SUMM

Introduction of Akshay Agrawal from Marimo and the context of notebooks and AI/data workflows
Marimo is an open-source Python notebook built from scratch (not based on Jupyter)
Focuses on reactive execution and built-in UI elements for new ways to interact with data and AI models
Notebooks are stored as pure Python, can be versioned with git, run as scripts/pipelines, and offer SQL/database connections
Marimo serves as a modern replacement for Jupyter, Streamlit, Gradio, and Papermill
Noted for rapid prototyping and smooth transition from prototype to production
Product sees strong traction: 300,000+ monthly downloads, more GitHub stars than Jupyter Notebook, and used at organizations like OpenAI, Hugging Face, Cloudflare, BlackRock, Stanford, and Berkeley

Inspiration from issues with Jupyter notebooks: hidden state, reproducibility, inability to version with git
Goal: keep interactivity but add software-like guardrails (reproducibility, no hidden state, code reuse as apps or scripts)
Emergence of generative AI increased the need for effective data/model interaction tools
Marimo designed to enforce best practices (e.g., no redefinition of variables, functional code structure)

Marimo notebook demo shown with Python code, markdown, and visual outputs (scatter plots with EMNIST dataset)
Unique interactivity: hiding code to present an app view, live selection in plots updates visual outputs downstream
Code remains simple; reactive model auto-updates outputs based on UI/code interactions
Different file format from Jupyter but supports conversion via command line utility (with some code restrictions auto-fixed in conversion)
Enables easy, interactive data exploration with less repetitive code compared to Jupyter

Unlike traditional linear notebooks, Marimo supports customizable layout (columns, flexible code arrangement)
Example demo: custom widget mimicking Microsoft Paint, outputting reactive drawing data
Integration with multimodal models shown (e.g., sending drawn graphs to Gemini, returning mermaid code), highlighting new workflow possibilities in notebooks

Built-in AI-native utilities for generating code and working with models
"Generate with AI" feature allows code completion and generation based on notebook context (variables, dataframes, etc.)
Dependencies can be managed inline in notebook headers via integration with UV package manager (from Astral), enabling reproducible notebooks with embedded dependencies
Smooth handling of missing packages: prompts for install, manages packages automatically

Data annotation workflow demoed, featuring custom widgets for input via PS5 controller
Uses UI and hardware integration for efficient, interactive tasks (e.g., classifying search results from vector DB using different embedding models)
Live updates and visualizations (e.g., beta distribution plots of model performance) boost feedback and productivity
Highlights Marimo’s extensibility (custom hardware/extension support)

Demo of starting from scratch: empty notebook with "Generate with AI" to help prototyping data analysis
User can prompt for transformations or visualizations using natural language, with model inferring context from data and variables
AI completions are context-aware: see variable states, column names, sample data, database connections, etc.
Multi-turn chat interface for more complex workflows, supporting various model providers via "bring your own key" (OpenAI, Olama, local models)

Marimo can run locally (pip install) or in the browser (WebAssembly) for lightweight, shareable notebooks
WebAssembly playground is browser-based but has some Python ecosystem limitations (e.g., cannot use Torch)
Restrictions in WebAssembly led to creating "MoLab," a cloud-hosted, Colab-like environment for running any package without such restrictions
MoLab supports importing notebooks from local files or GitHub, configurable CPU/RAM, and uploading data to Cloudflare buckets

Marimo notebooks are pure Python files, decorated for use as notebooks or executable scripts
Can run as scripts in terminal or as read-only data apps in browser (similar to Streamlit)
Potential to build full-featured, interactive data apps that can read/write to databases

Plans for deeper AI assistant integration (e.g., agent cells that accept prompts to generate code)
Encouragement of functional code style, prohibiting variable redefinition to improve code quality ("gentle parenting" of code habits)
Marimo project is open source with 100+ contributors, active engagement encouraged
Easy onboarding: pip install, try tutorials, or use the MoLab cloud service

Call for new users and contributors
Marimo seen as reinvention of a core workflow for research and engineering
Instructions to get started: pip install Marimo or visit MoLab for the cloud version
Users invited to engage on socials and contribute to the project

The Future of Notebooks - with Akshay Agrawal of Marimo