# How To Optimize AI Lead Generation with Einar Vollset

Channel: Rob Walling
Video: https://www.youtube.com/watch?v=bhGyFA1eTrs
Duration: 35 min
Language: English
Words: 7176
Transcript page: https://viewrankai.com/tools/youtube-transcript/bhGyFA1eTrs

---

[0:00] what I'm not going to talk about is um some of the sort of what I call almost the obvious stuff obviously like you know AI by AI usually we all mean chat GPT or at least llms and it's it generates text right so the obvious the obvious solution is you know what do I hate doing but I need know I need to do is generate a bunch of text right I need to write Outreach emails I need to write uh marketing content I need to write you know personalized emails I need to figure out you know repurpose in my blog post into um social media postings

[0:34] fitting it to Instagram doing all that stuff I'm going to talk about absolutely none of that whatsoever partly because it's sort of straightforward and obvious and also because it is getting kind of crowded that space if that's what you think of when you think of how can I use AI to help my business then I fully understand why you're sort of on the other side of this crested wave where what you're thinking to yourself is yeah this this is not that big a deal this is not going to change my business so instead what are we actually going to talk about I talk to Xander actually and

[1:06] he does a survey about what is the reason why um what is the reason why people come to microcom or are they hoping to get out of it and the number one thing is this group of conss here which relates to I need more top of funnel I need more leads I need more customers um just more stuff for my business I wrote all the code but nobody showed up effectively and so that's what I'm hoping to help you guys with here and I'm going to show you how to do it okay so my standard thing is just like if you're starting off with G with chat

[1:37] GPT or with llms or with AIS particularly with chat GPT my usual default go-to is I have a problem let's just ask chat GPT and see what it thinks right and so just for some context in addition to you know working with Rob at Tiny Seed I also a partner at a sside Investment Bank so what that means is we helped sell uh mostly B2B SAS businesses sort of in that 2 to 10 million AR range that's what we do mostly the private equity and so I was like okay let's just do it let's just ask CH GPT and see what it can do hello I run a salside m&a

[2:11] Investment Bank I'm hoping you guys can see the text by the way because otherwise this is going to be quite a boring talk at the back um I've run a salside m&a Investment Bank I'm looking for B2B SAS businesses with 2 to 10 million ARR that want to sell please list 25 and actually that's like a trick for me too if you're looking for something ask for at least 25 if not 50 of something because usually the first five qu five answers are absolutely super boring and predictable and then it gets interesting after that anyway so then it goes up and it's absolute garbage as you can see it says I can

[2:41] assist you in finding directories or platforms where they are listed for sale however identifying specific businesses within the requested Revenue range may require more indepth research or potentially direct Outreach to the companies right and then it goes on to make even I mean to a degree at least it wasn't hallucinating so badly that it started just giving me names right it could have just done that which is what he was doing before with version like chat gbt 3 it would have just gone Oracle Microsoft and just like made up seven German software firms that didn't exist you know but at least now is doing it now it's completely wrong because

[3:18] like for a number of reasons first of all bis bu sell they don't even list software companies like if you want to buy a chain of you know gas stations in Oklahoma bisb sells your place it's an amazing for that similarly flipper flipp is great but if you have a 5 million AR SAS business like if one of your friends is listing their 5 million AR House business on flipper then they're making a mistake and they should call me instead so and and finally like if they're already on Flippa or bis by cell or you know some of the other places well then they probably already that's

[3:51] too late for us you probably have a similar problem right you probably know when a big potential lead for you becomes a customer of someone else and and sort of picks your competitor but at that point it's too late right they've already made their choice you wanted to know that they were in the mood to purchase before they actually made their purchase decision so that you can influence it in that way all right so why does this fail in general this fails because llms are extremely extremely bad at like all the questions if you ask it things like give me all the whatever it is some sort of a

