บทถอดความฉบับเต็มของ Data Science Happy Hour 97 | 16SEP2022

on? Antonio, Antonio, good to see you here again, man. It's been so long. Antonio Hope you doing well. Vin Fascist. What's up? Our bell? Keith McCormick, Jan Duarte and of course, Russell Willis. Everybody chilling out in the in the in the room without the camera on what's going on. So keep the camera off. Shout Jason Faried dot Mexico and sinker. All right, let's get a pop, man. Anybody got anything they want to open the discussion with? Do let me know if there's any questions. Any burning questions y'all want to kick off the topic with? Do let me know. I had an interesting question that I was asking. I here at the dawn. During the the episode of yesterday. Right. And, you know, he's primarily a a creator. He's a creator. But he does all this stuff through the Internet right now. Harpreet: [00:03:38] It just got me wondering, like, you know, if the Internet was not around, how would how would your life be different? Right. Or you know what? Let's let's ask another question. How about this? Like another thought experiment I was doing was, who would I have become if my social media outlet of choice when I was in my young twenties was something different? For [00:04:00] example, Twitter, right? Like what if I what if I spent most of my time on Twitter versus like Instagram or Facebook? Because I only really started messing around with Twitter just a year ago and my mind was blown as to how much knowledge and wisdom is going on in Twitter. Just the ideas and thoughts I could've been exposed to at a much earlier age, how different life would have been. So I'm curious, man, if you know, I'm curious to hear what you think your life would be like if you had picked another social media drug of choice. When you're in your formative years or, you know, in the early years, social media. Let's go to Abdelhamid. Let's hear from him. If anybody else wants to jump in, please let me know if you got questions. Please do let me know. I will keep you up. Thank you. Go for Al. Speaker2: [00:04:50] You're going to go for the old man in the room and my like. They invented the internet in my freshman year at Ohio U. Harpreet: [00:04:55] What are you talking about? Speaker2: [00:04:59] Man. I don't know. I mean. Like, especially with social media. Speaker3: [00:05:04] Like certain certain avenues seem to sort of appeal to our baser instincts. So it's like, yeah, I don't know if, if I'd had Instagram around early on that that would have been net negative for me. So yeah, probably if I'd. Speaker2: [00:05:23] Had LinkedIn. Speaker3: [00:05:24] And YouTube around, which both of which I use for relatively positive things. Yeah, I think I'd be further ahead than I am today, certainly, if I'd. If I start and started upscaling and started watching like educational things on YouTube a few years earlier. Probably could have upscaled a bit quicker on my way out of the Marine Corps. Yeah, yeah. Harpreet: [00:05:52] Yeah. You know, surprising. Like, I didn't even know that

websites like, for example, Coursera or Udemy or [00:06:00] Udacity like these, these type of courses like, you know, websites that have all this knowledge and wisdom. I didn't even know this was a thing until I was in my mid-thirties and I was like, I've been reading textbooks the entire time. Getting bored is fun. But yeah, that's a trip, man. Let's hear to let's hear from Antonio. And if anybody else wants to jump in. Do let me know if you know, if you guys want to chime in on this question. I'd love to hear what you guys say. Just simply use the hand raised icon. A lack of hand raises will mean that. I'll move on to the question that we got from Gina. Gina's got a question about Pre-trained models, but Antonio, let's hear from you. Speaker2: [00:06:36] I mean, a couple of things come to my mind. I think if I went on Twitter other than LinkedIn, I. Speaker3: [00:06:41] Think the writing. Speaker2: [00:06:42] Style would have been a lot different. I go on Twitter and I see all of these, like, what is it called? Is it called copy copyright or something? Is all this concise stuff? So when I went from LinkedIn to Twitter and I have like, I don't know, I mean, even LinkedIn is short format, but not as short as Twitter, obviously. So trying to turn everything into a thread has been like, pretty difficult. I'm like, Come on, I just need, like, three more words. I don't want to start another tweet even now. Right now. So it's kind of. And I think it's very useful for people who are like the Justin Walshes of the world, who are pitching ideas and trying to earn money with being a creator. I think that's a. Speaker3: [00:07:22] Very valuable. Speaker2: [00:07:23] Skill set that I am missing. Now, that's that's like the one thing that came to my mind. I think the second thing was from a very young age, I was probably ten years old. I started I posted like videos on YouTube. I was doing statistics comparing different creators, but I didn't even know back then that you could actually, like, make money off of this. I had posted about ten videos and it was like comparing these different people. And then, you know, people I remember somebody commented on my thing like, Oh, well, you're, you're like, numbers might be wrong or you're not getting it. You know how people always like like to [00:08:00] shit on other people's work. So I kind of like backed away from that channel. I was, I mean, it was probably too young for it. But I think if I knew that you could actually make money and where it would be today, I would have definitely kept going. I think that's the one thing that I would I would change. I don't I probably wouldn't be anything crazy right now, but I definitely would have a lot more followers on YouTube and especially because I love that channel like for learning like you guys are saying, Yeah. Harpreet: [00:08:28] Antonio Thank you. Ralph We got a valid comment here saying maybe it wasn't a thing until your mid thirties talking about those online education courses and stuff. I think you're probably ready. That also has something to do with it. But let's hear from Gina then from Gina will go to Keith. If anybody else wants to jump in on this topic, let me know or if anybody has a question whether here in the Zoom room, whether you are watching live on LinkedIn. I welcome or YouTube. I welcome all of your questions, your comments, your concerns, all of them. I go for Gina. Speaker4: [00:08:57] Yeah. So do comments on this topic. I for one, am very glad social media was not around when I was young in high school, even younger. I think there's too much pressure from it. And of course, especially I think for women, there's so much pressure. I mean, it's bad enough with the fashion magazines you felt like you had to look a certain way. And it was heartening to hear guys, at least some guys say, I don't necessarily want to be with a super skinny girl, but that's but that's what we were kind of that was the message. And I think now with all of these social media platforms, there's just so much, I think, psychological risk for young people. And at the same time, if they can learn to navigate it successfully, I think as we see especially millennials and especially Gen Zs coming of age, they seem to have a certain ease with it that a lot of older people don't. [00:10:00] Yeah. The other thing is though, somebody mentioned brevity and just putting ideas out there. And the one thing that coming of age when the Internet wasn't a thing, shall we say, is that there was a time when if you were going to publish something, there were a lot of gatekeepers and you just couldn't publish wherever you wanted, whenever you wanted. Speaker4: [00:10:27] You can just say whatever you wanted to say. And because there were gatekeepers, stuff had to be of a certain quality. There was also a feeling and I think, Antonio, you kind of referred to this a little bit in YouTube, right? Like if you just put something out there, it might be great, but you're going to be subject to a lot of criticism. And at the same time, if you do put something out there that's really good in the current environment, it's difficult to necessarily rise above the noise. So the so the downside of that for those of us who grew up in that era, is that we can feel quite hesitant to just put start putting stuff out on social media. So I think certainly those who have grown up with it have an advantage. And I see Costa just joined and I know he on a day of silence happy hour I don't know some months ago maybe had some thoughts on the challenges of a younger person, particularly coming of age with Facebook chats and disagreements and all the rest. Harpreet: [00:11:37] Jeanne, thanks so much. Let's go to Keith. Thank you very much, Keith. By the way, if you guys got questions, comments, whatever, please do let me know. Monitoring the chat on all platforms. Go for it. Keith. Speaker3: [00:11:49] Yeah, I'm one of the older folks here too, so I

