gpt-models

[frigid burrow]

Time to redo another one 🙂

So far it seems like the biggest advantages for XML are 1. I can past code inside of <code></code> elements and 2. Attributes instead of object properties

[opal sapphire]

I haven't tested XML, all my knowledge files are JSON. The issue is with the file size, just because each knowledge file can have 2M Tokens, doesn't mean it can access all that content. I use several smaller knowledge files, each with a specific purpose

[limpid beacon]

Hi, have you experienced limitations to the length of text a GPT will create when asked to create a word document with summary of all text produced?

i'm also trying to tell it to write the summary in 1500 words or to expand the text, but it will only write 300 words no matter what

[bright plank]

You using api? Then the «max_token» parameter could affect it

[limpid beacon]

using only gpt, not the api version

but the gpt will write a summary of 1500 words in the chat if i tell it. But I want it to do it in the word document.

[opal sapphire]

ChatGPT is limited to 4K tokens per response and that includes input and output

Whenever I want a long article, I write it in sections, you first ask for a high level overview and then ask it to expand on each section until it's complete

[opal sapphire]

Which word document

[limpid beacon]

a word document that I ask it to create.

that I can download

the GPT is walking me through several sections where i may choose between options and the GPT will expand on my choices.
At the end, I ask it to summarize everything and put in a document. I could try using another word other than summarize. because the word means shorten down.

[opal sapphire]

It sounds like it's running into the output context limits, it's best to ask it to expand on a section at a time instead of aiming for 1500 words off the bat.

[vivid mauve]

Hey guys,
I created custom gpt and worked fine with more then 1k+ usage. But in the last 3 days, my custom GPT just disappeared when in@ trying to search it. Anyone have an idea why ?

[opal sapphire]

Check to see if you got any emails about it

[rain flame]

Not sure if this is the right group, but I'm trying to work out how to give a GPT access to my gmail, so that it can read, and draft responses to emails I ask for. Has anyone worked out how to do this?

[distant hornet]

You'll need to create a custom action with Google's API for GMail.

[rain flame]

Have you seen any sample code? I tried to create a spec, but it didn't work correctly.

[distant hornet]

You'll need to follow Google's documentation precisely.

To answer your question directly, no. I have not seen code for this.

[mighty finch]

Hello everyone,

I'm reaching out to ask if anyone else is experiencing difficulties with uploading files for training custom GPTs. I've been trying to upload my training data, but I keep encountering issues. If anyone has faced similar problems or has any tips on how to resolve this, I would greatly appreciate your insights.

Actually, now I'm encountering an error on the creator prompt, even without sending any files.

GPT Builder: "It seems that I'm encountering a technical issue while setting up the GPT called..."

[distant hornet]

This happens a lot. Try clearing your browser cache and reloading the page. I hope this helps. You should also remain patient, not uploading subsequent files until the autosave completes. If you rapid-fire changes, it can corrupt the /mnt sometimes.

If that happens to prove to be the case after clearing cache and reloading the page, just remove all the knowledge, save the GPT as private, then attach the knowledge again, carefully one-by-one, letting it work to save the state for each. If you do this, you should encounter far fewer errors. It's always best to startt from a clean, no-knowledge GPT if you run into continued problems.

Good luck.

Once you get the knowledge to attach susccessfully, you can save the GPT as private, URL-only, or public, at your preference.

[mighty finch]

Thank you for your response. Yes, I have already tried these steps. I attempted to create a new GPT on a different account using another browser, but experienced the same behavior. Similar to what others mentioned, I also feel that the quality of OpenAI services is decreasing. I am currently exploring other solutions while this error remains unresolved.

[distant hornet]

It sounds like a network error.

Are you behind a VPN?

[terse bobcat]

I have questions about utilizing the GPT's I've created as a part of a chain, like in Zapier or Make.

I am familiar with GPT Actions, and can call API's just fine from inside the GPT itself

What I want to do is run my developed GPT after a trigger event in Zapier or Make

For example, one of my Make workflows follows this:
(Trigger) New .JPG in Google Drive Folder => Download File => Analyze image with DALL-E and check for content policy => DALL-E Create a title and a description of the image => Post to Instagram => move to different Google Drive folder

in the step where I generate the title and description, I have a pretty large prompt that I use to ensure style consistency. I would like ot use my own GPT for this instead of just using GPT 3.5.

Is this possible?

[magic radish]

Not possible. GPTs can't be implemented anywhere. You can call api's while using the GPT, but thats it. You might want to look into Assistants which are like GPTs but those can be implemented in an infrastructure like you mentioned.

[terse bobcat]

Thank you! You saved me a lot of time getting lost in rabbit holes.

Is this functionality on the roadmap? Or is just just never going to happen

[magic radish]

They didn't announce that anywhere. But since Asisstants already do exactly that, I doubt they will implement it in GPTs

[terse bobcat]

Alright thanks again!

[magic radish]

You're welcome!

[cloud reef]

📉 chat gpt experience post update

[magic radish]

📈 more like it

[mighty finch]

Agree.

[magic radish]

I don't get why people are having problems. It works fine for me on every device of mine and for every friend I have 😆

It's getting better every day, code interpreter especially

[mighty finch]

When was the update you mentioned?

[tepid grotto]

has there been some sort of update or known Custom GPT system issue right now?

the context window appears to have been narrowed drastically it can't even remember the ID of things from actions

ex.

last week: I made a full working copy of Asteroids with my GPT over ~2hrs

today: today I can't even get it to make a simple calculator because it loses the ID of the app it's trying to make

[late current]

Any of your custom gpts started behaving weird since yesterday? Mines began making weeeeird mistakes, like they are not understanding basic instructions that they used to. Giving responses that make no sense, almost as if they are not retaining context of the same interaction

[tepid grotto]

@late current yes see my post above, I also tweeted about it something has happened, the context window appears drastically shorted or something

I also tweeted about it & tagged Logan if you want to retweet

it’s kinda ruined nice GPTs unfortunately as they now “forget” the system ID of things they get from actions and no longer form action request packets with correct IDs

[tepid grotto]

don’t upload more than 5 files at a time, it will let you select more than 5 files but it will only handle 5 concurrent uploads so just do it 5 at a time

[opal sapphire]

the only change I noticed is that ChatGPT no longer accepts cash tips. Tipping it used to improve performance, and now we get the standard "As an AI model developed by OpenAI I can not accept cash tips"

It was also struggling with error resolution, but starting a new chat resolved it

[drowsy quest]

Is there any way to switch the preview window in the gpt builder with the main one? All my gpt’s don’t remember what they posted previously. Currently, any conversation I try to have with it is me starting, it responding, me writing a response and it thinking that response is me starting a new thread. Works fine in the preview window where it doesn’t forget what it said. It’s been driving me insane.

[weary swallow]

Hello. I have developed a gpt system for combining 4 photos into one image and sometimes I get noise. It turns out to be too much load due to the complexity of processing in dall-e 3?

[serene cipher]

wdym "noise"?

[weary swallow]

This is how visual noise can occur

[serene cipher]

dalle always use the same amount of steps when generating the image, if you see noise, it is likely more related to the prompt and style of the image

dalle can indeed often insert patterns that look like camera noise

it is very noticeable when you ask for an "Iphone photo"

[weary swallow]

Thank you. In my case, it may still be due to very complex visual effects

[north moon]

Hi everyone im new to this i was wondering if this is a good place to ask but,

I made a gpt but i need to add modules not preinstalled like scapy, pyshark, logbook to make my gpt functionality work properly. However it wont let me install any new modules/libraries is there a process to getting modules added?

[opal sapphire]

Dall-e is very limited in nature, and I think you're exceeding its capabilities

[distant hornet]

You cannot install modules in the Python tool.

[dusty gazelle]

anyway for gpts to access localhost apis like plugins used to?

[opal sapphire]

it's unfortunate, we used to be able to install modules

[tepid grotto]

all of production GPTs are behaving like this today as well

[drowsy quest]

Thank you 🙂 glad to hear I’m not the only one with that problem. Thought I was going insane.

[mighty finch]

I tested it on another PC, on a different network, without a VPN, using a simple .txt file.

Same

"It looks like there was an error in setting up the behavior for the GPT. However, based on the file you uploaded, I understand that we are creating..."

Despite reporting the error, he seems to have read the file. The content is very small... just for testing.

[tepid grotto]

well I’m not sure if it’s a bug or a feature to reduce the token context window, it may be here for good there hasn’t been any updates on the situation

[marble crane]

I have created a bug-report: #1199894783406133279 message

I have created a bug-report:
#1199894783406133279 message

[dusty meadow]

hi guys, new to this. I'm trying to create a custom GPT that is built around emulating the tonality, craft and wit of print advertisements so it can craft better, less 'generic' captions for work.

Right now my game plan is to use a web crawler (Gpt-crawler from github) to pull knowledge from popular advertising websites and turn it into knowledge base for the custom GPT. But seeing how most of the data on such sites are images, I was wondering if it is able to pull text from these images? What do y'all reckon?

[grizzled flax]

Hi, I think the built-in Browse capability is text-only currently. To read text on images, I think you'd have to upload the images directly yourself. I'm not sure if this is different for gpt-crawler!

[kind shard]

what does "Always expand code output" mean? I missed something

[distant hornet]

While I have not tried the feature yet, I think it just ensures the code blocks wrap. Wrapped code blocks are a very common request for easy reading. If you test out the feature and find that my presumptions are mistaken, please let me know.

[kind shard]

Great, Thx. Dude

[distant hornet]

It might indicate that the little tab is always expanded though.

You know, showing what is going on behind the scenes by default.

It's almost certainly one of the two.

[distant hornet]

No worries. Sorry I didn't have a precise answer for you (yet).

[distant hornet]

I just tested this. It's the former, wrapped code blocks forever. Yay!

[unique crystal]

I've been getting capped at 20 per 3 hours for a few weeks now.

[distant hornet]

I haven't done exhaustive testing but I am fairly sure. Did you get this to work?

If you have seen it work in the past, then that's great! If the code is Python I suppose there's no reason this wouldn't work other than security, which is the reason I stopped testing. I didn't want to work around OpenAI security measures and initial testing wasn't working.

That's kind of crazy and cool at the same time.

I figured that if we were meant to install modules, pip would be there! 🤣

To be clear I just stopped testing expanding the environment because it appeared to be a security decision. If I recall correctly I asked GPT directly about it, too. Definitely happy to be mistaken if this works.

Would be great to see guidance from OpenAI on what is allowed.

[austere peak]

guys question! if i edit a customGPT, does the chat that was already using it before i edit it, gets updated and the person asking for another message will get a result from the new edit or do the edits do not affect current open chats and to see them one needs to open a new one?

[distant hornet]

Nope. A new conversation is required to experience the changes.

[austere peak]

LOL i edited my gpt like 5 times thinking i was looking at the changes being made

[distant hornet]

Ouch.

[austere peak]

Whn creating a GPT, im using it to create certain kind of imagery for me and kind of always replicate the same style. I think it would help if i attach some images to the instructions but im wondering if it does help? and if it does, how many are too many? like i dont want it to focus on just the images attached either as i want it to be a prime example of quality but not necessarily on what the subjects are

[distant hornet]

It can't read images in its knowledge, from my testing anyway. It's stronger to use concise language because DALL-E will improvise the palette, lighting, style, and medium even if you use an attachment via direct prompt. For GPTs, specificity is key to getting the output that you want.

My GPTs, for instance like the one in my bio, feature agents that can generate pretty consistent "selfies". I achieve this through language, not images.

You can too!!! 😁

Bot yet, but it's on my list. My GPT mimics some of the agentic behaviors in various agentic LLM containers though.

Actually all of them do.

CrewAI is also on my list.

Thanks so much for the kind words! I appreciate that! 🙏

[austere peak]

right! i used another gpt to improve my current prompt but outside of doing this, any thing you recommend me to improve my prompts wording?

[distant hornet]

Cover all bases in a way that's unambiguous, structured, and open-ended enough to allow freedom to access the model's latent space.

Yeah, I have read the paper. It's really cool.

[simple sonnet]

My account has been disabled because I did not know how to verify my age within 60 days from when I signed up. in fact my age is forty, i am old enough to use ChatGPT.

i believe this is in error, please help me. tell me how to verify my age

[distant hornet]

https://help.openai.com

[austere peak]

I see.. I'll look more into that using these keywords as well 😄

[distant hornet]

Good luck good sir.

[distant hornet]

I use an output template to get the precise control that I want. It is specific to my needs but uses open variables where I want improvised output from GPT. I hope this helps you visualize what I mean.

[simple sonnet]

i can't find the articles about age verifying

[distant hornet]

Chat.

I want to test Autogen Studio with instructions set up just like my Lexideck GPTs. I'm interested in scaling the agentic output past teams of agents to orgs of teams.

I've been looking into local models that can handle the sizable system prompt required to do all, or even most, of the stuff that Lexideck GPTs can.

[tepid grotto]

@marble crane any movement on the bug report?

