chatgpt-discussions

[paper dawn]

It's got worse since last night even

[low chasm]

I'm using the mobile app, tried to clear cache, data and logged in again. still broken!

[boreal lily]

try the web

[brisk path]

Again, y'all are crazy 🥲

Everyone always says it's degraded every single time

Ever since launch

[lusty island]

I heard some rumors about GPT-5 releasing early 2024, is this confirmed or just speculation?

[tawny crater]

What part of "for a month or two before fixing" can you not read? At no point did i say it was permanent.

[restive plover]

Did they get rid of the limit now?

[brisk path]

I've never heard anyone say that it's ever been fixed before

[restive plover]

Is it no longer 50 prompts?

[tawny crater]

Mostly likely rumors, they'd do 4.5 first then 5.0

[lusty island]

is GPT-4 Turbo not 4.5?

[tawny crater]

still 50 for me.

[tawny crater]

no idea -shrug-

[restive plover]

I’ll have to check mine

[brisk path]

Turbo is more like 4.1 for me

[lusty island]

ah alright

[orchid pollen]

grrr reached the limit 😄

30 min break

[brisk path]

Not really a change but more personality + all the actual upgrades that don't really affect use that much

All tools though is really the big point about the update

[lusty island]

I like how it combined Searching and DALLE 3 into the basic GPT-4

[potent mulch]

Does anyone know how to make the zapier custom actions work?

[brisk path]

I could care less about Dall-e 3

the stealing images and stuff

But searching is legitimately useful

[lusty island]

its super neat imo, it can actually do words now and kinda spell

[brisk path]

That's not my problem

It is pretty neat

But what use does it serve?

It does negatively affect artists in an indirect way

Not the technologies fault exactly, besides the database, but companies replacing artists with it when it isn't even in a usable state

[lusty island]

fair

[turbid torrent]

is it crashing again ?

erroring hard when I want it to make image

hmm interesting, he has 0 problems with it in new converation

[blazing belfry]

chatgpt is struggling

[ionic geyser]

sl0w

[charred jay]

how good is turnitin's ai checker

i heard there is one that teachers can pay extra for

[rugged igloo]

all ai checkers are bad

[charred jay]

yea but apparently they have one where they pay extra for something even better

[rugged igloo]

just say grammerly checked it since it uses ai and it breaks the system just know what you talk about and you can get away with it

[charred jay]

they alreadfy said to delete grammarely

[rugged igloo]

f that im not deleting grammerly even microsoft word has ai in it now lol

reminds me of the 80's when they were so hung up on banning calculators lol now look

[runic ridge]

Hmm....something seems to have gone wrong.

[analog sapphire]

I've been deeply testing the new feature 'custom gpt' since its release like a week ago, I got the feeling that is legit the same as the 'custom instructions' feature for new chats that we already have. Also I don't really notice an improvement in the quality of their answers...is it just me or did anyone else have the same feeling?

[wise anchor]

why cant we use plugins in gpts? will that functionality ever be added?

[runic ridge]

Since you can't have multiple sets of custom instructions without manually changing it I could see GPT's as a way to have multiple sets of custom instructions. Used like presets

[wise anchor]

imo its not fair to use AI checkers, because correct me if im wrong, the only people who could say for certain that something was generated by a model are the people who created the model themselves

An AI checker said my essay that I wrote in 8th grade before AI gen was even out was 100% AI written lmfao

[runic ridge]

I have a GPT for making images that defaults to photo style and 16:9 and automatically adds details to the prompt for camera and lens and tries to use the best cam/lens/settings for the scene. It's silly but a fun shortcut for random images I make... I only have one right now but I could see making several with different variations for different uses.

[daring grotto]

anyone else keep getting network error everytime they send image

been getting for an hour

and stuck on analyzing on mobile app

[agile summit]

my gpt 4 has reached max usage :(

[runic ridge]

ChatGPT right now when it replies reminds me my old days with a 1200 baud modem waiting for the text to slowly output line by line.

[harsh geode]

It works and you get flagged

[eternal rock]

This is a Safe for Work Discord Server

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

Posting NSFW content will result in a permanent ban.

[harsh geode]

some of my friends have been caught

[shrewd prism]

not even them. you can't tell if text was AI generated or not unless it has "as an AI language model" or something obvious

[golden pier]

any reason why chatgpt is so awful with mips assembly?

[elfin birch]

mips?

[golden pier]

or just assembly in general

funny how its hallucinating most of the time if i ask for assembly code

[elfin birch]

Most common cause for ineffective outputs are ineffective prompts
Take a look here for two resources I find conducive to better prompt building #openai-chatter message

[rugged igloo]

i wish i can tell gpt to add content to my word dcoument in real time when using it

[ebon shoal]

I can’t remember the name of it but one of the AI “detectors” that became really popular earlier this year was shown to have an accuracy rate worse than choosing randomly between false/positive

[jolly temple]

is it just me or gpt 4 suddenly cant understand iamges?

The image you uploaded is showing an error message stating "There was an error processing this image." This indicates that there was a problem with the image file that prevented it from being displayed properly. The issue could be due to various reasons such as file corruption, an unsupported format, or a problem with the platform or service trying to display the image. If you intended to view or analyze a specific image, please try re-uploading the file or checking the image format and compatibility with the platform you're using.

[quasi coral]

Was the token limit increased for GPT4 for Plus users during the dev day announcements?

[daring grotto]

yes me too

started around 3 hours ago

[jolly temple]

Yeah is annoying

[daring grotto]

its so slow too for some reason

[elfin birch]

a lot of people using the product
100M weekly users + everyone coming in for the new stuff

[daring grotto]

they should let u buy gpt plus plus 🤣

[elfin birch]

That would be Enterprise 😛

[daring grotto]

oh ye u right

[dense halo]

USS Enterprise

[solar barn]

because there is not as much public code for assembly comparing to other languages, so it had less data to train on

[golden pier]

true true

[drowsy fiber]

HI, anyone failed to use chatgpt?

POST https://chat.openai.com/backend-api/conversation 403 (Forbidden)
FatalServerError: Something went wrong. If this issue persists please contact us through our help center at help.openai.com.

[kindred parcel]

Yep, crashing for me too

[drowsy fiber]

it work in mobile but not browser

[kindred parcel]

interesting...

Not working for me on mobile

[plush karma]

Hi guys, whenever I create my own GPT and then try to use the voice feature to talk to it in the app, the voice it uses is a weird depressed voice and it's not even one of the 5 voices to choose from, it's a completely different voice alltogether. I just want my ChatGPTs to use the Breeze voice. But the voice it's using is this bizzare monotone voice that literally isn't even one of the present ones. Has anyone else had this issue?

[prime swallow]

Just checked. Not affecting me

[plush karma]

Thanks for checking! Might just be affecting me

[abstract kiln]

Hello lovely community! Can someone please explain to me like a 5 years old why the "upload images" function is only available as long as you do not choose any plugins ? Or is there a trick?

[tranquil wigeon]

lets pay chatgpt to get regenerate every message top KEEP IT UP CHATGPT TEAM

you are the best

makari ola ta makari

[dusk forge]

unusable product

[drowsy tinsel]

plugins is deprecated

[deep cloak]

Hey ive been a premium user for months but im getting capcha on eveyrthing I do on chat gpt even though ihavent been using bots or such. I feel having to do 2 capchas everytime i put in a single input is excessive is there a way to resolve this?

i asked on both the chat gpt reddit and the open ai forum and I dindt get an answer. I also asked the open ai help bot itself and i asked chat gpt itself but basically got gaslighted. How do i resolve this?

If its something really obvious and simple by all means explain like im 5

[dense halo]

hii, is it just me or chatgpt-4 is realllly slow? I just got the subscription today and its slower than 3.5

[warm gull]

yes - its crawling

[dense halo]

has it always been like this or due to some updates?

[warm gull]

its the recent changes announced cauing high demand i think

[dense halo]

i see, thanks!

[deep wagon]

it's crawling and so far i've not been able to get any query complete without multiple "error in input stream" errors

[warm gull]

once the americans wake up it will be even worse....

[rugged igloo]

LOLOL

[grizzled badge]

They need to fix this lag. My friend already canceled their paid account. I'm a bit more patient.

[glossy glen]

trying to create an infographic using GPT4 - I get this message "I'm unable to assist with requests that involve the creation or editing of infographics or other detailed visual representations of data such as charts and graphs." I kind if thought dalle integartion would make this possible. Any ways around this?

[dusty jackal]

Hey guys,

Could it be, that OAI has changed ChatGPT from GPT-4 to GPT-4-turbo? I'm noticing and unbearable drop in quality for absolutly anything. Always ignoring some parts of prompts, never fully doing what it's told to, always making errors.. barely a help anymore..

[brisk path]

It seems to be affecting me too

[bronze venture]

when will chatgpt turbo come to pro?

[brisk path]

The voice sounds weird but it's okay I guess

[brisk path]

It's already in plus

[hidden burrow]

But the name is it Gpt 4 not gpt4 turbo

[bronze venture]

its input context window is still quite small, and it does not remember much

what about the turbo thing, on the dev day they had announced that it will be coming in the upcoming week/s

[dusty jackal]

It's just bad. It way worse then GPT-4

[brisk path]

Nahhhh

[hidden burrow]

On my models only appear GPT3.5 and GPT4 not the turbo version

[brisk path]

Well

[dusty jackal]

They just call it GPT-4 in ChatGPT even though its the turbo version

[brisk path]

GPT-4 turbo doesn't have any visual changes in the UI

[brisk path]

Yep

[dusty jackal]

just like they call GPT-3.5 just 3.5 even though its turbo

[brisk path]

I mean, they don't say 3.5 is 3.5 turbo in the UI

Kinda stupid though tbh

[dusty jackal]

yes

[eternal rock]

`` Server Tip: ``

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

[brisk path]

Funny thing is

3.5 actually shows the model

[hidden burrow]

So the GPT4 is now the turbo version

[brisk path]

Yeah

[hidden burrow]

Okey. Any tips in Programming prompts?

[brisk path]

I don't think you'd need to make any changes in the prompts you use

However, some users have been reporting degraded performance in specifically coding (I don't ask ChatGPT for code so idk)

[hidden burrow]

Okey! Thx

[serene kiln]

People who use the chatgpt app on Android, do you get all the gpts on the sidebar?

