Pod Insights Gradient Dissent: Conversations on AI Building the future of collaborative AI development with Akshay Agrawal About this episode In this episode of Gradient Dissent, Akshay Agrawal, Co-Founder of Marimo, joins host Lukas Biewald to discuss the future of collaborative AI development.
They dive into how Marimo is enabling developers and researchers to collaborate seamlessly on AI projects, the challenges of scaling AI tools, and the importance of fostering open ecosystems for innovation. Akshay shares insights into building a platform that empowers teams to iterate faster and solve complex AI challenges together.
Follow Weights & Biases:
https://twitter.com/weights_biases
https://www.linkedin.com/company/wandb
Join the Weights & Biases Discord Server:
https://discord.gg/CkZKRNnaf3
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:02 Lukas Biewald? You're listening to Gradient Dissent, a show about making machine learning work in the real world, and I'm your host, Lukas Biewald. Akshay Agrawal is the CEO and co-founder of Meremo, a new kind of notebook that's been exploding in popularity. I myself am a big fan of the product. He's clearly passionate about building developer tools for AI engineers, just like I am, and it's super fun to nerd out with him. I hope you enjoy this episode.
0:32 Lukas Biewald? All right, so Akshay, yeah, I'm a fan of your product and user of your product, Meremo, and I'm sure it's going to get like, super in the weeds for like, Meremo users only, like, shortly. But I kind of wanted to keep like, at least the first few minutes of this recording accessible to more people. So I feel like maybe we should start with, you know, what is a notebook and why is a notebook important in the AI field?
1:02 Speaker 01 Yeah, 100%. So I view a notebook, and in particular, most notebooks these days are on Python, I view a notebook as a programming environment that allows for interactive computing. So it lets you have blocks of Python code and to visualize the output of each block of Python code. So you can have plots, you can visualize your tensors, you can visualize training runs, and you can interleave your code and your visuals with Markdown that documents the exploration, experiment, or whatever you've done. And for data work in particular,
1:44 Speaker 01 whether you're training a machine learning model, or whether you're hitting databases,
1:50 Speaker 01 it's really valuable to have a programming environment that lets you see your data while you work on it. And for this reason, notebooks are widely used across the sciences, but they're also widely used to train machine learning models, and they're the central interface for compute in a bunch of products like Google Colab, Databricks, and AWS StageMaker.
2:15 Lukas Biewald? And I think it kind of speaks to the fact, and this is one of the thesis is behind
2:21 Lukas Biewald? Weights and biases that the workflow of someone, you know, building an AI model is sort of different than a developer workflow. It's more exploratory. You kind of, you know, you write a lot of code that might not see the light of day or need to be like production ready. A lot of the code you write is to kind of learn, okay, what's going on with my data, what's going on with my model and things like that, right?
2:41 Speaker 01 Yeah, 100%.
2:44 Lukas Biewald? And, you know, Jupiter notebooks are kind of the, the place where, you know, I started, collabs are a little bit different, Streamlit's a little bit different, Meremo's a little bit different. Maybe you could talk about how you thought about Meremo and how it kind of fits into this world of notebook variants?
3:01 Speaker 01 Yeah, so I guess maybe to set the stage, I can briefly say what Meremo is and how it's different from Jupiter.
3:10 Lukas Biewald? Perfect.
3:10 Speaker 01 So Meremo is an open source Python notebook that's built from the ground up to, to solve what I view as key problems with Jupiter notebooks in particular. And so unlike Jupiter, Meremo notebooks are reproducible, they're Git-friendly, you can deploy them as interactive web apps, and you can execute them as Python scripts. So with Meremo, what me and Miles, my co-founder, are really trying to do is blend the best parts of interactive computing with sort of the rigor and discipline of traditional software engineering, as well as the developer experience associated with, with engineering. And you can think of Meremo as like a modern replacement for Jupiter, but also for, for Streamlit.
3:56 Lukas Biewald? And what do you mean by reproducible?
3:59 Speaker 01 Yeah, so there's two types of reproducibility. So there's an interesting study by Jetbrains, where they took 10 million Jupiter notebooks from on GitHub, downloaded them all, and studied whether if you reran them from top to bottom, whether you would get the same results that were serialized in, in the notebook. And over a third of them weren't reproducible, meaning that like they just, you would get different outputs. There was another paper from 2019 that did a similar study that found similar results. And like in my own experience, like when I was at Stanford doing my PhD, you know, my co-authors were amazing, great people, but not software engineers, would hand me a notebook. They're like, here's my code, and I would run it and like-
4:45 Speaker 01 I can't reproduce your science. So I think there's like this reproducibility crisis in terms of the code on the page doesn't necessarily match the outputs you see in Jupiter notebooks. And that's specifically because a Jupiter notebook is really kind of a dressed up REPL, like a Python REPL. And so it's just code that you executed, the output at the time, but it doesn't capture the execution history in a meaningful way. So that's one kind of reproducibility, that's the biggest kind. The other kind that we can talk about later is also package reproducibility. Like you may run an experiment, but not have all the packages you use documented clearly, so then others can't reproduce your environment.
