Embrace or Fall Behind: How Generative AI Is Redefining Creativity, Customer Experience, and Competitive Advantage
Guest: Michael Todasco, Author and Generative AI Strategist; Former Head of Innovation at PayPal
Hosts: Seth Earley, CEO at Earley Information Science
Chris Featherstone, Sr. Director of AI/Data Product/Program Management at Salesforce
Published on: April 9, 2023
In this episode, Seth Earley and Chris Featherstone speak with Michael Todasco, author of "Artificial America" and former Head of Innovation at PayPal, about how generative AI is transforming creativity, productivity, and customer experience. Mike shares his hands-on experiments using GPT-3 and image generation tools to produce AI-written books and comic books, explains why the democratization of prompt engineering levels the playing field, and argues that organizations must understand where human connection remains irreplaceable. The conversation covers AI authenticity, copyright risks for artists, and why the most urgent advice for anyone - regardless of profession - is simply to start using these tools now.
Key Takeaways:
- Your job will not be replaced by AI - it will be replaced by a human who is actively using AI to do your job better.
- Generative AI democratizes creativity by removing technical barriers, allowing anyone with curiosity and good prompts to produce remarkable outputs.
- Prompt engineering is an iterative, experimental skill - the best way to learn it is to play like a child, testing inputs and refining outputs over and over.
- Organizations must map customer journey touchpoints carefully to determine where automation adds value and where authentic human connection is non-negotiable.
- LLMs trained on generic data cannot serve enterprise use cases without a proprietary corpus - organizations must build and protect their own knowledge assets.
- AI copyright and IP questions are unresolved and urgent, particularly for living artists whose styles can be replicated without consent or compensation.
- Every professional in every field should start using generative AI tools immediately - familiarity today directly determines competitive advantage tomorrow.
Insightful Quotes:
"Your job is not gonna be replaced by AI. It's gonna be replaced by a human who's using AI." - Michael Todasco
"There's a huge opportunity for people just to test and just to play and be experimental - be childlike in the things that they're doing. Nobody knows how to prompt engineer from the start. It's not necessarily intuitive, but just see: if I put in X, I get Y. And iterate again and again. This is how human beings learn." - Michael Todasco
"You cannot outsource your competitive advantage. That is something you have to own and manage. Otherwise you're going to be held hostage - it can't be a black box." - Seth Earley
Tune in to discover how embracing generative AI today - through experimentation, proprietary knowledge building, and smart human-AI collaboration - will determine who leads and who falls behind in every industry.
Links:
- Twitter: https://twitter.com/todasco
- LinkedIn: https://www.linkedin.com/in/todasco/
- Website: https://medium.com/@todasco
Ways to Tune In:
- Earley AI Podcast: https://www.earley.com/earley-ai-podcast-home
- Apple Podcast: https://podcasts.apple.com/podcast/id1586654770
- Spotify: https://open.spotify.com/show/5nkcZvVYjHHj6wtBABqLbE?si=73cd5d5fc89f4781
- iHeart Radio: https://www.iheart.com/podcast/269-earley-ai-podcast-87108370/
- Stitcher: https://www.stitcher.com/show/earley-ai-podcast
- Amazon Music: https://music.amazon.com/podcasts/18524b67-09cf-433f-82db-07b6213ad3ba/earley-ai-podcast
- Buzzsprout: https://earleyai.buzzsprout.com/
Thanks to our sponsors:
- Marketing AI Institute
- CMSWire
- Earley Information Science
- AI Powered Enterprise Book
Podcast Transcript: Generative AI, Human Connection, and Competing in the Age of Creative Machines
Transcript introduction
This transcript captures a wide-ranging conversation between Seth Earley, Chris Featherstone, and Michael Todasco about the practical realities of generative AI adoption - from Mike's firsthand experiments writing AI-generated books and comic books, to the organizational challenge of knowing when to automate and when to preserve human touch, to the unresolved copyright and authenticity questions that every creator and business leader needs to understand.
Transcript
Seth Earley: Welcome to today's podcast. I'm Seth Earley.
Chris Featherstone: And I'm Chris Featherstone.