just looked it up. If I'm not mistaken, Internet was 1991, right? Is that the agreed upon [00:12:00] birth of the Internet year, give or take? That was undergrad graduation year for me, so this was definitely not something during my formative years. So I just want to comment a little bit on how I ended up being fairly active on LinkedIn. I've been blogging for a while, but I never blogged that consistently. So I would use Twitter to say, Hey, I just wrote a blog post because, you know, this is back when they had the character Women. So you couldn't say much more than that. But the thing about Twitter is I never found myself consuming it that much because it would alert you to the fact that there was a link to an article, but no one was kind of sharing their $0.02. There wasn't so much room for people to do that other than saying, Hey, this is great. Check it out, you know? So I started using LinkedIn, I think mostly because the the LinkedIn learning connection. So naturally it just felt very natural to say, hey, here's something I did or whatever. But the reason I think I'm on LinkedIn almost every day is because I like consuming it. There's just a lot of good stuff on there. If I spend 30 or 45 minutes on LinkedIn, part of me wants to say, Oh gosh, I got to get started with my day. But I usually don't look back on that half hour or more and feel that it was time wasted, you know? So I think it's because I'm constantly consuming stuff that I feel that that's that's the place that I want to be. That's my $0.02 on that. Harpreet: [00:13:28] Thank you very much, Keith. I'm curious, man. So let's. Let's go to the event. What? Back when social media first popped off. I mean, I know for us back then, it was like. Like Friendster, MySpace, things like that. But what was what was your social media kind of drug of choice? And how do you think your life would be different if you chose a different social media? Speaker3: [00:13:57] I think I had the weirdest path into social media because I [00:14:00] started on social media to do professional stuff like I never got I didn't do the social part of it. I got on Twitter originally to just start doing talking about data science, talking about analytics and big data and all of that, because there were a whole bunch of creators on there. I mean, we weren't creators back then. We were. We were called many things, and not all of them were nice. So we were all on there talking about big data and trying to get in front of all of the just BS that was out there. And that's kind of what jumped me on to Twitter was being able to refute some of the garbage that was out there. And so I got really used to. All I need is this short 128 character, you know, and then here's an article underneath it. And that's really what it was. It was sharing other people's content and adding just a little bit of a blurb to Here's what you needed to know, here's a little bit of context that you needed. And that was my relationship. That was the first relationship I had with social media, was using it professionally and then started looking at it as a personal connection tool. Speaker3: [00:15:09] I tried Facebook and I said, This is dumb. I don't want this. I didn't want anything to do with it. But now I have so many personal connections, friends that I know, and the only way we would have ever met was with social media. So I went backwards. I still don't have a Facebook account, but I use Twitter like I actually have friends on Twitter. I have a couple of alt accounts so I can be myself and be human. And I got introduced to Reddit four years ago. I think Reddit's awesome. I think that's the future of social media. When you look at things like Reddit, I think there's going to be some sort of an evolution of Reddit to something that's more. I don't know, not team, but I mean, more structured than Reddit is right now. And I think that's where we go with social media because it's so much, it's so much richer, the communities [00:16:00] are so much deeper. The connections that are made are so much better. And I think that's the that's the cool thing about where social media is going now is we literally went to the absolute worst place possible on social media. Speaker3: [00:16:13] And people realize they didn't like it for the most part. And then now we're going and figuring out what's good social media. What's healthy social media. It would. Which is kind of interesting because when we first had the Internet, everyone was talking about this thing that was going to connect people. And it wasn't until MySpace showed up and then Facebook and then Fark. I'm old. Fark was a website that was like a message board, free form, message board, troll of like redditors. The old redditors all started there. That's where they learned how to be. Redditors is on Fark and on other types of message boards like that. And you know, I think we've finally gotten to the point where we understand how to use it to connect each other. And we're not just using it to advertise to each other and we're not talking at each other. So I think that's where we're going. And it's really interesting from a career standpoint, I don't think anyone's career will be the same in 2 to 3 years as the type of career that I had before social media came out. I just I don't think it's possible anymore. Harpreet: [00:17:22] Yeah. I'd like to see more just niche communities where people who are just interested, you know, like, like if you think about 10,000, 50,000 people like that sounds like a lot of people, but that's such a small fraction of humanity. But if you get like ten, 50,000 people together, all interested in a small topic, imagine the wonderful conversations and just joy they'll be able to bring into their lives from that. So we got a couple of hands raised up. We got a cute up question from Gina that we'll get to Keith. There's a question waiting for you on LinkedIn that I want you to start noodling about also,

man, like, if anybody has any insight to share about the Ethereum merge, [00:18:00] I would love to hear about that. If anybody has any insight, please do let me know. But let's go. In the order of hands raised, we'll go to Serge than Mark than co sub. Then we'll hit Gina's question. Then we'll hit the question from Reem, a UBI that she has for Keith and then Will, hopefully somebody can enlighten you on what the hell the Ethereum merger is all about. Go for it, Serge. Hi. Speaker2: [00:18:24] Yeah yeah my social media is also history is I'm I've always been late adopter to like the major sites you see today. You know even even if I joined early I didn't become active until later each one of them because I'm kind of apathetic to the idea. And strangely enough LinkedIn I opened my account like, I don't know, like 12 years ago or something like that, a long time ago, but I didn't become really active on it until like four years ago. So it's just one of those cases. But LinkedIn is actually the closest, I think, to what? My best kind of. Engagement with social kind of sites are because early on if you wanted to do social media kind of things, you had chat rooms, you had forums and you had, you know. Other kinds of sites like that, I fail to remember, but there were just a ton of sites where you could engage with people that formed part of a community. And oh, blogs as well. You do comment on blogs. So it was like really disconnected. It wasn't all in one place. And actually, I built a social media site back then, you know, before it was kind of a term. It had a lot of the precursor stuff that you would expect from a social media site. It was [00:20:00] for electronic music events. And and the thing is, whenever social media sites, as they exist today, were created like general purpose, things that were not about communities but about connecting like high school classmates and like, you know, your parents and their family, your families and everything, you know, the entire community. That kind of sucked because like to me, like the discussions that I would have informs those are like kind of discussions I would have, you know, with classmates from high school and, you know, my cousins, it's just not necessarily the case, but like LinkedIn is like the closest. I've seen in social media sites, too, that kind of discussions. I would have. Back in the day. And that's why I've come to like it. Harpreet: [00:20:57] Yeah, it is the I can literally attribute everything good that has happened in my life in the last four years to LinkedIn the time that I've put into the platform like I've literally or 4 to 5 X my salary since 2018 I have met so many cool people, started doing all this crazy stuff. I wouldn't have been able to do it if it wasn't for this, for this platform that you started. But I remember like this one. A form that I was a part of. I think this is back in 2008, 2007, 2008, when maybe 2009 when I was like studying to be an actuary. It was called an actuarial outpost. And that's where I used to hang out. Whenever I get stuck on, you know, problems, practice problems would be working through it async. Think that was a good experience. So thank you very much. Let's go to Mark. Let's go to coast up then after Mark and Coast, go to Gina for her question on a Pre-trained model that after Gina's question on the pre train model, we're going to go to reems question that she has for Keith that's talking about TLC [00:22:00] compared with covariance based SEM and Bayesian networks. That sounds like a dope ask causal inference type of type of conversation. So I'm here for it. Mark, go for it. Speaker2: [00:22:12] So I think the question was like, was the first time social media kind of got on to for me it was MySpace. Who are your top five friends? Very important decisions to signal out to your network at that time. And then Facebook became the main thing for me in high school and early, early college. But one thing I think hasn't come out in this conversation yet and I feel like I completely missed out on and I wish I did more was gaming like online gaming, like people playing Call of Duty or Halo or those different things, the amount of close ties, very close relationships people made during that I completely missed out on because I was like, I was just like too socially anxious to like, talk to people online like that. And so, but I remember like, I'm ready, I hear people talk about like, yeah, I played this game for five years straight every single day and we finally met in person. And I'm like best of friends, right? And so I think there's something really special to that where even those that this mass reach kind of like Facebook showing up and meeting people from across the world and doing a shared task that's really fun to do. There's something really powerful in that. I think that's something you might take for granted, like how crazy that is. That like we had a time span even now where you can literally link up with anyone in the world and play games and form these deep relationships and not even see their face. It's just a character and listening to them in your headphones, maybe cussing at them. And so I think that's something really cool that I completely missed out on that I wish I got more into. I guess I could get into it now, but I just don't play video games now. But I think it's so cool. Maybe Carly might be a great person to talk to about that because she's a call of duty. Harpreet: [00:23:52] Yeah, that that was one of the first games where I started playing and building community like you're describing. Again, this is back probably 2009 ten, the [00:24:00] original Call of Duty, Modern Warfare two. We had like clans have like this clan. And then I remember once they had a clan that was the same name as another clan that had to go battle a leader to, to claim ownership of the clan. Actual tag. Yeah, I was wild bad back in the days. Mark, thank you so much. Let's go to a

