 Hello, thank you for joining us. My name is Kevin Mulhall. I am a Senior Technical Customer Success Manager here at TechSoup. This session of Ask the Experts, focusing on AI Prompting, or AI Prompt Engineering rather, we are joined by two wonderful people. Forgive me, Josh, if I pronounce your last name incorrect. Josh Peske and Kim Schneider, of Round Table Technologies. This is going to be- Keep it slow, and see if that is congelado. Okay, hold on one second. I'm going to get in here and start muting some folks. Just an FYI, as a warm-up to this, is cameras have been disabled, but microphones are on. If your microphone is on, we kindly ask that you mute it. If the opportunity to come off of mute and ask questions presents itself, we will encourage you to do so. We do have a series of CAN questions as well that were asked in advance, and we do want to address those first. So with that, I am going to hand it over to Josh and Kim. And Kevin, just real quick, I don't yet have screenshot permission, so I think I need to get promoted to organizer-presenter, I think. That you do. Just let me find it here. Give me one second. But in the meantime, what we can do is say hello to everybody, and I kind of start jumping in. So hello, everybody. Thank you so much for coming today, and the last to ask the experts we had a few weeks ago was fantastic, just incredible questions and great conversation. And that was ostensibly about ethics and privacy and transparency around AI use, but it sort of went off in a bunch of directions, which was great. Today, we're focusing on prompt engineering and prompts, essentially how best to interact with our language models in order to the results you want, right? Get the output you want. And thanks to Kim and to Microsoft, there's actually a huge update that we get to show you today that I only learned about a half an hour ago, thanks to Kim, but have been playing with, and it's incredibly cool and offers some really cool new functionality within Microsoft Copilot. So we're excited to show that as well. And it looks like I do have the share permission now, so thank you so much for that. And let's jump in. So I'll probably have to resize this, Kim. Let me know if... Are we seeing... No, it looks good. Okay, great. So I am Joshua Pesty, hello everybody. I'm at House 3 CPO, which I'm happy to explain if anyone cares. Basically, I do a bunch of C-suite technology stuff. I work with round table technology. I've been there for a very long time and Kim and I also have worked together for even longer. So I've been with Radhjil about 10 years. Kim and I have been working together at various organizations for 30 years, believe it or not, all the time in nonprofit technology. In fact, one of the very first projects we did together was create the first ever website for a nonprofit that we were working at at the time, Fountainhouse. And that was back in like 1996, I wanna say. And so that gives you a sense of our history. Anyway, let's let Kim introduce herself. Go ahead, Kim. Hi, Kim Snyder, VP of Data Strategy at Round Table Technology. And yes, Joshua and I go way back. And let's just say for the last 30 years or so, we've been, our passion is introducing nonprofit organizations to technologies in a way that's digestible, let's say, or demystifying tech, because it does tend to move very quickly. So I think we can go right ahead. We do a lot of stuff in the world of data and data privacy, which does come up in this, even though the topic is AI. But so just a quick, today's agenda, we're gonna go over a couple of slides just to kind of set the context and groundwork around what do we mean by prompt engineering? And that'll be about, let's say the first 15 minutes. And then at that point, we are going to transition into answering questions. We do have, people have asked great questions. And I think even some questions that may help answer questions that are along on other people's minds. So we've gone ahead and picked out a few of those. And we may even have a demo or two for you while we answer those questions. And then it opens up to your questions. The team at TechSoup is excellent at reviewing questions and kind of getting them through to us as we're doing this. So please by all means ask your questions as we're going along. But, and we will hopefully get to answer a lot of them. So I think with that, I think we pop back into- I'm not sure about other folks. I'm not sure I'm seeing your video or at least maybe, I don't know, Kevin, maybe you can let me know if I'm wrong and I'm just seeing the wrong screen. But I'm not sure if I'm seeing your video, Kim. Yeah, don't see the video, it is enabled though. Really, that is odd because I see my video. Okay. Do you see it now? No, let's see. All right, we'll truck along. But- Do you see it now? Or am I just- Yep, there you are. Success. That's weird. I wasn't sure if it was possible. Okay, so let's talk first about AI Prompt Engineering and the kind of focus today. So very quickly, we just wanna go through some kind of basics of prompt architecture, prompt engineering. Those are fancy terms. Simple term is talking to an AI, right? How do you talk to an AI so that you get instructions out of it? One of the things that's been really a great lesson about working with AI over the past couple of years is it's not so unlike working with people but because there is no room to blame the other person for not understanding your instructions, it forces you to really think about what is it that I am asking for in a way that I find really helpful and it's actually probably improved my ability to communicate with humans as well and understand when I don't get something I want, probably it was because I didn't communicate very well. So things to do are first, tell the AI what role you want it to play. Are you a psychologist? Are you a nonprofit consultant? Are you a physics researcher? Are you a mathematician, right? Give it a role and even a very specific role. Be very clear about what topic or task you need. If you're asking it for outputs and want a particular tone or particular style, tell it. This is the style I want. It can be very helpful to tell it the steps that will lead more likely to a successful outcome and then be very clear about what the output is and if you have a sample of that, which we'll get to later, providing one really good sample of the kind of output you want, at least currently