Saroop Bharwani: How to Leverage Generative AI for Your Work

Hello, hello everyone. I hope you're all having a great day. Thank you so much for joining us again. And if you're here for the first time, my name is Daniel. I'm a part of the FI marketing team. And if you don't know about FI, FI is essentially the world's largest proven network that helps turn ideas into fundable startups and fundable startups into global businesses. And before we really get into what the event is all about, I wanna know where you're all coming in from. So why don't you all let me know in the chat, tell me where you're coming in from. I'm based out of Pakistan. I can see that we have people from Austin, Israel, awesome, Hamburg, Lithuania, Denver, Croatia, New Orleans, Pittsburgh, India, Ethiopia, Romania. So we have a pretty international audience. Everybody, welcome to a very exciting event. I'm personally really excited about this because as a marketer, I use generative AI and it's always good to brush up on your skills and because it's such a new field, there's always so much that you can learn. Today is all about leveraging generative AI for your work and the speaker we have with us today is Saroop. Saroop is a thought leader in AI and a tech entrepreneur. He's also the co-founder of senso.ai. If you don't already know, this event is a part of a much larger bootcamp that we're doing with Saroop where he'll really go in depth about how to leverage generative AI for your organization, right? Akshay from our team will be sharing a link to the boot camp. So if you want to sign up, I would encourage you to sign up for the waitlist. Again, spots are limited, so, uh, I encourage you to sign up, right? And yeah, before we really bring Saroop on stage, uh, let me kind of go through the agenda for the event, right? Uh, Saroop will do a presentation, and then we'll move on to a Q&A portion, and then right after that we have a networking session happening. So I encourage you all to stay back for the networking session. You can talk amongst yourself, you can talk about the event, you can talk about anything and everything related to startups. The FI team will also be there to kind of speak to you, and if you have any questions for me, I'll also be there too. If you look to your right on the screen, you should have a Q&A button right under the chat. So if you have any questions for Saroop, when he's presenting or afterwards, I encourage you to click the button and ask your questions. And once we're done and we move on to the Q&A portion, I'll bring your question up on stage and you'll hear from Saroop live and in person. And if you look towards the bottom, you should have an option for emojis. So if you like what Saroop is saying, if you like what Saroop is talking about, show him some love, right? You should have a heart, a thumbs up, all that good stuff. So Akshay, could you bring Saroop on stage? And while we bring Saroop on stage, right, we want to know, we want to know something about our audience, right? Everybody's kind of really getting into ChatGPT. So we want to know, how often do you use ChatGPT? Do you use it multiple times a day, once a day, a few times a week, or you haven't signed up yet, right? Let us know where you're at with this. Okay, awesome. It seems like most of our audience, uh, is already kind of familiar with ChatGPT. Awesome. Yeah, so we have a mix, right? Essentially almost everybody knows what ChatGPT is all about. They're all using it. Uh, hopefully you'll kind of learn how to really get the best out of ChatGPT from Saroop. So yeah, Saroop, over to you. Would you want to unmute yourself?

Great, thanks Daniel. Great to be here. Thanks for having me. I'm just gonna share my screen.

Awesome. So I'll move to the back of the stage. I'll see you all again during the Q&A. Saroop, thank you for doing this.