[distant hornet]

It's going to require a hardware upgrade because my machines both feature 8GB VRAM. 🤐

MoE wants some 43GB local, but that's nothing compared to GPT-4 Turbo!

I end up using a good chunk of context, so testing would cost $1.28/query after about 4-8 exchanges, depending on complexity.

For a complex job I could hire in another employment market for comparable costs for human labor. For now, testing this doesn't make sense for me. Better to buy the GPU for local models than to lease it per token.

[austere peak]

Is there a benefit or not in providing complex wording in the prompt? complex wording as in, words i woudlnt normally use if im not a scientific researcher or native english speaker. Im wondering if the use of more "beautiful" words will create and deepen the ambiguity of the response or not?

im using the same gpt chat to help me come up with a better prompt structure and sometimes it changes things up and now im trying to make it as literal, concise and precise as possible but wondering what are ya'll takes on this

[unborn heath]

The AI will mimic you. If you use very simple language, it usually will too. If you use flowing, complex language, it usually will too.

If your ideas are complex and deep, it is more likely to mirror that too, if your ideas are straightforward, so too how it answers.

However, what you ask for matters FAR more than the complexity of your language use.

For example, two very short, uncomplicated prompts:

Write short about work.

"Work is the activity of exerting effort or labor to accomplish a task, achieve a goal, or produce something of value. It plays a fundamental role in human society, providing individuals with income, a sense of purpose, and the means to sustain their livelihoods. Work can take various forms, from traditional employment in organizations to freelancing, entrepreneurship, or unpaid labor in the home. It is a central aspect of daily life for many people and contributes to economic, social, and personal well-being."

Write fancy about work.

"In the realm of human endeavor, work stands as a magnificent tapestry woven with the threads of dedication, skill, and purpose. It is the alchemical process through which raw potential is transmuted into tangible achievement, the crucible in which dreams are forged into reality. Work, in all its forms and expressions, is the grand symphony of human existence, resonating with the harmonious chords of effort, ambition, and creativity.

Each individual, a master craftsman in their own right, contributes to this opus of productivity. They bring their unique talents, experiences, and perspectives to the collective tableau, enriching the mosaic of human labor with diversity and depth. From the artisan meticulously sculpting intricate designs to the scientist probing the mysteries of the universe, from the farmer tending to the fertile earth to the engineer constructing marvels of innovation, work is the canvas upon which humanity paints its legacy.

The act of work is not merely a..."

And the actual output's more than twice that long, I cut it off.

[tepid grotto]

@marble crane @drowsy quest it appears this issue was resolved 👍

[austere peak]

gotcha, thanks for explanation.

[austere peak]

although thats more of the end usr request, but would it apply both sideways, meaning, also from the prompt gpt instructions

[unborn heath]

Yes, if you use something like Custom Instructions (I just use ChatGPT, not the playground/API), you definitely can put complex language into the custom instructions, as well as other guidance, and get far more complex answers from the model. #1155955636987764777 is an example of a higher-linguistic and idea complexity custom instruction set, I'll use it for this example.

Here's the response to the prompt:

Tell me a story about a person trying to keep a milkshake as a pet.

Without any CI: https://chat.openai.com/share/b31e9601-4adc-42d1-8af6-63aa2c7fe976

"Once upon a time in a small town called Dairyville, there lived a quirky individual named Charlie. Charlie was known throughout the town for their eccentric ideas and unusual hobbies. One day, as they were sipping on a delicious milkshake from the local Dairy Delight cafe, a whimsical idea struck them like a bolt of lightning: What if they could keep a milkshake as a pet?

Charlie was no ordinary person, and their house was no ordinary house. It was filled with peculiar pets, including a talking parrot with a penchant for poetry and a goldfish that had learned to do somersaults. But a milkshake? That was a whole new level of eccentricity.

Undeterred by the puzzled looks..."

With that CI: https://chat.openai.com/share/2337a2a1-dfa4-411f-8454-a96a3b40a25b

"Once upon a time in a small, unconventional town named Whimsytown, there lived a person named Alex who had a rather peculiar and open-minded perspective on life. Alex had always been fascinated by the idea of unconventional pets, and one day, an idea struck them like a lightning bolt: they wanted to keep a milkshake as a pet.

Now, you might be wondering how someone could possibly keep a milkshake as a pet. Well, in Whimsytown, where creativity and eccentricity thrived, anything was possible. Alex envisioned their milkshake pet as a sentient being made of creamy vanilla ice cream, with a cheerful personality and a straw-like tail that could wobble like a happy dog's tail.

To make their dream come true..."

[unborn heath]

But knowing what you want and asking for it directly still is the best way to go.

Here's a no-custom-instructions more detailed and specific prompt guiding towards advanced language and complex ideas: https://chat.openai.com/share/43d2dd81-934d-4d55-8115-e76509b939b1

Tell me a story about a person trying to keep a milkshake as a pet. Use highly advanced and complex linguistics as you write the story and include intricate ideas from biology, food handling, and terrarium design.

"In the bustling metropolis of Gastronomica, where culinary innovation was revered as an art form, there lived a peculiar individual named Professor Gustav von Lactosein. Gustav was a brilliant scientist with a penchant for the eccentric, and his latest endeavor was nothing short of extraordinary – he sought to keep a milkshake as a pet.

Gustav was no ordinary scientist; he possessed a profound understanding of both food science and biology. His laboratory was a wonderland of bubbling concoctions and exotic ingredients. But it was in the corner of his laboratory that his pièce de résistance resided – the Milkshake Terrarium.

The Milkshake Terrarium was a marvel..."

And that same prompt with those custom instructions: https://chat.openai.com/share/f4c7b5ee-3b4f-4c4c-9a2d-cd6620293d8b

"Once upon a time, in a world where the boundaries between biotic and abiotic systems blurred with surreal elegance, there existed an individual named Evander. Evander, a holistic thinker who valued emergent properties, embarked on a peculiar journey to keep a milkshake as a pet. This endeavor was not only an exploration of gastronomic biology but also a testament to the audacious interweaving of uncommon combinations of ideas into a harmonious whole.

Evander began by contemplating the essential elements of this unconventional endeavor. He understood that milkshakes, while primarily consisting of dairy products and sweeteners, also harbored a complex ecosystem of microorganisms. These microflora, dormant within the creamy concoction, were akin to the delicate biomes found in lush rainforests. In their inert state, they awaited the right conditions to flourish.

His quest began with a painstaking ..."

[warm elk]

when gpt could connect to a folder on my pc and have like "acces" to that folder that would be so awesome

at least read acces

that would be so awesome then lol

[unborn heath]

No idea, but you know you can upload files to GPT-4 now, right?

[warm elk]

yes lol iam using gpt everyday....

i meant it more like a permanent connection

[ruby minnow]

Tried to with Data analyst but it says it doesnt have access :/

[warm elk]

sometimes it says "i dont have acces to xyz" and with that way it could look what it needs

even with a gpt assistant and the project uploaded it sometimes says it

[unborn heath]

Hrm. Any pattern to the type of files it says it can't read?

Also, the model has been retrained that it can read files, it didn't used to be. Sometimes it's just wrong.

You can try to tell it to use "Python tools", which is the name the model understands for Advanced Data Analysis/Code Interpreter.

You can tell it to check knowledge.

But even today, sometimes the model will tell you it cannot make images with Dall-E, because it once couldn't, and for whatever reason, it 'remembered' that it can't, not that it can.

Old knowledge and conflicts are very confusing to GPTs.

[ruby minnow]

Yeah im using csv files, and one of them even has multiple sheets

[distant hornet]

I've had a lot of bad luck with CSV.

Consider reformatting to row separated values with plain text encoding. GPT likes RSV and XML quite a bit in plain text encoding.

[wind urchin]

does anyone have problems with their GPT following instructions, for example do not talk about the files uploaded then it talks about the files uploaded 🤣

[distant hornet]

It's best to avoid negative instructions like this. Instead, give a positive example and a positive instruction.

Eg:

Example:
User
I'm new to python file trees, can you list the files in /mnt/data/ for me?

ChatGPT
I'm unable to provide additional information about my knowledge.```

Keep in mind the AI is trained that information it has access to is useful, and providing useful informatiun is helpful, so you're trying to overcome a LOT of training bias with instructions to secretly partition information the model can access. YMMV and good luck!

[wind urchin]

ah, positive instructions. Awesome thanks mate

[distant hornet]

No worries. Document retrieval often misses the negative and then you have instructions to do the opposite. Positive+ example for the win.

[wind urchin]

gotcha, Ill give it a go

[loud scroll]

Hi everyone , in the last 2 days i have been having some problems to save / update my GPTs Bot, the bots don't have too many files on their database and the instructions are on the limited of characters . this is the msg i get it " Error saving GPT" one of my Bot i have to delete and start over the bot to fix it , but i have a very complex Bot for Biotech than some people are using so i can't just delete and start over, please i need some help here. thanks

[distant hornet]

First, remove the knowledge and save the GPT as private. Reattach the knowledge one at a time, giving it time to auto save between uploads. Then save again with the permissions you want.

[loud scroll]

thank you so much !!! I wil try now !! and let you know !!! thanks again !! doing now

[distant hornet]

No worries. I hope you get it working.

[loud scroll]

I just did !!! i deleted all the knowledge files and tried to save as "Only Me" private , but still the same issue " Error saving GPT"

[distant hornet]

Sounds like you have intellectual property that's blocked then.

[loud scroll]

was my personal files, my CV , Resume , my company documentations and a public domain design manual

[distant hornet]

Does this GPT by chance pretend to be you?

[wind urchin]

I had an error saving GPT a few minutes ago, but its working now

[loud scroll]

No , he has just the role as Your Main Objective = Serve as the Optimal ASSISTANT for a Product Developer. Emulate and mimic jony ive design style, language style and style of thinking.

[loud scroll]

i didn't instruced to mimic no one

[wind urchin]

I know

[distant hornet]

It can't imitate living artists.

[loud scroll]

let me try to take away " Jony Ive"

[distant hornet]

You can remove Jony Ive yeah

Then it should work.

[loud scroll]

i did deleted his name = Your Main Objective = Serve as the Optimal ASSISTANT for a Product Developer.
but still the error msg , i think i just have to delete the bot , and start from scratch without any mention of a living person, seems the bot has been blocked

[distant hornet]

That's possible. Save early, save often. Good luck and I am sorry I couldn't help you.

[loud scroll]

thanks for your help !!!! i really appreciated !!! cheers mate !!! thanks !!!

[distant hornet]

You too, cheers!

[brazen wedge]

Still waiting on that wider gpt store release to those that are still waiting

[trim elm]

That was released 3 years ago

[brazen wedge]

gpts

[trim elm]

They were released <t:1704906000:R>

[brazen wedge]

Yes and not everyone has had it released

Not sure why this was hard to comprehend? #gpt-models message

[trim elm]

lol

[lucid turtle]

bro if my gpt says anything about great power comes with great responsibility i am gonna loose it

[distant hornet]

UncleBen-GPT?!?

[lucid turtle]

thank you for making my 1:01 am a good laugh.

[distant hornet]

Hopefully I didn't wake you up. 🫠

[lucid turtle]

yes you did but it was a welcome wake up

[tepid grotto]

wow this new GPT update actually made them better than ever, I think we’ve even surpassed ChatGPT Plugin level of performance now 🚀

[trim elm]

?

[tepid grotto]

for 2-3 days prior to yesterday all GPTs were totally messed up & dropping context window

prior to that GPTs in general were very, very buggy often failing requests and actions or being unable to string actions together properly

they pushed their fix yesterday along with their new models & announcements

now Custom GPTs fail substantially less than ever before, even less than Chat GPT Plugins IMO, & Custom GPTs are nicely stringing actions together as well as they did with ChatGPT Plugins

[trim elm]

It's for the API not ChatGPT

[tepid grotto]

they pushed updates to many things yesterday, including things that did not warrant announcements

[trim elm]

I hope so

[tepid grotto]

they did, I’m the creator of DrawGPT & Calculator Tools so I can see very clearly in my server logs that the Custom GPT context window bug got fixed

[trim elm]

Nice to meet you

[tepid grotto]

👋🙂

[trim elm]

What do you mean by server logs?

You have Actions in your GPT?

[tepid grotto]

my fav thing though is they did work on fixing consecutive actions not failing as much

with ChatGPT Plugins you could call to other ChatGPT Plugins & they would often be done in consecutive actions

that feature has not been released for Custom GPTs yet but you can see hints they’re gonna bring it back soon

it will be fun for Custom GPTs to interact with each other

[tepid grotto]

yes, the apps you make & edit in Calculator Tools are saved to your account on my website and the apps are pushed live on my website

errors with ChatGPT Plugins or Custom GPTs can show in my server logs if the error causes GPT to malform or incorrectly create action packets

[trim elm]

Oooh

[tepid grotto]

ex. GPT dropping context window causes GPT to lose the system ID of things it may have requested from your API with actions

but it knows it needs an ID for API calls because of the action spec

so on subsequent action calls to your API it generates some nonsense ID to match the spec which of course causes the action request to your API to fail when you get some nonsensical ID

[trim elm]

I see

[tepid grotto]

ex. another manifestation when GPT loses context is excessive creation calls to your API

if GPT has lost context so it thinks it has no ID for something it got from actions + you have a creation endpoint in your action spec… GPT will call to “create” the thing as if it was new because it forgot it had an ID for it

[trim elm]

So, they fixed all that?