5:27 Lukas Biewald? No, I mean, the initial experience of Marimo feels a lot like a notebook. What is different about what's going on that adds in reproducibility?
5:38 Speaker 01 So when it comes to, Marimo provides the user a guarantee, which is the code on the page matches the outputs you see. And the way that we do that is that
5:50 Speaker 01 Marimo looks like a notebook and it feels like a notebook, but it actually has like code intelligence built in. And like, when you run one cell, all of those, all other cells that referred to the variables that the first cell defined are automatically run. So it's reactive, not unlike a spreadsheet. And so that means that like, if you edit one part of your notebook, you will see downstream cells updating to stay in sync with your code.
6:20 Lukas Biewald? But it's only the cells below, right? It doesn't, does it update the cells above also?
6:23 Speaker 01 Above as well. So actually, how it works is that we statically parse all your cells to find the dependencies across variable definitions and references. And we form a DAG basically on them. And so actually the order in the-
6:38 Lukas Biewald? How do you keep it acyclic? I mean, couldn't there be circular dependencies here?
6:43 Speaker 01 So there's a contract basically between Marimo and the user. And one of the rules is you can't have cycles. And we can detect that. And so if you type in some code that introduces a cycle, we won't let you run it. And we'll say, hey, you have a cycle. It's across these cells. Here are some suggestions on how you can fix it.
7:01 Lukas Biewald? I see. And then what about this other stuff? So I guess like, what makes notebooks not compatible with Git? Is it the sort of like giant files with the output included that make it annoying for Git? Is that what you're talking about there?
7:15 Speaker 01 Yeah, yeah. So they've typically used JSON files, right? For storing the notebook code, but also the outputs. And you get this giant blob. So you make a small code change and you get a giant diff. So Meremo stores your notebooks as pure Python and we guarantee that small code changes yield small diffs.
7:37 Lukas Biewald? That makes sense.
7:41 Lukas Biewald? And so what's your thoughts on Streamlit? I mean, I know that, you know, when Streamlit started, it was kind of an exciting new thing that I think at least added interactivity to notebooks in a really cool way. You had the Streamlit founder on the podcast and you sort of talked about, you know, what he was thinking at the time. But what's your take on Streamlit? Like, why isn't that like a satisfying solution to some of these problems?
8:13 Speaker 01 Streamlit is a really cool project and I think there's a reason it took off so quickly. Because data scientists, machine learning folks hate writing any front-end code. So I think that interactivity was really delightful. It's not though, what's special about notebooks is, you know, as you mentioned earlier, that they're great environments for just exploring your data and interactively prototyping algorithms, querying databases, et cetera. Streamlit is not designed for that. Streamlit's when you're finished with all those things and you just want to create an application.
8:50 Speaker 01 And what I noticed is that a lot of Streamlit applications, if you go to their GitHub repo, there's an iPython file and then there's a Streamlit file and it's like a direct port. So Meremo bridges that gap, that every notebook can be seamlessly run because we have that, we understand the relationships across cells and Meremo ships with a bunch of interactive elements like sliders, data frame transformers, et cetera, from the command line. And I took this inspiration from Streamlit. You can say Meremo run my notebook.py. It'll just hide the code and run it as a web app.
9:25 Speaker 01 So I think to your original question, though, I think the reason Streamlit doesn't solve this full problem is that you can't start your data work in a Streamlit file, whereas you can in Meremo.
9:38 Lukas Biewald? Like you type one line and like it's going and you're seeing the results. Is that what you mean?
9:43 Speaker 01 Yeah, more like, so in Streamlit, if you interact with a slider or something, it runs your whole script from top to bottom. In Meremo, we have a really granular DAG. So you interact with the slider, it'll only run the cells that depend on it. And so you still have that fast iterative, interactive programming environment that I think many people have become familiar with after using Jupiter for so many years.
