Artificial Intelligence Study Group
|
Contacts: jdberleant@ualr.edu and mgmilanova@ualr.edu
Agenda & Minutes (162nd meeting, May 9, 2025)
Table of Contents
* Agenda and minutes
* Appendix: Transcript (when available)
Agenda and Minutes
- Today: ET's MSIS defense of book project.
- Book: https://docs.google.com/document/d/1tOhGzVFYnwmkTlb0_brlRSY6Uol6vy3MNFf3Q98ZNDc/edit?tab=t.0
- What next? Put it online, self-publish, both, neither, ...?
- Masters project on using AI to write a book.
- A source of information about the problem:
- https://spectrum.library.concordia.ca/id/eprint/993284/1/Smart_Mdes_Spring2024.pdf
Transcript:
0:02 - Unidentified Speaker
Hi, everyone. Hello.
0:45 - M. J.
Hi, everybody. Hello. Hi. Nice to see you, J. Good to see everybody.
0:54 - M. M.
A small group today because, yeah, this semester.
0:59 - D. B. & M. M.
Yeah, busy time.
1:01 - D. B.
Alright, well, so today is E. T.'s defense, and she'll be talking about her book project. And the way these defenses work is so first, the student will present their work and we'll take audience questions. Then after that, the audience departs and the candidate meets with her committee. And after that, the committee meets in a closed session and then reconvenes with the student to give feedback. So with that, I'm going to just go ahead and turn it over to E. to present her project.
1:47 - Unidentified Speaker
Hello.
1:48 - E. T.
Are you able to see my screen.
2:00 - Unidentified Speaker
Yes.
2:02 - E. T.
Alright. Good afternoon. As both an educator for the past seven years and a graduate student in information science, I'm excited to share my master's project today. It is evaluating the limitations of generative AI in expanding simple instruction content, writing a book. This research conducted under the insightful guidance of Dr. B., explores the potential and pitfalls of using generative AI to create comprehensive instructional materials using the practical example of writing a gardening book, which is my passion. Let's explore what happens when generative AI pushes us to write a full-length book. Today, I'll walk you through the background and motivation for this project, the research problem and gaps in the literature, the methods I used, my experimentation with various large language models, evaluation results, key findings, and finally, my conclusion and future directions. In recent years, generative AI models like ChatGPT, Gemini, or Claude have evolved into powerful tools capable of producing coherent, informative, and sometimes complex-ridden content. As an educator, I'm constantly looking for innovative ways to engage learners and deliver meaningful instruction. With these tools at our fingertips, a fundamental question arises. Can AI not only explain a simple process like how to grow vegetables from seeds, but also expand it into a full-length, structured, and engaging book? This project stems from that very question. I chose the topic of gardening because it is something I'm deeply passionate about. For the past seven years, I have been doing gardening. As a hands-on gardener, I wanted to see how well AI could mirror the care, detail, and personal insight that come with nurturing plants from seed to harvest. So this wasn't just about generating words. It was about guiding AI to build a coherent structure, sustain a narrative voice, and even integrate visuals, essentially simulating the entire process of writing an educational book. Through this lens, my project became a case study in both the capabilities and the limitations of AI when applied to long-form instructional content. While large language models like ChatGPT, Claude, and Gemini have demonstrated exceptional skills in generating short-term content, things like summaries, answers to prompts, or snippets of dialogue, their ability to handle long-form instructional writing remains largely untested. These models are often praised for their fluency and responsiveness in conversations, but writing a 100,000-word educational book is a very different challenge. It demands sustained coherence, factual narrative progression, and most importantly, continued reader engagement over a long span of content. Most of the current literature and industry focus has been on using AI for brief or segmented tasks. Think emails, essays, blog posts, or code suggestions. What's still missing is a deeper exploration into whether AI can handle complex structured content at scale. This is the gap my research aims to address. My project investigates whether, with proper human guidance, large language models can produce not just long content, but coherent, accurate, and instructional long content. Essentially, I wanted to test the real limits of these models when we push them beyond their comfort zones in full-scale authorship.
6:11 - Unidentified Speaker
Before starting this project, I identified four key gaps in the current research on generative AI. First, scalability.
6:20 - E. T.
While we know LLMs are good at generating short snippets, we don't yet understand how well they perform when stretched to tens of thousands of words. Can they maintain clarity, structure, and interest across that much content? Second is the comparison of models. ChatGPT, Claude, and Gemini each have different architectures and capabilities, but little research has directly compared their long-form performance side by side.
6:56 - Y. P.
Third is the role of prompt engineering and human curation.
7:05 - E. T.
Should I continue? Yeah, go ahead.
