---
id: "workshop:what-the-heck-is-chat-gpt-how-generative-ai-will-impact-government"
type: "InnovateUS workshop"
title: "What the Heck is ChatGPT? How Generative AI will Impact Government"
url: "https://innovate-us.org/what-the-heck-is-chat-gpt-how-generative-ai-will-impact-government"
date: "2023-05-25"
duration: "90 minutes"
language: "English"
instructors: ["Alexis Bonnell"]
recording: "https://www.youtube.com/watch?v=LjxUjdsXgHQ"
transcript: "automatic, from the recording"
source: "InnovateUS (innovate-us.org)"
licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org"
---

# What the Heck is ChatGPT? How Generative AI will Impact Government

InnovateUS workshop · 25 May 2023 · 90 minutes

Join Alexis Bonnell, Former USAID Chief Innovation Officer, and Google Emerging Technology Evangelist and Artificial Intelligence specialist to help demystify AI, and to explain generative AI tools like ChatGPT and Bard and what they might have to do with government. In this workshop, you will not only understand and discuss these AI tools, but will also get a chance to experiment with them, and become more confident in how they work, their boundaries, and where they might apply to you and your work. We will review key AI use cases for government and explore how generative AI might shape government in the future. Bring the questions you have, the rumors you heard, and join others in an informative and lively talk.

## Learning goals

- Understand and discuss AI tools, experiment with them, and become more confident in how they work
- Review key AI use cases for government and explore how generative AI might shape government in the future

## Speakers

- **Alexis Bonnell**, Former Head of Innovation, USAID

Recording: https://www.youtube.com/watch?v=LjxUjdsXgHQ

## Transcript

Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/what-the-heck-is-chat-gpt-how-generative-ai-will-impact-government.txt (plain text) and https://innovate-us.org/what-the-heck-is-chat-gpt-how-generative-ai-will-impact-government.json (every segment with its start and end second).

**Alexis Bonnell**

[0:00](https://youtu.be/LjxUjdsXgHQ?t=0) Thanks so much. So, y'all, it is so neat to be here. I'm already recognizing some public servants that I so deeply admire in the chat. So, just thank you so much for coming and for sharing this time together. As was mentioned, I too am a recovering public servant and someone who's actually going back to public service. I'm not allowed to say where yet, but in about two months, I will be rejoining your ranks. And really, the thing I love about our time together today, some of you may have, we're going to have probably a real breadth of understanding of AI. Some of you may be deep technical experts, and I will tell you up front, I am not going to get into the deep technical elements of AI. This was much more about me trying to contribute as someone who is not a coder, not an engineer, right, who just did kind of the business of the mission of public service in a variety of ways. It's really about my journey with AI. So, about three years ago, I had run some AI projects, but quite frankly, was still very much a novice, did not know very much. And it was one of the reasons why I went to Google for three years, because I wanted to peek behind the curtain and I wanted to understand what was real and what was hype and what needed to cost $200 million and what should be free and all of those types of things.

[1:12](https://youtu.be/LjxUjdsXgHQ?t=72) But more importantly, it was to be able to learn and become comfortable with tools like AI. So, today, I'm going to take you a little bit on that journey. And my hope is not only that you feel more comfortable kind of experimenting, playing, and interacting with these tools, but that you also really have an opportunity to inspire others, you know, to kind of demystify and lower the anxiety level maybe a little bit on these tools. So, welcome, everyone. As I mentioned to Hannah, I'm really happy for you to, you know, add in questions or comments. Everyone here has probably really deep and rich experience with the subject matter. So, please add your observations. Feel free to stop me and Hannah will help draw my attention if there's a particular thing before we move on, we want to talk about. And then hopefully, you know, a goal is also just to be able to answer a lot of specific questions or to have a great dialogue. So, most importantly, mainly my goal is that you don't fall asleep and that you know a little bit more than you did when you started.

[2:13](https://youtu.be/LjxUjdsXgHQ?t=133) So, with all of that, I'm going to go ahead and share my screen. Hopefully, I do this well. It is always a great test to figure out if the ex-Googler can run technology well. So, how is that, Hannah? Is it there? We can see it. Perfect. Okay. Well, we're going to jump right in. And my ask is this, you know, as someone who is just doing this because I really care about y'all as public servants, please help me by being engaged in the chat. I will often stop and not move forward on a slide until I get some type of answer in the chat in order to make this dynamic. So, please do participate and help me out with that. With all of that, I would love to ask you all, what do you notice about this picture? So, there are no right answers, you know, to what it is that I ask you all, but what do you notice about this picture? Go ahead and put in the chat. What do you notice? And this isn't a trick question.

[3:13](https://youtu.be/LjxUjdsXgHQ?t=193) No cell phones, everyone's busy, crowded, chaos, rush hour, train station, absolutely, world clocks, in the UK, yep, absolutely. So, you're seeing, you know, as a human, you're taking in all of this image, you're kind of being able to see things, a lot of waiting, male-dominated, etc. Now, what do you see that's different in this picture? And I know the fashion will be one of them, for sure, the fashion is different. Yep, everyone's looking at cell phones. Everybody's on the cell phones. Exactly, Ashwin, yep, on the cell phones, right? And I highlight this because a lot of times as public servants, you know, we question, right, whether we question change, right, or whether people can adapt, or is there too much change happening? And the answer is probably yes, there probably is too much change happening, but I highlight this because it might surprise you to know that the difference between the first picture and the second picture was only five years, right? So, if you think about it, in five years, what I think is interesting is to highlight how our relationship with knowledge changed, right? And they're going to highlight a couple of different ways, but in only five years, right? So, let's also think about what are some other major elements right now. So, if we look at this, one of the things, and I want to ask you all, how many of you feel like change is happening faster? And if so, how much faster? Is it twice as fast as it used to happen? Is it three times? Put in the chat kind of the relationship with change.

[4:54](https://youtu.be/LjxUjdsXgHQ?t=294) In my last job, one of the things I did was actually study change in government, and one of the things that was really interesting about COVID is that it changed the change paradigm in government shifted in the last four years. What I mean by that, yeah, too fast, pretty fast, exponential, et cetera, right? I say this because I want us to have a moment of empathy for ourselves as public servants, as public contributors. And what I mean by that is that the rate of change in government used to be an average of five years to 15 years. That used to be the average change horizon. And what I mean by that is the time that a policy might change, or a program, or a budget, or a lot of different things, a platform, a system, a technology, you would usually have about five to 15 years where it would stay somewhat the same or evolve slowly. But what has happened in the last three to four years is that change paradigm actually went down from five to 15 years to six months to 1.2 years. So if you feel like change is happening faster, more constantly, et cetera, you're right. And I want us to take just a moment of empathy for like, wow, that means that as public servants, like this job has gotten harder, right? It was hard enough to begin with, but now I've got to do all these things and respond so much faster than I did before, right?