coast to. Speaker5: [00:24:21] Right. Let me cut this off with the most cynical thing I have ever said on one of your podcasts. Harpreet: [00:24:28] You also said. Speaker5: [00:24:28] The last while I know, I know, I'm just going to pile it on and I'm going to talk myself each time. Right. So Oscar Wilde matters least himself. When he talks in his own person. Give him a mask and he will tell you the truth. The fact is, social media is society's mask, right? That's why you've got anonymity. That's why you've got read it as how red it is. Right. We're slowly lifting that mask because we're connecting everything together. We're connecting different platforms together, different the same login and ten different places now. So you can kind of start to lift the mask if you look deep enough. But yeah, still got the opportunities for a master. Harpreet: [00:25:14] Coast up. Thank you so much. All right. Let's go to a great discussion. Thank you so much. I just love how the conversation question is more of one answer. And it's just, you know, one question and answer trails and stuff. It's amazing. Thank you. Thank you all. Let's go to Gina's question. Gina had a question about pre-trained models for a particular use case that she has. Gina. Go for it. Speaker4: [00:25:38] Thanks. Yeah. So I am working actually with a nonprofit coding organization and we are using the starting off with the taco data set to track annotations in context. And the idea is to use computer [00:26:00] vision models to help identify trash in images. And some of you may have heard of Glitterati, and I believe that some of what they do, but this use case is completely open source giving people the ability to upload batches of photos. The actually the I can tell you I can put it in the chat trash. I you can go check it out if you would like. I think it's trash dot org. I see something in the chat. Go onlyfans account. Okay. That's I won't go into that there. Anyway, so through a chain of events. So. So this group kind of came about, they came to this project trying to solve this problem of how do we identify what trash is reported by just users, people, the general public and where it's located here in California, Caltrans, one of the sponsors of this is his day job is with Caltrans and has to do with there's a lot of money now, thank goodness, being allocated towards beautification projects in California, especially during the pandemic. It seems like litter and trash just got so much worse. And already in their first year, right, they've collected over a million cubic yards of trash across the state. And I don't know if that's over and above what they typically do, but in any case, a tool like this, like trash, A.I. or something similar, could be very useful for municipalities, for community organizations, anybody who's interested in identifying where trash is located and litter [00:28:00] hotspots and all the rest. Now, in their efforts to kind of get this thing kicked off, they found on Kaggle that I think there was a challenge at one point. Speaker4: [00:28:11] And so some people had used YOLO v five as a pre-trained model to and they use the taco dataset to train the model and come up with various annotations. And what we're finding is that the accuracy isn't great even if you run it on lots and lots of epochs. I should also add the taco dataset. I think it's a total of 3000 images. I'm not 100% sure it might only be 1500, so it's not a lot. So another down the road use case is to allow users to upload photos, annotate them, etc., etc.. So with that preamble, my question is and think about what a photo with litter might be. It might be the height of a person taking the photo five feet up or whatever that is, and the litter is pretty prominent in there. Or it could be a shot where litter is 10 to 20 feet off in the distance. One eventually could imagine aerial photos. So then in those cases the litter would be a much smaller portion of a photo. And so I'm looking for any advice people have of other PRE-TRAINED models or even YOLO v5i think. Is it roadblocks? I can't remember who they're working with to do to improve upon yolo v five or as some sort of an additional tool to use with it. And so I see Costa put something in the comments there, but I guess more broadly, I don't know if if certain models that are out there like [00:30:00] Resnick or what's the other one now I'm forgetting one of the other big name models, if those are general purpose, even for something like this, or are there particular models or approaches that. People would recommend for this particular type of problem. Harpreet: [00:30:19] Let's go to Costa resident, computer vision expert and then search if you're around, if you want to chime in on this, I'd be happy to hear from you. Anybody else who's got experience with computer vision, please do. Feel free to just raise your hand to come into the queue coast. Go for it if you're still here. Yes, you are. Speaker5: [00:30:34] Yeah I am. So been a total deer in headlights day for me. So. Yeah. So I guess you've got 3000. You got 3000 images, you got a variety of representations of what you're looking for. Right. The first thing really that I look at is how do I go about and it's a bit harder with images, right? Is how do I go about categorizing? What kind of images do I have? You said distance is one of the things I'd scale is one of the things. Can I get some idea of how big? What's a distribution of tiny things off in the distance versus big things up close? Right. Understanding that is going to help you try to improve the dataset either through augmentation or through whatever other means. Right. So how can you understand that? First, understand the nature of what's in the dataset, what's not in the dataset that you might be missing? And then are there ways that you can augment and improve your dataset so that you can cover for that? You might only have a few examples of things off in the

distance that are clear, right? Or what if all the trash in that system is you're going to go 3000 pictures? I can't imagine that's a huge deal of variety unless you're literally pointing to a garbage dump. Right. So, yeah, it's really about how can you take what you have and improve that dataset to maybe [00:32:00] synthetically cut and paste cut mix, all sorts of different approaches in computer vision, right? See if you can build up datasets that are somewhat realistic to make that a bit bigger or a bit more balanced in terms of the situations you're going to see, at least. That's a starting point, I guess. You've got a number of different models that you can try, but really, I mean, yeah, like, I mean, whether you're talking to Electron, whether you're talking to Resonant, whether you're talking YOLO, V five or any of the YOLO family or any of the YOLO, we are not actually part of the family YOLO models. Speaker4: [00:32:37] I do not know all of these. Speaker5: [00:32:39] So that's YOLO. There's YOLO, which is the Joe Redmond one. Right. And YOLO v two is also part of the family. I think YOLO v three was the first ultra lyrics one if I'm remembering right. Harpreet: [00:32:56] Yeah. Speaker5: [00:32:57] Right. Yeah. Yeah. And then YOLO v fall was then part of the original family again and then YOLO v five is kind of probably one of the better, better performing ones out of the out of the whole lot. But that's again ultra lyrics and it's really easy to serve. It's really easy to pull down from Tor.com they made the usability for that extremely good. There's now V six and V seven that have come from totally different people and don't necessarily actually use the same structure as YOLO, so they might perform differently. There's plenty that you could do to experiment with different, different models, but really start with your data set. Start with exploring how you can make that more representative of what you're likely to see in the real world. The other side of it is essentially. What what can you do in terms of learning features from pre-trained models? Right. Can you pre-trained instead of training [00:34:00] a model from scratch? Obviously, if you get something that's been pre-trained on the whole ImageNet dataset that at least gives you some idea of a baseline of general objects, gives you baseline of general weight. Where do you start to retrain? Where do you start to retrain from? Do you just retrain the whole thing on top of it? Or are you saying, okay, I'm going to freeze the weights of this layer and then continue to retrain up to a particular layer? Experimentation that you can do with that and that will improve or make your performance worse. Speaker5: [00:34:34] The other thing you can do is if you find that it's I mean, this comes to experimentation, but if you find that up to a certain layer is really good at developing the features that you need, but then after that, it starts looking at too many cats and dogs, then maybe

you can essentially retrain entirely from that layer onward. So there's a bunch of different approaches you can use. It depends how long you have, depends how much money you have for GPU compute and depends on really how accurate do you need it to be, right? Remember, the nature of trash is transient. So really, as long as you can tell someone that, hey, there is some trash within this, you know, ten foot radius over here, you could quite literally turn around and say, hey, go send someone or something, a robot to go pick it up. So it doesn't necessarily need to be a detector. The other option is a trash or no trash classifier. Right. What did you do? Something like that. The trick is you need to get your patch sizes of that image small enough that you're not worrying about overfitting to things like, Oh, there's sidewalk, there's a lot of trash is next to sidewalk, right? So there's considerations like that in the real world that you've got to think of. But maybe that's an alternate [00:36:00] approach is like my question is, is transportation at a distance actually useful to anyone at all? Right. Speaker5: [00:36:08] It just becomes a fundamental question. What are you trying to do with it? In my mind, either you're saying at a distance, hey, in this ten foot radius, there is some trash here, so I'm gonna deal with it. That would be my first thing and say, Hey, there is trash, there's no trash. Then you send a robot along that gets up close and personal, gets into that ten foot radius and starts scanning the area. Then you're within five feet, four feet, and you're basically looking at how do I detect it? And at that point, your perspective is no longer top down. It's going to be front on for two reasons. One, you be close to the ground anyway to pick up the trash. Drones are unstable closer to the ground. So you're probably not going to be wanting to use a drone for that anyway. Second thing is drones have really, really powerful rotors. So even if you did get a drone to come down and try and pick the trash up 90% of the time, the trash can get blown away because you got rotors. Right. So from a practicality standpoint, you're probably sending something on wheels, right? So you can assume that the perspective shift is going to happen. So I don't know. I would approach the problem a little bit differently, but I guess that doesn't really answer your question in terms of learning about this dataset and all that, but. Speaker4: [00:37:12] Well, no, that's really helpful. So there are several suggestions and I might have to listen to the podcast again to capture all of them and also make Echo put in the comments about data augmentation. Yes, absolutely. And I think I'm fairly new to this group. It's all volunteers. There's I don't think there's much budget to speak of. And up until recently, I think primarily they've had coders and one coder has just done an amazing job of doing the front end and back end to get the things so that it can upload

