How To Use AI Like A 7-Figure Agency Owner

what's going on everybody it's Nick and for those of you that don't know I've been running a Content writing company for the better part of the last two years called 1 second copy we were one of the first businesses in the world to use AI at a scale necessary to make a real dent in the service industry we started way back in 2021 before chat GPT or gp4 had come out and we were using the predecessors of this technology to make a lot of money we managed to scale this to about $90,000 a month in about 18 months or so and the entire way that we were able to do that was by leveraging understanding of still nent technology like large language models Ai and that sort of thing and the reason why I bring that up is because in my time at this business and now developing make. com systems for other companies I've noticed that a lot of people are using AI but they're using it very poorly and they're using it for maybe 10% of the total outsized result that you could realistically get using this technology if you just knew a little bit about how it worked so what I want to do in this video is I want to show you guys how it works I'm going to run through four simple ways that I use AI in make. com to make large amounts of money on the internet and at the end I'm also going to run you through some technical bits about how these models work under the hood and how you can optimize them for better performance so if that sounds like something you're interested in stay tuned and let's get into it okay so I have a notion dock here

that I'm going to link down in the video description below and it talks a little bit about my rationale behind some of the AI model uh systems that I've built and you know some of the ways to use them that most people don't really realize so I'll cover that in a second and then at the end of the video I'm also going to cover what I call the technical bits which are probably lower Roi but they're still really important and beneficial and useful to know if you plan on using these am models in your day-to-day and let's be real in the year 2024 and Beyond uh in order to keep or grow a job we will all probably need to be using these Technologies in some way shape or form so uh maybe that's actually pretty high Roi too any who what I'm going to start with is I'm going to cover the various workf flows that I've developed that allow me to leverage artificial intelligence is basically like a little helper in the cloud instead of me paying a human being to do some of this work now I can just pay AI it's about 1 100th of the amount that I would have to pay and uh it does things basically instantaneously so a lot of benefits there and um yeah let's dive right into it probably the first

and the simplest workflow that I'm going to start with is uh image analysis now the reason why I bring up image analysis first is because basically nobody uses image analysis is an incredible feature and because nobody's talking about it and I'm like what the hell I figured I just start with that I'm going to use this in the context of you running an email campaign as you guys know that's one of my favorite uh examples to bring up constantly just because I've used it personally to generate outsized returns and what I have over here is I basically have like a scrape from LinkedIn I've gone into LinkedIn sales Navigator over here and I've pulled out a big long search let's say I'm looking for just some random search Consulting uh like this and then I fed it into a scraping platform and there are like five or 10 that are out there the one that I like to person personally use is called Phantom Buster and then basically um all you do is you just drag and drop this into this scraping function you know you feed it the URL which is up here and then at the end of it you get a big spreadsheet like this um I'm also using another platform called drop contact that's just to enrich the data and give me some email addresses but that's actually not the important part of this video and you could just delete this whole thing uh it wouldn't matter the important part of the video what I'm really curious about is I'm curious to see whether we can gain any useful information from the profile image URL column and what I want to do is I want to build out a brief little tool that will take as input an image from one of these columns here I think it's just image URL yeah there you go profile image

image URL and so we got a nice handsome fell right over here and what I want to do is I want to feed this into artificial intelligence and then I want to have ai tell me something about the image so that I can use it to personalize the first line of an email I want you to imagine that you receive an email that's like hey how's it going Nick I hope you're doing well here's this business service that I'm providing and here's why I want you to do it and then I want you to compare and contrast that with like hey how's it going Nick dude like your profile pick looks like the perfect day in the whole wide world also I own a shirt just like that um my name is Peter and I want to help you do X Y andz with whatever system right uh the point is by personalizing the email in the latter case at minimum you'll usually at least have a little bit more interest or at least um you know some curiosity behind who you are because you've implied that you've taken the time out to like actually look through the person's profile and you're not just copying and pasting the same response that millions of other people are and at maximum that sort of personality can actually substantially improv conversion rate and sort of be the dividing factor between a business that succeeds or fails and so what we want to do is we just want to feed an image like this into Ai and then have ai tell us something about the picture and then we want to use that to personalize an email campaign so I've actually gone ahead and

