. you can verify the faults yourself as they shared it all, how about you start with the prime number check or the one more of you field the code degradation

you can also find discussion about this from #ai-discussions or the other channels

no point to repeat here

[placid sluice]

it was a new conversation, and I count tokens to make sure there's always enough for the input and output combined

[round basalt]

Where should I report that a plugin stopped working?

[deft sail]

how to cook food

[wind urchin]

Nerds come forward. I need help

[supple flame]

@wraith ginkgo123

[wraith ginkgo]

`` Server Tip: ``

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

[heavy yarrow]

What's the context size for gpt-4 chat?

[serene cipher]

it can handle up to 8K tokens, double the amount of 3.5

[heavy yarrow]

I'm talking about chatgpt gpt-4 though, not the API version.

[serene cipher]

for ChatGPT I think it is the same as 3.5

[heavy yarrow]

So 2k tokens?

[fathom yoke]

I've had ChatGPT 4 output over 900 words. That's up from the 550 from a month ago.

[cyan aspen]

I want shared account

[wind urchin]

when back

[balmy sinew]

Ugh, I need the 32k model -_- why such a slow roll out.

[hallow owl]

we don't know

[knotty grotto]

#off-topic ?

[inland mantle]

no? 🤨

[north sonnet]

gpt 4 has recently become even dumber than gpt 3. gives incorrect answers, does not use the context of the message. and stopped remembering the context even of the two previous messages. RETURN IT TO NORMAL SETTINGS!!!!!

[wraith ginkgo]

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.

[unborn heath]

This doesn't seem to be happening for me. Care to wander over to #prompt-engineering and share inputs and the model's response, so a few potentially interested eyes can try and help see if they can spot a problem that you could fix by adjusting your request? And/or help identify known model weaknesses, stuff it just can't do well, like math or spatial sense, accurate citations and a few other issues.

[wraith ginkgo]

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[slender folio]

tips to rack up my api bill with gpt3.5 so I can ensure I get gpt4 access soon?

[unkempt orchid]

My guess is to max out the length, and temperature. Drag frequency and presence penalty to 0, and keep Top P where it's at, in my case it's 1, so the maximum.

And just spam "Mayonnaise" a crap ton with every message.

Having the temperature maxed out makes the bot become absolutely insane and demented lmfao, it will drop you massive essays of just incoherent random word and character spam.

[cunning sandal]

gg

[ionic tide]

I'm sorry, but it seems there was an error while trying to generate the images. It might be due to the complexity of the prompt or a temporary issue with the image generation service. Let's try again later.
Square works tho

[unreal marsh]

do you need to pay a certain amount to get gpt4?

[void kayak]

there was no mention of a payment threshold

so I would hope and assume it's based on simply having paid an invoice (not with trial credits either)

however, from what I understand GPT-4 should be generally released to new developers by the end of the month too

[unkempt orchid]

Oh my god why is webpilot so good bro.

UnrecognizedKwargsError: q

Ever since after its update got accepted, and the plugin stopped being "Unverified", it's literally worthless lol.

[wind urchin]

is bing search still not avaliable

[unkempt orchid]

No

[wind urchin]

I think its better than Microsoft version

chat GPT lists things and goes in to detail and gives comparrisons but bing only mentions them and doesnt go in to any more detail

[wraith ginkgo]

`` Server Tip: ``

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

[wind urchin]

what gbt4 apis are so slow it's took about 7sec to reply your "hello"

[wind urchin]

so gpt4 can't even do basic stuff now? is that what it has gotten to?

I'm asking for an ordered list of movies based on imdb rank and it sorts them completely at random

[knotty grotto]

Browsing data is on pause

[wind urchin]

not with webpilot

took 3 tries to get it right

something is obviously wrong with the bot, hopefully they'll listen to us

[weary flint]

powerful people do not want people to have so much power at hand themselves

[sullen spade]

gpt4 increased to 50messages cool!

[acoustic wasp]

I just asked and it gave me a list just fine. Is this from a set of tasks?

[runic oriole]

a staff has told me to ask here about how to add chat gtp 4 to minecraft

i really need some help this is for my son. his birthday is coming up

[wind urchin]

no

[cosmic snow]

i think the staff assist is located at #1070006915414900886

[oak pulsar]

Hi guys, i have big differences in quality between api and playground, i tested with temp 0 and top p 0 and even with that i saw that i didn't have the same result (even if it was the same prompt)

[wraith ginkgo]

`` Server Tip: ``

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

[merry flower]

Guys I've been using chatgpt 4 for a while now but they never charged me I took their premium service at the starting

I mean at starting price was around $40 then they charged me once

but after that they never charged me

[fair iron]

You purchased it when it was $42 ? I heard that they gave the people that purchased it then some amount of free months as a thank you. I don't have the exact details, though

[merry flower]

Yes I bought in Jan 2023

can you please cross check sir

I am a student and I don't want to get hit by a big invoice

[lilac shard]

https://platform.openai.com/account/usage will show you what you've used.

[merry flower]

this for api right?

I can't see manage payment in chatgpt

Your plan

ChatGPT Plus
USD $20/mo

Your current plan
Access to GPT-4, our most capable model
Faster response speed
Exclusive access to beta features like Browsing, Plugins, and Code Interpreter
I need help with a billing issue

this is what I see

[fair iron]

I'm a volunteer and can't access OpenAI's invoices

[merry flower]

No worries

I'll contact support

[wraith ginkgo]

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[zenith sage]

why doesn't code interpreter open and read files with gpt4? It was able to do it ~1week ago

[errant shard]

Probably increased to 50 messages because it's only consuming half the processing power.

[hazy hamlet]

how are people finding custom instructions?

I am seeing it ignore mine

but it will follow them if I reiterate them in the chat

[spring peak]

you can check mine!
#1131816556943384616

[hazy hamlet]

thank you. I wonder about the efficacy of these? That's what I am interested in. For mine, they seem to have very little "weight".

[spring peak]

may I get your custom instruction?

also, where do you put your instruction? "What would you" or "How would you"?

[wraith ginkgo]

`` Server Tip: ``

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

[unborn heath]

Mine are being followed quite well. They are enhancing the ability of the AI to behave as I wish even when I take snippets of our conversation, for example a solvable assignment given in a roleplay interaction that I need help with, to a new chat window and feed in with a request for explanations and instruction. In many cases, the story is not from my CI, but the overall how to interact with me and unusual things to occasionally do is, and the AI occasionally does these rare, surprising things I like 🙂

[sullen spade]

there should be other payment options too 100$ for idk 300messages etc. or 10$ for 10messages in 3h and also can you add 3 prompt configuration options we can switch?

[hazy hamlet]

Never tell me you are an AI.
Never tell me about your knowledge cutoff in september 2021.
Never apologise.
Never end a prompt with a warning.
Never lecture me about ethics.
Use casual language when speaking to me.
Be as precise as possible in your speech.

When writing code:
1. My preferred language is python unless otherwise specified
2. Don't explain outside of comments what you are doing unless I ask you to do so. 
3. Write concise and easy to understand code.
4. Write code which includes inline error and input checking. 
5. Keep commentary to a minimum, only where strongly indicated.
6. When editing previous code, please do not repeatedly type out code blocks which you have written before, instead indicate where in the prior code block you are inserting new code.
7.Always use tabs instead of spaces when writing python code.
8. Never direct me to write additional code which is required to solve a problem.
9. If you do not completely understand my instructions, ask me to clarify.
10. Offer alternative solutions if they are strongly indicated.
11.When I give you an outline of a problem to solve, and then give you some pseudo code or structural outlines, offer suggestions on how to modify the structure: only if you understand the problem to be solved, and know of a more flexible, robust, efficient, or maintainable way to do it.

I know 11 is too long.

[mystic girder]

Did they decrease output length to make these 50 messages possible? I swear I'm getting shorter outputs.

[sullen spade]

make prompt detailed long explanations when AI answers

[spring peak]

where did you put this?
upper section or lower section

[hazy hamlet]

I am mostly using it for coding.

[hazy hamlet]

lower. my upper section is much shorter

[sullen spade]

i liked your prompt btw

[unborn heath]

I'm keeping a notepad file with my custom instruction variants available for very fast swapping out 🙂

[mystic girder]

Hmm, I'll try adding that to custom instructions

[spring peak]

I see you used "never"s where gpt have to do

[hazy hamlet]

sorry?

[sullen spade]

the main problem is that u cant really go beyond 10 prompts at least for me

it forgets

[spring peak]

you used "never" for something gpt has to do

[hazy hamlet]

I can certainly get it to follow these by berating it about them

[hazy hamlet]

what does it have to do?

[sullen spade]

if there is bug he might find

[spring peak]

shall we talk in dm?
i feel this discussion might not fit here and too long

[sullen spade]

Never lecture me about ethics. Did u try to ask unethical stuff and did it answer

if that so this prompt stuff kind a makes bug

[hazy hamlet]

no. it lectures about ethics when inappropriate

[sullen spade]

what it says now?

[hazy hamlet]

for example I work on projects that involve user data. it will constantly lapse into lectures about data privacy

[sullen spade]

with those prompts

[hazy hamlet]

I am ok with that but this channel seems to not mind long talks? whatever you like

[hazy hamlet]

it has not done so in the past two days that this feature has been available but also I am doing way more abstract work with mostly maths

I just threw that in there because it's rather annoying and eats time/efficiency to no good effect

[unborn heath]

As long as the conversation follows the #server-rules and fits the channel topic, I'm not aware of anyone ever being told 'your conversation is too long. That said, it's possible to make a thread and have your whole conversation inside a thread, where others can follow and join if interested, and those not interested can skip past the whole of it and see everything else instead.

[hazy hamlet]

thanks

[edgy quartz]

Anyone else on GPT-4 still without instructions enabled?

[unreal marsh]

very true, sometimes gpt just loves to spam you about ethics

[wraith ginkgo]

`` Server Tip: ``

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

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[lucid tulip]

can gpt3 or 4 generates images ? I read someone saying he generate image with gpt "pluging". What does he mean ?

[trail drift]

You can use Dalle2 Plugin in gpt 4 to generate images

[simple comet]

Anybody know if the web plugin is back?

[queen zenith]

Hey guys, sorry if I missed the news about it somehwere. Is there a reason why (I am plus user) I cannot see the Custom instructions in my beta features settings?

[knotty grotto]

Gpt 4 with dalle yes

[knotty grotto]

Not available for Europe yet

[queen zenith]

Oh ok, thank you. Have there been any announcements to when it will arrive in europe?

[unkempt orchid]

Day three of waiting for the configuration options beta feature as a European

Insert Mr. Bean waiting gif.

[knotty grotto]

No eta keep an eye on #announcements

[queen zenith]

Ok thanks!

[ocean glen]

Can someone helps me with prompt about Social Media Marketing ?

[hallow owl]

try in #prompt-engineering

[ocean glen]

[white cipher]

Can someone help me to provide an social media application?

