Introduction to Custom GPTs
What are custom GPTs?
ChatGPT has been trained on a vast amount of content and can provide answers about almost anything. Recently, OpenAI released a new feature that allows a user to refine and customize their own language models known as custom GPTs. These models have the benefit of focusing on a specific task, and when properly trained, they can perform their specialized task better than a general model such as ChatGPT.
Comparing ChatGPT to custom GPTs
ChatGPT analyzes the prompts given to it and, based on the context provided, tries to produce a helpful response. If insufficient context is given, it will "guess" what the user is looking for. Any time ChatGPT is forced to make a guess, it increases the risk of hallucination, or a response that contains inaccurate information. Moreover, even if sufficient context is given in one chat, ChatGPT will not remember it in any future chats.
Custom GPTs solve many of these problems. Instead of having to provide context and instructions every time a new chat is created, a custom GPT remembers its specific instructions and refers to them every time it is prompted. This generally results in more helpful and consistent responses. As the custom GPT is used, its instructions can be iteratively improved to further fine-tune its behavior.
Another powerful feature of custom GPTs is the ability to load files into their knowledge base. The model will refer to these files before it responds to the user, employing a simple form of retrieval-augmented generation. Custom GPTs can also make use of Actions, a framework that allows the model to interact with external sites and services.
The use cases for custom GPTs are endless. Some popular applications are programming, document editing, and image generation, among many others.
How to use custom GPTs
Using GPTs from the GPT Store
Custom GPTs are created by OpenAI and community members. Creators can choose the level of visibility for their GPTs, choosing to make them private or publishing them to the public on the GPT store. In January 2024, OpenAI reported that over 3 million custom GPTs have been created. Other sources have reported that over 150k GPTs are public and available in the GPT store.
To use a public custom GPT, navigate to the "Explore GPTs" section in the sidebar, search for a model, select one that matches your needs, and click on Start Chat.
Creating your own custom GPT
Custom GPTs are easy to create and require no development experience. In its simplest form, creating a GPT involves chatting with the GPT Builder, a custom GPT that helps users create other custom GPTs. Further fine-tuning can be done by loading external data, configuring additional features, and customizing the conversation starters that are suggested when creating a new chat.
CREATE A CUSTOM GPT WITH THE GPT BUILDER
Navigate to the "Explore GPTs" tab in the sidebar, and then click on the "+ Create" button.
A split screen view will come up. The left side of the screen is the configuration for the custom GPT (we will call it the builder), and the right side is a live preview for the custom GPT (we will call it the preview). As instructions are given to the Builder, you will be able to test the GPT in the Preview screen.
- Describe the desired behavior for your GPT to the builder. Remember good prompt engineering practices: the more descriptive and specific the instructions are, the more helpful the model will be.
- The builder will ask for a name and a logo for the GPT. The logo can be uploaded or generated with AI. It may also request more information about the expected behavior of the model.
- Start interacting with your GPT on the preview screen. If the model is not behaving as expected, continue to offer instructions and feedback to the builder. Go back and forth between the builder and preview until the model is preforming how you'd like.
- To finalize the creation of your custom GPT, click on the "Create" button. Choose a privacy setting and the custom GPT will be created.
- At any time in the future, a summary of the custom GPT's configuration is available in the "Configure" tab of the builder. Further changes can also be made in this screen.
Here is the link to the English to Spanish Companion GPT that was created in the image gallery above: English to Spanish Companion GPT.
CONFIGURING YOUR CUSTOM GPT
- If your custom GPT needs to be trained with custom files, upload them in the "Knowledge" section of the Configure tab in the builder. Don't forget to tell the builder how to use the files that you upload, either by conversing with the builder or including it in the "Instructions" section.
- If you'd like, you can customize the "Conversation Starters" to control what a user will see when they create a new chat.
- You can enable and disable certain capabilities of the model, including web browsing and image generation, in the "Capabilities" section.
Here is the link to the General Conference Study Buddy GPT that was created in the image gallery above: General Conference Study Buddy GPT.
While OpenAI does not use the documents uploaded to a GPT to train its models, it is important to remember that uploading sensitive data into a GPT can still pose major security issues. If a GPT is meant to be used internally and has sensitive information, strict attention needs to be paid to who can access it. Never share a custom GPT trained on sensitive information with the public.
Beware of Hallucinations
Remember that a custom GPT is simply an instance of ChatGPT with a specific set of instructions. Because of this, it is still prone to hallucinations. When creating your own GPT, you can avoid hallucinations by instructing the model how to appropriately handle prompts that are outside of its scope. However, hallucinations are always possible with even the best-trained custom GPTs.