How-To Automatically Research Anyone (Advanced)

It. Sam. Okay. Hello, Everybody. Oh, boy. 100 people. A lot of people want to see us automate things away. Let's go. Okay, so warm welcome to another stream in our Black Friday special edition. A lot of streams, a lot of tutorials, a lot of AI talk, all of that good stuff. Welcome. We're going to hang out and do some productive stuff today. Okay, so let me start things out by saying thank you so much for 300k. Holy shit. Can I swear on you, I don't think so. Seriously, all of you made this possible and this is a week of sort of like giving back. That's how I see it. So we're gonna bring a lot of the things that were requested by you in a live format. What we're gonna do today is we're gonna edit the automation that I showed you how to do on the channel a few weeks ago that basically uses a mix of perplexity and chatgpt to scrape the web and scrape LinkedIn, scrape a LinkedIn profile and then runs a prompt on top of it. What we're going to do today, we're going to extend it. We're going to extend it with custom data. Okay, so the concrete plan for today is the following. First things first, I'm going to welcome you. That's what I'm doing right now. Secondly, we're going to look at the automation and what it does in the extended version. Also, we included some error handling. There was an issue with it throwing a no access error sometimes. We improved it a little bit. We'll talk about that in a second here. Thirdly, we're going to go ahead and we're going to rebuild the new automation not from scratch, but from the blueprint points that I shared that I will share with you here today. There's a link in the description below and we're going to add the ability for you to add your own custom data. So for example, if you want to research somebody who's in LinkedIn, but you also have their CV that they sent in, you're going to be able to add this CV to the automation and all of a sudden your report will include the data that you included. I hope that makes sense. So you're basically going to be able to customize this with whatever you have on the person, whatever information. Maybe it's just a sentence, maybe it's a full cv, maybe it's more, maybe it's a Wikipedia page, Whatever it might be going to show you how to do that. I'm going to show you how to Implement that step by step. As soon as we do that, we're going to take some questions from you guys and we're going to show you some potential next steps that you could be doing here or that you could be doing within make. Because within the community we use it a bunch to teach various automations. We use it ourselves in multiple ways. And for that I'll have Michal from the community coming on here to answer some of your questions. Also, he's already hanging out in the chat and in the background. So if you have any questions that pertain to automating things, he's one of our automation experts in the community that teach it and that work with it. He also architected this automation right here. So, you know, if you have any questions, this is a fantastic opportunity to ask them. We'll do our best to answer your questions. If we don't have an answer, we can get back to you in the next stream on Friday. So that's the rough plan. And then I basically, I'm not on a schedule today. I'm sort of like open end. So as long as this is fun, as long as I have a little bit of energy left inside of me and I think there's quite a bit of it, we can hang out here and we can build different things. So we're going to get started with the step by step of making this automation, extending this automation. What am I talking about? Basically I'll just switch it over to my desktop here. I'm talking about this one. Here's my camera. I'm talking about this concrete automation that I showed you a few weeks ago. We extended it with a loop here that runs it multiple times if it gets no access. And we also added extra steps where it now considers extra data. Okay, I'll show you all of the step by steps of this in a sense here. Don't worry, no matter what level you're at, this stream is for you. I think everybody's going to be able to get something from it. And then in the end I want to do two things. One, I want to play around a little bit. Okay, so I had this, I have this idea of implementing a custom prompt in here and I want to customize the report a little bit and I want to show you what that would look like from a prompt engineering perspective. We could also do things that you guys come up with in chat. I think that's going to be quite fun. We can work in tandem and then in the end of the stream, towards the end, or maybe in the Second half. I want to show you something that we've been building over the past. I think it's three months at this point because we had so many iterations. It's super cool. It includes some very complex automations in the background. We're not going to focus on that, but it's basically one. Okay, I'm not exaggerating here. It's probably my favorite way to use AI as a content creator or as maybe like a public personality that I've discovered over this entire year. And we basically finished the whole thing. And I'm going to share it, it's completely free. I'm going to share it in here and then we can talk about how it works on the backend and stuff. I don't want to give too much away. It's going to be cool. Okay, so if you're into using LLMs, if you're into using LLMs and automated workflows that you don't have to touch, if you're into. If you're in the business world and you care about, you know, gathering information and gaining a competitive advantage on your competitors or co workers or whatever, I think all of those target audiences, if any, if any one of those hits you, I think we're going to have a great time today and we're going to look at some practical, practical things to kick things off. Okay, so first we're going to start. Very practical. I also wanna, I wanna hit two more points. First of all, this is one of five streams this week, counting, counting also next Monday. So I can briefly show you the schedule we have in place here. This is the Wednesday stream. How to automatically research anyone, the advanced version because we're customizing things and doing Q and A and stuff. And then there's also three more streams happening, one on Friday. That's again gonna be a very fun one. There's gonna be a lot of. So first one was like a bird's eye view. This is going to be maybe a bit more technical and focused on the automation aspect. Friday is going to be really like more prompt engineering and fun stuff and unique things we did with AI. And I have a few specials prepared there for you. And then on Saturday we're going to have some guests and we're going to be discussing the future of ChatGPT 2025. And then next Monday we're going to be evaluating the results of the challenge that I issued on Monday. Basically the challenge asks you, like, how did AI influence your life? And the first prize winner there gets $1,000. Second price 500, third price 100. So if you haven't seen that, you can check that out. That's Monday's stream. There's just a link in the description. Basically, I think in two minutes you can catch up on the entire challenge and you can see how AI changed their life, or you can submit your own story. Okay, so now that we have that covered and you know all about the streams, I think we can get to the practical part here which is going to be this automation. Let me just check on, check in on chat here. Guacs assemble that. Okay, so let's see any questions? If any questions come up, you can always share them. We have multiple mods from the AI Advantage team in the chat here. What's up, Matthias? What's up, Philip? With the Mickey Mouse profile picture? These are all technical people, so if you have technical questions, you can throw them into chat here. And also Michal is around that architected this automation and hey, okay, okay, okay, this is looking great. Okay, we have Ryan Gosling in the chat here too, so that's good. Hey, what's up, doc rock? Yeah, chatgpt birthday party. So again, thanks for the 300k and also thanks to make for making an amazing application. I want to point out that yes, they did sponsor us to do content on them, but they basically reached out to do a sponsored segment. I was like, hold up, what if we do multiple full videos? Like we legitimately teach with your product, we use your product, like heck, we have an active paid subscription to your stuff. It's great for beginners. Let's just do more together. And they were like, yep, we would love that. So here we are doing more with them. But genuinely, they're the best entry level software for no code automations because you can create a free account and try