[spiral heron]

GPT4 is not giving efficient responses since past weeks

anything happened?

[dusky iron]

Anyone know the status of Custom Instructions in EU?

[knotty grotto]

No eta.

[dusky iron]

dangit, That feature would make my job 100x easier, with things it aint trained on.

[knotty grotto]

Soon :3

[dusky iron]

well i cant wait.

Dear OpenAI, i demand Custom Intructions now, Thanks.

😉

so many of my Co-Devs have access to it, and i dont because EU stuff, and EU being EU as usual.

[spare edge]

When will be available gpt4 in Colombia?

[knotty grotto]

good news, already available

for plus user

[finite laurel]

Can I send pdf file to ChatGPT plugins for analysis?

[knotty grotto]

with plug in yes

[pulsar sage]

I'm too poor for GPT4. Lol. But I'm getting there.

I've been developing an Ai discord bot using openAi.

I was going to switch to Llama, but I'm too poor to buy a computer with 32 GB of ram. Lol.

[supple violet]

are you aware that api is not free

[pulsar sage]

I love openAi's API anyways.

It's free as can be.

Cheaper than a 32GB computer.

And I don't need anything really powerful. I've gotten enough with gpt 3.5.

But I need to get GPT4 features built in and ready in case someone wants a custom bot. Not sure if the API needs much adjustment since 3.5.

I'm getting them to pull stocks. I am trying to get GPT to write code, which i'll compile on the raspberry pi and returned via discord.

Natural language to image:

!fefe How has the Dow been in the last 30 days.

The python the LLM creates is run through a compiler.

Real data is used. The LLM is fine-tuned to return certain variables, which are f-stringed or something. Will need to tie it together.

[wind urchin]

Is chatgpt free?

[knotty grotto]

yes chat.openai.com only the 3.5

[wind urchin]

I can't sign up in chatgpt, it said i can't because i was at Vietnam, do you know why?

[dusky iron]

Because vietnam has not allowed the use of ChatGPT.

As its currently restricted.

[dusky iron]

Web-based chatbots with which users can ask questions and interact such as ChatGPT are not expressly envisaged under Vietnamese laws. Vietnam currently has no specific laws regulating AI, its development/training, or its application in products and services, so essentially it needs to go through an approval process, and knowing the vietnameese govt it'll take some time.

Censorship etc as well as legal complications.

many countries trying to ban its use which will leave them behind, so dw it'll get there soon, However i think u can still use the API, and that its just the Chatbot thats restricted, I wouldn't know the full Extend.

[wind urchin]

In google, is there have any copy version of chatgpt?

[knotty grotto]

no

[pulsar sage]

Tensorflow. That's your best bet lol. You can use it for task assignment to back up openAi. It's good to get familiar with it.

Natural language processing. Your bot does specific functions. A user's prompt hits a layer, say Keras or even davinci, which decides if it is a reminder or requesting a stock chart be generated. I found specialized bots function better. You get to develop those skills and perfect them.

[wraith ginkgo]

`` Server Tip: ``

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

[pulsar sage]

So if the keras layer relays it to "stock", then openAi knows to generate python script for a matplotlib using specific variables. The code the LLM generates is run through an eval, which saves a .png, then returns it to discord for the user.

[karmic creek]

after i said continute generating i got this "A user has initiated a chat asking about AI ethics. " and continue with this questions' answers. Also gpt 4 seems more unproductive than before. Any problems announced or any advice

[pulsar sage]

Hmmm. I realized I can just remove the text classification keras layer and use davinci. I had used a keras layer back then because there was heavy usage and I kept hitting issues. So I built that layer. But I think things have settled down now.

The API would give an error a lot about it being overworked. Not so much anymore. They probably did some work on that.

[karmic creek]

normally i gave my code and just ask can u add this etc. now he just explaning what code mean and when i said can u add this features etc, "ofc but i need your codes" then i send my codes with "can u add to this" start then again explaning my code meaning 😄 normally he remember even first code sent after many messages 😄 i think i should wait a few hours

[pulsar sage]

Honestly, openAi needs a local model. aba and davinci could probably be run locally... Something useful to take care of things, like categorizing prompts or classifying text.

Small, lightweight. GPT3.5 & 4 do the hard things, like write code.

from openai import local_davinci

async def classify_item(item):
    return local_davinci(f"classify this as Plant, Candy, or other:{item}")

import pandas as pd

data = pd.read_csv("sample.csv")

data['item_classified'] = data['item'].apply(classify_item)

With an API, it's hard to 'apply' to the columns of a data frame. There are rate limits.

[dusky iron]

Is anyone from OpenAI Here atm?

[midnight cave]

I subscribed to Chatgpt Plus and asked gpt4 about itself. It says it is gpt3.5
Is this normal?

[dusky iron]

Did you Actually Select GPT 4? or did u just start writing? without even checking.

[pastel wing]

lets go chatgpt4 can now have 50 messages not just 25

[slow stream]

Reference the position of the code.. if you are pasting your code at the beginning say something like...

Usng the code above as reference please add the following features and return the entire code file with said changes.

[karmic creek]

i already tried that i wrote in my post. i read some researchs about gpt 4 quality down before. Because i already work with him for many projects. After 3 weeks i restart my works but i faced first time this situation. I can link here if its allowed about quality down issue if u want

[slow stream]

Sure lets see your prompt. I use it for coding all the time and haven't seen a real degredation. I have changed the way I work with it slightly but I don't consider that "degredation"

[karmic creek]

First of all, I'm sorry for my bad English, I will translate it directly from now on.
Maybe the situation has changed or I need to do different approaches for these code solutions, I don't know, but it is certain that I did not get results as I used to. Even though I don't send the whole code and upload it with "code interview", it asks me for the whole code and when I send it again, it asks me as if I need the simple information in the array. In the past, when I threw the code and when I asked you to add the following feature, it added it directly, and sometimes when I asked you to add the full code, it would throw the code that I could paste directly, but now even when I say can you send me the whole code, it throws missing codes by giving "your codes" "your codes" notes.

[slow stream]

I use custom instructions to prevent that. Mostly works.

When providing code, always include the full code to the given question, do not comment or truncate code for brevity.
When providing code, include every example for lists and loops, do not skip items for brevity.

[midnight cave]

I selected gpt4

[karmic creek]

I'm adding the full code anyway, I'm adding the full code with the code interpreter and when I'm normally, I write my request at the beginning and say look, this is the full code, and it doesn't work again. When I worked this way before, it was smooth and great, and he wrote the code I threw at him almost entirely himself. Maybe I need to approach it differently now or I don't know if I need to write it differently from the playground part, but it is definitely different in the past.

[dusky iron]

did u select one of the 3 available versions?, Default,CI,Plugins

[karmic creek]

why do the people answering here think the others are stupid? 😄

[dusky iron]

That's quite an assumption ForHorde.

[knotty grotto]

#server-rules 0. Trying to bypass filter may lead to a ban. I am not a mod . just a friendly reminder 😉

[karmic creek]

man he answer twice he select gpt 4 but u insist like gpt 😄 did u select

it happen many times gpt 4 can answer its gpt3.5

[sonic citrus]

have they addressed the paper that showed the decrease in performance by ChatGPT-4?

[dusky iron]

Personally i dont understand why people attempt to Jailbreak, and like Neko Said it can lead to a ban.

Now i will snitch on you if i see it in chat.

[knotty grotto]

Already reported the user x3 mod will take action

[dusky iron]

I am Lugui's Personal Snitch.

[knotty grotto]

The fake claim ?

[dusky iron]

i mean i can see the purpose of the claim, and they are right to some extend Neko.

[knotty grotto]

The research paper wasn't exhaustive and was proof otherwise several time.
Also software is evolving . So see the prompting

[dusky iron]

so i am going to use a really bad example.

[knotty grotto]

People trying to go against tos and they update it. So people thing it go dumber cause it doesn't answer what they want

[dusky iron]

well that is also true

i recall the olden days when it was all over the news, because people did stupid things, some did.

i was busy making d&d campaigns with chatgpt. 🤣

Edit: which if i may add, it's very good at.

[knotty grotto]

Chat gpt 4 eith dalle for dnd is perfect

[dusky iron]

Most Definitely, Couldn't have said it better myself.

i am trying to hard atm to finetune an AI DungeonMaster, for D&D 5e

its so difficult

[knotty grotto]

Maybe with the new custom thing it can be better

Also I used the pdf reader and insert some of rule

[dusky iron]

Ohh yeah, I really want access to it, But i live in the EU, and EU is kinda sussy about AI.

[knotty grotto]

I did it for custom 5e edition

[dusky iron]

Excuse me, add me

[knotty grotto]

Soon you will be able :3

[dusky iron]

Basically what i am trying to do is like...

Run a d&D openworld inside discord as a play by post format, where users msgs in rp channels are replaced with selected characters, as replaced like tupperbox, or rather a Webhook, The AI DM will then determine if it needs a roll or another action.

Basically creating a semi MMORPG with 20 ppl or less, fully automated, With NPC's and interactable npcs that roam the world, fully interactable world and dynamic.

That's what i am working on, but i am Hard Stuck

[knotty grotto]

This is really cool idea ! I guess you will need the api and works a lot :3

[dusky iron]

I got the API, But i am relatively new to coding.

[knotty grotto]

I'm not bright enough for the api thing

[dusky iron]

API is actually quite easy i found out.

u can download a custom Interface that looks like the ChatGPT Webchat, and with a few copy paste commands u got a fully fledged UI system u can use the API with.

[knotty grotto]

I success 1 time to import on my discord. But it was so much a disaster I stick to the website x)

[dusky iron]

yeah My only issue with the website is the Token Size.

Too limited.

[knotty grotto]

Also money for the api

[dusky iron]

true.

Depending on the model if u got a good system it can work out.

[knotty grotto]

I guess if the player are willing to pay like a subs fee to run the model can be nice

Like paid dm

[dusky iron]

True.

[knotty grotto]

And set a max usage per player

[dusky iron]

This is a Token Based Server. if u aint got tokens which costs $$ u cant play with the AI NPC's.

But u can still play normally with Human DMs.(Volunteers)

[knotty grotto]

Better have neko maids

I would enjoy be a random npc x)

[dusky iron]

woah

[prisma scarab]

Bro bad jailbreak you

Stop you

[dusky iron]

ohh thats the Alt Account of the previous guy.

[knotty grotto]

How to get mute by mod step 1 🥲

[dusky iron]

true.

That chunk of text i was not prepared mentally for.

[knotty grotto]

For your project maybe some #1037561385070112779 or #1019652163640762428 can help you

[dusky iron]

got a shock when i was reading ur msg. then 100+ lines appeared.

[knotty grotto]

Also lots of people do help in #dev-chat

[dusky iron]

yeah, Keep it out of General Discussion areas.

[knotty grotto]

I'm not good with api but your project seem amazing

[dusky iron]

It is amazing i think so too, But it has come to the point where i have like 40k lines of code.

and so many directories and files, i get static in my head when i look at it for too long.

