S4 #8 Calling all prompt engineers to tickle machines in the new AI era

[00:00:00] KRISTINA: Deep learning, at its essence, learns from examples the way the human brain does. It's imitating the way humans acquire certain types of knowledge. What does the future business look like, and what does this mean for your organization? We'll find out today.

[00:00:14] INTRO: Welcome to The Power of Digital Policy, a show that helps digital marketers, online communications directors, and others throughout the organization balance out risks and opportunities created by using digital channels. Here's your host, Kristina Podnar.

[00:00:30] KRISTINA: Hi everyone. I'm happy to have you with us again on the Power of Digital Policy. Thanks for joining us. With us today is Michel Ballings, an associate professor, a James and Joanne Ford faculty research fellow, and the director of the JTV Center Intelligence Lab at the Department of Business Analytics and Statistics, Haslam College of Business at the University of Tennessee. He is the President Emeritus of the Informed Social Media Analytics Section and a research scientist at Amazon through the Amazon visiting academics program. Michel has many engineering and deep learning teaching accolades and is highly regarded in the industry for his focus on improving long-term financial business outcomes through the development of end-to-end machine learning systems for large-scale data. Michel, thanks so much for coming today and just hanging out with us. It's great to have you here.

[00:01:22] MICHAEL: Kristina, thank you for having me. I'm very excited to be here.

[00:01:26] KRISTINA: I'm so excited because recently, you published an article titled "The possibilities and risks presented by generative AI," where you said something along the lines of "the term game changer is often overused and shouldn't be taken lightly." As I was reading you also mentioned that it's an appropriate term to use right now in the context of generative AI, let's just dive in. Tell us why right now is the time to use that term.

[00:01:54] MICHAEL: Well, Kristina, I think that's a great question. I think if you look back at major disruptions in the past, technological breakthroughs, you would think about electricity, and maybe even before that agriculture, you think about cars, I remember a story where there were like 300 000 horses in New York, and then they started replacing that with cars. It's not that far on the go and we started getting into information technology, for example, smartphones, and then you had social media. And right now we have this new thing, called generative AI. And generative AI is really different from the AI that we have known in the past decade. In that, it is producing open-ended outputs. So we're not talking about what is the credit score of this person. We're talking about open-ended unstructured information and because it is open-ended and because each time when we make that prediction or when we generate that, that outputs, it is also changing. So it is never the same thing twice. And that's why we call it generative. And so, why is this a breakthrough? Because, well, these models are so good that it is nearly indistinguishable from what humans would produce. So that means that are gonna get lost and new jobs are going to be produced. And, it has the potential to influence the entire economy. So that's why I say, it is a game changer. One example, you might have a whole department of copywriters, creating social media content for a company. Well, right now I can't imagine that department is going to shrink. And then the company is going to hire people who can write prompts or even, co-developers or AI developers. So that's just one tiny example of how AI is going to influence the economy in a big way. And so that's why all that it can change.

[00:04:16] KRISTINA: And so it's a game-changing time for a lot of organizations as well. Yeah, I've seen a lot of organizations scramble, quite frankly, not knowing what to do at this moment in time. What do you think they should be doing? I mean, is this a time to panic as we see some organizations do, or is this a time to lean in?

[00:04:34] MICHAEL: Oh, I think you should definitely lean in. I think if you don't lean in, you're going to be left behind. Companies that are just trying to ignore what is happening are going to be left behind. There are a lot of things, while technology is improving rapidly, there are also lots of issues with the technology, which we also talk about at some point. You could think of the technology as having issues of status, privacy, and toxicity. There are all sorts of problems with intellectual property. So I think if you're not thinking about this technology. If what it can mean for you as a company, and also what are the potential pitfalls, then you're just gonna going to be completed out of there.

[00:05:21] KRISTINA: And so that's an interesting point, right? I love talking about digital policy because I feel like it really does a great job of balancing out the risks and the opportunities and creating that ultimate push point of forcing you to deliberately think about the choices you're making and coming down on one side or the other, how does an organization go about understanding what are those risks? You mentioned privacy, which I think we've defined somewhat what privacy is. But when we get into things like ethics and what's ethical and what's right, How do we start to define those things in the context of generative AI? How do we understand what those things are?