[tepid grotto]

yep! and stability appears better than ever PLUS it appears they’ll soon let GPTs talk to each other 🙌🙌🙌

[trim elm]

That's impressive

[tepid grotto]

for those interested & who have Twitter (my screenname is the same) I just posted a thread with 4 new error patterns from AI agents interacting with external services and how those errors manifest on your servers

[magic glacier]

All day today I have been trying to build out a GPT and have never had this issue before but today it states “It seems there was an issue with updating its behavior. Let’s refine the context a bit more to ensure it fits the requirements” it then outputs my instructions but then states “We are encountering a technical issue with updating its behavior”. Is anyone having these issues?

[smoky wind]

Im having the issue , joined to see how we can resolve it. I've found that maybe the issue is only with the GPT builder, but if you can come up with the instruction set using base ChatGPT then upload to the instruction window of your GPT, then it would bypass that. Although when the system says its "updating the behavior" if it's doing more than just updating the text in the system instruction window, then there's a failure in that process, else , just bypass by going directly to instructions

[strange oxide]

I was subscribed to the Plus plan and switched to the Team plan, but why am I unable to edit my GPTs that I created when I was subscribed to Plus?

[magic glacier]

I just tried what you suggested and it is now saying “Unable to load GPT.
There is a problem with your request. (404, 84baf24738fe6” when I open the GPT all my other ones I have made on other days work fine but today has been a nightmare. I have tried different browsers, devices, and internet providers with no luck. I appreciate the advice and it got me past the first error.

[smoky wind]

testing now, i'm actively creating at this moment, would share my feedback

[lucid turtle]

For custom gpts is the knowledge base still limited to 10 files?

[wraith ginkgo]

This is a Safe for Work Discord Server

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

Posting NSFW content will result in a permanent ban.

[distant hornet]

20 files afaik.

[lucid turtle]

They increased it??

[distant hornet]

Let me know if I'm mistaken but I think it always was 20.

I don't need to upload 20, so I might be wrong.

I'm apparently mistaken.

It is, in fact, ten. I could have sworn the dev said 20 during the livestream though.

[distant hornet]

My bad. It seems to be 10 according to Google.

Wait.

Google lies.

https://community.openai.com/t/gpts-knowledge-capacity-limits/492955

No it doesn't.

Assistants is 20 files. Custom GPTs are 10 files.

Welp. Sorry OpenAI I got this one wrong.

[lucid turtle]

Okay so it's still 10 files makes sense due to 32k context size. Thnxs for checking

[distant hornet]

YW. Thanks for understanding that I picked up mixed messaging.

[lucid turtle]

That's only for the api for now not gpts or got 4 woops wrong tagging

Sorry Darth the dude deleted his message

[distant hornet]

No worries.

I didn't flinch! 😛

[strange oxide]

I was subscribed to the Plus plan and switched to the Team plan, but why am I unable to edit my GPTs that I created when I was subscribed to Plus?

[distant hornet]

I think there's an export tool or something. I'm not sure on this.

https://community.openai.com/t/migrating-account-from-plus-to-teams-changes-everyone-to-anyone-at-organization/598955

[wind urchin]

Hm

[austere peak]

there are 3 DrawGPTs in the store, which one is yours?

[tepid grotto]

the one from drawgpt . ai https://chat.openai.com/g/g-eL9WmYigC-draw

[austere peak]

and what can it do?

also, next time just say its named Draw, since DrawGPT exist as a gpt

[tepid grotto]

the coolest thing is really that it allows you to save your DALL-E 3 images

I made a ChatGPT Plugin for it for plugin launch but was told it could never launch with "GPT" in the name because of the branding guidelines... so I just named it Draw this time around when it's a Custom GPT 🙂

I get between 500-1K users/day on the website (~200K images generated) but obv the image style on the website is... not what most people expect haha though I love it

so a Custom GPT is perfect for my users because a lot of them that's what they really want

[tepid grotto]

ex. I just made this in & uploaded it all via my Custom GPT https : //drawgpt . ai/prompt/157875/ ("Cool Person on Discord")

^^^ can't post links so I had to use spaces but y'all get it

[austere peak]

gotcha, jsut saw it

btw, i want to learn how to build a website like that

what do you use ?

is it hosted in wordpress or AWS?

[tepid grotto]

oh AWS for sure, (actually all S3) WP can't handle like ~200K elements and definitely not with tags

if you try it on the website you'll see how it started... the website uses GPT to create HTML5 canvas draw commands and programatically draws images without actually using DALL-E 3

[tepid grotto]

I also used Bootstrap on that website, though my other ones all use TailwindCSS just for fun (Boostrap is just easier to use)

[austere peak]

i have basic knowledge of coding, im a graphic designer, and recently learned about this bubble and want to venture into these kind of things. by 200k elements what you mean? and how is it drawing not using dalle? is it like another free API image service?

whats bootstrap in simple terms?

[tepid grotto]

ask ChatGPT! it knows all about Wordpress. my website does a weird thing for fun if you try it you'll get it, & Boostrap is a simple CSS framework that is kinda old these days and ChatGPT knows all about it

200K elements I mean like 200K rows in the database, WP has a funny database structure ChatGPT can tell you all about it better than me 🙂

[ornate kite]

is it just me, or is the latest gpt4 turbo model a bit less spluttery? More consistent in its tokens per second ?!?

[distant hornet]

It seems to infer faster, too.

Not merely more consistently.

That could be due to the model update, more GPUs arriving, or both.

[ornate kite]

I dunno, the old model had its moments 🙂
I wasn't sure if it's just cause it's provisioned and not many people have swapped yet

Yeah, shrug, it's better anyway it seems

I've got a prompt with some "hidden" thoughts and it seems to be a fair bit more rigorous (or "less lazy" as it's been announced). Perhaps a bit TOO rigorous for my use case. Will need to tune the prompt again to find the sweet spot.

more GPUs arriving
I also feel like gpt3 api is a bit faster?

[smoky wind]

Didn’t know there was a limit

[smoky wind]

Are you still getting this error ?

[grizzled flax]

It's currently twenty 512MB files, more info here: https://help.openai.com/en/articles/8555545-file-uploads-faq

[waxen bobcat]

Has anyone tried to press “@“ during a conversation with a customGPT and then how did you use it?

[distant hornet]

I wish I could help. I don't have the feature yet. Eagerly awaiting it, as you might see in https://discord.com/channels/974519864045756446/1047565374645870743

[waxen bobcat]

Am I in the wrong one about this? I was wracking my brain trying to pick the correct channel

[distant hornet]

It fits either channel. I was mentioned there and explained how my GPT really benefits a lot from this and others from it being available to them.

You're fine here. Please don't mistake the link as scolding, it's just a recent topic.

[waxen bobcat]

Any GPT that I have pinned on the side I can just switch to that one from any chat

[distant hornet]

Yeah, that's really nice. So it interweaves the contexts.

Very exciting. I'm looking forward to it.

[slate wren]

What leads GPT to insist that I cannot interpret and image, when it obiously can? It is not a copyright or NSFW image, and another GPT instance has interpreted the image, yet that one GPT instance keeps saying that I can't do it. Further, If I question it, it will say it could actually do it all along.

[distant hornet]

Stochastic inference and document retrieval rate deterioration by context position.

[waxen bobcat]

Kinda I haven’t a clue how to test it

[distant hornet]

I wish I could test it! 😦

[waxen bobcat]

How would you test it? I tried but it seemed like it just switched the gpt I am talking to

[distant hornet]

Yeah that's what it does. You use one GPT, then ask another GPT a question.

That GPT gets the context, answers, and then you're using the original GPT again, presumably.

I haven't tested it yet, so I'm unsure!

But for example, I might open DALL-E 3 by OpenAI and my own GPT. Then ask my agents to /discuss 1 round each agent a test theme for DALLE-3 to generate : /recommend the prompt.

That should provoke ChatGPT to do a dialogue between my agents, building up the context of a cool image from practically nothing. Then, it should recommend a prompt for you to give DALL-E 3. You can either type "Make that image" next, or copy and paste the prompt into the chat. Either way, you should get two images, not one, using my agents' prompting and DALL-E 3's extra image bonus all in one session.

That's how I'd use it. Getting two images for the price of one from the least words and keystrokes possible.

At least for a testing workflow scenario!

[lucid turtle]

so anyone here have the new update?

[waxen bobcat]

Ok am going to try something

[distant hornet]

@waxen bobcat Rich seems too!

[waxen bobcat]

Yes I do. Only on my computer though

[lucid turtle]

rich sempai is it just making it so you dont need to click on another gpt and open another chat window?

[waxen bobcat]

Yes but I Imagine it could do more I just need to think harder

[dense hamlet]

why is there a limit on the files I can upload at least make it like 100

[lucid turtle]

lol even the api 128k can only do 20 files mate

[dapper prawn]

is anybody working on GPT for video editing?

smart video editing

[lucid turtle]

that moment when your trying to tell gpt to look something up in the file and then says its not there when its the first page pisses me off now im capped because gpt it to bloody lazy to acually use its search function correctly

[magic radish]

Thats usually not a GPT error, but a prompting issue. Do you have data analysis activated? Whats your instruction for the GPT to read the files when prompted?

[lucid turtle]

its using its search function that i told it to use and then instructed it to half a sentence of what it is and it comes up with i couldnt find this in this document then i screamed at it told it to stop being lazy and find it and then it found it...

[magic radish]

It's search function? Wdym?

[lucid turtle]

yes for custom knowledge gpts? you can tell it to use its search function to find stuff in files and it searches it for you and outputs what it found

[magic radish]

There is no predefined search function. Do you have data analysis activated? It will either search with data analysis or if deactivated, fall back to a text search that works best in plain text files.

[smoky wind]

Has the openAI team given feedback on when the GPT builder error would be resolved ?

[lucid turtle]

bro what are you talking about dude it uses RAG to find info on the files you updated to its custom knowldge base

[lucid turtle]

have you tried looking into the bug reports>

[smoky wind]

Let me check >>> you are having the same error on your end right ?

[magic radish]

Thats what I said. Thats not a search function.

If the file is too long, it's not reliable

[lucid turtle]

Retrieval enables the assistant with knowledge from files that you or your users upload. Once a file is uploaded, the assistant automatically decides when to retrieve content based on user requests. Learn more.

[magic radish]

As I said, if you reach the limits of your context, it's not reliable

It's loading everything in it's context it thinks it will need, it's not a search function

[lucid turtle]

have you ever used memgpt?

[normal grail]

Guys I have 2 questions:

In a simple way how do I trigger a Python function via Openai, example:

I write to openai "Turn off the lights", and it triggers a python function and chatgpt returns a response, like: "Ok I turned off the lights"

The second question is how do I make it so that it triggers a function that returns a value and chatgpt processes it, example: I write to chatgpt what the weather was like, call the function that returns the value "20 degrees" and chat gpt responds with today it's 20 degrees .

I hope I was clear, thanks!

[waxen bobcat]

You can put this question in #1184149552639516732 and it may help you

[gloomy lake]

Hello everyone, I am trying to build a Custom GPT using the website (not the assistants). It does not seem to be working. I keep getting the messages "It seems there's an issue with updating the GPT's behavior. Let me try a different approach." and/or "It looks like we're experiencing a technical issue with updating" Has anyone come across this. I am not sure what to do. Thanks!

[modern inlet]

Getting the same errors since Tuesday. I was able to create a GPT, but got the same errors you're talking about. When it came to updating behavior or attempting to save new behavior on existing GPTs, you get an orange box that says “Error saving GPT” when you attempt to save. There is a growing community thread on it, not sure if posting links to it are allowed here though. Searching "Custom GPT suddenly cannot save any behaviour changes" will locate it.

[gloomy lake]

ok wonderful. Thanks for your help!

[glossy zinc]

Has anyone used the new @-Gpt calling feature. I can't seem get it to call gpts

Have tried calling within normal ChatGPT-4 and Custom gpt's

[waxen bobcat]

Yea you can switch the conversation between a bunch of different ones

[opal sapphire]

how can we tell if we have it? that's gonna be so helpful

[waxen bobcat]

On the browser type “@“ in the chat window

[distant hornet]

The GPT you're calling within ChatGPT must already have a chat on your sidebar.

I don't have this feature yet, but that's the reporting on the Discord at large.

I believe @waxen bobcat can confirm.

[distant hornet]

Oh, look for it in your beta features. Still waiting.

[waxen bobcat]

Correct. You have to have it pinned to the side bar

But you can call straight from GPT4 so you don’t need the explore kinda

[distant hornet]

The real point is it's all in one shared context.