Over to you. Perfect. How's that Daniel? Can you see that? Perfect. Great. Hey everyone, um, again, thanks for having me. Really appreciate it. I'm excited for today's session. Um, just, uh, I'll start with a little bit of background about myself. Um, the session today is on how to leverage generative AI for your work. I am Saroop Barwani. I run an organization called Senso, which for a better part of a decade has been building, deploying, uh, AI models and AI products, uh, for the financial services industries and, and now various other industries as well. And throughout the years I've been deploying these AI products to enterprise, you know, a lot of executive teams have come to me and really kind of wanted to dig in around how to really get the most out of AI. This was even before, you know, ChatGPT launched, years before that. So I launched an organization called First Principles AI, which aims to educate executives and staff in the enterprise on how to really get the most out of AI and now tools like ChatGPT and large language models. And ever since, you know, ChatGPT launched on November 30th, 2022, you know, there's been a surge of demand for this. So my objective is to really educate and prepare anyone in corporate for what's coming because it's gonna be really exciting times. So in today's session, I'm gonna talk about the upcoming bootcamp in September. There's going to be 3 days over 3 weeks. I'm going to talk about objectives and structure of that. Today's theme, I'm going to give you a real taste into an intro in generative AI. Like, I'm going to pretend that no one knows anything about it, and I'm just going to start from the basics, start from scratch. I'm going to handhold you through the basics, and this is going to lay a foundation for what's to come in September during the boot camp. So I encourage you to sign up for that because we'll be going into a lot more depth in the actual boot camp on these topics. Uh, so we're going to start with the basics, the amazing possibilities ahead of us, uh, give you a little bit of a taste into prompt engineering, a skill set that I'm sure many of you have started to dive into, and, uh, and governance, a little, uh, how corporate is really thinking about, um, embedding this type of capability in, in their workflows. Uh, and then we'll do a Q&A. I'm happy to answer any questions that you may have, so please do, um, add to the chat. Alright, so the objectives of today and the bootcamp are really one, to educate you. Every employer, every corporation I talk to, and I generally do 2 or 3 of these a week, every one of them has all of a sudden shot this up to the top of their list from a priority standpoint. They're looking to embed generative AI and large language models in their workflows, and they're looking to hire people who understand how to use these tools. So my objective with today and the boot camp is to provide you with the knowledge and skills necessary to prepare you for the future. Um, second is really kind of like share with you the amazing possibilities of this technology, but also the potential dangers. Um, any technology as a tool, uh, comes with both, you know, good and bad, and people using these tools can use it for either of those. So really, the boot camp is going to have a deep dive into some of the amazing possibilities, but the risks as well, and how to really be responsible when using this technology and bringing it to the world. And lastly, to enhance your productivity. I certainly have seen a massive surge in my productivity, and day by day as I continue to use these tools and embed it in my workflow, I am automating many of the tasks that my team and I do, and I am completely sharing this with everybody in the bootcamp. And I'm gonna show you how I do this internally and help my clients do it so that you can do the same. So the bootcamp will be focused on text generation primarily with a little bit of image generation too and multimodality. You know, one of the things I'll show you later on is what we're gonna do in the bootcamp is actually create a marketing campaign using text and image and showing you how you can really do those creative tasks in an automated fashion. We're gonna do future bootcamps as well on things like code generation or audio generation, but that is not in the scope of the bootcamp coming up in September. And yes, all of the images that you've been seeing and will see in this presentation are AI-generated, and I will completely show you and handhold you how to actually create images like this in seconds using prompt engineering and various templates that I've created, and I will share with you. All right, so in terms of boot camp structure, as I mentioned and as Daniel touched on, this is going to span 3 weeks, so September 5th to September 19th. And there's gonna be 3 days. The first one is really on large language models and prompting, really the basics. I'm gonna give you a glimpse of that today, and we're gonna be doing a deeper dive into that in day 1. I really want you to understand for the first principles foundations of this, the edge of it all, so that you can be as educated as possible when you talk about this in your corporate environment, to your friends on social media and everything. There's a lot of information flying around right now. and there's a lot of hype, but I want you to be educated and have first principles understanding of everything. That is what day 1's about. Day 2 is about rolling up our sleeves and being really practical. We're going to have live examples, and I'm going to share my repository of prompts I've developed and continue to develop with you so that you can sit in front of the screen with me and actually input, these prompts in ChatGPT to actually go through live examples, marketing examples, workflow examples, many, many examples we'll do live. So that one is going to be really practical. And lastly, really how to think about risk and governance. So in your organization, I want each one of you to be the smartest person in the room sharing the risks versus the rewards and really educating people in the organization how to responsibly deploy this so your corporation has competitive advantage. So that's the 3 days, and all of these are super, super important. So why are we all here, right? So, you know, as I said, I've been, you know, building AI products, uh, for a better part of a decade. I've been building tech products my entire life, starting when I was a kid. Um, I started writing code when I was about 7 years old. So, you know, we're here because late last year a product called ChatGPT launched. All of you have been using it, uh, everybody has been using it, and it's the fastest growing application of all time. And this has created a massive, massive surge of demand for this technology because it's actually useful now. It was not useful when I was building it for the past 10 years up until last year. All of a sudden it became useful thanks to OpenAI. Not only OpenAI, I want to give credit to Google too. The reason why ChatGPT exists is because of a