Ep. 69 - With Brian Julius (Learning About Model Context Protocol - MCP)

Speaker A: Foreign. Hey, everyone. Hey. Welcome to today's live recording of the Data Ideas podcast, joined by, uh, an esteemed member of the data community and friend of mine, Brian Julius. Brian, good to see you again.

Speaker B: Great to be here. Great. I really appreciate the invite.

Speaker A: Yeah, absolutely. It's good to catch up and talk with you again. You've been, been busy. Um, we last spoke, I think, um, at the last Data Career Summit where we had a hiring manager. I think it was like, impress. The hiring manager was the panel that you were on. And yeah, yeah, you've managed a lot of analysts over your career and hired many and um, but you've been up to a lot since then. What, what have you been up to and how are things going?

Speaker B: It's going great. Yeah, thanks for asking. Um, you know, as we were talking, yeah. I was saying I kind of feel like an extra in like, the most interesting movie I've ever seen. You know, just with the way data is progressing and AI is, is moving so rapidly, you know, I feel like, I feel like the whole field, you know, kind of changes on a weekly basis and, you know, so I've really been, you know, kind of leaning into that and, um, you know, really kind of following it closely, experimenting a lot with, you know, the tools that are coming out and kind of getting a. Trying m. To get a real sense of how I think these tools are going to affect the way we do analysis in the future. And so it's been, it's been a lot of, you know, it's been some, some client projects, it's been some personal projects. It's been talking with a lot of people in the community, kind of getting out, experimenting. So, you know, that's, that's really been the focus for, you know, about the past two years.

Speaker A: Cool. Awesome. And you've got this concept, Data Mad Science. It's on the banner on your LinkedIn profile now. And there's a community around this. Can you tell me more about that?

Speaker B: Yeah, it's funny, it kind of came out of a post I did a, uh, couple years ago and, um, at the time, you know, we were talking a lot about, you know, kind of how do you get started in a career in, in data. And you know, the thing you would hear all the time is, you know, you've got to learn these tools and you've got to build a portfolio in this way. And um, and I thought, you know, what if, what if you're just not inclined that way? And I'm, I'm not, you know, I'm not somebody who, you know, likes to do things necessarily in the, you know, kind of the, the tried and true way. And um, so you know, it was kind of this, this discussion I had with some folks and you know, Greg Decler being, you know, being one of them who's, you know, kind of if, if there's a, if there's a standard way of doing things, you can be sure Greg is not going to be doing it that way.

Speaker A: And I do remember uh, uh, we, I was on the data, uh, viz. Um, contest review panel, I think for Enterprise DNA with you and Greg. So I remember meeting him.

Speaker B: Right, I remember that. Yeah, yeah. And you know, so we were kind of talking and you know, a number of us really got in this discussion about if you, you do have to kind of provide social proof of your abilities, but is there, is there a different way than kind of taking the normal path? And we, we really came to the conclusion that there's a good way to do it if you're, if you're so inclined, which is this, this kind of mad scientist route where instead of, you know, kind of presenting polished projects, what you're doing is you're, you're out in public kind of constantly experimenting and you're trying new things and pushing the envelope. And you know, some of those work and some of them don't. You know, that's kind of the mad science part of it. You know, some of them kind of blow up in spectacular fashion, but some of them really work and you really advance the, the understanding of, you know, what, what the tools can do and um, you know, what some of the, in some of these kind of off label uses, um, are. And you know, so we've, there's kind of a, there's kind of a group, probably about a dozen of us who are just constantly out there, you know, kind of doing strange things with, with the tools that are available and seeing, seeing how it works. And um, and I think that, that, that's really been, for me that's been incredibly motivating and one of the best learning experiences I've ever had.

Speaker A: That's cool. That's awesome. That sounds like a ton of fun.

Speaker B: It is.

Speaker A: Can you give us an example? Just maybe like, like yeah, just like one thing that the group has produced or been working on or even if it's just like conceptually what it is,

Speaker B: you know, I mean a lot of it these days is um, is kind of centered around the, the intersection of AI and analysis and visualization. And so um, one of the things we've really been talking about is with visualization there's been this shift from, for the past 10 years Tableau and Power BI have been really UI focused. It's click and drag, um, kind of all focused on making it quote simpler. But m. As you move to an AI dominated era, what's simpler for a computer and an AI is totally opposite of what's simpler for a human. And so AI does great at coding. It does really pretty, pretty poorly at clicking and dragging. I mean you can, you can replicate that to some extent but it, it handles problems much better when it can just work in code. I think you're seeing that in you know, in Claude code and kind of the move to um, you know, kind of terminal based and you know, almost going back, you know for me like you know, 30 plus years. The you know, DOS 3.1, you know it's, it's fascinating how those commands and the, the look and feel of what I do these days is um, is really much closer to what I was doing 30 years ago than what I was doing two years ago. And um, and so a lot of us, we've been experimenting with okay in that kind of cloud code, terminal all code environment, how do you build visuals, how do you build dashboards, um, how do you um, kind of automate things like the um, IBCS structure, um, and rule set for data visualization. And so we've just been out there, you know, really, really just playing around and you know seeing are there, are there ways in which we can speed things up in a pretty dramatic way. And I think we, I think we, we've we've shown you know there potential is you know like a 10x you know, efficiency improvement in this and then you know, kind of are there other things you can do that you, you might not have been able to do? Um, even with the additional time in the, in the, in the past.