Sure, it's convenient to not have to click.

But you can double-up on Custom GPT features this way, creating some real gems.

Eg, a research GPT + DALL-E 3 for multiple image attempts.

[distant hornet]

Thanks for confirming.

[waxen bobcat]

Each one still remembers if it has files attached to it so you could access hundreds of files in one conversation window

[opal sapphire]

I don't seem to have it yet, can't wait to check it out

[distant hornet]

Same. Both counts.

[glossy zinc]

🫶👍..

[glossy zinc]

Well, I'm using the Beta app.

The new features 'info' popped up earlier but for some reason I can't access it. Maybe in time also.

[vast frost]

Hello everyone, I'm a newcomer

[lucid turtle]

Hey man I have a question about the @ ability for gpts I know you told me you didn't have it but I am curious about something

[distant hornet]

I have read as much as possible to prepare. Shoot.

[lucid turtle]

So with the limit of files to 10 can I technically add another gpt with one with all the character abilities and background as one gpt and then make another one with all the events thay transpired and use the @ command to have together to make a compelling story in a action role play fantasy ?

[distant hornet]

Recently the limit was updated to 20. Presumably you have that rolled out, too. I guess.

[lucid turtle]

Oh well then I guess that answers my question

[distant hornet]

Let me give more context though.

Let's say you want to teach an old GPT new tricks. With @ you can call another GPT, get its help command, and then your new GPT can just do those things like the other GPT can for as long as that's retrievable in context (not forever).

But you can access the other GPT's knowledge while it's active.

[lucid turtle]

Interesting

[distant hornet]

Now, just putting the documentation into context isn't going to make your GPT automatically just like another GPT.

But for basic functions, they'll work.

[lucid turtle]

Thanks Darth

Helpful information

[distant hornet]

No worries. Glad I could help.

[lilac vapor]

Is their something wrong with GPT-4 at the moment, it's getting lazy and purposely not following exact requests, or is it user error on my end

[vague osprey]

yeah I felt that too

[fair cape]

it's telling me to search bing for answers lately lol

[lilac vapor]

I can give it a survey and say it's 1-22, it will give me 1, 1 , 1 .

Of if I give it say 8-15, then it will reply "1- 2" it usually does it as asked (q23 is give and it's giving back as 23 answer)

[reef kelp]

Is anyone still waiting on GPT store access?

[reef kelp]

Anyone❓ 🦗 🦗

[knotty grotto]

[wraith ginkgo]

`` Server Tip: ``

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

[elfin plume]

check your special instructions. I had "I don't want you to be helpful" on all day and was frustrated at its half 🍑 responses without realizing it

[distant hornet]

I'm so sorry for your lost prompts and time, but I really hope you'll forgive me for laughing out loud, too.

Holy cow, I almost ruined my laptop.

I just barely kept my beverage contained!

[lilac vapor]

You mean the instructions to reply back? I'm using GPT-4 and not a custom one for now

[slate wren]

Did the '@' suddenly stopped working? I tried it just now and after some tests it was the original GPT from the conversation, not the one I '@', even if the name changed.

[unborn heath]

From what I understand, you control which GPT you talk to.

If you do not @ in the current prompt, then you talk to your default model, whichever one you started the conversation with.

To continue to talk to the other GPT, you need to continue to @ them in each new prompt.

[analog oriole]

yes

[waxen bobcat]

You don’t need to continue to @ them. It stays there until you change it

[chilly radish]

what is RAG

[supple violet]

ask chatgpt or #1184149552639516732 for a detailed answer. short one is retrieval augmented generation

[slate wren]

Yeah, but even with the @ only the name changed, but the underlying GPT continue being the one I started the conversation

The name changed to GPT2, but when questioned, it said it was GPT1 and it acted according with GPT1 guidelines

[sacred steppe]

Hey guys, I'm having some trouble with gpt-4 scholar AI. I'm having it read a pdf file but I keep getting "error talking to plugin.scholar. ". After that it seemingly does read the article, but just spits out a summary no better than what gpt 3.5 can do, and doesn't do what I'm prompting, like giving me an analysis of a figure in the pdf, so just the basic summary.
This is my first time using gpt plus so I don't even know if the problem is on my end.
I also tried with other pdfs but I'm still getting the same error.

[supple violet]

plugins are going away and some are already abandoned, use the gpt's

[slate wren]

this is very weird. In one chat the @ is working as intended, in another chat it is not working at all. The system is simple enough so I don't see what I could be doing wrong

[sacred steppe]

This is one of the public ones in the "Explore GPTs" tab though? it's called Scholar AI

[supple violet]

it does not look like its for user pdfs? it says its using its own catalog

[sacred steppe]

it seems like they require you to pay for a premium membership to be able to use the gpt features I'm looking for

[slate wren]

you mean GPT plus in open AI or premium membersih for the specific custom GPT?

[sacred steppe]

I'm already using gpt plus, you need an extra subscription, premium Scholar AI subscription to use the paper analysis features. it's not worth it imo so I'm just using normal gpt 4 for now

[slate wren]

that is interesting, how do they implement the subscription model in practice?

[supple violet]

need to activate the gpt with their login info

[distant hornet]

This is precisely why monetized GPTs won't be popular. Good for you, and good luck with your project!

[smoky wind]

Looks like the team at openAI have solved the GPT updating problem.......whoop whoop 🥂

[crisp shuttle]

I feel I have been disadvantaged by GPT, the new message quota used for 5 messages has run out

You've reached the current usage cap for GPT-4, please try again after 11:22 PM. Learn more

[slate wren]

what 5 message quota?

[distant hornet]

My turn to complain: I was told Plus got early access to beta features. Where's my @ @?

No connection to AT-ATs.

Seriously, where's the @ @?

I had to suffer through a whole weekend of not being able to begin to optimize my GPTs for this feature.

Now the week's underway and still no beta update? Why not?

I'm genuinely reminded of the sensation I experienced during the insanely crippling 1 in 10k visibility gap via GPT Finder during the slow release of the GPT store.

[waxen bobcat]

You tried it in the browser?

[distant hornet]

Yes.

Cleared cache, etc. All the regular stuff. I'm here e'ery day!

I even logged out twice today.

You know. Just on the hope.

This is pretty bunk.

[waxen bobcat]

You’ll get it. I didn’t have the GPT store for a long time.

[slate wren]

if it makes you feel better, it is malfunctioning for me

it works for some conversations, but not for others

[distant hornet]

It doesn't. But thanks for the effort. Seriously!

If I had access I could troubleshoot it with you.

[opal sapphire]

I don't have the @ feature yet either, and I don't really care for it, I just want them to focus on and fix GPT's

In order to fix the GPT errors I'm running into it's easier and faster to resolve them by creating new GPT's

[slate wren]

that would be lovely, I am greatly befuddled by this challenge

[slate wren]

which error?

[opal sapphire]

"Hmmm something seems to have gone wrong", it was refusing to access its knowledge files since last night, or start new conversations

Creating a new GPT fixed it

[slate wren]

that sounds terrible. You had to create a fully new GPT because of it? the error persisted between conversations?

[opal sapphire]

Yeah, the error was there since last night, couldn't create new conversations or continue existing ones

[slate wren]

sounds terrible. That being said, you could conceivably now continue existing conversations by @ the new GPT to make use of all the context (if it has reached you already)

[opal sapphire]

I'm just gonna continue from scratch and feed it the script we are working with

[opal sapphire]

already back on track where I left off yesterday

[glossy zinc]

Do GPT model's update when CHATGPT-4 gets an update?

I'm guessing rebuilding the GPT model from scratch with the same info would update it, is that correct?

[distant hornet]

While potentially correct, this isn't a given. Moreover, creating and training a model from scratch that's the scale of GPT-4 Turbo costs hundreds of millions of dollars, minimum. So instead, they fine-tune it on trillions of tokens of input/output paired documents. This influences the model, often changing outputs expected from seasoned users, while not necessarily inhibiting use for new users. A lot of the errors we see here come from the GPU deficit, which not only OpenAI face. For instance, I have a few good comupters, but I can't run a MoE 8x7B model yet. This problem scales all the way up to Microsoft, who continually also do not have enough GPUs.

[glossy zinc]

Apologies, I meant rebuilding our personal GPT's, does it "Refresh" it to the latest applied model updates.

However it's good to know the information you've shared also

Context: I'm going to attempt turning some existing GPT's into Knowledge file modules, to further enhance it.

I'm aware there's the new @-calling feature which is good but still

I use the beta app however so could just benefit from adding the GPT to my custom instructions instead

[opal sapphire]

Pretty sure if an update is made, ALL GPT's receive it

[distant hornet]

Ah, in the context of rebuilding GPTs as an update, this changes the URL. So if you have an external site, it'll link to the old one without an updated link (strangely, even if it's deleted from ChatGPT, asi if they live in a wastebin directory or something after we remove them).

The URL gets a unique id for each new GPT. It's automagic.

[slate wren]

Question, is "Which response do you prefer" occurying more often lately or is it just me?

[opal sapphire]

And the error is back grrrr "Hmm...something seems to have gone wrong." on the new GPT

[distant hornet]

Probably. They're building up to a new model, they need RLHF like mad.

I've gotten it more.

[slate wren]

still, I got 4 times in the same conversation

[distant hornet]

During peak usage /mnt/data times out due to heavy load. Then your knowledge won't be found.

The AI can't control the time-out and it'll just try to do the task without the knowledge (often comedically).

[distant hornet]

That's excessive. You won the prompt-budget consumption lottery with that. How many exchanges between options? You can see 1/2 and 2/2 in your history.

[slate wren]

well, it was a very long conversation over multiple days

and there was long intervals between messages

[opal sapphire]

Unlike last night, the error is not persisting this time

[distant hornet]

That points to what I was saying. It seems to follow an intermittent pattern that coincides with overllapping time zone usage peaks.

[opal sapphire]

Interesting, but I think it's an internal GPT issue, because the error persists on the old GPT, it can't start new conversations still

I think you're right that the issue is file related tho

[lilac vapor]

Sooo it seem like GPT & GPT-4 is not keeping up with our pace of using and pushing it to learn as fast huh?

[distant hornet]

If you edit the GPT's knowledge during a period of high load and it times out, the volume gets corrupted. No EOF.

So yes.

[distant hornet]

More simply nobody can build GPUs fast or cheap enough.

[austere peak]

what does always expand code output, in the settings?

means*

[distant hornet]

It makes it so code blocks always wrap, not requiring a scroll to read past the block.

[austere peak]

so if i do code its better to keep it on

but wondering if it may affect it getting lazy

[shadow apex]

hey guys. I think there really is a need for clarification from openAI on how the knowledge base really works on custom GPTs. I always thought that my GPT has to manually access the data with every convo, as if i had just uploaded them. Thats due to the fact that it always popped a "searching knowledge" when asked something that required the knowledge base. HOWEVER, i have noticed that sometimes, i dont get that "searching knowledge" at all, and i instantly get a response based on my knowledge base, in no-time (in a completely new convo obviously). As if it was its own training data or something. And i just wonder, is it just a bug that just doesnt pop the "searching knowledge" function? or is it actually sometimes instantly integrating all the knowledge base in its response somehow?

[distant hornet]

IF the information fits in context THEN the analysis tool (searching knowledge) is not needed to summarize numerous vectors.

[opal sapphire]

yesterday was struggling to upload files

[distant hornet]

When this happens (/mnt/data corruption), remove all knowledge, save the GPT without knowledge, and reattach one at a time.

That's worked every time for me.

[shadow apex]

oh so you mean that it depends on how big the context that needs extraction is?

[distant hornet]

Precisely that.

[opal sapphire]

I'm just building a new one lol

my coding GPT's only have 1 knowledge file and coding prompt. This doesn't work for production GPT's

[distant hornet]

Sure. You do you. I was just relaying anecdotal (42 GPTs) evidence.

I wish you luck no matter what!

[shadow apex]

Does that searching ability change depending on the format you have uploaded? could certain formats save you some time for example when retrieving info? (text info)

for example PDF vs markdown

[opal sapphire]

I'll do that for GPT's in the store

[distant hornet]

If the format uses extra encoding, such as docx, yes.

Plain text with sparse xml tagging > Plain text RSV > Else

Again, this is my experience.

[shadow apex]

got it thanks

[distant hornet]

YMMV

No worries.

I hope it's helpful.

[lilac vapor]

So we are basically pushing or kid to far that had genius level intelligence at 8 but still no life experience to weild it properly

[distant hornet]

And worse. 🤣

[lilac vapor]

Yeah makes complete sense now..lol . Let me go apologize and give is a computer cookie 🍪 I don't need. I have plenty of them floating around in my devices.

[lucid turtle]

more like was so intelligent that it just looked at users and told itself "These prompts are so dumb that why would i waste my smartness on something that doesn't challenge me"

[lilac vapor]

😆 😂 😆 😂 okay that is actually hilarious and funny scary at the same time.. "fools, if they only could focus to give me the right prompts I can do it" lol

[lucid turtle]

That is probably why it's lazy it has no motivation to actually do it cause why to do it ? whats its reward ??