I've built this and what I'm going to do is I'm just going to explain how every step of this works in make. com we have the action module here as the Google Sheets sech search rows so if you go to add and then you type sheets Google Sheets is going to pop up here uh mine is not because I've already added this as a favorite but when you select that you'll have a big list of items that you can choose from and then what I've selected is selected the search rows and uh this is going to be the trigger to my flow this is what's going to initiate the rest of it what I have here is I've gone through and I've selected a spreadsheet with my specific data and then what I'm doing here is I'm just returning one single row at a time because I just want to test this next I have an open AI analyze images or Vision module the way that you get that is you search up open AI once you find that go up to the very top here and it's this is like the first module that nobody ever uses um select that and then you'll have

a module here where you can add a prompt that's sort of like mine what I'm doing is I'm telling um GPT 4 Vision in this case it's the name of the model to analyze the image and then write a oneline email Icebreaker using the following Json format this is interesting and it's also very important if you do not force the outputs to be in Json a lot of the time if you're using make. com they're much more difficult to work with and you can't really like incorporate them into the rest of your flows and so what I always do as a design pattern is anytime I generate something with AI I'll always feed it into a parse Json module and then I can use that and I'll show you that in a second any who the actual prompt says analyze the image and Rite one line Emil I breaker using the fall on Json format I then give it an example of what I want it to do and then I also provide an explicit two or three examples showing the sorts of tones of voice that I want it to provide um you know the format that I want it to stick to and that sort of thing and I'll cover how all that works in a moment um that's part of the technical bit that I was talking about here this is technically called fuse shot prompting um but it's one of the ways that I can ensure consistency across models next all you need to do is you need to supply an image URL so I've gone here and then I've clicked image URL and then I've copied in profile image URL which is one of the fields way down here um near the bottom of this and what this is this is basically this URL up here and so inside of this gp4 module they sort of like a built-in image um parsing module and then for the model I've selected gp4 Vision preview Max tokens 300 temperature 0. 7 I always like to pick the temperature make the temperature lower than default because I find that the um responses that are being produced by AI these days are sort of silly and then because I have a transform as the last module in the red I'm going to get a warning but I'm still just going to run this and what you'll see happens when we

feed in the image URL is you get a uh you get a result that looks like this Icebreaker that's a great shirt in your profile pick blue stripes are my go-to as well if we look back at here what's this handsome fellow wearing well he's wearing blue stripes and so uh you know if you get a message like this and if we're just not kidding ourselves here you are probably like five to 10 times as likely to respond I mean to be fair the base rate of you responding to any cold email is probably pretty low so five times that might just be like 20% or something like that so it's not like a guaranteed sort of thing uh but you know having a degree of customization and basically implying that hey I'm a human being who actually took the time and energy out of my day to read through your profile and look at your picture um that goes a long way now we have a parse Json module right over here which is actually taking that and then what you can see is um we've converted adjacent string which is just a bunch of text and then we've turned that into like a make. com uh key value collection pair and so now we actually have a variable called Icebreaker that we can access the text within and you can imagine how if you wanted to add this as part of your email flow uh you might just add like an email module here um probably the more intelligent way to do this would be uh you would go through and then like add this to instantly or add this to some other cold email platform but maybe um hypothetically you just say you know quick q and then let's just add the person's name Quick Q Lopez hey Lopez and then let's add a couple spaces now let's insert this Icebreaker let's go and then you would say something like I wanted to tell you a little about dot dot and I'm just going to use this as an example and then let's go back here uh I'm sending the email here unfortunately I should have chosen the draft module but whatever I'm just going to send it to let's do just my base email just so I can show you guys what this might look like save that puppy we're going to run it oh sorry I'm not um sign into the right one here okay we're going to run it then I'm going to go to my email address and then I'm just going to see this live I want to see this as this comes in just to show you guys what sort of quality we're looking at so you can see um the text that you got it was different obviously than the text we had before because a is flexible and it comes up with its own uh with different variations every time but it was hey first name that's a great selfie those trees make for relaxing setting I wanted to tell you a little bit about whatever now in this case what I might do to make this higher quality is I might go back here and then I might say that you know make sure to explicitly reference a profile pick in your Icebreaker the reason why is if we're sending it through email it's probably not entirely evident um what we're talking about here so if we just say that's a great selfie that might be construed as sort of weird alternatively you can imagine how you might just send that over LinkedIn directly and saying like um you know hey that's a a great selfie that would probably be a lot more obvious and uh and attributable to that fact that you're reaching out to them over LinkedIn but this one looks a lot better I like the strip shirt in your profile pick makes for a great outdoor look these three dots here is just a Gmail thing if you have the same text in multiple subsequent emails it just tries to like hide everything that's a duplicate but um just pretend that this is back over here and then it says I wanted to tell you about whatever the pitch is so very quick and easy way to you know customize an email sequence based off like visual data which basically nobody's doing and I think that way more companies are probably going to be implementing this over the course of the next year or so might make sense you guys to get ahead of the curve and just see how you might be able to do something like that on your own so before I forget I'm just going to save this blueprint got it and then I'm going to move on to the next flow that I have so