[00:06:02] MICHAEL: Fairness is a big one, and it's, it's a hard one. Going back to our branded scoring example or default prediction, one AI model. If you think of an AI model predicting whether an applicant is going to default on their, and for that model, to be fair, you would want that model to not make differential decisions based on gender. You want to have a male applicant and a female applicant receive the same prediction, on whether they're going to default or not. And the model should not take into account gender. The same thing could be about, for example, age: older people versus younger people should receive the same prediction if they have the same credit score, same job history, that sort of the same trajectory leading up to this four-number decision. So that, fairly straightforward, Kristina, we could say we could limit the model to all the include relevant predictors, and you should include predictors such as gender and such as age, same for race, all these important predictors well, you would not include these predictors in, into the default prediction. You would only include the important predicts. I would say then you have the generative AI, generative AI is ,trained on a massive amounts of data, unstructured texts, they basically have gone to the internet, downloaded the internet, all the texts that they could find, and then they are putting that in a model. Now, think about that for a moment. You cannot go and look at all that text and go and remove biased data or toxic data, and so if you're then going to make a prediction you might get unfair generated outputs. For example, you might have a prompt saying on a beautiful afternoon, the Dr. X was about to, and then you give that prompt to the generative AI, and the generative AI is going to complete that, going to write a paragraph. Now, what would you think if the generative AI would always talk about Dr. X as being male, that the generative AI is basically perpetuating the idea that all doctors are male, and that is unfair. So what you could do is, well, you could then say, well, we want to put a constraint in the model that would make the model each time when it generates, it's kind of, you're flipping a coin. It should be half percent of the time it should be female and half percent, it should be male. You could do that, but, but then what would you do when you're not talking about doctors, but about firefighters, for example, or about nurses or about professors, teachers? It's very hard to define that it's not like that default prediction problem where it's very easy to say what is unfair. So, it's a very hard thing to do and it's even more subtle than that. What if the model, the generative model, when it talks about male people, it is always slightly more enthusiastic and more positive. And when it talks about females, it's always slightly a more serious tone or a more negative tone, it's quite a hard to detect that and to correct that. And so at first sight, it appears that the cost of having these open ended outputs and having these open ended outputs change each time when you run the same problems, the price of that is that we will have to tolerate some unfairness because it's not immediately clear or it is at least prohibitively expensive.

[00:10:37] KRISTINA: So does it become a situation where organizations need to be thinking about cost in the context of balancing out the technology with the human aspect, if you will, because I can actually use generative AI, and then I can layer on a human aspect to detect maybe some corrective action, and I can use a human perhaps to correct for some of that action and get us to the rest of the hundred percent that I need to get to, or how do we address that aspect, or do we need to always address all of that aspect, or does it depend on the use case?

[00:11:11] MICHAEL: Well, that's, of course, context is really important. Now, I think definitely you're right when you say it is about aspect, there's so many use cases and each time the output changes. So for the same prompt, you would have to have a human go through thousands and thousands of thousands of iterations of putting in the same prompt and getting the generated text and going through that and determining it's this particular text toxic or fair, or is there an IP issue or is this model hallucinating what, what's going on? So having humans do that manually is even though they have done some of that, but doing that at large scale is practically impossible. There are some other solutions, more automated solutions, for example, you could have guardrail models. What you're going to do is you're going to basically make an other model that is going to take as inputs, the output that's generated by the model, by the generative model, and then it's going to predict, well, what is the center? Or you have some sort of definition, even though it's going to be hard to define that definition, but let's say you have some sort of definition about fairness, you could have that guardrail model to predict whether the output of the generative model is being unfair. That's, that's one way to approach it. Another way is filtering and blocking. You could have models detect in the input text. On one of this chain, is there any toxic, toxic content? Is there any unfair content? And then we're going to remove that from the chain data and change the model again or refresh the model. So that basically is going to stop doing that. And then you could also do that at the output level. You could say, well, this is unacceptable. We're going to block that. We're not going to return that to the user. Another thing you could do is filter out prompts. So what the user of the model puts in the model. You're going to filter out unreasonable requests. You would not allow the model to engage in what the user is asking the model to do. So that's a, that's another way to block filter information. And so it is really about guardrails and building additional technology that can help us rein in that beast of a generative model that we're doing.