everything. I'll show you today. And everything just works. You can try it out. If it works for you, then you can upgrade for $10 and get 10,000 actions a month. There's a link in the description below if you want to try it out. And with that being said, let's get into the practical part. How about that omnipotent media? We'll talk about Sora and everything. We can talk in the second half. Okay, now I want to keep it practical and now I want to focus on the schedule here and then I'm happy to discuss whatever, whatever questions or stuff you guys might have. So with that being said, let me go into screen share mode and let me go into teacher mode a little bit more. This was this was the intro. Now let's transition into actually making this happen. So what's going on here? And let's catch up everybody who might not have seen the dedicated YouTube video showing how to make this happen for yourself. This is not going to be a comprehensive step by step, going from the first to the last step. That's why the YouTube video exists. Spent a lot of time crafting that, spend a lot of time editing that. Everything there is in place for you to get to a point where we'll be continuing in this stream. That's why it's called advanced. Okay, so it doesn't start at zero, but it starts at the point where the YouTube video leaves off. So where the YouTube video left off was essentially this version of the automation. Yeah, it's this. This is where the YouTube video left off. And it's very straightforward, I think, you know, learning by showing it is the easiest. So if you have a LinkedIn link on someone, for example, here's my LinkedIn profile, then you can create a new line in this database. You paste the link in here, you say initialized, and that's all you do. Now this automation will go in and actually run multiple perplexity searches, take all the data, process it through ChatGPT and create a report on that person, on whatever you might need. Okay, so that's basically what this thing does in action. I can show it to you in action. I guess in this case I would run it manually because it's set up to run every 15 minutes. So I would have to sit here for 15 minutes and wait till it runs. I'll just say run once right here and you'll see all of this kind of go through the motions here. This is the extended version already. What is the difference between the first version and the extended version? The extended version we'll be building here in a second here. The difference is mainly this up here where it takes in an extra field and then uses that here in the ChatGPT steps. In the ChatGPT step, the extra field is anything you put into extra information. And one more thing changed and that actually it showed up right now. So there's one error that we weren't able to fully iron out of this. We got closer to making it work every time. But sometimes when you run this, I would say at this point, one out of three times, maybe one out of four times, it will throw no access. So instead of switching to done and doing all the work, it just goes to no access. This is a problem with perplexity and if anybody knows exactly how to iron this out and how to remove this, let us know. But just to be clear, this is the Perplexity API that we use here. And sometimes it throws a no access error, so in case it does that, you can just run it again. Plus we can talk with Michael later on on what you could potentially do to make changes here so that it doesn't happen. I don't know, one out of three or four times, but even less. Okay, so basically I hit, oh yeah, I have to switch this to initialized and then I can run it once. The error fixing we implemented basically right here is it loops over it, I think three or four times. And if it gets no access, it just tries again, tries again, tries again, and that most of the time fixes it. Sometimes it still gets stuck on no access, as I just showed you. Okay, so let me see. Did it just happen two times in a row? Okay, that's very unlucky. Okay, let's give this a final spin. Initialized running and then you can see it go through the different motions here. Yeah, that's the thing with Live Stream. So this. Let's see how it goes. Let's see how it goes. Okay, so that's the basic basis of this. Makes sense. You put a LinkedIn profile link into your notion, or you can also alter it to use Google Sheets or whatever, and then automatically uses all these AI tools in tandem to create a report on the person. Yeah, now it's running. Okay, so let's get into the step by step. I just want this to have like, you know, worked once so I can show you the result and then we'll get to the step by step of adding the extra data and what kind of extra data you could add. I think this is actually the most valuable part of it because everybody can kind of like look at the LinkedIn and run it through ChatGPT. But I think in the extension, that's where a lot of the power of it hides. Because you might have some info or some data that is not publicly available. Right. Maybe you have a CV that you received, or maybe you have just some personal knowledge. You could type it in there. Maybe you have a PDF, maybe you have a call transcript that you can run for analysis prompt, whatever it might be. Adding extra data. Powerful. How does a Perplexity Pro account affect this, if at all? Miles? I don't think it does, except that you receive $5 of free credits, I think, with the API. So this obviously uses the Perplexity API here and with the Perplexity Pro plan, you get $5 off the API. So that's what I have right now. I don't even use more a month. And, and yeah, there you go. So I think let's just reset this because I've been. I have to multitask here. I'm basically running the stream, reading chats, and also trying to make this work. So I'm going to focus in on the step by step here in a second. But we're just going to do a new line. I think this is going to be simplest, say initialized and then we're going to remove this last one and then all you do is this and then you. And then you run the automation once and hopefully it works this time. Oh, okay. Something must have changed. Okay, one second, one second. So, Michal, you're in chat. Any. Any idea why? Why? We've been, we've been testing this over the past few hours and it ran like all the times and now that I go live, something seemed to have changed. So maybe okay either way adding custom data. So. Okay, so next up I'm going to show you these two steps and how to alter this. I just really want this to run now. Just really want this to run now. Initialized play. Maybe we can use the old version or maybe we can build it from scratch and I'll show you on that. How about that? I found I have to go into the Perplexity step and resave it. Yeah, now it worked. It worked like a charm, you see. So there you go. But we'll rebuild it from scratch and we can troubleshoot and change things afterwards. So you can see right now it's in the second, in the fourth Perplexity search step here and it's working it out. It's in progress. So this is how it's supposed to go. I don't know. I guess you'll just have to trust my words then on three out of four runs, this thing works. Yeah. And it takes a little bit. Okay. So it takes about, like about a minute for the thing to complete because it does pull a lot of data together here. And also some of the APIs are rather slow. You know, perplexity isn't instant. There you go. It ran the chatgpt step and now it's in the last steps of formatting it and adding it to the notion. And I think right now we should be able to see it like streaming in here. So it's adding the different pieces of text. Right. Like this is fully handoff, just automated. All comes in Here, but this is why we're doing the live. Okay. We want to show the reality of it because yeah, with edited videos you kind of just, if something goes wrong, you can just edit it out. Right? Like here we're going to handle everything live, step by step. So, you know, might take a little more time, but we'll get there. And there you go. It's finished. Yeah, it switched to done. And here I have the full CIA report on Igor Pagani. And I did this by also altering the prompt, which we'll also do in the end. So glad that work out. And yeah, you can ask your chats in chat multiple people answering it there. Do we approach these tools from the backend using APIs? Yes, 100% big fish. So here we have different connections set up with the perplexity API, with the ChatGPT API and also with the Notion database. I showed how to set up those connections in the original video. So let's get to the step by step finally. Okay, this took maybe 3, 4 minutes longer than I intended, but glad we got it to work. What we're going to do here is a step by step on how to get to