is a very good technique for getting outputs you want. And then one caveat we'll kind of say about all this is I'm not gonna say it every time we talk about it, but I will say everything we're talking about is interacting with these large language models or AI systems as they are right now. They're changing very rapidly as we will talk about. Copilot, which we'll be focusing on today, Microsoft Copilot has some different conversation styles. They have creative, balanced and precise and these are basically, they're well-named terms, right? So creative it's gonna be more creative and a little bit more variations in its output and we'll go out on limbs a little bit more. Balance just kind of in between that precise and precise is a very conservative. It's gonna try very narrowly to do the things that you want it. Very good to experiment with different styles with Copilot based on the kind of thing you're doing. If you're trying to do a really funny and out there blog post, probably creative mode is gonna be good. If you're trying to write an AI policy statement for your organization, probably balanced or even precise might be a better way to go and those are things you can experiment with. Some really basic ideas where generative text prompts are really, really helpful and generally very low risk in terms of doing things. Are they fantastic for brainstorming, for being a thought partner? So I need to come up with 20 ideas for something. I have three ideas for something and I wanna get some feedback on them. I've written this short essay or blog post that let me get feedback on it. Fantastic at these kinds of things. I'm getting started, getting rid of that blank page for those of you who can have writer's blocks sometimes. And any kind of short question can be a starting point. Really good for conversations. I absolutely amazed my folks and my daughter the other day. She was asking a question about dimensions and was asking how time functions as a fourth dimension, what that means. And I was trying to explain it and realized I was kind of struggling. So I opened up an AI on my phone and this was one that I can interact with in an audio way. And I asked it to explain four dimensions so it could be understood by a fifth grader, which I use all the time, this prompt, like explain it to me like I'm a fifth grader as a way of forcing the AI to break down the concepts in really understandable ways. It's incredible. I really encourage you to use that, all right? So Kim, anything you wanna add to any of this that I've gone through? No, I mean, one thing I just wanna point out, one of our questions actually, someone asked about technical skills and using AI. So one of the things that Josh was been talking about is mostly about asking questions. This is largely prompt writing, even the prompt engineering is very technical sounding. It's actually a question asking exercise. And as he said, you mold it in many dimensions but it's about asking. I forget what famous scientist was, Einstein or Feynman or got attributed to one of those and said, if I have an hour to solve a really hard problem, like I might spend the first 50, 55 minutes just figuring out what are the right questions to ask because if I can figure that out and working with AI, I find very much like that. It's about figuring out the right questions. Yeah, of course, big disclaimer here. AI's will make stuff up. They'll be wrong about things. If you don't know this, it's very important you know this. It's very important to understand that you have to validate information that comes out of AI's because it might just be a confabulation that it just made up, okay? These are probabilistic statistical engines. They are just predicting next words within a context and those predictions can produce some unexpected results and I'm sure we've read plenty of stories about that. And here's a lovely sonnet composed by ProPilot. All right, Kim, off to you with image prompts. All right, so generative AI is also great at visual, at creating visual kinds of content, right? So again, like Joshua said, brainstorming, experimenting, quick rapid testing of visual ideas can be great for testing out stuff around social media campaigns, promotional materials. You can also use generative AI to generate pictures of people that look pretty real most of the time, okay? Okay, the whimsy is very much there in pictures and there was a lot of that on the news. So anyway, prompting for images in CoPilot, here's a kind of a formula that I think you give it information. So describe what kind of thing you're looking for. Is it doing something? Is there, and then is there a particular visual style you want it to use? So in this case, an adjective plus a noun plus a verb plus a style, right? Curious frog, analyzing data in a comic book style, okay? And with that style one, maybe avoid blatant copyright violations like in the style of a Banksy portrait or something like that. Yeah, that comes up in- Oh, there we go, sorry, Kim, I jumped your slide. That's fine, it's worth mentioning twice. And in fact, AI doesn't let you do it anymore. So the types of things to include in a visual type of prompt is, you know, vibes, right? So a mood or a feeling that you want, right? You can ask for different media. You know, you can ask for photography. You can ask for a high resolution photograph. You can ask for taken by a Nikon, DSLR, blah, blah, blah with, you know, f, two point whatever, right? I forget my photography, but you can ask for close-ups, wide angles, things like that, illustrations and art history. So here's the big caveat. And actually some of the AI, and as Josh mentioned, it does, it's changing and learning all the time. So it changes a little every day, right? Or as we're working on it. So one of the things that's become much more restrained is the ability to request a specific artist style. Like give me this in the style of a Picasso, right? It's not just living artists. It's also certain artists' state to really come up against this. So you don't wanna say like with me this image as Picasso would do it. As much as saying, you know, if it's Cubism, you're after, let's say, you could say in a Cubist style or in the, you know, do something in the style of a certain period in time, like 17th century watercolor. You can also ask for the medium, right? Watercolor and wood. You can ask for claymation looking visuals. So, and it's