911: The Future of Python Notebooks is Here, with Marimo’s Dr. Akshay Agrawal

28:23Jon Krohn:And that actually is a great segue to my next question, which is around, you've previously said that, that there was a joy to rethinking a system from first principles, as you said about creating Merimo. So, you know, you get to have an opinion on the two, on how the tool should be designed. And so it sounds like if our research is correct, you began Merimo with a 2500 word design document so yeah tell us about that experience of you know coming up with a with this opinionated tool from scratch what that was like oh yeah it was a ton of fun um so this was like right after I finished my PhD so uh 2022 at that point like we talked about like I used notebooks a ton during my PhD there were still some gaps and so when I got the opportunity to do is that you know i had some free space and like i got to just like study the landscape you know got to you know really dig into like what do people like about notebooks interactivity seeing your seeing your outputs live and also like what are people doing like in other ecosystems like what kind of innovations are you know what kind of ideas are people playing with but what are they innovating on and so that that process took me through um uh i did research basically right and like i that's when i discovered the pluto notebook for the julia programming link which would felt like something out of like the future to me because like it has these concepts so pluto predates marima and it's a big source of inspiration it has reactive execution it has ui elements it's stored as a julia file um it doesn't have the web app thing but i looked at that and i'm like well wow that's like really close to a data app and you know we can build that functionality in.

30:12So that was a huge inspiration for me. Python has its own challenges of like making something like that work. So then the next step was, yeah, like writing like a design doc, like, well, how do you, it's like, you're threading this needle, like, how do you design a batteries included experience where you have a notebook, it doubles as an app, and it triples as a script or as a module right and like what i didn't want was like a bunch of like kludges to like you know it had to feel seamless and it had to feel frictionless and that's why there was like this design iteration and i was lucky though because i wasn't designing in a vacuum so i guess part of what we didn't chat about so like before we got funding from aix ventures um marima was built sort of with like close input from researchers at stanford slack national laboratory department of energy laboratory as well as with some like uh industry research researchers at a couple of startups um and so like i would have ideas and like throw them at them and they would give feedback and like you could see some of these early design documents and like some of the ideas were just totally horrible but like i got to work with them to like refine them and make them a lot better i think that was that was really special time nice that does sound like a great experience

31:29Jon Krohn:and yeah, hopefully something that's kind of inspiring for any of our listeners out there that are thinking about building a product. It sounds like you had a great way of getting going on that, thinking about the best parts, the worst parts of kind of the development experience for data scientists. You've put a great product together. It's really cool. Something that we talked about earlier, right at the onset of the episode, and that would have come up, I'm sure, as you were developing the product in that kind of Word document is this reproducibility issue. So the Jupyter notebook has been a very standard tool for a decade now for data scientists, but you've cited a study previously showing that only 4 % of Jupyter notebooks reproduce their original results due to issues that you called a hidden state trap, among other kinds of problems.

32:23Jon Krohn:Do you want to dig into these kinds of problems with reproducibility that Jupyter Notebooks have and how Merimo resolves them? Definitely. So just to caveat this discussion, I think every notebook does have its place and with Jupyter Notebooks, I think they are well suited to REPL workflows where maybe you're not shooting for reproducibility, but you're quickly really hacking things. The issue is though then, I think, because it's such a great environment for interactive coding, people started using them and really relying on them for issues where first use cases where reproducibility is paramount like data engineering like science like machine learning and so the question was dig into some of those reproducibility issues right and how marimo solves them um so two main classes so one is the issue of hidden state where in a jupyter notebook you run a cell jupyter is not going to run the cells that depend on it like the cells that use its variables and it's up to you to remember all the complicated dependencies that might flow through your notebook and you're going to forget some i i often did right and so now your notebook is in some weird inconsistent state so that that's one issue i to make that like really like pronounced i guess is like so in a traditional notebook like jupiter you can like say like you have a cell that like it's got a bunch of code in it um and maybe one thing that that cell does is like create some like PyTorch model class, you know, like instantiate some class.