work. Get this to work from scratch. So what you're going to do is you're just going to open up a new scenario. Okay. And then in here you're going to import a blueprint. What blueprint are you going to import? If you go into the description of this video, you should have access to this folder that I included. There you're going to see one make blueprint researcher with today's date. That's one blueprint. That's the one we'll be working with right now. Or if you want to skip the step by step that I'll be doing over the next 10, 15 minutes, you can just go straight to you straight to the second one. That one has the added extra info steps in it already. But if you want to follow along, or if you, if you just want the state of it as it is right now, you can get the first one. Our endpoint will be the extra info one. Okay, so let me know if that works. I believe these files should be publicly available now. Let's get to it. Okay, so I'll choose a file, I will take the original one, the researcher one, the starting point, and I will save it. Voila. And instantly you get to the spot. Okay, so you have this. So as I talked in the video, as I talked about in the video, you can import this in a second on a free account and you're basically set except of the connections in the video. I show you exactly how to establish the different connections. So we'll just go through this step by step. I already established the connections, but I have to like reconnect these things. So I'll start with the perplexity nodes. So basically this, my Perplexity Connection 2 is the newest one that I set up. And just. There you go. You just have to do this, you know, once you establish the connection. So establish them in advance. As I said, if you haven't, if you have never established a connection in make, then just go check out the tutorial video. There's. It literally includes a step by step on what URLs to go to and what to copy, where you don't need any technical knowledge to do this. Okay? This is very, very straightforward. That's the beauty of make. Okay, so we do this, this, then we go over here. We also pick the connection, we say, okay, I got this one right? Okay, I have to concentrate a little bit. So we make it through smoothly here. My Perplexity connection two, okay, and right there, excuse me, right there we have all the perplexity connections set up. So now that we have the perplexity connections, we can go to our ChatGPT connections next. So there's one node here, OpenAI Connection 2. Perfect. Just wait until okay becomes purple. Say okay. And then in this one we also do it. And this is one of the beauties of make. This is one of the reasons. And like, honestly, this is one of the reasons we love to teach with it because it saves all the presets. There's some other no code automation softwares out there where if you provide a blueprint, it doesn't save all the prompt presets. So then you would have extra steps of having to, you know, set all of this up, linking with the variables and putting the prompts in here. This is all just preset. Like this is all I needed to do, right? I gave you the blueprint you imported with one click and then you just have to enable the connection so it links to your OpenAI account and not mine, right? So there you go. OpenAI linked, perplexity linked, leaving one more thing. Okay, that's the notion step. So instead of notion, what we want to do is we want to link it to this database. I called it Igor's Autoresearcher for you. In the video, I provided you a template that was called DB Researcher template. So you would be looking for this or whatever you renamed it to in my case, I named it Igor's Autoresearcher. So I'm going to just take that name. And here in all the Notion nodes, I'm going to Igor's Autoresearcher too. And here as I showed you in the video, right again, we're just repeating what was in the video partially. I click search and here in the query I just search for it and Notion will find the database by itself and link the database id. I don't have to do any fancy, whatever parsing of the URL and the database id. I just search for it. I say okay, and here I say okay again. Next one, we do the same thing and we're just going to repeat this step a few times. I'm going to say okay here and say OK again. And then here's. There's four more. So pick my connection, search for my database. Okay. And as you can tell, this is basically the process of transferring this automation from my account to your account. Okay, so obviously I had this set up already, but I wanted to show you the step by step. So you can literally see which buttons to click as you're going through this. There's three more, one right here. Auto Researcher tool. Okay, Very nice. There's two more. This is funny because, you know, if we were recording this it would be like, hey, I'll see you in a second. As soon as I connected all of them. And here in the live, we actually go step by step. It's different, but I like it. Actually, I think I prefer. Even though things can go wrong, I really enjoy live streams because you guys are here, you know, the chat is here and I get to. Yeah, we get to do this together. It's much more fun. Excellent. So, you know, quick double check. This is Linked. This is linked. This is linked. This is linked. This is linked. This is linked. Just to make sure I've got all of them. And that's it. Okay, we linked everything up. So there you go. So now this thing is like fully functional. Basically all I need to do is I can say save once and I can turn this on to run every 15 minutes, change the intervals. I'm just going to keep this at every 15 minutes. And if I wanted to, I could also rename it. And I'm going to do one thing here and this is just for me, so you don't have to copy this, but I'm going to turn off the. I'm going to save the changes here and I'm going to turn off the other scenario. Right, because now I created a duplicate of this. Right. I showed you the first demo that was already set up. I'm going to turn that one off. So we're only going to use this version two that we just worked with. Excellent. Okay, so this is fully functional, right? I have the Notion Linked, I have the Perplexity API linked, I have the ChatGPT API linked and everything is place in place. So now I could just close this and just disregard this completely and only use my Notion database. So on a day to day, if I want somebody want to research somebody, I would take their LinkedIn profile like so, just the URL. Okay, so just copy the LinkedIn URL and I would put it in here and then guess and then if I would put it in here like so and then I just switch this to initialized. Okay. And then within 15 minutes this automation will check in the background and see, hey, is there anything new inside of this Notion database? Is there a new link that has been initialized? If yes, let me run the full automation that I just showed you. Okay, so we're not going to wait for 15 minutes here. I mean unless you want, unless you want to. We can kind of like just play music and hang out for 15 minutes until this runs by itself. No, I think it's better to just hit this button here and we can run it manually once. Okay, so let's see if it works. This is the new version, version two. And yeah, as you can see it's got pasted, it got the info, everything is running. This is completely hands off. Let me read the chat meanwhile and yeah, Okay, Okay, it's still running. Yeah. So thanks so much to the mods for answering some of these technical questions here in chat. It's much appreciated. Also Michael, maybe you could, maybe you could look at. What is Perplexity providing? Why use it over Serpent? That's a good question. This compares to standard Perplexity Toyo. Yeah, so it's completed. So I can just briefly show you here. But yeah, it's completed. And here is the standard reports on Igor Pagani based on all the LinkedIn data. Now if you want the formatting you can just, yeah, copy and then Command Shift V, paste it back in and you have the Markdown formatting tool. And there you go, custom reports on me with Perplexity. So Voyo, to address your question, yeah this basically uses fine tune of Llama that Perplexity uses. They also give you the choice on their side and it runs various search prompts in the process. We can look at those later and we'll be doing some prompt engineering here to alter this. Okay. But basically now this ran and it gave us this custom report. Now the next step here, as I talked about, is going to be extending this. So here we get to the interesting part, in my opinion. Here we get to the part where we're going to be adding some extra steps to add your custom data. And what custom data do we add? We'll look at that later on. Let's now look at the