When it's time to scale your business, it's time for Shopify. Get everything you need to grow the way you want. Like, all the way. Smack more sales with the best converting checkout on the planet.
Track your teachings from every channel right in one spot. And turn real-time reporting into big-time opportunities. Take your business to a whole new level. Switch to Shopify.
Start your free trial today. Evangelism is a role that's coming out and becoming much more popular and important at tech companies. It's really just deeply understanding your product at a technical level and then working with technical people to help them understand it and implement it effectively. Today on the podcast, we have a pleasure of being joined by Andrew Madsen, who is, you know, aside from being a LinkedIn influencer with, like, 80,000 followers, so if you're not following him there already, you've got to, he is a tech evangelist over at Dremio.
His work focuses on bridging the gap between theoretical knowledge and the real-world application. He is huge into data science and a really big figure talking about a lot of things happening that I think are very influential on AI and everything happening in the field of data science and analytics. So welcome to the podcast today, Andrew. Thanks for having me, Dan.
Okay, so super excited to have you on the show. I'd love to, first off, like, start by talking a little bit about what got you into the space, into data analysis or analytics, you know, all of this stuff that's happening now. How did you kind of get started? Yeah, I used to be a finance executive focused on legal and regulatory compliance.
So I really had to dive deep into statistics, which then led me into data analytics, and I just loved it, so then I worked on making it my full-time job. Okay, so had you done, I know, like, with data analytics and analysis and all this kind of stuff, it feels very much more, like, kind of, like, technical almost. I don't know, like, are you a developer? Would you call yourself like a developer, or are you more just kind of working on, yeah?
I was, yeah, I started as a developer learning SQLR, Python, and doing that for several years, and then moved into leadership. But they don't respect you unless you keep those skills up. Yeah. So I still try to stay technical.
That's hilarious and awesome. And actually, something I've heard quite a few people talk about is the challenge of moving from, when you're technical, going into management. But, like, you don't really have a lot of time to be doing the technical stuff. How do you stay up-to-date on stuff?
Like, do you have, like, projects or things you work on or play with, or what do you do to stay? I found it easier because I came from being in leadership before I went into data. So leadership was always easier for me. And when I was learning data, I had to learn at night anyway.
And so going from being an individual contributor to then leading data teams and data organizations was much more natural for me. But I always keep my hands on the keyboard when it comes up to projects and staying up-to-date with those. So the projects really are internal. I don't have to go look outside and figure out, like, oh, here's a new skill, a new library.
Let me figure out a way to implement this. I just make sure that I stay involved at my companies. Okay. That's so interesting.
So you mentioned, you know, you kind of started not exactly in this. You're in finance doing, you know, things there. Moving in, what were some of the biggest challenges you faced? What were the kind of, like, things that you're like, oh, my gosh, this is harder than it looks?
Learning to code was hard for me, but it is just like learning a new language. I had zero experience with it and decided to start with just getting a master's degree in it, which wasn't a great idea because they assume that you have a strong foundation before you start the program. But it's also moving from one industry to another is difficult because I already was coming from a high-paying role and didn't have any experience in data analytics, so I couldn't just switch right over. I couldn't afford at that time to just take, like, an entry-level role.
So I took, like, a mid-level data job at night, and I literally worked two full-time jobs for years. So I was building my second resume, my data resume, to the point where I could get a data job that could fully replace my income. So that was a challenge for a few years, which I don't recommend. But I have had a conversation with several, you know, high-level folks in other industries who are trying to transition into data and do run into that problem of they can't – they don't have the experience to get a data job that can fully replace their income.
How do they do it? Well, this is how I did it. Not recommended, but if you really want to break into something, if you really want to do something, there's going to be sacrifice on the sun. That is a fascinating idea.
I've talked to a lot of people that are, like, you know, I'm sure you get this in data a lot, but, like, with AI, they're, like, you know, like, I really want to get into, like, the AI industry. I want to become, like, a developer or researcher and, like, work on this stuff. But I have, like, no experience. And, like, yeah, I have a job somewhere else, sometimes a great tech company or other places.
