Notebooks: Tools for LLMs

Lightning talk about Notebooks for LLMs at AI Dev Day, hosted by AICamp.

This talk shows how we integrated computational notebooks, specifically Jupyter, with Large Language Models (LLMs). This talk will explore how the Noteable Plugin has been designed to enable LLMs to write literate computational notebooks, allowing models to write prose, code, and plot. An unexpected discovery, once GPT models had access to notebooks, they unlocked the ability to debug and reason in a paired interactive computing environment. Through some live demos and videos, we will walk through doing data science with our LLM partner. We will also discuss the tradeoffs and risks associated with temporary and durable computation for chat models. This session aims to inspire the audience to explore the untapped potential of LLMs and notebooks in their data analysis tasks, paving the way for a future of collaborative analysis that includes not just human peers but also artificial intelligence.

Transcript

1. Kyle Kelley at AI Dev Day, July 25th, 2023
   Notebooks: Tools for LLMs
   ⚡ Talk at AI Dev Day
   

   View Slide

2. I'm Kyle Kelley
   • Chief Architect at Noteable
   
   • IPython & Jupyter Developer
   
   • Previously Notebooks Platform
   at Net
   fl
   ix
   
   • Containers, DevOps, and UX
   Hi!
   

   View Slide

3. What are Notebooks
   Play
   
   Build
   
   Share
   

   View Slide

4. View Slide

5. In[1]
   :
   import openai
   

   import noteable
   

   View Slide

6. https://www.youtube.com/watch?v=Xd-VPCC_WKI&t=584s
   Choose your own adventure
   Video from I versus AI Live Demo!
   

   View Slide

7. Make literate computational documents
   Data Driven Documents
   

   View Slide

8. How?
   

   View Slide

9. Headless Execution
   From Scheduled Jobs to Realtime Notebooks
   • Original scheduled notebook jobs:
   
   • Interactive compute with realtime updates
   
   • Both interactive and batched
   

   View Slide

10. Headless Execution
    Close your tabs
    

    View Slide

11. View Slide

12. Comparisons
    Noteable Plugin, OpenAI Code Interpreter, and Jupyter Extensions
    • Ephemeral compute: Fast to spin up
    
    • Durable Compute: access to data
    fi
    les, environment variables/secrets, and
    databases
    
    • In all cases, ability to debug, reason about outputs, and run code
    

    View Slide

13. From: Please use the IMDb Data Source
    with the `imdb` schema to determine
    which years had the most horror
    fi
    lms.
    To: Which years had the most horror
    fi
    lms?
    

    View Slide

14. Conclusion
    Untapped Potential of LLMs and Notebooks for Data Analysis
    • Longer form blog post of this talk:

    

    platform.noteable.io/blog/notebook-tools-for-llms
    
    • Augment Notebooks with app.noteable.io
    
    • Try the Plugin with ChatGPT Plus!
    

    View Slide