technical part of implementing this. All right, So first things first. Where do I have this? Yeah, see my screen? See everything? So first things first, we're going to be adding two nodes. Okay. So we want to add them right here in the beginning. But basically you can add nodes here with this flow control and these tool buttons. Okay. This is where some of them are hidden. So we're going to add one Iterator up here and basically this can run loops for any more technical people. I'm going to try to keep this super non technical friendly because I think that's the whole point of this. You can import blueprints and I have to deal with variables and what kind of loops you need. But we're going to add one Iterator and we're going to add one aggregator right here, like so. Okay. And then what we're gonna do is we're gonna, we're gonna connect this. My bad. Nope, It auto connects like so. So as you can see, we have these two extra steps here on top that we connected in here. Okay. Now we still have to fill these out, but basically what this is going to do is the following. This Iterator is going to look at my text field in the auto researcher database. So you can see this extra column here called extra information. I can put anything in here. Okay. Any text I can put in here and it's gonna look at it and then it's gonna. I don't know how technical I should go, but it's gonna like iterate over every line of the text and it's gonna compile the different lines that are split by line breaks into a string and that string it will pass back into the ChatGPT summarization step here in a second. In simpler terms, it's just going to take whatever you put into that field and it's going to feed it to ChatGPT when it creates your summary later on on top of all the Internet research. Okay, I hope that makes sense. So what are we going to do exactly? We're going to modify this Iterator so I'M going to click on this and I'm going to give it the extra information from the notion. How am I going to do that? I'm just going to click in this field, okay. And then I can basically scroll down and inside of here I can give it this extra information right here. Extra information, See that? So this takes the extra information column here from the notion, right? And it selects it and now it can iterate over it. Okay. Basically I say, okay, here and this one I will go to the 55 step. I will take everything this iterator created and I will set this to plain text. Okay? So this one will turn all of the different lines the iterator collected into something that we can later use on as one unit. Okay? So these two steps, very simple but very powerful because now we copied everything that we inserted. So you can like put out as long as it fits into the context window. You can put, you know, multi page CV into the extra information field. And now it will pick it up. But we have to do one more thing. We have to go down here, we have to go into this, this step right here that combines all the variables and all the different information that happens throughout the automation. And we're going to add, we're going to add a new piece of info which we can do like, so we'll just add a little delimiter here. And what I'm going to do here is I'm going to add all. I'm going to add the step from the previous, from the, from this, from this aggregator. Okay? So the way this works is I just got to make sure I have the number right. So you can see the aggregator was tools with the, with the number 56. So I'm going to look for that in here. Okay, it's right here. But basically the way you can do this is it always goes from the last node before what you're editing. So right now we're editing this so it goes backwards. Okay? So if I go in here again, I will want to go, I will want to go all the way backwards to the variable with the number 56. Okay? So basically I just click this. Okay, text and I believe that concludes it. This way we edited it so we can have extra context in our process. And now what happens here later on, like we don't have to edit anything anymore. And I gave you the preset with these edits that happened already. I just wanted to make sure you know how it is, you know how to, how do you say, how the cake is made. But basically now, the next ChatGPT step will not just pull in the search output and all the other context here, but it will also pull in the extra field that we, that we added to our automation here. Okay. And then ChatGPT can use all of that context to do whatever it does. So a little bit of zooming going on here. My bad. Now we're set up. Okay, we could click save scenario here. And what will happen now is that if I go to my little auto researcher and I redo the process. So basically I just add a new line. I set this to. Well, maybe not yet. Now I can go in here and add extra info. Okay, I'll show you a simple example first and then a more complex example later. So one that came to my mind immediately is you could just add extra facts about the person. Okay. So if you're creating a meeting preparation summary and you know something about them, like, hey, they just, you know, their wife just had a newborn child, then you can add that in here and it will be included in the summary. In this case, I'm going to do something ridiculous and fun because I don't know, I guess that's where my mind goes. He's the male equivalent of a cat lady with a house full of cats. He also likes to put hats on the cats. There you go. So now we have some extra info, right? Something ridiculous about me. It's not true. I don't even have cats. I actually have a bit of a cat allergy. Not too strong, but I do like those little creatures. So this info will be considered on top of what Perplexity finds and what it finds on LinkedIn. So I'm going to set this to initialized and I'm going to run this once and if I did everything correctly, then we should be good to go. Let's see. Let me look at chat here. Additional variable fields, example users. Absolutely. So this is what I'm going to do next. So this was a very simple and kind of fun one, but I'm going to show you something more practical with a CV next here. Yeah, so I think this is a pretty awesome flow and like, especially by learning how to add this extra field, you can add multiple extra fields, Right? That's why I really, really wanted to show you this live and also show you the step by step live, because you could just copy these two steps and add them a few more times and instead of this one field, you would just pick another field that you add in your notion and yeah, and off you go. Let's see if it worked. Unlucky. Unlucky. Okay, let's do all of this again and run it. But yeah, you could basically add multiple extra fields. And I think this is the awesome thing. Okay, so you could have one for extra information, but then you could have another one for, let's say CV and another one for, I don't know, gender, age, whatever. You could really customize this and you could make somebody in your company or maybe somebody in your HR department fill out all the different fields and you could create customized reports on somebody based on a database of info. Of course, if you take this further, you could directly feed that info from a CRM. Right? So like here it is set up that you manually enter the info, but you could go out and even if you take this all the way, you could have an agent that scrapes the web and performs specific actions and like a whole flow of actions and all of that. The result would need to be data being filled into this part. I know, let me try, let me try one more time. But basically you could infinitely customize this. And I think that's the awesome thing about this, especially adding this extra step. Okay, Pray for me. Okay. There's some great answers from Mikhail and chat there to anybody who has some questions. If you have further questions, we're happy to handle those too. Mikael, maybe in our chat in a second we can have. So now it worked. We can have a look at this no access thing. We can make the loop run more often and we can talk about some potential ways to troubleshoot this. But yeah, even if it's a coin flip, I think it's pretty amazing to have this at your disposal, especially with the custom data. So as you can see now it's working perfectly and it's just running, running, running. And yeah, enough we go. So I'm just not sure if I talk too much and if I did, I successfully pasted in here or was I still in editing mode and then it doesn't have the data. Nevertheless, we can, we can let it run one more time just to make sure. But like what, what happens now is it considers everything in here. Okay. That fits within the context window of like. Yeah, I wouldn't go. Because it adds a lot of context from the perplexity