paper launched at Google in 2017 called Attention Is All You Need. And Attention Is All You Need introduced the transformer architecture. I'll be talking about more of this in the bootcamp, but I just wanna understand in 2000, end of 2017, this paper is essentially what started this wave and Google released this paper completely open, right? So they created their biggest competitors, but this is good for the world. The fact this is open source is an amazing, amazing thing for the world. I wouldn't imagine a world where this was closed, where one organization like Google had access to it. It would be very, very dangerous. I encourage all of you to read this paper, and in future sessions, I go through this in depth. So we are in a platform shift now. I'm sure many of you were around for the mobile internet platform shift. That was a lot of fun. I was in it for early days building apps in the App Store and on the Android Store. I was in the early days of the internet too, you know, like I was building websites when I was in high school, but I was a little bit young back then, but mobile was there. And of course, we had the PC server era. We are in the beginning stages of the next platform shift. Shift, and this one is going to be much, much bigger. It's rooted on everything of the past. It wouldn't be here possible if it wasn't for the internet and mobile, etc. Um, but you know, we're hearing a lot about macro these days, but I want to share we're going through a simultaneous event of a new platform shift at the same time. Um, so it's a very interesting time in history. So just to set the stage of where we are in history At this point in time, my belief is in the next 2 to 5 years, the world will have access to broad-based intelligence. So in other words, intelligence will be completely democratized. Every person on the planet with a smart device will have access to agents which can act on their behalf. Imagine having a specialized group of chief of staff specifically focused on planning for us, booking our trips, educating us, treating us for illness, prioritizing meetings, advising us on life, guiding us through negotiations, and helping us prepare our meals. Um, this is all now possible, and we're at the ground stage. Builders like me are diving in, and every single day we're improving this collectively as a unit. And this is getting really exciting, but, um, but this is only going to get better. This is not going to stop. So that brings me to what is generative AI, and let's just define it in its basics, right? Generative AI is a type of AI that creates new content based on what it's learned from existing content. The process of learning from existing content is called training, and that results in the creation of a statistical model. Think of this as the new way to code. When given a prompt, generative AI uses this statistical model to predict what the expected response might be and then generates new content. And I'll get into specifically what a large language model is doing, but, you know, it's probabilistic. And you'll get different results each time, right? Um, you know, uh, image generation models predict the next pixels. Text generation models like large language models, uh, like GPT, predict the next word. So large language models, so LLMs, understand connections between text really, really well. It's trained on the vast corpus of text that we've amassed across the internet for many, many years and decades. And a large language model is a prediction engine which takes a sequence of words, your prompt, and it tries to predict the next likely word. And it does this continuously over and over again, and to a point where it seems magical to all of us. That's all it's doing. It's not conscious. It's just predicting the next word over and over again based on probability. This is an infographic I really love. It shows all of the data that GPT-3 was trained on. We're on GPT-4 right now, and there was GPT-3.5, which is still still in production. GPT-3, as you can see, was trained on a lot of the common crawl data, um, web text. Wikipedia was a big contributor to this. So all of the data we've amassed over the years has been used to train these models. And there are many types of large language models, both centralized and decentralized. On the centralized side, I'm sure everybody's familiar with OpenAI and GPT, There's another organization called Cohere, Anthropic, and of course Google that have their own language models and are rolling out their own chatbots built on top of these language models. And the open source community, decentralized, which is where I'm diving in, super exciting. There's many contributors who are pre-training LLMs and completely launching the weights and the code completely open source on GitHub. Meta just launched one called Llama 2. The other day, which is equivalent to GPT-3.5. Super powerful at the fingertips of developers. Anyone can download these and actually fine-tune them for their use cases on device, on-prem, you know, anywhere. And it's super exciting. All right, so, you know, I'm sure that, as I said, I'm sure that a lot of you were around during the mobile internet platform shift, you know, when we— when, you know, Jobs and, you know, Apple consolidated all of this hardware, the camera, GPS, and phone into one consolidated piece of hardware, you know, we knew there would be like a camera, we knew there would be a chat app, but it was much more hard to predict that there would be photo sharing and Uber and Airbnb and all of these apps that we kind of like use on the go. And this is, this is sort of happening again today. The same thing happened with electricity. Before electricity, many— we're all, you know, we weren't around during this time, but, you know, scattered energy sources was a big problem. Life was very difficult back then. And electricity, or the consolidation of power, you know, of course the light bulb was very, very, very obvious, but the EV wasn't, the chip semiconductor wasn't, the phone wasn't, maybe the fridge wasn't, or the dishwasher, right? So all of these innovations we didn't even think about And the same thing is happening now, except right now it's scattered knowledge. And what we're doing is we're consolidating this knowledge in models through weights and probabilities, and we get apps like ChatGPT, which have completely, completely caught the world by storm. What are the new use cases? I'm going to share a couple of possibilities with you. And some of these possibilities, like, we're just starting to surface because we haven't thought of the next Ubers or anything like that, but So, you know, what are the new use cases? Let's think about this. So imagine waking up one morning and being able to watch your dreams. This technology actually exists today, not at scale, but people in Japan have developed the technology where you can do an MRI and actually it'll essentially leveraging these technologies, generative AI, it will actually be able to create an image of your dreams and eventually a video of your dreams. Imagine doing