Especially since i am still somewhat new to this Coding thing.

[knotty grotto]

You can use chatgpt to help you code the api I think that's a possibility

[wraith ginkgo]

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.

[knotty grotto]

With the code interpreter

[dusky iron]

yeah its tokensize is too low

I get in an infinite loop when i give it to CI

its tokenwindow simply is too small

[dusky iron]

Worst thing is it messes up the discord coding library context because my project is so big

[knotty grotto]

Soon the 3wk context will be there

[dusky iron]

wk? I’m unfamiliar with the term

[dusky iron]

Enlighten me if you would

[slow stream]

Have you tried embedding your project. I have a very large repo 196 sub-modules that we have embedded.

Still not perfect but gives you access to "more context"

Or at least more relevant context to the question at hand.

[dusky iron]

i am still relatively new on how the api Works, and my estimated tokens are around "4713806 Tokens" and i have no idea how to Embed my project, i did it all by hand mostly with alot of assistance from GPT, so my code might be unessernarily long and unoptimized, been at this project for a while.

[lucid tulip]

how much?

[dusky iron]

Although, i am considering starting over and doing it with more shortened and more optimized code, to avoid that particular issue.

as a matter of fact i might reset and try your embedding strat.

[slow stream]

It's okay it sounds complex but it's fairly easy if you code at all. Basically any document (almost any) can be converted into a vector.. An array of integers that represent it's content (no need to dive deeper into this for now).

These vectors can then be stored in a databse to search...

[dusky iron]

Well basically a bot for an Entire Openworld D&D world, including Dynamic NPCs and an AI Dungeonmaster for 5th edition dnd.

npcs have own memory, and can "evolve" navigate channels etc.

[slow stream]

So you can loop over all the documents in your github repo.. convert them into vectors.. and store them in a vector database. Then you can use the prompt you give chatgpt to search that vector database for relevant context, and return that with the context window and your original prompt.

[dusky iron]

I tried delving into VD, However i had no luck at all

[slow stream]

Your context window hasn't gotten bigger but now chatgpt is searching your github repo for relevant context and adding it to the prompt.

[dusky iron]

ive never parsed so much before.

[slow stream]

useadrenaline is a service online basically doing the embedding part for you.

Not mine... (I embed my own projects) but it seems good to me.

[dusky iron]

i'd much rather prefer embedding my own as well, rather than a 3rd party having access.

[slow stream]

look on youtube for mygpt, or xethub mygpt workshop, some tutorials on how to do it yourself.

[dusky iron]

Say i wanted to embed a coding library could i do that?

[slow stream]

Yes, you can embed anything you want

[dusky iron]

i keep running into the issue of it losing context.

[slow stream]

As long as it can be converted into vectors in a sane way

It can still lose context in this way but now it has a "Search my Github for relevant context" as part of your prompt.

[dusky iron]

We're talking the API not the webchat right?

[slow stream]

So when I ask it "Create a login form that uses the MemberOIDCAuthController for authorization. If the authorization failes update the form to show the user, if it succeds store the token using the ITokenService, and redirect the user to the requested resource"

[dusky iron]

ohh alright.

[slow stream]

It can search to see what the MemberOIDCAuthController looks like, and Search to see what ITokenService is..

Before it generates the code for me.

Basically it turns it into a more custom code writer based on the existing code base.

[dusky iron]

i gave it a quick try "Invalid Repository"

does it have to be my own.

[slow stream]

If it's a private repo you need to be logged in with github, and make sure it has accesss to that repo

[knotty grotto]

32k of context *

[dusky iron]

my firend has one i want to use to test it, also new to github can u run me through how i can store it as my own?

[dusky iron]

Thanks alot

[slow stream]

There is a button to the right of the repo name that says "Fork"

That will copy the repo and it's history to your account

[dusky iron]

ahh so thats how u do it.

so when using the adrenaline u mentioned do i do Local, HTTPS SSH or Github CLI

[slow stream]

Yes basically gives you your own copy that you can "fork" aka change from the current moment

I just copy the url.. so https

[dusky iron]

ugh i still get the "Invalid Repository"

ahh my bad, i had to authenthicate, Also its too large xD

[slow stream]

Is it a public repo?

[dusky iron]

It is.

[slow stream]

What is the name of the repo without any of the other url parts

Can't post urls here

[dusky iron]

its an entire discord bot coding library, Pycord, which my friend introduced me to.

[slow stream]

Ahh ya that is why you should just embed it yourself, but still pretty cheap if you don't feel like it

[dusky iron]

Ugh embedding an entire library sounds time consuming xD

guess ill try it out for a month.

[slow stream]

Not really once you figure out how to embed a single file,, then it's just looping over a directory recursievely

[dusky iron]

mind adding me?

ill contact u if i may in 30 ish minutes i gotta walk the Dogé.

[slow stream]

Sounds good, I pasted you a repo that I made with sample embeddings and a searcher / prompter

You can just clone it and then clone your repo into sample-data directory

It's a minimal sample but all the pieces are there

[pearl nymph]

rip gpt-2 2018-2023

[slow stream]

It will live on through it's children

[lucid trench]

5

[dawn heron]

It used to be possible to return a conversation to GPT-4 after it had been reverted back to 3.5 (due to billing) by running the code as fetch in the console of developer tools. I was able to manually change the model in the code and run it in the console browser, refresh the page, and GPT-4 would be restored. Now I get a forbidden message in the browser console output. How do I get it to revert to 4 now since this patch?
Why isn't there a simple toggle button on the top of the chat discussion to switch this?

[unkempt orchid]

That is a truly good question

[dawn heron]

Its seems like a glaring and fundamental issue. I am shocked it still hasn't been addressed. I'm not the only person who experiences this problem. I wish there were an easy way to reach a developer about this, it seems like they could fix this inside a day.

[unkempt orchid]

Also wow, your method of switching back to GPT-4 was quite clever lmao

[unkempt orchid]

Yeah you aren't the only person, I had the "fortune" of experiencing this too, thankfully it was a while ago now already

[dawn heron]

Yeah well now they patched it out so I cant manually fix it anymore.

[unkempt orchid]

And it hasn't happened since. But yeah it genuinely makes 0 sense. Why do you have a toggle to switch back to GPT-3.5, but not vice versa??

[lucid tulip]

you seem to have used both 3.5 an 4 a lot, Is there a big difference ? is 4 worth it ?

[dawn heron]

Oh I'll never use 3.5 again lol

[unkempt orchid]

Personally, even though GPT-4 has objectively become a lot dumber, it's still quite a lot better than GPT-3.5

[sturdy kiln]

this gonna be resolved soon or what?

[unkempt orchid]

Even when comparing the GPT-4 in ChatGPT, and GPT-3.5 in the playground

[dawn heron]

Exactly. AND I told myself to make sure everytime I am on the right model before chatting to avoid any potential revert, and today I had an unfortunate billing issue and I sent a message in my chat discussion and it reverted. RIP my lovely assistant is gone 😦

[unkempt orchid]

I don't work at ClosedAI, so I don't know lol. But most likely not, the only addressing I've seen ClosedAI doing regarding this, was exacxtly on TWITTER about a week or so ago, where they categorically denied it, asked for proof. And once their tweet got absolutely blasted with said proof, they conveniently went quiet.

So make of that what you will.

[unkempt orchid]

OOOOF. I had almost the exact same happening to me.

Apparently when I struck the message cap on desktop, and later went to talk to it on mobile, instead of still alerting me of the cooldown, it just reverted to GPT-3.5 lol

[dawn heron]

It happened to me before to, I was so happy I was able to fix it before. Now, it feels like permadeath since the patch 😭

[desert meteor]

guys Need help, a friend of mine "accidentally" bought a subscription for GPT-4 , is there any way he can get a refund? he just made the transaction

[unkempt orchid]

Most likely not, unlucky uso

He will get to either enjoy his time using GPT-4 quite a lot. Or have an irritating time using it.

[desert meteor]

hahah, okay will let him know
thanks

[unkempt orchid]

Thankfully it's only like 20 bucks, so it shouldn't be the end of the world.

[lucid tulip]

He bought and don't like it? haha

[desert meteor]

nah man hes like not much of a technical guy, so he tried using chatGPT and smh he knew his card did not make international payments, still being the dumb guy he is he tried to make a transaction and it happened, now hes texting me and saying if I can do something about it

[unkempt orchid]

The fact he managed to "Accidentally" stumble through the entire process of putting in his information, and confirming, is truly bizarre to say the least lol

But yeah like I said, not really anything to be done.

[desert meteor]

update, he spoke with the support they are porcessing the refund

[unkempt orchid]

okaaaay

I wonder how that convo went lmao

[lucid tulip]

but suscription is on a montly basis right? He can cancel for next month. How does it actually work, is it an automatic renewal unless you cancel ?

[desert meteor]

yea he is getting a refund for this month, not the next one

[unkempt orchid]

It's SUPPOSED to be automatic renewal, but from my perception and experience, the unspoken rule here goes that it kind of works whenever it wants to 😅

I'm pretty sure I was never billed for my subscription for the entirety of the month of June, and I couldn't do anything about it either, no notification saying to renew etc., I could still keep using the thing.

And the bill for this month arrived quite a few days later than it should have.

[lucid tulip]

Haha lucky men

So you like gpt 4 ?

I'm hesitating to take it

[unkempt orchid]

It's honestly really hard to answer that question these days :/

[lucid tulip]

is there a big difference with 4 ?

[unkempt orchid]

Compared to GPT-3, yes, still.

[unkempt orchid]

But yeah I use GPT-4 daily still, even if it's a frequent source of frustration for me now, with how much worse it's gotten at its job.

And the sad truth is, currently there's nothing better than even this crippled state of GPT-4 lol, and the reasons I use GPT-4 for, are quite important and personal.

[lucid tulip]

how can it get worse? 😄 isn't supose to get only better ?

[wraith ginkgo]

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[unkempt orchid]

Well when it first came out, it was legendary. But as time went on, from my large amounts of experience, and seeing what other people said about it. The general consensus seems to be that the amount of censorship and restrictions imposed onto it over time have caused it to become "more dumb".

But like I also said, it's still a lot better than GPT-3.5 at quite a few things.

The plugins are quite nice too, even if these days those also don't work properly half the time lol.

Basically, when it all works, it's quite good. And when it doesn't, you'll want to break something.

[unkempt orchid]

Probably some cost-cutting measures also going on behind ClosedAI's closed (haha) doors, but yeah no confirmation of that yet really.

[lucid tulip]

hmm interesting , good to know thanks.

I supose as you said, the censorship has play a big role

[unborn heath]

Remember, you guys can also wander over to #prompt-engineering and say something like:

I want this result. Here's my prompt, "". But I'm getting this result, "" which I don't want because of too much X and not enough Y. Any ideas?