searches. So I would say like somewhere around you're safe if you don't go over 12 to 15,000 tokens in this field, which is a lot. But yeah, so this is done. Let's see if I copied it into it. I don't think I have. I Think, Yeah, I think I was still copying it. That's why I did it again. But okay, this one's in progress. It's running again. Let's see, let's let it work. But yeah, there you go. So now that you know the process, you can add your own data to this and let me show you a more practical use case to use this rather than talking about preferences for cats or whatever. So I can already get started in this. This runs in the background, Let's see. So I think you get the point here. Like so run it, run it one more time. Okay, so it seems to be working great now. Good. So how do you use this practically? So practically you will be in scenarios like this. For example, you will have a cv, right? And this is not even. So this is from my folder here. This is not even a written cv. This is just a JPEG of a cv. So what do you do? This automation obviously only takes text and this is where like LLMs come in. Like one of the main properties that I talk about all the time and that you're probably aware of is it's amazing at transforming one format into another, whether it's text to music in the case of all the music generators, or it's your short prompt into a blog post, essay, email. It's really good at transforming one thing into another thing based on everything it learned. That's really the essence of it. So I could give it a JPEG here and tell it transform, I always like that word this into text. And then basically it will turn this CV JPEG of a CV into text. And you could do the same thing with any document, but as soon as you turn it into text, I can then copy paste all of this text, put it into my database and do some LinkedIn research on Olivia. And as you can see here, it ran multiple times in a row. So it's a bit of a coin flip here, but as you can see, and by the way, Again, this is 100% related to perplexity. I just want to make that clear. We still like using it though. And yeah, but it worked here. So if I open this newest entry, Then it should, then it should consider all that. Okay, maybe we can pull in Michal here in a second and troubleshoot the stuff. But basically back on track. So here we have the full CV in written form, right? You can now either copy it like so, or copy the whole thing even with the intro. Doesn't really matter. Voila, copy, go into the notion. And then in a new entry I could just Paste this in here, right? And all of a sudden I have a full CV in here. And then I don't know, are going to be able to find Olivia Wilson? I don't think so. That was just a placeholder on, that was just a placeholder on X. But I could also leave out the link. Right, but in this case, what do we do? We leave out the link or we do a mix of Olivia Wilson and me. It will be a bit confused there, but I think you see the point here. You can put any useful information in here, put a link in here, and I'll just, I'll just skip the link and see how this goes. So we're here to experiment today, right? Like that's the point. Like if we run into things, if we run into questions, we'll figure them out. But yeah, basically has no link now. So I think, I think it does require a link. So this is the point. You could copy, paste CVS into here. You can copy, paste information, you know, yourself. And one thing that is like even more interesting, and this is the stuff I would like to discuss in the second half. We're going to be experimenting a little bit. You could do something like this. Okay, listen up for a second. So you could take the transcript of, for example, this live stream and then you could run some specific analysis prompts inside of ChatGPT to extract information from that live stream. Okay, so this works well with, for example, phone calls. You can pull transcripts of the phone calls and then you can create specific psychological types of psychological analysis based on what it sees in the transcript. Okay. Every single prompt you might have that is some sort of analysis prompt works like a charming combination with this workflow because there's a lot of subtext in what people say in the way they formulate their sentences. Right. I'm going to talk, we're going to talk about the concrete example of the CIA analysis prompt that went GIGA viral a few weeks ago. But basically that prompt takes somebody's profile or some knowledge about a person and creates a CIA intelligence style report on somebody, including, you know, their psychological traits and their vulnerabilities and whatever else it can pick up on from the data you give it and then you can add that as the extra info in your little researcher flow. So basically this goes as deep as you can prompt ChatGPT, I would say, you know, and it's, it's really good at analyzing different documents or transcripts, live stream transcripts, call transcripts for psychological profiles of people. And then if you add that Obviously you're going to get more relevant and more in depth information. Right. If you have psychological traits of somebody as a part of your report as opposed to just what they decide to share on their LinkedIn. So that would be my, that would be sort of my pro tip. You can, you can really extend this with some relevant data and you can really use ChatGPT to extract some relevant data from transcripts if you prompted correctly. The CIA prompt is a fantastic example of that. So we're going to be doing that in a second here. Loving that the chat is fully functioning here. Also, I would love to hear some of your ideas. If you have any concrete ideas on what you would like to add in this flow, we can do that. But right now let me go back to, let me go back to a little automation here. So let's see. So we have the extra context here. This is quite long. We can always add. So, yeah, at the cv. Okay. So I think it's time to slowly but surely transition to the playing around part because that's what I've been personally looking forward to the most. Because we can do a little bit of prompt engineering and then alter this node in here and make custom things happen in this automation. But let's see. Yeah, this is a bit frustrating live if it doesn't work every time. But hey, that's okay, that's okay. We still have a lot on the docket here. So next up we're going to talk about different automations that we use in the background and we're going to answer some of the questions you have here. So Mikael, I would bring you on here soon, if you're ready to go. You can kind of like give me a thumbs up in the, in the back room here, just let me know. But basically, next up, I want to, I want to have a little chat with Michal and explore some of the automations that we built inside of make for the AI Advantage community, the AI Advantage YouTube channel. And just a few things that we did recently with this, we used it in various creative ways. This is just one of them. Also we have one more automation flow coming up on the channel and I would love the community's feedback. Everybody here. I would love your input on what you make of that and how we could extend it. And if you want to, if it's actually something that's interesting to you. So we have a few more. Could we get this to iterate for a file containing multiple LinkedIn profiles? Absolutely, yeah, absolutely. So you could just. Yeah. So we could customize this to add multiple perplexity steps where it does the same thing to multiple LinkedIn profiles. And then you would vary this tool so it adds multiple search outputs. And then the ChatGPT step will pull in everything that's in here and it will Simply consider multiple LinkedIn profiles. Yeah, you could do that, but it probably, you know, like mixing multiple LinkedIn profiles is probably a bit messy. So this is really set up to like research one person. So that's why it would be good here if we, you know, had Olivia's CV or my cv. I don't really have a cv. You can maybe quickly generate one, but I think you get the point. So, next up, we're going to talk about some of the questions here in chat and we're going to talk about some of the other automations we have running in the background. So that should be quite fun because some of them were just like a brainchild of mine. I was like, hey, we should do this. And then we found a way to do that. And the others are something that's actually quite convenient and that I'm grateful for the fact that it's running on an everyday basis. Michal, are you ready to go? Do you want to come in? Do you want to hop in on stream? I can dial you in right here. Very simply. Let's see. Okay, so everybody, welcome. Michal. What up, my man? Everything all right?