really fun to experiment. So I think, you know, and that's one of the things that is also really important around just AI, prompt engineering is not so much about technical skills, right? But it is about experimenting and getting comfortable experimenting back and forth and seeing what you get. So I think that's it on images, similar principles. Yeah, and just, we're almost done with the slides. We're gonna get right into the questions. We just wanted to kind of give some basics and we're also actually answering through these slides a number of the kind of questions that in various ways were asked multiple times by audience members. So we're trying to get some of the most asked questions just covered in this. So advanced prompt techniques, there are literally thousands, maybe tens of thousands, maybe millions of very, very sophisticated prompts that are very, very specific for purpose. So we're not gonna obviously go into all of those. So what we are gonna share are, what we would consider to be some of the more advanced prompt techniques that are very flexible, that can be used for lots of different purposes because they're effective and can be used in lots of different ways and will provide you with samples of all of these that you can go and use and play with. So one is a chain of thought. And there are a couple of different ways of doing this. One is super simple, which is basically just telling the AI, before you start doing this thing I'm asking you to do, I would like you to break it down step by step and maybe even tell me what your plan is and let me say, yeah, that sounds good before you proceed. It's been pretty well demonstrated that this will tend to produce higher quality outputs most of the time, although it does take a little longer because the AI has to go through that and you have to review it. So that's a chain of thought. And you can also actually give it the steps yourself if you know what those steps are and that can be part of your prompt architecture. Another term to be familiar with is the idea of zero shot, one shot and few shot. And this is super simple. If I want it to produce a blog post for me, that is 500 words and follows a specific, let's say five paragraph format, introduction, three basic points, and then a conclusion, a zero shot would be, I don't give it any samples. I just tell it, this is the kind of output I want. Good luck. A one shot is I give it exactly one sample that's exactly the way I like it and say, here's a sample of the kind of thing I want. All right, and a few shot is I give it maybe three or four or five or 10 or 100 samples. And then it will look at all of them and try to do that. My experience, Kim, tell me what you find and also the audience, please share with us your experience. I find one shot most of the time to work the best out of these. If I have a really good sample of the kind of thing I want an output for, the AI does a really good job replicating that for some new thing that I'm asking for. Kim, does that... I would agree. Yeah, and the general consensus that I hear, right? You think more shots the better, right? The more examples. But one of the things that I keep hearing and keep experiencing myself is if you load it down with too much information, you can confuse it. So if you're better off giving it one really good example or really close example to what you want. Yep. And that's what I've found. And then the last one that I want to talk about is this idea of Tree of Thought. This is probably, even some of you who've used AI may not be familiar with this one, even if you knew the other terms we're talking about. Tree of Thought's very interesting. Definitely takes longer than the other ones. And this is where you actually, and we have a sample of this as well, you actually ask the AI to create personas and then have those personas kind of debate amongst themselves which of the solutions is best or even generate ideas and then debate them amongst themselves. So for example, I might say, I want ideas for how to incorporate AI into my nonprofit organization. I would like you to generate five to 10 ideas. And I would like you to generate them and one person is gonna be an expert outside consultant in AI. Another person is going to be one of my nonprofit program staff and another person is gonna be a board member of our nonprofit. And I would like you to have each of those personas generate three ideas. I'd like to then have you have those personas debate about those ideas and provide rankings of which ideas the best based on each of their different perspectives. This is a very interesting way to have the large language model kind of take in these different perspectives. So that's one that I've played with with some success at different times. We've got, Kim, will you drop that in the chat for everybody? The prompt template that we're gonna share. I think one of the TechSoup crew will drop it in. And I'm gonna show some other ones as we go through and then we'll take it from there. So I think I will stop sharing at this moment and then I'll sort of reshare my screen, Kim, as we go through other questions. Yeah, so actually since you're on the topic of that template, it's probably a good time to go ahead and show it because one of the, we received a number of questions along these lines but this is probably the most clearly asked and kind of general. How to tailor the most impactful AI prompts that align with our organization's mission? Okay, so let me, oh, go ahead, Kim. Yeah, and I'm just saying so. And a number of people ask this of kind of like, how can we create a prompt that gets what we want and has our tone to it, right? It has our, so I believe our template is designed to do just that. It is, and let me, let me reshare my screen. So let me come back here. Let's see, maybe I'll just share, can I just share the whole Chrome thing? I can't see that. All right, we'll just, I'll just bounce it around then. So let's go right into co-pilot. Okay, so here we are in co-pilot and I'm gonna jump right into this new feature that for those of you who don't know about it and use co-pilot, I'll be very excited, which is notebook, which many of you will see, if you go to co-pilot in your browser, you will see this little notebook up here on the top and you can kind of click over to that, which is just new. This is the thing I literally just discovered a half an hour ago. And the first