Speaker A: That's amazing. And looking forward to digging into this a little bit further here later in our discussion. Um, before I jump to kind of our next talking point, did just want to encourage for those that are joining live, if you wouldn't mind just saying hello in the chat. Let us know who you are, where you're coming from. It's always good to see who's joining live. Um, and welcome to those that are listening to the recording. This uh, this podcast will be published out uh, later today on all of our usual podcast uh, platforms. Um, so Brian, what. So for those, first of all I should say for those that might not know, um, I think Most listeners probably know or have heard of Brian or followers of you. But um, for those that, that don't know, Brian was. Brian had a long career in, in analytics and was a analytics leader and then got into the Power BI world and was a leader um in that space and worked for a um training platform enterprise DNA for Power bi. And um has now kind of completely again reinvented um himself as a data mad scientist but also expert in um MCP and um, I think one of the leaders in content around mcp which is something we definitely want to dig into. But um, what prompted this kind of latest shift from kind of um. I would assume you're still a little bit focused on Power bi, um but like the shift from that into focusing specifically more on mcp. Like what, what prompted that pun intended for you?

Speaker B: You know a lot of it was, a lot of it was my conversations with Sam Makai. Um, as these tools started coming out we started playing around with them a lot and you really started seeing the potential for initially, initially for um, you know, data modeling, data cleaning, um, and then you know kind of further into analysis and we just really you know again just started experimenting and you know when MCP came out and this was really not even a year ago, you know, xanthropic drop that paper in November, the end of November of 2024. And it just based on what we've been doing with you know testing large language models, it just made a tremendous amount of sense and you know, kind of having something that instead of kind of keeping you locked within that chat box, that the ability of your, your AI to go out into the world and do things, um, you know, to, to browse, to take screenshots, to um, to you know, access your local files, to write you know, code and then execute it and you know, look at the results. You know, visually once you kind of give your AI ah the power to do some of those things, you just begin to see that the possibilities everywhere. And I think we really you know, kind of did, did some initial testing and there wasn't, when it first came out there wasn't a lot of tools. And then there was this boom, you know, probably about February or March where you know, thousands of tools just started proliferating and they're starting to get, become this organization around it where they're you know, sites that are kind of aggregators and almost kind of fill the role that um CRAN does for R. So for those who are, you know, R fans, CRAN is, is kind of this huge public Repository where they actually kind of impose a documentation standard and you know, an update standard. And so you can, you can be pretty certain that when you pull up a package from Cran, you know, that's been, that's been reviewed, that's, you know, up to date, it's got full documentation. And so I think what we're starting to see is people on the MCP side kind of building that sort of, you know, governance and, and you know, kind of quality control. And as that's happened, um, we've just continued to kind of expand our, our experiments and you know, what we found is that we've always talked about, uh, an analytics stack in terms of SQL, R, Python, Excel, Power, BI or Tableau. What we found with MCP is it gets down to a granularity where everybody can individually have their own stack that's specifically matched to exactly the type of work you do, the types of analyses you perform, the types of visualizations you create. And m. It kind of opens up this, this whole world where you're, all you're doing is you're, you're talking to the AI, the AI is calling the tools that kind of fit your stack. And it just, it's a completely different experience.

Speaker A: That's really interesting and it's actually, it's, it's. I, I want to share this with you. Before we went live, I was jotting down. There was a project that. So I've been following Brian's content at mcp and as a, uh, longtime practitioner when he first started posting and writing about it. And then I went and I was interested, you know, in what he was writing. And I went and listened and watched some videos on my own and I was like, okay, yeah, this seems pretty transformational. And I was just jotting down, um, on a piece of paper in front of me, you know, I was thinking of where this could have been useful in my prior life and trying to tie it to a practical example. Um, I had a project as an example once where we had this hypothesis which started out in just an office discussion, where we thought, you know, a product or some products that we had were economy driven for the most part, you know, and um, there's all these economic indicators out there. And we were curious, you know, how did our sales of this product or basket of products correlate well with these, the movement and economic indicators? And I figured you would appreciate this example as an economist as well. Right. But, um, then we thought, well, guess what, there's consensus forecasts for where these indicators are going. And if we could figure out which indicators our product sales are tied to, we could then use the consensus forecast for those economic indicators to potentially forecast, um, our product sales. Right. And um, and it turns out that, you know, our hypothesis was true. Um, there were some economic indicators that were really, really good indicators, um, for where this product would go. Um, of course we were reliant on the um, accuracy of the forecast for the indicators, you know, in terms of the ability to project the product sales. But as I think about what I had to do, you know, so I could have just written up that idea and that, and a prompt to go, you know, do all the steps that it would take to get from analyzing the economic data, comparing it with our product data, bringing a data source together with those, doing some analysis, um, figuring out which indicators were correlated, building a predictive model based off of it, and then visualizing, you know, where our product sales were going. Right. Based upon the, the um, analyst consensus forecast for those indicators, all of those steps. It feels to me like, you know, MCP could potentially, if I'm understanding it right. And Brian, by the way, everyone has a thousand times the expertise on me than this, so that's why I'm asking him. But, um, could I, practically speaking, put together a prompt in the right way, in such a way where um, we could connect the LLM that I was prompting to go and connect to all of the different tools to scrape that data, put it into a data source, do the analysis and then publish a visual visualization, essentially? Am I thinking about this the right way, Brian?