Yeah, thanks.

Okay, this is excellent. So just a quick backstory and introduction to Michael. We actually met through the AI Advantage YouTube channel. And actually when I first published the first version of the course and started doing the bi weekly live events, I remember Mikhail was one of the regulars that was always there, always following up with thoughtful questions or suggestions on what else we could do. Then later on, when we started the community, he came in and I think he won one of the first technical challenges with a GPT. It was a challenge. It was the Nerd without Glasses challenge, if I remember right. Remember that?

Yeah, yeah, yeah, yeah.

And basically, basically the challenge was asking, was saying that, hey, okay, here's a picture of a nerd without glasses. Now find a way to prompt ChatGPT to create. To recreate this image. It's a really tricky. It's a really tricky challenge because every time you use the word nerd, it's never associated with the fact that it's. That doesn't have glasses. Right. In all popular media, every nerd has glasses. And Mikael won the challenge because he had the most innovative idea and approach that. Do you want to just Briefly share what you did there, if you remember. I don't, I'm just.

It was a long time ago. But what was very interesting the challenge was when I joined the community and I remember I created this GPT just before I joined the community and that was when I was testing the GPT to write, to generate the consistent images, trying to regenerate from the input and that was how it was trying to recreate this image. It was trying to describe the image, what it see and after that change the prompt to generate the image to get this without the glasses.

So essentially everybody else was trying to like to recraft the prompt and to figure out a way around it. And Mikhail was like let me automate this, let me build a GPT that looks at this and comes up with the prompt by itself as you shared there. So that was pretty awesome. And then over time we've just seen more and more of your work. And now recently you joined us in the AI Advantage team and you help with a lot of things that relate to automations. You already answered some pretty technical and interesting questions in here. But let me ask you something that doesn't even relate to this automation. What other flows have we recently built or what other flows have we recently worked on together? And then maybe we can have chat ask questions about them. But I think two of them I think really stand out that we recently created. Right?

Yes. One is for assigning the users to the district, the simpler one I think so because it was really depend on which the user choose the answer. So it automatically know to which district should it connect without even integrating this with any LLM. So it was really simple.