33:51And then say, so you delete that cell. And then you continue coding elsewhere in the notebook. You deleted that cell, but say you didn't realize that that was the cell that defined the model class. And you really wanted that model class around. But in Jupyter, if you delete that cell, that model is still in memory for the time being. And so your rest of your notebook will work as you want it to work. But you come back the next day and you run the notebook from scratch nothing works and you're like what happened and it's like oh crap i deleted the cell that defined the model and like i really needed that so in remod if you do that if you delete the cell that defines your model variable it'll tell you it'll remove the model from memory and like it'll invalidate the other cells you'll be like yo that that variable is no longer around like you can't do this anymore and so it'll catch bugs like immediately when when you introduce them uh so that that's one big way and the second big way is package management and So Marimo has a special sort of, I guess, opt-in package manager, which I think is really neat.

34:54So basically, if you start Marimo in a mode that we call sandbox from the command line or however, every time you install a package, Marimo has a very nice slick package installation UI. we will save the package that you installed and the version as like a comment or an entry in the notebook file itself. And now when you come back to the notebook a second time and try to run it, and you do run it, Marimo will create an isolated virtual environment for you that has just those packages. So that means that like you can just send a single notebook file around and like people can just run it without even thinking about what packages they need to install, which makes them more reproducible, but also they're just like a lot more portable.

35:39It's like really easy to just create these single standalone tools and share them.

35:42Jon Krohn:I had more aggressive head nodding there while you were describing all that. I'm going to have to start using Marimo myself. Very, very cool. So something else that I think is really cool about functionality that we've learned to that I want to, it's something that's so visual and so easy to understand. It's something that would feel to me like magic as somebody who has really only used Jupyter Notebooks before for this kind of script development. So it sounds like it's possible, correct me if I'm wrong, that because of the way that you've thought about data analysts, data scientists, data people's experience from the ground up with developing this tool, it allows you to do things like highlight, select the data with your mouse in a scatterplot and then get it back as a data frame.

36:35Jon Krohn:Is that right? Yeah, that's right. Yeah, that's an example. That's crazy. Yeah, it feels really nice. It's like, you know, and I really needed this during my PhD and I didn't have it, right? So it really enables like really, really tangible like data exploration, data analysis workflows that were like really, really hard, if not impossible to do before. And so how do you decide when you're doing product development development how do you decide i guess to put something on your roadmap at all or to you know then prioritize highly a feature like this where yeah like how do you how do you decide you know which of these kinds of force multipliers to include in your product there's two angles i guess especially earlier on right like when we didn't have that much feedback and it was me and my founder just developing and jamming like with our built-up experience of like seeing like okay like here are some issues that we've hit with working with traditional notebooks and we know a lot of others have like we just we have strong opinions and we think this will land and we're just going to trust our gut so there was a lot of gut trusting in the beginning and there still is um now with the product being you know more mature um there's still gut trusting for like big new you know, features that we're working on, but we have a big community too, right?

37:57And so we can chat with them. And actually they come to us. They're very vocal about what features they want and they don't want. And I think like the way we think about it is like, is this something that like enables a broad class of users? So like data scientist, data engineer, ML engineer, AI engineer to be far more productive than they would have been otherwise in their sort of previous tool of choice. And I think like that, that is like something we call them like big rocks. Like what are the things that really moves the needle for folks? And it needs to move the needle a lot. And it needs to move the needle for a lot of people.

38:38Of course, we care a lot about craft and design and visual design and usability. And so those things are just like always top of mind and they're like ongoing. but for like the big new features that they have to be big rocks sort of how we think about it

38:50Jon Krohn:nice i like that big rocks and so you talked about community and how you can leverage them so open source projects often promise community but not every project earns it so what do you think separates tools that spark tons of devotion like the pluto project that you were describing there in Julia. What do you think separates those from open source communities that fade? And what are you going to do to ensure that you're in the first category? Yeah, it's a good question. I think, honestly, it sounds basic, but a lot of it comes from the maintainers just being like really kind and open like you gotta you gotta be open like if you if you want community you have to you know you can't just say i want community but like there are people out there who are excited you need to encourage them to like contribute and like be really really vocal about like how much you appreciate your community and also like help them make prs if they want to be you know super responsive on like your discord Like we try to respond really quickly to issues and especially like bug, bug reports.