Speaker B: That is exactly the workflow. The thing about MCP is if you think about the traditional way you would do that. You would have to go in your browser, search for the data, maybe scrape the data, bring it back into um, Power query, clean the data, then you've got to model it. You've got to basically be moving, moving data in and out of all sorts of things that you've got it from the browser, you download it, maybe you look at it in Excel, then you put it in the Power Query, then it goes in, you export from Power Query into CSV to pull into R or Python to do the statistics and then back out somewhere to do the visualization. MCP just takes care of all that. It makes it, it makes it almost seamless in the sense that I, I spend 80% of my day at least within one MCP client.

Speaker A: Yeah.

Speaker B: And, and I. This whole thing about jumping back and forth between apps and all, it. It's not even relevant in the way that in the way that kind of analysis under this framework works, that even the idea of apps isn't even a thing. Um, that you almost, in a sense, you almost change your whole operating system, your operating system becomes the MCP client. And so um, what you're doing is you're just sitting in that client kind of calling tools from your, um, from your AI. And the really interesting thing about this is, you know, people express that um, you know, their, their biggest concern almost always about AI is the dependability of what the LM can produce. And you know, it, it's definitely gotten better. You know, the, the models are getting better and better, but there's always a random element to it. Even if you crank the temperature all the way down to zero. What people who are experts in this have explained to me is that you're still going to have some element of randomness because what, what cranking the temperature to zero says is you're always bound to the next most probable token. But what happens in, in a lot of cases you get a tie and to break that tie there's still some element of randomness. And so you're probably never going to get um, an LLM to completely eliminate, you know, to make completely deterministic, where two people running the same query in the same way are going to get exactly the same answer. You know, that's I think, unlikely. And so what you want to do is instead of kind of putting the burden on your, your AI, the analogy I like to use is you want to take Einstein and turn him into the world's smartest receptionist.

Speaker A: Gotcha.

Speaker B: And you know that basically instead of having your AI doing the analysis, what you want it to do is to say, okay, all it needs to do is it needs to call the right tool to do the part of that analysis that's next. It may be that you've got a web scraper tool. And the AI calls the web scraper tool with the parameters saying I want you to scrape these data for this year range and um, across this geography and it pushes the parameters into that tool and then the tool runs validated proven code. And so you're eliminating to a large extent that random element. And what we found is that um, AI can call tools really quickly and accurately. You know, a uh, top notch frontier model like you know, Gemini Pro or Sonnet, you know, Sonnet 4 opus 4.1 are great tool callers. And so you can really much more dependably do analyses once you've got it kind of routing out to the proper tools than if you kind of burden the, the AI to do the whole thing itself.

Speaker A: And taking a quick step back, we talked about an example, um, a few minutes ago. Uh, but what, in your own words, how would you explain MCP to someone that maybe is listening, that hasn't even heard about it, or maybe they heard about it but hasn't done any investigating on their own? How would you describe it? Just kind of in simple terms on your own?

Speaker B: Yeah, you know what I would say? I mean, the analogy that gets used a lot is it's, it's kind of USB C for AI tools that it's this common set of connectors. And the way it used to work is if you tried to link a tool to an AI model, you had to do it through the API, and every API was different and you had to write all the code around that. Um, and then getting them all to interact properly was a nightmare. And so what Anthropic did was they, they kind of set this standard where as long as you're adhering to the MCP protocol, you can just take and plug those tools in. Almost kind of a, you know, plug and play. I mean, it's not quite plug and play yet, but it's getting really, really close, um, where you can just take it. In some cases, just double click on a, on a, on an MCP package and it'll install into your, your client and link to your, you know, your AI model. And, um, now you've got, you know, in some sense, like, you know, a given MCP Server might have 25 tools in it. So, you know, one of the things I use is Supabase, um, which is a Postgres, um, postgres database, um, that is very smartly MCP enabled. When I'm hooked up to my MCP client, I can then access that database. I can execute SQL commands, I can manipulate that data. I can pull all the metadata down. I can, um, create other tables. I can create relationships. The thing about it is that again, it, it's got this, you know, this list of 29 tools. And if I want to swap that out, if I want to say, okay, like it doesn't do X, I can then say, okay, find me. You know, I can even, I can even ask the AI find me a tool that does this, and it'll go out and search and say, okay, I've come back with, you know, three tools to do that. Uh, and then I can say, I can say, okay, go ahead and install that tool for me. If you give it access to the local file system. It'll actually write the configuration in your config file, set the thing up, and then you can even get it to say, um, write me a two page uh, cheat sheet with all your best tips on the 10 best ways to, to use that. And I can, I can show you, you know what I do is I have it, I have it, give me that. And then I laminate it. And so, you know, I've got the AI basically teaching me to use the MCP tools.