batch photos and, and basically does it with browser caching. And so they, they are managing to do very well, not burdening the, I [00:38:00] guess, S3 bucket. So that's fantastic. Speaker5: [00:38:06] Is this the context of a this this in the context of like a personal project? Or is this in the context of a work project that we're trying to work on? Speaker4: [00:38:14] Yeah, it's a work project, but I'm a volunteer with this group, but I just think it's really awesome. Yeah. So it could so imagine one use case is Caltrans. So our Department of Transportation or municipality or county trying to identify where is their trash, what kind of trash is it? So A, they can go out and pick it up more efficiently and B they can identify what's what is that content. And so how might they target, let's say, public education programs and the like. So to your point, while you may possibly have drones potentially going over rights of way taking images, it would be picked up in in probably as far as the bulk of the benefit by the highway crews and the like. But you could also imagine, in fact, there's a group sponsoring it. And I don't know if they're actually giving a certain amount of funds or just brought it to the group to work on. But they're seeking to reduce plastic pollution. And so it wouldn't only be a potential use case for, let's say, Caltrans. So that's just a bit of an overview. Harpreet: [00:39:30] Go for it. Speaker5: [00:39:31] Yeah. That makes a lot of sense. Look, and there's another complication there is that drones within areas near transportation and things like that is quite difficult from a legal perspective. Right. So it's quite unlikely that you'd be using that anyway. You're probably looking at traffic cameras, you're probably looking at cameras on other mobile equipment that already exists within the system, on trains and things like that, purely because you can't fly that close to infrastructure and people, right? Actually, I don't know what the laws are in the USA, but that's what it's [00:40:00] like in Australia, right? Pretty sure most countries just copied CASA's laws anyway. So Australia was so in this like everything we can go saying the chart is just so solid in terms of like creating synthetic synthetic data that documentation and things like that, right? That is the cheaper way of bulking up your data set fast as opposed to annotating new images because that means having to collect new images in the first place. Very, very expensive way to do this. Annotations can cost you both from a platform like how much you're paying for an efficient platform, as well as finding cheap laborers and finding good laborers that actually get honestly get it done the way you do it. So yeah, I think Mexico's got some great ideas in there. I think we want I want to hear from her, too. Harpreet: [00:40:54] Mexico. If you are there, do shut out. You know, I'm also curious, this is actually a question that was asked, you know, did an AMA session last week with might have been two weeks ago. I can't remember but it was with that some of the experts over at NSC and then Matt McFarlane, who's who who is asking the question there. I thought it was interesting question. He he's asking about the use of like, you know, things like stable diffusion to generate training data. Would that work? What would that be like? You know, instead of getting real synthetic data, we can type in a prompt and generate variations of the type of data that we want. I thought that was an interesting question. But Michael, if you're a. Yes. There you are. Go for it. You have some good comments. Talk to us about your comments. Speaker4: [00:41:42] Yeah, I guess I was just saying that in terms of like. So like Andrew Yang, he had this talk like two years ago about like data centric AI versus like model centric AI. And I thought it was interesting because a lot of times, like our first sort of and it's not like our wrong intuition [00:42:00] necessarily, but our first like I think steps is data scientists or ML engineers a lot of times is change the model. Well, it's like do feature engineering and then change the model architecture. But his talk was kind of like getting at the fact that like a lot of times there's just things you can do to like additionally add to your dataset that will like get you far greater returns as opposed to changing like the model architecture itself. So like the two suggestions that like I've been given when working with computer vision projects is one, try to do like data augmentation. So for example, if you have like images of like objects or items like try to deform them in some way or change like rotate them, change the shape, all that good stuff. So like from a single image that's already like annotating all that, you could potentially get like ten like images out of there, maybe even do like a dirty filter on it where it's like, you know, like little. Instagram, yellow snapshot vignette style or what have you. So there's a lot of things ways you can kind of add there. Another option is using like weak labeling. There's like a tool snorkel. Speaker4: [00:43:17] Ais very famous for that. There's other ones, too. And then another option is also to either scour, well, connect with like companies, for example, like Unsplash. So whenever you like, look for an image to add to your medium article or what have you. Most of those images, they would have actually already had the meta meta tags in there. So there are also companies that offer those sort of services where if they provide you with the metadata tag, that could be another way. And then another option as well is honestly, Kaggle is great. Like people are constantly putting up their own like data sets, [00:44:00] like the ones I've used for some fashion recommender projects, like they have like five or six of them where it's like thousands of images each. So you can find ways to like find datasets of like other objects that you

can then sort of like co-opt to add there. So I would, yeah, like I would explore those options because if you think about like a model architecture, you could almost just do like a very quick swap and swap out and you'll see that like. Like other mall architectures, they might only get you an incremental like one or 2% lift and that might not be good enough, right, for the KPI for how the mall operates. So those are just like some suggestions. I always like not spending or giving money to AWS or any of the cloud vendors, so I'd be cheap like that, but I would try those out. Harpreet: [00:44:57] Michiko, thank you very much. I appreciate that. Let's move on with the question. Hopefully, Gina, that has some good answers there. On the question from Reem, I see Reem joining us from LinkedIn. She had a question for Keith and I'm sure Veon might be interested in this question to give me both the guys an interest and causal inference. Reem, how are you doing? Speaker4: [00:45:14] Good, thank you. Can you hear. Harpreet: [00:45:16] Me? Yep. Loud and clear. You have a question? That was a good question. I'd love to hear more about it. Go for it. Speaker4: [00:45:23] Sure. I'd like to give you a little background. So I'm switching from academia to industry and trying to see where I fit. And so I've been taking Keith's courses to understand the language and how things translate. And so that's been super interesting, great courses. And I started to wonder how some tools that I use to relate to Keith's material, [00:46:00] that's his teaching because we use statistics, but machine learning has a slightly different set of vocabulary. And where is it the same or is it different? And so let's, let's see. So for my dissertation and also before that I've been using a structural equation modeling using smart plus is the tool, but the approach is partial least squares. So the reason I went with that is because of modeling the latent variables as formative. And in other fields a formative variable is called causal sometimes index maybe, but with the index I think everything gets thrown in there. Whereas when it's a formative construct, you see each, each one of the indicators separately modeled separately. So anyways, I'm trying to figure out, is it considered good enough to understand, sorry to call it causal inference in the sense of Judea pearls approach, but then I discovered the Bayesian networks and I think that's a super interesting approach to try as well with the data. So basically my question is how is it related? Is it used? Have you seen it used in industry and can one [00:48:00] call it causal inference? Does that am I summarizing things as well enough? Speaker3: [00:48:06] Well, there's a well, there's a number of interesting issues. So I think what's let me see if I'm clear on the big picture. One of the big picture questions is as you shift from academia to industry, would would the two would those two tribes, so to speak, have somewhat different preferences, perhaps, and what they find persuasive? You know, I think that's that's also kind of in your question. Right. And then I think that I think you also have a more specific question about given a particular project that you're working on now, which one possibly is the better fit to the problem seem or Bayesian, right. Speaker4: [00:48:46] Yeah. Speaker3: [00:48:47] So let me see the if I can't give the best background context in about 45 seconds for a minute here. So the, the course that ream is is talking about and of course thanks for checking out the course started as what was originally going to be is I get I've always been intrigued with some software called Baja Lab which is software for doing Bayesian networks. So I said, well, you know, I'm going to try to figure out how to put a course together. And part of it was and a lot of us as content producers have done this, it's like, okay, if I if I draw a line in the sand and I put on my calendar that I'm going to teach how to use the software, I've got to get up to speed on it too. Right? So we've we've all done that. So that was part of the motivation. So then I sat down to say, okay, what would people need to know before they really get patient networks? You know, what would they need to know to know that and what would they need to know? What would be the prerequisite to that? And working my way back, well, the list was was a lot longer than I thought it was going to be. So it ended up becoming like a two part course. The next thing you know, the course that you found is basically a five hour prerequisite. Speaker3: [00:49:59] List of list [00:50:00] of prerequisite topics to get people ready for vision networks. Right. So given that. The basic class of problems we're talking about is a way to systematically say something like A, predicts B, predicts C, and hopefully in a way that you could convince somebody there's some causality going on. So in terms of the two tribes question, I would say that for the academic audience probably seem is going to be the more persuasive because, you know, academics love their P values. So. So you're going to have goodness of fit tests and things like that or just more consistent with what they're used to. Now, I will say up front here, claim some ignorance on the difference between partial least squares scheme in this context and regular SEM. I'm familiar with partial squares and I'm familiar with a, but doing PLSA is something I haven't tried. Right. So but I don't think we have to go down that rabbit hole to address the question in terms of which would be more persuasive. I think for academics, probably the scheme would be more so. The other thing that I would say is from the brief description that I saw, it seems like you have latent variables. And of course by latent variables we mean something that you're not measuring directly. So you might ask somebody five questions and you'll know this screen, but to help everybody. To