[00:14:04] KRISTINA: Some colleagues of mine have suggested that data providence is a way to solve this issue. And as you're speaking, I'm wondering, is that even a potential just because of the necessary magnitude of data? It just seseems that the guardrails are possibly the only way to his problem because data providence wouldn't even start to chip away at the issue. Is that right?

[00:14:28] MICHAEL: You're absolutely right with the default prediction problem, it's very easy. You could run some queries on it, you could get some statistics and you see if there's something wrong with the data, how the model is going to behave. But with generative models, the data is so massive and unstructured, you would have to go about and read all that! Who has the time to read the entire engine? You would have to have thousands and thousands and thousands, maybe 10, 000s of people doing that. It's already very expensive to have these models and create these models. To go there is really prohibitively expensive. So, I'd agree if you could fix the input data, the problems are gone. There's just no way to do that manually so you'll have to create some walls, some garbage walls to do that.

[00:15:23] KRISTINA: So Michel, I think you just gave us the answer. All we have to do is clean up the internet. Problem solved.

[00:15:28] MICHAEL: Yes. Maybe we should fix that. Yeah.

[00:15:33] KRISTINA: So if you could solve that problem for us, we could be finished by Tuesday.

[00:15:38] MICHAEL: Yes. I think you see other thing is watermarking, watermarking. If we would have a way, one of the problems is, we're going to just run these models and we're just going to flood the internet with generated content. And so if that content is toxic or unfair, we basically have an unlimited way, a much cheaper way to produce that content than just having humans. So we could flood the internet right now today. We could double the size of the internet in a couple of years. So it's going to be exponential. The size of the internet is that is, that's a major problem. Now it would be less of a problem if we would watermark all of it. And you, you may think of watermarking when you think about images, when you have this image, you would have somewhere in that image. You would have this mark where you can clearly see, well, this image is generated by generative AI. But it is much harder to do that with the text. But there are some suggestions to do that. And so you could think of a generative AI as you would have a sentence. And then the goal of the generative AI is to predict what is the next word. So you would have let's say you have 50, 000 words. And so you have the sentence, for example, in a galaxy far, far away, then you have 50,000 words to choose from. We're going to choose one of these words, whatever the best word is, according to the model, we're going to plug that into the next. So in a galaxy far, far away, the evil then takes the next word out of 50,000 words. And buyer and so forth, thinking about Star Wars. And so you would continue doing that. You would roll that out. And so you get a whole paragraph. Now, how are you going to then watermark that? Well, one way is to each downloading have to pick out of those 50,000 words. You're going to randomly divide these 50,000 words into groups. You have the group, the green group, and you have the red group, and the model is only allowed to pick from the green group. So that means that the entire paragraph or text per sentence that you have generated is from green words. Now, if you think about that, when a human produces text, he or she is not going to only restrict themselves to the green words. He or she is going to select from the red words and from the greens. If we have that restriction, when we generate it through the model and it is going to be a different story and we can detect that. How do we detect that? Well, we just seeing the text that any texts. And we want to know, what is the probability that this text has been generated by an AI? Well, we just compared it at each position of the first word, what was the green and the red list? Does it come from the green list yourself? We go to the second word in the sentence. Does this come from the green or the red list? Or that the second position, we go to the third position, and we do that. And so we get a probability that or we get the challenge of how many green words are it and how many red words. And if we, if we create that, we calculate how many times there are words off the green list. We are basically able to tell whether this text has been generated by you or by an AI. Of course, there are two restrictions here. Well, the first one is. The AI, whoever produces that model, let's say OpenAI, has to remember what that green envelope list was and has to provide an API where we can put that text in that API and get that, the count of words in the green envelope. The other restriction is of course, what if somebody just has a model, creates a model somewhere that's not, that does not come from a system provider. And then they don't have that green or red list. So there are restrictions, but I think since these models are so expensive, I think the watermarking would be definitely a step forward. As long as we can identify what comes from AI, what consuming humans. I think that would definitely help in making things better.

