GPT reset

VGMoose

Yeah, it's 30 minutes from the last message (or if you message it "reset" without quotes), so you could drag a conversation out by messaging it repeatedly every 29 minutes for example. But the limit is mostly because it can't keep track of the whole history. If you talk to it enough, you'll eventually hit an error too.

VGMoose

to phrase that another way, every time you message it, the entire chat history (from the "starting" point) is sent to OpenAI, and each follow up message it re-sends this entire history again. So, a longer chat without a reset will eventually be too many tokens for it to handle.

dafunkymonke

Okay that makes sense, is there a way to extend the reset timer past 30 minutes? I feel like that having a meal and coming back for it to automatically reset is very tedious in terms of trying to be productive with it

VGMoose

@dafunkymonke Okay I didn't test it too much, but if you send:

timeout: 120

(needs a colon, and one space before the number, just like that)

That will set the reset timer to 120 minutes from the last message. If a custom timeout is set, it will notify you at the start of each conversation history. To undo, send:

reset_timeout

And it will remind you of this command at the start of the current history. You can set this as high as you want, but be aware that it will probably error if the chat goes on too long without a reset.

dafunkymonke

Thats extremely helpful, thanks alot!

ㅤㅤㅤ

@VGMoose If you put a decimal number in there like:

timeout: 1.75

then BrowseGPT breaks, he wont respond to anything and needs to be readded to the chatroom to function again.

idk if this is useful information or not but i thought id point it out

Edit: he also breaks when using letters instead of numbers, for example:

timeout: abc

seems like everything that isnt a number will break him if put there

Edit 2: it seems that after breaking BrowseGPT you dont need to readd him, you can also wait some time, not sure how long exactly but im guessing that it still does the chat timeout thing that was set before breaking him with anything

VGMoose

oopz, ok now that should return an error

ㅤㅤㅤ

@VGMoose just tested it, is fixed now

dafunkymonke

Also, another question. How did you manage to get GPT into your own website? I seen it was in GPT--3 instead of 3.5, do they release old versions as open source to use? Or is there some sort of deal for it to be used

VGMoose

It's using openai's chat completion API: https://platform.openai.com/docs/guides/text-generation/chat-completions-api

With the "gpt-3.5-turbo" model (It doesn't seem to be aware of this fact, but that is the model it's using). This model has very reasonable pricing ($0.50 per 1M tokens) so I don't mind the overhead

VGMoose

Also! if you have an openai account, this python script is an example of how to use the API: https://gist.github.com/vgmoose/a54408a28189b19501ed1afb7ee8d4e1

There's an initial prompt, and then each API request sends the chat history to the end point, so over a conversation it builds out.

On Line 44 is the main infinite loop. Line 47 prompts for text input from the user, Line 49 actually submits it to openAI, and then 50 prints the response, before saving it and looping back around.