Yeah, there you go. So that's this one, I have a picture of it up. I threw it up on our faces. I hope you don't mind, I have it in the middle of the screen stream right here. But yeah, basically just, just as Michal described, basically this automation does the following. If somebody new joins the community and they fill out the onboarding form, it looks at the answers that they provided and it assigns them to one of three districts that we have in the community. We have the LLM district which is all things, you know, chatgpt and prompting and all the new releases. We have the creative district which is all things image and video creation and we have the no code district which is all about any tools that might be using LLMs under the hood or any other form of generative AI in a no code interface to do things like this. And basically depending on what answers you provide on the onboarding, this automation is like, I think we joked around, it's like the sorting hat from Harry Potter, right? It's like when you come into Hogwarts for the first time or any Harry Potter fans, fans out there, they put a sorting hat on every kid, right? And the hat basically speaks up and sorts them into one of the houses like hey Slytherin. It's like Hogwarts. Oh no, you're Hufflepuff, whatever. And we basically made the same thing happen with this automation. So before we used to do this manually, we used to review all the results of the survey. We still do, but this sorting happens automatically. So if somebody joins the community they share their preferences with us and then the automation we have here basically decides, hey, you belong in the LLM and the creative district and don't bother with no code for now. You can always join that later. But I think that's pretty amazing and it's something that we used to do like on a daily basis anytime a new member joined we kind of had to do it manually and now it just happens like this because the make automation runs 24 7. What interval did you set that one to? How often does it check?

I switched a few days ago because it also allows to run. This automation starts each time when it detects any changes in the database. So I just switch instead of checking every 10 minutes to an hour because sometimes the database maybe will be with the same person will be changed because person added his name description. But we don't want to check every time for this. So it's just running every hour and after that it's assigning to the district if it detects that the user filled the form and what's really important, switching from 10 minutes to one hour saved. It uses five times less the usage.

Ah, see so right there, a pro tip, you know if you, you could set the interval to every minute but then some actions will run every minute and it will run up your account faster. So if you set the interval to be lower, you know, for us we don't, we don't get multiple like we don't get multiple members an hour at all. So like it's totally fine if it just checks once an hour and if somebody joined an hour ago then it goes into this. In the case of this automation that we just looked at, we set it to every 15 minutes. That's the default and sort of seems to make sense if you add something you can check back in 15 minutes and you have to research on the person if you up the interval, it will just eat through your actions faster. So this is one thing that you can play with inside of make to preserve the amount of actions you have per month. They reset every month on the plan. So yeah, Carsten Lee is asking webhook, I guess, was it a direct type form integration or was it a webhook? Michal, how did, how did you bring the answers over from the form to the automation?

How did we get. We just had this form with the information with. With possible answers and specific answers response for specific districts. And that's all because the automation in the background, there's another automation that is filling the notion database. When user fills the form, when user adds any information about himself. So there's just only for assigning to the districts after the user filled the form.

Okay, I got it. So you split it up into two parts. One part is hey, the form gets transferred to the notion and this automation once it sees the data and the notion makes the decision. Okay. And then the other automation that gets the data from the Typeform. By the way, hint, hint, this is relating to the thing that I wanted to show you later. Typeform and automations. It's going to be cool if you take the data from the Typeform. How did you take it? Is there a Typeform integration or did you use a webhook or how did you get the data into. Into Notion? If you remember, you can pull it up. By the way, none of this is scripted, right? We're just kind of like chatting here and as questions come up we, we handle them.

It's hard to say right now.

Okay, no worries, no worries. Maybe you can, maybe you can just look that up meanwhile in your, in your question. But yeah, Carsten just asked. But yeah, I can tell you Typeform can throw web hooks. You can add an action at the end of every. Every either result or outcome in the end and it can trigger a web hook that you can then catch with make and yeah, so you can wonderfully link it like this. I think there's also a native Typeform integration though. But webhook is probably the most versatile way to do it. Yeah. Okay, there you go. We have one more automation right like that we wanted to chat about maybe Mikael, don't worry about it. Like whatever like I know, I know Typeform can do webhooks so so basically that's probably the Carsten. Yeah, you can totally do that. I can also show you how to do it if you really want to. Actually did it just yesterday. So the second, the second Automation that we have in place. That one is a bit more intricate and we're currently debating if we should do a tutorial on the YouTube channel for this. I would also love to hear in chat. If this is something you're interested in or you would like to see something else, let us know. But basically, can you summarize the second automation? This one is a bit more fun.

Yes. First thing in the community we also have some guides how to fine tune the GPT and one, one thing I did is just fine tuning the eager with the Igor style to, to learn the GPT to, to teach the GPT to answer in. In the ego Igor style. So it was really funny because we were testing this with the AI news you can use. So we just wanted to summarize in the Igor style the text. But after that I started this using after team meeting to summarize the content what we did from the transcription and after that also we have igor voice in 11 labs. So we just generated Igor voice with this summary in his style. And right now the AI Igor can summarize our team meetings.