[4:31] produce some sort of exhaustive list or a long list of things is extremely bad at it because that's not the kind of a system it's it's set up to to create it's set up to basically pleasingly provide the next you know token or text um or word that humans like that's what it's good for it's really bad at things like all the questions so don't ask questions like that and then finally even if it wasn't uh bad at this the information simply isn't out there what are we doing now okay so my first thought was all right so I have a list of SAS companies I can figure out if they're SAS

[5:07] companies that's reasonably straightforward right you guys can probably if you look at a website you can probably go like this and go yeah that looks like a SAS company like it typically has a pricing page it does all this stuff so finding a list of SAS companies is not all that hard but the problem is particularly with like design not necessarily correlating to revenue how do you even know um you know that somebody is doing 20 million in revenue or 20 $20 in Revenue it's very hard to do right you can't find out actually so I'm very small some people spend way too much time on design and not enough time

[5:41] on marketing and sales and so their website looks baller but they have like 500 in Revenue versus some really ugly websites I can tell you are making millions of dollars so that's not actually a good way to do it however my first thought was okay and actually I'm not going to show this because the talk was going to get too long um my you know I actually use chat gbt to help me out with this too the first one was like okay let's just do a heuristic let's just say okay let's use builtwith and builtwith is basically like a search engine if how many here know builtwith

[6:11] do you know what it is okay most of you good so for those of you that don't it basically just means like what are the technologies that are installed on the various websites it's like a search engine but not for content right and and one of the things that it does is it includes this spend estimate so it basically says all right we know roughly how much various software cost we know what you know is installed on the software um on the website that gives us an idea roughly how much money they're spending on software each month the thesis is you spend more money on software maybe you're a bigger you know

[6:41] AR Mr business so genius right more spend bigger SAS business seems reasonable um the problem is it doesn't work and so if you just straightforward do that um you end up in a scenario where basically at best the model of theistic you come up with is like 50/50 and so it's wrong half the time and the problem with that is at least from my uh business and I think for a lot of your guys case is that if you are wrong 50% of the time you end up spending way too much resources on leads that actually don't fit you know as your ideal customer profile it just doesn't it just

[7:22] doesn't make any sense to do so and so for example like particularly if you're doing like Enterprise sales or bigger ticker like items a lot of the time what you want to do is if you think someone is a very good fit for you you might spend a fair amount of resources on trying to to at least make that lead aware that you exist but if it turns out that you invest all this money all this effort into somebody and actually it's just you know a SAS company that has no Revenue then he wasted all that stuff so for us this is it just means that's way

[7:54] too much time and effort wasted on reaching out to two small um SAS firms so what do we do well the actual um hero of our story is something called chat GPT Advanced Data analysis how many people here have heard of chat GPT Advanced Data analysis all right keep your hands up how many have used it more than two times okay so three Heroes here out of hundreds that's about right I I thought five I got got with got with uh three so I I'll goope with that what it is is based basically first of all you have to be a paying customer of chat gbt and if

[8:33] you're not then you're being too tight and you should really do that um it gives you access you have to go into settings and then enable it and then you can when you hover over the GPT 4 you can do the default one which is just gp4 or you can do browse with Bing which I don't know what that does and then but this is the one that you want the Advanced Data analysis and effectively what it is it's chat gpt's most capable model plus a python interpreter and it gives gives you this magical button right here this magical button right here means that you can upload files to

[9:05] it up to like 100 megabytes and it will do analysis on it so now it has the ability to upload files it has the chat GPT and it has a python interpreter environment that it can write and run coded and evaluated and this thing honestly is extremely capable and like if if your last interaction with chat gbt was like chat gbt 3 or 3.5 like 6 months ago you owe it to yourself to go and play with this thing because it is mind-blowing and basically for our purposes it's a little bit like having a seconde university stat student at your becking call which is obviously extremely valuable the only