Seth Earley: Our guest today spends a lot of his time thinking about AI and how AI is going to change how we create and how we innovate. He has a lot of experience in product innovation and product launch planning. He worked for a number of years at PayPal, so he's been fresh out of the tech space. He is author of a book called Artificial America, which was mostly written by artificial intelligence. So please welcome our guest today, Mike Todasco.
Mike Todasco: Hey Seth, I love to be here. I'm glad you're recovering alright, and I'm glad to be able to chat with you both today.
Seth Earley: So Mike, give us the lay of the land. What kind of stuff you've done, what got you to this place. I know you really focus a lot on generative AI, and of course that's all the rage these days. But how did you get to where you are?
Mike Todasco: Yeah, good question. So going back to 2021, I got the bright idea - I was still working at PayPal running innovation for them - I got the bright idea that I wanted to focus more time and effort on writing. So I decided to go back to school. I'm 43 years old, or whatever it was at the time, and I'm like, well, what the heck, I need to go back to college. So I started doing that part time.
Seth Earley: Youngster, by the way. 43. I call that 33 - that's a very late 30. So you're really in your 30s.
Mike Todasco: I love it. I just really wish I was born on like leap year, you know, on February 29th, because then it's really just one birthday every four years. But so look, I decided to go back to school. And in everything I do - I read about 50 books a year. At PayPal I would interview authors all the time coming in. I just love this world. I just love being absorbed in it, and I think I wanted to do more creating myself, more writing.
Seth Earley: What a great job heading up innovation is - you get to talk to cool people about crazy ideas all the time.
Mike Todasco: That was an amazing job which I did for probably 6 years at PayPal. It was brainstorm sessions, bringing in authors, bringing in entrepreneurs, creating crazy stuff - kind of doing anything I wanted to do to inspire innovation across the company and bring forth more innovation. Anything tangential to innovation, I got to do. So it was pretty awesome.
Chris Featherstone: So one day what - you were eating lunch and like, I want to write a novel? What kind of writing did you want to do?
Mike Todasco: Pretty much everything I read is non-fiction. Every 50 books a year - it's psychology, it's business and leadership, it's a lot of biographies and all that. So that was the initial thought. But I think the greatest writers of non-fiction are people who do have roots in just being able to tell a story - people like Michael Lewis, who can craft a narrative and make it feel like a different world.
Seth Earley: Right now you're writing a lot about ChatGPT and generative AI. Tell us about Artificial America.
Mike Todasco: Yeah. So to finish that point and tie it exactly into your question, Seth - one of the first things I did was I started playing with Sudowrite, which was built off GPT-2 at the time. I remember I was actually writing a piece where a computer became President of the United States - an AI effectively becomes President. This was the AI's inaugural address. And I felt it was perfect to have an AI write this. I remember just running upstairs, screaming and telling my wife like, oh my God, it's writing for me. Look at this! This is garbage, but this is golden - just knowing then and there like this is something different and magical. So that was my first experience with generative AI. And I've been active on Twitter, just seeing the things people were doing with DALL-E 1 and Craiyon and all of these other tools, just being blown away. The thing that excites me about prompt engineering and generative AI is really the democratization of it - the fact that you don't need to be a data scientist, you don't need to be an engineer. These are tools that you can prompt engineer using real language and get the outcomes that you want. You're only limited by your own creativity.
So I like to do a lot of experiments in this space. The book Artificial America was kind of a Christmas holiday project in 2022. I was like, well, can AI write a book? What's it going to take to do that? So I used GPT-3, created a series of prompts, and the hook was that I wanted it to write a local story about every US state and territory - hence Artificial America. It still has difficulty doing long form, which is why short stories work well. For long form it's forgetful - there's a lot of ways to get around it, but it's still not good at tying something it mentions in Chapter 1 back into Chapter 14. That's something humans are good at that makes writing special. So that was the impetus of creating Artificial America - working the prompts through there, having it also generate the images, putting it on Amazon and just starting to sell the book and seeing what happens.
Seth Earley: I loved one of the things you mentioned - that people are afraid they're going to lose their jobs to AI, but instead they're going to lose their jobs to people who use AI.