And that is such an interesting path. You say you want to recommend doing, like, a side job pretty much to do it. Is that just because, like, it's so time-consuming? It was.
I literally was working from 5 a.m. to midnight every day, six days a week for two years. And I think that's the, if you want to accelerate your path to getting in, I think that's the amount of time it takes to really get in. You can do volunteer projects.
If you just want to be associated with it, 100%, you can just do less work than that. You can volunteer for free. You can get an internship. You can do your little projects.
But if you really want to get in and make it your full career, it does take quite a bit of time and experience. Yeah. So you're doing this kind of, like, part-time thing. You get a bunch of experience.
What was your first hop into, I guess, moving into, like, a full-time role in data? I took over a data team. I can't remember what company I was at. I don't know.
I've been at so many companies. I've switched jobs, like, 20 times. I was looking at your LinkedIn. I was like, oh, my gosh, this guy's getting everywhere.
Yeah, but I took over some data teams. And I took over a small team at first. And that was good because it keeps your hands on the keyboard. Like, there weren't enough people.
There were probably, like, three analysts. And so then I also was having to do full-time work while managing the team. And so it keeps you fresh. It keeps you sharp.
But then I was able to introduce, like, project management and privatization into that organization. A lot of data teams struggle with because they're a limited resource. And then they get into, like, the order-taking role of, okay, here's a project. Let's work on it.
Rather than taking the time to think strategically. Talk to me a little bit about you. You know, you had that job. You've been in companies doing this for a while.
And then I don't know if you switched or I noticed you've done a bunch of stuff with academia. You've worked at a bunch of universities, been in the structure there. What's, like, the differences or what are some of the nuances between, like, doing this in an actual company versus teaching it? Yeah, it depends on the company and on the course.
A lot of the courses are dated. They're definitely not at the forefront of the technology where a lot of companies aren't either, if we're being frank. But most tech companies, the big tech, are at the forefront of the technology. So what you're learning in school, even in grad programs, is really more foundational.
But you're going to have to do more learning outside of that to really excel in a forward tech company. But for 99% of, like, the Fortune 500s, which aren't at the forefront of technology, you'll be just fine. Grad programs really focus on, like, the fundamentals, clustering, regression, transformers, where these days, like, I'm not teaching RAG at all in my courses. But you really do need to, if you want to be an AI, you do need to know how RAGs work and how to build them.
At the very least, their architecture, which isn't being covered in most school programs. So where do people learn that? YouTube. YouTube?
I think if you want to learn new concepts, you should always go back to the initial academic paper. Whoever created the technology and first wrote about it, that's the best place to start learning. That's fascinating. And also, I feel like that's a trend in university and school for, like, literally anything I know.
You know, I study marketing. And I'll say 90% of what I took away from, like, college that I learned about marketing was, like, I'd be doing a class. And it's, like, whatever. But then someone was doing, like, a side project while I was in school.
And I was just using YouTube to learn everything to do with it. And it was coming from the classes. I was, like, a good 80% of it at least, probably. Oh, 100%.
Josh Starmer, who runs StatQuest, he worked at Lightning AI for a couple of years, figuring out iTorch Lightning and how to build all those models. But he will literally on YouTube show you how to build a RAG model. So you can just build it yourself. That's amazing.
What's something that you're most excited about that's coming out that's a new innovation in the field that you're most excited about? I really love where I work at Dremio. We're a data lake house company where you can do analytics and AI just right on the data lake rather than having to port it into a data warehouse. You've got stale data.
It's expensive. All of that. At ASU, when we were kicking off AI with OpenAI, they have copies of the data in the data warehouse. The data quality is fairly poor because there's changes in upstream systems.
It's decentralized data. And now they're trying to build AI on top of it. So that's one of the challenges. And then to build AI on top of it, now they're making another copy.