Exactly. So that's the result right there. Okay, so we hold team meetings once a week and the automation pulls in the full transcript of the team meeting and then uses a fine tuned model on my most recent 30 videos that basically speaks exactly like I do like it alters chatgpt in the deepest way possible, I would say. And then the automation uses the fine tuned model and on top of the meeting summary and runs another summary. And then it also uses 11 labs that has been trained on my custom voice. And basically the result is it summarizes a one hour long meeting into a two minute long voice memo in my voice, in the style I speak in. It's. It's actually quite funny. I, I just realized maybe we should show one of these. But you don't have to do anything. Michael, I can, I can pull this from one of our recent conversations. Like I'm not saying they're perfect, they're not perfect by any account. And also I don't want to like share every detail of our team meetings here, but maybe we could, we could have a little listen listening to this. There's a question here from Zombie Gamer Realm. Do I need my own API key for ChatGPT to run a makeflow? Short answer is yes, absolutely. The video that covered the step by step shows you how to do this. So basically it's linked in the description below. You can go to the YouTube video. There's a step by step you need to go to the OpenAI playground, get your API key and set up a connection inside of make once the entire process. If you've never done it before, it takes five minutes tops. Maybe minimum one minute or something. I could show you here in one minute, but I think the video covers it so no need to go over it twice. But I think this like custom custom thing is pretty cool. So I'm sort of. Okay, this is very unprepared. So we're going to pull up a random meeting summary. I'm just going to look into our Slack conversation here. Pull one of the recent ones and we're going to play it to the stream and you're going to see what team meeting summary on a fine tuned version of Igor with a fine with a custom voice in 11 labs sounds like. Okay, that's a mouthful basically. Long story short, this is what you look, this is what it sounds like if you use AI to summarize voices. Okay, how can I get it into stream here? Can I just open this up inside of. Inside of QuickTime Player? I think this should work. Let's see. Let me briefly switch on over to the desktop view. I think here we get the computer audio. Sorry Michael, you're invisible for a moment here. And let me just see if this is audible also. It's not perfect, right? So there's like hiccups here and there, but I think it gets a tone really well, the fine tune and I think also that the voice just sounds like me. Let's see. Hey team, welcome back. This is Igor with the recap of the AIA weekly team meeting from October 28, 2024. Okay, let's get into it. Is that audible? I think it's audible, right? I think the system audio is coming in and clean air chat. Just please, please confirm if you're getting this. I'll play like, I don't know, 20 seconds of this and hopefully nothing, you know, like, I think this is fine, but this is amazing. It's all done with make. Also just automation that takes all of this, pulls all of this together. I was a bit late joining the meeting because there were two calendar links that confused me. By the time I got there, Raluca, Matthias and Daniel were talking about food and local customs. Let's move on to last week's events. I asked for feedback and Matthias shared his positive impressions about the presentation, mentioning that the vocabulary was strong, though pronunciation could be improved. Then we had an interesting discussion about whether American English is actually native English, which. Well, it is in my opinion. Raluka. But see how, see how it like uses the way I phrase things. Instead of this is trained on news you can use. Okay, so the Friday show we have, we pulled in I think over 30 episodes and turn them into a data set for fine tuning. But like, see how it speaks like me. This doesn't speak like generic chatgpt, it does this. I don't know, it just has this way of speaking as I do. I think this brought up the need for standardized PowerPoint templates for events with consistent branding. Matthias supported the importance of color consistency but noted that structures can vary. I confirmed that we have templates on notion and suggested we collaborate on a new one. Hey team. Just to summarize topic four of our meeting, Matthias reported back on the success of our no Code ethics talk and Ripple at Agents events. Meanwhile, Dominique mentioned the low turnout for their event. We then circled back to how we can improve event attendance. Different formats and marketing strategies were explored and I suggested we could try prerecorded mini courses to set the stage for future sessions. Like, isn't this amazing? I think this is pretty amazing. So like the voice definitely sounds like me and the fine tuned model also speaks as I do. This does not sound like ChatGPT. Right. So I don't know, I think this is pretty awesome. So this is the next video we want to do. So we partnered with MAKE to do two videos on the channel. I was like, hey, we want to show off multiple flows. The first one was this research automation that I showed you and the second one could be this feedback seems to be really positive in chat so far. If anybody has a. Any feedback or suggestion or what we could put in here, then, then we could do that. It sounds like a spy reporting back. It does, doesn't it? It's like, yeah, somebody listening in. But yeah, I can basically show you how to do this full flow and then you can use it for your own meeting reports. Or you could, you know, you could do something like take your boss's voice and train a Lava Labs model on it and then all of a sudden you can give meeting reports to your boss and his voice. I don't know, you can use this however you want, you know, as long as it complies with your local regulation, I suppose. But yeah, this is something pretty cool that I didn't even plan to show here, but like, there we go. So, yeah, pretty amazing. Ryan saying that is crazy accurate. Yeah, I think it's surprising. I think it's. I'll say this about fine tuning Michael, like Maybe I would love to hear about your experience with fine tuning because I know you created a lot of fine tunes with us lately. But in my experience, nothing communicates and transfers the style as deeply as a fine tune. Right. Like you can prompt all you want, you can give all the examples, you can use all the prompting tricks, but nothing cuts as deep as a fine tune. What do you think about that statement and what do you make of fine tuning?

Yes, I think so. Fine tuning is the best way to get this thing. For example, we have your voice, your writing style from the AI news you can use. But also we have a different fine tuning with Daniel from our team that's responsible for creating the newsletter. And also we starting generating part of the newsletter with this fine tune GPT. And what's really important, English is not my native language. So for me it might be really hard to describe exactly to GPT what I want, what I expect from it, how it should respond in this specific style. But when we fine tuning, we not writing to GPT what it should do, but how we're showing how it should write. And that's I think sometimes better is to show than just to write.

That's amazing. That's a great insight because like, just like with multi shot prompting, you give it multiple examples. In this case you just give it 30, 50, 100 examples of how it should speak and you don't really have to describe it, you just show it. Right. Which is an amazing way. Yeah, that's a good insight that I haven't thought about. Yeah, totally. So fine tunes, this is what you can do with finetune with a combination of different AI tools. Right. We use 11 labs for the voice, we use fine tuned GPT. Was this a GPT4.0 mini model?

I think remember right now I use more 4.0 because it's still slightly better results. But also I Test comparing the 4o and 4 mini for the newsletter and it's similar but I still think in my opinion the 4O is better.

Okay. Okay, good. So increasing price but recommended. Yeah. So fine tuned 4o model and all brought together in make, right. Like you can have all of these standalone things you can prompt in ChatGPT. You can use the 11Labs platform to create the voices. But there's a lot of copy pasting included, There's a lot of manual work included. Just like with this automation. Right. I could totally go on a LinkedIn profile, I could select everything, I could copy it, I could paste it inside of ChatGPT, I could summarize it. I could run a web search, I could pull in the different information and put it into ChatGPT and summarize it. But along the way I will have, I don't know, spent five minutes, maybe 10 minutes. We'll have spent 10 minutes and I don't know how many clicks. And all of that I can just set up inside of the automation once, make it scan every 15 minutes, set it, and forget it. And you know, worst case scenario, you're going to get no access error in the automation and you have to run it one more time. But that is basically seconds of your time lost versus every single time having to spend five to ten minutes. And let's be real, like most people, including me, just.