7:08 - Unidentified Speaker
OK.
7:09 - E. T.
We know prompts affect output quality, but we don't fully understand how these strategies evolve over time and sustain project-wide work. And finally, we're missing insight into multi-model workflows, specifically how AI performs when we combine text generation with image creation to support a unified educational narrative. To address these gaps, my research focused on three main objectives: 1. Evaluate the ability of LLMs to generate long, coherent instructional content that minimizes redundancy and factual errors. 2. Compare their performance using structured criteria such as coherence, factual consistency, engagement, and repetition. 3. Develop and test prompt engineering techniques that helped maintain stylistic and thematic consistency across multiple AI-generated chapters. Together, these objects guided every phase of the project, from outlining and drafting to editing and evaluation. For this study, I used qualitative descriptive research design to explore how effectively AI can produce a full-length instructional guide, specifically a 100,000-word book on growing vegetables from seeds. To do this, I collected a wide range of raw data from AI-assisted writing processes. This included the actual content the AI generated, the prompts I used to guide the writing, and my own reflections throughout the process. I approached the project as a study where I played the dual role of author and observer. This allowed me to evaluate not just the end product, but also the process and effort involved in collaborating with AI tools over time. In total, the data collected included the full AI outputs, both rough and revised, prompt logs showing how I refined requests over time, revision notes detailing where and why I curated, and a reflective diary capturing challenges, insights, and moments of cognitive load while navigating between different tools and platforms. Additionally, to bring in a layer of external perspective, I invited a small group of readers, which are my co-workers, to review selected chapters and provide feedback on clarity, tone, and engagement. This methodology allowed me to evaluate the AI's capabilities, not just from a technical point, but also from a user experience and instructional design perspective. One of the most important phases of this project involved experimenting with different large language models. I worked with ChatGPT 4.0, Gemini, and Claude especially, each offering unique strengths and challenges. The primary goal was to test the limits of these models by asking them to generate the maximum amount of coherent instructional content in a single prompt. But this wasn't just about word count. I also evaluated each model's ability to maintain coherent factual consistency, engagement especially, and to avoid redundancy over long stretches of text. This process helped me better understand the capabilities and limitations of each tool. For example, Claude showed strong performance in sustaining longer uninterrupted responses. ChatGPT was particularly versatile while guided with clear structure prompts. Gemini had a more balanced style but often required additional clarification or follow-up prompts to stay on track. Through this experimentation, I gained valuable insight into how these systems can and sometimes cannot scale up from generating to producing book-length materials. These tests became the foundation for evaluations that followed and helped me determine how best to assign tasks and optimize prompts for each model. Ultimately, this phase of the project confirmed that while LLMs are capable of impressive feats, they still require significant human oversight and editorial judgment when it comes to producing long-form instructional content. In terms of sheer output, Claude stood out as the most efficient large language model. It constantly generated the highest word count in a single prompt without sacrificing coherence. So these are not actually token counts for these models, but those are the word counts for a single prompt. In contrast, both ChatGPT and Gemini were able to reach similar word counts, but only through follow-up prompts or manual extensions. This required more back-and-forth interaction, which made the writing process slower and more fragmented. Claude's ability to surpass the word target in one go was a key factor in making it the most practical in generating full chapters at a time. For example, if I gave Claude a 4,000-word count, it would go above like 4,500 or maybe close to 5,000. This slide shows a visual summary of how each large language model performed when asked to generate content using a single prompt. As you can see in both the table and the chart, Claude outperformed the others significantly, producing around 4,200 words in one go. ChatGPT 4.0 averaged about 2,700 words, while Gemini reached around 2,900 words. While those are still impressive, they often required additional prompting to continue or finish a full chapter. Claude, on the other hand, was able to sustain its response much longer, allowing for more efficient content generation and fewer interruptions in tone and flow. This difference really matters in a project like mine, where consistency across long chapters is essential. It also affected my editing workload. Longer coherent outputs from Claude meant less stitching together of fragments and fewer transitions to manage manually. So in terms of raw performance and usability for long-form generation, Claude proved to be the most powerful and time-saving option for me. In addition to text generation, I also evaluated how well different AI tools performed in creating images, particularly those used for book covers and chapter illustrations. This chart and table compare four AI tools, ChatGPT 4.0, Gemini, Leonardo AI, and Microsoft Designer on two main criteria: image realism and text accuracy, meaning how well the tool could include legible and correctly spelled text within the image. Leonardo AI ranked the highest in image realism, creating highly detailed and visually appealing results. However, it had some logical