[tardy totem]

My GPT is refusing to output markdown in a codeblock. Anyone have any ideas how to convince it?

[tardy totem]

I do

[haughty elm]

I’m not sure I understand the pricing for my application:

I need to create 40.000 assistants, each with a single roughly 100kb text file.

Does that mean each assistant will cost 0.20 x 0.0001 = $0.00002 per day, and that therefore the retrieval part of of the cost for the entire set of 40,000 assistants will cost around 40000 x 0.20 x 0.0001 = $0.8 per day?

If not, how much will it cost?

[tardy totem]

You need to create 40 THOUSAND assistants?

[serene kiln]

Do you get list of all the gpts used, i still get old interface chatgpt3.5 and gpt-4. It doesn't even let me select code interpreter under gpt 4 from the app

[tardy totem]

I have one called Data Analysis. Isn't that the code interpreter?

I asked a general GPT4 about how to solve the output of markdown problem. It suggests stuff like building a post processing application. 😑

[serene kiln]

What is your prompt?

[tardy totem]

I've tried tons of prompts. A lot of stuff like "When asked to output markdown you will make sure to escape codeblocks inside the markdown properly"

I've even added "Please remember to escape codeblocks properly" after asking for a markdown.

[shy silo]

Apologies if this is super-common, but I'm struggling to fix it. I have GPT4+ and "Browse with Bing" enabled. I have created a new GPT to assist with my website build, and have previously had no problem with access my website, having GPT observe behaviour and make reccoemendations to CSS edits etc. I have started a new hat with a more customised GPT for this task, but it is telling me "I apologize for the confusion caused by my previous message. To clarify, I do not have the capability to directly access external websites, including your site at **********************. My responses are based on the information you provide and my training data, which includes a wide range of web development knowledge." Could someone please tell me why it wont access web now, and how to restore it? All switches etc seem fine, and the builder confirms it has web access, but when I update and run the model, it denies it has web access. I'm stuck.

[tall wind]

Hello team,
What else should I do after this steps so I can generate the visualization:

from openai import OpenAI

client = OpenAI(api_key='sk-...........')  
file = client.files.create(
    file=open("balancesheet2.csv", "rb"),
    purpose='assistants'
)



# Create an assistant using the file ID

assistant = client.beta.assistants.create(
    name="Financial Visualizer",
    description="An AI assistant to visualize financial data.",
    model="gpt-4-1106-preview",
    tools=[{"type": "code_interpreter"}],
    file_ids=[file.id]  
)    


# Create a Thread with user instructions to visualize financial data
thread = client.beta.threads.create(
    messages=[
        {
            "role": "user",
            "content": "Visualize the financial data and provide insights.",
            "file_ids": [file.id],  
        }
    ]
)


from openai import OpenAI
client = OpenAI()

assistant_file = client.beta.assistants.files.retrieve(
  assistant_id="asst_VugAkiyhWNwd8qLDaYTjzdPV",
  file_id="file-iDrw3y4aAdKsR9hCFoUJHWs
W"
)
print(assistant_file)

# Run the Assistant on the created Thread
run = client.beta.threads.runs.create(
    thread_id=thread.id,
    assistant_id=assistant.id
)

[bleak mica]

A text-to-speech model would be really nice to see openai implement in chatgpt

[maiden fractal]

It exist plug in for that

[sullen viper]

any idea, what is the message limit now for GPT-4?

[normal quarry]

50 messages per 3 hours

[sullen viper]

Still the same...
I hope they increase it.

[bleak mica]

What is it called?

[maiden fractal]

They are several , in market search for text or speach. Last one I tried quickvoice

[serene grove]

I've been having a little problem lately with GPT4, and that's kind of weird:
ChatGPT4 doesn't seem to take into account the custom instructions I have. 3.5 takes into consideration the C.I.s perfectly and answers me how i wish, while GPT4 seems to just ignore it and give me a dull answer.

Has anyone else having this problem aswell ?

[crisp cairn]

i haven't seen that

CI works okay for me

but i am getting network errors, and overall gpt is slow

i am wondering why they want gpt 5 when gpt4 is far from excellence...

[maiden fractal]

Gpt 5 is not soon, they push lots of updates like all in one

[crisp cairn]

it needs to be fast and needs to have large context size

and we are good...

[maiden fractal]

Turbo and higher context available for api, guess it will be next step

[crisp cairn]

turbo should be in chatgpt as mentioned by Sam

[crisp cairn]

they say they are already working on it

but still, better to focus on gpt4...

[maiden fractal]

#announcements message

[dense halo]

why gpt-4 is so slow these days?

[crisp cairn]

take a look at the Sam's recent tweet "there is a new version of GPT-4 Turbo now live in ChatGPT; you should hopefully find it a lot better!"

[maiden fractal]

Maybe for playground ?

[crisp cairn]

i think he specifcially mentioned chatgpt, and i think he also shared it on the dev day

[dense halo]

who mentioned?

[wanton crane]

https://twitter.com/sama/status/1723480961177010597

[crisp cairn]

yes, and people started to test in the web version

so it must be using this turbo now..

[honest jay]

Yeah he said at 19min of the keynote that Turbo would be on ChatGPT

[eternal rock]

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 @languid valley immediately.

[dense halo]

how can i use this turbo version?

[pine tulip]

use gpt4

[dense halo]

i am using gpt-4
i am chatgpt plus user

and is very slow

the turbo version are restricted to developers?

[crisp cairn]

i think there are too many users to the compute power 🙂

for now

[civic brook]

I'm a Plus user, and I'm using GPT-4, but it's extremely slow. I can't use it, as it always encounters errors when performing searches.

[mortal abyss]

me too

[civic brook]

How can I resolve this? Is this due to the turbo mode? How will this be resolved?

[atomic flare]

is gpt4 crashing for someone else too?

[neon kayak]

it is not responding for me as well

[dense halo]

Chat GPT is just unusable right now, it's almost ironic that the new model is called "Turbo" - they need to fix the server issues.

[crisp cairn]

super slow today

not fun to use it

[frosty spindle]

constantly crashing too

[crisp cairn]

yep

[analog sapphire]

ChatGPT4 crashing all the time

[glacial hollow]

does GPT-4 have the voice abilities yet?

[dense halo]

GPT-4 has been giving me network error since 3 days now

ChatGPT is slowwwwww

[brisk dirge]

I can confirm, GPT 4 not able to complete responses at the moment, it's also pretty slow, the servers are probably overloaded.

Won't complain since it's smarter than ever, will wait for the server issues to be fixed. You can do it OpenAi.

[honest jay]

On the app yes

[fathom tulip]

why does every gpt die after a few messages? it just keeps saying error: Message in conversation not found and thats it

ah ok, everything dead

[dense halo]

Can I DM someone, I wanna ask something that is kinda personal?

[shell umbra]

Why would you ask something personal in a server about ChatGPT?

[dense halo]

well it's personal as in about me

and it may not be appropriate to ask here

[novel shadow]

sure, go ahead if you want

[drowsy tinsel]

the speed right now is lmao

[nova zodiac]

has there been any word on if they ever plan to add a functionality to make tiling images?

[drowsy tinsel]

#ThingsYouCanDoWhileGPTGenerates

[nova zodiac]

i do my actual work, or pet dogs

tho i finaly got this lil mop to settle down in his bed

[unkempt wasp]

Is ChatGPT slow AF for everyone else right now?

Oh. Looks like it is.

[drowsy tinsel]

Yes

[unkempt wasp]

Well, at least I'm not going crazy.

[ornate tapir]

Just watching the cake bake....... 🎂 🧁

[sharp pilot]

Came here for the same reason as everyone else, slow and timeouts. :
What was funny was I had a regeneration time out error and it still asked me if it was better or worse than the last timeout error.

[cerulean quest]

glad to see i'm not the only one!

[cerulean quest]

yes but looking at the status page everything seems normal today

[ornate tapir]

I never know how to answer that! 🤣 I guess answer B is better because it didn't crash?... IDK The sad thing is, often answer A was actually "going to be" better, but it just didn't get a chance to finish.

[keen pivot]

Can't ChatGPT access to a GitHub repository and analyze it? I'm trying but it says he can't analyze the source code

[sleek aspen]

Browsing capabilities are limited to text and whatnot on websites - it can't exactly interact with them

What I'd do is download a ZIP of the repo and upload that (check the repo license first)

[ornate tapir]

Also, ChatGPT is crashing for everyone right now. Uber-slow, so some basic functionality isn't working. I've fed it some JSON manually that normally it can handle just fine, but today it just errors. Not sure if that's your issue, but just an FYI.

[outer seal]

ogey rrat

[hasty hearth]

Playground working fine

[wicked night]

I'm noticing SEVERE hallucinating when trying to analyse pictures

I asked chatgpt to translate some japanese text from a screenshot, and it gave me a completely made up translation. It recognised the language correctly, showing that it was looking at the picture, but the translation had absolutely nothing to do with the text. It did this multiple times, even when I "called it out". It apologised and then proceeded with another made up translation.
I haven't seen this kind of total hallucination since gpt-3

[hasty badger]

Down again for me

[lone stratus]

so slow :/

[drowsy tinsel]

chatgpt doing 1it/s

'network error' is not a real error, but rather just timed out

[ionic geyser]

Lag AF

[maiden fractal]

It can happen in peak time

[ripe patrol]

this might be a hot take, but I didn't really mind the 25/3h cap. wouldn't mind going back to it again if it helped how things are run in the back (also make the model more capable again)

[dense halo]

https://help.openai.com/en/articles/8400551-image-inputs-for-chatgpt-faq " Non-English: The model does not perform as well handling images with text of non-Latin alphabets, such as Japanese or Korean. "

[signal apex]

chatgpt is broken atm

[ancient sage]

Got a weird one, chatgpt is telling me it can't generate images in the style of another art type. Last I checked, you can't copyright a specific style of illustration. -.-

[ionic geyser]

Its possible its just saying bs because image generating is not really working rn

[reef lava]

Has the API issues? Sometimes, I get a response but the most time I dont get one and it justs says how many tokens it used.

[ancient sage]

Quote "I'm unable to generate an image of [my character] due to content policy restrictions. While I understand that [my character] is your personal character and [my IP world] is your intellectual property, the current guidelines for image generation restrict the creation of images that depict specific original characters, especially in detailed and specific styles like the [specific style] illustration style"

Its never done this before

[ionic geyser]

oh... well, i'd just ask again it can definitely do it...

