How does Marimo handle dependencies and reproducibility?

Marimo integrates with the UV package manager to store dependencies inline within the notebook header. This allows a single notebook file to contain its required packages, making it highly reproducible and easy to share.

Can I convert my existing Jupyter notebooks to Marimo?

Yes, while the file format is different, you can copy-paste cells over or use the `mimo convert` command-line utility. Marimo's conversion tool also helps resolve restrictions like redefining variables across cells.

What are some of the interactive features of Marimo?

Marimo allows notebooks to react to UI interactions and vice versa. For example, selecting data points in a scatter plot can trigger updates in downstream code, and custom widgets can be created for unique interactions, even controlled by devices like a PS5 controller.

How does Marimo's AI assistance work for code generation?

Marimo has a 'Generate with AI' button that can leverage context from your notebook, including dataframes and variables. It can generate code for tasks like creating plots or performing data analysis using natural language prompts, without needing to switch to external tools.

Can I run Marimo code directly in my browser?

Yes, Marimo can run entirely in WebAssembly, providing a browser-based notebook playground. However, certain complex packages like PyTorch may not be compatible in this environment.

What is MoLab and how does it improve upon Marimo's web capabilities?

MoLab is a free, cloud-hosted notebook environment by Marimo that runs on cloud instances. It supports any Python package, configurable resources, and is designed for easier sharing, addressing limitations of WebAssembly and fulfilling user requests for a Colab-like experience.

How can I get started with Marimo?

You can install Marimo locally by running `pip install MIMO` and then `mimo tutorial intro` to get started. Alternatively, try the cloud-hosted MoLab at moab.marimo.com for a quick experience.

⚡️The Future of Notebooks - with Akshay Agrawal of Marimo

Latent Space Podcast

Science & Technology6 min read29 min video

Jul 18, 2025|1,867 views|63|6

Save to Pod

Key Moments

On this page

TL;DR

Marimo: An open-source Python notebook reimagining data analysis with reactive execution, built-in UI, and AI integration.

Key Insights

Marimo is an open-source Python notebook built from scratch, offering a modern alternative to Jupyter with features like reactive execution and built-in UI elements.

The notebook prioritizes putting data front and center, enabling users to interact with data and AI models in novel ways through live previews and selections.

Marimo supports a transition from prototype to production, with notebooks stored as pure Python, versionable with Git, and executable as data apps, scripts, or pipelines.

The platform integrates AI-native utilities, allowing for code generation via natural language prompts directly within the notebook and offering rich context for AI assistance.

Marimo addresses reproducibility issues by allowing dependencies to be stored inline within the notebook file, leveraging tools like UV for package management.

The ecosystem includes Moab, a free cloud-hosted notebook environment offering a Colab-like experience with broader package compatibility and easier sharing capabilities.

INTRODUCTION TO MARIMO AND ITS CORE PHILOSOPHY

Marimo is presented as an open-source Python notebook developed entirely from scratch, aiming to address the limitations of traditional notebooks like Jupyter. Its core philosophy revolves around enhancing the data analysis and AI workflow by making data more accessible and interactive. Key differentiators include reactive execution, where the notebook intelligently determines which cells to re-run based on changes, and integrated UI elements that allow for direct user interaction within the notebook environment. This approach allows users to work with data and AI models in entirely new ways, truly placing their data at the forefront.

ORIGINS AND MOTIVATION FOR DEVELOPING MARIMO

The genesis of Marimo stems from the creator's experiences at Google Brain and during a PhD in machine learning research. While recognizing the invaluable interactivity that notebooks offer for data and AI work, many pain points with existing solutions like Jupyter were observed, including issues with hidden state, reproducibility, and version control. The rise of generative AI, particularly around late 2021 with ChatGPT, further highlighted the anticipated increase in data work and model evaluation. Marimo was conceived to retain the interactivity of notebooks while incorporating the robustness and guardrails of conventional software development, such as reproducibility and the ability to reuse notebooks as apps or scripts.