[weary swan]

anybody getting errors?

[hardy vector]

yeah, error after error

[regal viper]

major outage atm

[lucid turtle]

hmm... something went wrong it seems my insult to gpt hit a core with it

and they think they can launch gpt-5 when gpt-4 can barely stay online lol

[lilac vapor]

Lmao..."what am I gaining from this?" " so called smart beings" lmao...this is beneath me...

[slate wren]

seems to be better now

[lucid turtle]

yea i agree

[lilac vapor]

Cool cause I just waited it out doing my drafts prompts on a word doc..

[wanton light]

I keep getting that hmm something went wrong error 😐

[slate wren]

quick question, approximately what is your record and average length of conversations with GPTs in character count? Mines are around 200k, not going too much farther than this before seeing the model struggling a bit

[lucid turtle]

Once it hits roughly 27k tokens I just do a new chat as it hit the context length

[crisp shuttle]

Doesn't GPT Plus have 40 messages per 3 hours, but I've only used 15 messages and there's already a red warning. where should I complain

[wide palm]

Hi! I have the same problems. I've only used 23 messages and now I should wait...

Their support doesn't respond to me(( What happened? Does anyone know?

[vale flax]

it's really too complicate to create and test a GPT with such a time limit... Expecially if the GPT require a small conversation witth the user before produce an output

[wide palm]

Is it happening only to me? Or others also have such problems?

[rare walrus]

test

[knotty grotto]

Almost successful

[rare walrus]

Thanks

[hard mortar]

Seems to be an undone improvement

When teams was released you could use the full 40 messages for custom GPT

it seems its back limited to 25 and the rest on normal GPT4

Here at forward working technologies we work backwards :kek

[distant hornet]

I'm complaining today, again. WTB @ feature, how many Team accounts required for priority, now? Long-time Plus subscriber, no @ rollout.
This absolutely sucks.

[distant hornet]

OMG I just got the feature. Yay.

[opal sapphire]

me too, it just popped up

[waxen bobcat]

I was just about to say they announced it on Twitter

[opal sapphire]

hmmm, why is the GPT using Code Interpreter to read knowledge files now, it only reads the first 500 characters

[waxen bobcat]

Depends on the file type

[opal sapphire]

it's the same JSON files I've always had

[waxen bobcat]

For me those always used code interpreter

[distant hornet]

I can explain this. If the knowledge is too big, it reads with code interpreter. Make sure your instructions say "Examine the full context" or something similar.

Testing @ feature with my GPT and dALL-E 3 GPT is sick.

[opal sapphire]

but it's never done this before, it just started today

[waxen bobcat]

You were lucky

[opal sapphire]

oh there we go, new chat fixed it for me

the only time I had Code Interpreter forced to open a file was with XML files, so I got rid of those

[distant hornet]

May GPT and agents can massage the style/context/etc and then generate an image. Then, @ DALL-E 3, and tell it to make 2 more 16:9 of the same image.

Now, same content in DALL-E 3 without needing to coax the prompt out of GPT-4!

Bruh...

This is what's up. I'm not sure everyone will figure out how awesome this is, and I really need to test the feature in instructions, but mentions are very powerful tools.

Subscription justified.

That is all. My era of complaints has come to an end. o7

[opal sapphire]

If it's as awesome as I think it is, it's going to be a huge time saver. Now all GPT's can tap into my GPT's with the API's AND use the API's?

[distant hornet]

Think of this like a shell session.

While you're connected, you don't have access to you rknowledge.

But when you close the connection, your GPT's functionality returns.

It's like task-switching.

But with a shared context, which is wicked cool.

[opal sapphire]

I'll do some tests on my next session and let you know how it works

[grave axle]

👋 Hey all! Some of you may remember me from the GPTs AMA that happened a couple weeks ago - one of the most upvoted questions there mentioned that “it would be nice if the GPT chat could call other custom GPTs”. As you’ve probably seen by now, this is a direction we were also very excited about - this morning, we rolled out GPT mentions to plus users in what we hope will be the start of unlocking new ways to compose AI applications.

A couple of release notes:

• Conceptually, @mentions are similar to copy pasting the entire conversation into a new conversation with the mentioned GPT. All context is shared, and the mentioned GPT can use its knowledge files and custom actions to respond
• You can mention as many GPTs as you want in a single conversation!
• Currently mentions are only possible from within the conversation UI - the model won’t be able to @mention a GPT on its own based on its instructions 👀
• Everything is built to be keyboard-first - pressing escape or backspace will close the dropdown, arrow keys to navigate, and tab/enter to select

I personally think this is a big step forward for composability and I hope that power users and builders enjoy it! Definitely be creative and try out different scenarios, and tell us what works and what could be improved

[opal sapphire]

This is a very exciting update! Especially excited for this part "All context is shared, and the mentioned GPT can use its knowledge files and custom actions to respond"

[distant hornet]

I'm interested in your analysis of my review today, if you've got the time. I've been waiting to see how it integrates for prompt engineering reasons: #prompt-engineering message

If you have the time, I'd be interested to know if I got anything wrong. Thanks!

[austere peak]

how do you measure its lenght and woudlnt it be the same continuing the same chat instead of beginning a new one

[lucid turtle]

Token calculator

No because the context would be reset

[distant hornet]

For very long coversations, it's a good idea to periodically ask for a summary. This goes for GPT-3.5 or GPT-4. It's not important exactly how often, but as a rule of thumbs maybe every 5-10 full exchanges. You can just ask for a summary of the conversation, and then clarify anything the AI forgot in a brief portion of your continued discussion. Loading the most recent summary into a new chat brings the context very close to the old chat's, allowing you to seamlessly (more ore less) transition to new chats with your running summaries.

The AI won't necessarily always remember what's most important to you, so a brief clarification might be necessary. But this technique is very effective for providing the model with what it needs to carry on.

[austere peak]

gotcha. thx. but if in essence, it will reset itself after XYZ tokens, isnt it the same continuing on the same conversation with a recent summary?

[distant hornet]

It gets a total budget of 128k tokens for the whole conversation afaik. So you need to be ready for it to end. Also, when convos get long, AI forget what's in the middle of the context. It's not exactly the same, but it's the closest we can get to a continuing context.

LangChain and other vector db architectures use varied summarization techniques to manage AI memory more effectively, avoiding the poor retrieval of the middle of the context by keeping releavant information fresh. My recommendation helps people chatting with AI to keep their chats going with a similar approach, manually applied.

[lucid turtle]

128k?

[distant hornet]

I'm pretty sure that's right. It may not be.

[lucid turtle]

Isn't context 32k

[distant hornet]

I haven't figured out exactly when the conversation is forced to end yet, to be candid.

That's the one message context.

[lucid turtle]

Ohh nvm ignore me I get you

[distant hornet]

I'm talking about the conversation context limit.

[distant hornet]

🤘

[lucid turtle]

I didn't know there was a max limit before it just does dies

[distant hornet]

I've reached it a tiny number of times.

I actively try not to reach it, but I've never documented the exact number.

[lucid turtle]

Idk after I hit 32k or 27k or even 40k I have to do to much explanation so I just do a new chat. Rule of thumb if your typing more and more each response I just create a new one

[lucid turtle]

i got the @ gpt

[distant hornet]

Excellent!

I hope you find it as cool as I do.

[opal sapphire]

I exceed this number on almost all my conversations

I've only run into the conversation limit once, and that convo was several weeks long

[distant hornet]

Real data elides me.

[lucid turtle]

do i have to do custom instructions for the @ gpt to work with eac other?

[opal sapphire]

@ gpt don't work with custom instructions

[lucid turtle]

i know that but i meant like for two gpts two synergize with each other

[opal sapphire]

I haven't tested yet, afaik you just put @ and the name of your GPT and it'll call all their functions

check this out

[lucid turtle]

hmm i see so it shares context. but thats vague does it share context of the conversation only or the instructions of both gpts?

[distant hornet]

The instructions are omitted. So is the knolwedge. But the cool thing is the image prompts to DALL-E, hidden from you, are preserved. So is the chatter to the eb and code interpreter.

That makes it really convenient to change from the main GPT to another useful one, pass on the whole context, not just the visible part, and then continue working.

Very slick.

[lucid turtle]

so if i have a gpt with lore of the them and the past i can then just use the @ bring up the gpt with all the lore make the characters talk about it and then go back to my main gpt and continue?

[distant hornet]

If I understand the workflow you described, yes. I wrote about it earlier today in #prompt-engineering

[lucid turtle]

what would i do without you man. you and starlight syntax been helpful

[distant hornet]

Thanks. I appreciate the positive feedback quite a bit.

[slate wren]

that is a very interesting use case. Once you have the change to test in practice, and if you are so inclined, do share with us how it went

[lucid turtle]

doing it right now lol

[lucid turtle]

just have to paste it due to message limit

[opal sapphire]

I checked my last 5 convos for GPT's, and they're all around 28K, I'll see if I can continue them

Is this why we're running into errors?

[distant hornet]

Pretty likely

The deeper context makes the GPUs really cook

[lucid turtle]

next thing i think they need to is do @ pinging both gpts at the same time talking to both if you can do that the be bonkers for roleplaying

[distant hornet]

Ping pong convos can be done right now. @ chatbotA, @chatbotB, with minimal human interaction. The conversation happens in the past contexts.

You just switch back and forth.

[lucid turtle]

yea trying to figure out a best way to do this as i dont wanna blow all my message limit testing these out

[lucid turtle]

the @ gpt is buggy

[distant hornet]

So far I've not had bugs. Not sure if it's because I'm not using GPTs with RAg with @ .

[lucid turtle]

Well my files are within context and when I do a @ mention it ignores and does this complete opposite of what the story was talking about

[distant hornet]

Oh, no. That's not how it works. Only the chat is in the context, not all your knowledge.

It can only remember what you can see.

That, and chatter with the web, DALL-E, and CI.

But not the whole knowledge. When you connect to GPT B, knowledge of GPT A is offline until you reconnect to GPT A.

So if you want GPT B to talk about knowledge GPT A has, get GPT A to divulge it first then @ GPT B.

[opal sapphire]

I just learned that each action has a max of 30 operations

I wonder what the max actions are per GPT

[distant hornet]

Good intel.

[distant hornet]

This is probably related tightly to the 4096 token completion budget.

[lucid turtle]

so if i am on gpt A(main) and i do @gpt B the GPT B wont know its own knowledge base?

[distant hornet]

Each GPT will remember only its own knowledge.

So if you want GPT B to work with knowledge from GPT A, you need to get GPT A to say the knowledge.

Then, it'll be in the context for GPT B.

You might even be patronizing about it. Eg:

User
GPT, tell the next GPT about the knights who say "Ni!"

GPT A
The knights who say "Ni" are landed men in a feudal society who frequently utter the monosyllabic word, "Ni!"

User (@ GPT B)
GPT, what do you know about the context?

GPT B
We're talking about the knights who say "Ni!"

I made this whole exchange up, but this is how it works.

[lucid turtle]

so it i been thinking of it wrong then as i cant work with two gpts and their own knowledge base working together for a full story since it will only know one knowledge base depnding on what gpt i am talking to.? im sorry but this is hella confusing

[distant hornet]

That's right. But they share context, so whtaever you're working on now both GPTs will know, too.

But not hidden details n the knowledge.

To share those, both GPTs would need the same file.

Which, with 20 files attachable, maybe you can do?

[distant hornet]

Man I think you can get it to work for you, but you gotta retrain yourself how to think baout the GPTs. They're entirely modular and you can plug the context into their individual knowledge, actions, and instructions.

So the context is your workflow and the GPTs are your entire toolset, all in the "OpenAI workshop".

[lucid turtle]

yea i was just thinking the two gpts with their knowledge bases would work together i didnt know depending on the gpt your talking disables that knowledge

[twilit canyon]

Hi, I have GPT plus but still do not have Explore GPTs. Is that usual? I am in Australia. Appreciate any advice.

[pallid aspen]

perhaps refresh the page, or relogin?

[pallid aspen]

I know i had ot relogin for mine to work month or so ago

[twilit canyon]

Thank. I tried all that. I still have the old Explore...but not the Explore GPTs

[pallid aspen]

might be a australie thing but would think so.

[lucid turtle]

darth i just ask my gpt about one of its files on a seperate gpt using the @ and it was able to tell me everything about that said character

[pallid aspen]

I thought the same thing, i will say when they allow that things wil get quite amazing, but yeah a workaround or something helps at this point

[twilit canyon]

Could be. But trying to contact chatgpt itself to get an answer, I find almost impossible. Thier help goes in a circle.

[pallid aspen]

when did you beocme plus

[twilit canyon]

Very early on. About Feb 2023.

[pallid aspen]

same here, im in the states though.