10:10 Lukas Biewald? Yeah, that makes sense. Anthony Goldblum was talking to me about his use of Meremo, and I think he's got the same workflow where he's got, you know, like notebooks, and then he would put them into Streamlit. So he was like so excited to kind of basically avoid that step in order to publish stuff. So he was telling me, he's like, I went from like six hours to like five minutes to deploy this stuff. And I was like, how is that possible? And then I realized, oh, because you're actually translating your notebooks like into Streamlit. And now like clearly you don't have to with Meremo to deploy them as web apps, I guess.
10:43 Speaker 01 Yeah, you just get it for free.
10:49 Lukas Biewald? Well, cool. I mean, tell me, like, you know, is, for you, I would think, it seems like adoption has been, has been pretty impressive. Do you, are you intentional about growing adoption? Or do you feel like, hey, I make the best product and it's sort of, you know, naturally viral because people share it as a notebook and then they use it? Or do you have kind of like, do you have sort of growth hacks that you're using to, to drive that adoption?
11:17 Speaker 01 We're trying to do both.
11:20 Speaker 01 So we're trying to lean in recently, in the past couple of months of like getting the natural virality. And people have started to share their notebooks more. We recently did an integration with Hugging Face, so it's really easy to deploy your Murillo notebooks as apps on Spaces. And we have like a WebAssembly-powered playground, so you can create unlimited notebooks on our playground for free and share them with a link. I think early days, we really started to spread of a post on Hacker News. Some of those really took off. Our showHN is like the second top Python showHN. You drill down and you eventually get to it. But so we got a lot of users initially through Hacker News. We haven't
12:04 Speaker 01 done too many growth hacky things. Maybe not, probably just because we're not expert growth hackers. But yeah.
12:14 Lukas Biewald? I feel like the showHN, when it works well, the comments can be bean. Was there kind of pushback from the Hacker News crowd?
12:23 Speaker 01 We were actually lucky. No, people were really quite nice.
12:26 Lukas Biewald? That's good to hear.
12:28 Speaker 01 Yeah. I think Simon Willison, creator of Django and now doing a bunch of cool projects, dataset, he commented and it was really positive. And maybe that set the tone or something. I'm not sure. People were pretty nice.
12:41 Lukas Biewald? You know, one of the things I noticed about your background is that you spent a little bit of time at Netflix. Is that right?
12:48 Speaker 01 I did an internship.
12:49 Lukas Biewald? Internship, yeah. And I was kind of thinking, you know, like I feel like Netflix famously like, you know, runs a lot of their production stuff on notebooks, I think, or at least that's what they were saying, you know, four or five years ago. Was that still the case when you were there? And did that inform any of your thinking on Murillo?
13:09 Speaker 01 So I believe it was still the case when I was there. But I actually wasn't in that part of the org. I was more on like algorithms engineering side. So actually I probably used, is your own notebooks that internship? It didn't really inform my thinking too much. I think they managed to make it work, but I think like...
13:33 Speaker 01 The way that Jupiter is set up is that you have to like jump through a bunch of hoops to make that work. And it is cool that they were able to. But most of my inspiration actually came from, I think, two projects. Mostly, actually, it came from a project from the Julia language called Pluto JL, which is an alternative to Jupiter that Miramos shares a lot of similarities with. So Pluto, observable by extension, Pluto is modeled after that, as well as Streamlit to an extent.
14:04 Lukas Biewald? Interesting. And what did you pull from these projects? Tell me about that.
14:08 Speaker 01 So Pluto I pulled a lot from. Pluto is a reactive notebook for Julia that has the same spreadsheet-style automatic execution. It lets you create sliders and other UI elements seamlessly. It has built-in package management like Miramos does too, so that your notebook files are a self-contained reproducible unit. So I pulled all these ideas from Pluto. I saw in, I think, late 2021, I had just finished my PhD. I'm like, this is an amazing project. In the Julia language, people are switching from Jupiter to Pluto. It's the highest, the most starred GitHub repo for the Julia programming language, aside
14:53 Speaker 01 for the Julia programming language itself. So to me, I saw this and I thought we immediately, I immediately thought we need something like this for Python. And then from Streamlit, what I pulled was, you know, I saw Pluto, I'm like, wow, these things look like web apps. And then I saw Streamlit, like, these things are web apps. And so I realized that, you know, you can kind of merge the two together and have a single notebook-like thing that lets you do the notebooking and then also do the creating of the data apps.
15:24 Lukas Biewald? Are you a fan of the Julia language? Like, how did you find yourself, you know, working with Julia notebooks?
15:32 Speaker 01 Oh, I admire what they're trying to do. I used Julia a bit, it's been a long time ago, during my undergrad. I did like math and machine learning at Stanford as well as computer systems. So I was, the idea was really appealing to have a single language that bridged the gap in the two-language problem between usability and performance. I think it's still a relatively new language, so it has some sort of...