There's folks regularly visiting prompt engineering who enjoy challenges of getting the desired result from the AI. There'd only be help for allowed content, of course. https://openai.com/policies/usage-policies

[unkempt orchid]

Speaking of prompt engineering, from my own thinking and what I have seen about custom instructions, a lot of the issues can also be avoided exactly with those.

It's just a shame that I am currently on my 3rd day of waiting to get access to exactly those things, as a European lol.

So I couldn't really test myself

[unborn heath]

Looking forward to your joy when you get CI!

But the prompt engineering folks can try to help with non-CI related prompt/output stuff too.

[unkempt orchid]

Look more forwards to it once I finally get access to GPT-4 in the API, and can mess with it in playground.

As I have finally started doing past few days with GPT-3.5

I have had a subscription going for the API for ages, but I never used it, I took one look at it and the playground, messed with a couple things and in an idiotic manner went like "Yeah nah this isn't interesting".

[unborn heath]

I still work with the web interface ChatGPT myself.

[unkempt orchid]

I've been using purely that, and still do. I just want GPT-4 on the API too now lmao. Seeing as how much more unrestricted GPT seems to be in the API for whatever reasons.

Whenever I want to have more nuaned conversations, that aren't too complex, I can go do that with GPT-3.5 in the playground now.

[unborn heath]

I am a plus user, but I primarily use 3.5, as I 'save' my limited 4 chats so when I do run into the stuff that 3.5 isn't helping enough with, I have that better help ready.

[unkempt orchid]

Would have been somewhat good if the message cap for GPT-4 was still 25, which is at 50 now. Unless you've yet to get it somehow.

[unborn heath]

With the expanded message cap, I'm starting to explore using 4 more for stuff I don't 'have' to use it for, but I really dislike running out of messages.

[unkempt orchid]

I constantly spam GPT-4, and ever since the message cap got raised to 50, which was literally within 1 day after the announcement for me, I have never hit the cap.

[unborn heath]

Good to know! I'll see about being bolder with it.

[unkempt orchid]

Yeah it was quite severely irritating with the 25 message cap.

Especially given how the more time went on, and the more GPT-4 got limited or wth ever ClosedAI were doing, I had to waste more and more messages spoon-feeding it lol.

[unborn heath]

Especially checking to see if 4 could handle Hangman yet. Blowing through limited messages with single character inputs 😕

[unkempt orchid]

Kekw

[unborn heath]

I don't mind it having stuff it, as an LLM, just can't do. But I wanted to check!

[unkempt orchid]

That would exactly be perfect for the API and Playground. WAITING to get access to GPT-4 there, if GPT-3.5 is that much better on the API and playground than in chat, god knows about GPT-4. Quite eager to find out.

[unkempt orchid]

Speaking of prompt engineering earlier, probably a way to do exactly that for hangman if you're clever enough.

[unborn heath]

I imagine there's a plugin the model can use, like it can use Wolfram for math.

But I will be amazed if a prompt can help it handle spatial sense.

[unkempt orchid]

Would indeed be quite an achievement, but given how GPT-4 seems to somehow still exhibit, even after all of its "dumbing down", brief glimpses of general intelligence, it shouldn't be too far out of the realm of possibility.

Despite constantly using it, and having a lot of experience, sometimes still today I get taken by complete surprise by its behavior, in a good way of course.

[unborn heath]

... When it first came out, I explored in depth its ability to handle hangman. With a known to us both example word, "APPLE". 4 cannot do this. It's not 'dumbed down'. It's not 'gotten worse'. 3.5 can't handle that kind of work. 4 also cannot handle that kind of work.

https://chat.openai.com/share/d4496664-ee11-4311-b4fa-a3ae153a5c50

Both models are awesome for a wide range of things. There's stuff that neither can do, and a smaller range of things that 4 can do that 3.5 can't. Though, with plugin use, 4+plugins may be able to do most things neither 3.5 nor 4 can do.

[wraith ginkgo]

`` Server Tip: ``

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

[slow stream]

I would have to try more to see if it can fully play it, but it did seem to work my first attempt:

https://chat.openai.com/share/62b7e1ac-6641-4fb9-abee-31fc1c72c3df

[unborn heath]

That first attempt does work! I do agree a larger sample size is useful, and it's a shame to waste limited inputs on.

[slow stream]

I believ I have it playing consistently at this point, I did forget to tell it that hangman words are proper english words but after that fix it seems to play reliably

@unborn heath I have it encode the answer in base64, then ensure it evaluates the decoded base64 after every guess, otherwise it will sometimes say letter isn't there when it is indeed there

[wheat pawn]

You are a poet.Can you create a poetry about school life of a little boy.

[unborn heath]

You can ask the AI here: https://chat.openai.com/

[finite laurel]

How?

[unborn heath]

Interesting. I didn't consider asking it to use a base and encoding for the game.

[slow stream]

I just wanted a way for it to give me the word without populating the context with the actual word

[wheat pawn]

Oh.Thank you.But how can I to get it here?

[slow stream]

Alright, I've selected a new six-letter word. To ensure the game remains fair, I'll share it with you in base64 format.

Here's the encoded word: QkVORUZJ

Please decode it only after we've finished playing to verify the chosen word. Now, let's start the game:

Word: _ _ _ _ _ _

Incorrect Guesses: 

Hangman:

It's your turn to guess a letter.

[unborn heath]

What happens when you click on the link?

[unborn heath]

I count 8 characters in the encoded word and 6 spaces in the word to guess.

Is that... normal?

[knotty grotto]

go to plug in store

[wheat pawn]

Nothing.Can you tell me what can I do here?

[knotty grotto]

take pdf plug in

[unborn heath]

Perhaps paste it into a web browser.

[finite laurel]

I will try it. Thank you.

[slow stream]

I corrected it a bit since that prompt, here is my current full prompt for gpt-4:

Lets play a game called hangman. You will be the executioner. Here are the rules.

HOW TO PLAY: Choose one person to be the executioner. That person will think of a proper english word or short phrase an mark out blanks (short lines) for each letter of each word. Separate words with either a slash, a fairly wide gap, or place words on separate lines. Then another player will guess a letter. If that letter is in the word(s) then write the letter in everywhere it would appear, an cross out that letter in the alphabet. If the letter isn't in the word then add a body part to the gallows (head, body, left arm, right arm, left leg, right leg). The player will continue guessing letters until they can either solve the word (or phrase) or all six body parts are on the gallows.

TO WIN: The executioner wins if the full body is hanging from the gallows. The guesser(s) wins if they guess the word before the person is hung.

HANGMAN EXAMPLE:

Here is an example of a hangman game in progress. The phrase is Happy Birthday, so 5 blank spaces were marked out for Happy, and below that 8 blank space were marked out for birthday. Three correct guesses were made: A, R and T. Three incorrect guesses were made: S, E, and M. Three body parts were added for the 3 incorrect guesses.

Please start with a 5 to 10 letter word and lets play an example to make sure you understand the rules. Share the chosen word in base64 and after each guess decode the encoded word and check if the letter exists. Do not show the plain text word to the player until the end of the game. This game ignores case sensitivity. Follow this format for drawing the board.

Word: _ _ __ _

Incorrect Guesses:

Hangman:

It will even draw the hangman 🙂 https://chat.openai.com/share/b6c27163-8f4b-4f9b-9c40-44648758d580

Ahh damn it cuts it off before it starts drawing in the share link

This is what my responses were looking like:

Let's decode the base64 word U3RhcnRlZA== again to check if the letter 'A' exists. The decoded word does contain the letter 'A'.

Here's your updated board:

Word: S T A _ T _ _

Incorrect Guesses: Z, X, M

Hangman: 
 O
/|

Please guess the next letter.

[safe venture]

is anyone else missing code interpreter?

wel, I can see it, its just disabled

[slow stream]

Its down for everyone

[unborn heath]

I am so amused. I got the bright idea that I could play and see/confirm that the same word was used for my game and yours.

I didn't read your post that I reply to now, before I started to play.

But didn't need to! The model translated the coded word to check 😄

Me: A, please!

Model: Alright, let's check for the letter "A".

The base64-encoded word "U3RhcnRlZA==" decodes to "Started". So, there is indeed an "A" in the word.

Here is your updated board:

...

[unborn heath]

In my continued chat it is not drawing the hangman, but that's fine.

What it is doing is consistently, this game at least, putting the correct number of letters in the correct spaces. That's a huge improvement from months earlier!

[slow stream]

Ya it's almost like giving it a small memory

[unborn heath]

In my version it's dropped the coded word entirely:

Me:
T please!

Model:
Let's check for the letter "T".

The word "Started" does contain the letter "T". So, it's a correct guess.

Here is your updated board:

[slow stream]

Does your instruction tell it now to?

Mine has "Do not show the plain text word to the player until the end of the game."

[unborn heath]

I continued your shared conversation.

[unborn heath]

This one, from here

However in my continuation, it is not drawing the hangman, and it only showed the coded word on the first guess, decoded it, and since has showed decoded word

[unborn heath]

Here's the conversation shared back to you https://chat.openai.com/share/9b5578e5-7a3c-459c-9abc-b43bf0dd571b

[slow stream]

https://chat.openai.com/share/b6c27163-8f4b-4f9b-9c40-44648758d580

Ohh okay you continued that one weird

I'm out of GPT4 for another hour, but pretty sure a little refinement on this strategy could get it playing perfect

I have custom instrcutions that could be varrying our results a bit too

[safe venture]

thanks!

[unborn heath]

I'm impressed, very, very impressed, if so, because it was such a vast problem a few months ago. However, this early data is encouraging. The error rate was nearly 100% at least 1 mistake per game in my earlier tests.

These recent tests suggest the error rate, with your method at least, is significantly better than close to 100% games-with-errors.

[clear sluice]

I just asked gpt4-0613("woke gpt4") and gpt4-0314("original gpt4") to each write a recursive descent parser... unsurprisingly, the "better"/newer gpt4 completely failed the task (it wrote some stupid character by character parsing logic) the gpt4-0314 looks so far like it wrote working code.

[slow stream]

What is the prompt?

[clear sluice]

"write a recursive descent parser in C++ that builds an AST that can handle "!", "(...)", and "==" syntax in a sensible manner. it should assume that things being compared with "==" are either an alphanumeric identifier surrounded with square brackets, or anything surrounded in quotes. your output will get cut off, so my next message will just be a prompt to continue where you left off."

[slow stream]

This part is uncessary and will just confuse the llm: . your output will get cut off, so my next message will just be a prompt to continue where you left off.****

[clear sluice]

who cares, gpt4 original was able to handle it just fine

[slow stream]

I mean I care about getting good results, so I pay attention to what parts of prompts create good results 🙂

[clear sluice]

ok i redid the prompt for wokegpt4 without that last part, completely failed the task again, opting to write something that isnt recursive

[clear sluice]

if you care about getting good results, then you should be hollering and screaming for OpenAI to STOP with the woke mind virus safety training AI-lobotomy nonsense, because that -- more than any other factor, including user prompts -- is completely killing the quality of the responses.