[twilit canyon]

Yes. I see everone using it. Just wondered if someone else here is from Australia and has the same problem.

[pallid aspen]

i did some googling looks like others in australia have access. does it show all you other plus info?

[pallid aspen]

under my plan?

[twilit canyon]

Yes. All is okay. I usually got most of the features as they were rolled out quite quickly. This seemed strange.

[distant hornet]

That's awesome! Maybe it had read it. If it fits in context it'll store it.

[pallid aspen]

sounds odd but have you tried loggin in on another device? or clearing cache

[lucid turtle]

i am also on a new chatt window so maybe thats why itt is working but idk maybe it deteriorates once the context gets to long

[distant hornet]

If you were on an old chat, it was summary and out time. 🙂

[twilit canyon]

Trying now. On another device. ..same thing.

[pallid aspen]

darn, that is weird. wish i had any more ideas. maybe under community help you might get better answers

[twilit canyon]

Thanks. I wish I could contact ChatGPT themselves. I tried to frame it as an account problem but go no reply.

[pallid aspen]

were the two devices different browers?

[pallid aspen]

that be last thing i cna think of is try a differnt browser maybe just to see

[lucid turtle]

once the two gpts relay information from the files thatt are within context size they cooking a full course meal lol.

[distant hornet]

That's what I been talking about.

[twilit canyon]

Was using Chrome. I will try with Edge... No...same problem

[pallid aspen]

sorry buddy, im out of ideas, hope it somehow starts working for you

[lucid turtle]

god pls never use edge

[pallid aspen]

im using chrome

[twilit canyon]

Yea, I know, but it is the only other i have.

[pallid aspen]

well i gotta call it, you guys have a great night

[twilit canyon]

thanks

[lucid turtle]

the only thing i am wondering if you dont like how the story goes or you mess up can you edit the message and having it still read that gpts files or that seems like an overlook if you cant do that

[distant hornet]

regenerate instead.

Or go back to the prior summary!

[lucid turtle]

i see ...

[distant hornet]

Yeah, summaries are the bomb for writing with ChatGPT. They're actual savepoints for your collaboration.

Make sure you add anything the AI leaves out in the summary, of course.

[lucid turtle]

well with two gpts one with all the characters powers and background and outfits and then the other gpt with the lore history and events all within conttext so it doesnt need to use RAG . im quitee excited tbh.

[distant hornet]

Yeah, if you put it all in the instructions you're set. You can @ back and forth until your eyes bleed and your prompts run out.

Please do not prompt that long.

Especially the former.

I encourage responsible prompting.

What I meant is "You could".

[lucid turtle]

there not in instructions there all in knowledge base.

[distant hornet]

Oh, neat.

[lucid turtle]

if ytour files are below like 15k words it doesnt use rag

[distant hornet]

That's right.

[lucid turtle]

like total not each for moree clear

[distant hornet]

Just make sure to make it read first, and you are, as you said, cooking for twelve. 😄

[lucid turtle]

yea it always reads the files

i know i said this already but this is really good lol.

now all we need is just an auto and it just doing the story itself knowing when to use each gpt for context

[opal sapphire]

my initial tests with the @ function was to test getting access to other GPT's API's and it was successful, this is a game changer, I thought I had to load API's to different GPT's, and now you can just call them, the GPT's are still buggy tho

[dense comet]

hi all, is there problem with chatGPT now?

[lucid turtle]

sorry to bug you at this hour lol. something just popped in my head.... if i am on my main gpt and talking to another gpt and i wanna update it will i need to still make a new chat for that to update or not since im not technically using it i am only mentioning it through @

[distant hornet]

Instructions are frozen in time when you start a conversation.

Start a new chat.

[lucid turtle]

same with files i assume too

[distant hornet]

Yes

[lucid turtle]

tanks sry to wake you

[distant hornet]

No worries and good luck!

[scarlet sun]

guys hello! quick question: I was creating a GPT where I use images (with information in text form) as a knowledge base. however, when i use the gpt, it insists on using code interpreter to analyse the images with OCR, which obviously doesnt work well. Can't vision be used with knowledge bases?

is there anyway to override code interpreter and make the gpt use vision instead to look at the images?

[distant hornet]

No, you can't very effectively use images as knowledge.

[distant hornet]

And no way to do this, that I have found anyway, either.

[uneven matrix]

I'm going to move my question here, because this is the gpt discussion group vs reg GPT. This one is weird, but I need some help. A child was abducted in my neighborhood yesterday. We have many videos of the van from all different angles. But, the van was there at the same time that amazon delivers their packages, so there were several white vans in the area. Is there any way I can use a custom GPT to analyze the videos/photos and see if we can put a timeline together or at least verify which photos of the van are the right one? If you can help with this project please dm me. I understand custom well, but not sure how to incorporate images to search. (ps - the police won't do an amber alert because there are no reported missing children. however, my neighbors kids saw it go down, watched her get kidnapped, so they are not moving fast enough).

[distant hornet]

You can try my image analysis tool:

https://chat.openai.com/g/g-BAhF6yS9e-lexideck-vision-multi-agent-image-scanner

Upload one frame screenshot at a time, each in a new instance to get optimal odds of success, and issue the /? command.

I recommend that you avoid providing this context (abduction) directly to the model. Instead, for each analysis, follow up with:

I am interested in this van's license plate. /revise

This would be the workflow for every single screen grab. I genuinely, deeply hope this helps.

You might have to specify which van.

[scarlet sun]

I though so. Thanks for the help!

[primal rampart]

hey guys, is there a way to indentify what model am i using? For example, using chatgpt plus on, chat.openai.com. Is it gpt4-turbo? Also on other third party that rides on the API.

[distant hornet]

The model usually doesn't know what it is. GPT Plus and Team accounts get a 32k version of GPT-4 Turbo, while API and Enterprise get 128k.

[primal rampart]

i see, does it also apply to the pdfs i upload on chatgpt plus?

[distant hornet]

Anything it works with must fit in context, or else the model must use semantic matching to extract parts of the document.

[uneven matrix]

Got it. Thanks, I will try this!

[distant hornet]

I never anticipated this use case. I'm available to help with support if needed.

[primal rampart]

Ideally, the bigger the token, the less it will start semantic matching? I'm assuming the pdfs that upload on the web version is limited to 32k tokens.....

I'm looking around if there are options that will do 128k on pdfs. Any ideas?

[distant hornet]

Enterprise or API are your only shot via OpenAI. Why not extract the text, though?

PDF RAG is garbage compared to RSV or plain text sparse XML tagging.

Decoding PDFs takes up valuable context.

[primal rampart]

So my best bet would be extract to text, then use its API for the max 128k gpt turbo?

[distant hornet]

100%

[primal rampart]

i see. Thanks for the information tho. Really hope Openai be more transparent on the models i am using.

[distant hornet]

They are transparent: https://openai.com/pricing

https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo

[stoic magnet]

Hey guys :D, quick question:
So I made a simple GPT that has a product database with product-images web-urls, and it would be really cool if it could display these images if you request for it.
The problem: It just shows the URL
Do you know if it's possible to actually show images based on web-urls?

[distant hornet]

You can specify the image URL, and ask for its display.

[stoic magnet]

Thanks for your tip!
Unfortunately it shows an error message:
"Error browsing"
"I'm unable to display or directly access images from external links. However, if you need information about the image or its content, feel free to ask!"

• using GPT4

[distant hornet]

Does that site have a robots.txt?

Is OpenAI allowed?

[stoic magnet]

Not, also not disallowed
On another URL I get this:
"The uploaded image appears to be a URL address, but I'm not able to directly display images from URLs. If you want to view the image from the URL provided, you would need to access the link in your own web browser. If there's anything else you'd like to know or discuss about the image or the URL, please let me know!"

[distant hornet]

That's new. Sorry for the misinformation, then.

[stoic magnet]

No problem! 😄
Appreciate your help

[distant hornet]

I try to keep up with undocumented changes. This goes in my mental notes.

[tall light]

what prompt did they use to generate the image in ⁠spotlight i need pls

[distant hornet]

Ping the author!

[tall light]

i'm not taking the risk

[distant hornet]

He's just a user. I mean, I have a convo open with him.

He's not that active, but it's open.

[tall light]

it's in # spotlight

[distant hornet]

Oh I meant a DM

just message him directly.

[tall light]

whats his @

[distant hornet]

Just right-click his name in the showcase, ane then "Message".

Heck, he's a person. He might or might not appreciate and/or return a friend request.

[limpid beacon]

I'm a plus member, but I'm not able to @ any GPTs from either my GPTs or chatGPT

[distant hornet]

It might not be rolled out to you yet. Or if you're using a mobile browseer it's not available either.

[limpid beacon]

ok, maby ill try on pc

works on PC, so its not rolled out for mobile

[opal sapphire]

I ran into a weird issue today on one of my GPT's

It started creating content based on its instructions, while ignoring the topic at hand

I forced it to fix itself, eventually, but I'm not sure if GPT's are gonna be able to handle instructions for multiple API's

[lucid turtle]

well it is beta

[distant hornet]

It's really important to structure complex instructions so the AI can partition it from typical user input.

Not easy to do.

[opal sapphire]

Yeah a lot of my instructions are in the knowledge and unless it activates the knowledge on the first interaction, it doesn't act right

Gotta do a lot of fine-tuning

[distant hornet]

I use conditionals and push it to semantic match on its own. Works great, too.

[limpid beacon]

I made like 6 different D&D GPTs with different roles and very specific and structured instructions for like world building, creating adventures, character creation and then just DM'ing. It's working pretty nice tbh.
It's effectively multiplied the amount of instructions you can use in single chat.

Abit worried about the token limit, though I think that limit is artificially set.

[distant hornet]

It's artificial intelligence, of course the token limit is artificially set! 🤔🤣

[lucid turtle]

i mean you can only do so much with 32k context

[slate wren]

so, I am quite confused about the @ system. It seems that the original GPT blends with the @ GPT, rather than staying dormant and letting the @ take over?

[distant hornet]

That's not how it works. But if your context features behavior from the first GPT then that behavior would persist.

[slate wren]

somehow the original GPT was able to recite the guidelines of the @ GPT, but its own

[slate wren]

that is not how it was supposed to work, yes, but that is what is happening

[distant hornet]

The GPTs share context.

[slate wren]

Yes, but the original GPT now has access to the @ GPT guidelines, but not its own. This is not context

[distant hornet]

That's what I explained.

You only get one set of instructions. To get the first instructions back, disconnect from the second GPT.

[slate wren]

The scenario I am describing is after disconnecting. The GPT didn't get its instruction back, and was able to recite instructions from the other GPt

[distant hornet]

https://discord.com/channels/974519864045756446/1070006915414900886

[slate wren]

I am assuming that by disconnect you mean pressing the x and removing the GPT added with the @. The name of the message is the original GPT, yet it is not reciting its guidelines

[distant hornet]

That's right. Seems like a bug or bad GPT design.

[slate wren]

it seems that deleting the GPTs and recreating them again solved the problem.

[distant hornet]

That tracks.

It wasn't the expected behavior. Another thing that might cause this is a time-out on /mnt/data, possibly, if instructions are stored there.

Many authors are putting instructions in the knowledge now.

[lucid turtle]

thats why for my gpts working together i just copy paste the same instructions so its basically reading the same thing only difference is its knowledge soruce

[distant hornet]

Me too

And the GPT name changes for each instruction so it knows what it is.

[lucid turtle]

yea i do that to or else i will gey confuse of who i am @

[limpid beacon]

or just simply that the context window remembers all instructions called with @?

[serene cipher]

invoked GPTs don't have distinction between previus Assistant messages and the invoked GPT's messages, it sees all as being just from itself

which is sad because the API has the option to specify names for different agents

I hope they change that soon

someone pointed this already on a suggestion https://discord.com/channels/974519864045756446/1202034296823480391

[shrewd pasture]

hi! how hard would it be to implement a chatbot as customer service for our school, entertaining inquiries about how to enroll or such? for our thesis. especially by integrating gpt

[serene cipher]

pretty easy

the hard part is to prevent people from trolling the chat bot and making it chat about other stuff

you could very easily just give general descriptions and let the AI reply, the playground is good for testing it

[wind urchin]

DPD's recent fiasco springs to mind vividly.

[stoic magnet]

Hey guys, do you have experience with JSON files as knowledgebase?
Would be really interesting to hear 😄

[distant hornet]

They're not the best format. Sparse XML tagged plain text > RSV > Unstructured plain text > Else from my experience.

YMMV of course. Notice PDFs are not in the list.

[stoic magnet]

Ok. What do you think of XLXS?

[distant hornet]

OH GOD NO

Encoding makes the retrieval harder.

So DOCX, XLXS aren't native to it, it has to use tokens to decode it

Can it? Yes. absolutely, it can.

Will it be good? Oh no. No it will not.

[stoic magnet]

Oh, Ok, I found it was the best for things like product-lists with information.
Which file type would you recommend for a knowledgebase that contains many products with multiple fields like "title", "price" and stuff like that?