errors in the pictures. For example, when I asked to create an image with several vegetables, it created one single plant that has tomatoes, cucumbers, and eggplants, all these vegetables, sort of vegetables coming out of one plant. ChatGPT 4.0, on the other hand, provided a strong balance. It performed well in both realism and especially in text accuracy, especially as the tool was updated during the course of my project. At the beginning, if I asked ChatGPT to put text on the picture, it would mostly spell some words wrong. And even if I asked to correct it repeatedly, it would still put the wrong spelling. It was the only one that reliably placed the correct title and the author name on the cover image, which made it my top choice for final book visuals. Gemini produced decent visuals. The pictures were very, very realistic. However, it still had that issue with putting the text correctly on the pictures. Microsoft Designer lagged slightly behind in both categories. Overall, this evaluation showed that even an image generation tool selection matters. ChatGPT is improving over time, and its ability to render accurate styled text gave it a unique advantage in producing professional-looking book covers and supporting illustrations. Once I completed the book, I wanted to go a step further and not just rely on my own editing, but actually ask other AI models to evaluate the book's quality. I uploaded the manuscript into four different large language models, again, ChatGPT 4.0, Gemini, Claude, and Copilot. Each model was asked to assess the book on key performance criteria: coherence, factual consistency, redundancy, and reader engagement. Let's start with ChatGPT 4.0. It praised the book's seasonal structure, strong transitions, and accurate terminology. It highlighted that the content was factually correct and used clear language. The main issue it flagged was moderate redundancy, especially in sections involving personal reflection or repeated discussions of root tools. Gemini agreed on the logical flow and strong transitions. It especially appreciated the solid grounding in ecological and scientific principles and noted that any repetition was purposeful and useful for learning. Claude went a step further in highlighting the book's emotional tone. It praised the use of personal stories and the way expressive language helped make the content more engaging for readers. Claude confirmed that the information was accurate in terms of gardening science and that the content flowed clearly from seed to harvest. However, it did point out some repeated themes, particularly around soil care and observation. Finally, Copilot echoed many of these observations. It found the book to be coherent and thematically consistent and noted a strong personal connection in the writing. Still, it also flagged a need for tighter editing, as some themes and phrases were revisited too often. These external evaluations were valuable. They confirmed that the book was not only informative and well-structured, but also emotionally meaningful to the reader. At the same time, they helped me identify areas for improvement, especially where repetition could be trimmed in future versions. As I wrap up this presentation, I want to highlight the key takeaways from my research project. First, this project clearly demonstrated both the strengths and limitations of generative AI when it comes to creating long-form instructional content. While these tools are incredibly powerful, they still require thoughtful guidance and human oversight to produce material that is not only coherent, but also accurate and engaging. Among the models tested, Claude stood out as the most effective when it comes to generating large amounts of content in a single prompt. This efficiency was especially important in a project of this scale, where I was aiming for a book-length manuscript. In addition to text generation, I also explored AI-generated visuals. These played a significant role in enhancing the overall look and feel of the book. Tools like ChatGPT 4.0 showed notable improvement over time, especially in terms of adding accurate, legible text to images, something that was a challenge in the early stages. And finally, the evaluations conducted by other language models helped reinforce the areas where AI still needs improvement. Redundancy and maintaining long-term coherence were two recurring issues, especially when models were prompted repeatedly across different sessions. These are important insights not just for researchers, but for educators like me, and developers working to integrate AI into creative workflows. Ultimately, this project wasn't just about pushing AI to its limits, it was about understanding how we can collaborate with these tools more effectively. And as these technologies continue to evolve, I believe this kind of human-AI partnership will become even more relevant in education, publishing, or beyond. Looking ahead, there are several exciting directions for future research. First, there's a lot of potential in exploring more advanced prompting and generating techniques. By refining how we interact with LLMs, we can likely improve the structure, clarity, and creativity of AI-generated content even further. Another important direction is to examine the impact of long-form AI content on actual readers, specifically how it affects learning, comprehension, and engagement. This would help us move beyond technical evaluation and better understand how people experience AI-written materials. Finally, I believe there is a growing need to develop clear guidelines for human-AI collaboration, as AI becomes a more common tool in education, publishing, and other fields. We need thoughtful frameworks to ensure quality, ethics, and creativity remain at the forefront of content creation. Thank you. I appreciate the opportunity to share this project with you all. If you would like to see some parts of the book, I would be happy to show some parts from the book. This concludes my slide. Thank you.
22:27 - D. B.
Okay, thank you. So we're open for questions from the audience.
22:32 - V. W.