understand the context here. Speaker3: [00:51:25] If they're not familiar with it, you might ask a battery of questions that are trying to get anxiety. You know, and you might say something like, if someone's experiencing anxiety at work, are they more likely to have more turnover? Are they more likely to leave that job and go to another job or something like that? Right. So in the same model, you have this variable anxiety, but it's not measured directly. It's measured indirectly by the battery of [00:52:00] questions. So it's kind of like factor analysis and regression in a blender. So that's what's going on with this. But again, it has a lot of these elements that academics like on the Bayesian network side. I think what's interesting about that option is I think in industry, I think a number of you would probably agree basement's kind of hot right now and I think it would be hot in some academic circles, too, but it's increasingly embraced. I think it's going through kind of a renaissance. And the other thing I would say about that is that JDA Perl is big on this idea of counterfactuals. What if you made a different decision? One of the example I have in the course is the prior recall is what if we didn't buy the Super Bowl ad? What would our revenue have been if we didn't buy the Super Bowl ad? Right. That's a really nice trick, which I think is persuasive to what management wants to see. Speaker3: [00:52:55] Because they're all about making. Should we have chosen a or should we have chosen be? Give me a sense of what my what the what the road not taken would have been. I think the road not taken argument is more persuasive to management and industrial setting in an industry setting. Then P values would be especially when it's the weird P values you see in SCM which are you want it to be above 0.05 because if it's below 0.05, you reject the, the null hypothesis that it that it fits and the whole the whole craziness of of that, you know, so, so my initial reaction would be is that because you have latent variables, I think going with Sam was probably the right thing to do. And since you were an academic setting at the time, I doubly think that it was the right thing to do. I don't know whether or not you have a scenario that could work with a counterfactual, but if you did, that's why I would be intrigued. To see what Bayesian networks might be able to do for you is if you could articulate a counterfactual that you thought [00:54:00] was interesting, but the latent traits make me think you were better off in Sam. And just one final comment. The reason that these two seem to be in this cage match at the end of the. Speaker2: [00:54:09] Course. Speaker3: [00:54:10] Is that at the end of all this time of meeting, all these prerequisites, there was only so much time. So I figured, well, what we're going to be what are the ones that that I can explain in a relatively briefly. That would seem like two interesting options that people might want

to pursue further. So I did a very basic SEM example on a in a very basic Bayesian network. I mean, these are very short examples, but some others have commented that they enjoyed the course, but they said, Wow, I hope you weren't trying to imply that these were the only two choices, you know, and they're not, of course. Right. There's there's a bunch. So but I would love to hear what others have to say about all that. Harpreet: [00:54:50] You've got to make time to go to that course of yours, at least skim through and come up with some questions on one side. Once I do that, we're going to have a deeper podcast episode conversation on that. Then any insights here? I know there's some that you're big on or you know, or many follow up questions or anything. Speaker4: [00:55:08] Yeah, I would mention that with plus ACM, the p value thing is like normal. If you get a p value below 0.05, it is a good fit. And so it works differently from the kind of that it's called covariance based. So that's how we distinguish between the two covariance based SEM as the one you mentioned in your course. And it has a lot more restrictions and assumptions that should be respected than plus ACM. Plus ACM is good with small size samples. Not that small, but it doesn't [00:56:00] require a huge sample size and it's very forgiving in terms of assumptions and normality and that kind of thing. And it is it allows a different way of modeling these latent constructs. So it's pretty practical in that sense. Speaker3: [00:56:21] But you reminded me another question I think was what would folks be convinced of causality from PLSA as opposed to regular covariance based SEM? I don't I can't think of I can't think of any reason that somebody that's comfortable with seem would be persuaded by covariance based SIEM and not pls SIEM. And I feel comfortable saying that even though I haven't played around with PLSA. Right. However, there are some folks that just in general don't think you can prove causality with anything. You know, and if you're an SM fan and you run into that kind of critique, you know, that someone says that you're going beyond what your data supports because they seem to be making this. You can never prove causality argument. Then I really like the article that JDA Pearl wrote with Ken Bolan. If I'm remembering right, it's called something like Seven Myths. About Causality and Sam or something like that. I said it in the course, but that that would be one that I would if anyone's familiar with those two topics, I think you would enjoy that that article. It's basically saying what what kind of critiques do people make about SM in general and how they would answer those critiques? So again, it's some kind of seven, I think it's 7 minutes could be eight myths, but it's myths about Myths About Sam and Causality by Pearl and Bolan. Speaker4: [00:57:50] I'm looking forward to reading that article. Just downloaded it. Do I have another minute to say something about causality? Harpreet: [00:57:59] Yes, absolutely. [00:58:00] And and the events also to this stuff, too. So then if you want to chime in on the discussion, do let me know and go for it. Speaker4: [00:58:08] Okay, great. So if if we want to prove causality, there are a few things that you went over at the beginning of the course. One is randomness. Another is the time sequence of events. Right. And then another is help me out if I'm blanking out a bit here. But so the experimental design, so that's a factorial design, for example, which is which comes from experimentation and so on. So you can set these things up with the research design to get at causality with them. So depending on when you collect your data, how you collect your data, when you put your interventions into the the process. All of these things can come together. To. Develop a causal model. It's like you said, not just the data speaking for itself, but the whole logic and the design behind the research study or the analysis that we're doing. It all comes together to create a causal argument. So that's what I wanted to say about that. Thank you very much and thank you so much for inviting me to come on and have this discussion. I really appreciate it. My research was on social media, by the way, and the impact of social media on well-being. So I could always explore that sometimes. Speaker3: [00:59:52] Could you give could you give a room? Could you give everybody one example of a latent trait in your research just so that they can kind of picture. Speaker4: [01:00:00] Sure. [01:00:00] Well, I can give you a couple. Very briefly. So, for example, perceived usefulness of software. My research was about social media. I asked about Facebook specifically, and it had nothing to do with the company Facebook. So do you find Facebook useful? And there is a validated measure that measures perceived usefulness from the technology acceptance model. And then one that I developed myself for the study I had too was perceived Facebook functionality. So what do people perceive as a functionality of Facebook? And so I had to do a whole qualitative portion of the study based on literature and then interviews and developed items for that. And I had many, many items that split into different dimensions. And so, for example, can you I can make friends on Facebook and it would be on a scale I agree to. I disagree. I can meet someone I've never met before on Facebook again, like scale. I agree to disagree, that kind of thing. So all of those can be factored, analyzed, put grouped into latent traits or latent variables and then put into the structural equation model that way. Does that answer the question? Speaker3: [01:01:39] Yeah. No, I just wanted to give a give everybody a little bit of a flavor how you might have multiple items that together are one latent trait. Speaker4: [01:01:47] Yeah. Yeah. And we would have an idea of what, what sort of factor structure it will sort into. But the factor