thing that is super interesting about notebook, and I know my little annotation tool, if you look down in the bottom, if you see that blue kind of button in the left top of words, this new topic, just to the right of that, you'll see I have a whopping 18,000 characters that I can put in here. Look at how long of this prompt that I put in over here on the left. And I've still only used under a quarter of what's there. So I could have something four times as long as this. And so what I'm doing here is I'm telling it, this is who I am, right? I'm telling it, here's an example of my writing. I'm telling a little bit about my audience. I could obviously be much more specific here. Okay, I'm telling it who the AI is. I'm telling it, when I give it tasks, what I want it to do, there's that chain of thought, take a deep breath, consider the request and full, convert that into a step-by-step plan. Say, if you have questions, you're gonna ask me up to five questions, tell me what assumptions you're making. This is something that's a prompt that's evolved for me over time that I just found to be very effective at doing complex tasks. Then I tell it what format I want it in, which is that I'm asking for it marked down, all right? This is a really cool thing that I like, which is after a response, tell me for further explanation and give me great follow-up questions that are thought-provoking. I find this very helpful. And then I've given it the chain of thought where I say I wanna develop an effective strategy to be able to be using AI. I've told it, I want a board member, an executive director and an outside consultant, a sign of probability of success in the confidence level and free solution deep in the thought process and so on and so forth. And here's the output that co-pilot has given me. So the board member came up with AI-assisted content creation and outreach, kind of goes through this. And all this, of course, I can export as word or PowerPoint. And the nice thing here is that I can make changes over here on the left and then just click the little arrow and it reruns the prompt. So it gives you a little bit more of an iterative way of working through prompts. So if I was trying to get something that was really in the tone of my organization, I would take the time to create a prompt that includes all of the things I need about who my nonprofit is, who our audience is, or our audience for this particular task that I'm doing, an example of what our tone is, our style, those things, and then give it the task and be very clear on that output. And now that you have notebook, that could be 16,000 characters long, which gives you a lot of room to play with and you can iterate back and forth. Yeah, and it just went up from 4,000, which is a lot. So you would keep these kinds of prompts that have worked well for you. Like you can always just copy and paste those into a document that you keep and you might use certain prompts for different purposes. Right? And EJ asked if I could use co-pilot on my Apple products or not. EJ, I don't know if I should say this on the Microsoft TechSoup, but I'm on a Mac and it works for me. And I use it on my Apple products. So here's a question that I'll go ahead and answer that was also submitted, okay? So related, so hold on to everything you've learned up till now. Are you gonna want me to share anything here, Kim, or are you gonna share anything? No, I'm just looking for more just to talk through. All right, I'm trying to take, so this is a two-par question. I'm trying to take existing content, say a blog post, turn it into an email copy. What's the best way to feed that content to the AI tool? I would say feed it the way you saw Joshua feed an example of his writing. And you literally tell it. Like I want you to be, you know, to, you are a professional writer and copy editor. I would like you to transform this content that I'm gonna give you. And I would like it in an email and you give it the type of tone. If you have an example, if you have something about the audience for which that email is going, you know, you could say I need it to be more professional and, you know, sound more formal. You can do things like that or friendly. And then the second part of that question, which I think is really important, like kind of that I want to, a misunderstanding that I want to clarify is what about training AI tool on my organization style and messaging guidelines to minimize editing requirements to make sure the results are on brand? Here's the thing, you think as you're writing and working with this tool, this thought partner, right? It kind of feels like an intern that you've brought. The longer the interns there, the better it gets your style. It kind of done work that way, okay? AI tools are very session-based. So it's remembering in that given session what you've told it, right? So it's not gonna remember tomorrow. Oh yeah, this is Kim and she does this and that. So you, that's why it's really good to have a library of prompts that works well for different situations that you develop in your organization. There are tools coming out. I think co-pilot studio, which does, will cost more, but that will allow you to kind of build custom like kind of little AI modules to add. And those you'll have a better chance of being able to then get it to be shaped in a certain way for every question that you ask it. But right now the tools that we're showing them today, it's not remembering anything like, you know, about you, even though it seems like it should. So next question for you, Josh, this is another demo on the Tree of Thought Prompt. I think that's helpful. And this is great for like marketing campaigns and things. If you're, you know, if you want a focus group, I just saw someone ask that question. It's kind of like having a focus group. Yup. Yeah. Thank you, A, for noticing Clippy. Clippy's helping us out today. I'm really appreciate that. We also have the Ministry of Silly Locks for any MoneyPython fans. So let's see, I think we kind of demonstrated that one, right? The last prompt I had was kind of the chain of thought. I could certainly do another one. What I was going to do, if it's all right with folks, is I was going to just show another prompt that I will share with everybody. In the chat as well, which is an AI tutor prompt, which I think can be kind of fun. So I will drop that in as well. And then I will share my