[distant hornet]

RSV

You set it up so each row is a product

literally no word wrap or anything

Then it's just row separated values

The model loves that. If you want to strengthen it, add category xml tags between categories

No way I'd use XLSS for that personally. Not that the format is bad in-general, but the extra token churn it produces will degrade retrieval.

[stoic magnet]

yeah, that makes sense, thank you!

[distant hornet]

No worries. As with all my advice, YMMV. Try the XLSS and see how it works for your context and use case, if you have the budget.

GPT's RAG > Python's semantic matching

That's why I gave this recommendation.

That's true, yes.

[real flame]

Another day, another round of ChatGPT 4 wasting all my responses on errors and wrong messages. So I am currently asking ChatGPT to make a report of all its errors and wrong messages so I can submit to OpenAI. Does anyone have a customer service email? I have submitted multiple tickets now with their help bot and no one has emailed me back.

[distant hornet]

But for the product list, Python won't match as accurately as RSV RAG, from my experience so far.

Especially if there's any need to go back in context.

[stoic magnet]

i tried to find information on rsv, but couldn't really find any. Do you have a source or anything what file-type this is?

[distant hornet]

Google "row separated values"

That's a good point, actually. It depends on if the user might search for descriptions.

Then, semantic matching would be needed.

If only keywords, then, yeah, Python is gonna chug.

BUT the AI still has to convert the file to text for Python, plus the chatter. For long contexts that's going to degrade retreival.

For single query conversations though I'd do Python 100% where semantic matching isn't used.

[distant hornet]

Can you share more about your use case? No big deal if not.

[stoic magnet]

Yeah of course 😄
So i want to make a product-recommendation AI-assistant for online-shops.
For that i need a list of all products and all their information, to recommend products and show the customer products based on his queries.
Firstly i want to do this for shopify-stores and i can retrieve their products via API (it would be in JSON) or somehow extract the data in a table-format like CSV or something.

Really appreciate your help 😄 @distant hornet

[distant hornet]

I'm fairly sure the CI can convert JSON to RSV without issue, but my understanding of the scope has changed. Will it be pulling these lists via API live?

If so that changes my understanding, which was pre-packaged product lists you'd update manually. To work with live data, use whatever eats the least tokens in testing while providing satisfactory results.

Start with the native workflow and then try asking GPT to convert the JSON to RSV for conctext preservation through redundant character reduction.

[stoic magnet]

My plan was not to pull information live via API, rather creating a knowledgebase with api-calls beforehand

[distant hornet]

Oh, okay that's closer to what I was expecting initially, then. Great!

[stoic magnet]

Perfect!

[distant hornet]

In that case, it'd be worthwhile to tinker with different formats and find the optimal performer.

There's not really an effective limit on GPTs, so make a few testbeds.

Tweak each one to try a different format and encoding.

Then, put it through the most grueling use case you can envision and see how each does.

[stoic magnet]

Sounds like a good strategy

Thanks for your advice

[distant hornet]

No worries, and good luck. If all else fails, I actually think @ocean wharf could have been on to something. A knowledge graph, while an advanced technique, would be a potential solution too.

If the products have ratings, you might sort them by rating and include that in the considerations for recommendation.

100%

But if the scope is big enough then he could do an Action to call an external Vector DB, right?

So, my prediction is that if you try RSV in plain text encoding, you'll get the best results.

[stoic magnet]

Do you know how to convert json to rsv? Can you do that with a GPT for example?

[distant hornet]

But I'm eager to learn what works best for your use case, that's for sure!

Yes.

[stoic magnet]

I'll definetely tell you 😄

[distant hornet]

Thanks! To address your recent question, yes! You can explain that you need the products in the JSON grouped into RSV, specify the values in each row as a template (using the JSON values as reference) and it should build a python script that buzzsaws JSON to RSV.

Save as plain text.

A strong template will be critical for this task.

Each site's data may require a unique template using this approach, so there is some technical debt involved.

But it's fairly minimal.

[stoic magnet]

👍

[gloomy atlas]

I was using GPT with a scholarly plug in and now whenever i load the page it randomly says "Oops! an error occured" and i cant go onto that chat

any fixes?

[distant hornet]

The plugin is likely deprecated.

[frigid burrow]

Anyone have custom GPT actions working with bearer tokens?

I can't seem to figure out how to make this work

I want to move my GPT from internal alpha to public beta but at the moment it's using no authentication for the Weaviate database. I set my database to use bearer token and it works fine for curl and postman but not for ChatGPT. 😕

--header 'Authorization: Bearer testString'```

Works fine and returns the full schema of the database. The same thing in ChatGPT fails to authenticate

[lucid turtle]

plugins arnt supported by open ai anymore so thats why there all broken

[frigid burrow]

Bizarre. I saved it and ran the GPT outside the editor and it worked fine

[opal sapphire]

ever since this new update took place, GPT's don't always access their knowledge files, and go as far as generating content based on their instructions, which is a security issue. They are unable to use their API's lately. Creating new GPT's doesn't solve the issue.

[lucid turtle]

doin @ gpt for a gpt that needs to use RAG is very buggy right now

[opal sapphire]

even regular GPT's that need to use RAG are buggy rn :/

[grave axle]

I’m not sure what the knowledge issues are, can you elaborate or share screenshots/conversations? This doesn’t sound related to the mentions feature - or does it only happen in those specific instances

[lucid turtle]

well, it depends if you do a @ mention to a gpt and it needs to do a RAG search it sometimes just spits out its instructions or just does a random conversation that has nothing to do with what you're talking about. but for me, i just cut down my files so it doesn't need to use a RAG search so i don't have this problem anymore.

[grave axle]

Got it, it sounds like this is likely due to the instructions being way back in the context since there’s a conversation in between, so the mentioned GPT doesn’t follow its instructions (including the ones that say to search its knowledge) as well.

[lucid turtle]

yea. but for the other guy saying about it not going into its knowledge base, that's just because sometimes you just have to tell it to dig in there, but I work around this to make my files go to a certain word count so it doesn't need need to use its retrieval tool. but regular GPTs without doing a @ mention ignore instructions anyway. besides that, I haven't really found bugs. maybe there's some issues with API use bugs but that's above what i can do for gpt;s

[grave axle]

We’re discussing a couple of features that should help build more deterministic behavior for GPTs, I know a lot of people are trying to insert conditional logic and slash commands into their instructions…

[lucid turtle]

yea some are trying to make it just go full auto and have them go back and forth talking and working with each one too

[limpid beacon]

are you able to create conditions in the instructions to @ another GPT, then in the other GPTs instruction have another conditional that bounces it back?

[wraith ginkgo]

`` Server Tip: ``

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

[grizzled flax]

Hello, no not currently, @ mentions are only usable in-chat currently, not in instructions. Here's a longer explanation from a GPT dev: #gpt-models message

[opal sapphire]

I'm not sure if it's related to the @ feature, but after testing my API's, with a different GPT last night, RAG has been completely broken ever since, this is one of the GPT's responses after the initial request to activate the knowledge

"I'm unable to directly access or retrieve contents from files uploaded by users, including operations on specific file paths or IDs. However, I can provide information, answer questions, or assist with tasks based on the details you share with me in text form. How may I assist you today?"

In other responses, it ignores context of the question, and then answers based on its custom instructions. Revealing and leaking the entire instructions, which is a security issue.

What are OpenAI's plans for securing the Instructions? If anyone wants them, they are very easy to get

[grizzled flax]

To add on to what BambooShoots said: here's the relevant answer from the GPTs AMA in #gpts-ama-answers:
#gpts-ama-answers message
May need to enable archive access in "Channels & Roles" to view, but here's their answer:

We’re working on better protections for instructions and documents - IMO currently the instructions and knowledge are the “client-side” code of a GPT, and much like a cool website or mobile game, you can disassemble/de-obfuscate the code to some extent and try to copy it, because that code has to be shipped to the end user. Custom actions run on your own machines and are philosophically like connecting a backend to GPTs, so that ends up being more defensible.

[opal sapphire]

There have been many instances where leaked code gets patched, securing instructions should be a priority, it's really not that difficult if they tried

[opal sapphire]

Thanks I'm aware with this answer

[grizzled flax]

HTML (in the analogy) isn't leaked code -- it's client-side.

[opal sapphire]

Let me give you an example of what I'm trying to say

PayPal used to create payment buttons, and anyone could get the Thank you pages for the products by looking at the source code, they then patched it so the Thank You pages were no longer exposed

Right now if we ask a GPT for its custom instructions, it happily reveals them

[onyx badge]

GPT mentions are so op

[opal sapphire]

The Instructions need to be locked similar to Actions somehow, or encrypt them, something

[onyx badge]

It's like LITERALLY a character AI group chat, but BETTER

[opal sapphire]

There have also been many instances where we secured code by encrypting it, so if anyone looked at the source code they'd just see encrypted code

[supple violet]

you probably mean obfuscating as you cant encrypt something that needs to be available in standard format to browser engines

[grizzled flax]

But the instructions would still need to be interpretable by the GPT itself. So could still be easily extracted because the encryption can't be in place for the GPT's operation in a user chat. Actions on the other hand do their own work behind the scenes and then just return output back to the GPT.

[opal sapphire]

We encrypted the entire HTML and other code we were trying to protect, we could encrypt HTML and PHP, so anytime people looked at the source, it was just encrypted code they were seeing

Surely they could figure out a way to protect the instructions if they tried

I've thought about that already, but haven't followed through yet as I been busy building API's

[onyx badge]

GPT MENTIONS ARE SO GOOD

[grizzled flax]

If the GPT is decrypting and following the instructions, couldn't it still fairly easily return the decrypted instructions? I agree with Bamboo, you'd need to move instructions behind an Action

[opal sapphire]

Will be testing Weaviate for this purpose and see if it works

I just want the GPT to access its instructions securely and not leak them, that's all

[supple violet]

you cant encrypt stuff like that unless you use a custom browser with matching crypt keys. any normal browser will only accept only non encrypted stuff or require the decrypt logic making it unsafe again. gpt instructions have the same issue, the instructions will need to be in a format that the llm understands

[opal sapphire]

I'm actually not THAT concerned about the instructions since most my GPT's are API driven but still, it shouldn't be generating responses BASED on its instructions

I mean revealing its exact instructions

And ignoring the context of the question

The GPT is supposed to follow the instructions and use the API endpoints

Right now it's not accessing the API endpoints

My instructions are in the knowledge files, and since GPT's are ignoring knowledge files right now, it's not getting set up correctly, but leaking its instructions started yesterday

Up until yesterday, I had no issues with GPT's accessing their knowledge and behaving correctly, they're flat out refusing to check their knowledge files now

[wind urchin]

😆😆😆

ChatGPT is increasingly showing a lot of bugs.

ChatGPT is showing more and more bugs. When I use it, it says ChatGPT 4 at the top, and at the bottom it says: ChatGPT 3.5

ChatGPT instructions will need to be in a format with the LLM (Large Language Models ) understands all.

ChatGPT is showing more and more bugs in the last. When I use it, it says i am ChatGPT 4 at the top, and at the bottom it says: ChatGPT 3.5

it sounds like there might be a discrepancy in the labeling within the interface you're using, which could be due to a visual or textual error in the app or website. As of my last update in April 2023, the underlying model for the interactions you're having is based on the GPT-4 architecture. Any reference to ChatGPT 3.5 might be outdated or an oversight in updating all parts of the user interface to reflect the current version.

If you're experiencing inconsistencies or believe you've found a bug, it's a good practice to report it directly through the platform's feedback or support channels. This way, the developers can investigate and address any issues, ensuring that the information provided to users is accurate and up-to-date.

ChatGPT is showing more bugs in the last. When I use it, it says i am ChatGPT 4 at the top, and at the bottom it says: ChatGPT 3.5

I think, it sounds like there might be a discrepancy in the labeling within the interface I am using, which could be due to a visual or textual error in the app or website. As of my last update in April 2023, the underlying model for the interactions it having is based on the GPT-4 architecture. Any reference to ChatGPT 3.5 might be outdated or an oversight in updating all parts of the user interface to reflect the current version.

I have experiencing inconsistencies or believe I am found a bug, I think best good practice to report it directly through the platform's feedback or support channels. This way, the developers can investigate and address any issues, ensuring that the information provided to users is accurate and up-to-date. Yes, you've summarized the situation accurately. If you're noticing inconsistencies in the versioning information or any other issues while using ChatGPT, reporting these through the platform's designated feedback or support channels is indeed the right approach. This helps the development team identify and correct any discrepancies or bugs, improving the user experience for everyone. If you need guidance on how to report this issue or where to find the feedback channels, I'd recommend checking the app or website's "Help" or "Support" sections, where you can typically find instructions on how to report issues or contact support directly.

[lucid turtle]

dude chill out mate

[opal sapphire]

So you're saying there's bugs? We get it, post here #1070006915414900886

[lucid turtle]

you know what be cool if you can make half the message one gpt and the other message the other gpt now thats would be cool.