Speaker A: It's amazing in terms of tools, what, you know, I, uh, mean, maybe we can talk about some of the popular analytics tools and platforms that folks may be more familiar with. You know, what, what's out there that's compatible with MCP today? And then also curious, like if you've prompted, you know, put a prompt together. We've asked for recommendations on tools to connect to. Like, are there any like, surprises in terms of the tools that are out there that um, you know, maybe you didn't, you hadn't heard of them and they did something really cool or like, so what's, what's um, what's out there that's like popular tools that you can connect to and then what are some that are kind of more interesting to you?

Speaker B: Yeah, um, so the really popular ones are things like, um, some of the official ones, like GitHub is a great one, you can have it manage basically your whole version control, creating repos, um, branching, merging, um, writing the readmes, pushing documentation, all of that just natural language. And what I do is I just use voice. I use um, a program called Whisper Flow, which is the best um, you know, voice to, to text program I've ever used. And I've been using those programs for about 10 years. And so all I'm doing basically is you're talking to it. And um, and what the GitHub one um, that I recommend people start with is if you've got a personal knowledge management, um, application. So if you say notion or Obsidian, um, that is a phenomenal way to basically increase your knowledge and also the AI's knowledge. Um, so I've got a six gigabyte obsidian vault of, you know, every piece of content I've, I've ever written.

Speaker A: Oh really?

Speaker B: Yeah, yeah. All the content I've saved from other people, medium articles, LinkedIn posts, you know, anything that I think this is information I want to, I want to, I want to be able to search and retrieve for the future. So I've got, I said I've got six gigabytes of that and now my AI is trained on all that and it has immediate access through mcp. So I can say, okay, like for. I was working on a project the other day, um, analyzing airport, um, data, you know, for Washington.

Speaker A: I saw that analysis that you published. Yeah.

Speaker B: And so what, what I did was basically I said I, I had stored a lot of documents for that in Obsidian, and when I started up the, the project in the next day, I just said to my AI, go to my Obsidian vault, look in the AIRPORT project and review the files. And um, it does that and says, okay, like I, I understand what we're doing. Um, you know, I've, I've got, you know, all these, all these things now lined up in my memory. What do you want to do next? Yeah, and you know, it's, it's really kind of an incredible way. And the other thing it'll do is it'll learn. And if you give it the instructions in your, in your custom, um, instructions, what it'll do is you can say, if you come to a point in an analysis or you know, respond to a query where you feel like you don't have enough information, go to this part of my Obsidian vault and look for that information before you, you do anything else. And so.

Speaker A: Got it.

Speaker B: What you'll see when you, when you kind of watch the um, the conversation is it'll, it'll kind of come to a point where it's, it's stumped and they'll say, oh, okay, let me check the vault. And you know, it goes in, it's got, you know, six gigs of information now that it can, amazing, it can pull. So what I feel is you, you've kind of got your prompting knowledge and your context knowledge augmented by everything you've ever done in the past.

Speaker A: Mhm.

Speaker B: And the power that creates is remarkable.

Speaker A: That's amazing. That's interesting. Um, now I'm wondering, I wonder if we took the, uh, you know, I have the Friday jobs rundown post. I wonder, uh, maybe we could put all those posts into Obsidian Vault and maybe we could, uh, do some analysis on, um, who the movers and shakers are, um, in terms of doing hiring, what trends we see in titles, things like that. Um, that might be an interesting one. Maybe I should start saving those.

Speaker B: Definitely. I mean, yeah, that's, that's a perfect example. Is that what you can do, what you can even do with that is you can say, okay, read through the structure, you know, if the, if the, if. If the poster of similar structure every week, build that into a, um, you know, a CSV or a parquet data set that I can use. And then you can say, okay, upload that to my, you know, my, my database. You know, for my case, it'd be Supabase. You know, it can be fabric, it can be, you know, whatever, whatever you use. The really interesting thing I found is that, um, there's kind of a, there's kind of a great merge of kind of old school, new school here, which is the way I, the way I was taught to do analysis was you always start with an analysis plan and your analysis plan is what are the questions you're asking?

Speaker A: Right.

Speaker B: What, what's the data you have? What's the data you need? Um, what are the hypotheses you have with regard to those questions? What are the tests you're going to run? What are the visualizations you're going to do? And what I found is, you know, that all through my career and you know, for most of it I was working for the government where the work I was doing it was actually a legal requirement to publish your plan.

Speaker A: Yeah.

