Deepnote CEO: Why Notebooks Are the 'Perfect User Interface' for AI Agents

Key Concepts

• Data Notebooks: Computational mediums designed for data exploration, analysis, and collaboration.
• Low Barrier to Entry vs. Low Ceiling: The trade-off between tools that are easy to start with but limited in functionality (e.g., spreadsheets) and tools that are powerful but have a steep learning curve (e.g., code editors).
• Literate Programming: A programming paradigm that emphasizes embedding code within documentation to create a narrative explanation of the program.
• Operational Transformation (OT) vs. Conflict-free Replicated Data Types (CRDTs): Two different approaches to implementing real-time collaborative editing. OT was chosen by DeepNote for its stability and similarity to Google Docs.
• Stateful vs. Stateless Systems: Stateful systems maintain information about past interactions, while stateless systems treat each interaction independently. DeepNote transitioned to a stateless architecture for improved collaboration.
• IPython Notebook (IPYNB): The original open standard format for Jupyter notebooks, which DeepNote initially built upon but found limitations with for its evolving features.
• Blocks (in DeepNote): The fundamental units of content in DeepNote notebooks, replacing the traditional "cells" of code and markdown found in Jupyter.
• AI Agents: Artificial intelligence entities that can interact with users and assist in tasks, with notebooks seen as an ideal interface for this collaboration.
• Productionizing Workflows: The process of taking a data exploration or analysis and making it a production-ready application or service.

DeepNote's Journey: From Data Notebooks to Open Source

The Genesis of DeepNote: Bridging Computational Worlds

DeepNote, founded in 2019 by Yakob Yurik, emerged from a perceived need for a new kind of computational medium. The founders, with a background in developer tools, recognized a gap between simple, user-friendly tools with limited capabilities (like spreadsheets, described as having a "low barrier to entry, but also it has like fairly low ceiling") and advanced, powerful tools like code editors and IDEs that offer "the sky is the limit" but require significant technical expertise to use. This gap highlighted the absence of a computational medium that could effectively bridge these two worlds.

The concept of notebooks, with roots tracing back to the 1980s and early AI research, was identified as a potential solution. While early iterations like Mathematica existed, the advent of Jupyter is seen as a pivotal moment, signaling that the time for notebooks had arrived. DeepNote was conceived as a reimagining of the notebook format, initially focusing on collaboration and later evolving to accommodate the "AI era" where collaboration extends to AI agents.

Overcoming Philosophical and Technical Challenges

Philosophical Challenges:

A significant hurdle for DeepNote was the prevailing sentiment in 2019 regarding collaboration in software engineering. There was a strong inclination towards asynchronous collaboration, exemplified by pull requests and code reviews on platforms like GitHub. This contrasted sharply with the workflow of data analysts and data scientists, who often engage in exploratory tasks with less defined end states and require much tighter, real-time feedback loops.

Yurik explained, "When you are a data analyst, data scientist, you want to do some data exploration. The task is more like well here is a CSV file, here is just bunch of data and go ahead and try to find something interesting." This often involves immediate questions like "why is that?" when encountering issues like null values in a database, necessitating rapid, synchronous interaction. The traditional software engineering model of "every single chunk of work that I do has a nice commit message" was not suitable for this iterative, often uncertain, data exploration process. DeepNote's initial challenge was to convince users that a tool built with collaboration in mind was not only beneficial but essential for their workflows.

Technical Challenges:

Once the philosophical groundwork was laid, technical challenges emerged, particularly around implementing real-time collaboration. The choice between Operational Transformation (OT) and Conflict-free Replicated Data Types (CRDTs) for collaboration algorithms was a key decision. DeepNote opted for OT in 2019, citing its stability and its use in established platforms like Google Docs, as CRDTs were not as mature at the time.

This decision necessitated building significant infrastructure to coordinate multiple clients through a server. The process involved several iterations and the development of extensions for editors to handle the diffing and patching of code correctly, including overcoming numerous "line endings bugs." Over time, DeepNote refined its architecture to a "fully stateless" system that supports "very nice and intuitive collaboration." The scalability of this system was demonstrated by early instances of hundreds, and now routinely thousands, of users collaborating within a single notebook.

The Decision to Go Open Source

The decision to make DeepNote open source was a long-held aspiration, stemming from the founders' own open-source background. However, the immediate priority was to address critical user pain points such as stability, reproducibility, and collaboration. Initially, the team was also uncertain about the long-term viability of their chosen architecture, and keeping development proprietary allowed for flexibility and iterations without the commitment to backward compatibility that open-sourcing would entail.

After several years of development and overcoming these initial challenges, DeepNote has reached a point of architectural confidence. They believe they have a strong understanding of the constraints and needs for notebooks in the coming decade. This confidence led to the decision to establish an "open standard" that they are committed to supporting long-term, with future revisions expected to be evolutionary rather than revolutionary.

Business Model and the Evolution of Notebook Formats

The sustainability of open-source projects is a significant concern. Yurik emphasized that "open source is tough" and requires a "meaningful business model" to survive. Without it, open-source projects can stagnate, falling behind proprietary alternatives that can adapt more quickly to market demands.

DeepNote's journey also highlighted the limitations of existing notebook formats, specifically the IPython Notebook (IPYNB). While acknowledging its "surprisingly very well designed" nature from its inception in 2011, DeepNote found itself "abusing" the format, particularly the metadata field, to accommodate its expanding features. The fundamental difference in DeepNote's "blocks" (which can be of 23 different types and are expected to grow) compared to Jupyter's two cell types (code and markdown) made mapping them to the original IPYNB structure increasingly difficult and acrobatic.

This realization, coupled with the fact that IPYNB was designed "pre-cloud, pre-collaboration, pre-AI," solidified the need for a new format better suited for the modern technological landscape.

The Future of Notebooks: Beyond Data Exploration

Yurik expressed a deep affection for the notebook category, acknowledging the influence of Jupyter and other interesting experiments like reactive kernels. However, he sees the notebook's ultimate potential as the "perfect user interface for working alongside AI agents."

While chat interfaces have shown promise in interacting with AI, they often fall short when it comes to producing tangible artifacts. Notebooks, with their inherent structure of input and output, and their collaborative capabilities, are ideally positioned to bridge this gap.

The vision for notebooks extends beyond their current perception as mere "scratch pads" for data exploration. DeepNote aims for notebooks to be a comprehensive environment that can host workflows all the way to production. This includes starting with quick, messy queries, building simple pipelines, and ultimately having the notebook itself become a "whole data app" that can be scheduled, have an API endpoint, and be called externally. This evolution from a scratch pad to a full production medium is seen as the future of notebooks.