[9:43] downside is the thing is disgustingly cheerful as you will see okay so let's just do this and like keep in the context of what I was doing again I'm just taking this approach where I'm like I'm just talking to it so let's say hello I a sside m&a Investment Bank I'm looking for B2B SAS businesses with 2 to 10 million AR and then I say if I have a list of companies where I know the Mr and I know this because of you know Tiny Seed and discretion I have you know some stuff and then um and then another list of those same companies with various externally observable facts

[10:18] so these are the stuff that I just downloaded from builtwith right do you think you could help me predict the mrr of a given company is greater than some number let's see well first of all like a side note here so I actually hate the word prompt engineering because what it ends up doing is giving people writer block and actually programmers are the worst at this because programmers in their head have this view that like they have to like model things in their head and then precisely write things down on the page and that actually is is really bad for chat GPT and in general you should just

[10:51] start talking to it and just like interacting with it much more because if you have like this what I call the chat gbt writers Block it's it's actually worse than the normal writers block because with the normal writers block you're not producing anything but with chat gbt writers block not only are you not producing anything but you're also sort of missing out on chat gbt teaching you something that you may not know so rather than thinking about this like oh I have to perfectly figure out how to engineer this prompt so I get the right answer it's just like just talk to it just interact with it it's usually a

[11:23] much more productive way particularly when you're doing exploratory stuff so like I said disgustingly cheerful absolutely all right predicting the monthly Mr of a company based on various observable fact is a classic regression and classification problem depending on you do it here is a high level approach you see what it's done now I just gave you two things I was just like hey I have this problem I have this data what do you think can you help me and instead it lays out basically the six steps that are involved in uh running a classic regression and classification problem from scratch so it says let's do data

[11:57] preparation let's do future engineering let's do model selection let's do model training model evaluation deployment prediction and then obviously there's there's the other uh Improvement so let's dive into this and keep in mind like my background is fairly technical I have a PhD in computer science but it's been a long time since I did any of this stuff and so because it's it usually takes you a fair amount of effort to get up to speed to use these toolings directly yourself I haven't done it in like 10 years Okay so oh and I I'll highlight some of the stuff here the stuff in red is like me going like yeah

[12:32] I sort of kind of this is like think of every like red line as like two days of work for your competent programmer to figure out okay how do I do this how do I go to stack Overflow what the some the example of this cut and paste this stuff run it oh do I need to understand this blah blah blah blah blah so it goes through and data preparation if you've ever done any serious ml stuff this they'll tell you this is usually the most painful part of the whole thing so the fact that it just kind of does it is sort of mind-blowing so first it says

[12:59] I'll check for missing values which of course you have to do but like if you never thought about it then it's kind of hard to come up with um then we'll need to convert the Technologies this into a format suitable for machine learning models I mean obviously that's true you need to put it into a format that makes sense but again that line there is probably two days down a rabbit hole of stack Overflow and figuring things out and doing like what the hell is going on here and then it says all right let's create a binary Target variable and again like yeah obviously it's a

[13:26] classification task yes or no we got to figure how to do that but even like learning to understand what a binary Target variable is and how to do that is probably another half a day at least okay so it goes through and it just does it in this case it just let's start for checking with missing values and it says uh technologist has no missing values however Mr has 48 missing values um and then it suggests some options for you in this case I haven't shown you the entire stuff but it says you want to remove the rul or do you want to remove the row or

[13:54] do you want to impute them and I just say let's just let's just kind of remove them okay it moves on the feature engineering all right next we will like I said it's very cheerful next we'll work on converting the Technologies list into a format um that can be used for machine learning and then let's do it into uh something called one hot encoding another line there this is probably another days worth of my time when I'm like okay I have an array of texts how do I turn that into a format that makes sense for these models while it's one hot encoding first I'd have to figure