analysis, the exploratory [01:02:00] phase would put them into those factors using a machine learning algorithm. And then the second part of the study with a new sample would confirm whether that factor structure stays or if it's not valid. And we go back to the drawing board. Harpreet: [01:02:25] Thank you very much. Being an insider or comments here, I know there's something that heavily about on his newsletter stream. If you're interested in more, definitely check out Vince's newsletter. Speaker3: [01:02:38] I would love. I would just say, Oh, sorry, good. I'm just going to say I would love to hear more comments on it. I would really enjoy that. Yeah, I think what's because Keith pretty much covered more than I know completely about Castle, about the question that you ask. But what's interesting about causal is no matter what you say, no matter what answer someone would give to your question, someone else would disagree. And you could have a complete discussion where both people could very easily support themselves being right, which is where we are in Castle right now, especially when you talk about combining it with machine learning approaches. We're not in a place where we have certainty. We're in a place where a question like yours is almost philosophical in some cases, because, like I said, we're when I see Judea Pearl arguing with multiple other people that are smarter than I'll ever be, and they are both making legitimate points and they're both you know, it's that's the place that we're at. And the reason why I wanted to throw in a comment is because when you get into the business setting, no one in machine learning really has a strong grasp on data on the data science connection to causal. It's [01:04:00] something that's being worked out. You have companies that are finally publishing frameworks like Microsoft and Amazon have great frameworks that they've just recently published, a DUI framework where or it's got a new name now pi y sorry, not DUI, pi y where they're collaborating on it together. Speaker3: [01:04:18] And Amazon came at it from one completely different direction than Microsoft came at it, but they are complementary to each other. And you could see from their work and if you go to Microsoft's website on pi y there, it's worth reading because you'll hear Amazon and Microsoft's approach is contrast to each other as far as the original need. And I think that's what's important to read there. It's probably not going to be a whole lot of information for you, but you're going to hear Amazon needed diagnostics. Amazon needed that diagnostic causal. Portion of it to figure out what was going on with something as complex as us. If we make this change, what would the implications be? That's kind of important, but for them it's more something slowing down. Why? What went wrong? And so they were looking at it as a diagnostic use case, and then they would provide that information to smart people, engineers who could solve the problem, and then the engineers would pass their solution back to the model. And the model would say, yes, you probably will solve this problem. Or No, we don't. You know, there may be other issues and you can see what they're doing is they've got a smart person who can make that decision, who can figure out what the fix should be. But the complexity is really what went wrong and what am I going to do or what am I going to do to the system when I introduce the fix? And so I think those are the things that are going to be interesting for you since you're making the transition in, because in academia you're really asked to do something so much [01:06:00] more rigorous than we are in the business world. Speaker3: [01:06:04] Because when you look at the way that it us is implemented there causal framework, it would get torn to pieces in an academic setting. But for that engineer who's got to figure out what went wrong. It is a perfectly legitimate approach because it's mostly right and that's all they need. They'll figure it out from there. And so the types of questions you're asking right now are very, very rigorous, and you will end up backing off from that level of rigor because the business doesn't need it. The business needs are actually a lower threshold of certainty. And so that would be the piece that I would say with causal. You're asking questions that we'll get complete disagreement across three different spectrums, two of them in causal causal research, one of them in the AI version of causal trying to integrate machine learning and do causal discovery and those sorts of methodologies. And so when you ask those questions, be ready, you'll get three different responses and they'll all be very hard to differentiate between who's right and who's wrong. And in the end, the methodology is very important, but that level of certainty isn't so much important. Speaker4: [01:07:26] Thank you. That's very interesting to hear. And this is, I think, where the the exercise is going to be, how to back off the rigor and now to make it into industry and make sense of what I need to do. Harpreet: [01:07:44] Yeah. I think the important thing is like, you know. You're not trying to gain marks on an exam and up the marketing exam, are you're trying to, you know, make some more money. And it's interesting because sometimes [01:08:00] you get the sea level grade to maximize returns. I'm trying to begin analogy is probably not a good one. Anyways, Keith, go for it. Speaker3: [01:08:10] And I couldn't I couldn't agree more with what you were just talking about, you know, particularly the way that you brought up a pearl, you know? You got it. Here's a guy who won the Turing Award on the computer science side, but basically has spent 30 years fighting with academics. That that causality really is like a thing. Right. So, I mean, couldn't have couldn't have said it better. And the whole pi y thing is super interesting. One quick comment for folks that maybe have

never really thought about this stuff and just want a kind of a like an easy entry point. It's not it's not causality. Exactly. But on Google Analytics, they have this section in their help file on why they do a B testing with Bayesian instead of a B testing with P values. And it's only a couple of pages. It's absolutely brilliant. And I think anybody who's in the data could follow that logic. And it really is an example of what Vin was just talking about, you know, because there will literally be questions in the help file, where am I p values? I can't proceed without p values. I'm freaking out. Where am I? P values and it's addressed right in the help file. The reason we don't provide p values is because we're doing this Bayesian, and the reason we've decided to do Bayesian is because it's actually answering what your real question is, which is, is my B better than my A used to be? Right. But again, they walk through it very clearly. It can be an entry level discussion, I think for those of you that are intrigued but just haven't gone down this rabbit hole before. Harpreet: [01:09:56] Keith. Thank you very much. Yeah, I'm excited to, uh, to find [01:10:00] some time to go to that course, and we'll talk about this in a in a podcast episode, for sure. It might be fun to do. Maybe Vin and Keith together with me asking stupid questions because I feel like the audience could gain a lot from my stupidity. By the way, I don't know if I should head it out or not, but I did a episode on my podcast with Dana McKenzie, who is the coauthor of the Book of Why. So definitely check out that episode is Link right there in the chat. I think I just call it like the book of why. So definitely check that out. Dana did offer that I to to introduce me to Judea Pearl, but I was like, ah. I think I've passed up on that. I might have to hit a up and see if that offer is still valid after this conversation with. Speaker3: [01:10:46] Well, I've got to tell you, I've heard a couple of his podcasts. He's an amazing guest. And he is not as scary as you think to interview. He's he's a great guest. Harpreet: [01:10:56] Yeah, I've heard the episode with him and Lex Friedman. I was like, that was such a good get episode. Yeah. Speaking of Lex Friedman, did you know Lex Friedman and I have interviewed two of the same guests, Jon Verbeek and Jordan Ellenberg. So go listen to this episodes side by side. Me versus him, let me know who you think the better podcast is. Pretty sure it is me. Keith, any final statements or anything here? Speaker3: [01:11:19] Well, I was thinking you should invite Lex. Harpreet: [01:11:23] Yeah, I've tried multiple times, actually. Actually, a Lex Freeman put out a a call for like it was just like a, you know, looking for people to work on his podcast. And so I applied, you know, he said he was looking for research assistants for the podcast and I was like, All right, well, research this shit out. My guess, like, I'm amazing at that and I know how

podcast work and a little bit about them, so we'll see if he'll get back to me. One of the things he asked was how much you want per hour to put like an absurd number for a virtual assistant that does research, [01:12:00] but we'll see what happens. But he had this question and the question was, do you do you work hard or do you work smart? Then the follow up question was, explain your answer to the previous question. And so so my answer was, I work hard. And why? Because I'm too dumb to work smart. So I have to work hard. So that's those. Let's see if he buys that. Any of the questions coming in? Do let me know. Don't say anything on YouTube. Scott Taylor, by the way, was in the YouTube chatting it up. Scott Taylor did not forget about you. Hopefully you're doing good, man. It was good hanging out. I hung out with the. Was it earlier this week? Maybe last week. I don't know, man. Weeks just been running together nowadays. But I was hanging out on a call with Scott Taylor. Kate Christina it. Tom Bunch of the people. It was a great time to be hands up. Go for it, my friend. Speaker5: [01:12:58] Yeah. Just quickly on that. Work hard, work smart. I'm too dumb to work. Smart thing. The fact that you know you're too dumb to work. Smart is proof that you're smart enough to realize where you can work smarter. Right. The fact is, like, I'm in a similar boat where I work hard because I'm not the smartest bloke in the room and I like not being the smartest person in the room. If I am, that's the wrong room. But the fact of the matter is, there's a limit to how much effort overcomes things. Right? Because you've got to balance effort with balance. Right. And everything else. So, yeah, work hard, but don't beat yourself up. Harpreet: [01:13:36] And Michael, let's hear from you on this because you put there in the in the chat. Same. So talk to us about working hard versus working smart. Go for it. Speaker4: [01:13:48] This is going to sound so bad and vaguely hypocritical, but I'm a little bit over hustle culture, so I'm investing in the things that bring me joy and that is my meal service, [01:14:00] my quilting table and fabrics, knitting and doing content. So yeah, yeah. It was funny though because I like it's funny because like when I, since I'm like starting the new chapter of my career, right? I was like looking back at all the different twists and turns that like my career has taken. And I remember in high school, like most of my teachers were like, like I was, I was going to be too dumb to like, accomplish anything. I had like a few high school teachers. Like most, my high school teachers loved me, but there were ones who were like, I don't know. I don't know what the deal is. They were so intent on trying to break the spirit of a high school student. And it's so sad that some teachers are like that, you know. But I think it's just I think also to adaptability. It's something that you and I have talked about, right, about playing the long game, the

