Sponsored Session: Jupyter AI v3: Human + AI Collaboration

## Overview

Jupyter AI v3, developed by Amazon's AIML open source team, represents a major evolution of the Jupyter Lab extension that brings AI assistance to notebooks. The extension evolved from a hobby project launched in November 2022 (two weeks after ChatGPT's release) to a sophisticated collaboration tool built on three years of user feedback and open source community input.

## Key Topics & Evolution

The speaker outlined Jupyter AI's journey: v1 was published at Jupyter Con 2023 in Paris (January-May 2023), v2 emerged when Jupyter Lab 4 stabilized, and v3 represents a comprehensive architectural redesign. Critical user feedback identified three major pain points: ephemeral chats that disappeared on server restart, poor context management limited to copy-paste or RAG, and absence of agentic capabilities.

## Major Features in v3

**Persistent Collaborative Chats**: Chats are now stored as files with `.chat` extension and fully integrated with Jupyter's collaboration APIs, enabling multiplayer editing similar to Google Docs. Users can create multiple threads that persist across sessions and collaborate with both AI and human colleagues simultaneously.

**AI Personas Framework**: The solution to agent requirements, AI personas are humanlike agents available in every chat and invoked via @mention syntax (e.g., @jupyter-not, @claude-code). They retain conversation context and can be installed as packages. Jupyter Not is the default persona, designed to be conversational without excess verbosity. Multiple personas can coexist in a single chat, allowing domain-specific tuning rather than one agent with excessive tools.

**Enhanced Context Management**: Users can now drag-and-drop file attachments, use @file commands for automatic path completion, and select individual notebook cells for AI analysis—eliminating tedious copy-pasting. This approach leverages modern language models' support for 100,000+ token contexts, reducing the necessity for RAG (retrieval-augmented generation).

**Model Agnosticism**: Integration with LiteLLM provides support for over 1,000 language models without vendor lock-in. Users can configure custom API URLs, enabling deployment of proprietary models, proxies, or local instances.

## Architecture & Modularization

Jupyter AI v3 shifted from a monolithic structure to a meta-package architecture under the Jupyter AI contrib GitHub organization. Key subpackages include: - Jupyter AI Light LLM (model abstraction layer) - Jupyter AI Router (message routing) - Jupyter AI Persona Manager (persona discovery and lifecycle) - Jupyter AI Tools (real-time collaborative editing capabilities) - Jupyter AI Jupyter Not (default persona)

This modular design allows separately versioned APIs, enabling breaking changes without major version bumps and facilitating independent contributions. Contributors can focus on single packages rather than understanding the entire codebase.

## Developer Capabilities

Custom personas require only two methods: `defaults()` (defining name and system prompt) and `process_message()` (handling user input). Developers can use any Python framework—LangChain, LiteLLM, or custom agents. Hot-reloading via `/reload personas` command accelerates development iteration. The demo showed Claude Code integration as a remote persona installable via pip.

## Technical Achievements

The team significantly rewrote Jupyter Collaboration to enable real-time editing—agents type code into cells line-by-line as if another user were at the keyboard. The demo featured quantum circuit analysis (inspired by Amazon Bracket team's use case) where Jupyter Not explained quantum concepts and generated working simulation code.

## Future Vision

Planned enhancements include tool approval UI, MCP integration consideration, stable persona API, and ecosystem expansion. Envisioned personas span data science (leveraging HuggingFace's agent research), code review (separate model validation), and AI research (literature search tools). Beyond chatting, AI personas will proactively suggest optimizations, auto-fix errors, and manage background build commands—treating AI as true collaborators rather than tools.

## Governance & Community

The project operates under multi-body council governance with weekly community calls (Wednesdays, 8:30 AM Pacific). The small team of ~5 engineers, led by Brian Granger (Jupyter co-founder), manages strategic contributions to multiple open source projects including Jupyter Lab, LangChain, Dask, and MLflow.

## Takeaways

Jupyter AI v3 demonstrates how iterative user feedback drives architectural decisions in open source. The shift from ephemeral chats to persistent collaborative documents, introduction of flexible AI personas replacing monolithic agents, and modular design enable both users and developers to customize AI assistance. The vision extends beyond chatbots toward proactive, context-aware collaborators that augment research and development workflows.