[6:16](https://youtu.be/LjxUjdsXgHQ?t=376) And I put this slide up because if we look at it, we used to have 86 years before a major change, 99 years before industrial revolution, got down to 10. We're now in the six months to two year cycle, right? And I think that that's really important because what it does is it really forced me as a leader to think about, well, if I'm now leading, you know, a public mission in a time where the thing I can really count on is constant change, what does that mean about how I lead differently? What does that mean about the tools and technologies that are out there? Yes, Jeannie, our brains cannot keep up, right? And we're actually going to talk about that a little bit more. And I think a lot of us hoped that after COVID, it would like, it would slow down, right? It was like snap back. And the reality is that we're just not seeing that. And so this question of things like AI and other tools as well, okay, what does that mean in light of this? Yes, Frank, I love that comment. Faster than we have the ability to analyze. And this is exactly why we're going to spend the time today in the way that we did. So another thing I'm going to show you, and yes, this is very overwhelming. I don't know if you have like a gigantor screen, like I do. If you don't, don't feel bad if you can't read all the little bullets, but I want you to look at the left-hand side, because when we think about that six month to two year horizon, you know, here, some of you are in different types of roles, but like, if you look at the green, those are just the things around environment, et cetera, that had come up when I created the slide in the most recent six months. And if you look at the language on the left, what I've bolded, you know, really a lot, you know, of our job as public servants is to navigate how we incorporate discovery, right? And change into our mission, into our local, you know, state, national, global, you know, types of elements. And so, again, that is a point of empathy of not only are things changing, but they're changing because new things are being introduced, right? And this relationship with information, with knowledge is being strained. So the other factor, which we've seen a little bit

[8:19](https://youtu.be/LjxUjdsXgHQ?t=499) in the chat already, is that it's not just that things are changing quicker. It's that we're actually expected, you know, to be navigating in some cases, hundreds of thousands of times more information than we were only four to five years ago. So I want those of us who maybe have been in public sector for a while, I want you to think about, and I was reflecting for myself, if I looked at my role in public sector, even just eight years ago, the amount of information that I was expected to responsibly navigate in order to make a decision was so much less than I am now, right? And part of that is like societally, we have new things that we ask ourselves. So for example, we might be asking ourselves questions about equity that we didn't ask before. We might be asking ourselves questions about environmental impact or things that we didn't navigate before. And so we're introducing new contexts of information. But again, I want to take a moment of empathy where if we're expected to navigate change faster than ever, and we're expected to do that with more information than ever, then gosh, that seems like a challenge, right?

[9:28](https://youtu.be/LjxUjdsXgHQ?t=568) And if you take a look at this, these are just a few of my fun information kind of overload facts. But, you know, I would love you to take a look at this slide and pick the one, pick the one element that is most surprising to you. I have a favorite on this. But pick the element that's most surprising to you and put it in the chat. For those of you that haven't heard about it, the Rockefeller conundrum on the right-hand side is this idea of if you could be as rich as Rockefeller, and with all the resources and everything that he had during his time, but you had to live then, versus having what you have now and living now, which would you choose? And when you do the Rockefeller conundrum, I'll, you know, go ahead, actually, people put that in, go ahead and put your answer in as well. Would you choose to be Rockefeller or choose to live now? Yeah, Carl says now without a question. The 174 newspaper now, yeah. So usually when we do the Rockefeller experiment, like 90% of people say now, right? And which is really interesting when you think about, you know, we think about all the resources someone like that had, but we recognize that we live in a really incredible time. Yeah. And if you're looking at the left, my favorite stat on this page is actually, you know, the stat that over the last two years alone, 90% of the world's data was generated, right? So I want us to take one more reflection, and that is, you know, as a public servant, there's usually two things that I really, really feel like I'm held accountable for. The first is do not waste taxpayer dollars, right? We all know that one, that is just do not mess that one up. The second, though, is a real expectation that I'm going to make evidence-driven decisions, right? That I'm going to make good decision based on good information, good evidence.

[11:18](https://youtu.be/LjxUjdsXgHQ?t=678) And so again, if we're talking about an ever-changing state and more information, right, that feeds that decisional construct, this is a real challenge, you know, to keep up with all of that. So what does that mean for us as people? Unfortunately, for those of you that don't know Moore's law, Moore's law is relevant to the idea of computing processing power kind of doubling consistently. Our biology, Alexis's personal compute power in my brain here does not follow Moore's law, right? I can process about 120 bits per second. Now, I don't know, does anyone here have kids? And if so, put the ages of your kids, you know, in the chat. You will know that if one child is talking to you while another child is talking to you, your compute power is about maxed, right? And at that point, my, you know, my direction is mommy, you know, mommy wants to talk to one person at one time, right? And that's because attention is actually our most essential mental resource, right? And it's true for any organism. It's how we have survival, it's how we thrive. And so our attention system in essence protects us from being overwhelmed. So let's think about for a minute the conundrum of what it is to try to avoid being overwhelmed and pay attention to the right thing while at the same time being confronted with incredible amounts of information that we're supposed to be paying attention to to make decisions, right? And so that's really, really challenging. And so the way that I kind of started to form a relationship with AI, and I often think about AI as being augmented intelligence, meaning how do I augment my functional knowledge in order to make a decision versus artificial intelligence? But that's just an Alexis preference. So the way I look at it is AI is really our ally in information and attention overload. It can kind of serve, it serves as like my buddy, my like super processor so that I can make more evidence-driven decisions, you know, while still maintaining my discretion, my deep

[13:20](https://youtu.be/LjxUjdsXgHQ?t=800) years of experience, you know, all of that expertise that I have to exercise. Oh my goodness, so many great kids. Oh, Robin, 12, 8, and 8. You never have many people talking to you at once, right? They all just wait their turn, I'm sure. All right. The other thing that was really fundamentally game-changing for me is that if you think about it, about 20%, only about 20% of any of our organization's information is structured, meaning like in our ERP system, etc., we have actually said, here is the field, and here's what the field is, and we've cleaned it, and we've structured it, and we normally go through that process with information or data that we're going to use for reports or dashboards, right, or other types of things where we need a dependent, you know, reliable flow of particular elements of data or insight. But if you think about that, only 20% of our data, right? So where's the other 80%? The other 80% of our information, of our knowledge base, right, of all of the information we're supposed to be pulling on for those incredible decisions, it's in unstructured data, right? And so that could be in an email, it could be in a chat, it could be, I know in my case, coming from the public sector element that I was in, you know, a lot of PDFs, like think about project reports, you know, monitoring and evaluation studies, you know, other types of things, research, you know, all of those things are often locked in PDF or unstructured, you know, data. And so we're actually, it's really hard to bring them in. Another thing that was really interesting, and I'm fact-checking this, but I've had two people now tell it to me, so I'm sharing it with you, is that 87%… 7% of government data is never seen or used by more than three people.

[15:06](https://youtu.be/LjxUjdsXgHQ?t=906) Think about that for a minute, about how much information that we generate, right, in our work, and really valuable information that may never be unlocked. And it's not seen because people don't want to see it. It's not seen because it's hard to find, right? It's in that kind of unstructured, kind of locked away, you know, kind of hidden element. And so the question for us is, well, what does AI and generative AI do to potentially change that? So the final thing that I really had to navigate is really, and Hannah, I'm going to say this the wrong way, so please correct me if I do, but there's a term, I believe it's anamorphization. So anyone that knows that, can pronounce that or spell that better than me, please put that in. But in essence, it's this principle that we do as humans, where we assign human characteristics to non-human things. So for example, you know, I am convinced that my chihuahua is intending, right, to give me a guilt trip about something. And the reality is my chihuahua probably doesn't think and does not try to work in guilt manipulation, but I assign her that, right? Similarly with AI, you will hear people say things like, well, the AI wants to take my job, right? Or the AI is biased. But I think it's really important for us to understand that, you know, the AI does not have intent in that way.