16:01 Speaker 01 It has some things that make it difficult to adopt. But I think I actually, I came, I wasn't using Pluto notebooks. I, you know, I, at the end of my PhD, just did like a broad survey of a bunch of different tools for working with data. I just wanted to see what people were working on. And I don't know, I went down rabbit holes and somehow I found Pluto. And I was, yeah, I immediately fell in love with the project.
16:23 Lukas Biewald? Oh, interesting. So were you doing a survey because you're like, hey, I want to start a company doing tooling for machine learning?
16:31 Speaker 01 Yeah, that's right. So what happened was like, I, so I used to work at Google Brain, where I'd worked on TensorFlow. In my PhD, I did like machine learning and optimization, but a huge part of it was open source tooling for machine learning and optimization. And I realized the thing that I liked the most is building dev tools that let other people solve problems instead of solving the problems myself. And so I, yeah, that's why I did the, did the survey. So I wanted to see what, how can I, you know, use my systems background and my ML background to make, make the biggest impact?
17:04 Lukas Biewald? Cool. What, what else did you, did you like, were there any other, what other rabbit holes did you go down in that, in that survey?
17:11 Speaker 01 Oh, yeah, man, I, I should pull up my Obsidian Vault. It's, it's pretty gnarly.
17:19 Speaker 01 One of them actually was
17:23 Speaker 01 making it a lot easier to access cloud computing. And so
17:32 Speaker 01 in a few different directions. So one was like provisioning. And I think like actually Modal and like the folks from the Sky Computing Lab at Berkeley are doing this really well now. It's just, you know, every PhD student feels like, wow, it's absurdly difficult to run something in the cloud. I just want to ship this one function there and just run it on a GPU. And it's cool to see that happening. So that was one. The other one was kind of related to what I did at Netflix, actually, which was like studying like cluster management and like optimize, optimizing cluster usage to minimize costs and maximize utilization. I think a company called Anton Metal and probably a few others are, are doing this now.
18:17 Lukas Biewald? Interesting. And how did you choose notebooks?
18:20 Speaker 01 Notebooks?
18:21 Lukas Biewald? Yeah.
18:21 Speaker 01 You know, it was the systems part of me and who was like really attracted to these cloud computing problems. But then there was the part of me that likes designing for humans and individuals that was really attracted to the notebook project. And I think ultimately, after watching a bunch of people and myself, you know, doing research and like working with data, like, I saw that like, a lot of it starts in notebooks. And like, people love notebooks, but they also, they kind of hate them. Like, you know, you just, if you just do any like Jupiter notebooks site reddit.com, like you will find people
19:01 Speaker 01 just complaining about like hidden state, like, oh, God, the JSON file format. And like, people feel so strongly that something better should exist. And nothing existed. And once I saw Pluto, like, I was like, okay, there is a way to make something better. And this will have a huge impact on so many people. To me, it was clear that this is what I should do next.
19:26 Lukas Biewald? Yeah, it's funny. I remember, I think I saw an interview with the founder of the Jupiter notebooks. And he was talking about how he was inspired by Mathematica, which is something I remember from, you know, my math undergrad at Stanford. I don't know if people still even, even use it. But that actually had like some more kind of delightful features where you could, you know, kind of do more complicated inputs than just like a linear, you know, block of code. I'm curious if you have any experience with Mathematica? Do you see that as an interesting kind of notebook?
19:57 Speaker 01 Actually, I embarrassingly don't, probably because I was a CS major first and math minor second.
20:04 Lukas Biewald? Ah, I see.
20:04 Speaker 01 I do know that some of the folks in the math department are still, you know, or like who come from like more of a double E background, use Mathematica. But that, I'm, I'm sure Marie Ma has somehow has the imprint of Mathematica on it, but probably through indirection, through its impact on Jupiter and Pluto.
20:22 Lukas Biewald? It's kind of like a grandchild of Mathematica, I guess that is.
20:26 Speaker 01 Yeah.
20:28 Lukas Biewald? Cool.
20:30 Lukas Biewald? Well, look, I want to ask you, I mean, are you aware of some of the real world use cases? I mean, it seems like one of the really cool things of what you're doing is it's so broad. Like, what are some of the interesting things that people are doing today with Meremo?
