Generative AI Full Course 2026 [FREE] | Complete Generative AI Tutorial For Beginners | Simplilearn

What if one AI tool could help you create content, analyze the data, create images, summarize research, build workflows, design apps, and even prepare you for real world AI projects. That is exactly what today's session is about. Hello everyone and welcome back to Simply Learn. In this video, we will explore how generative AI is changing the way we work, learn, create and solve the problems. Today, AI is no longer just a tool for answering questions. It can help professional translate content, analyze customer feedback, generate marketing, copy, create, design, automate workflows, build product road maps, and even support data science and machine learning tasks. And this topic is important because the future belongs to the people who know how to use AI smartly. It is not just about typing a prompt. It is about giving the right context, asking better questions, checking the output and combining AI tools with strong technical knowledge. In this session, we will cover practical use cases of chat GPT, image generation, multimodel AI, custom GPTs, automation tools like Zapier, design tools like Canva and myro AI and also AI powered tools for research, video creation, transcription and data exploration. We will also look at important data science and machine learning concepts such as pandas data cleaning regression classification confusion matrix precision recall and assemble models and also imbalanced data sets. By the end of this video you'll have an understanding of how generative AI can speed up creativity business analytics and technical workflows and why human judgment is still very important. Before we move ahead, let me quickly share something exciting for business leaders and professionals who want to understand how generative AI can actually transform work. The Michigan Engineering Professional Engineering Generative AI applications for leaders is designed to help leaders go beyond basic AI tools and learn how to apply generative AI across real business functions like sales, marketing, customer services, analytics, product research, software engineering, and workflow automation. The program includes live online interactive classes, real world case studies, peer discussions, and hands-on exposure to tools like Chat GPT, Microsoft Copilot, Hugging Face, OpenAI, Zapier, Make, Figma, Power Apps, and more. You'll be also working on 12 industry focused projects including automation, content workflows, building generative AI powered support agents, designing AIdriven campaigns using generative AI data analytics, and planning product launches with Copilot. On completing the program, you'll also earn a program completion certificate from Michigan engineering professional education along with the Microsoft course completion certificate hosted on the Microsoft learn portal. So if you're a leader, manager, entrepreneur, or even a professionals who want to use generative AI strategically to improve productivity, innovation, and business impact, this program is worth checking out. The link is given in the description box below and in the pin comments. Go check it out. Before we move ahead, here's a quick question for you. Which of the following is most important for getting better results from generative AI? Is it using very short prompts? Giving clear context and instructions, asking only one question, trusting the output without checking. Let me know your answers in the comment section below. So, we're going to start today by doing this one and then as I said, we'll move to lesson two. Um so this practice is to uh leverage chat GBT and a little bit of prompt engineering potentially um to create a 2x two kind of prioritization grid. Um and the the background to this is to uh kind of analyze potential product features um based on two criteria. So namely their business impact as well as their uh feasibility like how possible is it to actually build those features. Um and we'll organize those into kind of this 2x two grid to determine which ones we should prioritize, which ones we should dep prioritize based on this grid. Um so it's an interesting kind of uh way to organize your thoughts maybe around different features, different ideas. Um and we'll use jet gbt to help us do that. Um so the again this kind of scenario is we are imagining we are uh working as the CEO of bills smart a SAS software as service company offering billion billing solutions for retail businesses. We we want to add a uh a a powered flyer generator to make your product stand out. um and it will dynamically create visually appealing flyers etc etc etc. So um our goal is to again kind of assess the priority and feasibility of potential features we could add to this uh flyer generator product. um and we will organize those into this grid and see um kind of how they get prioritized uh according to their impact and feasibility. Um so here's the potential features that we're going to rank um and they're not really in any particular order um necessarily, right? But they're going to be placed into this grid. So um these are potential features and of course we could have a larger list than this. We just have um several of them here which is basic flyer generation. So generate flyers of product images, prices, promotional details. So just the basics that's probably a standard feature that we should do obviously. Um and then personalized flyers. So tailor flyers specific customer preferences or purchase history. So um that one may require using more data more more um purchase history from users. Um uh and then we can try to you know build out custom flyers that way. Um seasonal themes. So automatically adjust flyer designs based on holidays or seasons. And then analytics integration provide insights into which flyer designs drive higher sales. So, integrating analytics there um to see which flyers are the best. So, those are some of the features. Now, of course, you might have some others you want to add to this list as we're going through this. That's totally fine. Whatever ideas you have, you can definitely add them in there. Just starting out with these four. And what we're going to do is um we're going to ask Chad GBT to help us generate this grid to assess these features and basically assign it to the appropriate uh quadrant of the grid. Um so and then uh at the end we can come back and answer some of these questions once we've created that grid and see uh you know see what it's produced for us. So maybe we'll come back to these uh questions, but let's take a look at um I wanted to go through and take a look at uh some of the um some of the the fe the basically the the requirements of the grid which are that um we want to basically have four different quadrants of the grid. So in the top right of the grid is going to be um this section that includes features that have high impact and high feasibility, meaning you know they're really doable and but they drive a lot of impact. So these would be um up in the top right and we could think of those as kind of like our quick wins um um category of of features. Then there's ones that are less doable uh maybe take longer time and but they still have a pretty high impact and those would be in our top left category. Those we could think of as kind of strategic investments like we should be working on those. Those are features that are important. They do have a lot of impact. They just take longer to do. Um they have lower feasibility. Then we have low impact but high feasibility. These would be the bottom right and this would be kind of incremental improvements. These are ones that are really really doable. They just don't have as much impact. Um so they're kind of incremental smaller wins. Um right, they're doable. They just are don't have as high of impact uh as a feature. Then we have kind of the the last category is the bottom left which is um you know low impact and low feasibility. So this is uh this is not having a lot of impact also not being very doable. So we should probably just depp prioritize those. So this would be kind of a bucket of features we should probably just depp prioritize and then of course here's the features we want to uh evaluate and we're doing them on those two criteria impact and feasibility. So, it's really, you know, a, if I were to draw it out, it's really um kind of a grid where we're measuring um, uh, impact this way and then, um, feasibility kind of this way, right? Feasibility. So um so exactly what this kind of describes um you know the low the the quick winds would be in this corner. High impact high feasibility would be uh up here and then the um high impact um low feasibility um would be uh um high impact low feasibility would be down here in this picture. I know it says top left, but in this picture um and and on and on and on, right? Okay. So, this would be our um this would be kind of the prompt we want to use as an example. Um so, just to uh organize this our thoughts around this, right? If we were doing this, this is something we would definitely want to do, which is provide a bunch of context. So, we wouldn't just jump in and say, "Okay, here's my features and here's the I want you to put in a 2 x two grid." You never know where that's going to take you, but this is providing a lot of detail. That's why it looks like a really big prompt and it's because we have um you know a a lot of information that's good as background. So for instance, we're talking about um the kind of product we have which is product flyers for retail customers. Um just to describe a little bit about what that product is and then we're talking about the an example. It's almost like a oneshot example prompt, right? So we're saying here's a bunch of detail about this grid as an example grid. So don't just build out any grid with any kind of criteria. It's actually driven based on these two things I really care about which are impact and feasibility. Um and so we have a bunch