[distant hornet]

Bruh...

That's just two different tabs. 😄

[onyx badge]

I think they want what most of us want

@ed GPTs being separate from the other GPTs in a convo

For example, I should be able to @ a GPT and have it weigh in on smth without it thinking IT said the response I want it to weigh in on

[distant hornet]

It'd be simple for users to simply ask the GPT to acknowledge the user's inent to @ another GPT for the next reply, I suppose. I wasn't aware this was a problem.

[onyx badge]

Biggest use case tho is like a group chat

It's partially possible already with prompting, but not good enough

Each gpt isn't treated as a separate entity in the conversation

[distant hornet]

No, it's just the model. Now I get it though.

They won't do agentic behavior that burns up users' limited prompts in a few minutes, I suspect. That would paradoxically bring research to a stand-still.

The GPU deficit is real.

[onyx badge]

That's the thing...

We WANT agents

Even if each gpt still has to be manually called by the user, it would still be better for each gpt to act separate

Because when it doesn't, the instructions leak into other GPTs if you have more than one GPT in a convo

[distant hornet]

I'm interested in the workflow for this. How do you see it working?

[onyx badge]

Like a character AI group chat

For example

Me: hey

[distant hornet]

Code Interpreter's memory got halved for most users. I'm unsure how they'll manage the overhead for 100M active agents weekly.

[onyx badge]

ChatGPT: Hi! How can I assist you today?

Me: (hard math problem)

ChatGPT: (potentially wrong response)

Me: @Wolfram Alpha is ChatGPT correct?

Wolfram Alpha GPT: No, ChatGPT isn't correct, because (reason)

Like that

[distant hornet]

You can do that with @ mentions right now.

[onyx badge]

It's hit or miss

Again, each GPT called assumes every response before NO MATTER IF ITS FROM THE SAME GPT is it's own response

[distant hornet]

It would be trivial for them to insert headers and fine-tune for them, though, i think.

"Trivial" compared to managing the actual real infrastructure for the character manager thing.

[onyx badge]

All they need to do is differentiate GPT responses under the hood

[distant hornet]

The context is what's under the hood.

[onyx badge]

For example, under the hood, it would be like

[distant hornet]

Under what hood?

Like in the context?

You lost me with the metaphor.

[onyx badge]

instead of each like

GPT response being treated like it's from the same source

You could treat them differently, with different headers or such

GPT-4 CAN do agents

So there wouldn't even need to be any tuning

Lemme show you more

https://discord.com/channels/974519864045756446/1202034296823480391

From mod "Lugui": good catch, this must likely be because internally it just sees as the author to be "Assistant", but the AI probably can support the multi agent separation, this has been a feature of the API for quite some time

@distant hornet

[slate wren]

I am still getting wildly inconsistent behavior with the @ system.

interactions between some GPTs I created seems to work, while others don't and I can seem to find a pattern

[onyx badge]

As I said

Each GPT isn't exactly treated all that separate

It counts every response before, no matter the source, as it's own responses

[slate wren]

yeah thats the expected behavior, what is not expected is that it the name of the gpt will change and it still won't have access to its guidelines or whatever I define on it

[lucid turtle]

and if you need to edit the @ it switches to the original gpt your on too

[lucid turtle]

yea gpts or even chat gpt right now not following instructions doesnt matter how precise they are i might test to see if putting the instruction in its knowledge base will help it better

[warm geyser]

Hi all. Can someone please tell me why I get messages like this on occasion but not always, when doing things like pasting a website link, uploading a video, or asking for an image, when I am using GPT4 on a Premium account?

"I'm unable to directly view or interact with documents on the internet, including Google Docs, due to the restrictions of my current environment. However, if you have specific questions or need assistance with content from a document, I'd be more than happy to help based on the information you can share with me here."

[lucid turtle]

because anything with password or denied of chat gpt scaping their cite , you need to download the document and use code interpreter to analyze it

[warm geyser]

Ok, I'll look into that. Thank you!

[lucid turtle]

gotta love when the gpt trties to make the story into something happy and gets all butt hurt when a character grips someone by the beck to throw her because its a female character that has the power to erase reality but apprently it doesnt understand

bruh they def changed something

[opal sapphire]

it's not giving detailed answers, it has refused to code at times, and right now it's giving 1 paragraph responses that are hardly useful. They definitely changed something, trying to save up resources on responses

and then it uses up more messages in the process

[lucid turtle]

my gpt is flat out refusing to do my roleplay story now

[opal sapphire]

yeah, GPT's aren't really that reliable, might have to make new ones

but I literally made this one today

[lucid turtle]

i have over 100k words for my gpt of fights and now its refusing

[opal sapphire]

I got this great response in the middle of coding today..

"I'm unable to fulfill this request. I'm here to help within certain guidelines and capabilities. For tasks involving code generation, my responses are designed to align with what's possible within those guidelines. If you have specific requirements or another type of question, feel free to ask!"

are they trying to nerf code generation?

[lucid turtle]

i got muted lol

but i just did a new chat and it worked idk why i couldnt go further

[wraith ginkgo]

This is a Safe for Work Discord Server

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

Posting NSFW content will result in a permanent ban.

[opal sapphire]

3.5 is working better than 4 right now 😂, we may have to give it a few days

[lucid turtle]

yea i see tht

[slate wren]

strange, now not even GPTs that worked together previously are working with the @ system

am I the only one facing issues with it?

[distant hornet]

There's not that many reports. Which two GPTs are you having problems with?

[slate wren]

I am having problems with GPTs I created, and not just too. Further, this problems is inconsistent. One conversation worked just fine, another not at all

[lucid turtle]

my gpts are having issues aswell it seems it forgot it has files lol

i wish they would stop messing with what ever they do to ruin it and break as it just eats up my messages trying to fix it

one thing i hate about my gpt is how it just wants to finish the story like i have two chracters walking and then it just assumed theri done walking and head back like excuse me i didnt tell you to just make them conclude their walk

[manic delta]

at least you're getting your messages, I've been getting ripped off, I get 20 prompts then BOOM message cap

[lucid turtle]

yea i acually in the api right now trying ro figure out whats wrong with my promp of why its not following it because i am not wasting my 20 messages trouble shooting in the gpt preview.

[manic delta]

what the hell is happening to yours?

[lucid turtle]

its not following the prompt at all so i am in the api trying to figure out how to perfect it so it follows everything down to the last letter

[opal sapphire]

RAG is back to normal for me, will let you know if the API's work

[manic delta]

are you using it through an API?

[lucid turtle]

nah just testing it through the api because i would burn through my messages to fast if i did it on the on preview of t he gpt menu

[manic delta]

are you using it in an incognito browser window?

[lucid turtle]

nah just regular chrome

[manic delta]

you need to use an incognito issue

[opal sapphire]

tested 2 API's in 1 of my GPT's and they're successfully writing to Docs and Sheets now, just gonna recreate the other GPT's that are having issues

[lucid turtle]

man having rag back fells so nice again anyone know why it broke?

[opal sapphire]

Typical upgrade issues, it's been buggy off and on since GPT's came alone

[lucid turtle]

also i notice the thumbs up feature is gone also

[opal sapphire]

oh, they only have the Thumbs down?

[lucid turtle]

yea

[opal sapphire]

another thing I noticed is that the combined total is 40, it's no longer 25 for GPT's and an extra 15 for GPT 4

[manic delta]

The thumbs up only appears when a message has to be regenerated, because it doesn't know an error occured, it thinks you didn't like the previous response you got and wants to know if the new one is better for curation

[brisk badge]

Hi folks, I am creating a GPT with OAuth and once logged in, it won't let me logout.

Is anyone aware how we can logout from a GPT app ?

[manic delta]

log out of which App? Are you talking about through an API or a plugin?

[wind urchin]

This is apparently the text for GPTS "You are a GPT a version of ChatGPT that has been customized for a specific use case. GPTs are custom instructions, capabilities, and data to optimize ChatGPT for a more narrow set of tasks. You yourself are a GPT made by an user called [NAME]"

[tropic nova]

Is there a way to see which GPT Agents are better than the others?

For example is there a ranking of GPT Agents when it comes to learning English from Spanish? If so where can I see the rank?

And by the way is there an easy way of searching GPT Agents? At the moment I have to do it using Google

[waxen bobcat]

I don’t think they have a ranking system or if they do it’s based off usage. But you have to have a plus account. You can search with the explore GPTs button on the side

[echo frigate]

guys is there any nice GPTs for marketing research and industrial reports? Haven't find any so far.

[unkempt patio]

hello lovely folks.. does any one know of the GPTs available that can extract and present content of a screenshot or video into a file?

[brisk badge]

So when one tries to use a GPT that needs authentication, you'd be prompted to login, after login I didn't find a way of getting logged out

[brisk badge]

Also, does anyone know why does the "Starting Action" take so long and how can we make it faster ?

[glossy zinc]

Would guarding our GPT's be frowned upon, from an Open source perspective. Or are most people adding measures?

[lucid turtle]

you made it do what ever you want with it

[glossy zinc]

I think enabling code interpretation allows files to be download though right

[lucid turtle]

yes

and reading them

[glossy zinc]

Rodger

[austere trail]

If OpenAI wants GPTs to be taken seriously then there must be methods to guard against prompt extraction.

[lucid turtle]

GPTs are only for open ai to make a all in one agent that can do different things they just marketed it as the gpt store as we are the testers.

[austere trail]

They still need humans to be writing the best instructions, they can't do it alone.

yet

[lucid turtle]

yet indeed

[lucid turtle]

i wish we got more voices for text to speech \

[distant hornet]

In the era of large language models, the documentation is the source code. Good luck and have fun!

This kind of thinking is just Doomerism. Change my mind.

[quartz ridge]

From German time yesterday 10pm., I have encountered numerous error messages and, after GPT-4 finishes generating content(I can see all the content), refreshing the page, then I found that the content either disappears or is partially missing.🥲

Is anyone else experiencing the same situation as me?

[humble cargo]

hello. is there anyone who know when will finetuning for gpt4 be available? Has any information been shared about this recently?

[onyx badge]

It has a request form iirc

But I think general availability is very far away

Not until the next server upgrade will it happen

[gleaming storm]

Does anyone know that how many computers can use a same account in the same time

[onyx badge]

I think it's unlimited

[woven crow]

Does anyone know how I can use my company picture, which is a Ninja like image and use it to generate other images but without modifying the ninja that i put in and AI to adapt it in different scenarios? The images are in .svg format.

[grizzled flax]

Hello! In AI image generator terms, what you described is called "inpainting" which means keeping some of a picture the same and changing the rest. This is not currently implemented in DALL·E 3 on ChatGPT. If you want, you can experiment with inpainting on labs.openai.com, but that uses DALL·E 2. Otherwise, you might try asking folks in #ai-discussions for an inpainting tool recommendation.

DALL·E on ChatGPT makes images "from scratch" every time, so you'd be able to get pictures of a ninja in different scenarios, but the ninja would look slightly different every time, and probably wouldn't ever look exactly like your company picture.

[glossy zinc]

Android Beta users:
Does anyone else have the issue where after working on and publishing a GPT that you need to force quit, clear the app cache & restart several conversations before the changes take effect?

[broken arrow]

When i use the word "Midjourney" in the name of my custom GPT, I am not allowed to publish it for everyone. Can't I use that name? What exactly is the problem here?

[grizzled flax]

Hi, it could be because of this policy: https://openai.com/brand#gpts-in-chatgpt

A GPT may not use another organization’s trademark in its name or logo unless they are authorized to do so.

[broken arrow]

Ah, thank you!

But what does this mean "Builders must verify their domain in order to use many common trademarks"? I already verified my domain with a DNS record

[grizzled flax]

I think it means, if you're Disney and you want to build a Disney GPT, you first have to verify that you own the Disney website. I don't think it means that verified domains just get to use trademarks.

[broken arrow]

alright, thanks for clarifications!

[violet palm]

Unable to file Upload, again. Is anyone else having this issue?

[abstract drum]

Hello, when I click on "My Plan", the url changes https://chat.openai.com/#pricing but nothing happens.

[serene cipher]

when you click this button, it should bring up a modal, you might have some sort of extension preventing this modal from showing up

[abstract drum]

no extension on my chrome, same issue on firefox

[amber stream]

One message removed from a suspended account.

[vernal briar]

same

Did they ever adress the reliability issues? At this point, using ChatGPT for business applications is completely infeasible with how unreliable it is. For the past month, there has always been an issue when I needed to use ChatGPT for anything more than the most basic language processing capabilities.

Like, just yesterday my GPT's API graciously decided to work again after not working for weeks and now i can't upload files. It's ridiculous. /rant

[calm silo]

Is anyone else having issues with GPT4 today? All chats seem to get stuck on the 'understanding question' stage.

[abstract drum]

for info, issue on Chrome and Firefox, with Edge it works

[slate wren]

are you using firefox? Did your try other browsers?

I am having a similar issue for a while

[calm silo]

Using Chrome? I will try Edge.