20:42 Speaker 01 Yeah, so there's a lot. That is one of the fun things because it is so broad. So you talked about Anthony Goldblum. So he identifies, I think, one persona, and there's many people like him who do this, who first use a Meremo notebook to explore some data, but ultimately, what they end up with is like a little mini reusable app, like a tool that they can use on a daily basis to do like some analytics work or some kind of internal task. So one of them is, and maybe that's like sort of the Streamlit-style application of I'm creating tools for myself and for my company to do analytics and other things.
21:25 Speaker 01 What's interesting is that they're not just like, Anthony's case, they typically interact with data. But I've also had people reach out to me saying like, I'm a software engineer. I've never used notebooks before. I never wanted to. But Meremo appeals to me because of the developer experience and I'm using it to manage and build a dashboard for my EKS cluster or something like that.
21:50 Lukas Biewald? I'm so surprised to hear that because I feel like the first interaction with Meremo is just like a notebook. And at least for me, it took me a little while to even notice what the difference was messing around. Like, I just, are notebooks so offensive to a software developer? Like, it's hard for me to imagine that.
22:12 Speaker 01 Yeah, it's surprising. Like, a number of people come into our Discord and they're like, "Oh, yeah, but I haven't ever used a notebook." And when we try to get the setting context and I ask them about like, "Yeah, they're like, "Again, I've never used notebooks. I never wanted to." I think maybe it's because we mentioned, I think it's the Git-friendly and the stored as pure Python part. Software developers like, "Oh, okay, cool. That makes sense. Like, I can work with that." But you know, actually, a bit of a tangent, I can come back to your use cases that you were talking about. But you mentioned like this, I mentioned Git-friendly.
22:46 Speaker 01 Pleasantly surprised, but really quite surprised by like the extent to which like that is the hook that has brought in a ton of our users. You know, I thought it would be like the reproducibility or like the super duper interacted elements like so. For example, if you output a data frame in Remo, you automatically get like a filterable searchable table. Can page you the whole data, not just like a tiny static HTML preview like you get in Jupiter. I thought that was the thing that would bring people. But I keep on hearing over and over again, it's it's the Git friendliness. I gave a talk at UIUC at a business analytics class last week. And so I talked about all the data related features. I talked about how we have built in support for SQL.
23:30 Speaker 01 And at the end, the professor was like, I'm going to use this because it's Git friendly. And so I
23:37 Lukas Biewald? I'm still surprised, like, you know, I've used notebooks for years. I've used Git for years. I feel like the execution model of Remo is like really cool. Like deploying to the to the web is like super cool. The Git friendliness, I didn't even notice it. I don't know. I guess some people are like freaking out about about big diffs in their
23:57 Lukas Biewald? maybe I don't even like look that up.
24:00 Speaker 01 So I think they it's something, yeah, I think they get scared by the diff. Yeah.
24:06 Lukas Biewald? So how do you actually handle it? So like, I mean, do you not like store the output like, because what if the output is a giant table? Like, how do you?
24:13 Speaker 01 Yeah, so we don't store the output, but we have an option to store the outputs like alongside the notebook. So you can like, yeah, so you they're just in separate files and you may not track your HTML with Git. Or if you did, then you would just be like, you would know to just like ignore the diff and you could at least see the code diff.
24:34 Lukas Biewald? Because there is sometimes a nice experience with a notebook of like looking through it without having to run it yourself. So that's not a bad thing. Yeah.
24:40 Speaker 01 Yeah, so that's why we have sort of that snapshotting feature. But yeah, so I think a lot of people come for the Git friendly, but then they stay for all the other goodies that they discover along the way. Like they discover the execution model, the interactive elements, and all those things.
24:58 Lukas Biewald? Now, do you even mention the Git friendly in your website? Is that just like a word of mouth?
25:03 Speaker 01 We do mention it.
25:04 Lukas Biewald? Somehow I didn't even register. I gotta look at your website now. Oh, yeah, Git friendly. You're right.
25:08 Speaker 01 Yeah.
25:09 Lukas Biewald? Wow, who in the world is people are getting mad about the?
25:13 Speaker 01 Yeah. So I don't come from a business analytics background. I come from more of the machine learning background. But yeah, so I was surprised that the BI people also like it. I talked to someone else who does business analytics, I think at Capital One. And she was also like, oh, yeah, the Git friendly. And I think she was traumatized by having to deal with like Tableau dashboards, where like you have no idea like the provenance of what you're looking at. And if an audit comes, then like you get really scared.
25:40 Lukas Biewald? Yeah.
25:40 Speaker 01 I'm gonna stop talking because I don't know anything about that field, but.
25:46 Lukas Biewald? Totally.
25:47 Speaker 01 Yeah, so that was a tangent.
25:51 Lukas Biewald? I'm sorry, we're talking about use cases. Yeah, yeah.