They were all trying to agree on, okay, what's the data? What does it mean? Let's now build out these tables or these cubes and have the AI train on this. Super costly and very inefficient.
Whereas with a data lake house, you can do the analytics and AI right on the raw data or on virtualizations of that data, which is super important for real-time AI or real-time analytics. And it's also cheaper, which is a huge benefit. So because Apache Iceberg and other columnar table formats have rolled out and made it possible to do this work on data lakes, it's really expanded and accelerated the pace at which we can build and train AIs on the source data. So I'm seeing a lot of companies being able to roll out AI and LLM programs super fast without having to build a ton of infrastructure around it, which historically, even a year ago, they really would have had to do.
So talk to us a little bit about the difference between data lakes and data warehouses and the pros and cons. A data lake can hold any kind of data. So it's unstructured. It can hold semi-structured and structured data, but you just drop it all into Azure or GCP or Amazon.
You just throw it in the data lake. It just lives there. It's raw data. Traditionally, what you had to do then, because you couldn't really use the raw data, it wasn't a format that analysts or data scientists could make use of.
So data engineers would go in and format it and then pipe it into a data warehouse, which was much more organized. It's purely structured data. Now you can run your models or your analytics on top of this data warehouse. Problem is, the data warehouse is now stale because you've had to move it from one place to another.
It's expensive because it's a physical copy of the data, so you've got it living in two places. So there could be a mismatch between what lives in the lake, which is where the source systems are feeding, and what's in the warehouse. Warehouses were really great for structuring data. They still are great.
They can play nice together, but there are some downsides to it. You couldn't really do analytics on the data lake because it was difficult to use. Also, accessing it. When you run a query on a warehouse, it's super structured, so you can access that data easily.
It's easy to find. You don't select from where. But with the invention from Netflix of Apache Iceberg, it's a table format where you don't have to read billions of rows of data. It's columnar.
So it can omit tons of the data when you're doing a query, which now makes it possible to efficiently access data within the data lake, which is where all the source systems are dropping that data anyway. It's just now you can use it without having the whole process for restructuring it, copying it, and moving it into a separate place. That's amazing. Obviously, it's some huge innovation that unlocks a ton.
Tell me about what got you excited about working with Dremio to begin with and what you're working on now there. I felt like it was a job that I made for myself. Evangelism is a role that's coming out and becoming much more popular and important at tech companies. They've historically had developer relations, which has been much more on the writing manuals, writing release documentation side, where evangelism is much more on the product promotion, understanding use cases for the hands-on developers, working with developers and speaking to them about how they can best implement the technology for their specific use cases.
But it's nice because it's not sales. It's not marketing. It's really just deeply understanding your product at a technical level and then working with technical people to help them understand it and implement it effectively. So a big part of that is, like you said, LinkedIn, YouTube, stuff I've been doing for free for years.
I'm like, oh, this is going to be a full-time job by just making YouTube videos and LinkedIn content for a product that I think is really great to an audience that I understand and have worked with a lot. So it's a really fun job. You see a lot more companies starting to roll out evangelism. And they're pulling folks off LinkedIn who are posting technical content and building communities because they're like, oh, this is what we want to do.
We want to build communities around our product. We want people who deeply understand the issues that developers are facing. And we need to convey that in a way that's not super salesy but also technically competent. No, I think it's so important.
And also the issue is, like, of course, me coming from my marketing background, yeah, like, if you give me a product and I understand it, it's kind of like direct to consumer and it's consumer focused, yeah, easy. I'll just make content about how to use it. It's great, whatever. But as soon as it gets to be, like, a really technical thing that, like, I don't personally understand, all of a sudden I'm like, yeah, I'm having to talk to developers every day.
Like, okay, how does this actually work? Who's actually using this? Like, what do they do? And they explain it.