of details about that. And then we have our features that we want to prioritize. Um so it's a it's quite a lengthy prompt. And in reality, what might happen is we it might take us a few iterations to land on this step. Like, you know, we might have to follow up if we didn't include all of this right away. We might have to follow up and say, "Oh, I really wanted to prioritize based on, you know, based on feasibility and impact or I really wanted to prioritize based on um these other these other factors and that would build a completely different grid." Right? So what we can do is um paste this in. So this is all of the prompt but just uh realizing that you know it we may not get this prompt right away. There may have there may have taken iteration may have taken um you know quite a bit of back and forth to land on this. But th this is a good example of including a lot of details up front so that we you know maybe don't have to spend a lot of time iterating. Not that that's a bad thing, but it does save us some time, right? If we can uh kind of put that all up front in the prompt. Um so we'll copy that and then I'm coming over to um CHBT and just entering that uh in here. So we have a bunch of details. Here's all of our background. Um, here are the criteria for the grid. We have all that detail in there. Um, if you want to try this yourself, by the way, um, you can use this text as well. Okay, so it says here's a grounded way to map these features. By the way, that was a pretty fast generation, right? Um, didn't take too long at all. Here's a grounded way to map these features in a practical sax lens. Um, I'll place them each in 2 x two. So our quick wins our basic flyer generation. So this is our it says this our foundation um and if it isn't strong nothing else matters. So this is high impact high high feasibility feasibility even gives you some examples of what we can do um to accomplish this. So this first basic flyer generation gets put in the quick wins category. Um then we have our strategic investments category which is high impact low feasibility. The personalized flyer go there. um that can have a really high impact. That makes sense that it categorizes categorizes that in that category, but um it takes longer to build that out. Namely because you need um you know data collection and to um maybe do some type of modeling there. This is where um you know we can use some of the things we've learned in the past, right? We can do things like clustering um to segment the customers into different groups uh in order to um you know maybe have like five different kinds of flyers for five different kind of groups of customers. We can do that and we we have like you know we've done that in terms of building those kind of models but we need data to do that right. So that would be the that's why this is in the low feasibility category because it just takes longer to build that up um to to get that data set going. Um in another example, we could build out recommendation systems. We've talked about how to do that using, you know, collaborative filtering or even more advanced um methodologies. Um so so that's possible. um you know so different ways to come up with what types of personalized flyers we could generate but that just is a lot more effort um than just a basic flyer so that's why it's in this category but it it could drive a lot of impact it could be really really impactful um just obviously takes more effort and longer more data collection as I said it's not trivial then in the um low impact but very doable is seasonal so that's very doable it's just driven bas bas on the time of the year. Um, but may not have as much impact as being highly personalized year round. Um, so it gets put into this category. Um, and then the low impact, low feasibility is analytics integration. Um, so it says that and the reason so what's nice is it gives you a reason for this, you know, putting it in that category. So says insights are valuable only after customers are actively using flyers. Um so low feasibility because it um requires a lot of tracking and and data gathering and pipelines. So in general like spinning up analytics is a huge effort. So it's takes a lot of time. Um but don't really need to do that right now. It's kind of low impact initially and doesn't mean like in the very very long run might you know could definitely have impact but um it's dep prioritized amongst these other things. So that's uh and so it gives you this final grid right here basic flyer generation um you know personalized flyers uh seasonal themes here in this bottom corner analytics integration bottom right corner dep prioritized okay so hopefully that makes sense I mean that's a pretty easy example with GBT but I one of the reasons to go through this practice is to see how the example and details in the prompt really really help you and I think that's something we have a good feel for especially based on when I asked you guys about your own use cases last time um I was really glad to hear that a lot of you are already using generative AI so that's pretty cool um so you probably have a good feel for this but uh just seeing it kind of again here a lot of these details if we were to take them way. Um, you know, it can go in any direction. Um, and so if we wanted to go in a specific direction, we got to provide those context deals in the prompt as much as we can, right? And realistically, it might take several back and forths to get to that point. Um, so that could be something else different than this is maybe we uh start out with just a generic grid and then we kind of have to add to that. No, I really want to use, you know, uh impact and feasibility. Um that could be a possibility, but uh it it you don't have to do as much of that iteration if you kind of include more of that up front. Um not that again, not that iteration is a bad thing. You may have to iterate. Um and there's nothing inherently wrong with that. It just takes more time. Okay. Any questions? Uh based on this um based on this demo, were you guys able to try were you a did you have access to this uh prompt? Were you able to run it through? And there's not, by the way, there's also nothing special about uh GBT. Um uh you know, you could run this through any one of the uh language models that are out there. and kind of compare and contrast what it comes up with. Um but um pretty simple example just just running through this prompt. Um and again it's it's sort of close to a oneshot example because we are we are providing details about the kind of grid that we want almost like an example. Um, you know, and maybe to take it one step further, we could have provided a previous example of like some other 2x2 grid we made for another project or another product. Um, that might be pretty valuable as well as like another example. Um, that could be something we add in here. But, um, you know, we're providing a bunch of context really about that. Um, so it's not quite a polished example that we're providing, but it's a lot of really good context about the grid rather than just saying, "Oh, just put it in a 2 x two grid." Right. Very good. So, let me um there's no other questions about that one. We can uh stop that uh practice and go on to lesson two. Pretty straightforward. Just um working with a chat GBT. Again, you know, there's times where that may have required a bunch of iteration uh to land on on the proper output that you want. And that's there's really nothing wrong with that. Um, but something you probably have a good feel for is like the better you get at including details up front, the less time you have to spend iterating. Um, which can just save you time generally. Okay, so now we're going to move to uh So, if you have um these notes and you want to pull those up, we're going to move over to lesson two, which is going to be a um lesson just full of demos. Full of lots and lots and lots of demos. Um more demos than we've probably ever done in one lesson. It's going to have a lot of them. Um mainly because this lesson's all about showing off various capabilities of generative AI. Again, a lot of them you're probably aware of, but if you're not and you haven't worked with it that much, I think it'll be a really interesting lesson to see some of those capabilities showed off uh here in our various demos. Um so, we'll work with a bunch of different tools. So, throughout the demos, um kind of like when we did the hugging face spaces, we're going to just show off a bunch of different um really cool Gen AI tools along the way. um that's kind of the focus of this lesson. So, a lot of demos, a lot of jumping back and forth between the slides and the demos that we'll do uh here. So, uh our objectives for this lesson are to really look at how generative AI is applied to a bunch of different real world examples, a bunch of real world tasks. Um we'll take a look at some of the multimodal capabilities. So this is kind of like um expanding just having text input which is usually just a prompt right into something like a uh a language model. And nowadays they can actually work with many kinds of data as input. So they can work with um various files like CSVs or spreadsheets. They can work with images. Um they can even work with audio and video. So um you know we'll explore those capabilities by doing a demo with that. Um we'll take a look at some demos with some various custom GPTs. So if you've never worked with those before, they're pretty interesting. These are um models that have been fine-tuned for specific use cases. So things like creative writing or educational research um or actually doing education tutoring. um if you've never worked with those, I think it'll be cool to to see some of those capabilities. So, we'll do that and then we'll um take a look at some workflow automation tools that um leverage Genai. So, um things that you may have heard of, Zapier, um Otter, a bunch of different tools that can uh um help us automate workflows and we'll do some hands-on