INTERACTIVE DATA EXPLORATION AND VISUALIZATION

Marimo facilitates a deeply interactive data exploration experience, exemplified by a demo using the MNIST dataset. Users can interact directly with visualizations, such as selecting points on a scatter plot of data embeddings, to see live previews of corresponding images. This reactive nature means changes in the UI or code automatically update outputs downstream. The underlying code for such interactions is often surprisingly simple, wrapping visualization objects in Marimo functions. This allows for seamless drill-down capabilities and data manipulation, which are cumbersome in traditional notebooks. Marimo's design allows for these interactive workflows to be built with minimal code, bringing data to life effectively.

FLEXIBLE LAYOUTS AND CUSTOM UI INTEGRATION

Moving beyond the traditional linear notebook structure, Marimo offers flexibility in how code and outputs are arranged, supporting multi-column layouts to better utilize wide screen real estate. Execution order is intelligently managed based on variable declarations, akin to a spreadsheet. A compelling demo showcased the integration of custom UI elements, like a Microsoft Paint widget, directly within the notebook. This custom drawing could then serve as input for a multimodal AI model, demonstrating how Marimo breaks the mold of static notebooks and enables complex, interactive workflows, including generating code like Mermaid diagrams from drawn input.

AI-NATIVE UTILITIES AND ENHANCED CODE GENERATION

Marimo incorporates AI-native capabilities to streamline code generation and interaction with language models. A 'Generate with AI' button allows users to prompt natural language requests, which Marimo translates into Python code. This feature provides context from existing data frames and variables in the notebook, leading to more relevant and accurate code suggestions. Unlike external tools, Marimo's AI assistance has direct access to the notebook's runtime environment, including data structures and potential database connections, offering richer completions. This accelerates rapid prototyping and analysis without leaving the notebook interface.

REPRODUCIBILITY AND DEPENDENCY MANAGEMENT

Addressing the critical issue of reproducibility, Marimo offers innovative solutions for dependency management. While traditional Jupyter environments often rely on external `requirements.txt` files, Marimo can store dependencies inline within the notebook header, compatible with package managers like UV. This feature allows a self-contained notebook file to be shared, carrying all its necessary dependencies. When a package is missing, Marimo can prompt the user to install it directly, simplifying setup. This approach significantly enhances the ability to reproduce exact environments and ensures that notebooks are more portable and reliable.

ADVANCED INTERACTIVITY WITH EXTERNAL DEVICES AND DATA ANNOTATION

Marimo's interactivity extends beyond the mouse and keyboard, allowing integration with external input devices. A demo featured using a PlayStation controller to interact with a data annotation workflow. This showcases the platform's extensibility, enabling custom applications within notebooks to be controlled by anything from a gamepad to other hardware. In the data annotation example, users could rapidly accept or reject paper summaries retrieved via vector search, with the controller augmenting the speed and reducing repetitive strain. This demonstrates Marimo's potential to transform tedious data tasks into more engaging and efficient processes.

STARTING FROM A BLANK CANVAS WITH AI ASSISTANCE

For users starting from scratch, Marimo offers an intuitive path to data analysis. An empty notebook provides a code cell with a 'Generate with AI' button. Users can populate data, for instance, using the Vega datasets library, and then leverage AI to generate code for analysis, such as creating histograms or calculating average statistics by category. By tagging data frames with an '@' symbol, Marimo passes column names and data types as context to the language model, facilitating more informed code generation. This eliminates the need to consult external documentation or switch between applications, enabling fluid, AI-assisted coding and analysis directly within Marimo.

MOAB: A CLOUD-HOSTED NOTEBOOK ENVIRONMENT

Recognizing the limitations of Web Assembly for certain Python packages, Marimo has developed Moab, a free, cloud-hosted notebook environment. This initiative is driven by user demand for a Colab-like experience that supports any Python package. Moab notebooks run on cloud instances, allowing users to import notebooks locally or from GitHub, select configurable CPU and RAM, and upload data stored in Cloudflare R2 buckets. This offering democratizes access to powerful interactive computing, providing a robust platform for sharing and running complex notebooks without local environment constraints.'Moab' is a direct response to user requests as the sixth most upvoted issue on Google Colab's GitHub page.