I don't even understand their explanation. And you have to have a person like yourself who's an evangelist who you understand from a technical perspective because your target audience is technical people. The marketing team isn't really going to understand that, right? And so, yeah, I think you're 100% right.
This is going to be a role that we see a lot more going forward into the future. Tell me about some areas in AI that you're most excited, just kind of, like, generally, innovation that you're excited to see. I think AI can help with a lot of areas that are kind of nuts and bolts. One is data quality.
We've seen AI be able to first categorize data and be able to build, you know, automatic semantic layers so that the technical side of the business can now speak the same language at scale, which is super important and incredibly difficult to do manually. Data quality. AI can highlight outliers in data quality, which when you're building an LLM system is super important because when you're training on billions of tokens, you can have malactors inject, you know, pretty nefarious things into your data that now you've trained on and they'll hold you ransom for some Bitcoin to tell you what it is, which is happening. But AI can help you identify that and pull it out.
Also, education is really making professors much more effective because it can take off a lot of their administrative tasks. It can help them re-envision their lessons, create visuals, be much more impactful with their students, and empower students to, you know, get a lot of the grunt work off of their backs. It can help them brainstorm. It's like a study partner.
So I think it's really going to transform education, as we saw with the recent open AI announcements with ChatGPT4O. Yeah, totally. 100%. Yeah, a lot of innovation going down.
What are some common misconceptions that people have right now that they're just getting wrong in this field? I think with something so new, there's so much hype, there's so many people talking about stuff, or some things we're just getting wrong on this. Well, AI's been around for a long time. Like, we say, like, AI, but what they really mean is generative AI, which has also been around not for a long time, but it's been around for a while.
ChatGPT3 has been out for, like, a long time. It's just 3.5 was accessible. It's also not magic, and it's really just math underneath. And so I think we get that wrong a lot of times, especially business leaders.
They're really looking for a silver bullet or a shortcut, which AI really isn't. It can be super helpful, but if you build it on top of a bunch of junk, it's not going to be helpful. So having a strong foundation that these processes, these math, this technology can pull on, it will be helpful. But if you have a bunch of flimsy junk, crappy data quality, business units that have different definitions for their KPIs, AI is going to make the problem worse.
That's so interesting, because, you know, so often we hear the pitch, like, AI's going to fix, like, literally everything. But you're saying, in some cases, this is actually making the issue worse. What are things people should be looking out for to avoid that? It's the business fundamentals.
It's do we speak the same language with our data? Is our data strong? Is it consistent? When we say, like, a customer, what do we mean?
Like, do we all mean the same thing when we say customer? Is it an account? Is it somebody who formerly had an account? Is it an organization?
Is it an individual? Different business units will define that differently. So if now the CEO is using an LLM and saying, you know, tell me how many customers we have, what does customer mean? Make sure that that's buttoned up before you start AI.
You know, this data is getting fed into AI. These models are, you know, becoming better and better. What's your opinion on, you know, companies like, let's say, OpenAI, who are working towards AGI? What's your opinion on when this arrives, how possible this is, and maybe some of the data needs that would be needed to actually achieve this?
What are some of the things people are thinking about? Yeah, this comes up a lot. I think the definition of AGI varies from company to company. And I think...