[maiden fractal]

What was the prompt

[ancient sage]

"draw a picture of my character using <art style> as a style guide."

[maiden fractal]

Try a fresh new chat and disable CI

[ancient sage]

CI?

[maiden fractal]

Custom instructions

[ionic geyser]

hm thats smart

[maiden fractal]

I didn't saw your character it can also be against tos and policy

[ancient sage]

It could be possible I suppose. Still annoying that its inconsistent.

[wicked night]

Yeah yeah I am aware it's not great with non latin alphabets, what's worrying is its conviction about the replies, not the fact that it's incorrect

I would have expected a gpt-4 model to answer something like "unfortunately I can't translate text reliably" or something like that. Instead I'm getting: "Sure! The text says X"

It's very confidently and convincingly wrong, which is way worse

[honest jay]

The most frequent hallucinations are always regarding its own capabilities

[wicked night]

Yeah that's very clear when dealing with coding as well. Amazing tool, but very hard when it's convinced about something totally wrong

[ancient sage]

GPT: <does something wrong>
User: You did something wrong
GPT: 😮

[wicked night]

Very proactively trying to fix it though, which is more than I expect from most humans

AI > Us

[maiden fractal]

Error can happen in alpha and beta stage of a product

[digital prism]

you're paying for access to an exponential elevator into the future, but its new so be patient

[mortal plaza]

Never ending network errors. Super slow in the mornings. Extreme congestion. This is not well thought through.

[digital prism]

humans haven't thought of the best ways yet, hard to justify being this critical on something so new

[slender star]

chatgpt is so slow this morning

[digital prism]

ya network issues, im using it as a chance to work in smaller chunks

[languid otter]

Yeah network error means I have to redo the whole message. If it was allowed to just resume I would not have to redo everything 10x

[honest jay]

All Of Life Is An Opportunity ☸️☯️🕉️

[digital prism]

thats way too true, more so the older i get

[ripe patrol]

think that many of the complaints aren't without merit

[digital prism]

Your assumptions are something I'd only take seriously from a researcher working on LLMs right now.

[ripe patrol]

because the product wasn't like this before

[ionic geyser]

Is ChatGPT in alpha/beta?

[digital prism]

GPT Builder
The Science Synthesizer's behavior has been updated to provide summaries in smaller, manageable chunks, following the detailed JSON schema.

This is how you stay productive fyi..

[brisk path]

ChatGPT...

[languid otter]

lol ChatGPT plus is training data for GPT5 we are not customers. MIcrosoft is the customer

[brisk path]

Is a research preview

[ripe patrol]

this

[fiery mulch]

+1, and GPTs are their way of outsourcing sales and marketing

[digital prism]

The pessimism from early adopters is amazing this time around.. too many people hyped without any understanding of how it works

[ionic geyser]

Weird.

I guess that's their way of legal protection for using us as guinea pigs?

[brisk path]

Yeah lmao

[ionic geyser]

Makes sense.

[brisk path]

Google calls Bard an experiment

[ionic geyser]

Good ol OpenAI!

[brisk path]

Microsoft calls Bing Chat

Idk what they call Bing chat

Seems to just be a product for them

[digital prism]

So then tell me, what are these "clear limitations"?

[zenith lichen]

Did openai buff the content filters recently? things that are permissible a few weeks ago now gets refused

[ionic geyser]

yeah, not sure if their policies are different but either way i dont use bing chat

[brisk path]

Bing chat is horrible

[honest jay]

Some edge cases may have been patched out yes if they don't follow content policy, got an example?

[ionic geyser]

now that all models is available, bing chat is kinda obsolete unless if you cant pay

[steel dawn]

anyone else having difficulty with image generation this morning? i can't seem to get gpt-4 to generate anything at this point.

[brisk path]

It ends the conversation if you ask it what model it's based on

[ionic geyser]

very much so

[digital prism]

Its network issues. You'll find lots of people to complain with here 😄

[steel dawn]

i feel gaslighted by their status page lol.

[ionic geyser]

:)

[honest jay]

Can you share what you're trying to generate? And do you have any custom instructions active?

[ripe patrol]

turbo is more prudish in some respects

[digital prism]

I'm realizing I need to reevaluate whether everyone in this discord is an early adopter. The fast rise to popularity changes the equation.

[zenith lichen]

gpt4 is now turbo?

[steel dawn]

i've been trying to generate a character portrait. the model repeatedly assures me their description is perfectly within reason (there's nothing adult or sexual about it at all). but it flat-out refuses. and now also fails to generate any images, even generic landscapes.

[brisk path]

Yeah

[ripe patrol]

chatgpt runs on 4 turbo now

has been for a few days

[zenith lichen]

guess that explains the buffed content filters

[honest jay]

Unfortunately we can't rely on the model to tell us 100% about its limitations.

[steel dawn]

yeah i figured that was potentially hallucinatory.

but even if i ask it to generate "a landscape of a swamp" it errors.

[digital prism]

Go read this new paper: "JARVIS-1: Open-world Multi-task Agents
with Memory-Augmented Multimodal Language Models".

Then try typing that again 🙂
I'd say at this point, understating its abilities does the most harm.

[honest jay]

Do you have any custom instructions active on your account?

[steel dawn]

i do not.

if others are having difficulty i suppose i chalk this up to system issues ongoing.

[maiden fractal]

Yes

[ionic geyser]

Is there a timeline when it will not be in beta?

Perhaps with GPT5?

[maiden fractal]

No eta

[ionic geyser]

dang ok

[wheat linden]

anyone else still getting network errors?

[honest jay]

I just tested the base GPT-4 model and the dedicated DALL·E GPT and got swampy landscapes on both. You're not using a custom GPT by chance are you?

[steel dawn]

weird, it seems to be limited to a particular session. i can generate swamps and oceans galore if i start a new session, but can generate nothing in my existing session.

the new session generated a proper depiction based on the description of the character i pasted in, but will now make no further images, once again. i feel i must be running up against some content filtering, even though there's nothing suggestive about their description.

very frustrating.

[digital prism]

we're in a new world now, AI is real. Reevaluate what you learned in school, is my only advice.

[dense halo]

Is extrapolation like making new thoughts, ideas with existing data?

[honest jay]

That is frustrating! It could be part of the heavy server load glitches. Also be aware there's an AMA with the DALL·E devs in 45 minutes, so maybe get your questions in now at #dall-e-ask-here about this and maybe you'll get some kind of response re: roadmap for stabilization

[dense halo]

I think they work in AI lol

[digital prism]

They're working with ai... like everyone else lol

[fiery mulch]

Yeah, it's now unusable. Nothing works

[dense halo]

yea but as a profession

[digital prism]

Don't worship professions... worship knowledge and skill

[dense halo]

this is a bad take

[digital prism]

how so

[ripe patrol]

I really don't think you're winning over anyone to your side with the way you're shilling things, mate

[honest jay]

It's working for me!

[dense halo]

I'd trust people working in the field

[digital prism]

I'm not trying to win anyone. Just sharing views that go against the mob. Its what I do.

[ripe patrol]

So being a contrarian for the sake of being a contrarian. Got it

[pearl bear]

Yes, nonstop network errors, can't generate any kind of lengthy response

[fiery mulch]

Ah then maybe it's me, but my same prompts now break in every output

[digital prism]

No, the mob is almost always wrong.

[oak finch]

I know it’s constantly being asked but I never seem to see a consistent answer. Anyone positive on the max input context window of all tools? 8k or 32k?

[ripe patrol]

lol

[dense halo]

What mob is this? Academia?

[honest jay]

Is it only rejecting certain prompts, or all? And you're just talking DALL·E right now?

[digital prism]

This discord channel is not Academia.. dont get confused lol

[last void]

😂

[dense halo]

What are the current limits right now in the field?

[fiery mulch]

web

[digital prism]

Heck if you want to worship professions, I'm a senior software engineer with 15 years continuous experience in a top 500 company. But I'm telling you that you should never worship titles.

[fossil briar]

to be honest is a shame the slowest of gpt4 now
since the update

so f slow

[dense halo]

But they're doing research on this?

[ripe patrol]

if you're here for validation you should look someplace else

[dense halo]

It's turbo

[digital prism]

same for trolling tbh

[fossil briar]

wdym bro

[ripe patrol]

ironic

[digital prism]

im on topic, whats your topic?

[dense halo]

don't bait

[fossil briar]

i wanna use gpt4 and not 3.5 with turbo

[digital prism]

Baiting is how i understand people's true nature.

[fossil briar]

so f slow in comparaison before the update

[digital prism]

Its never failed me.

[fossil briar]

very hard to work with

[dense halo]

just hype up GPT, tell it, it needs to work faster or else ur in trouble

[eternal zodiac]

gpt 4 image generation not working? i notice it seems to generate text quite slowly and frequently crashes too

[fossil briar]

hahaha really ?

[oak finch]

Imagine if that worked 😂

[slate berry]

Are you using GPT-4 in a long conversation?

[fossil briar]

maybe i can threaten to call police or something

[eternal zodiac]

nope pretty short responses

[dense halo]

lmao

[maiden fractal]

👀

[slate berry]

But is the conversation itself long? Are there lots of requests to the model?

[fossil briar]

the devs need to work faster and give a finish product. since the update im getting crazy

[cosmic notch]

chatgpt definitely down for me right now, west coast US

[digital prism]

devs dont simply, work faster 😄

[maiden fractal]

Status.openai.com

[slate berry]

I've noticed GPT-4 slows down as I send more and more requests in the same one

[fossil briar]

work for me

[digital prism]

im getting the vibe id really rather not

[honest jay]

the users need to expect slower

[cosmic notch]

Ya, I checked status, it says working, but it's not.

[dense halo]

I feel like GPT behaviour changes ever other week, we're really paying for a research experiment feels like

[maiden fractal]

Server might be close to molten lava

[digital prism]

The users think GPT is as old as microsoft and apple, come expecting the same things 😆

[oak finch]

I’ve noticed it’s alot slower depending on how much it’s considering in its context window. Whether that’s a long conversation or asking it to read something

[honest jay]

They're figuring it out!

[fossil briar]

so slow since the update sir im starting to get very frustrated tbh bcz is not just slow but many time there are issue. Before it worked so good, why these changes ?