with those tools, which I think will be interesting. if you've again never worked with them before um or never seen them think it'll be interesting uh to to take a look at those. So, a lot of demos um a lot of hands-on stuff today which I think will be great. Okay, so we'll the first place we'll start is expanding our horizons on some of the capabilities of chat GBT. So um I you know have um heard from you all that you use these kind of models quite a bit. So it's not surprising that they can do these kind of tasks like language translation um social media profile uh or post creation um even doing customer reviews uh sentiment analysis. Um they're all capable of doing this. And a lot of this stems from um going all the way back to transformers, right? They were uh designed to be able to handle language tasks, NLP tasks, which um really fall into these categories. Things like translation, sentiment analysis, text generation, and summarization, and on and on and on, a bunch of different language tasks. It's no surprise that that you know these uh these transformer inspired models can handle these kind of things. That's what they were designed to do. And so um we will take a look at some demos that can do each of these things just to show them off. Um and it really will be no surprise that they can handle these pretty easily. Um so we will do that. The first place we're going to start is um uh first place we're going to start is with doing a language translation. So uh we will pretend that we are um uh operating an online bookstore and we want to translate a bunch of our um snippets from our books like our our uh titles or various uh excerpts from the books um into various languages. And so if we wanted to do that quickly, we could use AI to do that, right? We don't need to um use kind of really old translation technology. Uh um we can use generative AI. It's really capable of doing language translation. So we'll take a look at that. Uh quick question on a different topic. Sorry, no worries. Uh we haven't discussed the projects for deep learning. No, we did we didn't uh I didn't go into details on those um those projects. Uh no. Um but yes, there are projects available for that course. Um and it's the same as usual where you want to do one of them to get the certificate for that course. Um, so yes, I would uh try one of them and um you know if we want we can talk about them uh later on if uh maybe on Monday um we can go back and discuss some of those if we want. Um but I have I didn't cover the specifics of those projects. Uh but there are options to uh multiple project options that revolve around um you know deep learning. Uh so would give those a try. It's this it's the usual. You only have to submit one of them to get the credit. but yes uh there are project options. So I would take a look at them uh see which one seems doable and uh give that a try. Yeah, I think if I remember there's, you know, there's a few options. One doing um computer vision, um one doing a more traditional uh modeling um like doing a binary classification. Um, so there's uh and then another one doing I think another kind of uh image recognition. Um, so there's a few options that are sort of aligned to projects that or examples that we've done before. Um, so so there'll be things close to kind of examples that we've worked on for sure. So inside of lesson two, there's a bunch of demo documents, uh, quite a few of them, and we're going to try our best to work through all of them today. Um, and so this is the uh book translation example. Um, and the so what we're going to do is just um read through the scenario and then really um kind of uh utilize this prompt uh into the uh utilize this prompt into chat GBT to kind of show off its language translation capability. Um so the background scenario is what I had said before. where we are an online bookstore that offers a bunch of books in multiple languages to cater to a global audience to expand our reach and provide readers with books in their preferred language. Um, we're going to use Jad GBT for real-time book translation. So our goal is to take some excerpts from the book and these excerpts may actually be you know advertised on our website for instance as you know pieces or snippets from the actual text of the book but um when someone logs in from Japan or from India or from Europe somewhere we want the uh maybe we want this to display in kind of the local language uh that is in that region whenever ever someone accesses our website. That could be a use case, right? Is we could use something like gener generative AI to do that for us um to do that translation. So um so our prompt here is to uh just to show off that capability, translate the following book excerpt into multiple languages to make it accessible to a broader audience. So we have this uh excerpt which is from uh uh this book journey through the echoes capture captivating exploration of ancient civilization. So this is just describing um a little bit about the book and it's an excerpt from it and we will um uh we will um so we'll copy this scenario and our prompt and our excerpt and we will paste this in. And again, a lot of background information, a lot of context provided, but mainly what we're looking to do is generate this translation of this text here into multiple languages. So, let's paste that in. So, here's our excerpt from the book, and here it is. So, what we're expecting to get are multiple translations. Now, something that you may have noticed is we're not providing specific languages that we're hoping it translates into. If you want to see specific ones, feel free to tweak this and add those in. Um, so where it says multiple languages, you can um specify explicitly which ones you're targeting. in a real application, you know, we may grab that from their um IP address location or something like that and then um supply that location uh to kind of infer what language we should be producing um and and um then use that kind of in real time. That would be a more appropriate way to to do that in real time. But we are um just seeing an example of how this can do translation pretty effectively. So, we'll see which languages it does, but um you know, if you want to see specific ones, feel free to edit that. Okay. So, uh here are the translations. So, it chose to do Spanish and here's the full Spanish translation of that exact uh excerpt. So, it's just giving us that. It chose to do French. Um so, it has that translation. chose to do German. So, it has that one. Um, chose to do Chinese, chose to do Arabic, chose to do Hindi. And then it even has a follow-up. It says we can tailor it for specific regions or even like maybe dialects of those or uh completely new languages if we want. So, this is just the sample of the ones that it gave us, but um obviously we could tailor this to the use case. Uh but this is just showing off you know these language models are highly capable at doing translation very very good at doing um and and also notice that it's zero shot right it doesn't need any examples it's just using the model capability it's just using the inherent model capability to do the translation um we don't need to provide any examples at all um it can just do that kind of out of the because it's been trained to do that. So there is that. Um all right. So let's see. So we saw a language translation use case. I think we're fairly confident that it can handle that. Um let's see a LinkedIn profile creation use case. So now what we're going to do is um ask Generative AI, namely Chad GBT. Of course, feel free to experiment. Like if you use other language models and you want to try it out in those, there's no requirement to use JGBT. So, feel free to use those other ones. I think that's totally fine. But, um, what we're going to do in this situation is imagine we're preparing to create our profile. Um, we want to create a personal brand. So, we will use Generative AI to help us do that. Um, so let's go over to that demo. Um, let me open that one up, which is this one. So hopefully you guys have this one. As I said, we're going to be doing a lot of them. Um, all the ones in lesson two. Um, so this one is to again help us see how JIGBT can help us create a uh LinkedIn profile. Um, and what we're going to do is uh take a look at utilizing the information from a sample resume. So this is kind of a fake resume here. Um, but the idea is we can provide that as background data as context, right? And what the model will do is kind of extract important details from that to help generate our profile. So we're using this uh resume here. So we have all of this resume details with uh basic information, objective, education, skills. This is all this is all fake obviously, but um uh bunch of uh certifications, professional affiliations, publications, etc. Um, and then what we're going to do is ask the model to create a LinkedIn headline that showcases expertise or sorry, experience, qualifications, and passion using industry relevant keywords. Then we want to write a summary of around 150 words that highlights personal uh professional journey and unique skills. In the experience section, provide detailed entries that showcase my contributions. Suggest. Now, this is an important one for LinkedIn in particular. Um, and again, this may be something we would follow up with. It's not that we have to provide this in this initial prompt, but it's it's good in that it produces a lot of content for us to get up and running with LinkedIn. You know, suggest content for starting conversation with new connections, focusing on recent industry news or or their achievements within a 100 words. So beyond just filling out the information that would be in our profile, let's have some conversation starters um focusing on news or their achievements and then help me craft a LinkedIn post. So we actually want to build a a relevant post for our profile. Um sharing insights on a specific topic with relevant SEO hashtags. So that's search