screen. Is that all right, Kim? If I just quickly kind of do that? Yeah. Because this is one that I find really fun and also potentially powerful for folks. So I just shared the prompt that I put on the left, which I got from a wonderful writer and professor at Wharton called Ethan Molek, who has a wonderful sub-stack that he writes about, which is one useful thing, could not recommend it more highly. So this prompt is one that he wrote. There's lots of these out there. So basically it just says, hi, I'm your AI tutor and I'm here to help you with any questions you have, so on and so forth. Now this one actually, I realized, won't work well in the notebook version. So I'm actually going to move that over here and put it into the regular co-pilot where I can have an ongoing conversation. So notebook, again, it's one prompt, so you're iterating it like that. You can't do a sequential kind of thing with notebook. So notebook is more for one-off kind of things but allows you to do it incredibly comprehensive. So if anyone have a topic they want me to ask the tutor about, so you put it in the chat first. I'll take any topic as long as it's a safer webinar. If you have a topic you want to talk about today, let's see if we give it a minute. All right. No one coming in with a topic? Event strategy, all right, thank you, yes. All right, so I will do a nonprofit event strategy. Let's see what it does with that one. All right, I'm going to think about it. It's a topic that covers that nonprofit organizations can plan, organize and execute successful events. Am I a high school student, college student or professional? All right, I'm just going to answer these for everybody, so I will say I'm a professional and it's going to truck along. Notice that Copilot is dropping little links to the resources that it is using to provide these answers. Really important to understand that the quality of the responses sometimes is dependent on how good the responses are. So I'm not going to spend too much more time on this but you can see it's just going to, from this point forward, now that it's gotten this first instruction, the AI tutor instruction, it's just going to, you know, keep going. Did I share the right tab, by the way? Oh, yeah. Okay, great, great, I think I'm just making sure. All right, so I'll stop there, but hopefully that's... So long as you are there and in, oops, and in creative... Someone did ask right here, an example of running a prompt for AI generative image creation and that person's name was Josh. Okay. So is there any particular image, Josh, that you'd be interested in seeing, not to put you on the spot? You're asking me or asking me Josh? No, no, no, Josh, another question. Josh, you're asking the chat, yeah. Any image you want to see, Josh? Again, save for a webinar please. We don't need to put you on the spot, you can always... And Kim, I might defer to you on this one. I consider you much better at image prompting than I am. So intergenerational mentor relationship. Oh, nice. That's pretty cool, all right. So Kim, help me out. Oh, okay. So I would say, okay, so for an intergenerational mentor relationship, okay, so that's a nice one. I would say now intergenerational, I'm assuming that's a young and an old person, so please make me... And we're gonna ask for a photograph, but we can ask that. Please make me an image for... Well, actually, why don't you send that one through first and just see what it does? Okay. Okay. And then you can start to get specific. You can start to ask for something like, give me an image of a young girl, of a woman in her 20s working with an elderly man or a elderly man at a computer or vice versa, not to be ages here, and do it in a photographic style. Sometimes I... Okay, repeat that. Okay, an image... Okay, image of an intergenerational mentor relationship. Go ahead. That's what it gave us. So let's say, no, we actually wanted somebody in their 20s image of an intergenerational mentor relationship showing a young man in his 20s Okay. Working with a woman in her 80s. Okay. You know, at a big table, let's say. Okay. You know, in the style of a Kodachrome photograph. Okay. Off we go. While it's doing that, we can kind of take a look at these ones in larger views. So this is how these came out. Kind of funny, I don't know. Black and whites. That's kind of... It made black and whites and it made them all females and all white people except for one Asian girl, which is interesting. And this is certainly a... All right. So now we got it a little bit differently. I wouldn't say the photographs. Well, I guess that's kind of Kodachrome. And now we got a bit more diversity, which is very nice to see. Okay. Okay. And so you can keep on iterating. You can also ask for an illustration, right? So here we're asking for photograph. You can ask for an illustration. You can ask for a watercolor, et cetera. Oh yeah, AIs can mess up people's hands. Although they're getting better. AIs also in images can mess up text, which I think is counterintuitive for a lot of people. It is getting much better, but it can really misspell words. And that's with all of the different systems. And Josh did get back to us. So we'll say a child being picked up by their parents in a small African school. So we'll see what we get from that one, Josh. Now you can get some bias warning, but... Yeah, all the AI systems are having various challenges around, reinforcing stereotypes in image creation and doing all sorts of nutty things. You can see we got really different results here before and quite a bit of variety. That one's lovely. Yeah, it's interesting that we got two that are kind of cartoon-style and two that are very photorealistic. Kind of interesting, right? Neat though. So Josh, I hope that you found those helpful and you can certainly do that, all right? So what's next, Kim? So I think, I mean, some people did ask and I've seen some questions float by in the chat. We're ready TechSoup team to start taking some of the live questions that might be flying by, but... We can do the fun one. Wait, well, before we do them, we can show you how they're, one thing they tend to mess up on. Someone did ask for an infographic, right? Yeah, well, I think we have to, Kim. I think we've got a idea. Right, so it's doing all that. This is gonna be a fail and it's gonna be a big one and this has nothing to do with co-pilot and everything to do with large language