uh the very first way was image analysis we're going to mark that as done the second probably most valuable way to use AI that a lot of people aren't using today is as a filter so not only can you analyze an image using the image mod you can also perform basically like some cognitive functioning on a set of data and you can determine where you want that data to go um in make. com scenarios and just a lot of automations uh usually you know you'll use either routers or use some type of switch module to determine whether some output is X or Y let's say um but you need to do so using procedural rules so keywords or maybe some mathematical function or that sort of thing the unfortunate thing about reality is that a lot of these decisions don't really work one to one with these procedural rules right they're a little fuzzier than that and the cool part about AI is it allows us to basically use fuzzy logic in order to determine whether or not something is in one camp or the other now if that sounds like an abstract example for you which it was um a better one is probably uh what I'm about to show you in a second which I've showed multiple times before which people are quite interested in but essentially there are a lot of these request for proposal websites out there and a request for proposal website is a website that basically just lists a bunch of people that are looking for um contractors or people to bid on a job so upwork is a perfect example of one of these platforms upwork has a giant list of people basically being like hey I want somebody to help me with make. com and zapier build a treehouse for my son-in-law right just a bunch of like requests for bids and then what you do as a contractor or service professional is you'll go in there and then you will bid for that job the issue with these platforms I find a lot of the time is that there's just so many jobs that it's difficult to filter out all the irrelevance and then focus on high signal to noise and so AI is fantastic at basically allowing you to do that with minimal issues and what I've done here is I've actually created a flow that goes on to upwork pretty consistently I think this do this once every hour or two yeah so hour and a half it pings a resource which in this case is called an RSS feed it parses the description and then it feeds it into Ai and a lot of people that have been watching my channel have been like hey can you break this down can you actually go through this exact flow in detail um and so I mean I've sort of done that half and half in another video but um this time I'm actually going to focus on the AI bit so anyway the first part of this module here so first of all everything works off of an air table the air table looks like this we have a title column a description column a status column URL column you can click this to go to the job URL uh and then a copyable message column and these are really the four or five most important columns we'll cover everything else in a second but this is just one example of something that you could probably implement on a dozen or more big requests for proposal platforms where you're taking some information from the resource adding it to a third party data store like Google Sheets or air table and then using AI to filter so any who let me just stop waxing poetic we grab the RSS feed items we parse the description here using an HTML to text module just because the description field is always coming in as this long HTML string and we want to minimize the number of tokens that we're going to have to pay for so we just want to eliminate that as much as possible then we feed it into basically our magical AI module and the way that this works is um we Define uh a bunch of prompts here the first is a system prompt which just tells the system or tells the module how to identify and so we're telling it to identify as an intelligent admin that filters jobs easy enough right secondly what we do is we Define a user prompt and then we Define a bunch of examples after that user prompt and so my user prompt is hey we're an operations agency that builds Outreach systems CRM systems project management systems no code systems and Integrations blow as a job description filter it for relevance true or false in Json some of the platforms we use include big list of platforms if relevant write a short introductory Icebreaker note API Integrations are good since we can develop them using no code we don't develop full stock apps we don't use Python telegram WhatsApp Facebook bubble soft or Google Apps scripts or chat Bots I add that line in there because I just hate all of those platforms we're not VA virtual assistants and we don't work on very tiny projects like Integrations that will take less than 10 minutes to do or those that require extremely fast same day or 24-hour delivery I then also provided some example client projects that I built out for other people and so these are meant to be social proof sort of big dick projects that you know get people a little bit uh interested or excited like end to end project management for a 1 million year copywriting agency that is my uh company one second copy right over here outbound acquisition system that grew a Content company from 10K a month to 90k a month and 12 month that's sneaky but it's also my content writing company uh monday. com CRM and project management for recruiting company doing greater than 1 million year that's a client automated content generation system for an SEO agency doing greater than $10 million a year that's another client and then click up CRM build for a marketing company doing $1 million a year that's another client so uh I've given it all these projects and now that I've told it what I wanted to do that isn't enough I actually have to