[00:20:47] KRISTINA: How long do you think we have before watermarking comes into the scene? Thinking about timeframes, are we two years away from that being a reality, or do you think it'll be sooner?

[00:20:59] MICHAEL: I think the technology is there. We know how to do it. It's not hard to do it. But I think it's going to be it's going to come about in two ways. It's either going to be some law that requires companies that run AI to provide that service or the companies are not required to do it by law, but they get to charge people more.

[00:21:27] KRISTINA: I think people are very excited, but it seems we're so far away from commercializing. generative AI, and yet we're teetering on the brink of it. Is it just another hype cycle from your perspective? What should I be expecting from a true timeline perspective? What does this mean for me and my organization if I'm listening to you?

[00:21:51] MICHAEL: Well, and, and is that question particularly concerning commercialization or is it still the watermark?

[00:22:01] KRISTINA: I guess both commercialization and watermarking because I think for a lot of folks are probably thinking, how soon can I commercialize this? And then the next question is really going to be. Watermarking and the regulatory aspect that's coming at me because of the moment I commercialize it. I'm either going to be regulated, or I'm going to have to self-regulate; it's just a matter of time. And I think everybody already has their EU caps on. It's just waiting on what will be happening from a regulatory perspective anyway.

[00:22:29] MICHAEL: I think regulation is an issue. You might think I don't want to be investing too much and then it gets regulated away. But at the same time, do you want to get that behind? So I would say, don't wait, just do it and we'll see what happens. I don't think it's going to be over regulated because of course there is that arms race between different nations, we want to dominate in AI that's regulated too much, then we basically lose out the race to, to win this game. In terms of commercialization, I think you , I was just talking last week to somebody. And it, it's it's already there. Everyone can do it. For example, just think about social media before, if you wanted to be an influencer in anything, you would have to go and produce content. You would have to go make videos, buy a camera, take pictures. Maybe think about the travel, travel agency or clothes, a fashion, and then you would post regularly on Instagram, Tik TOK, whatever platform you want and then you would get followers and likes, and you would interact. The cost there was creating content and you're only limited by how much content you can create. And so creating content is expensive and takes a lot of time. Now, what was that person doing every day? Sits at the computer, writes a prompt, create a picture from a beach in Thailand. People's snorkeling, shelling, having a good time. So basically she was just typing basic prompts as he was generating 10, 15 images every day, and it's virtually free for that person to do that. Post that on Instagram, Instagram account is, is growing. People are coming to whatever she's selling. So it's already there. You just have to be a little bit creative, but think about this is what I'm doing or want to do, and how can AI help me do that at scale and at a massively different price point.

[00:25:05] KRISTINA: And so what strategies can organizations employ from your perspective to educate employees and even their stakeholders about the role right now and the implications of AI and deep learning within their operations, because it really is a mindset change?

[00:25:21] MICHAEL: It is. I think lots of people have not even used AI once, so they don't really know how it works and they try to ignore it. So I think the first step is to do some training on this. What is AI? What can you do with it? And here are some use cases. I think that's the first step. The second step is: there should be a central document somewhere where people keep prompts. The game here is to who can write the best prompts. If you can write the best prompts, you will get the best output. And so the cost here is creating these prompts. The cost is not anymore creating the outputs. None, it's not about creating the computer generated content. It is about who can write the best prompt that will generate the best output. And so the game here is. Now, it's not the output, but it is the inputs that you want to control and that you want to iterate on. So you wouldn't have to have some sort of, for example, you'd have a Google sheet or an Excel sheet. You put in the prompts, you say, this is a good prong. This is what it, what it does. And these prongs can be very complex. They could be a whole paragraph long and it could be the difference of one word that delivers a great output versus a so so, so I think. If you do, if you have training where in your company and you have that central location where you have prompts and you experiment with these prompts, I think that is the best thing you can do now as an organization to be ahead of the game.