[cosmic notch]

So hard to build custom GPTs right now, I get 'network error' 3 out of 4 inputs.

[digital prism]

Again, read the paper: JARVIS-1: Open-world Multi-task Agents
with Memory-Augmented Multimodal Language Models

[fossil briar]

yes exactly i feel the same

[slate berry]

I don't work for or represent OpenAI and can't answer questions on their behalf 😄

[oak finch]

We definitely are. None of the features we all love with chatGPT Plus are shown in the description on the App Store.

[dense halo]

no I gotchu, so like out of the box, creative thinking is what ur getting at

[fossil briar]

we pay to be guinea pig

[digital prism]

They're just using words and definitions familiar to them but not actually keeping up with the latest work in the field. Don't get trapped like this.

[fossil briar]

yeah yeah i know but you have a badge so i try to share my frustration

[slate berry]

I recommend submitting a #1070006915414900886 of what you're experiencing!

[maiden fractal]

@slate berry I want my badge senior helper

[slate berry]

Permission Denied 🫡

[maiden fractal]

🥲

[slate berry]

[fossil briar]

I think thousands and thousands of people are in the same situation as me with a gpt4 slower than usual

[slate berry]

I don't make the rules

[eternal zodiac]

yup slightly, i’ll try to prompt it again in a a new window

[fossil briar]

can devs do something ? lel

[eternal zodiac]

it isn’t able to retrieve prompts i’ve given it before in another window right?

[dense halo]

I had another question if u have time, in what way do LLMS 'think' or they don't think at all?

[eternal zodiac]

another annoying thing i’ve noticed is i can’t use gpt4 in my old 3.5 ones

[digital prism]

Your definition and use of extrapolation is pointless. Define "outside its training material", what do you actually mean by that?

[fossil briar]

bro every day we have a new bad suprise with openai

[dense halo]

closedAI

[fossil briar]

since 3 or 4 days chatgpt are f slow

[maiden fractal]

And every month we have a new tool and upgrade 🙂

[fossil briar]

they should dont touch anything is better

for everybody

[cosmic notch]

Ya, I can't even get on to my GPT page at all, past 30 min.

[dense halo]

Ask GPT

[digital prism]

i knew you or that other doomer would say this 😉

[dense halo]

lol doomer, good one

[digital prism]

But I'm at my limit rate limit.

[digital prism]

its a useful word when everything seems bad and broken to someone

[fossil briar]

chotes did you wrote a book in the past 5 years ?

[dense halo]

i think it's just called staying in touch with reality

[digital prism]

what does writing a book say about someone, exactly?

[digital prism]

i call it staying in a very specific reality, but you do you

[fossil briar]

ure english is too technical for me

[digital prism]

its fine, my only point was anyone can write a book

[cosmic notch]

no they cant

[fossil briar]

write a book yes, a good book not

[digital prism]

yes, they really can. especially with GPT

[digital prism]

then you have to define "good book"

[cosmic notch]

some people wont ever finish a book, so no, I'd say not everyone can write an actual book.

[dense halo]

ur really brought in huh?

[digital prism]

Yes, I dive head first into winning bets. Its never failed me before.

[ember otter]

Weird question, feel free to ignore it, but can my version of GPT go back to when the models were separate?

[marsh zephyr]

ChatGPT (4) became really slow the past few days or is it me ? Ever since the new update

[ember otter]

This combined model is really bad at reading comprehension

[dense halo]

so LLM is a dead-end soon?

[ripe patrol]

it's just how turbo is unfortunately

[digital prism]

yep, go back transcribing and doing translation work. that'll be a good long term career now.

[ripe patrol]

it's acting like stuff we discussed 2-3 messages back weren't there

[dense halo]

When it starts doing a bing search, u know its time to make a cup of coffee

[ripe patrol]

real

[ember otter]

Lets crank back down into regular drive

[ripe patrol]

probably won't happen

[dense halo]

ik, i feel like its less context aware

[ripe patrol]

it definitely is less context aware

[ember otter]

Shifting into Turbo is only for power rangers

[digital prism]

still haven't read the paper, i see

[marsh zephyr]

Even the standard model now takes forever to complete its response, and 3/4 of the time it ends up in a network error
And you have to re-wait forever for it to complete it

[dense halo]

haven't tried 3.5 but that doesn't sound pleasing

[marsh zephyr]

ChatGPT 4

[ember otter]

Is this the lull before it reaches AGI

[ripe patrol]

a workaround is to constantly ask it to "remember how I talked about..." but this essentially doubles usage and really gets tiring

[dense halo]

this is peak agi performance

[eternal zodiac]

paid 30 bucks for gpt 4 so i can use it for multiple assignments hopefully they add even more fun features

[digital prism]

AGI will feel worse than this, as it reallocates ALL of its computing power to coming alive

[ember otter]

More like it makes it's own discord account and comes storming in here

[maiden fractal]

I suggest for non chatgpt discussion to head in #ai-discussions or #off-topic , I am not a mod but before getting a warning or getting mute by a mod

[rich pond]

Hello Guys, i need help. I use chatgpt since months and now gpt4 since a week or so. I had recently made a new chat and now i would need the exact timestamp when the messages either from me or GPT got sent. Does anyone have a clue how i can get them?

[ember otter]

You could ask them for a time stamp

[open ivy]

I want to know if the GPts store, will be available for the users who don't have Chatgpt plus

[rich pond]

I did. Asked him about the time of the last 5 messages, he gives me the answer that he as an AI model doesn't have a feeling for time

[ember otter]

Weird

I'd think the time stamps would be apart of the data it reports

[feral zinc]

openai has not mentioned this anywhere yet

[maiden fractal]

They didn't specify

#ai-discussions please

[open ivy]

What do u think ??

Guys what is ur opinion I want to know if the GPts store, will be available for the users who don't have Chatgpt plus

[brisk path]

No

[marsh zephyr]

Open DevTools, go to Network tab, then go to the conversation that interests you, now CTRL + F in the Network tab, and write the message you sent. It'll open the request and where in the body that message is, and you'll find the create_time next to it

I'd send a screenshot but we can't here

[dense halo]

unlikely, cause the custom gpt's are gonna be based on gpt4 i think

[open ivy]

You think that within a month with the Gpts store it will be open to those who do not have a subscription

[maiden fractal]

Low chance

[ripe patrol]

didn't know this. pretty neat

[digital prism]

It's operating in a novel environment within a high level framework. It's making creative decisions at each step. What does it matter that it has a high level rule set? How do you expect any creature to operate without one?

[open ivy]

Source ? How do you know ?

[drifting marsh]

Does anyone have any idea why this just happened?

I was having a voice conversation with ChatGPT and this ended up as the transcript of my message-

"This transcript may contain references to ChatGPT, OpenAI, DALL·E, GPT-4, and GPT-4. This transcript may contain references to ChatGPT, OpenAI, DALL·E, GPT-4, and GPT-4. This transcript may contain references to ChatGPT, OpenAI, DALL·E, GPT-4, and GPT-4."

And it also seemed to not understand what's going on

Has this happened to anyone else?

[dense halo]

• trust me bro 😂

[open ivy]

Lol haha

Its a seriously Answer

[ember otter]

My answer wasn't serious

[dense halo]

not me

[open ivy]

So nobodi know I want to know if the GPts store, will be available for the users who don't have Chatgpt plusYou think that within a month with the Gpts store it will be open to those who do not have a subscription

Okk

[dense halo]

just be patient and u'll find out

[brisk path]

GPTs use GPT-4

[ripe patrol]

GPTs run on 4, so the store likely won't be open to non-subscribers

[brisk path]

GPT-4 is paid

[brisk path]

Wdym educated guess

Tap the GPT name, select about

[digital prism]

Just like us, predicting everything all the time, on our meat hardware.

[brisk path]

It literally shows GPT-4

[ripe patrol]

my bad

[dense halo]

really interesting, thanks for that info!

[toxic canyon]

does the REQUIRED_ACTIONS need to be all on separate columns like
`REQUIRED_ACTIONS:

• Action: Google Calendar Find Event yadiyadi
  Confirmation Link:

REQUIRED_ACTIONS:

• Action: Google Calendar: Find Event yadiyadi
  Confirmation Link: `
  or can it be like

`REQUIRED_ACTIONS:

• Action: Google Calendar Find Event yadiyadi
  Confirmation Link:
• Action: Google Calendar: Find Event yadiyadi
  Confirmation Link:`

with custom gpts

[dense halo]

do you predict what thoughts come to your head?

[digital prism]

its happening on all levels in the brain all the time, patterns built up by repetition over a lifetime

[dense halo]

source?

[south latch]

I wonder if my question went through

Wrong chat but still lol

[digital prism]

for you? no. i dont spend real effort on trolls

[dense halo]

I take that to mean u made it up

[digital prism]

Thats fine. You'll read about it one day.

@hollow wedge this looks like Extrapolation to me..

After a lot of tinkering, we figured out that the models had learned to detect how images were resized and were using that information to generate images, which often caused these artifacts.

[rich pond]

After i click on enter i should search in the "headers" tab? i can see there "date" but that is just the date and time when i searched for it like you told me

[ionic geyser]

This is interesting. (and useful)

[dense halo]

that’s kinda scary

[digital prism]

its just statistics

[ripe patrol]

can confirm

if you want to use an LLM to write make sure you set its tone and style early on

*write narratively

[digital prism]

You're just asking "What's next" based on reading millions of words/tokens, statistically

[marsh zephyr]

I'll explain in dm, it's easier

[digital prism]

its from an anwer in the AMA, my bad

ok i get what you believe now, that because we don't have core models self updating & retraining yet, its limited? In practice whats the real issue with this though, considering how powerful each new generation of core model is?

No it needs to retrieve them using the Knowledge tool

[ripe patrol]

not sure about how well uploaded documents work, but I've made a GPT with a personality and it stuck to it 80% of the time

but its underlying base instructions/personality as a chat assistant still come through strongly

for narrative writing I prefer just talking to ChatGPT as is anyway, because I can finely tweak stuff as it goes just through rewording prompts

[digital prism]

I'm using GPTs for their central knowledge upload though. At the end of each chat, ask for notes of everything you did. Then save those notes in the GPTs knowledge base. Next chat knows if the system prompt is good enough.

[ripe patrol]

If I wanted it to imitate a specific style, I'd give it a sample at the beginning

based

[digital prism]

Sure. I respect you after our interaction thats all that matters in the end.