that. Wouldn't that be an odd world? Would you share that with your significant other? Would you share that with your friends? I don't know, maybe or maybe not, but this technology exists today, just not at scale. Um, AI to see people through walls— this exists. This is actually a company that a few of my friends invested in very, very early. But through, uh, through basically Wi-Fi, you can actually, um, actually kind of like mimic the positioning and the movement of individuals without any camera. You know, there was a documentary that was just released that brought Wilt Chamberlain back from the dead. Imagine being able to talk to a loved one who has passed away or interact with them in a video. You know, I certainly hope that that's possible in my lifetime to talk to people that I've lost from an early age. My dad, for example. So getting into stuff that's more real today, Google Bard, I'm sure that a lot of you have been using it. Actually, you can input video and image. So video and image in there produces an output. So this is multimodality at its works. It's not just text, it's multimodal. This is an open source model that was launched called DragGAN, and it allows— it basically takes the design methodology to a completely different level. So all the designers in the room, this is something that you can actually just use today. And this, this will be democratized into all apps, Photoshop, etc. Here's something that, speaking of Photoshop, that they're producing. So it's not just kind of manipulating current images, it's producing brand new type of images that never even existed. And this is— Adobe's doing an awesome job with deploying some of this technology out. So, um, the possibilities are endless. I could probably have 100 more slides on these examples, but I just wanted to give you a taste. We're going to talk more about them in the boot camp for sure. All right, so let's dive into a little bit about what a prompt is. So prompt is, as I mentioned, is the text that goes into a language model. So it is the text that you put into the input field in ChatGPT. And prompt engineering is the art of writing text to get the language model to do what we want it to do. Remember, if you input text in one instance of ChatGPT and it produces a result, it's likely to produce a slightly different result the next time you put that exact same text in. These models are probabilistic, and, you know, essentially the art of actually crafting these messages to get what you want out of it matters a lot. And I'm going to show you some examples. This is a structure that I use on a regular basis to really set the foundation of how I generate my prompts. I'm an engineer, I'm a computer scientist, so I like structure and everything, and I like reusability. So as I develop my prompts in my prompt library, I actually reuse them on a daily basis. I template them. And this is something I'm going to teach you in the bootcamp, how to reuse prompts. It's not starting from scratch every time. It's basically developing a prompt template and being able to use that over and over again in your workflow and automating that. That's something we'll talk about in the boot camp. But elements of a prompt, let's start with context. Context is external information or additional context that can steer the model to better responses. It's defining who the model is. Instruction, a specific task or instruction you want the model to perform. That's, uh, that's going to be highlighted in blue. The input, what do you want to analyze? What do you want to summarize? What do you want to consolidate? What do you want to extract information out of? An output indicator is basically the format you want things in. So let's go through a couple of examples. So this is a prompt that I only put in, put an input and an output indicator. So discuss the benefits and limitations of utilizing generative AI, and I wanted to summarize it in, you know, a couple of bullet points, you know, no more than that, and limit it to one sentence each. So When I did this in GPT-4, it produced some pretty good results. It was pretty random, like it selected what the benefits and limitations are, and but it was a pretty good result. Can't complain with this. It got me a good output. But again, like my objective here is how do you get more out of it? So with something like this where I added in context, really defined that the model is an AI researcher at OpenAI focused on generative AI models. And, you know, the extensive research and numerous interactions they've had with it provide them the ability to actually understand, you know, the challenges and limitations. And the instruction is actually, you know, detail out the advantages and limitations of using generative AI models. Then the input is just basically what I want, right? That is, the input changes every time. And as you notice in the input, I actually defined what the categories I want were. So creativity and novelty, efficiency and scalability, and data requirements. Those are the things I'm most curious about You can really interchange that with anything you're curious about. And then in the output indicator, I defined the columns that I want and I defined what each column should really entail. And again, it did a pretty good job of doing that. Cut and paste this table into my deck and, you know, away I go. So the sky's the limit in terms of really kind of like tweaking the nuances of these prompts to, to get the most out of it. So I encourage you to try that. Of course, as many of you are familiar, these models are really, really capable of doing a lot more than just, you know, extraction, summarization, conversation, etc. You know, these models are really good at reasoning, and one of the things I'm going to talk about and show in the course is how to generate your own prompts using these systems. Because what we don't wanna do is be sitting there like type, spending like 10 minutes typing a prompt every time we wanna do something. What we wanna do is get these machines to leverage, you know, their capabilities to develop the prompts for us. And I'm gonna show you how to do that through templates and how to create a prompt repository that you can use on a daily basis and get better every single day, get better every single minute you use this. And I'm gonna be talking a little bit about autonomous agents and how to enable autonomous agents in your workflow. These are user— these are LLM-generated agents that do work for you. That's going to come in the course as well. Those are some of the advanced things that we won't talk about today. All right, so one example I'm going to give, a technique that I use quite often, is called expert guidance. What happens is when you actually allow these models to think, break down the steps in which they do things, you can get better results. So I often review recommendations on, you know, my website, on my apps and my workflow, and I just find it really tedious. So what I've done is I've plugged in, you know, my API with all my recommendations coming in into large language models. And