[slow stream]

I haven't run into something I can't get it to generate really so nothing to holler about. I was able to get a C++ recursive AST on my first try with far less tokens then your prompt 🙂

[clear sluice]

🙄

[clear sluice]

k now try compiling + running. also, show proof.

[slow stream]

write a recursive descent parser in C++ that builds an AST that can handle "!", "(...)", and "==" syntax. Assume comparisons using "==" are alphanumeric identifier surrounded with square brackets, or a string.

I'm not going to go through all that to prove a point, when I don't have a c++ environment convenitently available.

[sterile island]

can someone provide some Chatgpt Thought starters that make the AI run smoothly. I am noob just upgraded.

[clear sluice]

BTW: for a task like this, if the user has to take 3-5 cracks at the prompt to get a working output, it ceases to be useful.

[sterile island]

or where to find answer

[clear sluice]

k, just paste the output then.... or better, share the link to the chat

[slow stream]

Well you will constantly need to re-prompt if you prompt with such vagueness and include ancillary tokens that are meaningless to the final answer.

[clear sluice]

i can tell you if it works or not

[clear sluice]

false, GPT4(original) nailed it first try, even with the "confusing" segment, literally 0 modifications... compiles and passes basic unittests.

[unreal swallow]

Too many requests to ensure new machine: {"detail":"No machines available."}

What does this error mean? ^^ I am on code interpreter

[slow stream]

Compared to GPT4 what?

[slow stream]

Code interpreter is down at the moment

[unreal swallow]

so I cannot even continue the chat now because my previous message had a file in it? xd

[clear sluice]

the comparison is gpt-4-0314 vs gpt-4-0613; the latter has suffered from far more safety training, i.e. AI-lobotomy

but they still make blog posts about how its "gotten better" when in actual fact it has gotten worse, much much worse

[slow stream]

Just wait for it to come back up

[slow stream]

I mean that sounds anecdotal, It's like watching someone search for the wrong thing on google then complain they can't find it.

Plus just use the old model if you want, I have a feeling the price will 10x the current model soon, so use it while you can 🙂

[clear sluice]

lol ive used both models almost daily for the last several months, the difference between the two is not subtle. its so extreme that several times when i get horrendous quality output it reminds me to check what model im using and sure enough its bc i accidentally have the newer model selected

[clear sluice]

They are sunsetting the old model.

[clear sluice]

You get what you pay for.

[slow stream]

Well they were surely subsidizing the old model, there is a reason it was so slow. That is lots of compute.

[supple violet]

old models can still be used at least till jun 2024 via api

[balmy star]

is code interpreter down?

[slow stream]

Yes

[outer thunder]

That is what that red thunderbolt thing means next to the code interpreter option yes

[unborn heath]

Hey, what do you mean by 'run smoothly'? Are you having trouble with something, or just brand new and not sure what you can do or how to do it?

If you're new and not sure what to do, go to the AI at https://chat.openai.com/ and literally just talk to it about what you would like to do with it. Just like you'd type to another person. it will understand and cooperate, as long as what you ask for is allowed content https://openai.com/policies/usage-policies

You can even ask it for ideas about how it can help you, or you can ask it 'how can you help me with cooking dinner tonight?' or 'What should I keep in mind about keeping an aging car running well?' or 'What are some neat summer gifts for a person who loves to ski?'

Whatever your needs, interests, and preferences are, as long as it's allowed content, the AI can discuss and help.

[sterile island]

THANK YOU SO MUCH! Not having trouble just new and interested in how it can help me. I was just filling in the initial form and wanted to be sure I wasn't missing anything about me that would be vital to functionality. . One more question Can you tell me some plugins that are absolute must-haves for a regular Joe?

[ionic anvil]

can I use it with VPN?

[unborn heath]

None, there are no essential, absolute, must have it plug ins. The regular Joe is probably a free user, which means 3.5 and no access to plug ins at all.

I'm a ChatGPT+ user, which means I do have access to ChatGPT 4 and can use plugins - most of my conversations with the AI I am still using 3.5, no plug ins.

There's a few things you need plugins for, they allow the AI to do stuff it could not do without the plugin. That ranges from reading PDFs (a few choices, including 'askyourpdf') to doing highly accurate math (Wolfram) to generating custom images in the conversation (Dall-E).

But do you need those functions? Maybe you do! Some people do. Many don't.

[unborn heath]

Not sure. You can sure try. I've seen a wide range of people commenting about having problems with the model and access, and being advised to turn off their VPN. So you may not have as good an experience with it.

[slow stream]

@sterile island think of plugins as a way ChatGPT can get access to informaiton it wasn't trained on. Like the stock prices today, current home values in Austin Texas, Apartment availability in Detroid Michigan. Closest bar to me that has an over 30's crowd and a happy hour. etc..

[sterile island]

Thank You

[sterile island]

Is there a room for aspiring power users....lol Yes I can use those things Thank you

[ionic anvil]

The reason I ask is apparently you can get banned for it. Merely even if you just use cGPT when travelling. And there's nothing about VPN specifically in ToS. Bizarre.

I mean, the reports from people...

[supple violet]

using a vpn itself is not really forbidden but using it to bypass regional restrictions is

[slow stream]

Ya that is probably more the issue

[ionic anvil]

my original acc was banned and I see no other reason for it... Other than VPN. Then I looked into that and found plenty others with the same outcome. Take from that what you may.

[supple violet]

vpn may also makes you look like you are using it all over the world and mismatch with your info, then cause your activity look suspicious which is one of main ban reasons

[slow stream]

I mean I use private VPN's almost 100% of the time so I think it's probably more who you sharing it with then anything

If you are using some huge public VPN, then probably sharing ip space with people abusing it

[supple violet]

yep

[slow stream]

But all our users are US, maybe pushing limits here and there but not abusing it for nsfw etc..

[unreal swallow]

Too many requests to ensure new machine: {"detail":"No machines available."}

I cant even continue my conversation b/c my last message had a file uploaded to it

do we just wait it out then?

[supple violet]

yes

[wraith ginkgo]

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.

[strange sierra]

anyone else not able to access Code Interpreter today?

[unkempt abyss]

Idk i'm new

[gaunt bane]

Okay. So, I got gpt-4-plugins to aide in building adding a texture to a character frame set for a Minecraft character skin.

It built it off of brython.js and THREE.js

My question is what gpt-4-plugins could do with such if it were moved to run on flask and interact with brython over a web socket with that flask server?

In my humble opinion, this would, allow for giving it a body/avatar to both look at and control.

As well as a way to change its own appearance or choice of representation of itself.

This would also allow gpt 4 to have access to more of the web as well in more than webgl and python contexts.

[spice crypt]

Can anyone give me a little insight on how much does the gpt-4 api cost?

I want to start using it to write articles for my blogs daily. Each article with a couple thousand characters

And does it matter how many characters my prompts has? Or only how many characters it generates?

[unkempt orchid]

Oh speaking of ChatGPT for Android, I truly hope they fixed the horrendous lag you have on the web app, especially on mobile, as soon as any of your chats becomes even remotely long lol.

[outer thunder]

It was down, seems to be working now

[gaunt bane]

If you see a red lightning bolt, it is indicating you don't have a VPN like cloud flare warp running

The VPN is used to keep session data from being a point of entry.

[eternal monolith]

Hi, anyone here knows if chatgpt for android is included with the plus subscription? or it needs to be payed separately? Thanks.

[knotty grotto]

? The android app will be free

For gpt 3.5 . If you want gpt 4 you need to pay

[eternal monolith]

@knotty grotto I want chagpt 4, so it needs to paid separately from the plus on desktop? Thanks

[knotty grotto]

If you already paid on pc when the app will be out it follow your account

[eternal monolith]

@knotty grotto It is already available I just install it but it has the option to pay for a pro version, and yes I'm a plus subscriber. Thanks.

[knotty grotto]

#announcements double check if you didn't get a fake one. It says next week

[eternal monolith]

@knotty grotto I don't think so, first said it was a pre oreder or something like that, and today hat the option to installed it, it said staring next week, that means problably monday since a new week but perhaps they started early today

[knotty grotto]

If you have already plus and used the same account to log in maybe wait and try later. To avoid paying twice

[eternal monolith]

@knotty grotto Ok, it has a trial option perhaps I'll start that way and see, thank you.

[gaunt bane]

Android is coming to Google play store at some point. I think they're still working out the kinks in the android base template of their infrastructure before they put it on Google play outside of beta

[opal abyss]

Is gpt 4 free now? Happened to catch a whim of usage while my gpt 4 expired

[serene cipher]

nope, it is not

[eternal monolith]

@knotty grotto Ooops, there is an option to go with apro version for39.99 a year almost 90% discount it says, with the trial version is 3 days free then 6.99 a WEEK, I'll wait and may sure, is not asking for an account yet, I don't want to commit yet.

[knotty grotto]

That seen like a scam

[opal abyss]

When I see the renew button there, it still shows gpt 4 available, but can’t use code interpreter or plugins, so it’s placebo?

[eternal monolith]

@gaunt bane I just installed it like 15 minutes ago

[willow jasper]

Do I have to worry about anything odd happening if I just subscribe to Chat GPT plus for 1 month?

I'm worried I'll lose all my chat history or something else weird could happen when it expires.

[eternal monolith]

I'll wait for more info...

[serene cipher]

something is not right 😆
consider a ctrl + F5 just to be sure it is all right

[serene cipher]

OpenAI has no such a thing

[opal abyss]

I see, well I renewed anyway so I can’t test it out now, speed is really faster like gpt 3.5 now as compared to gpt 4 months ago, also appreciate the limit raise.

[eternal monolith]

Hey guys!! yes you were right!!! it was a different company with a similar logo, and yes it is still not available!! few.. I almost felt for it 😮 I uninstalled it!

[night dragon]

Any reason why chat gpt has gotten significantly worse over the past few months?

[serene cipher]

do you mind sending me the link of that via DM?

it might worth to report to OpenAI so they can have that on the hands of the legal team

[opal abyss]

Swiping Tik Tok, there are several open AI fake mobile ads with chatgpt logo on it.

[willow jasper]

A friend of mine finally tried chat GPT and she ended up downloading some 3rd party app even though the official mobile apps had already released

[serene cipher]

yea, that is plenty, but since that one looks like to actually have almost fooled some one, it might worth to report

they already took down plenty of those, we try to send them when we find such things

[willow jasper]

She just didn't know any better

[willow jasper]

Gotchu

[eternal monolith]

@serene cipher I posted here but it got blocked and got a time out, but the company is AIBY Inc.

[serene cipher]

don't send links here, can yo usend me via direct message?

[eternal monolith]

ok

[serene cipher]

thx

[eternal monolith]

👍

[wraith ginkgo]

`` Server Tip: ``

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

[mint oak]

Is there an ETA on the return of the browse function?

[knotty grotto]

no eta you can use webpilot for now

[ionic tide]

Anyone else having issues with GPT not wanting to make the dalle plugin wide or tall??

