Week 3: Rooting for GPT-4

Hi everyone, and welcome back! As I discussed last week, I began creating a custom GPT using OpenAI’s GPT-4. The process for creating your own GPT is pretty simple, but I will go more in-depth on how exactly I was able to customize this chatbot. In this post, I will cover the steps for creating a GPT as well as what it can do.

The Steps

Installing GPT-4 is a crucial step, as OpenAI’s latest model provides tools to create your own large language model. I had a lot of fun playing around and exploring the GPTs, such as DALL·E, which generates an image based on the user’s inputted text. More impressive, however, is the ability to create your own GPT through the GPT builder.

Once you open the GPT Builder, it prompts the user to provide a gist of what they are trying to accomplish. For example: 

For this step, I asked the GPT Builder to create a virtual dentist focusing on dental issues and preventive care. From that, it generated some conversation starters on the preview side, such as: “How can I prevent cavities?”. However, I later found that when prompted, the GPT doesn’t have a very colloquial style of speaking, which is important when supporting human interactions. Since we are dealing with a very personal issue regarding medical history and health, it is important to handle patients with a sense of empathy. This is difficult to tackle with a chatbot since patients seek that one-on-one human interaction for a sense of understanding and comfort. While this is hard to achieve, I asked the GPT to emphasize a friendly tone so that the audience is able to understand what is being advised. With these changes, the description of this chatbot is now: “A virtual dentist focusing on dental issues and preventive care, using accessible language.”

Next, the GPT Builder prompts for the role and goal of this virtual dentist. I asked it to focus on providing general oral care tips and preventative care. This step proved to be a challenge, as when the user inputs their stated problem, the LLM provides a large amount of text. If the chatbot’s messages are difficult for the user to understand, it defeats the purpose of having a chatbot, as there is no effective way for the user to communicate with it. Additionally, if the user states a problem that requires more inquiry, such as their specific symptoms, I asked the chatbot to ask more questions so that it is medically accurate and doesn’t misdiagnose.

Even after prompting the GPT Builder to ask them questions one at a time and separate the text, it did not seem to understand this command. This is something I definitely have to fix when implementing my model in Google Colab as useablity is an important element.

One issue I came across when brainstorming this project is the risk of misdiagnosis as, since we are dealing with patient diagnosis, it is important that the algorithm is accurate in order to protect the health of the patients. Learning from the patient’s medical history is necessary as it could be harmful to recommend a treatment that could worsen the patient’s symptoms. To combat this problem, I placed a limitation stating that the GPT should not provide invasive treatments that have long-term side effects but rather focus on general treatments and over-the-counter (OTC) remedies. 

Summary 

To provide a summary of the instructions I gave the GPT builder, these are the configurations of the chatbot.

Now, with these instructions, my custom GPT is able to provide OTC treatments for a patient’s stated problem. This is a preview of what my model looks like in GPT-4:

Conclusion

Next week, I hope to fix these bugs in my custom GPT by finding a way for the LLM chatbot to be more communicative with the user. Currently, it asks the user for their problem and generates a large amount of text without truly going in-depth on their exact symptoms. This is harmful because if the chatbot is not medically accurate on the exact issue a patient has, it can worsen their symptoms by providing the wrong treatment. Once I implement this chatbot in Google Colab, I hope to solve this problem by breaking down the generated text. 

That is all for Week 3! Thanks for reading, and let me know if you have any questions. See you next week!