models in their current state. All large language models will do what we are about to see. So we are going to ask, can you create an infographic? Please create a detailed infographic showing ROI, operating cost of an income percentage of funds and operations, calculate the total cost of a project and output. So we'll do that for a multinational nonprofit organization. I don't know if this is like the best. Okay, so we'll see how this comes out. But someone asked it, I don't know if this is like the best, okay? So we'll see how this comes out. But someone asked this and Kim and I knew this in advance and we tested this beforehand. Hopefully it will do what we expected to do, which is produce some infographics that are not always that helpful. So, and we'll see what we have there. Notice, it's telling you what ROI is, which it's much better at than creating that infographic. It's stalling. Kind of chain of thought itself, right? Sort of saying, I need to think about this and make sure that you understand the infographic. Thank goodness it's going to tell us all this, all right? Okay, I'm not sure it's going to get to this. Another thing that that that AIs are really, really bad at and I'll try this as well is where's Waldo? It's not giving you the infographic in this case. Yeah, it's getting there. Okay, all right. Someone did ask, someone would like help making an image for STEM science camp for kids. It's so nice. Okay. We can work on that. All right, here are fantastic infographics. That's really good. Yeah, that definitely helps me understand ROI, NETVM. This one's even better. I think, and this is, I have not figured out how to do that. Josh, pardon? We don't see it close up. Oh, you don't? It didn't zoom in? No. Oh, interesting. I wonder why not? I clicked in on it. I'll stop sharing and I'll reshare and see if it shows you. Interesting that it wouldn't show you so why don't we do, I'll just do the entire screen and then hopefully that'll work. Sorry, Kim, you're not gonna see yourself for a little bit. We'll go back to the pilot. Now, does everybody see it? Ah, that one's, yeah, see. I mean, that really clears things up for me. I think we've got $2 of return on investment which appears to be 90% of, and then we have 9% of our OFI, whatever that is. And then it looks like there's a person sitting underneath $3, and this is exactly what I get every time I try to do it. All of them do it, right? Yeah, they all do it. Not an indictment of co-pilot at all. No, and someone did ask about, there's me again, someone did ask about what are some of the low tools and there are co-pilot is obviously the tool we're talking about today. There's also ChatGPT which comes in a ChatGPT plus version and, you know, Claude AI and Google just released newly named Gemini. What's very nice, and I would say this, even if this wasn't a Microsoft sponsored webinar is that with your, if you're a licensed Microsoft user on any of four licenses, three, six Microsoft Office E3 license, E5 license, business premium or business standard, I believe, all of those, all right, include the, hang on a second, you're protected. And so you can see, because I'm logged into a Microsoft account that has an E3 license here, my Chat is protected. And if you click through to this, it'll give you all the details around the privacy and what they're doing, but essentially it's saying, as I read this anyway, if anyone from Microsoft is on and wants to correct me by all means, but it's basically saying, you know, if you're already at the data that you're using to interact with the AI is already in your 365 tenant in OneDrive, SharePoint, Teams, Word, you know, whatever, you're not introducing it to anything new by using it with your licensed co-pilot version, right? It's protected by all those same things. That is a big deal, all right? And Microsoft is currently one of the only companies that's doing that on what's not quite free because you need a licensed 365, but since most of you already have that, it's essentially a free AI tool that's available to you and has data protection. That's a big deal. Yeah. And another question, this is Janet from TechSoup here. We did have a question from Josh actually, the other Josh. And maybe this is around the best practices, but so would you suggest starting with vague prompts and move towards what you were looking for with specific continual prompts? It's kind of like what you were showing us a little bit before, but yeah, maybe could talk about the best practices there. Can you repeat that, Janet? I'm not sure I understood that question really well, but Josh, did he write it in the chat and I can go find it? He did. It's around, yeah, starting with a vague prompt and maybe going and adding continuous details to create what you want. Gotcha, okay. So that's an interesting question, I see it. Thank you, Janet. So Josh asked, so do we suggest starting with vague prompts and move towards what you are looking for with specifics through continual prompts? Kim, do you wanna answer that first or do you want me to take that on first? My inclination would be no, if I have, look, here's the thing, I would start with a vague prompt if I have a vague idea, if I'm not quite sure. And I have certainly many, many times gone to an AI tool to ask for general vague ideas about something. But if I know what I want, I will give it that in the very beginning so that then when I'm fine tuning it and continuing to iterate with the AI that it gets me closer and closer to what I want. So that's my inclination on that. I agree with you, 100%. Yeah, it's a great brainstorm partner though if you're not quite sure, right? So any other questions, Janet? I also see from Sarah, can you create coloring sheets with co-pilot? Yes, I can. I just prompted that now so we can see how it comes out. Now, we're not seeing, you're sharing your whole desktop? Yeah, because otherwise it seems to not zoom up the things, is that okay? Yeah. Okay. It's just smaller. Oh, I can try it again. It wasn't, just when I did it that way it wasn't zooming the photos for reasons I'm not clear on. We'll see if it does it this time. Tell me when it gets larger when I click on it. So yeah, that's pretty good coloring page. I'm not. Now, if you wanted it for like a six-year-old you might have less detail. I can certainly try that. Well, while that's happening, one of the questions that someone asked had to do