[14:25] that all out understand that I needed to do something in the one hard encoding and then actually do it and like I said that's at least half a day if not two and this just does it and then it says we'll create a binary Target variable let's start with a one hot encoding and then it just kind of just goes away does a little work comes back all right it says the one hot encoding on the Technologies list has resulted in 2,161 binary features Additionally the binary Target variable Mr over 100,000 that was the threshold I sent has been successfully created amazing all right then and again again I

[14:59] just wanted to point out this is one session this is me sitting there for 45 minutes just [ __ ] around with it pardon my French and it says all right now that we have our features at Target Val we can proceed with trading classifier now I told it and this is me feeding this more data than you necessarily need like I told you that Precision is more important than recall in my case it suggests that I'm using a model that allows us to optimize for this metric and I cared about Precision more than recall anyway the fact is when I started this like 45 minutes half an

[15:27] hour before this this this point I I didn't really know what a random forest classifier was nor did I really understand that this was a decent model if you wanted to tune precision versus recall it just knows this and sort of just does the right thing while educating you in the process all right would you like to proceed with the training model and I go yes please because you have to be very polite to your model otherwise it won't behave as well the random Forest classifier has been trained and evaluated here are the results and then it spits out this stuff and I you know I like I said I I have

[15:58] PhD in computer science when I first SP this out I was like what does that mean I don't remember it looks good there's Precision there's an F1 F1 like racing that's good there's a confusion Matrix that's confusing okay but it's a little bit like it goes along and it says okay it tells you and you can just ask it this is why I'm saying like in general I see too many people like they get to this problem and they're like I don't understand this I'm going to Google it and then they Google the question I'm like what are you talking about just have it explain it to you and so it does

[16:34] it starts to do this and it actually says the Precision for the class is 100 which means when the model predicts that Mr is greater than 100 is always correct however the recall is quite low indicating the model missing significant number of instances which is fine and actually like the confusion Matrix I think is something along the lines of uh How does it go I think it's like uh true positive predicted positive true negative predicted negative something like that anyway but the point is if you know as little or less than me about this stuff you just at this point you would just spend 20 minutes having

[17:09] saying like Okay can you explain this please as if I'm and then whatever level you want five high school college whatever and it'll walk you through the model and do it and actually this side of the gate is pretty good you know it's comes along and it says Fair what this basically means is okay this is the ones I'm looking for right I want to look for Mr more than 100,000 um and and precision means like if it makes a call and says this company is more than 100,000 at least in the evaluation data set it says it's right all the time it's pretty good the downside here is um the

[17:40] recall is only 02 which means it misses 80% of the ones that actually are above 100,000 so I was like this is Sweet let's keep going okay so it then says would you like to try to fine-tune the model or is there anything else you would like to do and and again I'm sort of cut for times there and I say uh can we try an svm approach which is another model like in theory what you'd probably do here um if you didn't have short time and wasn't giving a presentation you'd be like you know tell me about the trade-offs of the various models and they would tell you about support Vector

[18:11] machines and random force classifiers and all the other uh different classification algorithms that are and the various trade-offs and how they were tunable and all this stuff and this is the interesting thing about this it sort of goes between being like an extremely capable tool that just executes for you and then at the same time when you need it at the level that you needed it becomes an educator which goes like oh you don't understand this piece let me explain it to you in the detail that you want um and then you can obviously ask it say hey I have specific example I have this this this you know can you use

[18:40] that as the example or the explainer for me so then it goes through and it says let's just do that let's just switch to an svm model and train it and it goes okay it's done and again it's been trained and evaluated again like this is me just saying yes clicking a button and it builds the svm does classification spits out the classification report report spits out the confusion Matrix for a wholly different model and if that doesn't improve you impress you then I guarantee you you've not not done a lot of machine learning or training um or statistics really in in in real life okay so what I show you could do from

[19:18] Zero to Hero basically in under 45 minutes so okay you so you think to yourself all right fine like this is great now I have a model but like how do I glue that how do I get it close to something act usable and here's another thing with this too because you can download files no you not only can you upload files to it like whatever it produces it can produce a download thing for you so what I said was okay let's proceed can you give me all the files I need in order to just run this like okay we we did this thing we modeled it out