40:07Like one of the feedback we usually get from our community who like file issues is that they're like shocked by how quickly we fix their bugs. Like if someone files a bug, you triage it and like fix it and ship a release the same day. Like you've kind of won a supporter for life. Oh, wow. Yeah. That's the thing that we've noticed something we've also noticed like from other popular Python projects, like from Charlie Marshall, Astral's UV project. And I think my community is many levels, right? Not everyone has to ship code to your project, although many people may want to, but just encourage all kinds of engagement.

40:45Yeah, and just make it a fun place for people to hang out. Nice.

40:49Jon Krohn:If we have listeners who would like to contribute to the project, how should they get started? There's a few ways, and it depends on how you want to contribute. So if you want to contribute code, you can check out our GitHub issues. And some of the issues are tagged as like good first issue. And like these are like great places for new contributors to kind of get their feet wet and learn what the code is like. And some of them are even like improved documentation. Which is, by the way, just generically a really good way to start contributing code to a project. Improve the documentation. other ways you can contribute is you can just like file a feature request or a bug a bug report we have a little check box saying are you willing to submit a pr for this and if you are and if it aligns with our roadmap then we'll like work with you um and then more generally we have a discord so you can get the discord link if you go to marimo.io slash discord and there like we have lots of free-flowing conversation.

41:50And I think that there's a lot of good touch points to get involved.

41:54Jon Krohn:Fantastic. And what if this is a listener's very first, you know, what if they've never contributed to open source before? Would you recommend, yeah, where would you recommend they get started? How should they get their feet wet with open source development? Yeah, so actually we do have, I think, a number of contributors who have made their first ever contribution with us. For that, I'd say like, you see a typo in our docs, something like that. I think that's a really good way to contribute. The docs ones are...

42:28I think I could be wrong, but I think if you're just making a simple change, you might be able to click edit in the repo itself, like online on github.com, and it'll create your fork for you and kind of simplify that process. But yeah, make a docs change. We have a contributing.md guide in the GitHub repo that tells you about, oh, you'll need to make a pull request. That's a pull request. And it'll walk you through that workflow.

42:55Jon Krohn:Nice. All right. And then that brings me to maybe one of my last technical questions for you, which is, this is completely beyond Merimo. And so this is, going back to the research that you've done, you focused on machine learning and optimization. and as an engineer, you've contributed to several open source projects from the deep learning framework TensorFlow that probably most of our listeners are familiar with to the vector embeddings computation library. You're going to have to tell me how people pronounce this, but it's PyMDE. Perfect. Oh yeah. Okay. All right. So it's, yeah, it's P-Y and then the letters M-D-E.

43:36Jon Krohn:And I'll have a link to that in the show notes. And so, yeah, so that vector embeddings computation library, PyMDE, as well as there's a convex optimization parser, which is also just letters, CVXPY. CVXPY, yeah. CVXPY, that makes sense. And so, you know, half of your published research relates to convex optimization. For our listeners unfamiliar with the topic, could you expand on how convex optimization differs from machine learning and the kinds of questions that they can answer? Precision, interpretability, scalability. Yeah, tell us about convex optimization. So mathematical optimization in general is like you have some variables that you're trying to make an assignment of values to.

44:29So, for example, say we're trying to choose a good stock portfolio. The variables is how much to invest in each stock. And then you need something that tells you whether or not your assignment to variables is good. So that's a mathematical function. It's an objective function that says, hey, what do I predict my return will be if I make this investment? And maybe it trades off. It factors in risk into it as well. And then you have some constraints like, well, I can't invest more money than I have. and maybe I'm not allowed to short stocks. So you have some constraints. And mathematical optimization is then the process of finding an optimal assignment of values to the variables to minimize the cost or maximize the reward while satisfying constraints.

45:22That's mathematical optimization. Convex optimization is just the subset of those problems we know we can solve super efficiently, super reliably, provably. And the use cases are somewhat different. So in some sense, like machine learning, a lot of especially classical machine learning, logistic regression, SVMs, all these are actually under the hood. They're using convex optimization techniques to fit the models. In terms of like use cases out in the wild, like what are people using CVXPy for today? It's well, well, one huge one is financial portfolio construction. So like many billions of dollars daily are like allocated through portfolio optimization problems, which are solved with convex optimization problems.