long game of building a career and a body of work as opposed to playing the short game of like sometimes it's like the short game is trying to get as much money as possible, try and get as many clients as possible, try and get as many projects as possible. But the longer game is building the body of work, building the quality relationships, building the quality community. You know, like marketing yourself at the price you feel you deserve for the value that you bring. And the right people will recognize that and they will come and like be a part of your tribe. So yeah, I mean, I, yeah, I totally get that whole like you have to work hardest or in point, but I'm also a little bit done over the hustle culture of Silicon Valley. So now I want to like work happy. Harpreet: [01:15:39] Now I feel you on that 100%. Probably. Mexico has made a career switching to developer relations. One by one. I will be converting everyone you see here to develop relations because that is what I do. I convert people to my side. But dude, yeah, man. Like. Probably the first time ever. I feel like everything is just lined up. Like everything's lined up, [01:16:00] right. Like. Like the stuff I was doing for fun on the side is now my actual full time job. Like, this is what I do for a living and it is amazing. Like, I spent half my week learning computer vision with no stress. Like, I don't have to worry about people asking me, like, why the fuck you spend time on computer vision? We need to get this done for this business value in that report and all this, that and the other. Nobody fucking cares. Like, yeah, it's been like two days a week learning computer vision, do it. And then I host the events. I write content, you know, strategize community. I bring people together like I'm doing everything that I enjoy doing this stuff that I would do in my spare time. It's just not my full time job. Yeah. Everything's just, like, lined up. It's amazing, right? It's amazing. Speaker5: [01:16:44] Yeah. I got to say, man, that's. I love I love hearing that. That you're doing exactly what you love doing. And that's huge, right? Because one thing that I've learned in the last few years is a well, it's not just doing what you love doing and the job that you think you love because I mean, I've kind of broken multiple points in the last seven or eight years. I've broken down the hey, this is the job that I'm meant to do. This is the job that I love. Breaking through is preconceived notions of, Oh, I need to go do a PhD from MIT, from a top tier tech school, to breaking through all of those barriers of saying these are preconceived notions and not necessarily in line with my, my, my true nature. Right. So getting into those jobs where you're doing that, that's powerful, man. And I love hearing that. That's the space you're in right now. And the other half is the general in general. The fact that we have a formalized role for something like this within the ecosystem of of tech, it just shows that maturity of we're bringing forward the human element of tech development, which for the largest part in the last 40, 50 years, it's just not been around. So it's really good to hear it. Harpreet: [01:17:59] Yeah. And to [01:18:00] that point about just this, like you're talking about, oh, I've got to do this, I'm supposed to do this, supposed to do that. And the other, like I feel like a lot of like, you know, people like online do this to like Justin Walsh, for example, like posting a picture of his calendar with no meetings, no appointments, saying, oh, goals like, like I cool man. Like, so what you're you're making your money teaching people how to make their money online so they can then teach other people to make like, I don't want any part of that anymore. Like I thought I wanted to be a part of that or do something close to that just because all the everyone said I should be doing that, that's what I should be doing. But really like that is not what freedom really is. I think I posted about this the other day. Right. And I said, freedom isn't freedom isn't having, you know, an empty calendar and just complete idleness. That's not what freedom is. And it's having the time and space to think about and work about the problems that you actually want to solve. Like these are things that I want to be doing and I'm getting paid for it, but getting like I made it in life, right? Speaker4: [01:18:58] That was the thing that to me was like really surprising when I transitioned into data science and then and then hopped my way over to like engineering was I thought I would be doing a lot more learning, building and sharing. And it ended up not being that a lot of times it ended up being like fighting with legacy code, sometimes fighting with legacy culture. And a lot of times like. Spinning the wheel on. I'm just really kind of Manute things like, I don't know, like how initiatives are named or whatever. And I ended up taking on more of like, like because like in my head, like devil is basically like you are helping like developers or folks or geoscientists use your tools and you're helping to make the tool like the product or service like better. And you're trying to make people's lives better. And I kind of sort of thought that was what engineering was, but especially like platform engineering. But I realized maybe that was a little bit of a naive [01:20:00] notion that might be like the aspirational intent. But I also think to I don't know. Yeah, it was just it was weird, you know, it's it's weird that like I had to kind of switch to like a non engineering title to finally do the work I thought I was going to be doing like as an engineer. Harpreet: [01:20:17] So yeah, I mean there's, there's still a dark side. Deborah We're still going to have to deal with the same kind of stuff from executives and people like that, but. It's just so much more fun. I don't think itself is so big. It brings me so much joy. It's so fun. Keith, do you

have anything to add here? I'd love to hear from you. Speaker3: [01:20:40] Well, the only I keep on thinking about something in Mexico said a couple of minutes ago about the whole high school teacher thing. I'm curious how many of us had at least one high school teacher that thought we wouldn't amount to anything? For me, it's kind of like a badge of honor. I kind of get inspired by that kind of thing. Harpreet: [01:20:59] All of them, all of my high school teachers that I was I was I was nothing but honors classes in high school. But I was like the bottom 2% because I kept skipping school to like, you know, go do drugs. But I'd show up and I'd take the exam and, you know, do decent on the exam. But there's one teacher in particular. He was my chemistry teacher. Ap Chemistry teacher, I think since 10th grade, ninth grade, maybe ninth or 10th grade. And I was an active student participating in class like, you know, asking questions like, you know, just asking questions that I was just curious about. And he just went off the rails on me. One day. He was like, you sitting here acting like, you know, all these things like that. You're smart, but you know what? What do you see? Your next next exam, the score you got? I'm just sitting there like, what the fuck blowing up on me, man? But he. Yeah, he genuinely. I got killed my love of science for a very long time. It was horrible because [01:22:00] it was here from you. Speaker5: [01:22:04] I mean, teaching is hard and teaching is a tough, tough gig. I just want us to step back and appreciate that for a second. Right. Like having taught at a university and having tutored a bunch of different students who all learn differently. For high school students, it is a tough gig and whatever else we may say, we may rip on them now because we're like, Hey, heads on a pole. Every teacher that told me I wouldn't amount to shit, right? Yeah. Just step back and appreciate it. Right. Like it's a tough gig and dealing with students that just think differently, they just approach things differently. The people on this podcast, let's be honest, those and this is not, we think, better or worse, it's just that we think differently. Right. People from who have that love for data, who love that love for we have we have an inherent love for chaos, for organizing chaos, and for looking into this abyss of things that don't make sense and trying to make sense of it. Whereas most a lot of other people prefer structure and prefer, okay, this is an organized way of working. And we follow those, you know, those rules and protocols that's already established. A lot of us love to dive in and fix the mess. I mean, sure make you go. I mean, I don't know your take on it. From the ML ops standpoint, it feels a lot more solid and probably a lot less chaotic than some of the experimental work out there. Speaker5: [01:23:30] But in reality it's fully chaotic and you're just trying to exactly like shake your head. You're digging your hands into this mud that everyone says, Oh, that's just engineering, right? And you're like, No, this is mud. And we're trying to build structures with mud here. We're trying to build brick houses with mud. So there's this love for chaos. The point I'm trying to make is all of us here, I think we can establish that we have this inherent love for chaos and teachers struggle to deal with that because it's so hard to teach. A class of 30 [01:24:00] kids has one person who's had barely any support, barely any mentoring. We expect young teachers straight out of universities to teach at this extremely high level, you know, and they're getting paid. Jack Shit. Let's be frank. They're getting paid, you know, barely anything compared to what they should be getting paid. And fundamentally. Right. What I found was two kinds of responses from most teachers. One that would essentially turn around and say, Hey, you're smart, but you're lazy, just do the work, right? The others that would just be like, Yeah, he's lazy. He's not very good at this. It's just going to drop the ball. So I'm not going to bother, right? Let him let him drag his feet long through the class and the latter killed my love for science. Speaker5: [01:24:47] I ended up hating biology and chemistry. The only thing that kept me going was like electronics, and I was always kind of obsessed with robots and aircraft. I kind of love that site anyway. So that love kind of overcame the boredom of science, right? But on the flip side, you have those teachers that I can remember maybe three or four teachers that A, they chose to, you know, throw their hands into the mind and try to figure out how do we teach this chaotic kid. Right. And they all had different approaches and it was fantastic. I hated I was good at English, but I hated English lit. And I had one teacher that realized, Hey, this guy doesn't care about literature techniques, he cares about philosophy, right? So he he managed to tailor every single lesson in the 11th grade and 12th grade into a tweak of like he'd talk about the structural lesson that he's supposed to teach, and then he dive into, like, young in psychology or something like that, just to throw in enough there for the two or three of us that were super chaotic in the class and wanted that additional input. And then you had like people that were just challenge you. I didn't [01:26:00] like being told, Hey, I still don't like, Hey, this is the thing we do. This is how we do it. Just follow the protocol. I'm a hi ID personality. Speaker5: [01:26:08] If you've ever done this analysis right. I'm a high ID personality, right? So very, very low compliance, like eight out of 100 kind of thing, right? You don't want me as your QC engineer. The fact is we had teachers that would challenge us in those ways. There was one teacher that literally, like, he would limit me to three questions per lesson, right? We all know that I can talk for days. I can do it right now, but it was like

