Transcript – How a Data Scientist Used ChatGPT to Curate a Book
SEASON 3 EPISODE 2
“WS: We are really at the beginning of an entire revolution that’s going to completely restructure every industry on our planet.”
[0:00:12.2] MM: Welcome to Veritone’s Adventures in AI, a worldwide podcast that dives into the many ways technology and artificial intelligence is shaping our future for the better. I am your host, Megan Mintchev and I’m here with William Steele, a constructor, designer, innovator and a data scientist. He recently worked on a nonfiction project which we’ll dive into on this episode.
Veritone disclaims any responsibility for any statement of guests in the podcast. The views expressed in this podcast are those of the interviewee and do not necessarily reflect the views of their tone or its directors, officers or employees.
[0:00:47.5] MM: Thank you for joining the Adventure in AI Podcast, William.
[0:00:50.8] WS: Thank you so much for having me, really a pleasure to be here.
[0:00:53.2] MM: Yeah, absolutely. So can you tell us a bit about who you are and what your background is?
[0:00:59.6] WS: So my name is William Steele. I guess, the best way to describe me would be I do design and development, most of it is self-taught. I did spend some time in university studying computer science and applied math but now I work full-time at night as a data scientist for a company called Maximus. I do data analysis and reports and generating critical reports for our leadership to make decisions on some of the programs that we do.
But on my spare time, I play with natural language processing models. I’ve used artificial intelligence into some of the workflows where I translate audio for different audiences on YouTube as well as the development of audiobooks and audiobook production using some of the artificial intelligence software that is available for text-to-speech.
[0:01:51.8] MM: I love it and as an avid reader myself, my ears just perked up when I head audiobook. I love listening to audiobooks. As a very busy person, I thrive off of listening to them. So that’s really cool, I’m very excited about this conversation. So since coming on the scene in November 2022, we have been hearing a lot about this thing called ChatGPT but of course before we really dive into that, I really want to back it up a little and let people know what exactly ChatGPT is.
So we have this company OpenAI, it’s an AI research and deployment company and this company trained a model called ChatGPT, which basically interacts in a conversational way. So it can answer a follow-up questions, admit mistakes, challenge incorrect answers even, also reject inappropriate requests. So before I have you tell us how you discovered ChatGPT and what it can be used for, I do have to say I actually had fun with it myself and use it to help me come up with the questions that are for today’s episode.
I do need to say, it made my job easier, much faster and way more fun. So William, can you tell us how you discovered ChatGPT and what it can be used for?
[0:03:09.5] WS: Definitely. So kind of being in the industry, I’m a huge fan of Sam Altman. He’s one of the founders of OpenAI and I kind of just like to keep tabs on certain people in the industry and kind of the ongoing projects that they are associated with and of course, OpenAI had been talking about GPT-3 and some of the… it’s a generative transformer model that they’ve been working on for a while.
And so when they made the announcement, of course I immediately logged in just to play with it and within five minutes, I was absolutely mind blown, I’m sure like everyone is, about the applications and some of the domains that it has such robust knowledge and so pretty much just following some of Sam Altman’s movements in OpenAI and keeping my ear kind of to the industry and what’s going on through Twitter, through newsletters.
It had already made some ripples by the time I found out about it. I actually found out, maybe it was 15 days after they had released it, so I was not one of the immediate first users but I definitely did decide to try and apply it in ways where I don’t think anyone had at the time.
[0:04:17.9] MM: When you’ve been using it and everything, how do you think ChatGPT has handled and understood context and natural language processing?
[0:04:27.4] WS: So in my experience for everything that I’ve used it for and I’ve done a little bit of prompts, I’ve done prompt request for writing parts of Python scripts. I’ve done prompt requests for writing screenplay scenes, I’ve really tried to explore using the tool and it is very effective at its current stage for most of the application, for most of the users. Of course, anything in its early stage could use some more refinement.
It will continue to evolve as time evolves but right now, I mean, for most people and for, I think, a lot of business use cases, it’s at a very good stage for performance. Of course, you need somebody to prompt it and you want somebody to look over the work before you move forward but take for an example, asking you to write a chapter for a book. It’s very effective at doing that and in English and in Spanish and in Korean.