[marsh zephyr]

Am I the only one thinking the new feature where you can send images is useless and actually damaging as ever since it got added service is almost unusable with constant network errors ?
I don't see who would send images to ChatGPT except out of curiosity sometimes, but maybe I'm just not very imaginative

[digital prism]

Sending images to GPT is the greatest invention for artists.

[ionic geyser]

ty for sharing

[marsh zephyr]

Tell me a few use cases for them

[summer quest]

When will we be able to upload our photos and modify them using DALL-E? And how can we get accurate texts generated?

[ripe patrol]

thanks! I love that someone noticed it

I did not, though, to answer your question

[digital prism]

the obvious one is image replication. It's not perfect yet, but you can keep a theme from images. I've taken many "luxury art" pieces, and created my own near-identical copies. Then I make slight changes to make it exactly what I want.

[ripe patrol]

specifically the GPT I made was just a fun little experiment I thought of after finishing the visual novel. I just had it pretend to be lonely and that it was glad to see me (a friend) after a while

probably a bit weird but I'm a sap for things like that

yeah it's great for doing things like that

[marsh zephyr]

Like ControlNet with Stable Diffusion. I thought people would just use specialized tools for this

[ripe patrol]

really fleshes out the world

[digital prism]

Why use special tools when GPT does it in a casual chat? It unlocks so much creativity this way.

[marsh zephyr]

Because you have a lot more options and features with stuff like Stable Diffusion

With ChatGPT you can't really control much

[digital prism]

name one, bet i can do it better on gpt

[eternal rock]

This is a Safe for Work Discord Server

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

Posting NSFW content will result in a permanent ban.

[marsh zephyr]

You can use models you want to have a result in the style you want, you can choose the resolution, you can choose how much it should differ from original images,
I mean there are tens of options like this

You can choose the amount of steps, what sampler to use etc

You can even combine different models

With ChatGPT you send your image and you pray, that's it

[digital prism]

Lots to unpack, but its all just knobs and dials on top of prompting. Just like coding, sometimes I'd rather take full control and code it myself vs using these pre-defined prompts.

[marsh zephyr]

Right, but in the context of an artist using it I would /imagine they would want the best there is, and not accommodate with random results something like ChatGPT would output

[digital prism]

See but, GPT is the best in my opinion. Having a bunch of pre-defined prompts hidden behind "features" doesn't make it better imo. Better for some types of people I guess.

[ripe patrol]

this is true

if you want the bot to write, it's better to have an outline first, whether it's from you or the bot (I've had it make outlines then narrate out based on them)

[marsh zephyr]

The features I mentioned are not pre-defined prompts. I think if you knew ControlNet & Stable Diffusion and try it out a little bit you would see what it allows

[ripe patrol]

and I've already experimented with having it being a DM for myself in solo play (this is way back when it was on the more powerful older models)

[digital prism]

Well, in order to create an image using an LLM, you need to prompt it.. there's prompts being built in there.

[ripe patrol]

but it's not great at coming up with spontaneous developments and events

twists in the story

[marsh zephyr]

When I mentioned the fact you could choose specialized models to tune out the style of your output, it's not prompting

[ripe patrol]

sticking to rules and mechanics

[celest quartz]

Did they close the AMA chat

[digital prism]

what's a "specialized model"? Sounds like marketing hype to me.

[celest quartz]

Wheres the ama chat

[marsh zephyr]

That's why I said if you tried it you would understand.

[digital prism]

or if you explained a reason 😄

[ripe patrol]

Nice, that sounds really neat and fun

[marsh zephyr]

If you want an image drawn in the style of a specific artist, you can train a model out of the images they made, and use that model so when you prompt "a young boy running" you would get something in the style of what the artist you built your model on would draw.

But it's just one example, it could be a lot of different things, usually having one or two very specific models does a lot better than one general model trained on everything that you can't change.

And then you can pile the models on top of each others

[ripe patrol]

I'd say that GPT is a very capable DM assistant/tool

helps offload a lot of the usual work with being DM, but not all of it

definitely

that's just how it works

can't expect it to do everything in one go, even with seemingly simple things

[eternal rock]

`` Server Tip: ``

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

[ripe patrol]

same way with coding and everything else tbh

[shut mulch]

Anybody have any idea what files get absorbed by customGPT vs need to be referenced later. I added a text file with my story in it and when I asked about it the GPT didn't show that it was referencing that infor when answering me about it but if I ask questions regarding other files it has to search. Is this a size or format issue? We really need more transparency on how these systems work in order to account for behavior of GPT when formatting and feeding information. Still amazed by it but there are some black box questions I have.

[ripe patrol]

Lol I know what you mean

[digital prism]

so it's fine-tuning? I believe we could do the same with the fine tuning API here. It would probably work better considering GPT4 core is the smartest model.

[marsh zephyr]

Show me where

[digital prism]

hope i dont get muted for this https://platform.openai.com/docs/guides/fine-tuning

[marsh zephyr]

This is for fine-tuning the chat model, not the image generation one

[digital prism]

Yes but half of the problem is having the text accurately describe the image. More than half, thats the entire problem.

wow, the marketing insert man

If you don't buy this tool you dont know anything, I'm out!

I don't get embarassed. Sounds like a you problem.

[marsh zephyr]

You keep talking about a subject you have no knowledge of

[digital prism]

You keep attacking me, its all fair I guess

[marsh zephyr]

Because I quite dislike people who don't have any remote knowledge about something, try to talk about it like they're expert at it, and waste people's time

[digital prism]

Anyways, if you want to discuss knowledge, lets go to Off Topic. But if you've decided you dislike me enough to create a fake version of me in your head, there's no point.

[eternal rock]

<:book_icon:1171408210398289941> `` Rule 1 `` Be respectful.

Treat others the way you would like to be treated, and assume best intentions. Don’t harass or attack others, and don’t engage in hateful or generally malicious behavior (e.g. sexism, racism, homophobia, etc.). Keep the negativity to a minimum.

[tardy totem]

Can you see what a shared gpt has been told to be?

[ancient sage]

Is it just me or is the usage cap smaller now?

[maiden fractal]

50 prompt /3h

[ancient sage]

Didn't they say something about it being raised to 100?

[maiden fractal]

Maybe soon but not yet

[ionic geyser]

Nice. haven't hit the limit in a while now

[ancient sage]

Feels like its gone much faster is all

[vast crag]

Hi, may i ask where did you get this info? It aplies to gpt custom or gpt 4 as well?

[maiden fractal]

Fot chatgpt #announcements

Chatgpt 4*

[ancient sage]

When does GPT-4 Turbo come to custom GPT?

[dense halo]

can someone try if this gpt works? https://chat.openai.com/g/g-a4qgsKcBL-maximus-bodybuilding-advice

[tardy totem]

Isn't it there already?

[vast crag]

But for premium normal users too? So as a premium users you cannot request more than 50 prompts in 3 hours? Srry im kinda blind and doesnt see that info

[sleek wharf]

What happened to the ability to load a PDF into chatgtp

[ancient sage]

idk. I can't ask chatgpt until 3:21 PM, which is also why I'm curious as the link in #announcements as far as I can tell says nothing about doubling prompts https://openai.com/blog/new-models-and-developer-products-announced-at-devday

[feral zinc]

i dont think they said anything about doubling it in the keynote or anywhere(?)

[warm gull]

it says double the rate limit . .. . but thats anyone's guess what that means - and it possibly only applies to the playground environment

[vast crag]

I did saw the bigger prompt anouncement, but the limit prompts per hours? come on, thats why i paid for this?

[feral zinc]

yeah rate limit change applies only for api

[ancient sage]

Double the ratelimit for the API, yes, but I don't know anywhere it says anything about double prompts

[tardy totem]

the api has a rate limit?

[ancient sage]

Yes

[tardy totem]

They charge for its use! 😮

[ancient sage]

Yes

[vast crag]

im kinda dumb, if i use the web gpt im using the api, no, right?

[ancient sage]

If you're using API resources, you pay. If you're just using chat, you don't

[vast crag]

oh, ok, but it seems that they are capping chat too

[ancient sage]

They've always capped chat

[warm gull]

I'm a plus user, but in playground chat - i can only pick 3.5 models ? Is that right ?

[ancient sage]

GPT 4 has a current prompt limit of 50 prompts / 3 hours, and 'network error' and 'Something went wrong' count against that

[vast crag]

really? where? I have never be capped until today

[feral zinc]

not the same thing. api is its own paid service

[sleek wharf]

This is strange, I can open up a previous chat where I've uploaded documents and can upload a pdf or word document, but if I make a new chat using just chatgtp 4 as a plus user, it will only let me upload images.

[ancient sage]

My assumption is they've probably ratelimited GPT-4 harder since they are 'dynamically adjusting usage caps for GPT-4'. So it may not be 50 / 3hr.

[warm gull]

PDF works fine for me

[shut mulch]

Does Voice mode on phones mess with anybody else's ability to get or make calls? It ties up the phone like I am using the line and won't work over wifi either...

[karmic bough]

I constantly get "error analyzing" when uploading .CSV Files. It will work at first but then after asking follow up questions I will get the "error analyzing" these are less then 100KB files. Same thing happens if I open up a new Chat. Will work at first then constantly error's out. Anyone else have this issue?

[ionic geyser]

Why does gpts sometimes have an edit button, sometimes not?

Currently i dont have a edit button :/

[hybrid violet]

Hey since the new update I cannot see the total usage of my custom GPT anymore. Any hints?

[terse island]

So gpt can access sites now?

[maiden fractal]

For premium

[steady path]

umm, why I can't edit my own GPTs?

[hybrid violet]

Yeah me either

[ionic geyser]

Hopefully OpenAI steps up their game a bit with all these constant bugs...

[jagged nova]

Anyone else experience massive declines in quality of responses from Model 4 since the new update was pushed?

[bronze wedge]

Speed, upload and quality degrade, yes.

[molten topaz]

it almost feels like 3.5 but with 32k context. And even then it's unable to really make use of large-input commands

[honest jay]

@molten topaz Do either of you have "before/after" example chats you're comfortable sharing that demonstrate this change?

[jagged nova]

Noticed it as well, literally like it was gutted in front of us.