Speaker B: And so, you know, it really reinforced this idea of, you know, analysis plan first. And what I found is in the MCP and AI context, that old school plan is golden because, uh, basically what it is, it's a giant prompt. It's a giant prompt that gives the AI vision into your whole analysis. What you can do with that is you can say, okay, here's huge flat file that I just had to create of, you know, all the jobs data. And here are the questions I want to ask about that jobs data. Now come back to me with recommendations for the schema that I should build. Sure, support that.

Speaker A: Yeah.

Speaker B: And it'll come back with a, it'll come back with a detailed, detailed plan. And then you can say, okay, I like that. Build that schema and you'll just watch it, you know, kind of magically rearrange all your data and you know, do what, what could be hours or even days of data modeling and right. Of, you know, edge function creation and testing all just happened right in front of you. It's, it's remarkable the first time you see it.

Speaker A: That's amazing. And it sounds really transformative. And um, I can think of. I think I mentioned this in one of my posts this week, but I can. My mind has been going wild with use cases for this and now that we're talking and having this conversation, it's going even wilder. Um, I think it's a good prelude to the next question and I think there's no question that this is going to have a significant impact on the scope of, you know, analytics professionals roles. Um, yeah. Um, I'm curious your thoughts on what that may look like, you know, in the nearer term. Um, you know, maybe next six to 12 months and then we can talk about longer term, but maybe we start with the short term. What do you, what are you thinking?

Speaker B: Yeah, you know, it's funny because I think the, um, you know, kind of the, the more limited view of what data analysis was, which is, you know, kind of the creation of reports, I think is, is gonna be kind of a commodity very, if it's not already. I mean, one of the things that I've been experimenting a lot with is this program called, um, Plotly Studio. And Plotly Studio basically will take that analysis plan or, you know, a large prompt or even a small prompt, and basically just turn it into a whole series of visuals. You know, and if you give it specificity, as you were saying, kind of that, that analysis, the economic analysis that you were talking about, you know, it can, it can do the correlation matrices, it can do the, um, you know, the heat maps, it can do the, you know, the time series forecasts, you know, all kind of automatically. And, you know, the visuals it creates are really pretty spectacular. And, and they're interactive, you know, so it's, it's putting slicers, it's putting filters, it's putting all sorts of interactive features in there. And, you know, I think when you look at what AI is capable of, the, you know, kind of what I, what I call the wrench turning, you know, uh, aspect of data analysis, you know, the, the coding, the visualization, you know, the kind of creation of those things, I think is really going to become less important. You know, that in some sense, I think we're, we're very close to the point where AI can do most of that better than almost all of us anyway. And so, you know, really, I think what it does is it puts us in a management role. You know, what I like to tell people is, you know, whether you wanted it or not, you've all been promoted to managers and that you're now managing these teams of AI. And um, it's remarkable to me as somebody who's spent a lot of time both doing analysis, my career and managing teams of analysts, how similar managing teams of AI is to managing teams of people. In a lot of ways, the, the parallels are really astonishing. And particularly when I'm interacting by voice, um, it is very hard to tell, I think, to Somebody, you know, my wife upstairs, you know, listening, whether I'm on a call with a person or whether I am just, you know, prompting and talking to my AI. If you don't hear the other side of the conversation, it's very hard to tell because, because it just sounds like, I mean, to a large extent it sounds like the conversation we're having.

Speaker A: Right, interesting. Um, do you think that. And so, I mean, for the last decade or so, you know, building a data product from a new data product from scratch, you know, going through all the steps and all phases to do that and release it, um, and get engagement and such, you know, this is something that has been for the most part managed by technical folks. Um, do you think that this is going to change things? Do you think non technical folks are going to be able to step in and start producing data products or, you know, uh, what's your take on that? How long is that going to take?

Speaker B: Yeah, I kind of think in some respect the idea of data products is going to change, you know, because I think one of the things, I think that's been a downside of the dashboard era is, you know, like think of it is that when it, when it came about, I think a lot of us, myself included, were like, wow, this is awesome. And this is now, this is the hammer that we're going to hit everything with. And.

Speaker A: Right.

Speaker B: You know, you, you talk to people and I, I did a lot of this myself. Is that every request that came to you became a dashboard. Right. And you know, I've talked to people who, you know, there was one conversation I had on a post where, you know, I said, yeah, I talked to somebody at 300 dashboards, you know, and people uh, came back to me and said 300 is nothing. You know, we've got 1200 or we've got 3000. And you know, when you, when you think about that, it's like how much, how much wasted time and effort and you know, resources goes into, you know, 1200 dashboards of which probably 10 of them get used regularly.

Speaker A: Totally. Yeah.

Speaker B: And yes, I think that when you now have the ability to really just query your data, live with natural language, um, it potentially I think really brings the idea of self service bi to the forefront in a way. I think it's never been. I think we thought it was going to be with, know, with the dashboards, but um, really hasn't been to a large extent. And you know, so I think, I don't think dashboards are going to go away. But I think what you're going to see is they become really focused on those things that are recurrent, widespread and need that single source of, of truth for everybody. And sure, you know, a lot of analyses are not that way. They're not persistent, they're a one time, they're a one time question or a periodic question. And um, you know, for your, you know, your point about, you know, which, which time series are correlated with the variables we care about. Right, Yep. Yeah, that, that, that's an in depth question. It's a difficult question. But once you, once you resolve it, you don't need to keep going back to it, you know, every week or every month.