engine optimization hashtags. offer advice on optimizing my profile to attract recruiters and enhance professional connections. So, a lot of information in this prompt and what I would say is there's probably multiple approaches to doing this. As I said, um you know, this kind of does everything in one shot um of including all this detail within this prompt. Do we have to do it that way? No. You know, we could break this down into multiple prompts that we follow up with, which would probably be more realistic of starting with, you know, writing out most of the details in the profile. Then maybe later down the road, we come back and ask for content to do uh conversations. Then maybe at some point down the road, we want to generate a post in which we come back and ask to do the post. And we may have some more context around what do we want to do the post about because this this just says any specific topic but um maybe we have one in mind you know at some point in the future and then um maybe you know we can ask for advice whenever it doesn't have to be within this uh prompt itself. So but there you know it's fine we can add all this detail in there up front and save ourselves a little bit of time in the future. Um but in all likelihood we'd iterate on this uh probably with multiple prompts would be realistic. Um but what's important is especially for this we need that resume information. Um yes exactly right. So exactly what you said wouldn't we need to upload our resume? Yes. Exactly right. So you read you read my mind is like yes if we're going to do this in particular we need to um provide that resumeé data with that prompt and that's what we do here. So it's actually um we could upload it as a file here. It's just here's the text of this fake resume. Um so we have it and we would include this this text um with our prompt. So, we're going to include all of that um with this prompt. Um but you're exactly right. We would definitely want to include that. Okay. So, here is our resume details. It's all up front. Um and then we have our prompt. So, create a LinkedIn headline that showcases my experience blah blah blah blah blah. Everything we just talked about. And so, there it is. and we so we could um upload it as a file. That's fine, too. Um but as I said, it's kind of just here in the text format and we're just pasting that in first. Either way is okay. Um you know, it really it would just extract that from the file anyways. Uh so either way works. So let's send that along. So uh when we do that it um you know uh generates a bunch of stuff for us. So we have all those elements that we asked for in the original prompt are broken down into this response. So we have our LinkedIn headline. Um so we have an AI research scientist um generative AI LM. So, so this is just um a bunch of different headlines uh um that we could put at the top of our profile or kind of in the headliner of our profile which are mainly kind of keywords like industry keywords um and and skills in some way PyTorch, TensorFlow etc. So we have that then we have the summary. So about 150 words which is what we asked for. Um so we have about me um you know a IML specialist strong focus on generative models large language uh models multimodal etc and then um it has some details about uh what we're passionate about um etc. So every all the details we include in our prompt show up here as they should, right? Um so that's good. And it even used information from that resume like currently working as AI research scientist um specializing in disabled diffusion blah blah blah blah. Um crossodal embeddings um these kind of things. Uh so um there we go with the uh summary experience. So uh if we were filling out our profile, wanted to extract a lot of the same information from our uh um that would be in our resume. Basically expanding on that a little bit here um in the experience. So, uh, this maybe builds out those bullet points a little bit more detailed than what they were in the original, uh, resume, but here they are. Um, so you see the different positions. So, this person worked at Dind and OpenAI. Pretty impressive. Pretty impressive. John Doe. Uh, and then we have our conversation starter. So, here's kind of a template. So, what it does is it builds out a template for us that we can use to start conversations. So, you know, it says, "Hi, uh, blank name. I saw your recent work on blank. Really impressive, especially your approach to blank." Um, so it gives us a nice little template we can use to, um, you know, start a new conversation with a potential new connection. Um, so that's pretty interesting. Then we have our first post, right? So again, this is something the post I would say I would say this starter like this and the post um you know are uh this and the post are likely things we could have followed up with. We don't really need the resume to do that. um you know, and we don't really need the uh um we don't really need the starters as part of our profile creation. They're just additional helpful things if we're trying to build up our LinkedIn, you know, brand, so to speak. Um they're they're just helpful things that we uh might need. So, here's here's a first post. Um and we have kind of a title of it. Um and then some uh details that are right around 200 words. And then importantly, it provides a bunch of hashtags. Now, these are ones that can help it get discovered, right? So, there's search engine optimized uh search engine optimized to um you know, uh be able to make it discoverable by the the everyone else's feed who's interested in these various hashtags. Then has some pro profile optimization tips. So, use a banner image. Um, uh, use a skill section, recommendations, activity, SEO tip, uh, repeat key terms across sections, um, etc. So, um, just some extra tips. Again, one of those things that we don't necessarily need I if we're just trying to build out our profile using the resume. Um, that would have been kind of these first three items. The rest of them are just kind of interesting additions to help us get up and running on LinkedIn potentially. But really cool, right? So really cool usage of generative AI. If you haven't ever used generative AI to help you with these kind of things, um maybe this gives you some inspiration to try that. Um so so maybe give that a try for helping you craft some posts or conversation starters um or uh enhancing your profile uh about me or summary those kind of things. You can provide your own resume try to get that um synthesized into this nice uh about me summary. It's it's an interesting use case. Uh, can we send the link to extract? Uh, I think you can as long as it is um public like you don't have it um private um I think that probably will work. Although I've never tried that. Maybe give that a try. Maybe sending along your own profile link um to see if it'll extract. That's a that's a good question. I'm not I think it should work as long as it is uh not private like it like it's you know you can click on it it as from a public view and you can view it but I've never tried that personally so I can't say if it does work or not so spells in strange languages Yeah, that's a good Yeah, good headliner or about me. Yeah. Okay. Any questions on this one? I think Yeah, I think the link is an interesting one. I I uh haven't tried it. Um but um be would be interesting to to try that out. Okay, so we have one more uh for you here of just basic GBT ones. Um I want to do sentiment analysis. So this is one that we've looked at before I think in terms of language models are generally good at this task as an NLP task, right? Uh sentiment analysis. Um so in this one we're going to take a look at um customer feedback um and see how that can kind of be um obviously automated by a language model that can extract based on reading that text extract the sentiment um of the feedback and if it's positive negative which is a very useful thing um for analyzing what kind of feedback you're getting if you're like an online store online brand um kind of uh they they do this they do this all the time where they read like social media posts or or reviews like Google reviews or Yelp reviews and they um will basically summarize it or uh provide some sort of sentiment analysis of it like how much of it is positive how much of it is negative and you can easily do that um through generative you don't actually have to read through every single one and apply a label manually kind of manually calculate that you wouldn't want to do that take too long especially with lots of post, lots of feedback. Um, so we can do sentiment analysis with um a language model like JGBT. So let's um practice that one. So let me go to demo 3 and I will open that one. Okay, so hopefully you have this one. Demo three, this is sentiment analysis of user reviews. Um, so what we're going to do uh is take a look at some various uh um uh reviews from customers that are kind of fake. I mean, these are just fake reviews. uh but um we're going to uh take a look at categorizing them into these various sentiments, right? So like a positive, neutral, negative sentiment. Um and that can give us an analysis of uh you know what the review is kind of expressing almost like a summary of it. We also have a mixed sentiment um contains elements of both positive and negative. Questioning sentiment expresses curiosity. Appreciative sentiment shows gratitude, appreciation or thankfulness towards the subject. So this is a great example by the way of um fshot prompting essentially because rather than just relying on the model itself to do the sentiment analysis which it certainly can do but when you rely on just the model it's generally just going to categorize things into positive negative neutral right it's going to be into what the model has been trained on and most of the time it's been trained against those three because that's the most basic type of sentiment analysis we can do. This actually has examples of what we're looking for across several