TRANSFORMING NOTEBOOKS INTO SCRIPTS AND DATA APPS

Marimo notebooks are stored as pure Python files with a `.py` extension, adorned with decorators. This design allows them to be executed not only as interactive notebooks but also as standalone scripts using commands like `uv run [notebook_name].py`. Furthermore, any Marimo notebook can be run as a read-only data app, similar to Streamlit, enabling deployment and sharing with non-technical users. While the data app mode is read-only by default, the underlying notebook can be configured to write back to databases or other services, showcasing Marimo's versatility in bridging the gap between interactive development and production deployment.

FUTURE DIRECTIONS AND COMMUNITY CONTRIBUTIONS

Looking ahead, Marimo aims to further enhance AI assistance, exploring concepts like agents that can generate and add code to notebooks based on prompts. Projects like 'Mimo Agents,' a fork by a Stanford PhD student, demonstrate this potential with specialized agent cells. The platform encourages community contributions and adoption. Users are invited to `pip install Mimo` and start with `mimo tutorial intro`, or try out the cloud-based Moab experience at `moab.marimo.com`. The project actively seeks contributors, underscoring its commitment to open-source development and the ongoing effort to redefine the notebook paradigm.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

●Organizations

●Concepts

Marimo Notebooks: Dos and Don'ts

Practical takeaways from this episode

Do This

Use Marimo for rapid prototyping of AI and data workflows.

Leverage reactive execution and built-in UI elements for interactive data exploration.

Version your notebooks using Git as they are stored as pure Python.

Share your notebooks as data apps or execute them as scripts/pipelines.

Integrate with your databases and data lakes using built-in SQL.

Utilize the 'Generate with AI' feature for code suggestions and completions.

Install packages easily using Marimo's integration with the UV package manager.

Experiment with custom UI extensions for unique interaction methods.

Try MoLab for a cloud-hosted notebook experience with full package support.

Run notebooks as standalone scripts or read-only data apps.

Avoid This

Avoid redefining variables across multiple cells, as Marimo uses a dependency graph.

Do not expect all Python packages to work in the WebAssembly playground environment.

Do not rely solely on traditional Jupyter notebook functionalities; explore Marimo's unique features.

Common Questions

Marimo is an open-source Python notebook built from scratch for AI and data work. It differs from Jupyter by offering reactive execution, built-in UI elements, and ensuring reproducibility by storing notebooks as pure Python files, allowing for versioning with Git and use as data apps or scripts.

Topics

Notebooks Interactive Computing Reactive Programming Reproducibility Data Apps WebAssembly

Mentioned in this video

softwareUV

A package manager from Astral that Marimo integrates with to manage dependencies directly within the notebook header, improving reproducibility.

softwareMarimo

An open-source notebook for Python built from scratch, designed for AI and data work, featuring reactive execution and built-in UI elements.

softwarePaperwork

A framework mentioned as being comparable to Marimo.

softwareModal

The cloud infrastructure where MoLab notebooks are currently running.

softwareMimo Agents

A project from a Stanford PhD student that forked Marimo and added agent cells for English-to-Python code generation.

softwareVega datasets

A source of datasets used in the demo for starting a Marimo notebook from scratch, specifically the 'cars' dataset.

softwarePolar

A data manipulation library that the speaker is learning, used within the Marimo notebook for data analysis.

softwareCloudflare R2

A storage service where user-uploaded data for MoLab notebooks is stored.

companyAstral

The company behind the UV package manager, integrated with Marimo.

conceptEMNIST dataset

A dataset of numerical digits used in the first demo to showcase Marimo's interactive scatter plot and data embedding capabilities.

softwareMoLab

A free, cloud-hosted notebook environment similar to Colab, built by the Marimo team, offering full package compatibility.

productPS5 controller

Used in a demo to showcase Marimo's interactive capabilities, demonstrating custom widget extensions that can be controlled by gamepads.

softwareJupyter

The traditional notebook environment that Marimo aims to replace or improve upon, highlighting its limitations in reproducibility and state management.

conceptradio

softwareArchive

bookTorch