limit to three questions per day. Right. And that made me have to think hard for, hey, what am I going to ask him? What can I answer for myself with the knowledge that I already have while still keeping up with the conversation? Because a lot of the time that I'm asking him a question, it's not exactly about this. It's because he's given me a seed of information that I've gone off to think about something else. It gives me that habit of writing that down and exploring that separately and coming back to that teaching me was tough. Right. And then I like that competitive nature of it. I'm a little bit competitive at this thing. I'm not gonna name names on which subject because everyone figured that out. But into the thing where there were like eight or nine of us in the class, he basically right down and we were all competitive like we were all that kind of competitive nature and we enjoyed it, which is why it was okay to do. Speaker5: [01:27:25] But a lot of classes may be not right. He put off every exam or assignment or assessment it. Basically we put all the scores. In the class, right on the whiteboard. They just put the scores down, but not our names. Right? Starting at the top. Going right down to the bottom. Right. And then he would ask each of us to come up and put our initials next to the score we thought we got. Right. And it's brutal because everyone in the class knows what you think you got and what you actually got. But it's a great way to teach us to recalibrate [01:28:00] our thinking of where am I actually from an understanding perspective. That was something that was so invaluable to me as a competitive person, as a person who needs to battle ego and humility all the time. One of the best lessons I've ever had, and he understood that that particular class of thinkers operate in that way and enjoy having that feedback to them. Those are some of the best teachers I had were the ones that stuck their hands into it, and it's very hard to find it. So you've got to appreciate, right? Yeah, sure. Heads on a pole. There was a lot of teachers that said I wouldn't amount to anything, but hey, all you need is those two or three that push you through, right? Harpreet: [01:28:39] Yeah, dude, I was actually a high school teacher for. For two years. Speaker5: [01:28:43] This was an you'd be a great high school teacher. Harpreet: [01:28:45] I taught I taught math, and I kind of just did it because, like. For whatever reason, I had no other options. I became a teacher and so I was a teacher for two. But I was teaching that this in California we have something called academic recovery. And there's this kind of a segment of schools or that's called options for youth, opportunities for learning, options for youth, whatever. There's one of those they're both one company and I taught there for two years that options for youth teaching the kids that were kicked out of high school just fresh out of juvie or for whatever reason, we're on academic recovery and at the same time they're in the same school. And by school, the schools were interesting. It was just you imagine a strip mall and like just a little like place inside the strip mall. This is what the school was. This is where I show up to work to teach kids math. But yeah, in the same class I have kids that are trying to catch up, but then kids that were on the other end of the spectrum that are trying to get ahead, get faster because those are your options. Like, let's say you wanted to finish high school quick, you got opportunities for learning, options for youth, and you just take your units. Harpreet: [01:29:52] Everything was unit based, you have a packet, your questions, you turn it in, you get your credit. Very interesting model. But yeah, I was dealing with [01:30:00] some it's dealing with kids that are just like me, man. Like, you know, like most Indian people you see in data science come from very cushy and plush backgrounds and they're all good kids and they all lived in nice neighborhoods with rich parents. That was not me. I grew up broke as buck in the hood and, you know, dealt with struggles that a kid would deal deal with in that environment. And I could just relate to them. I think that's what made me a good teacher. I don't know if I was a good teacher or not, but that's what made me maybe be an impactful teacher being able to help kids. I just understood them. I got them because I was them. But because of that teaching, because of doing that, a little bit of math teaching. I was teaching high school level math and that's when I was like, Holy shit, I'm actually like like pretty good at this. I really enjoy it. And so that's why I decided to go back to grad school and study math, you know, in graduate school and pursue a career in actuarial sciences. Harpreet: [01:30:57] And they wouldn't have had that if I didn't take a pit stop teaching teach math. Great discussion now. Man. I don't see any questions coming out LinkedIn don't see any questions coming on YouTube. I think it's a good time to wrap it up. And as usual, good discussion. I'm excited that you all are here. Good to see a lot of a lot of old friends. A lot of new faces. Hope you new faces swing by those you watch our LinkedIn you know, we had some good guys on LinkedIn hopefully guys enjoyed it. Do let me know. Smash that. Like let me know what's going on. I don't got any announcements on anything, man. Just. Just have a good weekend. Chill out. Enjoy. I'll see you all next time. Russell's got some good comments here as well. Russell saying empathy is one of the strongest skills to be able to connect with people. Russell's saying that we're all as challenging as a child is also often using the eye mapping. Well, talk to me about that real quick, Russell. I haven't heard from you at all, man. So let's [01:32:00] hear from Russell. Then we'll. Then we'll wrap it up. And meantime, I'll keep an eye out for questions and comments in the chat. Speaker2: [01:32:06] Okay. Speaker3: [01:32:07] Thanks. Yes, sir. I triple t if this then that a lot of people in the daily community will be familiar with that philosophy. And as a child, I was fascinated to learn anything that I could. Learn. You know, I had a great appetite for learning. So if a teacher presented. Speaker5: [01:32:23] Something that was really interesting, I would normally go to them with additional questions. Speaker2: [01:32:28] So that's very interesting. If that means that then. Speaker3: [01:32:30] By association there's something else that's connected to it mean this, then I might want to explore the further boundaries of the very fixed lessons that was on the the lesson plans that they had, usually to the frustration of the teachers. And ultimately. Speaker2: [01:32:46] That did result in. Speaker3: [01:32:48] Me being asked to sit in the corner to stop bothering the teachers. Some teachers, as you mentioned, you know, you get good teachers. Speaker5: [01:32:55] Some teachers are very good and tolerated me very well. Speaker3: [01:32:59] And I responded quite well in those. Speaker5: [01:33:01] But in others. Speaker3: [01:33:03] They used me and I used to be able to sit in the corner quite a lot in some of those. So. Yes, I think it depends very much in the institutions and facilities you achieve your education as to your your feelings about those and how they. Speaker5: [01:33:22] Can. Speaker3: [01:33:23] How they can optimize your ability for learning. So I don't think I optimized my ability as a young child. I think I could have learned more faster. But I'm sure that happens to many, many people. So I don't class myself as unique in that respect. Harpreet: [01:33:40] Thank you so much. Thank you all for hanging out. Let's go ahead. Wrap it up. It's been a good good evening with you all or morning depending on what side of the world you're at or late in the evening night. Thank you all for being here. Remember, my friends, you got one life on this planet. Why not do some big cheers, you know?