I'd first like to make a comment. The graphical style, and I put this in the chat, is very effective. It's clear, uncluttered, and it's very pleasant and consumable. I don't feel like I'm having to be stressed out. I'm just getting your message in a way that's very sustainable, and that's fantastic. We mirror your experience in word count with Claude exceeding expectations and accurate technical content. I hope you get your Ph.D. because this kind of clarity of thought is very valuable.
23:09 - E. T.
That's what I have to say.
23:12 - D. B. & Y. P.
Thank you so much.
23:14 - Y. P.
Hi D., this is Y. and I would like to concur with what Dr. W. said. Everything that you mentioned I just want to say but maybe you know a couple of more thoughts as the follow-up this round so congratulations on getting here and good luck for the next steps but a couple of thoughts that came to my mind is one of the things like as next steps you mentioned maybe you can create like a model to continuously test what you're testing because these models will continuously get updated so there could be like a technology solution that is continuously, you know, testing what you're testing or what you have tested. And also, so that was one thought that came to my mind. And the second thing is, you could actually use the model that you have for publishers. Now, I don't know how organized or how you have done that, but help people who want to build similar project to do that also, which further tests or continues testing of your model. So those are some ideas that came to my mind, but again, take it or leave it, but very good presentation and good luck.
24:42 - E. T.
Thank you. I appreciate your feedback.
24:45 - D. B. & M. M.
Can we see please some chapters or images or something from the book? Of course.
24:52 - M. M.
I think she mentioned she will share. I would like to see that also.
24:59 - Y. P.
Maybe I'll start gardening after that. Or I'll learn more about gardening. Two more things actually I wanted to share, E., is Claude. We use Claude significantly from a development standpoint, the backend especially. That has been my experience also. So again, just reinforcing what V. said.
25:23 - E. T.
And then the experience of cohesiveness and concurrency.
25:28 - Y. P.
I've expressed that across different models. So for front-end development, we are using WordCell and there are, I think, three, four models my team is using. We have the same problem everywhere, by the way. So maybe there is an opportunity to solve that on top of these models. So it is not unique to the models that you have tested. The other models that are used for different purposes, we have observed that problem. So what we essentially do is we restart in your case from zero, we just remove everything, create prompts outside and restart. So we are doing some manual processes to fix that problem. But I wanted to also validate and confirm your other things, which is not across the models you have mentioned, but other models too. Sorry, I'm done. And sorry, I have a quick thought that amplifies what Y. said.
26:26 - V. W.
And that is one second, there was a request for E. to show the book.
26:31 - D. B.
Did you want to share your screen and show the cover? Something about the book, E.?
26:37 - E. T.
Sure.
26:38 - V. W.
While she's pulling it up, I just want to quickly say that Y. opens the opportunity that the revision of a book itself could be an automated process as the large language models improve. So if you have your whole procedure tabulated in your prompts, your editing and your follow-up, you could conceivably just revise the book by pushing a button. And that's a pretty interesting idea.
27:03 - E. G.
And that's exactly where I was going with it, V. When you went through this, it sounds like you covered two sides of the equation, the development of the book and the editing of the book. Which did you see the strength in the editing aspect or the development of it? I'm sorry, would you repeat it?
27:30 - E. T.
One more time. Where did you spend?
27:32 - E. G.
Where do you think it was strongest in this process? Because it sounds like in the developing of the book, you found C. to be the preeminent tool. But in editing the book, you were able to find that you were able to get a lot of feedback. Which did you find it was stronger in developing the book? Not C. but AI in general, or editing a book based on your seven years of gardening experience? Does a better job what I see.
28:09 - E. T.
I mean, it does develop the book, yes, as long as you give the structure. Because if you just throw in something like, OK, I want to write a book about gardening, it just throws everything. It's not the way I wanted to structure it. So first I had to structure the book like saying that OK, I would do this. I would get the seeds, pick the seeds, the type I wanted, and then the challenges that I faced. After that, how would I transfer them to the gardening? I mean, the backyard garden, or either raised bed or container gardening? So as long as I give the overall structure, it develops the chapters and the parts of the book. With writing the book, I think it did a better job with editing. Thank you. We saw the cover.
29:13 - E. G.
Can we see some of the chapters?
29:17 - M. M.
Yeah, the index would be nice.
29:21 - E. G.
That's a great cover.
29:24 - E. T.
So this was generated by ChatGPT. And here's the content. It is around 317 pages, something like that. And it was, I think, it passed a little bit, a hundred thousand words. If I need to show. OK. Alright. So it's 100,666 words. And I wanted to, so before I started writing, I had to structure the way I want my book to sound, not just like AI-generated, I wanted to sound more human. So I had to give AI a very specific writing style. And here is the introduction.
30:19 - M. M.