25:53 Speaker 01 Yeah.
25:53 Lukas Biewald? It sounds like BI is an important use case for you. But I mean, more like practically, like, do you get stories of people using it in interesting ways?
26:01 Speaker 01 Yeah, so Anthony is building his little mini apps. He calls them mini apps. A lot of people use it as just a, but just a better Jupiter notebook. So everything you would do in a Jupiter notebook, whether it's like munging data frames or training a machine learning model, people will just use it for that workflow. And they'll find that the reactive execution and the UI elements, like sliders, dropdowns, et cetera, will just make them way more productive. I mean, the simple experience of having a data frame where like you can actually type into a search box and like get your filtered in real time, I think is just a big step up from.
26:45 Speaker 01 The experience you have in a in a G乎itter notebook.
26:49 Lukas Biewald? Totally.
26:50 Speaker 01 So that's one. And I think a lot of our users, when they first hear about Meremo, they get a little scared, especially the machine learning folks, by the by the automatic execution element. Because they're like, oh, I don't want to accidentally kick off a training job on GPUs, or
27:05 Lukas Biewald? Totally.
27:05 Speaker 01 I don't want to hit my OpenAI, you know, whatever endpoint.
27:08 Lukas Biewald? Yeah.
27:08 Speaker 01 So we do have, you can configure the runtime to be lazy, so that you can't, like, if you execute it. So the downstream cells will be marked as stale, and like, there's a visual queue and everything, but they won't automatically run. And then there's like a button to bring it back up, up, up to date. So you so you can, and people do use it for for the more expensive sort of computation as well.
27:35 Speaker 01 Yeah.
27:36 Lukas Biewald? Well, okay, so tell me about how you think about AI coding inside of Meremo. Another kind of notable difference that maybe you don't highlight as much is that, you know, you have code generation like built right into the to the application. It's kind of cool. It seems like the code generation actually looks at your data as well as the code, which is cool for, you know, a data centric application.
28:04 Speaker 01 Yeah, so we do have a built in like AI integrations. We really try to be batteries included across a lot of dimensions. So here, there's like built in support for GitHub Copilot and Codium. And then you can also, if you bring your own keys, you can generate, generate code using the context of your notebook. Excuse me, as well as the schemas of your data frames or your attached DuckDB tables. We should probably highlight it more.
28:35 Speaker 01 Yeah, we should probably, you know, I was when I was talking to Anthony at a table, and he was talking about some other product, and he's like, oh, yeah, what it's really cool is that like, that product, you know, I told it to generate some map plot lib code and then change it to plotly, and it did. And I was like, oh, you can do that in Meremo. And he's like, what? Oh, so, so, yeah, we should highlight it more.
28:55 Lukas Biewald? Do you find it funny, I actually started doing notebooks and Streamlit inside of Cursor just because it was so fast to develop it and I was kind of used to the code tools. Could you also develop Meremo inside VS Code or something like that? Is that a common pattern?
29:13 Speaker 01 Sort of. We do have a VS Code extension that does like a split pane kind of thing where like you can type in the code cells on the left and it like auto runs the notebook UI on the right. I think we have some more work to do to make that feel more seamless. It's a little tricky because our, I guess our, it's not quite an IDE, but it almost is. Like our editor has so many features, like from the Data Explorer panel to like live documentation, as well as the built-in AI stuff that I feel is best experienced. And I feel like a lot of folks went from, yeah, I feel it's best experience in the editor,
29:57 Speaker 01 but people do ask, like, oh, can I use it in Cursor? Can I use it in VS Code?
30:02 Lukas Biewald? Well, I'm feeling like the growth team telling me that this podcast is really kind of going off the rails, but I'll just take it further off the rails and just ask my own tech support questions here. Do you, like, I think one of the great things about Cursor is you're kind of chatting, which doesn't feel like Meremo is oriented towards like a chat code generation. Is there a way to sort of like ask it more broadly, like, hey, what's going on? And can you fix this up? Like maybe like Anthony was saying, like, you know, convert map plot to Plotly or whatever.
30:31 Speaker 01 You can do that on a per-cell basis. Like you can refactor the cell. I think as we build out our roadmap for the next year, like, part of it is IDE features and part of that will be more modern AI features, such as maybe like a Cursor-like experience.
30:47 Lukas Biewald? That makes sense. What else is on your roadmap? I mean, how do you even think, are you a two-person organization now?
30:53 Speaker 01 Yeah, we're two people. We did raise a seed round, which we just announced this week.
30:59 Lukas Biewald? Congratulations.
31:00 Speaker 01 Thank you. Yeah. So we are looking to grow the team, but we've been just two people from now, me and Miles. For things that, there's a lot of stuff we want to build. Some of them are Better integration, more thorough integrations with SQL. So I mentioned, I alluded to that Merima supports SQL. And so Merima is pure Python. It's stored as pure Python, but you can embed other languages in it. One of those is SQL. And we actually build a data flow graph across both Python and SQL, which could be a conversation of a whole other podcast. But the point is, like, in your SQL, you can query data frames and you get a data frame out. And this is powered by Ducky B.
31:43 Speaker 01 One thing we want to do is, I think, build out that experience more. So like, make it easier to connect to other types of databases as well, whether it's Postgres or whatever you're querying. Have richer, like, column previews so that you can, like, explore all your data at a glance without having to actually execute queries or very many. Other things on our roadmap are exploring some of the infrastructure side of things that I was mentioning that I was interested in early on. I think there's some interesting things related to hybrid execution. I think a common model is a lot of your notebook in code can be done locally,
32:23 Speaker 01 except for some really expensive cells, which you may want to ship off to remote servers. So that's one thing we're looking at, just making it easier to run expensive workloads. And then also better support, making it really, really easy to share Merima notebooks in two ways. Like, we're building out a community cloud, which is free because it's powered by Pydide and WebAssembly, meaning it runs entirely in the browser. And we may also look into, like, static site generation. So like, in your make docs or like Sphinx documentation, you'll have Merima notebooks running with live code, interactive widgets, all those things.
33:04 Lukas Biewald? And I guess that that tees up an obvious kind of investor question, I guess, which is, you know, if you look at, you know, I think, I mean, Streamlit obviously was was sold for a lot of money, but it was sold kind of pre-monetization or just as they were like starting to monetize. And I don't know if they ever really got it meaningfully monetizing. And Jupiter Notebooks, obviously, like an incredibly successful, you know, project that was never intended to be monetized. So it does sort of seem like maybe this, we don't have quite figured out the right, you know, monetization model for this type of thing. Like, how do you think about, how do you think about that?
33:43 Speaker 01 Yeah. So I think it's, the way that Meremo is different from Streamlit is that Streamlit usually really only came in at the very end of like, a data project, like for the, some small fraction of what you wanted to deploy an app. And so maybe this is, they were focused on just the app deployment. What's I think compelling about our, our product is that Meremo is where you start. It's where you first start in querying data, where you first start training a model, et cetera. And you can also carry it, carry it all the way towards the data app. So I think, I also mentioned earlier that like notebooks are the central
34:25 Speaker 01 interface to compute in a lot of commercial products like Databricks, SageMaker, et cetera.
34:31 Lukas Biewald? Probably.
34:31 Speaker 01 So I, I think there may be an angle for us to lean into, to there as well. Like the surface area of the user workflow that, that we can monetize is, is much larger than Streamlit. And Jupiter, I think intentionally started as a nonprofit and really had, had no intention to, to go out into any of these infrastructure type, type problems.
34:54 Lukas Biewald? Did you ever consider a structure like that? Or were you always, hey, I want to be a venture backed startup?
35:01 Speaker 01 I didn't really, you can just build a lot faster by just working on it full-time with a bigger team. So, yeah, I was sort of thinking about startup from the beginning. Although I did start with, this led me and Miles really focus on crafting, I think, a delightful product. We started with funding from Stanford Slack National Laboratory, which was enough for just the two of us for two years.
35:32 Lukas Biewald? This is a linear accelerator?
35:33 Speaker 01 Yeah.
35:34 Lukas Biewald? Cool.
35:35 Speaker 01 And so, there were some scientists there. I was having coffee with one of them one day. Should knew him back from my PhD. And he was like, "What are you working on?" And I was like, "Oh, I want to build this notebook thing that also lets you make apps and run a script." And he's like, "That's really cool. We really need that. We use Jupiter and it's so many problems. We'll pay you to do it." And I was like, "Oh, that, that, that sounds good." And so, we worked with them for two years. And they gave us a lot of early feedback before we open sourced it. It was, I think, really essential for us to make a product like this.
36:10 Lukas Biewald? You know, Marima feels like such this like kind of delightfully simple thing. But I'm sure there's so much like thought and work that goes into it. Like, were there, have there been like kind of tough product choices along the way? Or paths you went down that turned out to be dead ends?
36:32 Speaker 01 Yeah, I think there have been a few. So, there's definitely tough technical challenges. In terms of product challenges, I think the biggest one was like...