Mike Todasco: Yes. And I'll be honest, I don't know who came up with that quote originally - that person deserves a whole bunch of money. I can't take full credit for it. But I do say it all the time - you know how it goes. The first time you say it you cite the person, the second time you cite them, the third time you say "as I always say." I think I'm at the "as I always say" phase.
But look, I think the magic of this is that AI is making people better at what they do. I just saw someone send me an article talking about how teachers more and more are using AI for lesson planning, to do all these administrative tasks that take a lot of time and effort.
Seth Earley: I just did a workshop for a charter school - the Phoenix Academy. They do a great job with underrepresented populations up here in Massachusetts. I talked to them about exactly that. What types of things take a lot of your time? Writing job descriptions, writing lesson plans, doing the types of things that require some research, maybe a little bit of drudgery. But the most important part is the prompt - being able to ask the right question. And if you don't have enough knowledge to ask the right question at the right level of detail, you won't get a good answer. I showed examples of here's when you ask a generic question, here's when you ask a very specific question, and here's where giving the tool more context gives you a richer, better answer. Therefore you need to know the subject. And you need to verify what's in there - you can't just turn this stuff over to these tools. You have to be smart enough to know what to ask and then be able to judge and validate the results.
Mike Todasco: Yeah, I want to build on that point. I mean, the beautiful thing about this is there's a huge opportunity for people just to test and just to play and be experimental - be childlike in the things they're doing. Nobody knows how to prompt engineer from the start. It's not necessarily intuitive. But just see: if I put in X, I get Y. Well, I don't really want Y - I want Y to look more like this. So I'm going to change that prompt to do this. And just iterate again and again and again. This is how children learn, this is how human beings learn. And I think the one thing I tell everyone about these tools is just play with them. Literally put 30 minutes on your calendar and say, hey, I'm planning a trip - I want to use ChatGPT or Bing to help me plan that trip. That's something I've literally been doing over the past couple of days, and it's amazing at doing that because it's aggregating all this data and knowledge and bringing things together.
The one thing, if nothing else, for people who are listening - if they haven't played with the tools, or if they said "I tried it once and it didn't really work" - go deep into it. Just see what happens and explore. Large language models are effectively limitless in the types of things they can do. You will potentially be the first person ever asking this type of question or giving this type of prompt - and that's pretty amazing. We're still at the early, early stages of this, and this is a way for all of you to get in on the ground floor. Imagine if you don't have to spend 30 minutes figuring out how to phrase an email. You could just put a prompt into ChatGPT, say what you want to convey, get 80% of what works, and save all that time to spend with your family, do more work, or do whatever you want.
Chris Featherstone: Let me ask you this - at what point do you believe you start to lose authenticity? Not only in the content you create, but also when it comes to fact versus fiction. At what point do we start to lose the authenticity of it?
Mike Todasco: Chris, it's a great question. And I think that line today is not going to be the same place where that line is going to be in 5 years, as we become more and more comfortable with these tools. So I "wrote" these AI books, but I did them under a pen name. I asked the AI to create a pen name for me, and I made sure it had the initials AI. I said, give me 10 different male names with the initials AI for a writer that doesn't already exist today. Alex Irons was one of them, and I bought alexirons.com - that is my AI personality. But I would never feel right using that name myself. But I feel completely okay using Grammarly all the time - I don't necessarily disclose that. Grammarly says hey, you're using too many passive verbs, here's a way to fix that.
Seth Earley: Yeah, it's almost like it's going to get to the point where spell check doesn't affect your authentic voice. I'm a terrible speller - people used to know that. Now they don't see that.
Chris Featherstone: Part of this too is - do organizations necessarily care how the sausage is made, or do they care about the output? Who cares if something else helped create it - it was still yours. You created the process. So now it becomes your thing. The ownership and creation of it matter, regardless of what engine is underneath it. That said, there's a statistic I read this week - we've got 20% of 18-year-olds in this country who have less than a sixth grade ability to read. Taking out the non-native born population, we're still left with 43 million people who are actually illiterate. And what can we do as part of that? There's this growing gap between those who will lose their jobs to people utilizing technology and those who won't.
