ChatGPT and other general-purpose language models are capable and useful But creating a "custom GPT" can give huge benefits for little effort.
For example, as part of the creation process for my new book Use A.I. Stay Human. recently, I wanted to create a profile of what an ideal reader looked like, and the problems they were facing.
A target reader profile like this allows an author to ensure both the book's content and marketing are appealing to the audience. We want to make sure that the text of the book clearly addresses the problems and needs of the reader, and that that's clear to them when they see the book title, cover and description.
I don't write new books every day, but it does seem like something I'm going to be doing more regularly, so having a tool that helps me make a good reader profile relatively quickly would be useful.
And I know a lot of business authors who might find it useful too.
So I decided it would be a good "custom GPT" to make.
What's a Custom GPT?
Custom GPTs are fine-tuned versions of large language models (LLMs like ChatGPT or Gemini).
They usually outperform their general counterparts in specific tasks.
What is Fine-Tuning?
Fine-tuning is a technique used to adapt LLMs to specific tasks or domains.
LLMs are initially trained on vast amounts of general text data (think the entire internet - yes, even Reddit - plus the all the books in the Libary of Congress, plus Encyclopedia Galactica). Apparently not all with permission, either!
This general training gives them broad knowledge and capabilities.
However, for specialised tasks, these models can be further trained on smaller, task-specific datasets.
This additional training is called fine-tuning.
Fine-tuning offers several benefits:
- Improved performance on the specific tasks they're trained for
- Better understanding of specific terminology and concepts for those tasks
- More relevant and accurate outputs for those tasks
- Reduced likelihood of generating irrelevant or incorrect information for those tasks
Custom GPTs in Context
Custom GPTs are a practical application of LLM fine-tuning. In OpenAI's ChatGPT, each Custom GPT is a fine-tuned version of the general ChatGPT models. They're usually tailored to perform specific tasks or operate within particular domains.
By fine-tuning a general-purpose GPT on carefully curated datasets, we can create custom GPTs that excel in targeted applications.
For instance, a custom GPT designed for medical report writing could be fine-tuned on medical literature, case studies, and symptom databases. This specialised training allows it to understand medical terminology, and provide more accurate and relevant reports in a healthcare context.
Why are they better than general purpose LLMs?
Domain-Specific Expertise
Custom GPTs excel in specialised fields. When fine-tuned with domain-specific data, they acquire specialised vocabulary and concepts, resulting in more contextually appropriate outputs. This makes them particularly effective for tasks such as medical diagnostics or legal document analysis.
Enhanced Task Performance
By training on tailored datasets, custom GPTs become adept at understanding the nuances of specific tasks. This can lead to fewer errors and more accurate, relevant responses. For instance, a custom GPT fine-tuned for customer service can handle inquiries more effectively than a general-purpose model.
Efficiency and Effectiveness
Custom GPTs achieve high performance with fewer training iterations compared to building models from scratch. This efficiency saves time and computational resources. By focusing on a narrower range of topics, they reduce errors due to overgeneralisation, ensuring more relevant outputs.
Tailored User Interaction
Custom GPTs offer responses aligned with specific user preferences, maintaining a consistent tone and style appropriate for the intended audience.
Creating Custom GPTs
Creating a custom GPT involves fine-tuning a pre-existing model with specific data relevant to your needs. This may sounds complex, but anyone can do it, and you can get a huge return on just a small amount of effort.
It involves far less effort than building and training a model from scratch, think "years and $10B+" vs "minutes and free".
The process to create custom GPT in ChatGPT is simple
- Figure out what specific task we want to give the GPT
- Break down the task into instructions
- Collect any relevant input data (usually in some kind of text format)
- Feed the instructions and data into a new CustomGPT
- Test your new "fine-tuned LLM" and tweak instructions & data if needed
Future Prospects
We might see the development of an "App Store" for custom GPTs in the future, making AI more accessible and beneficial for a wider audience.
Custom GPTs could facilitate the creation of "organisational prompt libraries" or "AI grimoires" for companies.
Custom GPTs deliver better results in specific use cases, surpassing the capabilities of general-purpose models in targeted applications.
Custom GPT Resources
I've created some custom GPTs that demonstrate the simple and practical applications of task-specific AI. These GPTs are designed to perform single, well-defined tasks efficiently.
They are not complex, but they do streamline specific processes that I do regularly, saving me a lot of time and effort.
Here are some examples:
- Prepare Post for WP Blog: Prepares blog posts for self-hosted WordPress sites, providing formatted HTML with a table of contents and image prompts.
- Write a Compelling Opening Hook: Assists in writing engaging opening hooks for social media or blog posts.
- Timezone Converter: Helps convert times between time zones, accounting for daylight savings and other complexities.
- Emoji My List: Enhances lists with appropriate emojis for improved visual appeal.
- Clean HTML for WP: Outputs clean HTML code blocks for easy integration into WordPress text editors.
- Create Target Reader Profile for Non-Fiction Book: Builds a target reader profile for non-fiction books using website data, competitor analysis, or user input.
Each GPT serves a distinct purpose from writing assistance to time management and content formatting.
Have you made a Custom GPT? Let me know, I'd love to check it out.