go in then I have to give it a bunch of examples and so that's what this pattern is here where uh we have a system prompt first then we have a user prompt where we tell it what to do and then we have a bunch of user assistant Pairs and this is sort of training the model so have a user pair here where I feed in basically just all of the description of the job as an example and so I'm saying hey if we give you this job as an example here's how I want you to respond and so I feed it with an assistant prompt that says Jason um curly braces result false reason this is about writing nothing to do with systems Icebreaker I think if I were to be smart about this I'd actually go and I'd put the result after the reason I would probably improve the quality now that I'm thinking about of the result yeah and I can get into why in the technical bit in a moment any who then there's another user prompt here with another example and then there's another output and let's just do this consistently across this whole thing I love make. com because I tend to um I mean I'm always learning and so when I go back to an older scenario or one that's been serving me well um and I read through it from top to bottom I just see so many opportunities for improvement here's another prompt from User it's as air table expert needed and then here's how I want it to respond if the reason is true so I'm going to go here and then delete that and then paste that over here good and then here's another example and so you can see I'm doing this quite a bit the reason why I do this as many times as I do is because it's just worth the money that I'm spending in tokens realistically it cost me a couple of cents every time that it does filtering and then uh for those couple of cents I basically just like save myself I don't know um maybe like a minute each and so you can imagine how I mean this is just me and I'm more of a freelancer at this point than anything but if you were to do this at large enough scale and you had a lot of people working for you in a company um it probably makes sense to do this um and maybe even give more examples any at the very end after I give it like six or seven example pairis I have a user prompt and this is my final prompt where I just feed in the same information that I fed in before I just do it using variables and then uh I've defined Max tokens as 250 I've set temperature to 0. 3 because the temperature of a model is basically it's Randomness and I do not want it to be very random I deterministic and pretty simple I'm going to cover everything and anything need to know about temperature at the end of the video so if you're interested in the technical bits behind that stay tuned um and then I don't do anything for frequency penalty or presence penalty the response format I leave as text what I do next is I will then parse the output of the previous module just like I did in the other scenario here and that just makes it easy to work with and then I go through and I will just do a bunch of tertiary sort of like text processing function so I'll get the budget and then I'll split it up into an hourly or project job and then I'll add it to my a table and what I end up getting is a couple things so obviously I get the title description uh URL and the rest of the fields but then if you remember I also added an ice breaker and that Icebreaker I've added to an introduction field here the introduction field is basically me telling the AI that I wanted to write a short again introductory Icebreaker um that I can use in the job application to later copy and paste into whatever the request for proposal website is and so in this case you know let's look at a good one here um I've worked on similar projects involving API Integrations and automation I'm confident I can help you with this task it's not really the best maybe this one's better confident on the man you're looking for he's make a com to build automation compan doing greater than1 million doll a year I understand the need for human readable documentation would love to help you with this great so this is sort of an introduction and then in air table what I do is I set um I set a formula field that looks like this I feed in so I concatenate that just means add text together the introduction and then I also have like a template that's just the same text every time and I just concatenate the two I put that in a formula field and then this now is just like a simple copy and paste sort of thing that I can do maybe a higher of mine can do that sort of deal uh makes it a lot easier and simpler to apply to jobs in this way and the only way or only reason why this is doable is because I've now implemented like a filter just do all the hard work ahead of time for me if it identifies a job as being worthwhile well then it can go ahead and actually even pre-write some of the work that I would be doing otherwise so I like to think it's a little sneaky and a pretty intelligent flow um I don't see a lot of other people doing this basically I don't see anybody else doing this aside from I'm sure a lot of the people that have started building it after I showed it in my uh in my upwork video uh 3 weeks ago uh but you know like outside of that small little realm uh very few people understand this and I think this is probably one of the highest leverage use cases um so I would try and build filtering into any app that I'm building or any backend workflow however possible one of the reasons why this is so powerful and this is the last I'm going to say about this is because artificial intelligence by nature is extremely flexible and flexibility is both its strength and it's also its downside if you think about it you know we're trying to Output code reliably here and you know if I mean up until very recently and when I say very recently I mean like literally in the last four or five months it was very difficult to get one of these large language models to Output the same code reliably it's only very recently we've saw that and so you know the flexibility in Ai and its ability to do fuzzy matching and all that stuff that's its strength but it's also a massive downside when you want to build like an Enterprise level or small business automation that's supposed to do mostly the same thing every time right but there's a way that you can work around that and that's what I just showed you instead of using it as the end all be all just use it as something that filters and then depending on the output whether it's a or b or left or right you know you can build procedural logic to take that fuzzy match functionality and then do something with it that's a lot more reliable a lot more stable so hopefully yall understood that um but yeah that's how you use filtering