And models that you can capture your style and they mimic your style.
30:29 - E. T.
So I specified the style I wanted to use, like poetic, adding some anecdotes from my experience and at the same time informational because I didn't want it to just sound like a story. I wanted to add some information to readers.
30:58 - M. M.
Would also links or the references links also?
31:10 - M. M.
Some references? Probably. References, like, I'm sorry.
31:16 - V. W.
Do you have an idea of how many pictures are in the whole book, what the ratio of text to pictures or illustration to text is? Oh, sure.
31:26 - M. M.
There's a picture for each. Let me show one more time here.
31:30 - E. T.
So for part one, there's a picture that goes along with part one and part two, part three and part four and the back cover of the book. So if I scroll down, I think we all get here is one. So this is the illustration. For part one and not for each chapter though. I could see this accepting a lot more pictures.
31:58 - V. W.
The ones that you've chosen are really exquisite, but I would like to see some more how-to verb of planting, soil preparation, troubleshooting problems. I would see that also embracing the notion of more visual representations. I like this.
32:20 - M. M.
Yeah, I like this suggestion. Yeah, I think that it's necessary a little bit more.
32:26 - E. G.
A picture is worth a thousand words type of thing.
32:30 - E. T. & M. M.
Exactly.
32:30 - V. W.
And a movie is worth a thousand pictures. Yes.
32:34 - Y. P.
Yeah, I was about to say that, but after you said that, that with AI, you can actually make instructional also.
32:42 - V. W.
She could take her major illustrations and put them in Poe.com and with Flux or some other animating program she could actually create nice five-second animations which could create the opportunity for a web-based version of the book in video chapters where she would become an influencer, sell seeds, and become infinitely rich.
33:09 - M. M. & V. W.
Absolutely.
33:10 - Unidentified Speaker
33:10 - E. G. & V. W.
I like it.
33:11 - M. M.
We're looking at the finished product, which I think is phenomenal.
33:17 - E. G.
Thank you. What does it look like on the prompts?
33:21 - E. T.
Do you have the prompts that you used?
33:24 - E. G.
Basically, how did you get here? What road did you travel? I'd like to see how you organize the prompts, what you started with.
33:36 - E. G. & E. T.
And how you retooled it to get here.
33:39 - V. W.
The making of the sausage.
33:42 - E. G.
Yes.
33:42 - E. T.
So when I started first, I was like, my mind was everywhere. I didn't have much prompt engineering. I can't say I'm the best right now either, but I can definitely say I have a better idea of how to guide AI. Because when I started the project, it was like a simple question, okay, how would I write a book? Or what will be the chapters of a book? So I tried several different things. It's, it's like a mess, I need to show my AI accounts, I guess, for those my chats, I don't have the snippets. But if you want me to show those, I can log in and show those chats. So it started like that, like, very simple, how can I write then I started structuring my mind, okay, I need some outline and When I get the outline after getting the outline in chapters, I was like, okay, Well, I want to write a book, but I don't want it to sound like a robot. I don't want it to like, you know, something boring. So I started searching the famous book writers. I mean, famous gardening writers. So I read a couple of them but some and how they sound like or what is their style. I searched through those and then I made a kind of mix what sounds like engaging at the same time, not totally that writer, but a little bit of me, my experience and how my writing would sound. I mean, I'm not a writer by the way, it's just, you know.
35:26 - V. W.
You are now.
35:27 - D. B. & E. G.
What I'm trying to allude to is you didn't come into this without having some subject matter expertise.
35:38 - Unidentified Speaker
No.
35:38 - E. G.
So what I'm seeing here is all this was, was a tool to basically augment what you had bringing to the equation. So it allowed you to be more productive. Quickly, not to say, I want to write a book, boom, it regurgitates a book on some subject. And there's the notion of making a song your own.
36:07 - V. W.
So although you may have excerpted hundreds of writers' experience and gardeners' experience in the construction of this work, by retaining your own voice, you kind of make their song your own and become the best of all those worlds.
36:25 - E. T.
I hope so. Yeah, an amalgamation, but in your tune.
36:30 - D. B.
I have a follow-up to E.'s question about writing a book about a subject that you already know a lot about versus something that you don't know a lot about. Based on your experience, do you have any comments on whether it's better to use AI to write a book about something you already know about versus something you don't know about?
36:59 - E. T.
I personally would rather try on something that I know about because if I choose something that I don't have any idea how to do, I think it would get more time for me to get familiar with the topic and then structure the book because gardening is something that I already have an idea about. If I want to change some part, and if I feel like some part of the book sounds more technical than engaging, I have an idea to change that. But if it's something that I have no idea about, I guess I would just go with whatever AI puts there.