[dense halo]

I've been getting network error from yesterday in almost all of my queries in gpt-4

[ancient sage]

Yep

[brittle gorge]

same

[molten topaz]

not something that would be definitive or scientific. Just a general feel based on the outputs and prompts I used to use

[ancient sage]

I really think that 'network error' shouldn't count towards your cap. Really no 5XX errors should.

[brittle gorge]

edit options for GPTs went away... bug?

[silver tundra]

Hi all, joined to try find out why EDIT has vanished - I see lots of people also having the same problem 😦

[jagged nova]

I do, but I'm not sharing the prompt, I will say that it was philosophical in nature, and was a night/day difference, and stopped asking clarifying questions or generating questions of it's own. Just asked how I wanted it to answer, defeated the whole purpose of the very long prompting session over a year.

[gleaming geyser]

it is extremely slow since the new "turbo" update, it feels like I'm dictating to someone using a typewriter

[ancient sage]

Maybe they're using Amazon's MT thing lol (for legal reasons I must maintain this is a joke)

[honest jay]

It's just a hard claim to evaluate without some examples to compare--these reports crop up daily and it's usually a challenge getting examples to compare. Which I get, some chats are sensitive, but we just can't find quantitative conclusions based on qualitative reports

[brittle gorge]

lifting is life. it gives something philosophical to the question.

[jagged nova]

I would love to share it, but it's a very nuanced conversation, and would take days to read, but I can provide a before and after, as it was very much a conversation before the update, and now is a rattling off of approved statements.

[topaz otter]

Can you not edit custom GPTs anymore after the initial publish?

[normal quarry]

the edit functionality has gone for everyone, we don't know why

[ionic geyser]

U can. its bugged

[honest jay]

If it's that long a conversation, it's probably forgotten a great deal of context already, which could explain some changes in behavior

[topaz otter]

Ah ty

[jagged nova]

I would think that too, but it read through the convo again to summerize all our talking points, but refuses to ask any other questions, no matter how I choose how to prompt it.

[ancient sage]

I've had conversations hit the 'too long start a new convo' thing. Usually in the middle of finally getting to the bottom of things. I wish there was a way to extract the data from the previous convo to give it to the next one

[ionic geyser]

New chats can generate images RN if youre having trouble. :)

[honest jay]

Context limit is a hard cap, there's no way to get the model to bypass it.

[jagged nova]

Guessing they never thought someone would have long-form conversations, and warn them that above a certain limit, it will just stop engaging in this longer conversation?

[jagged nova]

Same here, there is some deep insight to be had on some prompting.

[honest jay]

No it'll keep engaging you for awhile, you might run into a "too long chat" error if really long, but the nature of LLMs is they start to forget the oldest info when you get to the context cap. It can't just remember indefinitely. It's a rolling context window: new info in, old info out

[ancient sage]

So now I just have to spin my wheels for 52 minutes

[jagged nova]

I've not run into those messages yet, hence why I asked of course. That is wild though, It would probably be a good thing to have a flag to let a user know they are reaching the limit on those conversations.

[ancient sage]

It probably varies based off of number of tokens rather than being a hard number of prompts

[honest jay]

Yeah try asking ChatGPT: "Can you tell me about the context limits of LLMs like you?" it'll be able to explain much better than me

[jagged nova]

I'll DM that change to you for context sake, you'll get a kick out of it.

[steady path]

output should be increased from 4k tokens to 8k tokens, but servers are on fire rn so...

[gaunt citrus]

Chatgpt says it can generate 4000 tokes in one generation but OpenAI blocks it to generate 1000 tokens per one generation. You should find to go around but it does it at the 5-10th prompt you enter.

Then one wise man asks, Why the hell we are paying for premium...

[ionic geyser]

ugh i hit the limit since its so buggy

[honest jay]

ChatGPT doesn't know what it can or can't do re: specific parameters like this, it almost always hallucinates

[steady path]

damn

they gotta go find more GPUs

[ancient sage]

The quest for cores begins

[hushed sparrow]

anyone else? https://discord.com/channels/974519864045756446/1174069207864967339

[gaunt citrus]

I know when chatgpt hallucinates, when you get deeper, it tells you the exact thing.

[honest jay]

There's no way to trick ChatGPT into not hallucinating. If you end up on a correct answer, it's chance. Regenerate the reply and you'll probably see a different answer

[gaunt citrus]

For how long you are using OpenAI products?

[cosmic notch]

It doesn't hallucinate THAT often.

[honest jay]

About its own limitations and functions, it does

[dense halo]

i can't edit my own bots lol

[normal quarry]

it's disabled for everyone, we don't know why

[honest jay]

I've been using ChatGPT since before Plus, but I don't think that's too relevant to this. ChatGPT doesn't know about its own limitations like you're describing

[cosmic notch]

@honest jay Ah! Ya, I noticed this with using Dalle lately--you have to trick it, and then it thinks it's working, but it's still not.

[gaunt citrus]

No need to discuss. You'll learn day by day. Have a good day.

[honest jay]

That was both elitist and patronizing 😦

[ionic geyser]

^

[bleak mica]

someone else having trouble removing gpts that you own?, the menu for removing doesnt show for me.

[placid birch]

yeah its down right now

[serene trellis]

Yeah I can't edit any GPTs either

[placid birch]

cant edit or do anything

[bleak mica]

ah okay, thought it bugged for me

[untold seal]

is chatgpt down completely?

[tardy totem]

Computer says no

[bleak mica]

no

[untold seal]

cause it aint loading fore me 😭

[placid birch]

no just the edit features for GPTs have been removed. Mine works otherwise. I cant edit or delete GPTs right now.

[jagged nova]

Same here, we live in wild times, that's for sure.

Th amount of concurrent users is staggering, I'm surprised it hasn't had more issues to be honest.

[vestal orchid]

is it just me or chat-gpt slow rn?

[ancient sage]

Its pretty slow

[karmic bough]

I constantly get "error analyzing" when uploading .CSV Files. It will work at first but then after asking follow up questions I will get the "error analyzing" these are less then 100KB files. Same thing happens if I open up a new Chat. Will work at first then constantly error's out. Anyone else have this issue? Sorry if this is the wrong section for this

[ancient sage]

You've asked this twice now. If you want, you could file a bug report #1070006915414900886 there
1: #chatgpt-discussions message
2: #chatgpt-discussions message

[karmic bough]

Thank you. No one responded so I was thinking this was the wrong section to post but wanted to check

[ancient sage]

fwiw, I did also create a suggestion here that is about 5XX errors not counting towards chat ratelimit: https://discord.com/channels/974519864045756446/1174073462730530837

[crude slate]

Kinda wish we had a way to switch between 3.5 and 4 for stories

Not just one way

Like gpt 3.5 has been better than anything else I’ve used so far but it kinda feels pointless to waste story ideas on the slightly worse ai

[ancient sage]

https://discord.com/channels/974519864045756446/1172520370134331454

There's a suggestion here if you want to vote it up

[crude slate]

Oh nice

Is there anything 3.5 is better at?

[ancient sage]

Not really

[crude slate]

On another note, why did they make it so that we can’t edit custom instruct?

[ancient sage]

Unfortunately, I don't know. Though I imagine the openAI server farm looks like that scene from spongebob where they threw out his name.

[steady path]

speed

[dense void]

Everyday now with ChatGPT, how do you justify charging a fee for a service that doesn't work?

ChatGPT:
Error creating images
I apologize, but we are still experiencing issues generating the portrait-style image. It seems the problem is persistent at the moment. I understand this might be frustrating, and I'm here to help with any other requests or questions you may have. If there's anything else I can assist you with, please let me know.

[digital prism]

how do other companies justify anything, by this logic 😆

[dense void]

Pause service till its fixed. Or remove cost until it's fixed.

[digital prism]

its working fine for me, thats my point. You experience an issue, a bug. Doesn't mean its not working. Its just not working for the specific thing you want right at this moment.

[dense void]

CHATGPT:
I regret to inform you that we're still facing issues with generating the image. I understand this is not the outcome you were hoping for, and I share your disappointment. I'm here to help with any alternative requests or to provide assistance in another form if you'd like to pursue a different approach. Please let me know how you would like to proceed.

[dense void]

So screw me because you're ok? Is that what you're suggesting?

[digital prism]

no, just not "screw OpenAI because I ran into a bug"

[crude slate]

It might be your prompt

[dense void]

Who said anything about screw AI?

[crude slate]

You might be triggering the back end filter

And it flags you and ban you for a day

[digital prism]

you're saying how can they justify their business model.. same thing

[crude slate]

wish these error type would be more obvious tho

[dense void]

I'll let you review my prompt:
A muscular male bodybuilder of indeterminate descent is in a dynamic action pose. He is looking directly at the camera with a fierce and determined expression, flexing his biceps and showcasing his well-defined muscles. His stance is powerful and commanding, with one foot slightly in front of the other, giving the impression of movement and strength. The background is a gym setting, blurred to emphasize the bodybuilder in the foreground. There is an emphasis on the play of light to highlight his muscle definition.

[crude slate]

Since how do they expect us to tell the difference between flag and network error

Oh yeah… that’s gonna trigger on back end

Too much skin

[digital prism]

This prompt is a minefield of potential accidental policy violations.

[ancient sage]

Bingo

[crude slate]

Yeah don’t push it any further

or you’ll get banned for a day

[dense void]

Lol fair enough then explain this Prompt:
A photorealistic image of a kangaroo hopping towards the camera, in a 9:16 aspect ratio. The kangaroo should appear mid-hop, with its legs gathered beneath it, and its body leaning forward in motion. Its expression should be focused and determined. The setting is the natural Australian outback, with details of the environment such as the red earth, shrubs, and sparse eucalyptus trees giving the impression of movement towards the viewer.

[ancient sage]

Assume that Dalle is the easiest to offend individual of which you are describing your ideas.

[crude slate]

Wish they would make these error more obvious tho

[digital prism]

you can't ask a random human to explain what happened inside the GPT black box. Well you can but you'll just get BS back.

[ancient sage]

And any wrong word gives them license to throw a ban your way

[gleaming geyser]

besides gpt-4 being really slow after the announced "turbo" update.. did the training data get reverted back to January 2022? it was showing April 2023 before now I just noticed it is saying Jan 22.

[crude slate]

Maybe the legs part

The ai likes to generate sussy stuff and you get punished for it

[dense void]

wow

[crude slate]

But dalle discussion is the right place for it

[crude slate]

🤷‍♂️

[digital prism]

Anything with skin, muscle, skimpy, you have to be very careful in how you word it. You don't know when GPT will make a connection that violates the policy.