the third thing that I'm going to show you guys how to do is content generation now in a previous video I've actually shown or I've sort of revealed one of my big prompts for generating articles on the Fly and these are like quite heavy you know 1500 to 2,000 word pieces and these are very involved and they're very sort of high value um what I'm going to do in this video is I'm not going to show that cuz you guys just watch that other video if you want I'm going to show how I generate um proposal and I've also shown this in a previous video but I'm going to dive deep into what the actual prompt looks like this time and so the way that this workflow works is basically um during sales when you're you know about to speak with a pro or when you are um finished speaking with a prospect uh we're filling out a quick little form and that form just has a couple of questions one of the questions in the form uh that the form asks you is something like hey what's the problem the client is suffering from another question the form asks you is hey what's the solution you want to build for this and the really cool part about AI is you can take a very tiny bit of basically top level human instruction and then using AI you can multiply it into paragraphs upon paragraphs upon paragraphs this is known as expansion and it's like one of the key ways that content is generated using AI reliably so what you do is you have a human provide some type of little snippet which in my case would be a problem or a solution and then you feed it into Ai and then have it blow it up and turn it into basically like a big proposal and so what I've done here is I have a discovery call and so when the discovery call form gets filled out we go to these type form modules which I've talked at length about in previous videos but this is just a simple design pattern to get you in the habit of separating the trigger um from uh basically like a testable module that you can like run whenever you want and so we're separating the source of the data from um I guess what's the best way to frame this that I haven't already done it just becomes a lot easier to test because then you can always just move this here and pull data from this module instead of that module and then just always have access to it anytime that you want on demand you don't actually have to go out and fill out the type form but if that doesn't make sense let me know and I'm more than happy to walk you guys through it any who um you feed the output of that into a gp4 module again and what I'm doing here prompt wise is I'm saying you're a helpful intelligent sales assistant which is slightly different rationale than I had previously um I'm then giving it the same pattern that I had before where I'm doing a user prompt to give it highle instructions and then I'm actually just giving it a bunch of user and assistant pair examples and so what I'm doing again is I'm saying hey we're an operations agency that builds out on Outreach systems CRM systems project management systems no code systems Integrations below is a loose scope turn that scope into a highquality proposal in Json our clients are Enterprise so write in a Spartan No Frills tone that implies intelligence what I do after then is I basically feed it my own JavaScript object which I've or I am going to compile in a second the reason why I do this is because feeding it in this format allows me to compress it a little bit I don't have to worry about any formatting and I think that this is just one of the many ways that you can communicate with a you don't have to do it this way but I just tend to like to structure everything into code blocks as much as possible so the input or the example in providing it is hey when I feed it with this code right which talks about some company that does X Y and Z then I want you to turn that into this huge much longer string of Jason that I'm going to use to feed into a proposal that I'm generating I do that again here with another user prompt where I talk about in this case this is uh some real estate company right and you know a lot of these are fictional um and then I have the same thing an assistant response over here where I basically blow this up into a bunch of variables there's like uh problem pitch one there's solution pitch there's title right so I'm getting it to generate me a title I'm getting it just bunch of information based off of some small input this is the expansion that I mentioned earlier feed in maybe 200 tokens and then get out like 700 in response in this way AI is a multiplier for human leverage and like human productivity and then after I give it a couple examples I then just go to the actual request which is another user prompt and then I say business description whatever the business description is Problem whatever the problem is solution whatever the solution is tools whatever the tools are um these are all questions coming from the type form by the way if that wasn't clear already any who then feed it into Json again then I'll generate milestones and timelines and so I have a very similar flow set up here where basically all I'm doing is I'm taking one tiny text field which is rough timetable which literally just says a week or two weeks or 5 days and then I use it to generate a massive or a much bigger list of Milestones uh on how that project might actually work out in practice so maybe AI is generating a milestone for a CRM build and I'm saying okay great you know if it's one week then generate me five Milestones but write it intelligently and be fuzzy about it and turn That Into You Know day one we got Project X day two we got project y day three we had project z um you know that sort of thing and so in this way it's a very flexible sort of and very simple way to take literally two words like the number and then the word week or day into you know a very detailed or supposedly detailed breakdown of what the Project's going to look like in practice anywh who uh then we'll actually go and we'll send the proposal and I'm not going to harp that to death cuz I think most of the alpha