[16:32](https://youtu.be/LjxUjdsXgHQ?t=992) The AI does what we learn it, what we train it, and it does it with the things we give it to do it. And so it's very important that we actually own that relationship with knowledge, right? And we just don't give away that power to the technology that you'll often see happen in conversation. Thank you, Carl. Anthropomorphic, okay. Now you see why it's really hard to say, right? A little bit of empathy for me on this one. Thank you, Carl, for grabbing that though. Thanks, Jennifer, as well. Exactly. They don't like it, Renita. They don't like it when we do that. All right. Now, this is really interesting. So one of the things you'll see is the other way that we often experience technology, and I'll just speak for myself. I'm sure no one else is guilty of doing this, but we have what we call the headline rule, meaning that the nature in which you see headlines about that technology determines your general attitude toward it, right? Even if you've never used it, never touched it. And so what I did is I took a couple headlines here, and I just, in essence, does anyone see what I did with them with colors? There was one I messed up, but the rest are pretty accurate. What do the colors infer?

[17:46](https://youtu.be/LjxUjdsXgHQ?t=1066) What do you see when things that are red are generally what? Dangerous. Yeah. Red, pink, they're generally bad, right, or negative. Things that are green, except for when I messed up, are generally positive, right? And things that are yellow are neutral. And I highlight this because it is not unusual, right? If I see kind of three negative headlines about AI, I might then have a relationship with a technology that is negative. And I think it's really important for us as public servants who are expected to navigate these technologies to really kind of put more attention than just the headline. And so one of the other things I really like to highlight for people, does anyone know why I put a picture of an old car? What does an old car have to do to remind us with, you know, what it is that we kind of go through in adopting technologies? Anyone have any guesses? How far we've come, was that a good one? Yeah. How things evolve, fear of the unknown. Yep. Jennifer evolved. Absolutely. So the reason I put it here was because as I was studying adoption of technology, what is interesting is that there was a 20 year, okay, in the media, if you go back to the automobile, there's 20 years of headline debate about whether the car is a good thing, right? So was the car a tool of the rich that is deadly and dangerous and killing people versus is a car something that's, and I say that because, you know, again, this is a relationship that we have as individuals in society with any type of technology, right? And so, you know, what is actually interesting is that it took 20 years, right? For that to play out because of the nature of information sharing back then. Now we have to navigate these tools very quickly, right? And as a society, we're constantly pressure testing them. And as public servants, we've really got to peek under the hood

[19:48](https://youtu.be/LjxUjdsXgHQ?t=1188) and say, well, what is good or what is bad? And what I will tell you is that it really, really comes down to the use case, right? And we'll talk a little bit about intentionally around that. And, you know, I think exactly, I think it's Hagan makes a great point that, you know, using headlines will often produce biased opinions as, you know, sometimes the media ends to get more clicks if they're reporting something bad than good. And I don't want to overgeneralize. I don't always, I don't think that's always true, but, you know, I think this idea that like, again, as people, we also are people, right? And our response to these types of changes are really normal and have happened throughout history. But this is my way of encouraging you, like look beyond the headline, right? And hopefully by the end of today, you'll feel a little bit comfortable playing with these tools and do those types of things. All right. Now, what is this? Who knows this tool? One of the most game changing tools of bureaucracy. Come on, people. This control H, which does, right, bind and replace, right? I mean, this was, I have a good bureaucrat, the ability for me to take Frank's good policy document, swap out, you know, swap out his agency and put my agency in that, like Frank, you advanced me perfectly, right? I was able to take all of Frank's work and his knowledge and just, you know, and I don't need to recreate the real, right? I could find and replace. And this is actually, I want to highlight this because this is a really interesting step in our relationship with knowledge, right? Our willingness to kind of share, curate, et cetera. And it's still a tool that we use today. And part of why I'm going to walk you through some of these is because we can often have this perception that a technology comes out of nowhere, right? But actually what happens is technology functionality stacks, right? So we'll start to do something. It'll turn out to be really useful. And then we'll start to want to do something else, right? Slightly more, slightly more. And so where we are with AI is really a lot of stacking around these. All right. So who knows what this one is? Another

[21:53](https://youtu.be/LjxUjdsXgHQ?t=1313) really popular tool, you know, still in common use. Anyone know what this is called? Word cloud. Absolutely. Word cloud. Now, again, I want you to think about the fact of how powerful this was, because not only in one kind of image in this case, not only are you seeing the things that people are thinking about, but you're actually seeing sentiment at the same time, right? You're seeing waiting. You're seeing the thing that someone cares the most about. And that in itself is a contextual relationship with information, right? So it'd be one thing if you had all these words, right, in Times New Roman font in 11 as a big list, right? We had that, but we needed to know more. We needed to know, well, which one is really the most important? Which one do people have the most feelings about? Are those feelings good? Are they bad? And so we came up with a new way of organizing knowledge to be able to kind of attribute that. And I show you this one because it's a really critical step in that idea of our desire to contextualize information, right? And to kind of have a more compound relationship with it. Okay. Now, what is this? I'm not even asking you what it is because most of you have probably, you know, done a Google search, but I will ask you this. What did this change? What was game changing about this coming out? What did it change? Place that said Helvetica. I love that. And come on, someone come off, come off, you know, use your real life. Yeah. You try to find the information online now rather than reading the book. We people used to find, go library and those ways. Now everybody has a tool on their fingertip to find the information. Absolutely. So, I mean, what I'm hearing is absolutely true. It democratized, right, access to information. It unearthed the access to knowledge that we had never had before, right, at that volume. But it also did a few other really critical things.

[23:53](https://youtu.be/LjxUjdsXgHQ?t=1433) The number one trait that I love most about people that I'm fascinated by that I try to exercise is curiosity, right? And that idea to be able to, to be curious, to ask things, right? And the way that search, you know, really advanced that. I want to remind us too, that Google was not actually the first search engine. Google was the 11th search engine to be introduced to the market. What Google did, in my opinion, that I think was particularly strong is that they understood the relevance of context in our relationship, right? So Google kind of guessed, we think this is what you want, right? And so algorithms and things like that were created to kind of serve us up, not just what we asked for, but what it interpreted we were actually trying to achieve with our curiosity, right? And there's an enormous amount of power in connecting someone's curiosity, you know, with, with information that they need. So I know Chris and I also remember like before Google, so, so don't worry. But I also think that this is really critical, right?

[24:53](https://youtu.be/LjxUjdsXgHQ?t=1493) So when we think about where initial kind of single line search now changes with generative AI, is that if you think about it with Google, I might say, what are the top 10 restaurants in New York, right? That's kind of that single line of query. What I can now do with BARD or ChatGBT or generative AI is I can say, I can have a compound relationship with the information. So I can say, not only what are the top, I might say, what are the top 10 restaurants in New York by local preference vote? I might say that have, you know, their, their main dishes between the cost range of $11 and $16 that are within half a mile of me that, you know, are not open or that are open on Mondays, right? So again, in one query, I'm able to have a much more, a much more compounded relationship with the knowledge, right? And so this is one of the big things that has shifted with, and just, again, just kind of stacked with the capacity we have with generative AI now. All right. So what are these things? And I promise I'm not going to do too much technical wonky stuff, but I feel like you're not allowed to do an AI 101 without what are these things? So the way that I think about AI is in essence, AI is where you watch the system. So what you're saying, can I, sure.