Mike Todasco: Yeah, and I see it. And you bring up literacy, and it's funny - I was listening just this morning to a podcast where two tech people were talking about white papers, and they're like, who can actually read these things? What's going through my head is, these are great things where large language models can engulf these papers, summarize them, or I could say to GPT-3, given this paper, explain it to me in one page at a fourth grade reading level. It's not going to get everything perfect, but it goes from almost being too high a barrier to having a much better understanding. And that's accessibility.
Seth Earley: I think that's right. And once you have accessibility, you have an entry point, and the entry point can lead to additional investigation. If you don't have that accessibility, you're not even going to have that entry.
Mike Todasco: Yes. And think of Google Translate in Mexico - none of us are native Spanish speakers, but it opens up a whole other world. Can you get other technology to take a white paper and make it so theoretically you don't have to be a scientist to understand it? The number of people who actually read through academic papers is astonishingly low. It's lucky to be a couple of editors of the journal, probably the person who wrote it, and maybe their parents. That's about it.
Seth Earley: I agree. And I think what we'll also increasingly see out of generative AI will be learning and courses that are attuned to the capabilities and knowledge gaps of individual students. Using this as an effective tool to meet people where they are.
Mike Todasco: And the comfort level question is interesting too. Where is the real value-added customer touch point? Where is it critical that we're human, that we're personal, that we get this thing right? And where - for password resets - okay great, let's have that automated.
I remember when I was at PayPal, huge customer service organization. We would always talk about those kinds of things. One day I was in a meeting and I said, we need to step back and decide who we want to be in customer service. I asked, who's ever called Zappos? No one had ever called Zappos before. They're legendary customer service. So I'm sitting in a conference room, I literally dialed the Zappos phone number. Someone picked up and said, "Hey, Zappos! How's it going?" I start talking to them - they're based out of Las Vegas. I said, hey, I'm going to Vegas in a few weeks. Where are some good places to eat? And I just kept going on and on. They never once asked me for my order number. They never once said, why are you calling me? Because for Zappos, they're just there to please the customer. I don't even think I'd ever personally bought anything from Zappos. They didn't care. That is not who they wanted to be. That touch point is so critical that now I've told that story to thousands and thousands of people after that 5-minute interaction. That is who Zappos wants to be. And it's not to say every company needs to be that in all circumstances. But you need to do a journey map - okay, what are all the customer touch points? Where is it critical that we're human, that we get this thing right? And where, for password resets, can we let that be hands-off?
Chris Featherstone: You're bringing up the crux of it - how much empowerment do you give that single entity the ability to solve the problem themselves, whether it be self-service or a call center agent? Because I've done the same thing with Zappos, and it was empowered from on high for them to be able to just solve the problem with the customer. Don't worry about what that means. Because 9 times out of 10 they'll come back in droves and tell a thousand of their friends about their great experience.
Seth Earley: And the question really is, how much will generative AI be able to fill some of that gap - give the feel of a human conversation - but within certain constructs. We need to leverage the language models with the curated language model and curated content of an organization. How do you retain that level of conversational fluency and ability to sound human and solve a problem, but stay within the guardrails of what that organization does?
You could build a bot that would act more like the Zappos customer service agent and talk about anything, but that's not necessarily what an organization wants. They want to handle the low-touch requirements, but do it in a more humanized way. And automate completely the things that don't require touch. So the point here is - how will we use those models with curated content, with IP that needs to be protected, with the specific processes of an organization? Large language models don't solve your knowledge problem. A bot and automation still needs to be able to access curated, classified, tagged knowledge assets. And organizations are very wary about depending on tech giants for their core functionality. We'll see more large language models being developed in house so there isn't that dependency. I've seen people try to outsource something they consider a competitive advantage around customer service, and you cannot outsource your competitive advantage. That is something you have to own and manage. Otherwise you're going to be held hostage.
Mike Todasco: You need a lot of data. That's the one thing we know. Why has GPT-3 been so much more efficient and successful than GPT-2? Part of it is that it's devoured exponentially more data, and the corpus of information it has is so much higher. I think you're going to see organizations take every single sales call they have, transcribe it voice-to-text, put it in a database, and be able to map those sales calls - these were ones that led to a successful close, these were ones that did not. What are the elements of that, in a way far beyond anything humans can ever possibly do? Maybe 60% of what it spits out will be really interesting insights and 40% will be absolute garbage, and it's up to human beings to say, okay, we could tweak the script this way, and just iterate and rinse and repeat.