in here was already gained that's the third um item and then the fourth item is scraping uh which is where we can use AI to basically you know we've pulled a bunch of text from or we HTML I should say from a website uh and maybe that's like uh you know like a listing website or something like you guys probably saw on my red fin scraper or maybe it's somewhere else and then instead of doing the parsing procedurally we use AA to fuzzy match all of the text in the document and then extract only what it is that we're interested in and so if that sounds complicated uh it's really not here's a very simple flow that I set up to do this I'm starting with my website so I'm calling an HTTP request module you can find this by going to add writing HTTP and then getting this blue one with a little Globe inside of it and then clicking make a request that'll pull up this what you want to do is just put in the URL click get and then leave everything else as a default and click okay and then I'm just going to test this to show you guys what the output's like you'll see we got a bunch of HTML and where we're getting that HTML from is the actual web page that you know we're calling and so in our case it's this URL left click. a which so happens to be my awesome website service you guys should definitely sign up for um but you know we can see a bunch of data here that's uh on my browser that's also represented in this page right like um you know find the perfect offer automate your lead acquisition you go back to the website find the perfect offer automate your lead acquisition so it's the whole website it's just an HTML form what we can do is we don't want to just feed this into AI to have it extract stuff because it's just a lot of tokens and'll probably cost us money we want to feed it into an HTML detex parser instead and that's basically going to take as input this super long HTML string that we just outputed from the module 4 over here and then it's going to strip away all of those tags that we don't need and turn it just into plain tag text and in that way we're going to eliminate probably 90% maybe not 90% we're probably going to eliminate a fair amount of the text so we don't have any of those tags anymore and then what we can do is we can feed this into AI to have it scrape or extract stuff for us and so as an example for this what I'm going to do is I'm going to pretend that I'm an SEO company and one of my favorite things to do is run parasite campaigns on uh other companies and so I have a big Google sheet with a million in one websites and all I do is I pump those websites into this flow right so maybe a Google sheet is the is the trigger here then I feed the URL into the get a website module and then for every website I call I parse the text and then I feed it into gp4 with a prompt that says hey below is a plain text straight from a website I want you to extract all the menu and navigation items and write them in this Json format and so I have a Json format down here that just says navigation items and then I have an array which is some examples home about Etc and what I basically want to do is since I'm this parasite SEO company uh my whole job is I just want to uh scrape all of the links on the page and maybe I could say this better maybe instead of saying I want to scrap all the menu items I could just say links but I'm just going to assume that I'm specifically interested in the navigation links CU some websites you know uh when you go to them up at the top they have nav bars and then inside of the nav bars there's like a nested other nav bar with like all the links and all the service pages right so this is something that a lot of companies I think actually do any who after that we feed into a parse Json module and then that's that so let's just run through this end to end and let me show you how powerful the extraction features can be the output of this text is like a million in one let me see here how many tokens or how many words okay I don't know oh it actually looks like there was an error probably outputting the Json weirdly enough H yeah there was an error there um it's probably because I'm using the new model yeah the new model is uh historically uh less reliable than the older models so I'm just going to use 0613 for this run this again okay great looks like that worked um so basically we took uh I don't know um how many how much text here we took about 1,00 or 12200 words and then we instantly within 3 seconds extracted only the items that were relevant to us which in my case is all the navigation item links this might be something that would take a virtual assistant like 3 or 4 minutes per page we can do it another 1 second for you know just a few cents and so I don't mean that you guys should be building scrapers for a living just because you can do stuff like this I just mean that there are probably tons of workflows where you guys can break those workflows down into systems that involve extract like uh scraping or pinging a resource and then another system that might involve extracting data from that resource and that might be enough to save your company or maybe the company whose systems you are building like hundreds of thousands of dollars a month maybe make them a bunch of money right a lot of options here a lot of opportunities and it's just for people that know how to harness and sort of Leverage AI uh the right way so yeah that's that for the systems here I wanted to frontload this video with examples what I'll do now is I'll talk a little bit about the technical bits behind um artificial intelligence and I feel confident talking about this because um like we've been using AI for better part of three and a half years now I would say um I was using it before I started 1 second copy I was obviously using it while I was running 1 second copy trained over 30 maybe 35 writers on how to use this technology in such a way so as to minimize our token usage but also maximize the perceived quality of the results so I'm very familiar with this I also know how to build models um this