I had very effective results just on the first take of the prompt without needing any translation services, you know, manual human translation services, which is quite incredible. It took the amount of time necessary for traditional writers to produce work to, I mean, like a fraction of what it used to be, as well as the fact that it just completely removed an entire workflow process of having to get a human translator to manually translate, verify, proofread and then set back.
So all of that is done in one step just by adding two words at the end of your prompt, you just write your prompt and then you put in Spanish or in Korean or in Vietnamese and it immediately spits out things that at first take that are, in those three languages, Korean, Spanish and English, they were grammatically and spelling correct 100% proofed already.
[0:06:18.3] MM: This is like really mind-blowing stuff. It’s obviously, I’ve dreamt on it. I’ve basically used it almost daily just to see what will come up and help me with like, every day, even things outside of work and stuff and it’s quite impressive the way that it just spits out and how fast it spits out questions and answers and sentences and paragraphs even.
So I can imagine how it was with when you just said like, writing that chapter. So where are you talking about giving examples of real world application and you’ve already mentioned writing a chapter, can you explain a little bit more about that? Are you talking about writing a whole book?
[0:06:59.9] WS: Definitely, definitely. So that was actually the first kind of big weekend project I undertook with ChatGPT. I had it generate 35 chapter titles for a book titled, An AI Wrote This Book: How the World was Changed Forever by Artificial Intelligence, Insights and Predictions Written by an Artificial Intelligence system.
So the title idea is something I came up with and then I asked it for 35 chapter titles, an intro and a conclusion and it produced those chapter titles and then I basically just plugged in those titles and asked it to generate a 4,000 word chapter on that particular chapter title.
So what came of it was a book that was entirely generated by the AI and its insights and predictions into what it thinks the impacts that the artificial intelligence will have on the world. So it was very interesting just to see how it broke up what chapter titles it selected and then it was very interesting to see how it presented the information in each chapter.
You can almost kind of tell where there’s more data points that was the model was trained on because of the way that it presents certain chapters and the information within those chapters. It’s very clear that there’s much more information on maybe, let’s say an example is, on the impacts of AI within the financial markets and then there’s a lot of information maybe per se on the impacts of artificial intelligence in the healthcare industry.
So those are just examples of like, where maybe it had a lot more data so it was able to make a much more robust answer and even with very precise case examples versus maybe an area where it wasn’t so clear, let’s say an example for how AI will impact government. There’s a less amount of information to be trained on or maybe the parameters are set and situate that it doesn’t talk too much about that.
There’s very clearly a spectrum of which the responses came back on and you can see that in the book, either by reading it or by listening to it.
[0:09:06.5] MM: I’m curious, did you ever find that it was repetitive in each chapter at all? Or again…
[0:09:11.5] WS: Definitely, yes.
[0:09:12.5] MM: Okay, okay.
[0:09:13.7] WS: There was repetition for certain things like the AI really like to talk about what it’s going to best at doing and what it’s not going to be best at doing and it kind of repeats that through multiple chapters. It’s kind of a reoccurring theme but the goal of the book, the goal of why I pieced it together and I coined a term kind of titled, AI curation, there’s a lot of different terms being thrown around right now with how humans and the job market will evolve with artificial intelligence.
I think that the term AI curator is probably the best big umbrella term for it as the way that we have to interact with the AI changes. Meaning like, maybe right now we’re doing natural language input. We’re doing text inputs but maybe one day, there will be an AI that’s trained on, let’s say for an example, you’re a musician in a jam session and you start to maybe make a guitar riff and there’s an AI that is listening and it starts to lay drums down in real time.
So the way that we interact with these systems will continue to evolve. So I think that the term that would be best fitting for how we begin to build our understanding and maybe potentially how we frame larger theories of interaction with these systems would best be the term AI curator. So as I curated the book, there’s a specific set of constraints that I curated the book with which is, I don’t at any point, edit anything that I was given back and I present everything in the exact structure that it is.