I'm going to show you an example of what I get the language models to do and flag for me. So this is a review. It's a 3-star review. And these are the ones I find really difficult to analyze. It's the ones where it's kind of like, yeah, he seemed happy at the beginning and then he seemed disappointed at the end. I'm like, I don't know what to make of this. All right, so without expert guidance, I basically just said, hey, analyze the emotional state and viewpoint of the customer in the provided example. The input is just the input that is literally API from like the vast amount of recommendations I get on a daily basis. And then, you know, provide a succinct sentiment analysis. And again, does a decent job. The sentiment expressed by the client in conversation is mixed. On one hand, it's trust and satisfaction. On the other hand, it's stress and uncertainty. So again, not bad. But me, as, you know, an engineer and a data scientist, I like things to be very structured, and I like to store information. I like to use information in the future to ensure that my next workflow is better. And that's something I encourage each one of you to think about So with this one, you know, the context is basically a customer has recently purchased a product on Amazon. I try to keep that generic to the type of recommendation and then really providing context that its job is to really kind of understand the customer sentiment. This blue is where the expert guidance comes in. So I said, perform a sentiment analysis on the following excerpt on Amazon. That is the input that it's analyzing, the same input. And this is the expert guidance. So number 1, it's identify the key sentences in the review and express the emotion and sentiment. Analyze each sentence separately to identify the underlying emotion. Summarize the overall sentiment of the customer on the emotions, because I want it to be really specific. I did a repeat. So the, the what I want in the table is repeated here and here. And do not provide a final judgment, right? These are all things that you can take and just reuse in your own prompts. And as you can see, what it did is it broke down the sentence into very, very small chunks, identified an emotion associated with that particular part of the sentence and the intensity of that emotion. And then it provided me a nice summary at the end, right? So I actually use this identified emotion, intensity emotion. I attach that to this particular recommendation and store it. As tags so that in the future I can provide— I can actually identify, automate, automatically identify similar recommendations related to this category. That helps me with my workflow. But again, this provides a much, much better breakdown of what exactly the customer was frustrated about. This data that I've created over here also allows you to train future sentiment models. So again, this is just something that just to keep in mind as you get more advanced. All right, one of the things I'm going to be doing in the boot camp is I'm going to be taking a template like this and a brief from like a mock manager at an organization, and we're going to be able to— we're going to be filling in the image and the text of this in literally seconds through tools that are available to each of you. We're going to be filling out the text using, you know, tools like GPT and we're gonna be filling out the images using tools like Midjourney and various other things. And I'm gonna hand you hold through the development of the prompts and how to really input your brand guidelines. So your colors and your styles and get the most out of it. So these, these are just a couple of examples I did live with a credit union that I worked with about 2 weeks ago, and they were amazed at these results. They're actually using some of the examples in their current marketing campaigns. So, you know, all of this is good. We know that, you know, GPT and GPT-4 and all these tools are getting smarter and better and everything. And, you know, I know this group of people is here because you want to learn, right? You want to learn. And, you know, I really wouldn't— you know, my advice to you is I wouldn't worry about all the, you know, doom and gloom stories out there. I would embrace this and try to be the best at, you know, adopting these capabilities, because it's, it's you that adopt these capabilities and roll up your sleeves in like the very first inning, that's going to really reap the benefits out of this. You know, I would encourage you not to worry, be very optimistic about the future and just dive in. The elephant in the room though is a risk. And, you know, a lot of these organizations that I come across, like that I go and educate without even kind of knowing me, is they decide to block it, right? So I don't know, it'd be interesting to know how many in the audience, like your organizations have blocked it. But when I go speak with an executive team or a board team at an organization, you know, I basically kind of make this argument and I, by the end of it, I convinced them why blocking isn't the best thing. So this sort of matrix is like, you know, blocking versus unblocking, the pros and cons. So sure, blocking prevents, you know, data exposure within your organization. But the lack of visibility on how your employees are using this out of the organization is an even bigger risk. And I flag that, especially with this being so accessible on like iOS and just like being able to pull it up. So whereas unblocking, sure, yes, people in your organization can, you know, put information into ChatGPT and, you know, Samsung had a breach. Someone put proprietary information into ChatGPT and that information was exposed. But it also encourages proactive governance strategy to mitigate that downside risk. And one of the things I educate these companies on is how is your data exposed, where are the risks, and what can you do to mitigate that risk? So I walk them through a framework I've developed. It's a risk-reward framework that really kind of like educates, you know, corporate and even like individuals like yourself that are kind of using this on a day-to-day. How to kind of get the high rewards and the low risk. My argument is that risk, the blocking, it has a low risk and a high risk and a low reward because of course, you know, you're not using the product internally and your employees are probably using it externally. So that lack of visibility is creating the biggest risk of all. I educate, you know, really these, these organizations how to get up here. And that's one of the things we're going to be discussing on day 3. In the bootcamp. So just a couple of key takeaways, and then I'd love to go into Q&A and networking. Remember, it's still day one. We're in the very first inning. I don't even think we're in the first inning yet. I think we're in just the dugout, just preparing for the game, and everyone is still trying to figure this stuff out. And again, I speak to like 2 or 3 executive teams a week on this, and they're still trying to just wrap their heads around this.

