How Custom GPTs works, What action they made to build a Custom GPTs in the backend? Anyone who is interested tell the behind theory of Custom GPTs?

Community
1

Everyone is talking about: GPTs are custom versions of ChatGPT created by OpenAI users. All you have to do is tell the GPT builder, in plain English, what you want to create, and the builder will take it from there . But how it is actually happened/managed in the backend by OpenAI.

2 Likes
2

If you need more flexibility, you can use Assistants API.

3

Thanks for your suggestion @TonyAIChamp :slightly_smiling_face: But I really want to know how this structure actually work.

1 Like
4

My feeling it is just an average Joe interface to assistants API (hence a lot of settings are pre-defined).

1 Like
5

Iā€™m a regular user, so my knowledge is incomplete. But from my perspective, this is nothing new, in terms of functionality. Rather, a combination of existing ā€œCustom Instructionsā€ and ā€œpluginsā€ functionalities. This is the ability to save multiple named configurations. If you have used the ā€œCustom instructionsā€ option before, the ā€œCustomized version of ChatGPTā€ is just an extension of this functionality and the ability to save it as a permanent link that you can share with others. From my perspective, these are different sets of ā€œCustom instructionsā€. Very convenient to use, but it is not some super new functionality.

2 Likes
6

@adam.suskiewicz So, we can say it as ā€¦ it is nothing more the instruction tuned version of GPT where the model is able to remember your provided instructions as system prompts do!

7

Aha ā€¦ @TonyAIChamp Can you plz tell me about your experience with Assistants API!

8

Yes, but you also see some backend. You can watch as the AI turns your instructions into an internal prompt. A very educational experience.
Additionally, you can add actions. But as @TonyAIChamp wrote, it is a GUI for functionality that was available via API.

2 Likes
9

@adam.suskiewicz what do you mean by ā€œsee some backendā€?

10

It is another step towards AGI (though nothing ground breaking; just giving higher level functions) It is still in beta. Useful for own tasks, not yet much useful for user-facing apps.

3 Likes
11

The instructions you entered in the ā€œcreateā€ tab are transformed into the internal prompt and then you can see the result of this transformation in the ā€œconfigureā€ tab, in the ā€œInstructionsā€ field.

1 Like
12

Why doesnā€™t anyone mention the fact that files can be given away as the GPT knowledge base ? So the GPT can do some kind of RAG (Retrieval-augmented generation) and integrate the data into its answers. Itā€™s an interesting feature, but I donā€™t think itā€™s any more advanced than plugins that can read files.

By the way, never upload a file to a public custom GPT that you donā€™t want users to have access to; itā€™s actually very easy to get access to files. With a little prompt-engineering Iā€™ve been able to download complete PDFs that are copyrighted.

1 Like
13

The same can be achieved through the standard ChatGPT 4 GUI. I can not see the difference. Files can be uploaded and analyzed, and incorporated into the response text. Itā€™s even more secure than custom GPT, because you didnā€™t share Your files. The main difference is the repeatability of the queries. In custom GPTs you load once and it remembers it. In the standard GUI you have to reload them in the next session. The difference is preconfiguration, which is very convenient, but, as you mentioned, also dangerous.

import pandas as pd

Load the Excel file to examine its contents

file_path = ā€˜/mnt/data/case_study_SFL.xlsxā€™
df = pd.read_excel(file_path)

Displaying the first few rows of the dataframe to understand its structure

df.head()

1 Like
14

Much better off not entering instructions into the create tab. Much better and effective for purpose to enter exactly what you want into the configure tab

1 Like
15

I agree. After using it for some time. At the beginning, it is worth seeing how he transforms himself.

1 Like
16

If a GPT does not respond as expected, you simply provide the answer you anticipated and ask it what you should enter into its configuration to get the expected answer next time. This works wonderfully. In this way, you use your Custom-GPT to improve itself. However, any changes to the configuration only take effect in subsequent, new chats, a behavior I suggest improving.

3 Likes