with using prompt engineering for automated replies. That's something I might actually avoid, okay? Because one of the, and in our last class on AI risk. It's extremely low risk. Like so low risk that it almost doesn't exist. That's still hard for a six-year-old. I don't know, maybe I'm not good at coloring, but. That does not look like a child. That looks like a dog child. Interesting. But that one's not bad. Not too bad. All right, all right. Anyway. Well, there's yes. I think you might need to do a little tweaking, but yeah. Any others? So yeah, short answer around automated. So credit to George Weiner of Pull Whale for this term, but real shorthand, okay? First, not final, human in the loop. So what that means is when I'm using AI systems, I am, is not the final draft that comes out of an AI and a human must be in the loop. If I am allowing an AI system to do automated replies that are not reviewed by human and those are going to humans, in my view, at this time, that is irresponsible unless it is so low risk and you are disquaining a million times that like, hey, this thing may give you bad information. Yeah, cause it can, you know, there are unhappy stories about that. Yeah. Janet. Jan, yes, you can download all the images. Don't know about dimensions and file formats for the image. I can certainly try that. Do you know that camera? I have no dimensions right now. I think with co-pilot, you get a square. Like any minute now, you can probably ask it for a 16 by nine, but right now that it is generating square images. Some of the other tools, you can give it dimensions. It's always gonna come in as a PNG file. Yeah, I'm asking, it says I cannot provide the images or download in any format. I can only create and split them for you. If you want to save the images, you can take a screen shot or download them. So, sorry. We also have another question from Chris Lynn and this is more about understanding how AI generates their images. But she's asking, if we type the same exact prompt to create an image in both co-pilot and open AI or chat GPT, will they create the same image? No. In fact, even if you paste the same prompt into the co-pilot over and over again, it will make different images every time. This is the kind of magic of AI systems. They're always generating it from this probabilistic statistical model that they're doing. So it's always new. And it's always changing because people are always using them. They're very dynamic. So I know we're flying without a net. I was playing with this just or a perform without a net. I don't know what the right term is here, Kim, but I'm going to play with, I'm going to show us something else again. So I've got, unless Janet, we've got other questions we want to hit, but I wanted to show now that we've got this notebook function in co-pilot, something else we can do that's pretty neat. So there's all sorts of free data sources you can download. Okay. What I have is something called the IMDB Movie Database, which is just like something I used from way back in the day to help me learn like how to do pivot tables and things like that. And it just has names of movies, has a bunch of data in it. So what I'm going to do is in my notebook where I have a lot of context. So this uses up the whole context. I'm going to delete some of it and I'm going to go all the way up to the top. You see, I've almost used my entire context window here. And I'm going to say, please tell me about this data. Stop responding and let's go. And when I was playing this, whoops. Okay, let's start over. Like I said, flying without a net, but I'll put it to work. All right, we'll try this again. It was doing a pretty decent job telling me about this information and then giving me some options to ask questions of this data. So just pasting in, and this is a CSV file. So I'm just pasting in basically text data. It's figuring this out all by itself, just looking at that data. Okay, so it's saying it appears to be a data set of movies. Each row represents movies with the following attributes. Really nice breakdown here, right? Of what this data looks like. And then gives me some ideas of what I could do. And now I could start to ask some questions. I could ask questions about like correlations between different data points or things like this. This was not previously something that I was able to do very successfully with the co-pilot because the context window was so limited and you didn't have to be able to upload giant CSV files. But now that I have this context window of 18,000, as long as your data set isn't too large, it's something you can definitely play with there. So I just wanted to show that because I think data analysis, the ability to do advanced data analysis using these tools is one of the most exciting aspects of them. It's really like a superpower. Yeah, and unlike infographics, it can actually do data visualizations really well. And that's those are the kinds of features I'm looking forward to seeing as Microsoft rolls out more capabilities in co-pilot for Excel and things like that. Again, these are all works in progress and it still is early days. Yeah, and Marina asks a great question. How reliable is that data you produce? So what I have found so far, Marina, working with large language models, there's something called RAG, retrieval augmented generation, I wanna say is what it stands for. It might be something else, but it's essentially a technique by which you're saying, here's like a set of data to the AI to co-pilot and saying, I want you to answer the questions I'm asking, but only of this data. I don't want you to get data from anywhere else when you're answering this question, because of course they've got their entire training thing and they can browse the internet and all that, right? So RAG are different techniques for doing that. This is on the more sophisticated side of prompt architecture, right? So in co-pilot studio, right, you'd have much more control over uploading the data you want and then configuring the prompts and testing it to make sure it's doing it. If you're doing it like this, Marina, I would say you definitely would wanna go do some spot check validation. So do some easy things and see if it seems to be giving you correct answers or just making them up or giving correct answers, but from some other data set that isn't going to be so uploaded before you went and published anything. Again, first, not