So that maybe people that are very busy, don’t have time to mess with it in this much detail or don’t care to could get very quickly and then one concise body of work, exactly how the AI spits out information and how it was structuring that information.
[0:11:02.0] MM: Okay, so once you had the entire book written, that’s when you back and edited it however you see fit, correct?
[0:11:08.8] WS: There was not much editing required. So I put all of those 35 chapter prompts in and what I got back, I put into the book and of course, I had to format like for a specific standards like it’s available in ebook. So I had to format the spacing correctly in the margins but I did not do any content editing whatsoever. What you get is exactly what ChatGPT, the model known as Text Davinci 003, what it spits out.
The main thing to note that I would say is, I was able to get the entire book, everything written and submitted for publishing within eight hours, which is I mean, if anyone’s every tried to… I got one previously published fictional novel as well and that took four years to write with this nonfiction. It took me eight hours, I was absolutely taken aback, it was an incredible experience.
[0:12:04.2] MM: And you're talking eight hours form the moment that you started asking these questions or writing these prompts into ChatGPT to the time that it was completely published?
[0:12:14.4] WS: Yes. From the moment that I came up with the idea for the title to the book being submitted so that it goes through quality control with distribution, it was eight hours.
[0:12:29.9] MM: Wow.
[0:12:30.4] WS: Yeah, so what most people get done in a work shift, I mean, I was able to literally curate with ChatGPT an entire nonfictional book.
[0:12:37.9] MM: Wow. So okay, you’ve written a fictional book so that probably explains why you felt compelled to use ChatGPT to write this nonfictional book, is that correct?
[0:12:52.7] WS: Yeah, definitely. Having experience in the field, it was one of the first places that I thought, “Huh, I wonder if there’s like some kind of application, creatively or business wise?” So I immediately jumped to, “Well, I wonder if it can really write a book” like if there’s a way to use it in such a way that you can create enough information and enough value to present to an audience and that’s what I came up with.
[0:13:13.8] MM: It is completely different from traditional writing methods. Are there any… instead of asking how different it was, are there any similarities?
[0:13:24.0] WS: I think that that’s project, that will be project dependent. So for somebody maybe creating fictional works using ChatGPT, which it’s very capable of giving you some ideas, each creative process in the curation process looks totally different. So for my nonfictional book, it was very cut and dry because the nature of the information was supposed to be unrefined and very raw to show the audience exactly what’s going on.
Maybe for say, for someone that’s writing a fictional book or maybe a screen play using ChatGPT, it will look much different because it builds based of what you feed it. So you could give a pretty broad set of constraints on the, let’s say for example, a scene in a fictional book that takes place, maybe an underground lab under the ocean. So that constraint, the setting is something that it will take into consideration if you give it that constraint but also the way that you propose the constraint affects how it generates the response back.
So it can be a very, very wild and fun creative process if you decide to try and create something fictional using ChatGPT. It will be interesting to see how different creative processes are generated especially because the nature of art history is like each time you go to create something new, it is going to be a totally different process than it was last time because you are not the same person.
[0:14:54.7] MM: It actually kind of reminds me of those choose your own adventure books in a way using ChatGPT, yeah, that’s pretty neat. So in your experience, what do you think were the limitations or challenges of writing or curating this book?
[0:15:11.7] WS: So really right now with the stage it’s at the only limitation and challenge that I could see that was immediate was you can’t generate too much text with one prompt as the system currently is. It limits it to 4,000 tokens, which pretty much means you’re not going to get really over maybe 2,000, 2,500 words back in a response in any one given prompt. So you really do have to parse out the structure of the whole idea.
Meaning like it was important that we had instead of just 10 chapter titles, 35 chapter titles because it gives it the ability to parse out the information in a way where it can give a good enough response for each chapter that you have a coherent total body of work that is fulfilling, you know? It wouldn’t be much of a book if it was just a short 5,000 word essay kind of with only five chapters.
[0:16:07.1] MM: Right.
[0:16:07.9] WS: You do have to kind of think about the best way to leverage the current output capacity of the system on a per prompt basis.