[19:49] you know I'm not going to call chat GPT every time I need to do this classification how do I just do this locally and it says absolutely oh certainly in this case to make it it easy I'll provide you with the following files model file serialized version of the train model feature names a way to basically one hot encoding the new data and a python script to run it and then it explains what the script will do and then let's write it and then it's been created here you are and you can hardly see the download link that's pretty cool that's like okay so I was just

[20:22] fussing around I gave it some data we played around with it it it teach taught me some stuff that I didn't know and then it gave me uh something that I could just run so now I can come along and I can say we have 40,000 SAS companies in our database I could pray built with to give me the list of Technologies installed of all of those and then just run this model on my local database and now I have a pretty good idea of which SAS businesses are over my threshold in this case 100,000 Mr for me that's obviously super important because of the kind of

[20:53] companies that I care about and the key thing here is like what actually just happened here like it's worth stopping and pausing and thinking what actually happened and what actually happened was chat GPT used Python and the files you uploaded to solve this problem right and like specifically how it did it you can just look like you just take a look and you can say Well it imported the SK learn models uh it imported the random Forest classifiers it got the classification port and the confusion Matrix and this is how it actually run it and this is important because a lot of people the reason why they're

[21:29] dismissing chat GPT at the moment or similar models is they're saying like oh um but it's just hallucinates things I can't trust it if you're an expert in the field and you ask it about the field you you'll know how stupid it is I'm like yeah but this is not the same as that this is an imp precise sometimes but very capable tool using a very precise tool and so the stuff that comes out of it the files that you download from chadis model are not we're not going to hallucinate they're sort of deterministic or at least more deterministic than chat GPT is more importantly though conceptually what

[22:08] happened conceptually we just spent 45 minutes is and 17 cents is to get results that usually I never see unless this business like a SAS business has 10 million AR or more in some cases I deal with 10 million AR businesses that don't have this kind of capability because what do they need to do it this isn't product features right so you can't like imagine taking your current developer even if you're a developer and say oh just start doing this stuff well they might figure it out in a couple of weeks but your feature development grinds to a halt so basically what you're going to need and usually is like you need a data

[22:46] analyst or two you need you know budget to P pay for for bi tooling and then it starts to get into like expensive sounding Enterprise software like ETL and data lakes and all this stuff and in fact in Tiny Seed we tell people people like it's a good indicator that you can charge people more if they're using this kind of tooling and that's true because it's a very expensive thing but also crucially you need dedicated time to solve this issue which if you're like 99.9% of boost sters is you don't have dedicated time to do anything other than sell the [ __ ] out of your product and

[23:15] build more features right and the average bruise Trapper has none of that but you don't actually need that now like you can have the equivalent output of like a half a million of a year budget data tooling for 45 minutes and 17 cents and a bit of knowledge which I think is pretty cool okay so you might tell yourself well I'm not an investment bank I don't need to predict SAS Mr um just because you're being obnoxious and I can say well okay but you already have like basically what you're looking to for this particular use case what you're looking for what is the hidden what is

[23:52] the hidden variables what you're looking for and a lot of the time like can you predict which of your Lees look like your best stickers or biggest customers right you already have customers like you you definitely have some that you prefer over others maybe they pay you more they stick around longer well can you build a model that predicts which of your big group of leads look more like those and focus your efforts that way can you predict and prevent turn based on user action can you predict and improve you know what is your freed paid the conversion Based on clickstream data can you uh predict what is the most

[24:23] likely upsell opportunities among your existing customers there's a lot of stuff where basically what you're doing is um there's some hidden information that if you somebody would just tell you your life would be much easier well with chat GPT and this Advanced Data model you can basically build a model that predicts that hidden data so in the T of just being able to say what's what's wrong here I'm just I don't know how much time we have for this but like another example of this is pricing Consultants right so pricing Consultants are people that go in and they help you figure out how to make the most of your