[26:15](https://youtu.be/LjxUjdsXgHQ?t=1575) Yeah. Is it a machine language? So artificial, so machine learning is a, yeah, machine learning is a subset of the larger practice of AI. And I'll go into more detail, but Adarsh, did you have another question? Yes. So you can ask a detailed question like you did in ChatGPT in Google? Yeah. And in ChatGPT and in BARD, in Google, you can try to do that, but it isn't, Google, the original Google search is not a large language model, right? It is a search algorithm. And that's why in the search algorithms, you tend to look for more singular types of queries, right? Like what are the top 10 restaurants where with large language models, you're in essence asking the system to apply a number of different factors or framing. And that's what makes it more powerful, right? And kind of the next generation of this. And this is why you're seeing a lot of organizations like Microsoft or Bing or Google start to integrate generative AI into their search capabilities. So search will continue to change pretty significantly in the next year because of this. Okay. Thank you.

[27:30](https://youtu.be/LjxUjdsXgHQ?t=1650) Sure. So in AI, what I'm really looking for is those things that I would have to train a system to do like me. And I'm going to give you some examples of that as we go, but where you want the system to actually be applying some level of contextualized understanding, human intelligence, interpretation, where machine learning is a subsector, a smaller group within that, where really what we're asking is for the machine to be able to not simply use rules, but to actually be able to, you know, to be able to interpret context. And I'll give you some examples that are helpful. So one thing I do like to anchor us in is the fact that machine learning AI is all around us already. And so can anyone kind of look at some of these pictures and put in the chat, what are some things you're seeing here that you realize now is AI? Weather forecasting. Yep. Absolutely. Siri. Absolutely. Broad detection. Yep. Weather forecasts are using regressions, spam folders, ride share, Google. Exactly. So the reality is AI is already all around us. You know, it's built into most of the things that we experience. And so whether it's machine learning, whether it's human intelligence, whether it's machine learning, AI is all around us. You know, it's built into most of the things that we experience. And so weather, as many of you probably know, the COVID vaccine wouldn't have been possible to create without AI. And also curbside, which I really, really appreciated, you know, during COVID, wouldn't have been possible without AI. And so those digital twins, all of those things, these are all types of AI that are around us already. So when we think about AI or machine learning, so you'll see AI ML, that stands for artificial intelligence and machine learning.

[29:21](https://youtu.be/LjxUjdsXgHQ?t=1761) There's really four main categories. And I want to give some thanks to some of the tools we'll talk about today who kind of gave me some of these really great slides that are wonderful. So the first is classification. And I'll give you some examples of that. The second is prediction. And this is something that we're going to see more and more used in government. This idea of scenario running, right, or prediction or things like that is really something that we're being asked to do more and more as public servants. And we'll look at some examples. Generation is both kind of what you might see in chat, GBT, but it could also be generating music, generating images. And then finally, understanding. So When we think about classification, you know, part of what we're really asking a system to do is to identify and recognize things. And I'm going to actually show you a tool that we're going to play with live. But in this case, on the top right side, there's actually an AI tool that was created in the Amazon to be able to tell and to determine when it heard a chainsaw, right? To be able to actually direct rangers to that area to, you know, try to stop illegal deforestation in the Amazon, right? So in that case, the AI was programmed to understand the difference between the sound of a chainsaw, a bird, and other types of things. Or you might see classifying images.

[30:38](https://youtu.be/LjxUjdsXgHQ?t=1838) Does anyone here have Google Pictures, where if you, for example, for me, I have two chihuahuas, and Google Pictures will actually organize an album for one chihuahua that is different than the other chihuahua, right? And that's image classification, where it does that automatically, because it knows, as a good chihuahua mom, I've got to, you know, I've got to pay equal homage to my two girls. Yeah, iPhone does that exactly. So that's an example of classification. So operations, and in this case, AI, here are a couple of tangible examples. And I think I saw in the chat a few of these as well, that some of you are already playing with these in your spaces. But, you know, we've seen a lot around, like, asset inspection. And so, for example, on the first case, a lot of things like windmills, right, who might be a couple hundred feet in the air, or wellhead inspections, things that are difficult to access. In the case of windmills, we're actually seeing drones go around. And I was fascinated to learn that a drone, you know, can take a video or a picture that can determine if a screw is loose on a windmill by half a turn, right? So when you think about maintenance, when you think about having to send someone up, you know, a 300 foot windmill, and being able to game change that, that's pretty cool. A lot of inspection, like photo inspection, one of the tools that I got to work on was actually a AI tool that would help detect rust on a ship, right, whether or not, you know, a ship had more rust than others. And that becomes really important for predictive maintenance, right? Because you may have a schedule of ships coming in to be maintained, but one just might not have actually had as much damage. And so you're going to take it out of service, when it really doesn't need as much as something else, right? So there's a lot of these just practical ways, fleet management. And then on the right, I love this as a Memphis example, where they put cameras on city buses, to actually take images of potholes and blight and be able to

[32:39](https://youtu.be/LjxUjdsXgHQ?t=1959) direct crews to fix those based on severity, etc. So a lot of really cool kind of classification work in government already. Another one, as I mentioned, is prediction. This is absolutely huge, it's exploding. So in this case, on the right, this is around flood prediction, right? So we can look at weather patterns, topography, historical, you know, types of elements and understand the likelihood of a flood and actually give people alerts right around the risks. This is one that for those that are into, I know we have some Department of Transportation folks, you know, some folks that are looking at issues like equity. And I think what was really interesting in this particular case was being able to take multiple geospatial sets, look at usage and really understand, you know, what the different impacts or access tools that people had to transport, right? And transportation challenges, whether that be pedestrian incidents, car dependence versus public transport. There's another really cool shout out to anyone from Hawaii, because Hawaii is doing really great work where they're using AI to map not only the infrastructure, but then the equity elements, but more, and even on top of that, climate. So in other words, they can say, well, if Hawaii gets six inches of rain, what infrastructure, what neighborhood, what group is actually most impacted, right? And understand how they might change or respond to that. So some really cool work out there. The other type of prediction for those of you that are in audit, accounting, et cetera, one of the most powerful cases that's really coming out around AI is improper payments or fraud. Many of you probably know or have had to navigate, especially over the last three years with the amount of resources that went out through COVID response. Unfortunately, you know, doesn't always go out in exactly the right way that we intend in our attempt to really, you know, bring people relief. And so we're seeing a lot of this also similarly on the

[34:40](https://youtu.be/LjxUjdsXgHQ?t=2080) finance side around security or other types of alerts as well. Generation. This is pretty cool. This gets into kind of chat GBT bar generative AI, but you know, this is about being able to interpret what something is hearing, what it's seen, and being able to generate new things. One of the examples that is one of my favorites, and again, shout out to the U.S. Patent Office, is in their instance, they really took and looked at their most valuable kind of talent, which was their patent examiner. And so, you know, those patent examiners are, if you think about it, deciding really powerful things, right? They are literally kind of giving, you know, giving people's dreams like a stamp, right? Someone's life is completely changed often if they get a patent or don't. And so that's a really important job. And, you know, one of the things that because of that, they have to spend a lot of time searching to create a case to understand whether they should issue a patent or not. And so in this case, the AI, again, kind of like an ally, you know, was matched up with the patent examiner and the AI in essence learned how the patent examiner searched. What types of information was interest, you know, was appropriate or relevant, what types of information was not. And so it became better and better and better at understanding the search that a patent officer needed to do. And so in this case, that search improvement was able to actually make it so the patent officer spent less time doing that perfunctory kind of searching and more time reflecting and being able to issue and make a determination, right? That really important human judgment around whether or not someone should be issued a patent. And if you think about it, one of the things I'm excited about this, and again, shout out to Jamie Holcomb and team over there, that same organization of information that a patent examiner might use may also be able someday to be pushed down to the individual citizen to understand if there are things actually really new or not, right? And make that more accessible and kind of change

[36:40](https://youtu.be/LjxUjdsXgHQ?t=2200) the projection of someone's, you know, kind of invention and dream. And so that's really exciting. Is the screen still frozen?

**Audience**

[36:50](https://youtu.be/LjxUjdsXgHQ?t=2210) Not for me.

**Alexis Bonnell**

[36:52](https://youtu.be/LjxUjdsXgHQ?t=2212) Okay. We'll keep going. Kayla, hopefully yours will become unfrozen. And the final one is language understanding, right? And so that is when you talk to, and I'm not going to say the name because I know if I say the name, it will activate many things in your house or offices. But if you ask, for example, you know, your digital assistant to help with something, that is language understanding. And one of the other types of language understanding that's really, really powerful is, you know, the idea of like chatbot or transactional assistance. And in this particular case, you know, really a state did an amazing job of putting together a virtual agent. They were overwhelmed as I'm sure many of you were and may still be with, you know, additional interactions and so they could not add all of the call center agents that they needed to add. In this case, they would have had to hire 4,000 additional human agents for one year to deal with the volume and they couldn't. But what I really want to draw your attention to is actually the bottom, right?

[37:55](https://youtu.be/LjxUjdsXgHQ?t=2275) Because many of us, we've all had bad experiences with chatbots for sure at our chat in ways that we haven't found helpful. But these are actually like quotes from people to the chatbot, right? In the sense that it was so good they really thought they were talking to a person, right? And so, you know, thanks so much for your help. Thanks for being there. You know, I just saw that. And I think that this idea of really being able to help someone to create delight and for someone to get what they need from us, you know, in government is on an exciting trajectory. Now, I promised to give you the Alexis definition in easy to understand, in my case, motherhood terms of the difference between maybe a normal reporting or rules-based system or the artificial intelligence. Oops, sorry. So if we think about it, and the way I want you to think about this is the context of the stages of information and our interaction with that. So if you look at the swaddled baby, okay, hold on. I'm looking at the chat.

[38:59](https://youtu.be/LjxUjdsXgHQ?t=2339) Oh, they're just wondering if I'm an AI robot. She's the best. She's our most advanced model, y'all. Hannah is like, it's Hannah 6.0. Like you're getting the good stuff with her. Oh, I love public humor. Bless you all. Okay. So when we look at normal report or rules-based system, I think about that like a swaddled baby, right? So the reason I say that is because when we have typically had a relationship with like the types of systems that give us reports, what we've done is we've told that data, we've told that baby, this is what you're allowed to do. This is what we want from you. And we've kind of constrained, right? And kind of limited and directed what we want from it, right? We want particular fields from a dashboard. We want other types of things. When you get into process automation, you're talking about a little bit less constraint, right? You're asking the system to kind of do a process automatically or move information or data or ingest it or do it differently. So you've given it a little bit more flexibility, still in the crib, right? Still constrained. When we get into machine learning, you're really talking about modeling, but within constraints. And what I mean by that is that might be auto-translation. So if you were in a report-based system, you might say, well, this word in English is this word in Spanish, and it can never be anything else. In machine learning, what you might say is this is the word in English. This is usually, you know, the word in Spanish, but if it's slang or an idiom or a particular, you know, dialect of Spanish, it may be slightly different, right? And we give the machine like kind of abilities or boundaries to work within that. And that kind of gives it a little bit more flexibility. The final one though is AI.

[40:47](https://youtu.be/LjxUjdsXgHQ?t=2447) And I really need you all, especially those of you who have maintained the survival of tiny, tiny humans to bigger humans, to think about this for a minute. What would happen if I, Alexis, took my three-year-old to the playground, much like you see in the picture, and just left the three-year-old there? And I went, and I went somewhere else and I did something. What would happen? Come on, don't hold back. Mass chaos, absolutely. What else? Problems, arrested, may not survive. Like the point is, is like not good stuff, right? Not good stuff would happen. And so when we think about artificial intelligence, what I really want you to understand is artificial intelligence means active engagement, active supervision. So in this case, my three-year-old or six-year-old or eight-year-old or husband, for that matter, may be fine to play on the jungle gym, but I'm there. I'm keeping them safe. I'm keeping them in boundaries. I'm directing, right? Don't go out of the playground onto the street, right? Or don't go on that slide. You're not ready for that yet, right? So this idea of really, really being engaged is important because I think a lot of people tend to have this idea that AI is set it and forget it, or we'll just kind of take things over. And actually, AI needs a lot of direction, a lot of care and feeding, a lot of attention.

[42:09](https://youtu.be/LjxUjdsXgHQ?t=2529) It's incredibly powerful, but it is not constrained in the same way, right, of our normal reporting swaddled, you know, kind of baby relationship with information. And therefore, we have to have a dynamic engagement with it. Is this helpful to kind of help break it down that way? I've been trying to find the ultimate metaphor and I'm still working on it, but hopefully-

**Audience**

[42:30](https://youtu.be/LjxUjdsXgHQ?t=2550) Alexa, I'm just going to ask a question that was in the chat that, I mean, part of it has already been explained, but if you could provide a little bit more of a distinction between AL,

**Alexis Bonnell**

[42:40](https://youtu.be/LjxUjdsXgHQ?t=2560) ML, deeper learning and generative AI. Sure. So, okay. So ML, machine learning, is a subset of larger AI. So it's more constrained and more focused. Generative AI usually involves what we'd call large language models or very large amounts of information that you're asking the AI to have a more fluid contextual relationship with. So to be able to

**Audience**

[43:10](https://youtu.be/LjxUjdsXgHQ?t=2590) ask that kind of compound question too, right? And what was the fourth factor, Hannah? Sorry.

**Alexis Bonnell**

[43:18](https://youtu.be/LjxUjdsXgHQ?t=2598) There was AI, machine learning, generative, and you're on mute. Sorry.

**Audience**

[43:25](https://youtu.be/LjxUjdsXgHQ?t=2605) Sorry. Deeper learning. Oh, so I don't, personally, this is Alexa's personally, I wouldn't necessarily

**Alexis Bonnell**

[43:32](https://youtu.be/LjxUjdsXgHQ?t=2612) parse deep learning outside of the process in which you are asking and training the AI, right? So deep learning is a lot about the information set you give it, the things that you're asking about it, but it's that actually like human machine learning interaction is the way that I usually think about it. And deep learning creates a lot of the basis of where generative AI and other tools come from. Yeah. I know it's always hard to describe, but hopefully I'll give you a couple more things. Yeah. And deep learning, it's a great point. It's for huge data sets, you know, a lot of predictions. And if you, if others have good insights or definitions, please put them in the chat. There is lots of great brainpower on this call. It does

**Audience**

[44:18](https://youtu.be/LjxUjdsXgHQ?t=2658) not need to be only me who are helping us all understand this. I'll just ask one more, if that's okay, if you could clarify where deterministic algorithms fit into this spectrum. So I want to

**Alexis Bonnell**

[44:37](https://youtu.be/LjxUjdsXgHQ?t=2677) make sure there is no shortage in technology land of a lot of different terminology. So when I think about a deterministic algorithm, I usually think about, and this could be a misinterpretation of what the person meant, but I usually think about an algorithm that is, or a, you know, an algorithm that is being asked to make a particular type of decision, right? Or. Or kind of combine information to a decision point. And actually, that's a great segue into this slide. So when we think about decisions, we really want to think about those within context, right? And so there are some decisions where we're really happy for the algorithm to make them. And I'll give you an example. And those decisions are usually on the left side of the screen, where they have a low impact on human fairness or safety. So I gave you the example of rust on a ship, right? So when you're thinking about determining, letting the machine determine whether or not what it's seeing is rust, that's actually a pretty safe thing, right? If you train it enough, it's likely to be able to detect the rust. But more importantly, even if it's wrong, and what they find is not rust, let's say it's an oil stain, there isn't going to be a huge impact necessarily to human lives on that, right? And so we might be more comfortable allowing the algorithm to be decisional, right, in that like low impact to human fairness or safety or the rust example. On the far right side, we might be thinking about something that has, that is a use case that has a lot to do with human fairness. So for example, it might be determining whether someone does get unemployment benefits, right? That actually really impacts someone's life, you know, whether they can pay their bills, it's really important to get right. And so in that case, we would want to do what you see in the middle, which is make a decision about where we would want a human in the loop. Now what that really means is when we want the person to be making the final call, right? So much like the patent examiner, you don't want necessarily that the algorithm to decide to be the thing that

[46:43](https://youtu.be/LjxUjdsXgHQ?t=2803) decides whether someone gets a patent. It's okay for the algorithm to help organize information to give that patent examiner a better relationship and a faster relationship with the total knowledge that they should consider. But you still want the patent examiner to kind of make the call, right? And so when I think about, you know, algorithmic determination, what I really think about is what types of use cases and where in that process do I want a human being to be the one to make the call? And in AI world, we call that human in the loop.

**Audience**

[47:14](https://youtu.be/LjxUjdsXgHQ?t=2834) Is that helpful, Hannah? We can always come back if we want to talk a little bit. Yes, thank you.

**Alexis Bonnell**

[47:20](https://youtu.be/LjxUjdsXgHQ?t=2840) Okay, cool.

**Audience**

[47:20](https://youtu.be/LjxUjdsXgHQ?t=2840) Alexis?

**Alexis Bonnell**

[47:21](https://youtu.be/LjxUjdsXgHQ?t=2841) Yes.

**Audience**

[47:23](https://youtu.be/LjxUjdsXgHQ?t=2843) So in the case of chatGBT, when you are asking a question on the system, where does the human involvement?

**Alexis Bonnell**

[47:33](https://youtu.be/LjxUjdsXgHQ?t=2853) So chatGBT is a great example of, and again, keep in mind that when you use a large language model or when you're asking a question, you are actually the human in the loop, right? Meaning you are the one formulating the question. And if you think about it, you would actually formulate the question very different than Hannah, right? Or Joan or Mitchell, because you have a set of your own experience, of your own expertise, of your own time on the planet, right? Which would guide how you ask that. So when you think about generative AI, the results you get are in essence based on how you ask the question. And so in that case, I would define that as you are actually the questioner or the person who creates the query is actually the human in the loop in generative AI. That makes sense?

**Audience**

[48:22](https://youtu.be/LjxUjdsXgHQ?t=2902) Yes, I understand.

**Alexis Bonnell**

[48:24](https://youtu.be/LjxUjdsXgHQ?t=2904) Great. All right. So hopefully this drives the point home. If you think about the Chihuahua on the left, and I do not mean to insinuate that technology does not have weight and heaviness and cannot go wrong. I'm not making it intentionally cute. Other than to say to the point, if you imagine that the Chihuahua, let's call the Chihuahua Milo, let's call the cow Claude. I want you all to tell me what would be the end result if I took Milo, the Chihuahua, and turned Milo out to pasture with Clyde, the cow, for three months? Because I may only have to check on Clyde every month or so, make sure they've got the water, they've got the food. But Clyde, the cow, is pretty self-sufficient generally. But what would happen to Milo if I said, well, Milo is the same as Claude, and so I can just put Milo out with Claude? What would happen? They would become friends, hopefully. Hopefully, in a Disney movie, Claude would look out for Milo, and they would live happily ever after and become the best of friends. But yeah, Milo would probably be swallowed. Rest in peace, Milo. Barking moose, you never know. But the reason I say this is because AI, if Milo is AI, if the little Chihuahua is AI, it's actually really needy. It's a learning technology, and that means that I have to nourish it, to train it, to give it the construct for which it can be most effective. I can't just set it loose, say, oh, I've written an algorithm, and now I just let it go into the world, and I never have to check on it, and I never have to train it. There's really quite a difference. The cow-based technology, and forgive me for the oversimplification, but I always kind of appreciate it. The cow is not the same as the Chihuahua, the cow is more of the information that you've constrained. So that report you've written, where you've cleaned all the data perfectly, and you know every field, and everything is aligned,

[50:24](https://youtu.be/LjxUjdsXgHQ?t=3024) you can kind of set that and forget it, because you know all of the elements that are going into it, and therefore, what should be coming out of it. This is our classic government system that we've had for a long time. When you introduce AI, you have to really want to learn as the human being, and to teach the AI, and to teach it well, and to develop it well, and to nourish it well, and you really can't just let it go. So that's particularly important when you're thinking about things like prediction, or organizing information, or large language models, or things like that. AI becomes a little bit more set it and forget it, when you think about some of the tools that have been quite well pressure tested. So for example, translation. Translation is a great example of a really powerful tool to bring equity to government, to make things more accessible, that has been really well pressure tested, and is probably pretty darn deployable in most cases, right out of the box. Similarly, document ingestion, document AI, is another tool that has tended to be very well designed. But when you get more into the scenario, the predictive, the interpretive, that's when you really want to think about the care and feeding of Milo, right, versus just setting it out in the wild. And this is my favorite way of thinking about generative AI, or chat GPT, or BARD. And the way I think about it is this. I think, and I have a relationship with it, and I use it, and I treat it like it's a really good intern, right? Meaning like a really, hopefully we've all had like at least one really good intern, where you can give them an assignment, right? And they come back, and they may have gotten 60, maybe even 80% of the way there, right? And like, that's awesome. Like 80% of the work is done. But let me ask you this. How many of you at least would ever admit to giving an intern an assignment, having them give it back

[52:28](https://youtu.be/LjxUjdsXgHQ?t=3148) to you, you never read it, and then you post it publicly as policy for your organization? Probably zero, right? Because if you simply posted and created a policy based on something you asked an intern to do, that would probably, you know, maybe not reflect the type of review and care and things like that that you would be expected to do. And maybe Tracy, we can send you an intern if you haven't had one yet. But I think it's really important to think about that this way, right? And the reason that I think it's also really, really important is because what I'd like to ask you to reflect on is what do these large language models, what does putting all of this amazing information, you know, at our fingertips do about our relationship with knowledge, right? And I'm going to let you know and share with you what it changed for me. So for me, up until really the point of good large language models or chat GPT or BARD, I took a lot of pride and personal identity in what I created, right? And so that idea of whether I wrote a press release or a policy or designed a new initiative, that idea of creation was where I found my value, that is where I exercised my expertise and my deep sectoral, you know, 30 years of experience and things like that. What I think I've come to realize with what generative AI does is that, again, if I'm making decisions, having to make decisions and change faster than ever, I'm having to use more information than ever, then part of what generative AI does, it allows my role to expand not just from creator, but to curator, right? As I mentioned, Frank may have done a really great policy, you know, over in, I don't know where you are, Frank, we're going to say Nebraska for the case of, you know, a point, I'm sure that might not be the case, but Frank might