This.

In the meantime, each one of you is here today, um, and, you know, hopefully you sign up to the boot camp, um, here today to learn and be the smartest person in your organization. That's, uh, that, that's, that's why you're here. Uh, the prepared mind, right? Um, this is, this is the beginning of a platform shift that's going to just kind of exponentially increase year after year. Um, and the boot camp is really designed to help you upskill yourself in preparation for the age of AI. And every employer, um, out there is prioritizing hiring with these skills. And lastly, you know, key takeaways are increasing productivity. Like, a lot of the things I talked about in this session are how to leverage this technology, automate it in your workflows, um, really not one-off usage. It's how do you create like a prompt library and prompt repository and actually attach these into your workflows so you could just click a button or make a request, um, whether it be a chat or your voice. And that's going to become more, more possible in the very near future to really, really feel like, you know, a magician or a wizard in your job where you're just getting stuff done just by asking and talking. And I think that's the dream that we're all looking to get to. So with that, I thank you for spending, you know, the last half an hour with me, and I am happy to answer any questions. I really look forward to kind of teaching you more in the boot camp over those 3 days. It's going to be phenomenal. So, uh, Daniel and team, thank you so much, and, uh, really happy to kind of spend the next 20 minutes answering any questions.

Awesome, awesome. Thank you so much, Saroop. I feel like, uh, I kind of learned so much just by sitting through that. Even though you, you know, when you use generative AI, you feel like you know the basics. It seems like it's just, you know, prompt engineering, and you type in a couple of words and you get a result. But like, there's so much to it, and this is what, you know, this is the type of content, this is the type of knowledge that you need to kind of realize, uh, like the rewards of, um, AI, right? And like you said, we're right, we're in the first inning, right? Like, who knows where it kind of goes, right? Uh, we have a couple of questions. I want to be mindful of everybody's time here. We also have a networking portion afterwards as well, so I'm just going to go through some questions. Uh, let me just end the screen sharing. Okay, awesome. Right, so Roman asks, what are the main risks for corporations when it comes to using ChatGPT? I know you kind of— you'll be going over it in the boot camp as well, and you kind of glossed over it right now too, but maybe a couple of things that you feel like are kind of important or on top of the mind right now in terms of this.