Speaker A: Right.

Speaker B: Um, you know, maybe you need to see, okay, over the years, do those, do those patterns remain stable?

Speaker A: Exactly. Yep.

Speaker B: Um, and so, you know, I think with the speed and the customization with which you can do those sorts of non recurrent analyses, um, you know, what I found is it really changed. Even if you go into an analysis with a, a really well refined plan, uh, what you, what happens, you get surprised by things and you say oh, that, that did not produce the result I thought it was going to. And okay, that now raises a whole bunch of other questions. And the ability to kind of dig in with potentially some very complicated methodologies very quickly, um, you know, in response to something you may not have anticipated is a, is a whole new experience for me. Because normally it would be like, okay, that, that isn't what I expected. Let me go back and you know, write two days worth of code, you know, to, to reroute that analysis to something else or to dig further into why that's happening. And sure, the ability to kind of do that, I mean Art Tenic is one of my, you know, kind of my favorite mad scientists. And he, he did something the other day where in five minutes he had AI create a machine learning response to a question he had. And the ability to do that and then to look at that data and analyze it in almost the same way you would do it, uh, in exploratory data analysis. Simple things of means and medians. The fact that you can now do machine learning or principal components analysis with, with a, with a voice prompt, it just kind of opens the door for levels of analysis we've never been able to do. Mhm.

Speaker A: That's pretty exciting. And I did want to just give a shout out, there's a few folks and if you are joining live, let us know, you know, say hello, let us know where you're joining from. Just want to say hi to Taki man then. And then also, um, Albert, yeah, he was just on the podcast a few weeks ago, but uh, he said he never thought he'd see the day of seeing you live.

Speaker B: Yeah, you know, I mean, yeah, I, I want to give a huge shout out to, you know, to him because, you know, his um, LinkedIn hard mode really taught me how to develop content. You know, four years ago I didn't even have a LinkedIn account and you know, I came on and um, yeah, he was really the one who taught me. You know, when I first started, I was, you know, posting three times a day. You know, I was stepping all over my own content. You know, I was kind of gearing to the wrong audience. And um, you know, he really, he really provided the structure and the discipline and the, the community to help me understand how you interact in this environment. So I, you have this huge, huge debt to Albert and you know, really, you know, really recommend that anybody who's struggling with kind of building a, a profile and an online person, you know, kind of proof of skills, you know, really talk to him because he, he is one of the best there is. And I, I really, you know, really thank him for, you know, kind of helping me reach an audience I never would have expected to be able to do.

Speaker A: Yep, absolutely agree with that. And uh, Albert has some really good advice to Brian's point for, you know, professionals, whether you're early career stage to mid or late career stage, just on how to promote yourself, market yourself and even like push yourself outside of your comfort zone, you know, to do things that maybe not everyone would do to get yourself in front of, um, whether it's hiring managers or just other interesting professionals in the field, um, where you can benefit from learning from them or having some connection with them. Um, Albert is just really, really good, um, at teaching those skills and tactics for, for building your network and getting your foot in the door, um, in places and pushing beyond your comfort level. So appreciate you joining um, Albert. If anyone does have any questions, drop them in the chat. Maybe we could take one or two. But, um, um, one thing I wanted to ask you about, Brian. It feels to me kind of, I've mentioned this to folks, it feels to me kind of like 2015 or 2016 where I saw some of the data viz tools that were coming out and started it. Uh, well, they were out, but I was starting to really think about how I could use them to produce things for the organizations I was working for that had never been done before. And I just had all these ideas of Use cases. And I could just see that in the next few years, like this is going to be a big deal. This is really going to change things and proliferate. And it was kind of like that calm before the storm at that point. And I've told folks it really feels like that point right now. Again, um, it feels like the calm before the storm. Things are really going to change. Technology, like what we're talking about today is going to proliferate. Um, what are some of the. And it's interesting because I've talked with other peers and professionals and, and I think a lot of folks have like, uh, maybe I shouldn't say a lot, but some have heard about mcp if you know about it, but like no one's really deep in it, you know, and, and certainly at the organizational level, I think a lot of organizations are not thinking about it or certainly have a roadmap for how it would be used. You know, I think for people listening, you know, that have technical and business acumen, they can connect the dots in terms of what some of the use cases or opportunities could be. But like, what, what do you, like, what would you be thinking about if you were a leader of an organization? Like, what are some of the opportunities that this might open up just at a high level?

Speaker B: You know, I mean it's almost, it's almost too big a question, you know, in the sense that it's a big one. Yeah, yeah. And, and it's really, you know, I think, you know, one of the things I like about you know, being in this community and you know, kind of posting, uh, content is you get challenged a lot. And you know, it's, it's almost always respectful and you know, smart, you know, smart challenges. And one of the things that you know, I think people have been responding to in my content is that they're saying, okay, like this is, this is cool. And you know, it's, it's really. You've kind of shown that this is possible, but is it feasible at, at an organizational enterprise level?