37:42 - V. W.
And being a domain expert helps you screen AI hallucinations and fabulation so that if something doesn't sound right according to your gardening experience you can catch it before it hits the press.
37:56 - E. G.
That's exactly where I was going with it because at that point you're also a reviewer, a validator of the information coming out.
38:05 - E. T.
Yes. OK, here's another question.
38:08 - D. B.
So if you look at the table of contents it's got, you know, page numbers and everything. I guess you did that yourself, right?
38:18 - E. T.
OK. The AI wouldn't sort of know what the page numbers are probably.
38:24 - M. M.
I'm so happy to see J. C. here. Thank you for coming, J. Do you have any comments? This is your first time here. We're having discussions and J. has a lot of experience with so many businesses. What do you think? About this one?
38:44 - J. C.
I'm just visiting and getting accustomed to the style of the meeting but I did have one question about the illustrations. I've dabbled with AI illustrations and last week one of my friends had a broken phone and they were unhappy about it so I decided to send a get well card to their phone. And the AI made me a wonderful, funny illustration, except that it spelled get well with an H instead of a W. So it was get hell. And I could not convince it to change that. Wait a second, wait a second.
39:28 - E. G.
That was a broken phone. You sure that the AI wasn't performing a Freudian slip?
39:39 - J. C.
I don't know. Anyway, I've had that experience trying to do illustrations with text on them before, and she mentioned that as a difficulty. I'm just interested in what tool you have to use to mix text and illustration well.
40:02 - E. T.
So I've tried all of them that I list over there. Gemini, ChatGPT, Leonardo AI, which I really like the realistic pictures that Leonardo AI created. And also Microsoft Designer. Yes, Microsoft designer. The one that goes with the text correctly spelling was ChatGPT. So Leonardo creates beautiful pictures, Gemini creates beautiful pictures, which I kind of sense like Gemini gets some pictures from the web and puts some stuff on it to make it generated by itself. I'm not sure. I just sensed that. So ChatGPT was good. I mean, it's not the best realistic pictures for ChatGPT, but it was able to put the text correctly. And I tried to put the back. So let me show the back cover. of the book as well. So I put a long text on the back cover of the book. ChatGPT struggled with this one as well. After a couple of tries, it finally brought the correct spelling and the correct text that I wanted to put on the back cover.
41:22 - E. T. & J. C.
So if I try something long, a longer text on a picture, I would say ChatGPT would still struggle with that as well, but for short texts, like a couple of sentences, maybe a couple of words, ChatGPT is the best.
41:37 - V. W.
One solution to this problem is to composite your nice background picture with prearranged formatted text, which you can do in programs like PowerPoint almost for free. You just say, send the picture to back and bring the text to front and choose the style of the font you want and the aspect ratio and the layout and the equalization of the picture. If you have Photoshop, you can also do it there, sometimes with slightly more crispness and quality, but I think that that would reduce the number of iterations that it took for your text to go. And you can also, in PowerPoint, adjust the color so the text really pops and can be clearly read against all, say, light-colored things like the path.
42:21 - J. C.
And so that doesn't cost very much and works really well.
42:25 - E. T.
Yeah, and I don't know if anybody else, the survey from Microsoft this week. But if you use Copilot, you're supposed to be able to do that sort of thing within PowerPoint now.
42:39 - J. C.
And the direction of their questionnaire was AI-assisted. So when I gave it an answer, it asked me for elaboration or commented on my comments. It was kind of interesting from that. But it seemed as if they're going to try to miss almost as if they're trying to get rid of windows and you work with an AI and the word processing, data processing, image processing products.
43:11 - E. G.
E, when you created this, did you hand it the picture in the text separately or did you give it the text and describe the picture you wanted?
43:25 - E. T.
First, I described the picture I wanted, and then I also put the text that I wanted to be on the picture, but it failed at that. So then I got the picture. The picture is from Leonardo AI, and I asked ChatGPT to add the text on the picture and kind of move it around or change the font, those sorts of things. And then it was able to put the text correctly after a few tries, because first it made the background of the text white on the picture, and I asked to remove it because it didn't seem nice on the back cover, those things. So after those tries, it was able to get the text picture with the correct spelling. I have a couple of questions.
44:22 - Unidentified Speaker
I'm sorry.
44:23 - M. J.
I have a few questions for you. Can you hear me okay?
44:28 - E. T.
Yes, yes.
44:29 - M. J.