is a skill that I thought I was going to have to use like three years ago when all of this stuff was just coming to a four and I was starting to see the potential of this technology so I actually went out and I learned like the math behind it I spent four months locked in little Sky Box in downtown Vancouver little shoe box apartment uh just reading all about back propagation and stuff and I got to the point I could build these language models and it's given me a very in-depth understanding of I think some of like the more mechanistic bits that most people that might consider themselves like prompt Engineers or whatever are missing um as an aside you do not need to know how to build these things I just I didn't waste my time I'm glad that I learned how to do this but it turns out that the uh the moat here is not your knowledge because anybody can learn to build this stuff within like a month or so the moat here is access to billions of dollars of gpus so if I had we' be calling this puppy GPT Nick but I do not have millions of dollars at gpus so we're calling it open AI or whatever

U anyway I feel very confident that these tweaks are essential basically if you want to be like a high performing prompt engineer uh so the first one is pound-for-pound the fewer tokens that you use in a prompt the better your model will perform basically car blanch this doesn't mean that if you have you know a prompt earlier like I showed you with like 10 examples then you have another prompt later that it show do with zero that the one with zero will perform better just cuz there are fewer tokens it doesn't mean you just be like yo write content good please and then it'll do a great job but what this does mean if you have a prompt that's 2,000 words and then you can find a way to keep all of the information preserve all of the instructions in that prompt but then cut down the word count to like a thousand then the large language model will perform a lot better with a th000 tokens than 2,000 and the reason why is because statistically if you think about it like this is a big math equation and in autor regressive models which is what like the GPT family the clouds the anthropics whatever all these models are all auto regressive models the way they work is um basically you know there's a bunch of text and then you want to generate the next token and in order you need to look back at all of the texts that's been generated and fed into it and then perform this big mathematical equation basically multiplying every single like numerical representation of every word and token together to get some number that can then help you evaluate the next token it's basically what's happening under the hood now because you know we're dropping zeros every now and then we're rounding numbers and that sort of thing the further on the text goes and the longer the prompt becomes the lower quality that next token Choice also gets and so if you have a choice between two prompts both that have the same information density but one that's 2,000 words that's 1,000 words always go with the one that's 1,000 words not only you're going to be saving twice as much in token costs but you're also going to be getting a lot higher quality response so that's one and that's something that I think that you know if the vast majority of people watching this just implemented that your prompts would probably be better than 90% of everybody else's on the internet

today the second is um the difference between F shot and then zero shot now I'm looking at this example now this isn't actually what I wanted to show you I think this is like a special paper um let's do this maybe graph I just want to show you guys like a visual here to show you um how the prediction accuracy works okay yeah that's pretty pretty solid uh this is basically like a breakdown of model performance um based off the number of examples that you give the model and so zero shot just means giving it instructions and telling it hey write me a poem one shot means giving it instructions saying hey write me a poem and then saying here's an example poem two shot means doing that twice three three times five shot means doing it five times few shot in this case just means doing it at like between one to maybe I don't know like 30 times in the literature and so what you find is the performance will go up basically with the number of examples that you provide you'll see that um it's not always a massive performance bump I mean here the accuracy on this specific task coqa was like maybe 5% difference but you can imagine now if you're writing content or you're doing some like Enterprise application where you're sending hundreds or thousands of responses per day that 5% can make a really big difference to your bottom line right and so what I would recommend everybody here get in the habit of doing is don't just do like zero shot don't just tell it to do something um always add a couple of examples you know like when we were generating the proposal for instance we weren't just like telling it generate a proposal based off this we showed it hey if we feed this in we want this to be the exam the result then if we feed this and we want this to be the result now I want you to try with my new data and the idea is you hope that it will learn the relation ation ship between input and output and sort of model some deeper like you know function so always do um more than less let me copy this puppy in here just so we have like a visual example of what's going on the reason the last image didn't work by the way is because it showed that the uh one shot performance was worse than the zero shot and I think that was like a statistical abnormality that was only present in that model I may be mistaken but I'm pretty sure that was that so yeah generally just add more prompts this does go direct at odds with having a longer prompt by the way so there's like a fighting sort of effect between the length of the prompt going up and then that having an impact on the performance versus the number of examples going up and then having an improvement in performance so be careful with it obviously the third is uh if