it's good to have that goal i mean if open ai isn't shooting for agi then what are they shooting for like if your whole model is to build the best ai and we define agi as the best ai then that's really what you should be working towards regard whether or not it's feasible i don't know but i think we can get pretty dang close and i think that what we're going to see is instead of these giant generalized models like uh you know chat gpt4 and 400 and gemini which you can ask at any question on any topic i think we're gonna see much more niche ais which know everything about a certain topic or a certain task i think that we're gonna see a lot more of niche ai tools strung together than more of these enormous general proprietary models and um correct me if i'm wrong but based on a bunch of leaks that i think happened last year for opening i when they were like leaking some models and wait some stuff for a gpt4 they were saying that gpt4 technically was like 16 has like 16 experts or something inside of the model that it's like sending different different places do you think like if we're doing these big general purpose models that's where it goes to where you just have more and more of these kind of like sub models doing things talk about what that looks like yeah that could be a way to do it where you've got niche models which are experts on certain things when you have an overall model which chooses which model to query that could be a way to do it but i think the the overhead of building that would be so hard that there's not gonna be a lot of competition in those areas where i've seen a lot of ai companies grow is in these niche models where hey we're gonna be the best python model that there is we're gonna be the best um someone the other day that takes textbooks and turns it into like quizzes or exams and that's all this ai does but it's super good at it i think that's where it could be going like the app store yeah no 100 i agree um that's why i'm building ai boxes because fundamentally i believe yeah open ai is great gemini is great there's gonna be a thousand probably at some point a million different ai models and they're gonna be good at different things specifically but like i just if i'm doing if i'm doing java or i'm doing python i just want the best python model and maybe you have access to proprietary data because you had some python community forum network that you were able to grab and you pull them offline people aren't training on them you have this data you can make this incredible model that maybe other people can't um monetize it and and i think that you'll be able to be but my one question or maybe like apprehension in this is like what about synthetic data do you think that um i mean yes you could train to be very specific good things at what point just opening eye just say forget it we're not licensing anyone's data anymore we're gonna create synthetic data somebody asked me this question the other day about the pros and cons of using synthetic data to train models um because they felt that there was a lack of training data i don't think there is a lack of training data my opinion on using synthetic data to train models it's like the blind leading the blind i think once you get some error in the synthetic data now we train a model and that model's gonna create more synthetic data and it exacerbates the error that's my opinion on it i just don't think we have a limit to the amount of data that's not synthetic that we could train on i wouldn't use synthetic data to train but i'm also not in charge of open ai but i have seen there is a large rise in in creating synthetic data to train not just ai but ml models yeah i don't i do see the need for ml models um just because there are times where you're trying to identify outliers where there's a very small sample size for instance financial services they try to find insider trading that's one of their like risk models but there's not a lot of examples of like here's a thousand insider trading right yeah like in a hundred years you found five and there's the data around that so maybe you build synthetic data around that you train on it to the best of your ability i have done that i think with ai there's not a limit to the amount of data you can train on that's my opinion do you think that there's untapped data like places that people aren't thinking about because we oftentimes it's like so open eyes training there's a video model and everyone's like do you guys train with youtube and like not gonna say anything everyone's like okay you guys are training with youtube like so there's just like public things like youtube right huge data set people are training off of it are there are there companies or places with mass amounts of data that's not public that are going to be really valuable and useful in the future yeah there are data marts out there that um specifically take data that's hard to get or hard to organize it's becoming more easy but you know video is hard to organize and to train on um a lot of this data is like scraped so the average person can't go and scrape the internet so a lot of these companies made their model by scraping the internet and organizing the data and then turning around and selling it um axiom is a big company which really collects uh all the publicly available data on people they want to know absolutely everything about a person um for better or for worse and they're super good at it and they sell that data to companies tons of companies that i've worked for have bought axiom data for marketing so like if your name matches something in the axiom data set now they know like a hundred different data points on you and they can build a model about like what you're most likely to buy i think that's going to grow and grow the better and more easily accessible some of this unstructured data is where do you think some of the biggest opportunities are in the market today based off of the players and landscape and kind of what's going on i really think that hugging face is doing amazing things with their open source models and empowering developers to to build their own gpt models for whatever use case that they have i think that we put a lot of hope into the huge