[24:54] money and they're really expensive like our mutual friend um what's his name profit F guy who sold his business for 200 million yeah Patrick Campbell that's pretty much what he did like that's why he sold for $200 million because he was charging like $150,000 per pricing consultant engagement but it's very variable right it's very valuable like if you nail the pricing you know that's the difference between potentially retiring as a multimillionaire or having to get a job with Google again right and it instantly made me wonder huh I wonder if chat BT could do this now the problem is I don't know anything at all about

[25:27] you know what a pricing consultant even really does so then I just started talking to it all right so I was like let's do this your highly paid pricing consultant focused on SAS educate me on the demand curve and then I get some stuff like clearly like I did this before okay certainly the demand graph is a graphical representation of the relationship between the price of a good good or service and the quantity demanded blah blah blah in the context of SAS the demand curve can help you understand how changes in price might impact the number of subscribers when you plot the price blah blah blah it

[25:56] goes through right and then it says here are some of the concepts that you may want to know luckily for you guys you'll get these slides so you don't have to read it all now but it goes into Concepts like elasticity of demand which is like how sensitive is the uh customer to price shifts in the demand curve due to various external factors determinance of demand prices rated Goods taste experience expectations number of potential consumers all that stuff for SAS companies understanding their demand curve is crucial it helps pricing decisions forecasting strategizing for example if SAS company knows a is elastic it might consider low prices to

[26:30] significantly incing prices would you like visual representation and I was like yes I would like a visual representation so I click the button and it just tells you here are the various I'm going to skip across this guy here are the various kinds of demand curves along with corresponding Revenue curves linear exponential parabolic logarithmic they basically look like this now if Patrick like this graph here was like the money shot in any report that came from Patrick's firm like you paid $100,000 you were getting this kind of thing and it basically shows here's the um let me get make sure I get this right the demand curve is the blue

[27:02] one it goes down um and then the price goes up and then the red one is the uh is the um um the red one is the what are you here Revenue curve so you end up in a scenario where like how do you Max where do you maximize revenue for this friend well the right price is 50 right it's a pretty linear the intersection here the curve the red curve is the highest okay like I said I didn't know very much about pricing consultant but even I can figure that out the best price for this kind if you if this was exactly what your demand looked like the

[27:33] price the right correct price is 50 to maximize revenue and then you have different ones like here's an exponential decay one where basically it falls off much faster in that case the right the right one is is 20 bucks um you come along because the top of the curve is here is now on this one and then there's different demand curves of different ones again now I'm being educated about pricing Consulting right I'm looking at different demand curves and you might again think to yourself that's great but that's still how many steps is that away from um you know how many steps is that away from actually

[28:09] doing anything with my data and my customers like don't I still need a consultant for this like yes I've learned some stuff but like this feel still feels like there's like a gazillion different steps before uh I actually uh get any useful data and the answer is kind of and so instead I said hey can you generate me a template spreadsheet with real data that would need to be collected in order to perform this analysis on an actual business please indicate how many samples you need certainly like I said it comes along to perform this kind of analysis we need to this it is price point number

[28:42] of units solds total revenue date time period other relevant data all of a sudden for me this is like this is an achievable thing right now I have a in by the way here's the spreadsheet with a template that I'm going to need to do analysis on your specific data all of a sudden now I've gone from do I spend $125,000 to Patrick Campbell or do I figure out how to fill out this spreadsheet and then F feed it back to chat GPT you know I think at least initially I'm going to save Porsche's worth of money and just fill in this spreadsheet that's going to be

[29:15] my goal

[29:26] okay I wonder if there are any questions is it safe to upload all your sensitive data to JB especially when you're uploading files so that's a good question it's quite com it's quite a common question the way that I think about it is as follows you're doing this all day any day anyway like you're you're taking sensitive data and storing giving it to third parties which have computers on the internet all the time it may feel different because you sort of feel like there's some sort of an intelligence that understands your data and like feels like it's more like giving it to this other person but