46:11Energy management, like a lot, there's a lot of usage there as well. Like real-time control, like controlling a vehicle, landing a rocket, like SpaceX uses convex optimization to land, I think it's like the Falcon or something, using software developed by our lab. And so these are cases where you can model the world and you have some understandable constraints and you can really exploit the structure. Machine learning, sometimes you, there's typically not many constraints that are kind of implicit. You don't really have as good of an understanding of like, well, what is the model doing? And also, you're just trying to find a solution.

46:51that's kind of good enough that you're going to test it out in the wild on unseen examples. And so the use case of finding the optimal assignment, it doesn't even necessarily always make as much sense. That said, there's a lot of overlap. So the vector embeddings library that you mentioned, for that to fit these embeddings using a GPU, even though the problem was in the machine learning domain, we use techniques from convex optimization to solve it really, really efficiently.

47:23Jon Krohn:On this podcast, I'm always going on about how Claude Code is mind-blowing, but now Claude Cowork is making my jaw drop as well. For example, I recently wanted to quantify how healthy my sales pipeline is for my AI consulting business. I simply asked Claude to estimate my sales for the coming quarter, and it brought info from relevant Google Sheets and my Gmail to create a professional spreadsheet of clients with estimated revenue for each one. Whoa, this might have taken me a day. Instead, it was done flawlessly with Claude Cowork in minutes. Claude is the AI for minds that don't stop at good enough.

47:54Jon Krohn:It's the collaborator that actually understands your entire workflow and thinks with you. Whether you're debugging code at midnight or strategizing your next business move, Claude extends your thinking to tackle the problems that matter. Ah, and you'll appreciate that I can ask Cowork to show me data, such as my sales spreadsheet, and it provides an interactive chart right in the conversation. For problems worth solving, get started with Claude at Claude.ai slash That's claude.ai slash superdata. And check out Claude Pro, which includes access to all of the features mentioned in today's episode.

48:24Jon Krohn:Claude.ai slash superdata. Nice. That was a beautiful explanation. As I've actually all of your explanations today, you do a great job of taking complex concepts and making them seem really approachable. And it isn't just the language that you use. You also have this really accessible tone. You make everything just kind of seem light and relaxed. I really like that. It's been a joy interviewing you. Actually, on the note of you making things feel so accessible, on your personal webpage, so akshayagrawal.com, which I'll have a link to in the show notes, it says that your goal is to make machine learning and math accessible and actionable.

49:09Jon Krohn:What does actionable mean in this context? What's the biggest gap today between mathematical tools and being able to take action with those tools? oh i love that question like when you you're in school or you're taking a course especially like a traditional course i.e most of them like you're like you know they're teaching you like okay this is logistic regression and then like you write down like the optimality conditions by hand you're computing some gradients and you're like okay great like i did a bunch of I don't know homework on my paper and I submitted a piece that that's fine but and that's good like I've done a lot of that in my life but like to make something actionable and to make like I think like what's really cool about our field is that like all that math you can do real things with it and like that's why like PyTorch and TensorFlow and Jax and like PyMDE and CVXPy all exist it's about like you know using concepts from math and and in order to like affect real change in the world um and so like that's sort of i guess been the theme of like the kind of projects that i've chosen to work on from tensorflow to cvxpy pymde and marimo like even though there is no math necessarily in the marimo code base like what it does is give you a really really tangible interactive environment to work with your data.

50:42And so it's making your data actionable. Like it's actually useful and like you can actually run your notebook as a script or share it as an app or reuse the code in it. And so like, I guess that theme is sort of what resonates to me.

50:56Jon Krohn:For sure. You're building a platform to make the change that you want to see in the world. I've loved everything you've said about the Marimo product today. And yeah, I'm so glad that Sean introduced us. It's been a great episode. Before I let you go, Akshay, I always ask my guests for a book recommendation. And actually, I can see and our YouTube viewers can see a huge bookshelf of books behind you. So what have you got for us? OK, I can recommend a couple. One that I most recently read that really stuck with me, as well as one that I'm currently reading. So I most recently read American Prometheus, which is the biography of Robert Oppenheimer.