[54:30](https://youtu.be/LjxUjdsXgHQ?t=3270) have done a really great policy, right? And it actually makes sense that when I think about a similar policy that I see what Frank and Joan and, you know, and Hannah and other people have done, and that I let that inform me, right? And so what I started to realize is that when I looked at generative AI, again, it was an opportunity for me to exercise my expertise and my deep experience and just my years on the planet, right? And all of the hard skills and soft skills that gave me, but I exercise it as a curator, meaning, you know, my query, the way that I ask information of large language models is going to be unique from what Michelle does, right? Or Jennifer does. And a great experiment that I really encourage you all to try is get a team together of about, you know, six folks on your team and give them a test to run with generative AI. So for example, you might say, you know, create a policy around, you know, or make the case for a policy around X, maybe it's to charge for plastic bags or anything like that, right? Pick something within a theme that resonates for you and give that team, give that team that assignment, but don't tell anyone how to do it, right? What you'll find is that everyone will actually shape their query and their relationship with that knowledge set in a different way. And because of that, everyone will actually come back with a slightly different answer, a slightly different outcome. And that's actually a good thing, right? Because as experts, as long-term public servants, that expertise that we have, that we wield is really valuable. And what we're able to do is instead of just exercising it as creators, always having to birth something from scratch, we can also exercise our ability to curate and to be able to bring in, you know, the best and also maybe not the best, right, insights out there, and then use our judgment and determination to figure out like that intern,

[56:32](https://youtu.be/LjxUjdsXgHQ?t=3392) what are we going to then, when we get those results, what are we going to refine? What are we going to use differently? You know, what are we going to edit or change to make it appropriate for our use case? Yeah, see, and Nathan, you're training your interns on CHAP-GBT. Perfect, right? Now you're getting the intern and the large language model in one. Hopefully you'll get like an 80% return on average. But this idea, again, for us as individuals, a lot of, I was worried of like, is CHAP-GBT going to take my job, right? Am I going to, and I started to realize, no, actually, like the way that I guide this tool is driven by my experience. What I choose to use or not use around what the tool gave me back, that's the same discretion I use, right, when I run a meeting to decide if someone's insight or experience is relevant, right, or should be adding to the policy or things like that that I'm supporting. So I just say all this because I think it's important that we experience these things and that we also continue to pay like honor and respect, you know, to the fact that like our discretion and our experience set as humans is equally exercised with these tools as it is before these tools. So let's talk about a couple tools. We promised to show you some cool things. So Hannah, I'm going to try to navigate like a tech-savvy person and you can tell me if I get this right. So I'm going to escape from this. I'm going to do

**Audience**

[57:58](https://youtu.be/LjxUjdsXgHQ?t=3478) a new share. I am going to go here. Alexis. Yes. I'm wondering if you would mind starting us off with CHAP-GBT because. Sure. We've heard a lot of people in the chat asking for examples of using

**Alexis Bonnell**

[58:14](https://youtu.be/LjxUjdsXgHQ?t=3494) CHAP-GBT. So would you mind starting us with that one thing? No problem. Okay. So I preloaded something in CHAP-GBT for you all. So if you go to CHAP-GBT and please go ahead and put in the chat if you've used CHAP-GBT yet or not. And if you haven't used it yet, what is it just because you haven't had the time? Are you concerned about something? But what I did here in CHAP-GBT, as you can see at the very top, what I in essence did is I took what I'm highlighting. I in essence said this was my command. Put the following content into more simple, straightforward, and easy to understand language. And then what I did is I copy and pasted from the SBA 7A loan. I copy and pasted their entire website page, right? Which when I read it, I was kind of like, I am a little confused by this. I don't understand. And this is something I've been doing with a lot of my government counterparts and customers is kind of saying, let's look at how you already talk about something and let's see if we can make it easier.

[59:18](https://youtu.be/LjxUjdsXgHQ?t=3558) So I literally uploaded this. This was all the content. Then in essence, this is what CHAP-GBT gave me back. So it took all of that, but it didn't just take what I uploaded from the website. It actually took other information sources of conversations or other types of guides or things like that that might exist out there around the 7A loan. And in essence, contextualized it, right? So for example, it actually started out, which was very different than the first, what I put into it, that the 7A loan program is the most popular loan from the SBA. That actually, I don't think was in the original website. That was actually gathered by looking at external information. And and being applied. But if you look at it, look at how short and simple these bullets are, right? Very human, very regular language, very, you know, human. Where if you go up here, look at this. Exact same information flow, right? But much more, you know, much more kind of broad.

[1:00:20](https://youtu.be/LjxUjdsXgHQ?t=3620) Now, in some cases, this may have detail that's really important and you don't want to lose it, right? In other cases, it might be something else. But let's do another one.

**Audience**

[1:00:30](https://youtu.be/LjxUjdsXgHQ?t=3630) Alexa, could you tell us what the prompt was for that one? I'm not sure if everyone could see it.

**Alexis Bonnell**

[1:00:34](https://youtu.be/LjxUjdsXgHQ?t=3634) Absolutely. The prompt at the very top, forgive me, was, and I'm highlighting it here, was just put the following content into more simple, straightforward, and easy to understand language, colon. And then I literally just copy and pasted the website content from the SBA file into this. So that was the prompt. So this is all stuff I copy and pasted in. But let me give you another one. Yes, actually, on the question, Nathan, that you had around states, municipalities, pooling internal data and reports. One of the things that, and I'm not gonna call out the particular states because I'm not sure if they're ready to be public yet, but there is, especially in some common, you know, common areas where multiple states do really the same function, where they're doing things like shared glossaries, shared training data, and other types of things to help have like a wider breadth to better train their AI models or other things like that. So that is out there.

[1:01:35](https://youtu.be/LjxUjdsXgHQ?t=3695) But let's do another one. And I'm just gonna make one up here. In fact, I'm gonna copy and paste it from my chat GBT example. So one of the examples, and if you see the video trainer on AI that GovLab did, you will actually see this one there. So I stole it from them. But here is a new prompt. Are you still seeing my chat GBT page, Hannah?

**Audience**

[1:02:06](https://youtu.be/LjxUjdsXgHQ?t=3726) Yes.

**Alexis Bonnell**

[1:02:07](https://youtu.be/LjxUjdsXgHQ?t=3727) Okay, perfect. So this new prompt, if you can see the bottom, is write a decision memo to a policymaker asking them to adopt a plastic bag tax, include evidence supporting the idea and a timeline for implementation. So I wanna call out, look at that query, right? It has multiple compounded things. Write a decision memo. So that's what it's told that it wants the output to be to a policymaker. So it's telling it the audience that will be consuming this that you're trying to convince. So we said asking them to adopt, right? So we've told the model we want to convince them to do something. What it is, a plastic bag tax, but then we're giving greater parameters, include evidence supporting the idea and a timeline, right? So now let's see what happens. So let's go ahead and look at that. So let's go ahead and look at the proposal. And here is your proposal.