36:45 Speaker 01 So, we've talked about how like Marima notebooks are reactive. You run a cell, dependent cells automatically run, or a Merk test still. And then you brought up the objection that like, "Hey, that means I can't have cycles."
36:59 Speaker 01 It also means you can't define the same variable twice across two different cells. Because if you do that, then Marima doesn't know which definition to run first, and your outputs won't be reproducible. So...
37:12 Lukas Biewald? I found that very confusing, my first time user experience, for sure. That was a
37:17 Speaker 01 Yeah, it's a big break from traditional notebooks.
37:20 Lukas Biewald? Totally, yeah.
37:20 Speaker 01 It's a big break from traditional notebooks. But we felt that was totally, we felt that was necessary for us to have in order to have the amount of reproducibility that we wanted. And there was precedent, like I saw PlayDoh did it. And PlayDoh is doing very well in terms of open source adoption now. So I'm like, okay, people can get used to this. So we stuck with that. We chose, okay, Meremo is a DAG. Essentially, Meremo is a DAG on notebook cells and it's going to have these constraints. We'll try to explain them to our users. And, but, but, but it's going to have these constraints. But we got a lot of people, and a lot, but we got a number of people coming into our Discord, basically asking us,
38:01 Speaker 01 can you not, or like, can you have an option to like, can you, can there be a toggle for me to turn off the DAG? So that it's like Jupiter again, and I can just, you know, type random code, whatever I want, and then toggle the switch back on later?
38:16 Lukas Biewald? Yeah.
38:17 Speaker 01 We, we considered that, but we, and you know, they're like, you know, their reasoning was like, that will only increase the amount of stuff you can do in Meremo. But I think it would also like lead to like a, if you look, think of like the larger ecosystem, then like people would share notebooks with you that were developed in, I don't know, dangerous mode or something. And then those wouldn't work. Like the DAG is necessary for the notebook to run as an app and be run as a script. So then all those downstream things that we enabled would have broken. So in the end, we decided to politely say, no, we, we can't, we can't do this. Sorry, like, we cannot let you turn off the DAG.
38:56 Lukas Biewald? Okay, but tell me, okay, now we're going into like personal Lucas tech support. But what do you do? Like, what's your workflow, like when you want to like quickly, like tweak one thing, but not necessarily like save it? So you want to copy into another cell, like some block of code? And just like, see what happens?
39:13 Speaker 01 Yeah.
39:13 Lukas Biewald? And all the variable names get copied?
39:16 Speaker 01 Yeah.
39:16 Lukas Biewald? How do you, how are you supposed to do that?
39:19 Speaker 01 Two, two, two recommendations.
39:23 Speaker 01 So one would be.
39:26 Speaker 01 I tell folks to try and use functions as much as possible. Functions introduce a local namespace, right? So variables defined in a function don't get added to global, so you can have as many of those redefined as you like.
39:40 Lukas Biewald? Yeah, that makes sense.
39:41 Speaker 01 First, I'd say, Lucas, wrap that code into a function and then, you know, call it multiple thing, which requires some more, you know, a little bit more engineering wrangling, but it can be done.
39:49 Lukas Biewald? Right.
39:50 Speaker 01 And then, but if you don't want to do that, what you can do is you can prefix a variable with an underscore. And in Merima, what that does is it makes it local to the cell.
40:00 Lukas Biewald? Oh, I see. Cool.
40:01 Speaker 01 So then you can reuse that name across multiple cells. But it, but yeah, because it's local to a cell, then you can access it in another cell. And so you don't have any reproducibility issues. So that is the escape hatch.
40:15 Lukas Biewald? Well, look, I, I feel really optimistic for you. I feel like you made a really beautiful product that's getting incredible adoption. So it's exciting to catch you like kind of early in the journey. It'd be fun to, to check back in in a year or two and, and see you right yet. But I really appreciate you taking the time to, to talk to me.
40:33 Speaker 01 Of course. Yeah, it was a pleasure. Yeah, working at Merima is definitely a dream job. So, yeah.
40:39 Lukas Biewald? Awesome. Well, good luck rooting for you.
40:43 Speaker 01 Thanks, Lucas.
40:44 Lukas Biewald? Take care.
40:45 Speaker 01 You, too.
40:49 Lukas Biewald? Thanks so much for listening to this episode of Grading Descent. Please stay tuned for future episodes.
Previous Evaluating LLMs with Chatbot Arena and Joseph E. Gonzalez Next DeepSeek, Stargate and AI's $600 Billion Question with Sequoia's David Cahn Page updated November 5, 2026 · transcript and summary are machine-generated