Yep, sure. Um, look, in context of leveraging tools like ChatGPT in the workplace, it's one of the biggest topics that's being discussed is, um, data protection data privacy. So inputting data into ChatGPT and that being exposed to OpenAI and enabling them to leverage that for pre-training and inference retraining, which happens much more frequently, that is the biggest thing that on, on these corporates' minds. And, and, you know, I talk about this in the boot camp, that there's a few ways around that, um, you know, and, and organizations really have to kind of be thoughtful and purposeful in instilling the— instilling ways to mitigate that, right? Um, for example, I'll just give you one example: using the API OpenAI's terms are they store the data for 30 days for misuse purposes and they purge the data. Whereas for ChatGPT, if you have a setting called history and training on, which is the default, you're basically— any data you input there is basically giving them permission to use that data for pre-training and for inference training.

Right, awesome, awesome. So we have a couple of questions in the chat about kind of the potential of AI, right? Because we have a ton of people who are also startup founders, right? And as somebody who's been really in the AI field for quite a while, right, in terms of potential for startups, where do you kind of see this going, right? We've seen a lot of generative AI content creation tools, right? Like what might be next in your opinion?

I mean, so this is a general, general purpose technology. So the more I actually learn about this, and you know, there's an infinite amount of learning, the more use cases I uncover. But the most common ones today are retrieval tasks. So if you think about it, and this is the one we're focused on with Senso. If you think about our day-to-day, right, um, you know, in retrieving information from our knowledge base— so many of us use Google Drive or SharePoint— think about the friction involved in that. You know, you got to open up your browser, you gotta, you know, log into your cloud environment assuming you remember your password, you got to dig through a bunch of folders, you got to, you know, shift through so much documents and information in those documents to get to the piece of information you want to need. Now the possibility of just being able to type a question or say a question and retrieve that, the possibilities are endless, right? Because there's so much that can be unlocked in an organization once you increase the bandwidth of information retrieval. And I would just think about that as a fundamental and think about how you can unlock use cases in your organization or, you know, any organization out there through some sort of SaaS product. That is on the information retrieval side. Then you have, you know, comparing documents in the legal world, comparing documents is is, you know, a huge, huge opportunity. You have marketing, as you touched on, Daniel, um, being able to do creative tasks like creating marketing campaigns, image, text, multimodal, uh, just based on personal information on, on the client or, you know, what products they do and don't have with the organization. I've seen that use case quite a bit. Those are just a few of many, many ones that we'll talk about in the boot camp.