51:38And so like most people I watched, like many people I watched Oppenheimer, like I want to learn more and I want to learn a lot more. So I read the book and it was really fascinating because it like actually like, you know, talks about. It talks about not just like, you know, the events that happened and not just like the witch hunt of Oppenheimer that happened after after the war, but also like what is like the social process of like doing science and like what are the social factors that influence? someone you know who is as much of a you know genius as Oppenheimer was to like make certain decisions that like led him to like develop the atomic bomb and I thought just like having that broader context was like super interesting so that's one another one that I'm currently reading that I would recommend because it's already making an impact on my life is the design of everyday things so a classic and just like product design it's taking me a long time to read it because I'll go like two pages and like, oh man, like I need to go and like improve something in Marimo.

52:40Um, and so like, if you want something, if you, if you, if you work in design or just are like observing the world, um, I think that that's a good way. It's already sort of made our product better.

52:51Jon Krohn:Fantastic. Two great and very different recommendations for us. Love it. And then for people, our listeners like me who have really enjoyed this conversation with you today, what's the best way to follow your work or connect with you going forward so that we can continue to get your thoughts after the episode? Definitely. So I am on the major social media platform. So me personally on X, my handle is Akshay K. Agrawal. I sometimes post under the official Marimo channel, which is Marimo underscore IO. I'm also on Blue Sky as well. Those are the best ways to follow me personally. And then Marimo also has various social channels.

53:40So we're on YouTube, we're on Blue Sky, we're on Discord, and we have a newsletter, marimo.io slash newsletter, if you want to subscribe, which I personally write once a month. Nice.

53:52Jon Krohn:Love that. Thanks so much for taking the time out of your busy founder schedule with us. Running an early stage tech startup like this must be exhilarating, but also very time consuming. So it means a lot to me and to our listeners that you took that time out. And yeah, thanks for providing us with such a great episode. Thanks, John. Really appreciate it. It was a blast.

54:16Jon Krohn:I hope you enjoyed that episode and I hope you liked the shorter intro to today's episode. Reach out to me on LinkedIn with a DM or a comment if you weren't happy or if you have any other ideas how we can improve the intro or any other part of the show. Really, I always love to hear from you. We assume that if you listen to the entire episode like you have today, that you'd probably still like the full outro that we usually do. So here you go. In today's episode, Akshay Agarwal covered the core reproducibility problems with Jupyter notebooks, such as hidden state traps where cells can run out of order, JSON file formats that break version control, and missing package dependency tracking that makes notebooks impossible to share reliably.

54:57Jon Krohn:He talked about how Merimo's reactive execution model that automatically runs dependent cells when variables change ensures your notebook state always matches what's displayed on screen. He talked about advanced interactivity features in Merimo, including UI elements like sliders that trigger real-time updates, the ability to select data points directly from plots and receive them as data frames, and instantly, an instant conversion from notebooks to deployable data apps. That's cool. Akshay also talked about the technical innovations that make Merimo notebooks stored as pure Python files, enable command line execution, function imports, and seamless integration with existing development workflows.

55:33Jon Krohn:and finally we talked about how you can get involved with open source projects like these Maribou notebooks if you'd like to as well you can get all the show notes including the transcript for this episode the video recording any materials mentioned on the show the urls for Akshay's social media profiles as well as my own at superdatascience.com slash 9-1-1 thanks to everyone on the Super Data Science podcast team, our podcast manager, Sonja Breivich, our media editor, Mario Pombo, our partnerships team, which is Nathan Daly and Natalie Zajski, our researcher, Serge Massis, writer, Dr. Zahra Karche, and our founder, Kirill Aramengo.

56:12Jon Krohn:Thanks to all of them for producing another excellent episode for us today. For enabling that super team to create this free podcast for you, we are oh so grateful to our sponsors. You can support the show by checking out our sponsors' links, which are in the show notes. And if you're interested in sponsoring an episode yourself, You can get the details on how by making your way to johnkrone.com slash podcast. Otherwise, please help us out by sharing the podcast, sharing this episode with someone who would enjoy this episode. Review the episode on your favorite podcasting app or on YouTube or wherever you watch it or listen to it.

56:48Jon Krohn:Subscribe, obviously, if you're not already a subscriber. But most importantly, I hope you'll just keep on tuning in. And I'm so grateful to have you listening and hope I can continue to make episodes you love for years and years to come. Till next time, keep on rocking it out there. And I'm looking forward to enjoying another round of the Super Data Science Podcast with you very soon.