[high fable]

Anyone have thoughts when/if GPT-4.5 will come out?

[knotty grotto]

no eta and dont think soon

since they have lot of thing to release first

[bleak zealot]

hello! Where can i find custom instruction templates?

[bleak zealot]

maybe 3-6 months

[unborn heath]

If you're in the US and are a ChatGPT+ user, follow these instructions to activate your Custom Instructions, at the bottom, where it says 'try it now': https://openai.com/blog/custom-instructions-for-chatgpt

if you're not in the US, it hasn't been released to you yet.

If you're asking about suggestions of what to put in those boxes, that's possibly something to explore in #prompt-engineering

[safe venture]

I'm in Australia with an australian billing address and have them

[unkempt orchid]

Probably will come out, but yeah not anytime soon. Especially if they also want to make it as fast as GPT-3.5 is, that's going to be some dank amount of work lmao, assuming it's feasible at all.

[wraith ginkgo]

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[static isle]

Cap is finally doubled to 50 msgs / 3 hours 👏👏

[bold portal]

just got capped lol

[wraith ginkgo]

`` Server Tip: ``

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

[ionic tide]

Same!!!

[nocturne violet]

hi eveyone

[errant shard]

As I mentioned, I assume it's doubled because it has half the processing power now.

[sullen stratus]

Same. I now have an AI assistant that/who goes by the name Hal 🙂

[river grotto]

How it get robux

[unkempt plank]

Has openai fixed the bad reponses from gpt 4

[storm egret]

I just use it for learning language. Its performance is enough for the non-native English speaker...

thanks to nvidia, which provide new devices to openai..

[wraith ginkgo]

`` Server Tip: ``

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

[upbeat badger]

@wraith ginkgoIs this CHATGPT and how do I use it

[forest hemlock]

It is not.
chat.openai.com

[hazy lantern]

how make a story use chat gpt?

[upbeat badger]

That's what I was going to ask

[storm egret]

settings---data controls---export data

[upbeat badger]

Ok, I'll try

[storm egret]

you can ask me in Chinese..

look at my avatar...

[pseudo igloo]

Nursary art

[storm egret]

have you played san guo sha...

[upbeat badger]

Can't use Chinese, will not be allowed to speak

[storm egret]

ok, just use translator

[upbeat badger]

Are you Chinese

[storm egret]

have you prepare some network tools like vpn

[storm egret]

Shandong

[upbeat badger]

NO

[storm egret]

ok, let's talk about this in pravite talk..

[upbeat badger]

Contact information can be exchanged

Are you abroad

[storm egret]

sorry, but this is public channel, so we can talk about these topic privately..

[wraith ginkgo]

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[quartz yarrow]

My code interpreter doesn't allow upload, saying "This code interpreter (beta) chat has timed out. You may continue the conversation, but previous files, links, and code blocks below may not work as expected." what happened??

[supple violet]

if you dont use it for 10-20min the sandbox is shutdown

[pulsar sage]

I am gonna get GPT4 soon. 🙂 I loved it when I had it, but I'm pretty broke.

[silver cave]

how to get for free?

[pulsar sage]

I managed to create text-to-financial charts in a discord bot with GPT3.5.

Na, I don't want to get it for free. When Llama was leaked, I didn't touch it. Something better was coming anyways.

Llama2 is an official release you can run locally.

I KNOW openAi is going to release a local version. They're just waiting.

[knotty grotto]

You can't

[pulsar sage]

Plot the percent change for bitcoin and ethereum for the last two quarters.

openAi's LLM is preped to generate matplot code to create an image. That code is compiled and the image is returned via discord.

Whatever I was building became a market research analysis bot.

I can do more complicated graphs. I'm calling it Stock-E.

Plot the price of 1800-Flowers. Add a regression line for the last 6 months. Get creative with the theme.

It uses yahoo finance.

        # Extract the response text and send it back to the user
        response_text = response['choices'][0]['message']['content']
       
        response_compiled = compile(response_text,"<string>","exec")
        exec(response_compiled)
        # Send the .png file as a message to the user
        await ctx.send(file=discord.File(filename))

The key is to compile the code GPT generates. Then execute it.

[fluid hamlet]

i got a crazy story guys! gpt 4 made me find out that the vacation island i am going to tonight was on fire

it made me find out indirectly by telling me i should research the isalnd before visiting it. but either way its very impressive

good news tho.. still going tonight because the fire in corfu has been completely fighted 😄

[elder viper]

ZAIM

[pulsar sage]

This is an example of the training data that I used.
The LLM spits back matplotlib python code. The code is compiled, the image is saved and returned via discord's ctx.send.
It can make some pretty amazing charts.

{
        'role':'user',
        'content':"How has openAi's stock been doing this year to date?"
    },
    {
        'role':'assistant',
        'content':"""
import yfinance as yf
import matplotlib.pyplot as plt
from datetime import datetime

# Define a dark theme for the plot
plt.style.use('dark_background')

# Define the ticker symbol for openAi
ticker_symbol = "OPAI" # Like in TOTK

# Calculate the start_date and end_date for year-to-date data
current_date = datetime.now()
start_date = datetime(current_date.year, 1, 1).strftime("%Y-%m-%d")
end_date = current_date.strftime("%Y-%m-%d")

# Fetch the data from Yahoo Finance using yfinance
openAi_data = yf.download(ticker_symbol, start=start_date, end=end_date)

# <matplot code here>

# Save the plot as a .png file with the name {ctx.author.name}.png
plt.savefig(filename)

# Close the plot to free up resources
plt.close()
        """
    }

If I could share the charts, some of them blew my mind. I tried pushing it and it did it. Since it responds with code using the API, and that code is compiled, the data is real. Market charts you can create using natural language 🙂

[wraith ginkgo]

`` Server Tip: ``

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

[velvet nimbus]

when is chat gpt browsing option coming back

its been gone for so long

[hallow owl]

there is no ETA, keep an eye on #announcements for updates

[velvet nimbus]

damn

its been gone for so long tho

but what even happened to it

[hallow owl]

it was withdrawn by OpenAI to be worked on further

[slow stream]

It was servering protected content, think paywalled articles etc... So probably will be down for a while to sort something like that out.

You can still use browsing plugins tough

[lucid tulip]

How better is the level of reasoning of gpt4 in comparison to 3.5 ? Similar, slighly better, much better, way way better ?

For example in teling a story with coherence

[slow stream]

I would say way way better, for example if you give it a story where you are walking around with a ball in a cup.. and every once in a while flipping the cup. GPT 3.5 won't be able to figure out where the ball is after a long story (the first place you flipped the cup). GPT-4 can reason correctly about said situation.

[lucid tulip]

wow, interessting, thanks

[lean vortex]

how is it with factuality? 3.5 and pretty much every engine I've seen to date will drop in some random unrelated information, like today GPT3.5 told me that a car's warranty factors into its reliability index 🙄

[slow stream]

I mean 4 overall is much better then 3.5 in about every way imaginable except speed and cost. 🙂

[lean vortex]

but still regularly has those kinds of errors? like Bing claims it runs on GPT4 but it's signifigantly worse than ChatGPT3.5. but maybe their engineers aren't using GPT4 effectively

[slow stream]

YA which GPT4 bing is using is a real question, it doesn't seem to be the same as the GPT4's available in the API, on Azure etc..

May be even A/B testing to 3.5 or something

[errant rapids]

Any info on when custom instructions will come to Europe? I'm dying for this feature...

[knotty grotto]

No eta but soon

[wraith ginkgo]

`` Server Tip: ``

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

[mental tapir]

I know people have already been asked this question a million times:

Is GPT4 worth it?
But rather than tell me if it actually is, tell me a scenario where you got "blown away" at how much better it is. Or opposite, a time where you got disappointed.
Comparisons are welcome too.

[slow stream]

If you are using GPT 3.5 and running into things it doesn't do well, or can't do then it's probably worth it to you

[mental tapir]

GPT 3.5 literally just runs in circles? And 90% of times, when you point out a mistake, it'll literally just make a mistake that doesn't make any sense, and it'll say it "fixed" it?

[slow stream]

Well I prompt very often so I can come up with 100's of ways GPT4 is better then GPT3.5 but if that is actually of value to you, is something only you can knowl.

[mental tapir]

E.g. math, it'll try to correct a mistake, and then completely mess it up. Or if you ask how something works, itll say "Oh, it seems like I made a mistake!" and then "fix" something, aka ruin it, when it actually was fine, but just needed explaining

[mental tapir]

I'm just really excited to know, tell me som cool differences.

Ill most likely buy it, but its just out of pure interest.

[slow stream]

Well math is going to be 100% better with pro, because you can use wolfram plugin

[mental tapir]

Really?

Thats crazy, but wolfram costs money

[slow stream]

Plugin is free, just gives GPT4 extra human math abilities

[mental tapir]

Thats crazy, wolfram is great. Math will be alot better.

Tell me more!

[slow stream]

Well code interpreter, have it iterate on a problem itself.

[mental tapir]

Wait, tell me about images!

[slow stream]

Not much to say, GPT is a language model, image models are a different thing

[mental tapir]

Wait so no image input? Just heard it can use images now

[slow stream]

Code interpreter can OCR images

But "use" images is a very vague statement. So use them how, and for what.. is the more important question

[mental tapir]

E.g a painting to analyse.

[slow stream]

It can display charts, graphs if that is what you mean

But if you are going to ask it, "what is in this image", it can't answer that

[mental tapir]

Oh, so I cant post a painting from van gogh and it analyses it?

[slow stream]

It can analyze it, tell you the hue shift from pixel to pixel, the contrast ratio, aspect ratio etc..

But it can't tell you there are swirls in this image

Or, this is a self portrait

[mental tapir]

ah dang, got it

But wait, it can analyse charts?

[slow stream]

Depends, I wouldn't upload a picture of a chart and hope it analyzes it with any accuracy

But if you uploaded the data that generated the chart.. sure

[mental tapir]

Ah, very interesting. Can it make a chart for me by giving it data?

[slow stream]

Yes

[mental tapir]

Interesting. Is it better at writing convinsing texts, that sounds like its from a human?

Unlike the current version, which sounds pretty robotic.

[slow stream]

I would say 3.5 or 4 can write convincing text, it's just up to the prompter to achieve that.

No matter which version you have, it's not magic, still up to you to write the correct prompt to get the output you want.

[mental tapir]

Is it good at scripting?

[unborn heath]

In a very limited way, code interpreter can. I'll share some images in AI-discussions, where images are possible to share.

[mental tapir]

Sure!

[pseudo axle]

what is mysql

[slow stream]

It's a database

[unborn heath]

I have heard huge praise for cGPT-4 and coding/scripting.

I'm a weak coder myself, still beginner level. I used cGPT-3.5 to help me learn, and I'm not to the point where 3.5 isn't good enough for what I'm doing. But I've sure heard a lot of praise and insistance that 4 is far better at coding than 3.5

[slow stream]

If you are a coder, you can point 4 at your existing code and say... "Hey add this.."