I think for it to work internally, you have to create your own corpus of data. If you're selling gummy bears and you're reaching out to retailers, GPT-3 by itself is not really going to be able to do that very well. It might have some sales techniques mostly from watching movies - it might not actually work well in reality. So you have to create your own corpus. There's just going to be more and more effort internally to actually create that data and then be able to use it in these large language models.
Seth Earley: That's been a lot of our focus at Earley Information Science - building out those knowledge bases and componentizing the knowledge so it can be reused, building the reference architecture and reference ontologies that can leverage that content. We still need to bridge the gap between a highly specialized language model and a generalized one. And it comes back to harvesting those assets, classifying and tagging them, and being able to surface them for particular use cases.
Chris Featherstone: I'd love to get a take - you said you generated some images for Artificial America. How did they turn out?
Mike Todasco: So the honest answer is I let GPT-3 and ChatGPT create all the text, and then I said, given this, give me a prompt I could put into Midjourney. It didn't necessarily always know what a prompt was - sometimes it would give me something that would look like an effective prompt, other times it would give me a summary. DALL-E is also owned by OpenAI but GPT-3 was never really trained on prompting into DALL-E, so that full circle was never fully there.
That said, I did another experiment more recently where I went into a large language model and said, can you create a comic book script with a twist ending? It did that, then I had it make the artwork. Basically everything created by AI. And I did that again just in February 2023, last month - and in that time I would say it was probably 4 times better than it was in October 2022. I don't know if that's the true acceleration - 4X better in 4 months - but the images coming out are so much better. There are still artifacting issues with hands and all that. But if you just look at the improvement curve we're seeing in generative art - it is incredible. Midjourney Version 4 has just been really amazing.
Chris Featherstone: There's a definite gap between where language models are and where pictures and imagery are. And there's a huge divide between the two. For instance, we were doing some of the same things and asked for a picture of a woman. As you can imagine the corpus of images out there for women is wide-ranging, but it brought back something akin to a magazine model perspective. And there was a bunch of discussion around - well, that's broken. But that's sorry, that's the corpus of data that's out there on the web right now. So if you want to make it better, generate more images that it can learn from that are more representative.
Mike Todasco: Yeah. And there was a really great YouTube video where they had 3 people basically generating the same type of image. One was new to this, one was someone who'd played with generative AI, and one was a true artist. Then they put the results on Amazon Mechanical Turk and had people blindly vote on which they thought was better artwork. By far the artist's work using the exact same tools was vastly better - because they know detail, they understand the style, they know how to describe what they want. And that's why this isn't AI replacing you. It's AI amplifying the skilled person. There is still so much work in saying "give me a painting of a woman" - what is she wearing? What does she look like? What is the intent? You have to know how art works, the style, all of this. The person running this on the other side really does matter as far as the quality of the output.
Seth Earley: I think that's true of any use of generative AI. You really do need to understand the topic, the right questions and prompts, and you need to know more about the topic so that you can judge the quality and refine it and edit it. People that are just going to be putting stuff in and using it as-is without enough background knowledge will very quickly find that content will not be effective or achieve the objectives.
Mike Todasco: Yes. So the way I think about this - do you know the Library of Babel? It's an online tool where you can type in any sentence of about 50 characters, and that sentence already exists in this library. Any 50 characters you can imagine - because there are only so many options, trillions and quadrillions, but still: you have enough monkeys on typewriters and eventually they'll write Shakespeare. True with images as well. These tools can produce any image that ever has existed or ever will exist. And part of it is just the magic of working these prompts to get it to match the vision that you, as an artist, as a creator, as an entrepreneur, as a business person - want to see exist. You just have to find a way to get it out of there.
Seth Earley: Yeah, and the combinatorial explosion is just truly vast. This stuff is trained on content produced by humans, and you can prompt it to produce art in the style of well-known artists. How does that happen? Because they've trained on that artist's work. And there are lawsuits right now about misappropriation of intellectual property and copyright. How do artists get compensated? How do they prevent their work from being appropriated?