[30:03] conceptually it's no different giving data to chat GPT than it is to giving it to S3 or Amazon S3 and you know you can even just straight with the API and stuff you can say Okay opt out a training data which is you know using your data for training which is probably the only way I think that data can leak out but equally like you can go instead of going on uh just chat GPT if you go on Azure you can have like the Enterprise version of chat GPT which can be like Hipp compliant sock compliant whatever ISO number makes you you know happy um so so actually no I don't think

[30:41] I understand the question and like it it it I get it often but realistically you're giving data to a third party with computers on the internet that's been going on for at least 25 years so no I wouldn't be too worried hi uh thank you very much for that talk it was really AES awesome thank you I want to know with this model that you created did you actually use it and has it been successful so yes I am using it at the moment and basically it allows me to prioritize like I said we have these 40,000 models and right now I I I haven't bought enough credits

[31:13] unbuilt with yet to to fill out more so we ran it on about 5,000 of them but yeah I'm using it and actually one of the things that the key thing really for this kind of problem is you got to figure out where examples are sort of and so I for example had like 200 companies where I happen to know the Mr you guys probably have like customers where you know how much they're spending and so you can use that kind of data another data point that someone pointed out was in the US there was the PPP loans during covid which actually tells you how much like uh what the company

[31:45] was and how much money they got which basically correlates to two and a half months worth of salary so then you can estimate how big they are right so that's training data that you happen to know this now you can build a predictor that predicts okay how how many employees does this company have um basically I have a question for you oh all right so well let's say that I'm running a SAS company and I have 500 customers and I want to find more like them do you feel like the same model like I would go because it's most of the audience here is in that space right

[32:19] that that going to built with and trying to match up maybe not all 500 of my customers but my 200 best customers or something like that right and trying to feed it into chat GPT and saying what technologies are they using is is that how you would think about it in terms of you would talk I would just start talking to chat gbt about it I would just be like hey and specifically the Advanced Data analysis I'd be like hey this is the data that I have hey this is the problem that I'm looking for like in that particular case you sort of need the positive and the negative you know

[32:51] like you need here are like 200 good customers here are 200 bad customers um and then it doesn't have to be built with it could really be anything like as long as it's publicly available data that you think in the larger scheme of things you'll be able to find for other prospects but built with is a good example um you would just basically tell it that like yeah this is the this is this is the class that I'm looking for this is the class that I want to avoid here's the data I have across the two can you build me a classifier then yeah I would do that so I'm just curious is

[33:22] there any like best practice that we should all be doing now to create the data that we can then later analyze like should we be writing every event to Red shift or whatever a system like it do you have any recommendations on that front it's a good question um I mean the answer is if you really optimize for this is yes I mean there's a reason why like you know the bi tools and the the the big companies that have you know teams of 20 uh doing this stuff full time why they use like data lakes and red shift and all this stuff and and that's effectively the approach like

[33:56] they just just sort of record everything and because later maybe you know that'll that'll be useful um I think in general if there isn't a giant amount of overhead then yeah I would I would just start recording a lot more data um and just like having available but that being said honestly I think most of the time like um most of the time particularly like in terms of finding leads like what you want to build a model on is not necessarily the stuff that's just internal to you you want to build a model and stuff that you can use on C people that aren't customers yet so

[34:29] there are things that are public data that you can find right because you want to build the reason why like you use an internal data set first is because you know the answer like it's like a cheat sheet like oh this is a bad customer this is a good customer the question is can you go from that sort of knowledge and build a model that uses you know publicly quote unquote unrelated data in order to predict whether they're going to be a good or a bad customer or whatever it is you're looking to predict amazing let's give a big round of Applause to Mr Anar

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt