Maybe you are wondering what a book enthusiast is doing with AI, or ChatGPT?

Well, when I started Joelbooks website, my vision was that I will search for great stories, and if I want to know the best new stories in first hand, I need to search them in form of books.

Since 2018, a lot of time passed. AI made a huge leap and everybody is seeking to find the right place for this technology, even myself.

Book Detective

Since I'm a techy, I already tried OpenAI GPT's capabilities and created What is that Book?, a book search engine that is capable of searching not only with keywords or categories but also themes or character-specific attributes.

If you remember a momentum or only the story outline of the book, you heard or read in the past, AI is capable to match the pattern and give title suggestions. Based on my personal test What is that Book? worked well with English books. However, the problem is that even the latest update of ChatGPT only includes books up to April 2023. What about the latest books we are all seeking?

To solve this problem, I taught ChatGPT about the latest books present on our website, and thus Book Detective came to life. (In November we also opened custom book submission to AI dataset.)

Atma

Book industry will change for sure. I'm almost certain that many new tools will coma out in 1-3 years for writers. Automatic editors, text transformers, that may quickly turn US English into UK English (even the style). Automatic book illustrators, that will quickly embed 1-1 image to your book, just by reading your character description. But my main vision is interactive books.

Consuming a book is a linear process, and was since the very beginning. But in the world of Tiktok we are looking for instant content, and instant answers.

I have always dreamed of compiling all my spiritual knowledge into one book, knowledge that transcends dogmatic churches and nonsensical spirituality. This dream led to the creation of Atma. Atma is an AI personality that holds all my sacred spiritual knowledge, and she is ready to speak about your deepest questions.

What Are GPTs?

GPTs enable users to create a version of ChatGPT tailored for particular needs, enhancing its utility in daily life, domain specific tasks, or more.

Since ChatGPT's launch, there has been a demand for more customization. While Custom Instructions were introduced in July, GPTs offer a more comprehensive solution for customization.

Creating a GPT does not require coding skills. It involves starting a conversation, providing instructions and extra knowledge. You simply choose its capabilities, like web searching, image creation and voilà.

What Will Be the GPTs Store?

A platform for sharing GPTs. The GPT Store, is launching in late 2023. It will feature GPT creations from verified builders, with functionalities like searchability and leaderboards. The store will also highlight notable GPTs in different categories and, in the future, offer the possibility for creators to earn revenue based on the usage of their GPTs.

You will find both Atma and Book Detective publicly in GPT Store.

How I Created Book Detective?

I know there are a lot of tech friends also on Joelbooks, so let me share some technical details.

There are two main parts of creating a GPT:

1. Give initial Instructions to the GPT, this is basically a prompt engineering task
2. Create a custom dataset that will serve the domain specific knowledge to your GPT, basically the inner knowledge not part of ChatGPT

I wouldn't say any of these tasks was simple.

Giving Instructions to GPTs

It seems very simple and honestly for a certain point it is.

You can simply talk to the GPT in “Create” tab which will generate a certain Instruction set. The problem emerges when you would like to achieve a certain behavioral pattern with your GPT.

I wanted to initially two important patterns:

• Formatting the book titles in answers in a way I prefer, I find it pleasant to read
• When the AI receives a task to lookup a book, it should primary search for the latest dataset I uploaded

You see part of the concrete instructions in the screenshot, giving you an idea of how I achieved a behavior pattern.

What I found during testing that Book Detective sometimes no matter what it drops the requests in the instruction, so repeating ourselves or using stressing language is a good idea when we are looking for consistency. Or at least it certainly improved Book Detective.

The overall length of the instructions in case of this GPT is only 1,114 characters and doesn't include conversation style.

Custom Dataset for GPTs

In case of Book Detective the more important part was the custom dataset. Good data means great results, so I spend several days to make the perfect dataset for Book Detective.

What I did?

First I created custom agent to Joelbooks website to crawl all the books I wrote about in the last years, there articles mainly include books from 2015-2023 (fresh books so to say).

I didn't even realize that I wrote about 4,054 books on my site, not a bad number at all 😊. For every book, I included the following information using the crawler:

• Book title,
• Author,
• Year of publishing (where applicable),
• Article title for more context,
• Category, keywords,
• Short description or review text.

Let me show you a sample from the export:

I created an upload ready document from this custom book dataset for Book Detective and uploaded it to the AI engine. The dataset size was about 3,8 MB in raw text.

After the upload, all I had to do is to match the dataset name in the instructions and click to deploy.

How I Created Atma?

After the above section, I must say the process was really similar, however Atma's personality is more detailed and complex in the instructions section.

I wanted her to reflect a young wise soul's personality, like kids in their early ages, they have incredible sharp insights.

The main difference is the knowledge base behind Atma. Creating this knowledge base is more like writing a book, but there are several parts where I also use prompt engineering techniques in the knowledge dataset, so called “hidden instructions”. These instructions may slightly change how a question is answered.

I primary using question-answer script format, where I focus adding the most important knowledge in the answers. In Atma's raw knowledge dataset you would find topics like:

• Our existance
• The nature of physical world
• The nature of non-physical world
• Life lessons

I would like to keep the dataset as dense as possible, since LLMs are tend to use the middle part of these dataset less than the beginning or ending part. So keeping the dataset small would slightly help Atma to focus better. At the end of the day Atma's knowledge will range from 10,000 to 50,000 words, a length of a short novel.

Atma is still in beta, and most likely I will need several months to finish the 1.0 version.

How Can I Test Atma or Book Detective?

At the very moment, you will need ChatGPT Plus plan to try Atma and Book Detective out, but most likely after the release of GPT Store, these GPTs will be all public and more or less free to use. Anyway I already set these two AI friend's status to public.

If you manage to try these AIs, I would appreciate any feedback or idea using the following contact form.

Joel Stafford

My profession is online marketing and development (10+ years experience), check my latest mobile app called Upcoming or my Chrome extensions for ChatGPT. But my real passion is reading books both fiction and non-fiction. I have several favorite authors like James Redfield or Daniel Keyes. If I read a book I always want to find the best part of it, every book has its unique value.