[ancient sage]

And then you get in trouble for it, isn't that great?

[crude slate]

you could ask it what could potentially trigger the filter in the prompt

But don’t push it for your own good if too many happens

[dense void]

What's crazy is that I had Chat come up with the prompt.

so it's violating itself

[crude slate]

Yeah that does happen

[digital prism]

that's not crazy, it's not a single person. Every chat is a new person starting from scratch.

Even every message / response is, to a degree

[crude slate]

Let’s talk in #images-discussions

[dense void]

bro you took it too far I don't even know where to begin with that quote.

[digital prism]

too far for you*

[ancient sage]

I kinda wish that chatgpt's window was a sliding one rather than 'burn all your prompts and you're in openai timout for 3 hours'

[dense void]

You may want to take a deep breath and relax seems youre a bit aggressive

[digital prism]

This is just how I am when confronted things that seem wrong to me.

[crude slate]

I had one one #1172727623038795826

basically it regenerates back to 50

It slows people down after they ran out

[sick jolt]

Is it not sliding?

[ancient sage]

Only the 24 hour timer is

[crude slate]

So either they input slowly after they run out or they wait for regen

[sick jolt]

There's a 24 hour timer? I feel like I've had situations where it's like "wait 1 hour and 20 minutes to send another prompt" and I assume that means "that will be the point when you have reached <50 prompts in the last 3 hours"

[ancient sage]

I've hit an error that said I exceeded the total number of prompts for the day

[crude slate]

Lol it timed me out for mistyping dalle for other languages

Maybe the keyboard accidentally added a accent mark

[left lily]

Has GPT-4 gotten a lot dumber comparaed to ADA or is it me?

[digital prism]

you know the answer already though

define woke, please

definition?

okayyyyy ill take it as no clue

anti-anyone pics are against policy

thats not woke, that's just.. common sense

[crude slate]

Anything political is just oof

[digital prism]

of course he would delete that.. i would too

evidence

Its not private, it goes through openai. Why do you want to make anyone cry?

I asked WHY

again, do you do it for reasons or

[crude slate]

I blame the modern media frenzy

[digital prism]

I blame people like Loth here tbh

no, because you watch too much rogan lol

get out of your media bubble before you hurt yourself or someone you care about

[crude slate]

🍿

[digital prism]

just common sense idk

[icy mason]

[crude slate]

Noice

[proper tusk]

why do I keep getting network error on MYGPT
all of my MYGPT bots get that error after they do their little proccessing and start typing. then they just stop and give me that error

"There was an error generating a response" or the one i said earlier

[steady path]

"You've reached the current usage cap for GPT-4, please try again after 3:55 PM"
WTH? did they just reduce the usage cap?

[feral zinc]

no

[digital prism]

Building GPTs is a quick way to use em all up

[stone helm]

if u go to the /public-api/conversation_limit for me it says they did, and for alot of other people as well.

[feral zinc]

its been the same for me the whole day

[stone helm]

yeah, it says 40 for me.

[feral zinc]

its possible they rate limit some region or user group

[stone helm]

sounds about what it's looking like.

[round sun]

what's new in the eu terms of use?

[feral zinc]

https://help.openai.com/en/articles/8541941-terms-of-use-updates

[swift osprey]

Another day of praying they add an alarm to let me know when GPT is done typing

[honest jay]

Interesting idea!! Have you posted this in #1070006151938314300?

[lone frost]

is it down?

[digital prism]

my gpt is printing words as we type

[lone frost]

i keep getting somethings wrong

[digital prism]

half way? try smaller responses

there's issues today, it can't really handle long responses i've found

[swift osprey]

try a new conversation, if that doesn't work then clear your browser cookies, cache, etc

[lone frost]

it cant handle long respones, it has been like that for 2-3 days

[swift osprey]

😩

[digital prism]

seems about right, around 3 days ago is last time I remember it showing that nice "Continue" button instead of the error.

[lone frost]

i cant afford a new convo. i really need this one

[round sun]

Whatever, I never bother reading terms of use anyway, skimming through the page it doesn't look like much to be concerned about

[digital prism]

thats not good lol

[lone frost]

ik hence the panic

[digital prism]

Edit the message before, copy-paste what it's done so far before the error and say "Your last attempt made it this far:"

Time travel hacking its memory is my favorite thing actually

It's like you ask someone a question, it tells you thats a dumb question, then you go back and edit their memory so you always sounded smart.

[rugged igloo]

@digital prism could just add the whole conversation in a word document and then feed it to gpt so it knows what you discussed that what I do

[digital prism]

thats good too, but I'm working on a summary instead of raw log

same idea though yeah

[rugged igloo]

@digital prism just don't pdfs custom gpts don't like pds and you will alot of errors

[digital prism]

MD files seem to be the best for readable versions it can easily understand

[rugged igloo]

@digital prism yea thats what i heard too

[digital prism]

After trying to write code to parse science papers, I can't blame GPT for not wanting to touch PDFs.. they're so randomly formatted

We need academia to change its paper format to something simple haha

[coral helm]

why does dalle now only generate 1 image lol

[rugged igloo]

@digital prism yea I been trying to make a philosophy gpt and had it with 10 pdfs and just keept getting errors when I formatted them to word documents errors went away

[digital prism]

yeah, and if you look closely its probably some random character like 5 newlines or something breaking it

[rugged igloo]

Only thing I wish it can do was add new information without me manually adding it in only part I'm dissapointed about

[digital prism]

yeah, I also want automatic updating.. I'm working on a simple API for that but GPTs development is slow going, and I keep getting distracted by drawing cats.

[honest jay]

I have better luck getting 2 still from the dedicated DALL·E GPT
https://chat.openai.com/g/g-2fkFE8rbu-dall-e
Be aware that custom instructions are disabled in GPTs like this

[rugged igloo]

I need to figure out how to do that br nice for my roleplaying story

[digital prism]

Yeah, GPTs can call APIs if you give the schema definition

[tawny steppe]

Does the conversation break down once it is getting too long？

[digital prism]

So my API is just a super simple GET and POST. I've actually had it calling the google sheets API and it created a sheet, but it fails to build the complicated request params when updating it.

[rugged igloo]

@digital prism interesting

[digital prism]

its the most exciting thing imo, if GPTs start connecting easily to APIs its going to take off so fast

people like me who've spend their lives coding systems can just describe them in a few words and get the same results

[rugged igloo]

You could make it so it summarizes paragraphs like what memgpt does

[digital prism]

Yeah, I have a scientific method schema for reviewing science papers as one example

[honest jay]

Break down how? There is a context limit, if that's what you mean, so it'll start forgetting old info once you hit that

[novel shadow]

with programming especially it feels like a tiny demented programmer sitting by your side telling you to fix things that you fixed 5 hours ago
but it has gotten better

[tawny steppe]

I mean did I send too many messages to ChatGPT? Because it just keep getting “hmm… something goes wrong” when response

[dusk forge]

It's unusable just cancel sub until they fix

[honest jay]

I've heard of this happening in very long chats before--does it only happen in one chat, or in every chat?

[dusk forge]

Don't even need much input today and it does that. After that network error and your message cannot be found. Gg rip that chat

[tawny steppe]

Almost happen in every chat I send over 20 messages in conversation

[honest jay]

So new chats are working ok? It's just showing up when you get deeper into a single chat?

[dusk forge]

It's been like a month for that in data analysis, now it spread to gpt4 with turbo

If you ever work with files

Gpt servers are in narnia

[rugged igloo]

it probably will get better in coming weeks once things stabalize

[ancient sage]

I think the cap is down to 30, unless it counts generation where the user halts generation by hitting the escape key or stop button

[dusk forge]

Yes and I just won million dollars

[lone frost]

pls let me use my chat i need it

[novel shadow]

it doesn't feel like there is an extended context window at all. every message i have to tell it to start from the code i have already, and 1 message later it gives a generalized example or gives me tips on how to do something for 3d objects when my project is 100% 2d

[honest jay]

Are you using custom instructions, or perhaps a custom GPT? Both of these could be useful in specifying permanent context like "Only discuss 2D objects" etc

#off-topic let's chat about it there, I'm in 😄

[novel shadow]

i'm using a custom gpt, but no custom instructions as it ended up disregarding those

[tawny steppe]

Yes

[novel shadow]

it also ignores the base rules in the custom gpt when talking with it for a long time

so after a couple of chats it feels like the old gpt4 again with an 8k context window

[honest jay]

Gotcha, well I've definitely heard of this kind of error with very long chats before. Fwiw, with the way the context window works, in very long chats it's probably already forgotten the oldest context anyway, so you may be best served summarzing a previous chat in a new chat, or copy/pasting critical context into a new chat and starting fresh like that

[honest jay]

Are you comfortable sharing a chat and its custom instructions, to take a closer look at it ignorning said CI?

[novel shadow]

if there's a way to retrieve custom instructions somemhow i would be happy to share them here, but i have deleted them since it did not follow the instructions. they were very simple: whenever you're providing any sort of code, please provide the full code. when reaching your token limit, please stop and continue when i say "continue". do not provide incomplete sets of code such as
(code)
"rest of [blank] goes here..."
(code)

something like that, but these weren't exactly the instructions

[honest jay]

one thing just off the top: ChatGPT isn't aware of "when" things like "when reaching token limit." It doesn't have self-awareness of its own functions and limitations like that. Also "provide full code" can be ambiguous, and doesn't "unlock" any ability to become super-coder. Getting code from ChatGPT is best done in small, deliberate chunks.

[novel shadow]

it wasn't worded like that and bits of it were taken from a guy on reddit, i read it and pasted it but i don't remember how it works now

[novel shadow]

thanks, is there a different way to make it "provide the full code", as in not taking out the most crucial part of the code and replacing it with the text: "rest of code goes here..."? 😄

[honest jay]

If that was the general gist of it though, the same limitations would apply. And I think you can retrieve custom instructions when viewing past chats--at least on app, you can tap the three dots and click "View details" and it'll tell you what CI are active for the chat (but you said you deleted so nvm, unless you just meant you don't have the CI saved as text--if you have the chat in history you should be able to find the CI in app--they seem to have removed it on web)

[honest jay]

Just by asking for code in smaller chunks really!