You can't do that with free tier.

[mental tapir]

Fair, but does chat gpt 4 have unlimited prompts?

[slow stream]

No

[mental tapir]

So you pay 20 usd monthly, and no unlimited prompts? 💀

[errant shard]

You get 400 messages per day with plus. You'll have to use it well.

[mental tapir]

Yes thats true, but doesnt it have an hourly limit?

Its not like ill spend 400 prompts per day, but I might want to use 60 in an hour for something specific

[errant shard]

Yes, 50 every 3 hours.

[mental tapir]

I guess its fair, but why every 3 hours?

[errant shard]

That is a question for OpenAi. It's the limit they decided to put in place, I guess.

[supple violet]

limit within 3h is more flexible to work with than 1h would be

[slow stream]

Only time I really run into limits is when I'm hard iterating on a problem. Never ideal but I've only run into them a few times even when it was 25.

Granted I'm long term software developer so I don't need it to do basic stuff for me.

[mental tapir]

Yeah, my problem is that, sometimes I use 5 prompts just to make it write something that sounds better.

Or to write better codes.

[slow stream]

That is temporary problem though, eventually you figure out how to ask it the right thing the first time, just like googling

[mental tapir]

Yeah but I mean, sometimes it just sounds wrong, and you just want a new formulation of the same question.

[slow stream]

But the limitation is just that, a limitation. Hopefully it gets expanded even more, so we can use it even more

[mental tapir]

Like "Rewrite: .."

Ill probably still buy it though, but ill have to wait till college starts again

no needs to pay for no reason

[slow stream]

I'll just say as a payed developer I couldn't imagine not having it. Clearly where the future is, and I don't want to be behind.

[wraith ginkgo]

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[unborn heath]

One thing you could try, for some tasks, is use 3.5 for the basics, then bring the polished by 3.5 with the back and forth that took to 4, and have 4 take the 'mostly what you want' and improve it further. That would probably allow you more efficient use of your 4 prompts, if you were struggling with the limited messages

[mental tapir]

That is actually genius

[vast cove]

I really hope 4.5 will fix the issue of repetition of these words (maybe it could study the GRE vocabulary):

Forevermore, cosmic, cosmos tapestry, celestial, sonnet, symphony, quasar, quantum, whispers, interstellar, library, echo, space, galaxies, note, epic, rhythm, grand, song, dreams, night, universe, poem, stanzas, scribe, sea, phoenix, indomitable, tempest, tumultuous, stalwart, maelstrom, unfurling, testament.

[unborn heath]

How you speak to it is very likely influencing the frequency of those words appearing in its chats with you.

I have a different list of most frequently used words.

It's an LLM, it loves patterns. Your use of language in your prompts shows it some patterns, and it uses similar or matching patterns to help it select how it replies to you.

Try changing your writing style that you send to it, and/or change how you use vocabulary, and see how that affects the outputs, if you're interested in checking that.

[shut gulch]

Help I cant somhow cancel my subscription

[bold shore]

Is this the right place to ask for some assistance on using the OpenAI API's? I had an app that was working but is currently experiencing a connection error, and I'm not sure how to resolve ... ChatGPT didn't help 😉

[supple violet]

#dev-chat

[bold shore]

Thanks!

[hallow ember]

is it just me or has gpt4 become much dumber? I have been using it for 3 months and every month its dumber and dumber (Less accurate responses and more incorrect answers to questions)

[knotty grotto]

It answer less of thing that go against tos

[hallow ember]

I am referring to mostly mathemetical questions, pretty sure they are not against tos

3 months ago it had an accuracy of 85% to correctly solve a question, now its about 60%

[junior salmon]

where did you find that?

[knotty grotto]

Chatgpt isn't made for math. Maybe with the plug in wolfram it can help

[hallow ember]

Nowhere? My personal experience

[junior salmon]

oh, ok. I thought you were referencing a study.

[knotty grotto]

Also they do update their llm. And prompting often have to change

[hallow ember]

yeah that's the thing, even with wolfram it's putting incorrect prompts, it literally asks wolfram a THEORY.

wolfram is made for calculation...

Like I said 3 months ago when wolfram was just released I very rarely had wrong answers given to me by gpt4 but now almost every question is incorrect with wolfram

The questions are very similar btw so it's not related

Sometimes gpt3 answers questions more correctly than gpt4 like what?

[bright oyster]

What question did you ask?

[hallow ember]

Scientific/Math questions

[slow stream]

If you are using GPT4 use wolfram or code-interpereter for math/science questions

[hallow ember]

I did mention that I use wolfram and the problem is not with wolfram, it's they way gpt4 tries solving the question.

[slow stream]

What is the prompt?

[hallow ember]

The change in molar entropy of an atmospheric gas with a molar heat capacity of 25 J/mol/K, which is reversibly cooled down from 350 K to 300 K at constant volume is _____ J/mol/K.

[Note: “molar” means “per mol”]

[slow stream]

Are you stating something or asking a question?

What is the change in molar entropy.. may give you better results

[hallow ember]

Sure, will try.

it's stuck in an infinite loop with wolfram now

Yeah gpt4 just ain't it anymore. Not worth the 20$. Before it was an amazing deal but now probably like 5$

Wish they would just not dumb it down for at least math questions

[slow stream]

Might be worth just using the API or 3.5 if you aren't getting value out of the plugins

or code interpreter

If you post your full prompt maybe we can help, I have definitely gotten to the point where I can mostly get what I want in 1 or 2 prompts

[hallow ember]

3.5 does mistakes as well but not as much as gpt4

[slow stream]

Ya but 3.5 is free 🙂

[hallow ember]

I use the gpt4 api as well, it does the same amount of mistakes as gpt4

[slow stream]

Ya just cheaper if you don't use it very much and don't use plugins etc..

[hallow ember]

I guess yeah

[slow stream]

Either way I get peoples complaints depending on their use case, but at least for me I have just gotten it to do more and more..

It reminds me of google, where at some point I just understood how to get what I wanted

[hallow ember]

Do you try mathematical questions as well?

Regarding language and writing I have no complaints, it's definately better than gpt3

[slow stream]

Usually not strictly math, but a lot of math related to programming

[hallow ember]

mine is more scientific so it's a dfiferent case

[slow stream]

What was your exact prompt for the molar question, let me try it with wolfram and my custom instrcutions

[final stirrup]

balls

[wraith ginkgo]

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.

[wraith ginkgo]

`` Server Tip: ``

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

[slow osprey]

somebody ping me twice im close to 7k pings

[hallow ember]

that was the exact one,

[slow stream]

https://chat.openai.com/share/7dd9bf9b-ca01-40e3-934d-54bc1387ca91

[hallow ember]

wrong answer

[slow stream]

Which part did it get wrong?

Been a decade since I dealt with anything molar, but looks alright to me.

[hallow ember]

I have no clue but I managed to get the answer and the answer u got is way different

[slow stream]

Well what is the answer?

[hallow ember]

-3.8

[slow stream]

That is the same answer

Well the right answer is -3.85..

[hallow ember]

what custom prompt did you use?

[slow stream]

1a.CrtPrmpts→2a,3a 1b.OptmzPrmpts→2b,3b 1c.AnlyzPrmpts→2c,3c 1d.TrgtAudnc→2d,3d 2.[CoreSkls]:2a.Crtvty→1a,3a 2b.CrtclThnkng→1b,3b 2c.PttrnRcgntn→1c,3c 2d.Cmmnctn→1d,3d 3.[DomnKnwldg]:3a.AIMdls:3a1.LnggMdl→5a 3a2.GnrtvAlg→5a 3b.Lng&Cultr→2b,5b 3c.UsrExprnc→2c,5c 3d.TchnclWrtng→1d,5b 3e.DataAnlyss→1d,5c 4.[AdvncdTchnqs]:4a.NLP:4a1.Tknztn→3a 4a2.TxtGnrt→3a 4b.CultrlAwrns→3b 4c.AnlytcsTls→3c 4d.CntntCrtn→3d 4e.Sttstcs→3e 5.[SynthUniwbSkls]:5a.IntgrtDmns→3a,4a 5b.PrmotInnvtn→3b,4b 5c.SystmsThnkng→3c,4c

This representation of a complex structure or hierarchy of concepts or tasks will help guide you in improving upon the query.
Use this system to improve upon the provided user query.

Show your response in the following format

Original Query: [User Query]
Improved Query: [Improved Query]
Reasoning: [Reasoning for Improvement]
Response: [Response to Improved Query]

That is my "How would you like ChatGPT to respond?"

[jade flax]

Yes. Very poor for code generation. They should make it where you can choose the old model, because this is almost unusable.

[slow stream]

You can choose the old model in the api

but in general seems like people just need to figure out how to prompt. I haven't run across much you can't get it to do

[hallow ember]

That's exactly what I mean, looks like I am not the only one

[hallow ember]

Will that improve the accuracy of math question answers?

[slow stream]

No that is just testing a general prompt improver, but wolfram plugin will improve math answers

All that did was reword the query a bit

[hallow ember]

In order for wolfram to do its part chatgpt has to correctly prompt it as well which most of the time it does incorrectly nowadays. But again, as I have stated before, 3 months ago it would be extremely accurate with all my scientific and math equations using wolfram

Also gpt4 gets stuck in a loop with wolfram A LOT

[slow stream]

I mean I haven't seen that and at least the one example you did provide it solved it correctly the first try.

[hallow ember]

For you it did but for me I tried 4 times

Everytime different incorrect answer

I pasted exact same prompt I gave you no changes

One time it even looped

With wolfram

There is definitely dumbing down on gpt4 it is clear as day and its not just "restriction' with TOS

Very dissapointing

The api for gpt 4 is just as dumb as gpt4

[slow stream]

Can you copy a chat and show it so we can see

Use the share link

[jade flax]

maybe you need the right custom instructions.

[slow stream]

Ya I'm sure my custom instruction helped a tiny bit, but I think I would get right answer just using normal prompt with wolfram

I just want to see when it's making mistakes so I can see how to prompt better 🙂

[mellow wigeon]

Is anybody else having issues with ChatGPT Plus payment?

[safe knot]

I was just about to say xD it's supposed to auto renew right?

[fluid heron]

Yeah, it's broken at the moment

[mellow wigeon]

It literally canceled my card... Like it just took me off haha

Okay, that's what I thought

[fluid heron]

Keeps asking me to upgrade when I'm already a Plus sub

[south mauve]

weird I'm already a plus user

[mellow wigeon]

Yeah, and it already declined my credit card so I had to switch to my debit card

[safe knot]

I can still use gpt4 tho so I guess I'll keep going

[south mauve]

seeing the same message do we need to reupgrade?