Okay, so did you do anything on the marketing side? Like are you thinking of actually publishing this book? Oh no. Okay, because I was just thinking, you know, from that point of view you can also do target audience research around who your audience for the book would be, who would buy these books. And you could tailor some of the content to it. And then I'm curious, because I'm also a gardener. I love gardening. Did you try to give any advice based on zones of where people live or anything like that?
45:10 - E. T.
Yes. I asked AI to put it where I mean I told AI that I'm living in S. so I wanted to get the zone and the climate and the frost and everything according to both S. as my experience and also guidance for the other zones and how they should pick, the reader should pick what type of plant they want to grow or what season they should start growing. Okay, cool.
45:44 - V. W.
You know, monetization is the logical endpoint of doing a lot of work. And I see this work with very few tweaks being very monetizable, just as it sits. So I would urge you, even if you go through something like the Amazon Kindle portal, where a lot of people will be able to read the book for free, it still gets your name and your work out there and forms the basis for subsequent works and author familiarity. So I would encourage you to go ahead and monetize this. I've written a bunch of advanced math books that absolutely no one reads, but it's been a very valuable experience because the people that do read them appreciate them. And that gets my name out there as being interested in these specific branches of endeavor. And I think that would be true for you as well.
46:33 - E. T.
Thank you so much for the suggestion.
46:36 - E. T. & Y. P.
There's one thought that came to my mind with E. and V. asking these questions.
46:44 - Y. P.
I think if you start also thinking in terms of solution as like an AI solution in your next steps, it seems that you will have to go multi-model given all the context that you shared. Problems and issues so although I know I heard C. as you know one of your favorites when it comes to content creation especially long text but then when we spoke about pictures or we spoke about a few other things we also realized that there are constraints and you might have to use other models so if you end up building a solution, I don't know which route you intend to go. But if you end up building a solution, it could be a multi-model solution that you could build. And I'm thinking only from a technological standpoint, if you want to go the technology route.
47:50 - E. T.
Is it making sense? Of course. I would love to go deeper into, especially as you said, picture generation, adding more pictures into it. A lot better as gardening needs more I mean everything needs more visual it makes more appealing for the reader especially and I would love to try those I would love to go deeper into those but I have because I had limited time to get the project done and everything this was the best I was able to do.
48:23 - Y. P.
And one more logical next step also could be thinking of like a rack model around this because you can build it for specific authors. So E. has one style and you mentioned like that human intervention, but then there could be focused on a particular topic.
48:43 - E. T.
It could be focused on a particular author.
48:46 - Y. P.
It could be focused on so many dimensions. So multi-model plus rack are some of the logical next steps from a technology standpoint that are coming to my mind.
48:59 - E. G.
Actually, I'm going to echo D. here from a previous. If you're not careful, you can actually take this and go into a Ph.D. because I could see you using agents to basically build out everything going forward. You tune those agents and you basically you start with something and let each one iterate out.
49:24 - E. T.
You know, I really like that, E.
49:27 - V. W.
And I also see another, there's another book in this book that we're looking at right now. If E. was to go back and document the whole process that she used with all of its faltering steps, with all of its misdirections, with all its attempts, out of that would emerge a process that was then refinable. And that would make an excellent primer for those of us who also want to write books using AI and encountering the various issues that she did. So there's at least one other book that we're staring at, and that making-of book would be extremely valuable to her, even if she never published it, but it would likely, just as this one is, be very publishable.
50:08 - D. B.
Well, I want to follow up on that question with one of my questions, which is, you know, we started out the semester with three students wanting to write books, and E. is the last person standing. So my question to E. is, how can I pitch this or organize this project for future students so that it works out as well as possible for them? I didn't give you any guidance. We met every week with this group, and people had some things to say. But I've never written a book with an AI, I wouldn't know how to guide people to do it. What should we do to make these projects successful in the future?
50:56 - E. T.
I would say encourage them to do the project on something they like because if I was forced to do this project on something that I had no idea about, I guess I would struggle and it would give me a hard time instead of finding joy in doing that because it also gives me insight and it also gives me knowledge on what I already do. I think that would encourage them to pick what they like to write a book on.
51:35 - Unidentified Speaker
And also giving, I think, I mean, the email was very appealing.
51:42 - E. T.
I received it at the beginning of the semester that you sent about the project. So I'm not sure why they did not follow up. Maybe more follow-up emails encouraging them to actually join the project would help. I'm not sure. Those are the only things that I can think of right now. OK, thank you.
52:17 - D. B. & E. T.
Alright, well, let me ask another. Well, actually, M. said she had several questions, so I didn’t mean to go around M. if you're here.
52:30 - E. T.
It's OK.
52:31 - M. J.