[1:03:13](https://youtu.be/LjxUjdsXgHQ?t=3793) So first of all, what I love is that it's actually putting it in a proposal memo format, right? So it's auto formatting, and it's just a simple fact. And again, this may or may not have a convincing or appropriately fully factual, et cetera, argument. But if you think about it, even if there's some semi good stuff in here, that might've saved me a little bit of time, right? As I think about moving forward. The other neat thing is that it's pulling, as I mentioned before, from what a lot of other material out there that might be in this case advocating for plastic bag tax. So I don't necessarily always need to reinvent the wheel, right? I can keep taking from my friend, Frank, all the good stuff he's doing, right? And bring that over if that makes sense for my county, for my state, for my mission. So literally I feel like my hands aren't doing anything, right? This is all the model doing this now.

[1:04:15](https://youtu.be/LjxUjdsXgHQ?t=3855) Yeah, oh, Sydney, you're here, hey. Sydney makes a great point. It's really this question. I ask myself a lot about the time value paradigm, right? And I think once I allowed myself to say, you know what, I'm still valuable as a curator, right? As someone who's gonna ask and frame and discern the relationship with this information and knowledge. And I didn't just say I was only valuable as a creator. It really allowed me to use the tools in more dynamic way. All right. Is it okay for me to show Teachable Machine? The different type of generative AI?

**Audience**

[1:04:51](https://youtu.be/LjxUjdsXgHQ?t=3891) Yes.

**Alexis Bonnell**

[1:04:52](https://youtu.be/LjxUjdsXgHQ?t=3892) Awesome, all right. So I'm gonna try this again. See if I can do, I too can navigate Zoom. It's gonna be exciting. All right. Now, are you able to see Teachable Machine?

**Audience**

[1:05:09](https://youtu.be/LjxUjdsXgHQ?t=3909) Hey, Alexis, this is Sydney. Hey, Sydney.

**Alexis Bonnell**

[1:05:12](https://youtu.be/LjxUjdsXgHQ?t=3912) Hey, Mallory Smith asked a really good question that I would love to know the answer to. Is it possible in chat GPT, I assume that's what she's referencing, to actually request the sources that the GPT used to provide the output? So that is a very interesting question. I will tell you that I have had at this point, like yes and no. Meaning that in some cases when I have said specify the sources, depending on the content, it's been able to do that. In other cases, it has not been able to do that. And you can imagine Sydney, I was less likely to use the sources, right? Or to double, triple backtrack them. I think one of the things you are gonna see, so this is also a great reason as to why when I do something like this, I actually always run it on BARD and GPT both. So for example, chat GPT's language in print stops in 2021. I believe BARD is in essence, taking from current information flow.

[1:06:15](https://youtu.be/LjxUjdsXgHQ?t=3975) So for example, if I was gonna put myself in my old job and be writing about evacuation in Sudan, right? Which has recently only happened, I would get a much different result, right? From chat GPT than I would BARD. There is like interesting nuance around facts and figures. I will share that. I think this question of sourcing, we're gonna see come out in a couple of different ways. You're already gonna start seeing, and if others have more information on this, I am not the end all be all expert, but you're gonna start to see things like data cards, model cards, sourcing, and other types of those elements come and be brought to bear more in the results framework of these over time. I think this is also something when you're seeing the larger technology companies actually ask for regulation and things like that. Part of the question is, it would be nice to kind of know what some of those features and functionalities are that we think are important so that they are able to prioritize building them in. And so I do think it's interesting to see really for some of the first time, organizations ask for those types of guidance structures from government.

[1:07:24](https://youtu.be/LjxUjdsXgHQ?t=4044) And I think the citing and the referencing is gonna be just one of the areas you see a lot more in. But my advice would simply be like test it, right? And then double check it. So my final comment on that is that if I'm writing something that is really factor or research-based, I will often go back and check them. If I'm writing something that is not that, right, then I don't necessarily have that same onus. But this is also really why I tell people to experience it, right? To test it, to play with it. You'll also find that the way you shape the query, right? You know, you might say, don't give me, and I'm literally making this up on the fly because I've never tried it, but like I would go in after this and test like, only give me facts that you can cite a reference to, right? Or it might be include a bibliography and footnotes, you know, to the resources. So I think playing with those types of things will give you, I would imagine, will give you different results.

[1:08:26](https://youtu.be/LjxUjdsXgHQ?t=4106) All right, so Teachable Machine. This is generative AI, but in looking at images, audio, and other things. So I'm gonna go ahead and turn off my video. So I can do this appropriately, but I'm gonna click get started. And these are all tools that are publicly available. So I'm gonna click an image project. I'm gonna click a standard image model, and I'm going to enter my first kind of category of image. So I have turned on my webcam, and I am literally taking a picture of what I call a tumbler. Maybe that's a like old school name for it. Then we're gonna enter class two. And then what else do I have on my desk? I have a passport. So I'm gonna open my webcam, and I am going to give it lots of pictures of my passport. Right? And then I'm gonna add one more category. And forgive me for this one, because it's gonna be a little obnoxious, but I'm gonna add Alexis.

[1:09:33](https://youtu.be/LjxUjdsXgHQ?t=4173) Now, one of the things you saw me do just now is I added in data, right? Now, did anyone see me actually tell the system or tell the machine how I was different than the passport, right? I didn't say that the passport was blue or that it was square or that it was something I held in my hand. I didn't say that my tumbler was metal or had a straw, right? And so what's really important is remember we talked about previously in our relationships with data, we would wanna be very specific. The field is this. It has this, right? And we would tell the machine exactly what everything was. In this case, I put a lot of information in the machine just by taking pictures, right? You could imagine this is how we did the rust spots, right? So those patches of rust on a ship, we literally just uploaded a lot of pictures and then we upload a lot of pictures without rust. That's in essence how we train the model. So now I click on train model. Now, I know you're all gonna be incredibly impressed because have you seen me type a single line of code?

[1:10:38](https://youtu.be/LjxUjdsXgHQ?t=4238) Anyone? Nope. Nope, no code. And why is that? And Sydney knows the answer to this. Sydney knows and Hannah knows my deep secret is I cannot write a single line of code. I could not code anything, right? I cannot do that. It is not my skillset. No code. But so in essence, I'm showing you this because a lot of times we think about these things as very exotic that I need to have a ton of data scientists or AI team or specialists. We're doing this together on the fly and I encourage you to use it as well. So now what we have is the outcome. And what I'm really asking is, it's looking at this using my webcam as the data source. And what I'm asking it to do is to make a determination when it sees one of the three elements I gave to it. So in this case, if you look at the bottom, the purple line, the machine is 100% sure that it's seeing Alexis right now. Okay, well, that's pretty good guess, right? What about now?

[1:11:43](https://youtu.be/LjxUjdsXgHQ?t=4303) Now, what about this? Right? So in essence, what you're seeing is that with almost no programming on my part, well, actually no programming on my part, it was able to tell the difference. So for those really light use cases, like detecting rust on a ship, or for example, you might have a question about whether a pest in a field is a locust, right? Or something that maybe is less harmful. That idea of being able to use AI to quickly train these things, this is also the type of things that you might use for diagnostic determination and health and other types of issues. So I show this to you because this, much like CAT-GBT or BARD, you can go and play with these things now, you can use them and, you know, hopefully you can play with them a little bit in the same way. I'm gonna turn my video back on. Thanks. Did not do that.

**Audience**

[1:12:39](https://youtu.be/LjxUjdsXgHQ?t=4359) Can we see you, Alexis? This is incredible. What I'm realizing in the chat throughout the whole session is that there are just so many questions and so much interest in this. Let's get into them.

* * *

Original page: https://innovate-us.org/what-the-heck-is-chat-gpt-how-generative-ai-will-impact-government · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org