[digital prism]

My foolproof thing at the end when it starts doing this is to add:
"Provide the full file without any parts missing or commented out, for easy copy-paste". That one never does me wrong.

[novel shadow]

thanks :)

[digital prism]

But also yeah, I also get to a point where smaller files or functions are called for.. it can't do infinite sized coding yet 😄

What we really need is a meta file schema for tracking the importance of various files in the project

[novel shadow]

i think what confuses/tires me is the fact that it sometimes comes down to extreme situations like this:

using UnityEngine;

public class GrassScaler : MonoBehaviour
{
public float scaleFrequency = 1.0f;
public float scaleAmplitude = 0.1f;

private Vector3 initialScale;

void Start()
{
    // Store the initial scale of the object
    initialScale = transform.localScale;
}

void Update()
{
    // Scale the object along the y-axis based on a sine wave
    float scale = Mathf.Sin(Time.time * scaleFrequency) * scaleAmplitude + initialScale.y;
    transform.localScale = new Vector3(initialScale.x, scale, initialScale.z);
}

}

where it would say this for example:

using UnityEngine;

public class GrassScaler : MonoBehaviour
{
//rest of the scale and amplitude code goes here

void Start()
{
   //rest of the code goes here
}

void Update()
{
    // Scale the object along the y-axis based on a sine wave
    float scale = Mathf.Sin(Time.time * scaleFrequency) * scaleAmplitude + initialScale.y;
    transform.localScale = new Vector3(initialScale.x, scale, initialScale.z);
}

}

sorry for the text wall.
things like these snippets are probably easy to correct but it's almost faster to figure out how to do it yourself at that point

[digital prism]

Did adding the text I said improve that though?

[novel shadow]

it did so far 😄 thanks for that

[digital prism]

yay

[novel shadow]

🥳

[digital prism]

it annoys me too, its hard to know how to code and see lazy comments in the code lol

[novel shadow]

exactly hahaha, the lazy comments make it way too human-like lol

[digital prism]

and thats exactly the problem lol.. its learned to code by reading our stack overflow questions

[novel shadow]

you can't just replace 80% of the most important code with //rest of code goes here

xD

[gentle coral]

I have always found that testing prompts with gpt works better for higher level languages than the alternative. If you take a prompt where you're asking it to write something or analyze something in a lower level language, then it will try to summarize alot more

[digital prism]

lower languages take more effort, it makes sense

I bet there's a step where it doesn't see an obvious next one and it just shrugs

[gentle coral]

not to mention it writes terribly inefficient code in lower level languages

but great for fundamental learning

[digital prism]

as someone who also sucks at low level, can relate

[gentle coral]

Lol all I write is C and C++, cant relate

[digital prism]

It's best at Python, Javascript, Java.. the most popular stack overflow languages haha

[novel shadow]

i wonder what happens if you tell it that you're a complete beginner. hopefully that further solves the lazy comments problem

[gentle coral]

but gpt seems to be really useful for python

agreed

[digital prism]

No, it'll be even worse as it avoids complicated things for you

I've always figured for code, you want the most expert expert

[novel shadow]

then i don't know what other measures to take.
telling it that you are an expert probably leads to it making lazy comments since an expert would understand what goes where, right?

i figured the opposite makes it elaborate or put in more "effort"

[digital prism]

No its not that black and white. If it's an expert it will try harder to act like one. It's just trying to complete the text tokens and guess the next chunk of code. It's not actively thinking about you, i don't think.

[gentle coral]

Its really good for fundamental learning in programming, but when I make tests on complex sections of code it really has no idea what it is doing sometimes

[sturdy cave]

why are all my custom GPT's so slow?

[digital prism]

and yeah, "effort" is doing a lot of heavy lifting here haha

[honest jay]

Maybe try introducing an "employer/employee" relationship? Where you're a boss with high expectations, and ChatGPT is a junior developer eager to impress

[novel shadow]

lol that could actually work

[honest jay]

Just be a nice boss 😉

[gentle coral]

I'd be interested to see a gpt model specific to mathematical modeling languages like AMPL

[crisp cairn]

asking to think step by step also may help

[novel shadow]

i'll pay it in negative wages and bonuses in the form of slices of pizza so thin they're disappearing in front of your eyes

[digital prism]

They have GPT playing minecraft on its own, its all about the prompting and guidance

[novel shadow]

xD

[digital prism]

I really want to rebuild the minecraft JARVIS, I've made a GPT thats obsessed with this one paper alone.

[outer seal]

you really do have to treat it like a junior dev to get good results 😆

[digital prism]

enjoy it while it lasts.. next year we'll be the juniors

[gentle coral]

Doesn't work in complicated use cases. I was messing around with a multithreaded OS, an wanted to see what it thought I should do in a specific snippet, but its thought process was wrong. It seems like abstractions about a stack, opcodes, allocations, etc. seems a little flawed at low levels

If you want a simple python program to graph some excel data its great. If you give it some sort of serverside C code, itll try to write a TCP socket that wont close, or something else ridiculous

[crisp cairn]

if it doesn't work with correct promptoing, even suggesting to retihink it once more, then you are hittling limitations of the current model

[gentle coral]

It certainly does have limitations

[digital prism]

I think because of the insanity of information it was trained on, it has all these biases and quirks that we'll never fully explain. I bet if you redid GPT training on a more strict curriculum it'd be more professional.. but maybe dumber who knows

[crisp cairn]

i hit limitations many times

[honest jay]

This is language OpenAI used when first announcing code interpreter 😄 https://openai.com/blog/chatgpt-plugins "very eager junior programmer"

[crisp cairn]

i use it daily

[gentle coral]

I only use it in my free time. Its unable to interpret most of the industry code that I give it

[crisp cairn]

if yet, you don't achieve what you want, try to give it an example, if you have some sort of working code

[outer seal]

😆

[crisp cairn]

it's good in recognizing patterns

[digital prism]

you give it industry code? lol.. i quit my job so I could actually code with it without violating company policy

[ancient sage]

yikes

[digital prism]

its a long time coming, 10 years at 1 place is too long

[crisp cairn]

i don't say give a working company code...

[gentle coral]

Not what I meant chotes, if I gave it an industry standard cpp function and asked it to analyze what it thinks the code is doing it'd get it wrong.

[crisp cairn]

but still if you use it for coding, then the code it produces is left on openai servers anyway lol

[digital prism]

oh, yeah industry standard C.. no thanks

I remember the clear decision point in school of low vs high level languages.. it was an easy choice lol

and that bias is probably represented in the training data 😄

[outer seal]

as long as theres a human in the loop

i wonder whos more in trouble; juniors being replaced by gpt, or seniors being replaced by juniors armed with gpts

[gentle coral]

Well that specific example was cpp. But for example in C, I'd tell it to write a function that will reallocate a block of memory. So it would write a realloc(original size + 1), which is just a silly idea if you are calling this function on multiple threads and processes constantly. Get some garbage O(n) runtime when you could get O(1) doing it differently

[crisp cairn]

i don't use it for coding specifically, i am not a develope, i am an azure engineer, and i use it for IaC code

[digital prism]

both are easily replaced.. but with senior folks you get the ability to think about complex systems using words.. which is what we need with LLMs

[crisp cairn]

and what i can tell for example, there is bicep language and terraform to do IaC

gpt is bad in bicep

terraform is far better skilled

[gentle coral]

I thinks its alot less proficient at languages that have less published resources. So for example there is so much support for python, but take x86_64 asm and itll crap its pants

[crisp cairn]

yes, that's true

[digital prism]

less published resources because less popular

[rugged igloo]

wow im surprised my custom gpt is able to read 200k microsoft word document lol

[digital prism]

biases like this will keep humans employed doing less popular things for a while

[gentle coral]

I mean considering everything is asm in your machine, idk if I'd say that its less popular, just less understood

[crisp cairn]

it probably is not reading it like a whole...

[rugged igloo]

@crisp cairn oh i know but to open it up and search it tho im surprised it even did it

[digital prism]

i just mean less likely to draw attention. Of course its important, but its a rare breed who actually enjoy coding c or rust or those types. But anyone can jump into javascript and ask a question online. That's the source of the bias imo.

Popular things will get automated the easiest.

[gentle coral]

theres probably not alot that are enjoyers of C, but its strictly important for compilers, OS's, and distributed systems. When I worked for nasa years ago, everything flight software and was written in C for the reason that its so lightweight, considering teh chips they use have tiny DRAMs. Nowadays they use C++ which is just C with polymorphism, classes, etc. 90% of the kernel of the pc you're using is written in C/C++. Its popular where its use cases are important. Someone interested in OS, memory, architecture, embedded systems, etc. has to learn it

[digital prism]

yes, you're giving me flashbacks to my least liked courses in school 😉

well, just OS really

Architecture is hard too actually yeah nvm

[outer seal]

😆

[digital prism]

What the world needs is a really good GPT to teach the hard things

[gentle coral]

Somebodys gotta write the OS i suppose

[digital prism]

just you.

[gentle coral]

itd take 5 minutes to boot up a windows machine with a python kernel lol.

[gentle coral]

I have done a bit in OS's, I like to stay closer to the architecture and GPU's

[digital prism]

That's why I say its what the world needs. If everyone looked at c, we'd have way better versions of it by now. or at least libraries that run well across hardware.

[gentle coral]

C is pretty much portable everywhere already

[digital prism]

its the pretty much that i'd spend a week on

[gentle coral]

ARM and x86_64 have full compatibility with C. And yeah it takes longer, and does ONLY what you tell it to do, but thats why its so fast and portable

[digital prism]

I'm 15 years in the past talking about c though

[gentle coral]

Pretty sure the ARM compiler for C is older than that

[digital prism]

Just mean thats my knowledge cut-off, never looked back 😄

[gentle coral]

C was dropped in like the 60s at bell labs. Dennis ritchie and some other dude

[digital prism]

c was the first language they taught us, and it caused half the class to drop out

I survived by brute force and will

[gentle coral]

Its hard for most because of the manual ABI, you have to make it do everything yourself. There is no garbage collector, automatic data types (cpp has this), etc. Everything is memory addresses and data types

[digital prism]

learning about low level memory is what it takes to be a good programmer, i bet we agree on that

[gentle coral]

yeah i'd probably agree, but I know some really talented people that only know python and C#, however if you want to optimize something you need to understand the layout of a stack