I feel like I got asked by others. I mean, I do think that kind of delving into some things like your personality and how you, you know, the tool with your personality, both for the writing and for the art style is really an interesting topic. And of course, there's all these new tools now for the graphic design that are really powerful. So, you know, you can, I mean, I think it's kind of fun to think what you could do now that you couldn't do just a year ago. I think it would be cool to customize or personalize a version of it depending on where you live. So that kind of thing could be handled more easily with AI than you could do it as a person, just a human, just a human. Let's see. I was curious if you entered in a reading level for your reading. Did you try to control the voice? Or personality or reading level of the writing?
53:40 - E. T.
I did control the voice and style, but I did not think about reading level, to be honest.
53:47 - M. J.
Yeah, because I think that the other cool thing you can do with AI is like you could do a book for kids, right? And you could do a book for adults, you know, you could create, you know, different content depending on the age of the person, teenagers, or whatever. And the pictures are really child-friendly too.
54:11 - V. W.
I mean, I just was thinking, I want a book like this for my grandkids, you know, this is fantastic.
54:18 - M. J. & V. W.
Yeah. And then it would be, it would be cool.
54:21 - M. J.
I mean, since you, you learned so much, you know, with your structure, you know, which you said was critical, you know, anyway, ways to give people tools. I like the idea of you saying guidelines for AI and human collaboration in writing a book. Yeah, to me, that would be a really useful set of instructions. And I don't know if you've done something like that also, or if that's part of this, because it sounded like maybe you had spoken to some students or something, some of D.’s students. Yes, I did talk to some of my students.
54:56 - E. T.
I didn't get their feedback on the book, though. I did talk to my students as well, that I will be writing a book and trying to finish it. But those are daily conversations we have in class. Yeah.
55:09 - M. J.
But in general, I mean, I just think it's really great. I'm working with some students who are creating tools to help children tell their stories using AI. So, I mean, I just think there's just so much opportunity. And I think that you picked a really great topic. Thank you so much.
55:29 - E. T.
Thank you. It was interesting.
55:32 - Unidentified Speaker
Thank you. Other questions?
55:34 - D. B.
Alright, well, I have one more question. If you go back to the slides, there's a slide on evaluation where you ask the AIs to evaluate the book. Yes.
55:49 - E. T.
If it's convenient to put that slide back up.
55:54 - D. B.
Of course.
56:07 - V. W.
You know, this presentation forms the basis of that book we suggested on the making of. You've almost got it outlined.
56:17 - D. B. & V. W.
Yeah, I mean, this is a one-semester project, so it's half the size of a full-size project.
56:24 - D. B.
The other half was an internship or a practicum. So I said, my suggestion was that the book itself would be the final report. But if there was another semester, and we can maybe have future students do it, another part of the report could be the process, and the guidance for others, and the excerpts of the interaction with the AI, and all those kinds of things. But anyway, my question about the evaluation slide here is, so these are how the AIs evaluated it, but I mean, how would you evaluate the AI's evaluations?
57:09 - E. T.
I do believe the redundancy part is correct for each AI. And I do see some redundancy about especially observation, taking notes, and using those for the future gardening experiences. I didn't see many things about engagement on ChatGPT's report. So, Claude was the one emphasizing engagement the most. And other than that, factual consistency, yes. I mean, it doesn't give some random information. Coherence, yes. Build up on. But considering that I structured the chapters, I didn't just leave the chapters to AI. I think structure is also built by me, kind of. It's more like collaboration, human-AI collaboration, I would say, for the chapters. So it's not just the AI part.
58:25 - D. B.
And yes, the most that I would agree is redundancy.
58:31 - E. T.
It's not like repeating the same thing over and over. It's more like re-emphasizing it or coming back to the same point in another path. More like that. OK. Any other questions?
58:48 - D. B.
Anyone maybe who hasn't had a chance to ask a question yet or someone would like to follow up on previous questions? Well, if not, we would go on to the next stage in the defense, which is that the audience drops out and E. meets with her committee members.
59:14 - E. T.
So thanks, everyone, for being here.
59:17 - D. B.
And we'll see you in the future as you like. We're going to adjourn for the public portion of the meeting.
59:29 - V. W.
Thanks for a great presentation.
59:32 - M. J.
Yeah, that was a great presentation. Thank you.
59:36 - R. R.
It was really interesting and inspiring.
59:39 - E. T. & R. R.
Appreciate you.
59:40 - Unidentified Speaker
Thank you so much.
59:42 - E. T.
Thank you, E.. Thanks, everybody.
59:45 - Unidentified Speaker
Bye.
59:45 - D. B.
OK, now I need to turn off readAI.meeting because this is not public. So...
No comments:
Post a Comment