you're writing content so if you're doing some type of content generation application always use what's called markdown language if you don't know what markdown language is uh it's pretty simple like let's say that we have a we're trying to make a Google doc and we want the Google Doc to look like this heading one some story introduction heading 2 some stor sub header heading three right this is what we want our Google doc to look like well if we're using Ai and this is just the open AI playground it's just a way to test the API inputs and outputs if we're using Ai and I want to write heading one and then I want to make it big how do I do it there's no lever available to pull to make that text Big right we simply lack the ability and so the way that you get around that is with markdown format is like a universally understood format in text where if you put a hashtag to the left of a line that just means it's a heading and so this is a heading one this would be a heading two and three so if I wanted to just reproduce this these two pieces of text are equivalent yeah so this now is technically equivalent just in markdown to this text and the cool part about make. com is make. com offers a variety of modules if I just type markdown here I can select a markdown to HTML module where I can then feed this Tex in here and then run just that and you'll see the output is now HTML and now that we have the HTML we can feed this in anything we could just like feed this into a Google doc and then use that HTML to like automatically create all of these headings if we wanted to and so yeah just you know if you're generating text for any sort of writing purposes just use markdown um it's a hack that you know we found probably two and a half three years ago to just be way more effective because otherwise you're copy and pasting stuff it's just not the same um if you want to like confirm this go to markdown live preview. com and then just like look at text with uh with these hashtags if I had a hashtag here and then a space this turns into a heading one and so this is basically like the HTML version of this um yeah so there's that and then the last

thing that I'll talk about and this is sort of sad very near and dear to my heart is that modern models are uh they're actually a lot lower quality in terms of output than older models like just in terms of the sheer intelligence of artificial intelligence Believe It or Not uh I would say these models were a lot smarter about a year ago to maybe a year and a half ago and that's because what happened between a year and a half go to now is open AI got very concerned open Ai and th opic claw they all got very concerned about uh obviously the content being used for poor purposes and uh you know they wanted to sort of like restrict all of the things that could be said with these models cuz they're finding that some of them were generating you know quote unquote toxic outputs and that sort of thing which I understand if you're a billion dollar company the last thing you want to do is get sued because you have some racist Nazi model right um but because the way that they trained it was called reinforcement learning through human feedback rhf because of the way that the rhf was done it's really dumbed down the quality of the outputs and so model outputs today are just really they all have this distinctive tone of voice and it's all very happy go-lucky and like hello I'm more than happy to assist you in a world where blah blah like you could see that from a mile away if you're even modly observative and You' know that that's Ai and so that can take away some of the fun obviously because obviously we're trying to automate humans out of uh existence right we're automating things that you know realistically should be automated and uh this sort of makes it more difficult to do so there are a couple hacks to get around that and the simplest and easiest hack is just to use like if you wanted to write me a story about Wales uh just say use a Spartan utilitarian tone of voice this just cuts out all that um it's going to be less interesting and less exciting of course but it will be logical it'll be intelligent it'll be very straight to the point so in a time before man Wales rule the oceans vast and size their dominance was in challenge migration dictated their life north to south following the food now imagine if I remove this it's going to wrate some happy go-lucky actually that's not that bad yeah it's not that bad at all maybe a better example would be write me an article about Wales yeah I mean I take that back I don't know why but it's doing a pretty good job here oh never mind it's uh yeah I guess because of my increased temperature it's just devolved into absolute Bonkers okay I'm going to stop that here before I get myself in trouble we generate a mini Nazi model um a Nazi whale okay yeah so that's more or less it those are my four tips um I would just encourage you to if you're using this for some type of content generation purpose make sure to use markdown make sure to provide a couple examples of what you want make sure to keep the token count as low as possible and the information density of your prompt as high as possible and then anytime that you're getting this to perform applications which you want to be logical and consistent add a token to a prompt like Spartan or utilitarian or something like that in output it'll substantially improve the quality of what you get awesome thanks so much for watching guys it was really fun to put this together I have made a fair amount of money with artificial intelligence over the last couple years and I've sort of been looking for an opportunity just to talk about some of the hacks and techniques uh that I've built on my way up here I'm almost done my 30-day video challenge just got a couple days left if you guys have anything specific that you want me to bring up or talk about in a future video now's your time after I'm done the 30-day daily posting challenge I'm still going to post on YouTube I'm just going to dial back the volum a little bit so that I can improve the quality of every video and then make sure that the things that I'm talking about are new and exciting and they're always provide some type of value so yeah if you got anything that you want me to talk about specifically leave it down below as a comment otherwise like subscribe do all that fun YouTube stuff and I'll catch you in the next video