Speaker A: Right. Mhm.

Speaker B: And you know, that kind of you, you, you and Sam and Greg and you know, some of these other folks have become kind of, you know, efficient one man armies, you know, but is that, is, is that extensible? And yeah, I think it's a great question. I think it's a very fair, it's a very fair critique. Um, and you know, with regard to mcp, it's a little bit of the wild west right now, but it's interesting because I think it really Parallels. What we saw with LLMs initially was when people started, you know, grabbing on a chat GPT and using it, everybody was saying, oh, you know, this isn't secure. You know, you're potentially going to be breaching, you know, organizational information. And that did, that did happen. I mean, I'm not going to sugarcoat that. It did happen. Um, you know, it happened with the web, it happened with email. You know, I think every, every major shift in technology comes with that rocky transition point. But I think what we're seeing is, you know, that pretty quickly after, you know, LLM's kind of hit the scene that you started getting, you know, the team versions and the enterprise versions and the versions where it didn't train on your data and you could keep things locked down. Um, I think that's really been one of the things Microsoft's pushed with Copilot is it's kind of integrated already into the enterprise system that you've got. And so I kind of see the same thing with MCP that people who are saying, oh, it's cool and you can do all kinds of new stuff with it, but it's not, it's not really ready for prime time in the Enterprise. And uh, sure, I think there's, I think there's, I think that's a valid point, but I think it, it's getting there very quickly. And you know, I think when you see the efficiency gain and the, the way in which it kind of extends the ability of every analyst, you know, that, uh, I can't, I can't see how business is not going to embrace that and figure out a way to basically get it within the kind of the acceptable bounds of, you know, privacy security, um, you know, kind of organizational control to some extent. Um, and I think you, again, you're already seeing that because in the, in the platforms like, um, you know, like anthropic, um, OpenAI, they've now got these, you know, kind of very well defined and kind of strictly limited on ramps to certain MCP tools. And so they, they're often called connectors. And so the ability to connect to your, your email, to your slack, to your GitHub, those, um, are becoming easier and easier and those are really becoming part of the enterprise tools. And so, you know, I think what, what's going to happen is you're going to see this kind of slow and steady progression. You know, some of the stuff I'm doing is definitely kind of out at the edge, but I think what you're seeing is each iteration of you know the new models coming out. And I think, you know, you look at GPT5, if you go into some of the setup for GPT5, it has pretty extensive MCP, um, tools built right in. And you know, so I think, I think what, what people are looking at now saying this is, this is kind of cool and you know, kind of one man army stuff is very quickly going to become corporatized and more tightly controlled and safer. And you know, I just, I, uh, think, you know, a year from now or m. Maybe much less, it's going to be something that every analyst is using to some extent.

Speaker A: And I appreciate you being on the edge with this because you're being on the edge and writing about it makes a lot of us better, uh, and more informed. It certainly makes me better. This is certainly something that's come to the forefront of what I'm interested in much more quickly than it would have been had you not been out there leading and writing about it. So definitely want to thank you on behalf of the community for all the, that you do, um, to keep us on the edge along with you.

Speaker B: Um, one of the really satisfying things, when I hear people say, hey, I tried this out and it really worked. And you know, now I'm doing stuff I, you know, I kind of can't believe I'm doing. And you know, that, that just, that, that that's a win every day I hear that.

Speaker A: That's awesome. And so let Brian know as you're experiencing those wins. Um, he likes to hear that. So shoot him a note and let him know.

Speaker B: Yeah, I mean, even better. You know, the thing I would encourage people is post your poster wins. You know, even if they're not, even if they're not total wins. You know, even if they've still got, you know, some rough edges, um, you know, post them, show them in public because, you know, to the extent that, you know, it's data you can share, because I think that makes us all better.

Speaker A: Totally, absolutely agree. Very cool. Yes. Definitely post what you're doing and then others can learn from you too and be inspired just in the way that Brian's inspiring others. So, um, there's a cumulative effect to that. So I love that recommendation. Um, one final question as we wrap things down. Brian, you mentioned that this is going to extend the ability of every analyst and maybe within the next 12 months or so timeframe, certainly sooner rather than later. If you were a data practitioner today, um, what would your advice be to them in terms of what they should be thinking about? Being prepared to make sure that they can be well positioned to lever this technology and maintain a solid value proposition for themselves as a practitioner within their organization. In light of, of, uh, in light of this, what's your thoughts?