proprietary models which are super good very impressive i have subscriptions to all of them so clearly i love them but i think open source is underrated right now and that the crowdsourcing power of taking these open source models and building something amazing is something that we have yet to see but i think it's coming do you think as far as the landscape going forward into the future we ever get to a point where some of these open source models are better than these closed source there's all the controversy of open ai and they were open source and they went closed source and whatever and people are like well it's because they need so much money and there's like google and that's so you have these like huge people are we ever going to be able to compete are people going to create open source things that compete realistically with that i think so i think it's similar to the machine learning paradigm that we went through when in like 2000 from like 2008 to 2000 i'd say like 16 like regulated industries like financial services healthcare only used sass and products like that because it was closed it was quote-unquote reliable and explainable and they wouldn't use python or r for their machine learning models because they're like open source we can't verify it or trust it but now none of them that i well in europe they still do but a lot of companies don't really use sass anymore one it's super expensive and two you can replicate um or make it better with r or python which are open source so i think we're going to go through that same shift with ai where right now like we have to trust the proprietary models because they're so advanced and there's explainability and trust behind it i think we'll get to the point where open source is just as good or better than the proprietary models who do you think are going to be some of the biggest winners in the next let's say three years like you say like five years on stuff oh my god that's too long next three years what are some companies you think are the best winners we've seen incredible run-ups with nvidia are they going to be challenged incredible run-ups with a lot of different companies apple has yet to really announce a serious jet to be competitive or anything like that are they going to jump in the race how do you predict the landscape what do you think the players are going to be doing yeah i think apple's an interesting one because we don't super know what they're doing but they tend to be late to market on almost everything that they do but they integrate it so well into their ecosystem that i would imagine that that's going to be the same for their llms whatever they do it's going to work seamlessly with their ecosystem and then everybody who has an apple product is going to use it that's just how apple rolls um i think we've underestimated google because of some of their uh issues that they have with their releases but i think google has some of the absolutely smartest people on the face of the earth cranking on this and um we've seen some folks have to leave google because of those uh missteps that they've taken so i think their fire is really lit under them even though they are big and move a little bit slower than some of the upstarts uh i don't think we've seen the best of google and i would love to see a challenger to nvidia i love nvidia but they kind of have a monopoly which i don't love i love that they're crushing it but i think that when there's a monopoly um there's no innovation yeah so i hope there's somebody out there who's going to challenge the video because then the consumer wins you know what something interesting you said that actually is a great point with google is like they had a bunch of missteps so they have a fire lit under them i actually would rather google be in that position and like let's say google just came out like last year with like a great gpt competitor that was kind of comparable and i was like okay cool and i got like 50 market share and then we just see like these two kind of like i would way rather have them just like miss that miss that because now like you know they're they're pushing so hard right now it's at every part of the company that's like a major focus i think that's where we're gonna get like that's where these insane budgets are getting dumped right like we talk about like oh it's amazing microsoft gave opening i 10 billion dollars like well look at what google's arsenal is at their disposal like they really are taking this seriously they're gonna put more than 10 billion dollars a year into um pushing some of the same projects so i do think that's gonna be exciting and uh i hope to see some great things out of google i'm slightly jaded like you mentioned because of some of their releases but uh yeah i think i think they'll have some great stuff um andrew it's been phenomenal having you on the show today talking about you have so much knowledge to definitely get you back on what's as we wrap this up what's one piece of advice you feel like you could give people that are getting into the industry whether that's ai whether that's data analytics um that you think is really important for them to to be focused on i would say to start small and not get overwhelmed and that everybody starts at the beginning right it's too easy to try to jump right into it and become discouraged and then say oh it's not for me it is for you like it really is for anybody anybody can learn anything really so i'd say start small and then incrementally build your success with those topics and then just keep expanding your abilities amazing absolutely love it if people are interested in following along or looking at the projects you're working on how do they find you just find me on linkedin andrew madsen andrew madsen you're not gonna miss him he's got 80 000 followers i'm subscribed i get every single update this man is an absolute legend if you're interested in data analytics or data analysis all of that is really phenomenal stuff thank you so much for tuning in to the podcast today if you enjoyed the episode you'd meet the world to us if you left us a review we'll leave the links in the show notes to go find and follow andrew and we'll catch you on the next episode