[00:27:11] KRISTINA: And so also you are in education. Your entire job is to educate as well as to innovate and create. As we think about younger generations, I have a teenager at home, and one of the things that he's been a little bit surprised at is at school, they've banned them from using ChatGPT. They're not allowed to use it. At home, I've been pushing him to try to use ChatGPT but to try to change prompts to be highly specific. If you give a very generic prompt to ChatGPT gives you back garbage. If you say, Hey, help me write an email asking for an internship. It just gives you junk. If you ask for a very specific problem, like help me write an email around an internship or for sports marketing in a software organization for this amount of time with this background, et cetera, then you start to get a more specific tailored email that you can take and iterate on. You're not going to copy-paste it, but it'll give you a really good base that you can work with. It'll save you time; it'll be helpful to you. It's a thing. How do you suggest that we go about helping? What should we be even doing for the next generation that will have to live in this world? That's not of my generation.

[00:28:21] MICHAEL: I think that's, that's a wonderful idea. And so there's the whole AI wave has created one, I mean, many new jobs, but one specific job, which is really interesting and it's called prompt engineer. So you now have people who are hired and pay really highly to just generate prompts and they get this feel about what are the best prompts to get the best outputs. And I think that's a really strange moment. I think it's also really interesting. And so that, that underscores the importance of the game is not who creates the best. Who is able to create the best outputs by using some software or generating images, or where's the best camera to take images or where's the the best method to do something. It is really about who is creative and in sort of tickling the machine to produce something. I think, I think that is definitely something to watch out for. So, I think every, every company could use a prompt engineer and sort of is the keeper of all the prompts and people can go to that person and say, basically like what you just did with with your daughter is saying, well, I need to get this. I need the machine to produce this. How could I approach this? Oh, easy. This, this, this, and this. I think that's definitely very, very interesting evolution.

[00:30:04] KRISTINA: Michel, I'm writing this down. I'm saying the job title is the keeper of the prompts and the job description is tickling the machine. So, for everybody who's listening, I think you have the copyright on that one. But I think that there's something there. What else should people be thinking about in terms of maybe new job skills or job types that are coming out of generative AI and deep learning? Because this certainly is a new area, the entire prompt writing creativity. And it's almost like a new art form; this keeper of prompts is a new art form. What else might be coming at us that we're not thinking about?

[00:30:41] MICHAEL: I think this is going to create a whole industry. It's not just about the prompts, but it is also about people who are, who can detect certain issues with the output. It is about developers who can create guardrail ones. It is about developers who can tune these AI models. Generative AI is just is this is it deep learning. And so we need people in deep learning and reinforcing learning technical skills, math I think these are all computer science. I think these are all very important. I think we shouldn't really analyze people who are using these technologies. For example, I tell my students, you can use AI if you want to. Even during the exam, I walk around in the classroom when I'm taking an exam, the exam is about deep learning and I watch students typing in prompts and I see the output that these that these models produce. Sometimes the output is really good and I see them literally copy pasting me the answer of the AI into the answer sheet. And I say, good for you. You can use the internet. You can open book. It's open notes. I will change my exams in function of that. So I would say whatever you can do to take advantage of your environment will make you successful. It is the survival of fitness. And so we should teach our children to use whatever they can to be successful. Of course, there are boundaries. For example, plagiarism is very important. If you are asked to produce a paper about something, then you cannot just go and copy-paste that from somewhere else. If you're going to be marked on it. It also means that you cannot use another human to make that for you. It also means that you cannot use AI to go and make that for you, be marked on that. So there are some boundaries there, but in general, I'm a big fan of survival with this. Use whatever you can to be successful.

[00:33:06] KRISTINA: I love it, Michel. That's a great way to sum up, I think, balancing the risks and opportunities while also just taking it full force in the new space, in the new realm of deep learning and generative AI. It has been such a tremendous pleasure, and with deep gratitude for being here today and for your time. Really appreciate it. Thank you.

[00:33:27] OUTRO: Thank you for joining the Power of Digital Policy; to sign up for our newsletter, get access to policy checklists, detailed information on policies, and other helpful resources, head over to the power of digital policy.com. If you get a moment, please leave a review on iTunes to help your digital colleagues find out about the podcast.

‍