Mike Todasco: Yeah, look, this is a legitimate concern. Before I joined PayPal I was actually running an artwork marketplace - literally selling pieces of comic book artwork from artists. This is very near and dear to me. I know lots of artists and creators who are very passionate about where this is going. You know, Greg Rutkowski was the epitome of the problem - you'd put his name into a prompt in earlier versions of Midjourney and it would give you these beautiful works in his style. It kind of became the de facto standard of like, I want something to look awesome, just throw his name in there. And he wasn't happy about this. He shouldn't be.
I think you have Getty Images and many more that have already opened up lawsuits against Stability AI, who runs Stable Diffusion. The fact that Getty Images is saying, hey, you can actually generate an image that has a Getty watermark - that tells us you were probably trained on our images.
For starters, the internet has a robots.txt file - if you don't want any website to crawl you and put you in a search engine, you just put this on your website and it's off limits. I think there should probably be an AI.txt for starters. But that doesn't necessarily solve the problem - these models have already been trained. And what if I'm an artist who's really influential, and actual human artists are just copying my work and making clones of my style? You could sort of reverse-engineer my style even without being trained on any of my specific pieces. These are probably problems where the artist should be compensated. I don't have the answers. This is going to play out in the courts. But it is a legitimate concern for anyone who is creating and putting things up on the web. If you're Rembrandt, fine - public domain. But for any living artist whose work is not yet in the public domain, this is a real concern.
Chris Featherstone: I think it goes back to the authenticity piece. And some of the next generation of the web is going to be about guaranteeing authenticity - hey, this is mine, I can shut you down from using it if I don't want you to. The biggest form of flattery is to copy, and people are gonna do it. It's just a matter of whether they want to gain from making money off those copies, or make fraudulent claims. It's an iceberg - we've only seen the tip.
Mike Todasco: No, we don't know. You're getting into questions of fair use and things like that. So for example, I love Alex Ross - comic book artist, does these beautiful paintings. I went in stable diffusion and wanted some self portraits, said do it in the style of Alex Ross, and my gosh, it did it - and it was amazing. But am I monetizing those? No. Am I selling copies? No. Is that okay? I don't know. I feel like I've given that guy so much money over my life. But he never explicitly said it was okay. These are tough, and there just aren't any easy answers right now.
Seth Earley: Well, we're coming up to the end of our time today. Do you have any final thoughts, Mike, in terms of where you see the industry going, or where organizations need to be paying attention?
Mike Todasco: Look, I would say everybody out there, no matter who you are or what you do, you need to become familiar with these tools. I heard about a Congressman who actually went back to school to study AI at 71 years old - getting a degree in AI. He said this is probably the biggest thing since fire. Maybe that's hyperbole. But look, this is real. The last few hype cycles with Web3 and other things - there were always naysayers. I see very few people saying I don't know the use cases of this, or I don't know how this is really going to change lives. This seems abundantly clear and obvious. So for whoever you are, wherever you are, just start using these tools. Use them in your work. If you work in a spreadsheet all day - there are amazing things people have done in spreadsheets using GPT-3 that can potentially save you hundreds of hours every year, from building formulas to bringing in data to doing all these other things. And that's just one little example. I don't think there is a single profession that won't be impacted.
And if not at work, in your personal life - I said before I'm using this to travel plan right now and it's doing an amazing job of it. I use it to do things with my children all the time. Sometimes my 9-year-old daughter is bored and she's like, hey, can we go into Midjourney and start making pictures? And I say yes, let's see what you want. I think it's really important just to get familiar with the tools. If I might give anybody homework out there - that is the homework. Do that. Because the more people we have playing with these tools now, the better we understand them, the better the outcomes are going to be. Let's just not have a bunch of pure technologists working these tools and making these decisions. We need masses of people, we need the public looking at this so we can all collectively say, okay, what is it that we want these tools to do and not do?
Seth Earley: Well, this has been great. Thank you so much for your time, Mike, and it's been a pleasure speaking with you. We'll look forward to staying in touch and seeing your writings and speaking about this very important topic. Thank you again today.
Chris Featherstone: Mike, thank you so much.