different sentiments. So, positive, negative, neutral. We have a a description of what we're looking for to label something positive, right? A description of what negative and on and on and on. We even have mixed questioning and appreciative. Um, and mix is a little different than neutral. Neutral just means kind of lack of any strong emotion one way or the other. But mixed means you do have elements of both positive and negative. So um you know it's mixed in that sense. Um so these are kind of additional sentiments that we're providing examples for of you know this is what we're looking for for these uh sentiments. So we have that. Um, if we didn't provide that, and I encourage you to try this, like if we didn't provide that and we just asked to do sentiment analysis of these reviews, um, it would likely just do positive, negative, neutral. It would stick with those basics. But here we have some additional, uh, sentiments that we are describing, uh, um, here. So let's do these sentiments and then let's do these reviews. So we have one review, two reviews, three, four reviews. So we're going to categorize four reviews into those various uh sentiments. So categorize the product the provided review based on the following input. So we have positive, negative, neutral, mixed, questioning uh and um uh appreciative. And then here's our review. So we should end up with four different sentiments, one for each one of these reviews. Um let's see what it comes up with. So number one it has it thinks should be mixed sentiment. So number one was a mixed sentiment. If we read number one let's see if that makes sense. So the number one was I recently watched eternal visions and it was decent overall. The story line had an interesting mix of mystery and romance, but it left somewhat somewhat but it felt somewhat predictable at times. The acting was good and the cast delivering solid performances um though nothing particularly stood out. The visuals are nice with some beautiful scenes, but they weren't as captivating as I had hoped. The score was fine, fitting the film's atmosphere, but it didn't leave a lasting impression. In summary, it was an okay, it might be worth a watch, but it didn't quite live up to my expectations. So, I think that's a fair sentiment, right? It said it was mixed and it even explains why it contains some positive things but also some negative things there too. Um, I think that makes sense. So, this gets categorized as a mixed sentiment. Um, number one, I think that's a pretty good categorization. Number two gets categorized as positive. So, if we take a look at that one, um, it says, "I just finished reading The Night and Gale by Kristen Hannah, and I was blown away. Storytelling was captivating. Weaving together historical events, deeply relatable characters. Found myself completely immersed in the lives of uh, Ven and Isabelle unable to put the book down. Emotional depth and raw honesty left a lasting impact on me. Highly recommend." So, very, very positive. I think that's fair. That seems like a fair. And again, it shows strongly favorable languages with no notable negatives. So, it's a positive sentiment um for that review. Now, number three gets a negative. Um let's see why. So, I said, I had high hopes for the new restaurant, but unfortunately, my experience was disappointing. Ambiance was nice, but the service was incredibly slow. Staff seemed indifferent. Food was mediocre at best, lacking flavor and creativity. Additionally, prices are exorbitant for what we received. Overall, I wouldn't recommend dining. So, yeah, that seems pretty negative to me. Um, pretty unfavorable experience. So, yes, this seems like a negative sentiment. Seems reasonable. And then lastly, number four is a mix. The latest updated app left me a mixed feelings. I mean, we literally use the word mixed feelings in there. That's not a shock that we end up with mixed sentiment, right? While I appreciate the new features and smoother interface, I'm frustrated with the increased battery consumption and the occasional bugs that disrupt my experience. So, it is a mixed sentiment there. I think that makes sense. Appreciation for improvements, but also frustrations. So, that leads to a mixed sentiment. So again the example I would give you give to you guys is in in reality um a lot of people are using generative AI nowadays to do this automated sentiment analysis. So they they typically have applications set up to route the social media posts that are about them. um you can kind of scrape that in real time um and then feed them through the generative AI to kind of generate a sentiment and you can quickly tell um how many posts you have in the last 10 minutes, last hour, last day are positive, negative, neutral, um whatever sentiment and you can gather up those those statistics relatively easil easily. Um which is useful data, right? useful data about about your posts, about your your brand. Um, it's relatively easy to gather uh from the reviews that people are leaving or from social media posts, whatever kind of text is being generated about your brand um or product. Um, it's pretty easy to pass through generative AI and get those sentiments. You know, in the past was a much harder problem. Usually, you'd have to have someone in the loop actually tallying that up. Or you'd have to have a very basic NLP model that would try to do this. That'd be more statistically based. Um maybe looking like a rules model that's looking for particular keywords um to label it as positive, negative, neutral. Um but now we have these really powerful language models that can do that pretty effectively. Okay, let me pause there. Any questions about this uh any questions about this demo? So, as I said, we wanted to take a look at some of the uh multimodal capabilities of of uh Chad GBT, but it's really true to any of these language models that are out there today. really has these um uh capabilities that go beyond just um working with text, right? Working with prompts that are primarily text based. Um so these days they can work with many different kinds of data especially as inputs or even creating outputs that are uh images. So, we've seen there's various different image models that are out there, mainly diffusionbased um that can generate an image, but they can also utilize imagery in the input. So, there's a lot of really cool technology involved there. Um, probably some of the most interesting is the ability to um analyze an image in particular and kind of uh map it into the same context space as text which is um known as kind of like a a crossodal or multimodal embedding. So it's a very interesting idea is to take multiple types of data as the input but understand the context across all of it in order to generate some type of text output let's say. Um and so that's a really strong amazing capability of these models is is that that uh ability to work with multiple kinds of data as input. Um, and oftent times like that may be uh a use case, right? You may have an image that you want to do something with or use as context. You may also have um some type of document like we just talked about the resume that could have been in a PDF or a word doc or something like that, right? Didn't have to be just a text. And so um uh what we should realize with this too is these um systems and they really are I would characterize them really as systems like a chat GBT are um you know the model in what you see is really just the tip of the iceberg. There's a whole system below the surface. So we just interact with um we just interact with this tip of the iceberg but really underneath there's this massive um system there to extract text from documents um so to extract text from documents to do um thinking algorithms um to do uh uh image generation there's all kinds of um it's almost like its own software application It really is. At the heart of it though is of course the LLM. So the LLM plays a really critical role in generating the content, but there's a lot of machinery there that um helps process this multimodal data or um you know uh basically do um prompt engineering almost automatically with thinking strategies uh to help um generate the best output possible. Um, but of course we don't really see all of that under the hood. We just enter our prompts here and then everything else happens, right? And and just works. Um, but there's a lot going on under the uh under the surface there. So, um, that being said, we're going to do uh some demos, of course. um uh we'll do those. But uh the interesting thing is this the the additional data just helps us provide context right at the end of the day that extra um extra data can provide a lot of context. But there's also a lot of interesting things we can do to that data. We can uh use it to build like hashtags or SEO tags or captioning for imagery. We can ask questions about the imagery specifically. Um we can do uh we can extract information from video or audio content. Um which is interesting. Uh we can also ask for image responses uh like build an image that matches this prompt. Um so these models have a lot of capability of working with multiple types of data. And this is a this is a relatively new advancement. So it hasn't always been that way. Um, but it's a really big capability that's been unlocked really in the past couple of years. I would say it's a really it's pretty strong within the last couple years. Okay. So, we're going to do a couple demos uh that will explore these capabilities. So um we want to uh dive into how G Chbt can work with image data in particular um and analyze an image. So just see an example of that. So let's go over to our demos. Back to those. This is going to be demo for exploring multimodal capabilities. Okay. So hopefully you can see that. So exploring multimodal capabilities of chatbt. Um so what we're going to do is uh we're going to look at two different use cases. We're going to look at analyzing a map. So we're going to have a picture of the United States and