final, human in the loop. Thank you, George Wainer. Oh, I did get it right. Retrieval augmented generation, thank you. Yeah, yeah. Okay, any, Jen, did anything else? Nothing from the chat, but we did have a few questions that came in previously that maybe we have some time to check in on. Let's see. How about, do you have any advice on getting the tone and language to reflect an organization's voice or language guidelines? You can show that. Absolutely. So I'll share, let's see, second. And some folks had commented that the template that we shared didn't have that full, it had the prompt that we wrote before we discovered notebook. So we will go ahead and add that to the prompt strategy template that then we'll use the same download. And Janet, you'll be sending out a followup email with this so folks can have the link and we'll give you both of those prompts. Correct, yeah. We'll be sending out the recording and then as well as this document as well. Okay, great. Cause yeah, I mean, Microsoft like quadrupled the capacity. Right before this. Well, actually it didn't happen right before this class when we discovered it then. So are you doing a... Yeah, so I'm giving a couple of ways that I would solve this problem. So example of desire, writing style. So I'm just pulling us from my other one. So if I... There. So, all right. So one way that you could do this, right? As to one shot it, so to speak, which is that within this larger prompt strategy template that we give you, right? I'm giving identity and purpose and saying who the AI is. I'm giving you the job. I'm then gonna go ahead and give it an example of a desired writing style, right? The other thing I could do is I could provide style guidelines, right? And then I could say, be witty, verbose, use lots of jargon, over explain and make things up a lot, right? And so obviously this is not what you would want, but you could use your style. So you, and by the way, if you're not sure how you would describe your style, this is another way to use AI. So what I could do is take my writing sample, right? I can move that into co-pilot. Okay, so let me share the co-pilot window again. Okay, do what you do. And we'll go back to co-pilot and I will say, please describe this writing style. That's it. And it's going to go ahead and describe that writing style. And then I could just basically take this and paste it into the writing style guide that I have. And you could provide lots of them and then say, please give me a summary of these writing styles. So you can use the AI to tell you the style you're writing and then feed that back into your prompt architecture to help it, all right? And you can see this is giving much and kind of clear things, okay? So it's really nice about my writing style. Nothing too insulting there. It's good. Someone did ask for an image if we had time. And so I think since we have a minute, they wanted an image for a STEM summer camp. All right, any other details about this STEM summer camp that they want? Get the name of the person. If you're there. Oh, images for a STEM science camp for kids. And if, let's see what it does. It's remain there, all right? Interesting. We can only generate an image for you deciding where the personal Microsoft account. Interesting, all right, let's refresh. Let's go back. I hope it's creative, although it does do it in balance. Yeah, it was made to be images this whole time of a STEM summer camp for kids, okay? Science, see if it does it. Maybe Destiny is off like implementing some of our new policies, Kim, and just like disabled it in a round table. But hopefully it'll work. Yep, now it's doing it. That was just a hiccup. By the way, these learning language models will hiccup a lot. So if when in doubt, refresh, close the browser, start over, nine times out of 10 and it'll work fine the next time. So are these going large again, Kim? Are they? They are, okay. So young kids are surprisingly happy. Wow. Five to 14 years old. With the name of the organization, that's where I see science fun for everyone. I might put that in with another tool because you could get some misspellings on what happened in there. That's improving, but not perfect yet. Yeah, I can try it again. Science camp for kids, call. All science fun for everyone. Science camp for everyone. Science fun for everyone. It's like those kids were already really aggressively having fun. But we're over time. All right, sorry, everybody. I have to go actually, because I have the three o'clock, but I'll wait till the image. But thank you, everybody. Really appreciate the time from TechSoup. Kim, of course, you're welcome to stay around if they're staying around. I'm gonna have to jump. But here's our science fun for everyone. Yeah, AIs are not great at spelling. A wonderful writer. Another great sub-stack, I recommend, is Strapatek. And I'll drop this link in here. She was specifically, this is Deb Sulegros, who's just a fantastic writer on AI. And she actually was just writing about improving spelling using images. So thanks, everybody. I gotta jump, I lost track of time. Thanks, Kim. Thanks, TechSoup. Thank you, everybody. All right, thanks, TechSoup. Thanks, Josh. Thank you, Kim. All right. Okay. Did you wanna take anything else, or are we good to wrap here, I think? I mean, I think we're good. Okay. If someone has another question, I'd be happy to take it, but... Yeah, I think just at this point, I think just following up with us for anyone that has additional questions, I'm gonna go ahead, I guess, and just toss my email in the chat. I've already actually received some emails from some folks, which is exciting. Be happy to field anything as far as with Kim and Josh and Roundtable. Certainly be able to communicate any information. I'm sure I'm guessing that's gonna come out and follow up email as well. A copy of this recording will be available in the coming days. Just give us some time to edit it, because you knew I was hosting it, that there was gonna be some glitches. So we're gonna get those out as part of this as well. I'm assuming presentation deck will also be included. There were some other items that some people had asked for, to what degree that can accommodate those, we'll certainly try. So with that, thank you very much, everyone, for joining us today. This morning, this afternoon. Have a great rest of the week, and we look forward to you joining us on another event.