What I Learned from Turning Myself Into an AI Chatbot
Today’s post is by author Josh Bernoff.
ChatGPT and its ilk are fine at answering general questions. Of course, in my little zone of expertise, which is coaching nonfiction authors, I’m proud enough to believe that any answer I could give would be more helpful, up-to-date, and discriminating than what you’d get out of a generic AI Large Language Model. So I began to wonder: Could I create a virtual book coach that would answer authors’ questions as well as I could?
As it turns out, there’s a lot more subtlety to doing that than you might think.
Let’s start with content. For the last ten years I’ve been blogging every weekday at Bernoff.com, 2.5 million words so far, and 90% of that content is about books and writing. Because my inspiration comes from my clients’ questions, the range of posts is pretty broad, covering everything from how to pitch a publisher to how to cure writer’s block to how to record yourself for an audiobook. In theory, a chatbot trained on all those blog posts, plus a couple of books and research reports that I authored, ought to be a pretty good resource for aspiring authors.
But I wanted to make sure that the experience for my virtual coach users would be simple and easy, and that putting up the chatbot wouldn’t create a flood of technical questions and support requests. I investigated ways to build it myself, but I’m a writing coach, not a tech wizard. So instead, I partnered with a company called Soqratic that’s focused on making chatbots for authors.
Soqratic took care of creating a simple interface so my potential clients could easily sign up. As it turned out, one of the biggest challenges was to make sure that the virtual coach’s answers were limited to my content. I didn’t want it pulling writing and publishing advice from Anne Lamott or Jane Friedman and passing it off as my own. The solution to this is a technique called RAG (retrieval augmented generation), which limits answers to content from a bounded data set (in this case, my books and blog posts).
I learned quite a few things that weren’t obvious at the start. For example, you tell the chatbot how to respond using a text file of system instructions that’s written in plain English. This is how you make sure that it does a better job than just asking ChatGPT, “What would Josh Bernoff say about X?” I told my Bernoff Book Coach to be authoritative and direct, avoiding the obsequious tone that’s typical of most LLM answers, and to answer only questions about books and writing. I also insisted that it include links to my posts and to chapter numbers in my books, a choice that both promotes my content and limits the chance of “made up” answers (commonly known as AI hallucinations). And until I tweaked the system instructions, the virtual coach had a bad habit of insisting it was actually me, rather than a digital simulation.
As cool as it might have been to make the Bernoff Book Coach free, that isn’t economically viable. Each query invokes a connection to an LLM engine (in this case, Anthropic’s Claude), which has a small cost associated with it. But I set the price low ($1.99 to start, $17.95 a month) to encourage as many users as possible.
From the user’s perspective, the virtual coach is far more effective than just searching my blog. It remembers each user’s previous questions, so if you tell it you’re writing a memoir about your recovery from addiction, it doesn’t make suggestions more appropriate for a technology trends book. It’s available 24-hours-a-day, and unlike me, it doesn’t get annoyed at potential authors’ repetitive questions or their outrage about the strange way that publishing houses work.
The coach has had surprising benefits for me. It allowed me to create an inexpensive, low-overhead relationship with potential clients who aren’t ready for more expensive coaching services. I can review what questions people are asking the chatbot, which informs what topics I should be researching and writing about next. When a client stops interacting, the system emails them and reminds them, not only that the virtual coach exists, but what topics they were asking about when they went dark.
And the virtual coach does not cannibalize my own coaching business. Realistically, people who are fully satisfied with answers drawn directly from my content aren’t potential clients. I’ve also found that the virtual coach often stimulates people to set up actual coaching sessions, so they can get a real human response and tap into parts of my knowledge that I haven’t yet documented in my blog or books.
Other nonfiction authors might be wondering if they could turn their expertise into a chatbot as well. If you want to develop more rewarding relationships with readers, it’s definitely worth a look. One question is where your training content is going to come from. It doesn’t have to be a blog; a virtual coach can be trained on your books, newsletter, Substack, ancillary training materials, or even audio and video like a podcast or YouTube series. Even at this early stage I found Soqratic pretty easy to deal with, and the path to standing up a chatbot based on your book is only going to get easier. Just as readers currently expect to be able to consume your book as an ebook or audiobook, future readers will expect to be able to interact with it as a chatbot.
I’m finding it oddly useful to have the Bernoff Book Coach as a companion. When I want to know what I’ve written on a topic, it does a fair summary with perfect recall. There are still things I know that it doesn’t, but unlike me, it’s got an unlimited supply of patience. Let’s just say between the two of us, we’re pretty likely to be able to give a good enough answer to address what any potential author might be wondering about.
To contact Josh about making your own chatbot, visit: https://bernoff.com/contact
Source link