ask different questions about it. And then we're also going to um see how it can help generate an image. So generate a visualization of a futuristic park. So uh following a prompt. So we're going to have a prompt to help us generate an image. So just show off some of those capabilities to work with image data whether that's in the input or generating it as part of the output. Um so uh we will take a look at um this first prompt. So what we want to do is analyze the image provided which is a map featuring various geographical locations including city, states and natural landmarks. Identify the country and suggest notable tourist attractions visible on the map. Describe these places briefly and explain their significance or appeal to visitors focusing on areas that are typically of interest uh etc. Right? So when we look at this image, of course we know what this is. It's obvious, but the computer what's what's fascinating is like a machine wouldn't necessarily know what that is. So, the ability to kind of analyze this um image and determine it's United States and here's a bunch of interesting things about it um that that uh you should visit or be interested in um is really fascinating. So, just the ability to work with that image um and we can include that in our prompt. Now, we could also download that image and upload it as a file into the prompt. that's also, you know, possible. Um, but we're just going to include it as the um as part of the prompt here, which it should, Let me double check. So, it should, if I copy that, but yeah, in the P maybe because it's in the Word doc, it's not in the PDF, but this works. Just creating a screenshot and pasting that. So, we say analyze the image provided. So, it gets it right. It says this image shows a map of the United States including it states, major cities, surrounding geographic features like oceans and neighboring countries. Here are some notable tourist attractions and regions visible on the map along with why they draw visitors. So we have New York City, Los Angeles, Chicago, Washington DC, natural parks and landscapes, coastal and regional highlights and then why this map matters for travelers. So really interesting. It was able to take even though I So this was my screenshot which was basically the original um original image. It's able to analyze this understand it's you know the United States there um and then take a look at the uh analyzing what about this this map is interesting. So um based on that we have these various cities landmarks um and then a little bit of description about uh um what what's there and it even follows up and says you know we can suggest a travel itinerary if we want um based on this map. So pretty interesting use case. Um, let me ask you guys, have you ever have you ever been able to work with uh images in Chad GBT or any language model? Has that has that ever worked for you guys? Any any use cases for that that you have that uh you've been able to work with imagery? Yeah, it's like I said, it's a relatively new uh feature of um these language models is be able to process imagery. that is you know a um recent advancement in language modeling is as I said the the challenge with this is often um you know we've talked about attention layers and what they do is they help us understand the sequence of text um al together to understand the context of what words matter to each other but when you have data that is not text that becomes a real challenge to understand context text. That's really hard. And so, but that but that's where the advancement has really been is to have these um uh embeddings that can map uh data like imagery into the same kind of embedding space as uh text and then try to also generate kind of context from these from a from that data. Um and and so that way you're able to work with all of this as input to the model. So not only just the text but the image as well or audio or video whatever. Um so really really amazing enhancements to again all based all inspired from the transformer but um especially attention and embeddings but um has been significantly adapted. you know, back when the transformer was created, 2017, I don't think they ever envisioned um working with these different kinds of data like image and audio and video, but we certainly can today, which is which is fascinating. Okay, so we have that example. Um now we're going to ask a a slightly different question, which is to create an image. So that example was using an image as input asking a question about it. Now we're going to prompt to actually create an image. So we are going to utilize this prompt create an image of a futuristic smart city park at twilight and we're going to have a bunch of context details about what we are looking for in that image. So this is really a detailed prompt. a lot of details captured here in what we're looking for in the image. So, we want an eco-friendly environment where technology and nature are seamlessly integrated. Solar panel walkways, bioluminescent plants, interactive digital information boards, really futuristic city, right? Atmosphere should be serene, soft lighting, sky should show transition from sunset to starry night. Background, there should be a skyline of sustainable skyscrapers with green rooftops and vertical gardens. So a lot of detail and that helps like if you want to get a very specific image all this context detail really helps and so let's see what we can build if we ask for this. So we can copy that in and we can ask to create this image. And by the way, this is going to do the diffusion kind of process. So the model it's using under the hood to do this is kind of that diffusion process which is um converting our our text details that we're talking about here into the right kind of um transformations on random noise that can um ultimately lead to the final image. So, it's it's kind of like erasing kind of like backwards um erasing noise and wiping that away to reveal what's left behind. Um almost like a scratch off uh ticket. You're scratching off something, leaving something behind. And that is the uh you know final image here. So, here is what it generates. Um and this looks pretty good. This follows a lot of the details that we asked for. So we asked for those bioluminescence the back the the background sky scrapers with that are sustainable um uh walkway um the biolum bioluminescent walkway we talked about um atmosphere should be screened with soft lighting. So a lot of those details we asked about um looks pretty good in this image. But again, there's a very sophisticated model there that stable diffusion process which is um learning how to generate pixels from text. That's essentially what we're doing, right? Learning how to go from text to pixels. Um and it's a very sophisticated model that's doing that. Do you guys get something close to this if you tried it out? So this is creating an image. Earlier we were analyzing an image but this is actually building one using the prompt. And again this is a very very new process within only within the last couple years has this really been um you know uh uh available to do is go from uh prompt to image. It's a very hard problem in general. Yeah, it's it's similar. Definitely very similar. Um but slight differences, but but yeah, very very in the same range. And and that's not too shocking. Mainly because our prompt is so specific. You know, the more vague that we left our prompt, I think that gives more avenues to be creative and be different, and we would end up with likely different images. um the the more vague our prompt was, but because we have such a specific prompt with very specific details that we're looking for, um it kind of narrows down what we're generating, right? Narrows it down a lot. Um so yeah, we end up with, you know, those very similar uh designs. Pretty cool. Okay. Any other questions on this one? So, yeah, if you've never uh generated an image before, you can uh ask it to create images. Um, by the way, all the other providers have image models, too. Google Gemini has it. Um, Anthropic has it. They all have these capabilities. Um, so it's not necessarily unique to GBT, but uh um you know GBT did have the Dolly model integrated from OpenAI early on. So it was one of the earliest that did provide the the text to image capability. Um and it's gotten better, you know, over time. So, we did that one. Now, we're going to take a look at uh um analyze customer feedback once again and uh take a look at um using it for a sales um uh e-commerce kind of uh use case and look at ways to structure um you know uh potential sales strategies and opportunities for improvement. So, kind of expanding on the sentiment analysis use case that we looked at before, but um categorizing different feedback into uh the segments and then um visualizing the results. So, I actually do a little bit of visualization uh which is where the multimodal capability comes into play. Uh which is having the best capabilities. Um, I really like OpenAI. I really think the Yeah, the CHBT one is really, really good. I think the Google one's really good, too. So, the Google um the latest like banana um yeah, I think the the banana model from Gemini is pretty good. The Nano Banana, whatever it's called, um is decent. I really think GBT is uh the the best probably because they've been doing it the longest um and have the most practice at it I think with starting with the dolly um which and they've just built off of that for a long time. Um so yeah, I think GBT is pretty pretty good at it amongst others that I've tried. Okay, let's look at demo five. Customer feedback analysis. So we are going to take a look at this scenario. So imagine we are a data analyst working for a large e-commerce company. The company has just completed a major sale event and they have collected a large amount of customer feedback through their website and app c. The company wants to understand customer feedback, identify areas of success and opportunities for improvement. So