[0:16:17.3] MM: Okay and then how did it handle grammar and punctuation?
[0:16:21.8] WS: Excellent. There was absolutely not problems with that whatsoever. I found no grammar and no punctuation and no spelling issues with ChatGPT with the responses that I got back. I’m sure that for a really productive workflow, somebody could take a really great artificial intelligent system like Grammarly and overlay that on ChatGPT to get different stylistic effects in their writing because I know Grammarly has some capacity to do that.
But I mean, then you’re just talking increasing productivity by adding another step to the workflow using another really well-trained system. I mean, Grammarly has been around for a while now and that’s more for style modifications, you know?
[0:16:59.7] MM: Yeah. Oh yeah, I use it all the time in my work.
[0:17:02.9] WS: Yeah, same.
[0:17:05.0] MM: All right, let’s talk about ethical considerations because I’m sure a lot of people listening are probably wondering about that. So when you are using this model specifically or even any AI generated text for writing a book or curating it as you say, I am particularly curious about authorship though.
So I know you had touched upon it a little bit but if a book is primarily written by an AI model, it obviously raises questions about who it should be credited as the author and then the AI does not have the ability to provide consent or to hold rights to the work of course. So it would obviously be important to define that role of the AI model for that whole writing process.
How have you addressed the ethical considerations for this book beyond the title? And go ahead and repeat the title for people so they know.
[0:17:54.6] WS: Yes. So the title of the book is, An AI Wrote This Book: How the World was Changed Forever by Artificial Intelligence, Insights and Predictions by an Artificial Intelligent System. So in terms of the ethical consideration and copyright all those types of matters, I can’t say that I necessarily might have a right answer for that question personally. I made it very clear in my – in this particular project that is mostly written by an AI.
It’s literally in the title. I can definitely see the value of using this generative model in any body of work to enhance any one’s creative process for any reason and there should be one thing to be said, which is kind of the idea where the tool is only as good as the hands of which it’s in. This is a technological tool. As time progresses, people will continue to see that there is a talent in it of itself in being able to effectively leverage these technologies in ways that really bring value to an audience or to a consumer.
So naturally, the things that aren’t valued won’t be purchased, won’t get attention, the things that aren’t bringing value they just don’t get traction. The things that do, do get those traction. So for the most part, it’s not much different than using a laptop to write out your manuscript versus using a typewriter. The process of getting to the final product is a little bit longer using a typewriter, you have to make by, you know, manual copies, things like that.
I look at AI as a tool in that same way, you know? It really comes down to if what the end product is, if it’s valuable to people then I think it’s net positive and the person that took the energy and the time to create the idea and put the work in is the one that’s due the credit. Of course on a personal level, if someone asked them like, “Hey, how did you do it?” and you know, someone is deceptive about how they created it, that’s a totally different discussion.
But I would see no reason for people to feel any shame in saying like, “Oh yeah, like 40% of my biography I used this particular AI software to generate the initial 40% of the manuscript.” I would’t see any problem with that and I think that’s still the work is there is a lot of work still left to be done by that author, by that creator. I think maybe in something like human degeneration, it gets a little bit – it gets to be a little bit more of an interesting discussion.
But I still think if you abstract what I just said, it still applies. There are people writing entire comic book worlds right now using DALL-E, which is another one of OpenAI’s tools and I mean, there’s some really, really amazing comic book worlds and they seem, the storylines behind them, the characters, the world’s very original and not just because they’re using AI. For me personally, I don’t see it as any discredit to the creative process.
I think what you are going to find is there’s going to become an increasing value on the depth of the essence of what you’re creating versus the technical capability of the individual to produce that. So really, really complicated storylines and ideas and themes for this new comic book worlds, the ones that become the greatest will be resonant stories and they’ll be very resonant ideas.
So it will be more about that ability to resonate and less about, you know, maybe a creator’s particular ability to produce a very articulate idea. Maybe the technical ability will be not really too much of a factor anymore. So maybe it will enhance very, very great writers and very, very great original thinkers and their ability to produce larger bodies of work by themselves or with a smaller studio and I think that that overall is a net positive.