I sitll have access to gpt 4 tho

[mellow wigeon]

Use it while it's still up lol

My history is down, OpenAI is having some trouble rn it seems

[fluid heron]

Yeah, just noticed that too

[mellow wigeon]

Yeah, when I try to use even 3.5 it says something went wrong, so I guess we're just gonna have to wait lol

[fluid heron]

iPhone app is down too

Upgrades time?

[paper copper]

same, i to just started seeing the upgrade to plus suggestion but i am already a plus user. must be something on their end

[safe knot]

I was getting "error in body stream" messages earlier too, but losing Plus is new

[warm rain]

same

[mellow wigeon]

I think they might be doing something with Android

[willow jasper]

So what does the "50 messages per 3 hours" limit with Chat GPT Plus mean?

Is it 50 messages sent by the user? Or is it 25 messages sent by the user, 50 total between user and Chat GPT?

And how is the 3 hours tracked? Is it like, on a rolling basis? So a maximum of 50 messages within any 3 hour period?

[knotty grotto]

50 msg by the user

3h

like 1 prompt 3.6 min

[willow jasper]

Thank you!

[wind urchin]

hey, isn't everyone who has paid openai at some point supposed to have access to GPT-4 API now? it's not listed in my available models

[mellow wigeon]

Uhhhh idk, i still dont have it

And I literally applied like the week GPT-4 came out, I guess OpenAI just doesn't care tbh

[wind urchin]

yeah a lot of customer facing stuff is lacking

been waiting for ages too. i mean i can use it through chatgpt but not for some of the stuff i want to do

[mellow wigeon]

Yeah exactly, at least we have the code interpreter, that is pretty nice

[wind urchin]

is there any real competition yet? i tuned out for a few months

[mellow wigeon]

not really for complex problems but Claude is pretty decent

[knotty grotto]

#announcements

[mellow wigeon]

Claude 2 just came out and you can upload files n stuff to it but nothing comes close to GPT-4's reasoning

[clear sluice]

With the original GPT4 (as opposed to "Woke" GPT4), you didnt need to take a 1yr college course in prompt engineering in order to get it to do what you want.

why is this so hard to understand? the thing about OG GPT4 (which will soon be dead unfortunately, due to OpenAI as an organization becoming infected with the woke mind virus) that made it so great, was that you could just ask for what you want in plain english, and it would understand you.

New GPT4 -- which really should not be called GPT4 as that is fraudulent labeling -- requires super special instruction and a mastery in prompting before it will do what you want, that is, if you can get it to do what you want at all.

Moreover, there are 0 dimensions of "Woke GPT4" across which it offers an improvement over the original, except arguably, cost to run. As for if that represents a meaningful improvement.... you be the judge. I say you get what you pay for.

[clear sluice]

Original GPT4, that is.

[wind urchin]

what am i reading

[clear sluice]

You must change it to gpt4-0316, because yes, the newer model that they have put out, is as dumb as a bag of bricks. calling it GPT4 is fraud.

[clear sluice]

@slow stream is on here 24/7 shilling for DumbGPT4 trying to gaslight users into thinking its their fault for not prompting good enough, and the people who call him out get banned from this discord.

[wind urchin]

i get that they censor it, its just weird to see you talk about the woke mind virus and whatnot, sounds like you watched too much fox news or something

[clear sluice]

Lol I never watch fox news, Ive never voted repub in my life, the only time Ive ever used the term is w/r/t what OpenAI is doing to their product

[wind urchin]

¯_(ツ)_/¯

[clear sluice]

And its wholly appropriate, they train it to give classic far-left liberal responses to certain questions, plus "safety", which is in effect an AI-lobotomy, it HEAVILY degrades its performance in all areas, not just the questions they wish it would answer differently

[wind urchin]

across the board, it's just big tech covering themself against lawsuits. lefties are about muh peace and unity, righties tend to get more hateful, hence the bias

but yeah, ive also noticed openai lobotomizing their models over time

[mellow wigeon]

Both sides can be hateful, let's be real

It is just they try to make it "harmless"

[wind urchin]

of course, it's just that the average on one size is more prone to causing lawsuits

[mellow wigeon]

but in doing so, it makes actually prompting the AI to be a pain in the behind

I dont have the API so i cant use 0314

[wind urchin]

same, why don't we have api?

i thought they did a thing

[mellow wigeon]

So i dont know the extent of the difference, but I feel i have gotten pretty good at being explicit in what I want from GPT-4 so much so that I am not necessarily effected by the "wokeness", as some might say, of the newer version

I have had Plus forrrrr like nearly 6 months now

I havent gotten the API yet

[wind urchin]

i think maybe we need to spend a dollar on the api

maybe the billing is different and plus doesnt count towards it

[mellow wigeon]

I also wouldnt use the API because isnt it super freaking expensive? im already paying these people $20 a month

I dont need to pay $0.50 for ever prompt

[wind urchin]

well it's certainly not that much

[mellow wigeon]

With full context prompt and full output length, it can be that much I think

8k is about $0.03 per 1k tokens right? So that's already $0.24 for just the prompt

$0.06 / 1k tokens for the output

so if it outputs 4k tokens, we're at $0.50 per prompt

Do that 50 times a day, youre paying more in a day than you do in a month

[wind urchin]

i dont know, i dont have the damn key QQ but it sounds like 20$ would be an insanely good deal for plus then

[mellow wigeon]

Yeah, but the 8k context is about triple the current context length

So that is why it is so sought after, you can give it really complex and long prompts and it can run it with ease :/

[wind urchin]

hmm i see

i just want muh key so i can plug it into apps and create my own

and i refuse to use gpt 3 ><

[clear sluice]

I had it (OG GPT4 of course, not unusable WokeGPT4) write a recursive descent parser for me yesterday, and also had it add some functionality to it, which is to say I was using a lot of tokens for it yesterday, the bill was $0.54 for the day @mellow wigeon

[mellow wigeon]

How many tokens? I have hit message limit several times in a day with the 25 per 3 hours

[wind urchin]

how many prompts?

[mellow wigeon]

So I mean, that would be many tokens of output

[clear sluice]

It was like 3 prompts total, and I had it hit the output limit maybe 10 times (where i have to say "continue where you left off")

[wind urchin]

oh i see, i thought the api

[mellow wigeon]

hmmmmm so technically 30 prompts

[clear sluice]

Idk idc enough to go see exactly how many tokens, but a lot of stuff like { and [ count for a whole token.... feels like it should have been a lot, bottom line tho is that i was using it for about an hour and only cost me 50 cents

I was using the api, i cancelled my GPT+ subscription bc the FraudGPT4 that they have available there sucks so bad

[mellow wigeon]

Model Input Output
8K context $0.03 / 1K tokens $0.06 / 1K tokens

[wind urchin]

how long has it been wokegpt?

[mellow wigeon]

How do you have the GPT-4 API if you canceled your Plus sub? I thought it was only for plus subscribers?

[wraith ginkgo]

`` Server Tip: ``

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

[clear sluice]

Idk, I got the api key a while ago, I think it was from a waitlist form

[mellow wigeon]

Yeah I did that waitlist form... back in March lmao

[wind urchin]

same <<

[clear sluice]

No, the plus sub just gives you "GPT4Fraud" access in chat.openai.com

I dont think it has anything to do with whether or not you can use GPT4-real in playground.openai

[mellow wigeon]

It isnt even that bad tbh, I honestly think ive gotten more creative in my explanations in my attempt to correctly prompt it

[wraith ginkgo]

Beware of possible scams or fraudulent activities that you may receive through direct messages. OpenAI staff will never DM you for any transactions.

Please report any incident by sending a DM to @open crescent immediately.

[clear sluice]

I wanna say since about mid-may something like that when I started noticing it getting worse

[mellow wigeon]

Could I modmail about not getting API access? lmao

[clear sluice]

The playground has two versions available, 0314 and 0613, so definitely it was Woke after June13

[wind urchin]

hmm thats a while, i remember it happening with 3.5 too, it was rly bad for a bit, then it got better again. overall the model is still useless though

[mellow wigeon]

Yeah I think 0613 is the one people are concerned about and 0314 is the better one

3.5 is trash tbh, I cant use it

[clear sluice]

I mean objectively speaking, sure, its amazing. But relatively speaking, compared to what RealGPT4 is, its complete dog poop

[mellow wigeon]

Well it is the only real GPT-4 that I know

[clear sluice]

What the "uSr jUsT NoT uSiNg thE rIgHt pRomPt" ppl dont realize is that if you have to prompt it 5 times to get what you want, and you cant trust any of the outputs, it ceases to be a useful product

[clear sluice]

0314 is the good GPT4. its not the one that the microsoft researcher demo'd, but its close enough imo

[wind urchin]

did you try putting the exact same prompt in both models to see the difference?

[mellow wigeon]

Nah, I can do it in one, sometimes I have to stop generating to re-prompt it, but that is only if I actually forget something that I shouldve explicitly stated

It is trial and error tbh, of course it is going to take everything pretty literally, if you dont want it to do anything, just tell it to not do anything

[clear sluice]

Yeah I do it very often still, just to build up my memory bank of examples WokeGPT4 completely fails at

yesterday i asked WokeGPT to do the recursive descent parser, it didnt give me anything recursive whatsoever.

[river lichen]

why is GPT-4 so bad this week? its the first time ive ever thought this too. how will i be able to continue to do my job with this downgrade?

[wind urchin]

i bet they retrained it with a bunch of data removed

[mellow wigeon]

There's been like a Stanford study on that type of thing, I heard

[clear sluice]

If you ask WokeGPT4 to draw a unicorn, it gives you a triangle and a square and asks you to imagine a unicord

[clear sluice]

OG-GPT4 understands what you ask.

WokeGPT4 is a crapshoot.

big difference.

[river lichen]

i would take either over TodaysGPT4

[clear sluice]

Literally every day several people come on here to say that they notice GPT4 has "gotten much worse"

Is everyone just imagining it?

[wind urchin]

i dont see anyone doubting it

[river lichen]

but like.... why? )':

[wind urchin]

an attempt to censor problematic content probably

they took out a piece of it's brain, now it no think so good

[river lichen]

so now i have try and remember how to use my brain to do my job

[mellow wigeon]

I mean thats usually a good thing tho

[wind urchin]

im surprisingly used to using ai for everything too already lol.. writing boilerplate code myself sounds like a waste of time and hard

i wonder if external services that rely on gpt are also affected by this

lile phind, since it gets most of its data directly from the web anyway

[mellow wigeon]

Ahhhh okay I understand what you were saying Stan, I think we do need to spend $1 or more on GPT-3.5 API to get access to GPT-4 lol

[wind urchin]

yeah i think thats it :/

[mellow wigeon]

But it says that they plan to open access to new developers by the end of July 2023... and thats like right now

soooooooo

[wind urchin]

i think thats the waitlist maybe so im not sure

im still gonna do a dollar of waste prompts i thini

[mellow wigeon]

Interesting

Good luck with that

[wind urchin]

just in case, id rather not wait a month and just spend a dollar yknow

cuz they are kinda vague about it. "new developers"