it's sort of similar to what we did before. Um we have uh um bunch of feedback data. we want to analyze it. Um, and what's interesting is rather than just doing the analysis, which we do ask to do, so given the feedback, categorize it into one of the following categories, positive, negative, neutral, and constructive this time. So, slightly different sentiments than we had before. But what's interesting is after we do that, we um ask it to do a little bit of analysis. So, after categorizing, provide a count in each category and build a graph. So build a pie chart that visualizes this sentiment analysis. So going one step beyond just just providing the sentiments, let's actually do a little bit of data visualization of that. Um which is interesting. so we have uh we have to grab the uh um feedback uh analysis which there's a lot of comments. I think we have to get that from the uh um from the CSV. Let me see if I have that quickly available. I don't think I do actually. Does anyone have that customer feedback from the data sets? It might be inside of the uh data sets. I think it's a CSV. Okay. So, what we're going to do is we're going to do uh we're going to upload that into here. So, we are going to paste our prompt. We have our scenario and our prompt we're going to paste in here. And then we're going to add that file. So, customer feedback CSV. We're going to add that file. So, uh we uploaded that file. So, thanks again for providing that. So, we uploaded that file and then we have our full prompt and we're going to um ask it to uh analyze this feedback data um analyze the data um and categorize each flag into one of the following categories. So, this CSV prov it basically has all that text. It has a lot of them which is more realistic of a problem. Like I said earlier, you know, if we had product reviews that were coming in rapidly or we scraped from our app and we dumped into kind of a file or a database, this would be more realistic of a problem. We have a lot of reviews we want to categorize at once and do some type of analysis on um and or maybe it was in more of like a stream from um from social media. That's another use case. Um, but let's send that along and see what it has to say. So, crunching on that. But you want to make sure when you do this, you add the CSV in there as the as part of the input data, right? So that we can uh um analyze the the all the reviews that are in that spreadsheet, which is there's a lot of them. Um so it gives us the count and the visualization. So it looks like it um basically produced uh um positive, negative, uh constructive or neutral. Uh and we ended up with 136 neutral, 131 positive, 10 negative, and three constructive. Um, and here is the uh graph of that. How many Yeah. How many neutrals did you get? 190. Wow, that's a lot more. So, it must have took some of those positives and thought they were neutral. That's possible. So yeah, it's it's possible to get slightly different interpretations of the reviews and, you know, have different numbers there. That's totally possible. Wow, completely different results. What models are you guys using? Same ones. GPT Gemini. Okay. So, it had a lot of uh Gemini seem to lean heavily on the positive. Any more in the free? Oh, I got you. Overused it. Yeah. No, that's a great question. So, could we add anything to the prompt? So, the data could should be stable in any model. Uh, that's a great question and yes, we could. So, I think what we could do is provide examples. So if we provided explicit examples on these reviews are positive, these reviews are neutral, these reviews are negative, that could give some guidance into what an example looks like. Um remember that um remember that you know these are probabilistic. So that's why we're seeing some instability and we get slightly different answers even across the same model. Um you know we get slightly different numbers and we may get that even on different runs like if I refresh this and reran it I might get uh I might get different numbers. you know, so what could really help is providing explicit examples on what we think is neutral, what we what we deem to be positive, what we deem to be neutral, what we deem to be negative. And those examples can guide the basis for predicting all the others and give us a little bit more stability. So that's where like the fot prompting really comes into handy there. So I think that would be something to do. But yeah, same same data giving you many different ones. I'd be interested to try I might actually try Gemini as well. Let's see. So, let's let's try it. Let me see if I get a different answer as well. Um, if we were doing anything new, we won't have the example. Um, that's where we have to come up with our examples. Yeah. Like unfort you're right like if we're doing anything new we we likely won't have an example but that's where we could spend some time creating our own labels and that might be helpful. So um but yes we would have to come up with those. We'd have to do some extra work to come up with the labels for anything new. Agree with that. Other than examples um I'm not sure. I don't think there is uh really because these models are inherently probabilistic. So there's no guarantee you get a stable answer every single time and certainly across models. Um there's no guarantee. So unfortunately I don't think so. The I mean the other the other aspect of this is if we had data that was way more decisive one way or the other. It could be that there's some data that's kind of on the fence where it has some mixed. So the the other thing we could do is change our labels to maybe there'd be something more appropriate like mixed rather than neutral that has some combination of positive negative. Um, so if we just change up our labels a little bit, maybe that could help categorize things more clearly and distinctly into those labels. So using the right labels um for the problem, like based on um maybe some preliminary analysis that we do of just a sample of the reviews that might be helpful to do as well. A little bit more stability. Yeah, I think the examples will definitely be the best. Um, now the other, by the way, another step to this is if we were always doing like let's say we were always doing the sentiment analysis, we could fine-tune the model. So, we could permanently alter the model against the examples to um, you know, longterm have better um, stability and better ability to to use this use case. But that's a lot of that could potentially be a good amount of work to fine-tune the model um you know that's that's a little bit more heavy lifting um and and takes more resources etc. Oh ch says if you want I can refine the model to better detect constructive insights that's usually where the most business value is hiding but then data analysis limit has been reached. Oh, how does this say? Yeah, that's sort of what I was saying is like we could fine-tune the model to end up with a uh yeah, to end up with a free or sorry to end up with an altered model that works good against this use case. We could do that. If we give data and everything related to data to GT write a prompt for that and then we specify the data should be stable in all types of AI models that that might be you could give it a try that might be okay I don't know if there's any guarantee that will change your results necessarily but so yeah let's let's see if we follow up with this and say um I want to improve the categorization especially so that it's stable across all language models. So this is a good so this is a good suggestion. So yeah like what you were saying um we can force a structure. So you are a strict classifier. So we give a role to the model. You're a strict classifier. We want to determine a sentiment, positive, negative, neutral, and if it contains a suggestion, yes or no. Sentiment, suggestion, final category. Um, suggestions override sentiment classify as constructive. So, anything with with a suggestion becomes constructive. And then be conservative only label positive, negative if clearly expressed. So, we have a little bit of rules here. And then we output uh everything into this kind of dictionary. Um and then we uh um have ex it also suggests having examples right a small labeled data set of examples. So it's kind of combining um having a a stricter prompt with some rules but also having examples. It's a lot of details here but essentially those two things can make it more uh steady is what it says. Okay. But yeah, really good thought to ask the AI what does it think it should do to make it more stable? And that's kind of what it says. Use examples and maybe make it more structured with some rules in the prompt. So rather than just say, "Hey, classified the feedback," we have a bit of a more structured prompt to attack Okay, perfect. Great. So that's uh that is that demo. Was there any question? Any other further questions? Really good ones. Any other further questions on that one? So, let's uh continue along and explore some GPTs. So, this is going to be interesting where we want to um take a look at some fine-tuned models. We were just talking about fine-tuned models a little bit um and we're going to explore um some of those fine-tuned model capabilities. Now, remember the fine-tuned models are ones that have been trained against specific data for specific use cases. So they've been updated from the basic um existing level of tragt and trained against additional data to be used for particular uses. And so we're going to try out some of those. Um now this table here by the way is a little dated. Obviously we are on the five series today. So this was um created a little over a year ago. So back then we only had the four series emerging at the time. The default for the free version of JHBT was the 3.5. Um much more limited especially its context availability much fewer tokens. So where we are today is you know much larger than this even. and of course we can use it for many different complex tasks as we're finding