Because who knows how many great, great, great stories and great worlds are in the minds of individual people waiting to be shared?
[0:22:00.2] MM: Yeah and as you were talking, I was even thinking maybe people who don’t have the ability to type on a keyboard but they can use their voice to say the prompts in ChatGPT to be able to create content or create whatever it is that they’re looking to make.
[0:22:18.1] WS: Exactly. Yeah, definitely.
[0:22:20.3] MM: Good point, yeah. You not only used AI to write your book but you also used it to create an audiobook version, so let’s talk about that. How did you go about doing this and what technologies did you use?
[0:22:33.5] WS: Yes, so I actually, once I finished writing it I was like, “Man, how awesome would it be to use some of the wonderful technologies that Veritone has been so awesome to allow me to use to produce an audiobook?” because I already use Veritone’s text-to-speech AI system for dubbing in different languages but then I was like, “I really wonder what it would sound like if I produce the whole audiobook that was written by AI using that same text-to-speech system?” and it all worked out really well.
Really, really well. I actually shared it with some friends and some colleagues and they couldn’t even tell that it wasn’t a human. They were really, really impressed and it’s these kind of combination workflows of different AI systems that I think are really going to be disruptive to a lot of big industries. I really, I think if you could empower individual people to be able to produce audiobooks because right now the cost of entry to produce an audiobook is very, very high.
A lot of times, the people that are producing the things that need to be narrated are creators. They are writers and they usually don’t have high budgets but with systems like this, they can create such a high value product very, very quickly and very, very effectively. It’s a very exciting time, you know? That a lot of things that seemed to be contained to certain groups of people are now becoming much more available to everybody.
[0:23:58.8] MM: Yeah and I listened to the audiobook and I hear obviously the Veritone voice all the time. So it’s always mind blowing to me still, I don’t care how many times I’ve heard it but I listen to the audiobook, when I heard it and I was like, “This is insane that this is not his voice or not a particular person’s voice.” It is an AI voice. So kudos to you for having the mind of, “Hey, let’s go ahead and try this.” That’s really neat.
So overall though, creating or curating this book using ChatGPT and then turning around and creating this audiobook version of it, what was your experience?
[0:24:37.8] WS: It very quickly showed me and gave me a firsthand experience at the incredibly disruptive nature of these systems and the fact that the timing is crazy because we are really at the beginning of an entire revolution that’s going to completely restructure every industry on our planet.
When you firsthand feel the capability and you are able to at the end of one weekend listen to an entire audiobook that you just had curated with the day before and now the production for the audiobook is done and it’s at the point of competing with all other, you know, high end value productions, it shows you that a lot of people have no idea what’s coming.
And if you can become an early adopter and really stay aware of these emerging systems, there’s new systems coming out every week it seems, it is a very exciting time for people with imaginations and with the ability to work hard and want to do something for themselves, it’s very empowering to the individual creator and to people with startup vision.
[0:25:45.4] MM: Of course, we need to know where can we find your eBook as well as the audiobook?
[0:25:52.2] WS: Yes, so the eBook is available on Amazon. It’s just the Kindle eBook, if you prefer reading. If you would like to listen, it is available for free. I went ahead and created a quick YouTube channel, the handle is @william.steele or you can type in the title of the book and yeah, if anyone is also interested in any of the other ongoing projects I have, you can visit my website, williamsteele.info or follow me on Twitter @will_e_steele.
[0:26:22.0] MM: Perfect and all of this will be in the shownotes of the episode, so people can easily see it and click through it as well. William, thank you. This has been really amazing. Again, I feel like I’ve said mind blown a hundred times on this episode but it’s true. It is like the keyword of the day for this specific episode but I’m just so excited to see what continues happening in the wonderful world of AI. Thank you William for your time today, I appreciate it.
[END OF INTERVIEW]
[0:26:51.3] MM: This has been another episode of Veritone’s Adventures in AI, a worldwide podcast that dives into the many ways technology and artificial intelligence is shaping our future for the better. Talk with you next time.
Data Scientist, Maximus