Speaker B: You know, this time I have really strong, strong feelings about. And it's not, it's really not based so much on the tools. It really is going back to building, building your fundamentals. And you know what, what I see, and some of the frustration I have is much of the analysis I see is just descriptive. Uh, you know, it's just kind of what happened in the past, you know, that, you know, here's the, here's the history of the line graph, you know, here are some KPI cards that, you know, show maybe, you know, year over year. And it, I look at a lot of this and I say, okay, like, who cares? You know, right. What is this? Answering and knowing what happened in the past is only the first step. And so, you know, the fact that we've now, as analysts all have these incredibly powerful tools that can get us from not just the descriptive, but, you know, the diagnostic saying, okay, why did this happen to the, you know, predictive, which is what's going to happen next to the prescriptive, which is, okay, what should we do about all this, you know, given that we now know, you know, kind of past, present, future. And um, you know, I think a lot of it is really kind of going back to building that foundation, you know, of statistics, of, you know, methods of, you know, research. Um, you know, there's a book I, I recommend everybody, called the Art of Statistics. And I kind of feel like when you become a data analyst, like, you should get a copy of that book handed to you. And the guy who wrote it was actually knighted, believe it or not.

Speaker A: Oh, really?

Speaker B: Yeah, yeah. And it, it, it's the single best book I've ever read in terms of figuring out the questions to ask and planning your analysis and kind of developing that, that mindset of, you know, kind of the, you know, the five whys, you know, of continual asking, you know, why did this happen? Okay, given that we know why it might have happened, you know, what are the things that are, you know, the, as you said, the other things that are correlated there, you know, that we can use knowing that knowledge and, you know, kind of the, the continual digging down of understanding the, the dynamics that, you know, that led to that result. And, you know, so I think it's a combination of you're really strengthening your methods foundation and really strengthening your domain knowledge, you Know, because I think that the thing that AI does not do well is kind of the big vision and the domain knowledge that the, the wrench turning and, you know, kind of saying, okay, once you've got those questions well defined and you know how you want to examine them, AI is a demon. It actually drops doing that, you know, that, that legwork. And so, you know, kind of using the time you gain in efficiency to kind of build your domain knowledge and build your, your methodological foundation for analysis. I kind of feel like those people are going to be absolute winners in this.

Speaker A: Love it. I love the advice. Um, Brian, this has been a awesome. I appreciate you sharing a wealth of expertise, um, with the community as you always do, packed into this hour, um, for this podcast recording. But, um, I did want to ask in closing, how can folks. I know you're super active on LinkedIn, um, but don't know if there's other, if you have any other, you know, blogs or anything. Is there anything, um, how can folks follow more of your content?

Speaker B: Yeah, you know, I, um, I get that question. And I, um, I started a YouTube channel. I got to admit, I don't really, I don't really like doing it that much, you know, because it's, it just doesn't feel interactive to me. And so I need to get a little better at putting my stuff up there. Um, I've also been thinking about doing some, some medium articles, you know, in terms of longer form pieces. But right now it's really just, um, it's really just the, um, the LinkedIn, you know, feed and the, um, the data mad science, um, YouTube channel.

Speaker A: Cool. Well, we'll be sure to put links to both of those in the show notes, uh, if you're listening. Yeah, you can find those links, um, under the show when you're done listening. But, um, I would recommend, if you're not following, Brian, already on LinkedIn. I know many people probably already are, but if you're not, um, one of the top people, um, in the data world to follow. I've learned a ton from you, uh, Brian, myself, I know so many others are very appreciative of everything that you put out there. To use your words earlier. You know, you're, you really are on the leading edge of things and you're making us all better. So I definitely appreciate that and I appreciate you joining, um, for an hour today.

Speaker B: Thanks so much. You know, I really appreciate those kind words. And yeah, I just, I really, I, I'm just very thankful to be part of this community. It really is a great community. And, you know, there's just some phenomenally smart, creative people out there. And I kind of feel like if we all kind of, you know, kind of contribute a little, you know, and kind of advance things a little bit, you know, forward together, we do some amazing stuff. And, you know, that I'm just so thankful for the people who, Whose work, you know, I've built on, um, and, you know, just welcome, you know, feedback and as I say, you know, kind of people who've been inspired just, you know, to. To check into this. You start putting your stuff out there, you know, don't be afraid because it's. It. It. It's not going to be perfect. You know, nothing I've put out there is, um. Some of it has been, you know, pretty. Pretty, you know, pretty shaky. But in, in kind of going through that process, it really, it really does refine things and it really, you, uh, know, the, the, the dialogue and the, the experimentation just, you know, really advance things for all of us. So, you know, just. I want to. I want to thank those people who've been really active in that discussion and really, you know, kind of the, you know, it's kind of, you know, standing on the shoulders of giants.

Speaker A: Love it. Totally agree. Wonderful closing remarks. Thanks so much, Brian, again, for, For. For joining. And I think I'll see you next. I think I'll. I'll see you next month. Um. Yeah, yeah, for sure. That's another benefit to Brian's point. If you put yourself out there, you know, um, you meet new people. Um, next month I'm gonna. I think I'm get to have a chance to have dinner with, uh, uh, with the legend here. So I'll be out in the D.C. can't wait.

Speaker B: Yeah, on me.

Speaker A: Sounds good. Cool.

Speaker B: All right. All right.

Speaker A: Thanks again, Brian, and we'll m. Talk with you soon.

Speaker B: Yeah.

Speaker A: Bye. Bye. It.