out. Um but this is a comparison of kind of the older uh models to see where it has evolved to. And of course where we are today is just in the 5 series just really really massive model that uh has gotten faster has basically every one of these areas has improved. It's gotten larger has more context length has gotten faster to process for sure. Um still doing really complex use cases. Um you know uh can do thinking. So it's optimized for for various things like thinking or it has it alternative version for efficiency. Um and certainly you can do multimodal. Um so but this is a comparison to kind of the older one. You can see like all the way back in 3.5 it couldn't actually work with images, right? It wasn't even multimodal. Only in the 4 series did that get introduced. Um and of course we're in the five now. Um so again um thinking about use cases like uh summarizing a research paper um you know uh if you want to go into detail you probably want to use the thinking models that we have today um which was aligned with more of the 4 series back then or if you want to do if you want it to be really really fast you want to use the the fast model which does less thinking um and doesn't handle the context quite as deeply. Um uh so um all that to say is like you can use different variations on the models today depending on what kind of tasks you're doing. So if you're uh doing something really really complex and you need all that context, then the thinking model can be very good. If you just care about speed, you want an answer really really fast, you can just use the fast model, too. But this this chart is a little dated admittedly. Obviously, we're on uh much newer models today uh than these guys. But you can see some different examples from like the same prompt. Explain polymorphism to me in a line. Um you can see there's some differences across the uh um across the models. They don't all produce the same exact words. They produce slightly different words. Um, and we see that even from like platform to platform or or I should say model to model, right? Like they're not always guaranteed to produce the same exact thing. And it may take different time to do that depending if you're using one model versus another. Um, so again, no guarantee in the exact same output, no guarantee in the exact same timing. They can take shorter, faster, it just depends. this leads me to the discussion of custom GPTs which are again these are fine-tuned that have been um uh produced by the community and are available to everyone to use. So people have taken the time to use their own data and to kind of customize the GPT um for particular use cases and they usually have um a description of what the GPT is used for um uh or what its special use case is. Um and so we'll take a look at that when we get into the demos. But there's a bunch of different GPTs out there. So there's ones for creative writing, there's ones for um programming, there's ones for productivity, ones for education research like uh um educational research like scholastic research. Um so these are just to name a few. There's a bunch out there and again these are taking the base GBT and fine-tuning it against data so that um it's been kind of specialized to that use case. it's a little more leaning towards being good at that use case uh because it's been adapted for that. Um and uh this is what a lot of people do. They take they take the basic basic GBT and fine-tune it. Um you can actually you can pay to have that done by OpenAI. They can uh you can upload your own data and they can basically fine-tune it for you and you get a customized GPT. A lot of people have done this. I've seen examples of this for a bunch of different use cases. Um, yeah, the there are some of them are I'm going to show I'm going to we're actually going to do some demos with some of the specialized ones for sure. They're they're kind of like the hugging face spaces in which they're like the community GPTs that people can use. Yeah, it's kind of like hugging face spaces in a way like anyone can use it. Yeah. So, I'm going to show you those and how to get to those um coming up in a moment. The first one we're going to look at is the write for me where they um have uh basically tailored the baseline GPT. So, they've fine-tuned it to be used for personal writing and creative writing. So, it crafts tailored content focusing on SEO audience needs um and really created creating like creative writing material. Um, so it's geared towards writing. That's what it's really good at. It's been fine-tuned to do that more so than the basic model. So, it has a bunch of data has been thrown at this that um emulates creative writing or is using creative writing sources. Um, so it's much better at content generation, editing, writing. Um, so if you have a use case for that, this would be a good G GPT to use. We're actually going to try that out. Um, we're going to do a demo where we explore the write forme GPT, which again is just a customized version of the baseline that is particularly been adapted for writing use cases. So let me go over to that. that'll be in demo six. So, content generation for eco-friendly reusable water bottles. So, we're going to take a look at that and we're going to see how to access those right something like the right for me GPT which is what we're going to use for this. So, we're going to do content creation for eco-friendly reusable water bottles um using the right forme GBT. And our goal here is to create marketing content for an eco-friendly reusable water bottle company. Obviously fictitious, but um mainly want to practice using this write forme GBT. Um, and so what we're going to do is go to chat GBT and then I'm going to show you how to navigate to the write forme uh GPT. I'll show you where those are. It's a little bit different than this demo. This is a little bit dated back to when it was a 4 series. So things were laid out a little bit differently on the website. I'll show you how to get to that now. Um, it's a little bit different, but once we find it, we're going to copy and paste this scenario and the prompt into it. So the scenario is we're preparing to launch an eco-friendly reusable water bottle u made from sustainable materials. Target audience is environmentally conscious individuals who are active and health conscious. Aim to highlight the benefits of our product. Um and here's our prompt. We want to create a compelling blog post for our uh reusable product. Reusable water bottle product. Blog post should introduce the product and its unique features. Discuss the importance etc. And this is really important context, right? The tone should be informative, engaging and persuasive, aiming to connect with healthconscious environment environmentally aware individuals. So let's grab this and then I'll show you where to find the right forme GBT and we're going to use this in the chat with that. So in order to get to the right for me GBT where you want to go is on this left side under more you want to go to apps. So underneath more go to apps. See how this apps window pop this this window pops up when I click on more. Right below codeex this pops up and I go to apps. So let's go to apps and it should bring me to this page. So once you're on this page, then you go to the top right corner where it says GPTs. Click on that. So click on GPT. So again, it's under more apps, then GPTs in this top right corner, GPTs. And this takes us to kind of the repository of custom GBTs that have been uploaded to uh the community. And this is where we want to search for the right or me. And it's the very first one. It's very very popular. Has 48 million chats. Um really really popular. Supercharge writing assistant. So, we want to access that right for me. And you can see um it's ratings. So, it has uh it's number 14 in writing, 48 million conversations. And then we we want to start a chat with this. So, we'll click on start chat and we can paste in our prompt, our scenario in our prompt. So, one thing I'll point out is, do we need to use this custom GPT to accomplish this task? No. We could have just used a regular GPT, right? That's fine. We could have just used a regular base one. That's totally fine. The hope is that these custom ones have really good capabilities geared towards creative writing and and um marketing material writing. And so it's worthwhile to use these custom ones because they're they're geared towards these use cases like essays, blog posts, any kind of creative writing that we would want to do. So um in this situation, we're hoping to create a product marketing blog post. So it's a good fit for this custom GPT. So, we send that along and it generates our uh our blog post here. So, it has a lot of detail. It says, "A better way to hydrate. Meet your new go-to water bottle. If you care about what you put in your body and your choices, your water bottle matters more than you might think. That's why we created a reusable bottle designed for people who want something cleaner, safer, and more responsible than a singleuse plastic. This isn't just a bottle. It's a small shift that adds up to a real difference. So um then a bunch of details about uh why it's worth rethinking meet the bottle design for real life with a bunch of feature details, health benefits, small change, why this matters, ready to make the switch. Um so pretty compelling blog post. And what I would encourage you to do is maybe try the same prompt with the baseline model and see how it's different. And and it's not to say that what the baseline model generates is wrong. It's just subjectively the right for me is geared towards this task. It's geared towards creative writing and so it it um has been trained against a bunch of creative writing use cases. So it can ideally perform better in these cases. But um we could see it for oursel. We could try it out on the uh just baseline model. and um see see kind of how it works…

Transcript truncated. Watch the full video for the complete content.