Awesome. So at the pace that ChatGPT and AI in general is going, right, I think there is some concerns about the quality of output people are getting from AI. And Andreas here kind of asked the question, I recently saw data that ChatGPT-4 effectiveness might be worsening over the last few months. Does that sound right to you? How does one handle that?

Yeah, look, we don't really have transparency around the safeguards, the safety layer that OpenAI is putting on GPT-4, and they're constantly changing that. As I said, inference, retraining, and that safety layer is something that they do internally. I think in terms of mitigating this, look, there's only so much we can do by leveraging a closed model, right? We can fine-tune it to our use case. We can do some sort of like very, very like slight like tweaks to how the output's displayed through prompt engineering. But I think that this is where my team certainly is playing around with these open source models, most recently Llama 2, which was just launched a couple of days ago, because then you can kind of like use modes of reinforcement learning to actually, based on human feedback, you can actually start to actually fine-tune the results to the particular use case, which, you know, I think that these, I don't think there's gonna be kind of like of just a few models. I think there's going to be thousands and, you know, tens of thousands of these models everywhere. And I think that on enterprise, like, these open source models are going to start to be really useful when actually getting the better quality of results for your specific use case versus, like, relying on GPT-4's, you know, trust and safety layer and how they manipulate that or tweak that to be a general purpose technology, which is good. There are going to be general ones like that. That is good. But the fine-tuned ones are going to be really important as well, right?

Awesome, awesome. So kind of just to wrap things up, right, um, what I really wanted to know is, are there any AI tools that you recommend for somebody, uh, who wants to, you know, increase their productivity outside ChatGPT or, uh, like the very obvious ones?

I mean, look, there's, there's new tools coming out every single day that are specialized. Um, and look, I mean, I generally like use GPT-4 on, for the most part, for general tasks. Uh, and I've really kind of like just increased my skill in prompt engineering. I leverage the API, uh, a lot because the API allows you to go from interacting with an interface on a one-time task to automating a workflow, uh, where you have like common information coming in that you just want the output on in an automated fashion. So I think that the API is another thing that you should leverage. But there are tons of tools for deck creation out there, for, you know, marketing content, for sales automation, and everything coming out every single day. And there are websites, you know, Daniel, um, I don't have it in front of me right now, but there are websites that you can go to where you can kind of see like the latest tools coming out, um, and there's emails that I can— maybe what I'll do, Daniel, is I'll send you a couple of those you can include in the notes.

Awesome, awesome. That sounds good. So there you have it, folks. Thank you so much, Saroop, for taking out the time to do this. It's been super informative. Guys, if you love what Saroop was talking about, if you feel like this really added to your knowledge, show him some love. We have the emojis right at the bottom. You're getting thumbs up, hearts. Yeah, awesome, awesome. The folks loved it. Saroop, thank you so much. Do you have any parting words? Would you want to maybe share some wisdom before we kind of end things.

Just, just thanks for having me, and I encourage everyone to continue sort of like getting your feet wet and really getting practical about this. Really, practice, you know, makes perfect. I think that there's an endless amount of best practices that are unfolding here, and I would just encourage you to not worry about, you know, losing your job or anything like that. Just get good at using this as quickly as possible. And you will be fine for the years to come. And that's why we've developed this bootcamp alongside FI, because, you know, we feel this is a skill set everybody should, should know. So, you know, thanks again and happy to— looking forward to the bootcamp.

Awesome, awesome. Thank you everyone for joining us. The link to the bootcamp is in the chat. I encourage you all to please stick around for the networking event if you have any questions. About what Saroop's been talking about, the bootcamp, or FI in general, I'll be here in the networking lounge along with other people from the FI team. Or if you just want to talk amongst yourselves, talk. Thank you everyone. I hope you all have a great day. We'll see you in the next one.