Spreadsheets AI

This is a project that I did with Hanson Wang as part of the South Park Commons hackathon in December, 2022. The goal was to prototype an experience where spreadsheet users could leverage OpenAI's GPT-3 editing and completion capabilities to bring new functionality into a spreadsheet context. We won the overall grand prize!

  • Prototyping
  • Human-Model Interaction
  • Large Language Models
  • React UI
  • Design Challenge
  • While GPT-3 has been used in many contexts and for many use cases, it is at its core a text completion model. The fact that it can be used for code completion specifically, or for spreadsheet completion specifically, is a testament to the breadth of its training data. But this corner use cases will always expose fragilities in the model output. The challenge of this project was to massage the user input to a format that would be most likely to solicit a useful output from GPT-3. Then this output, which is just a text string, needed to be parsed and applied back into the structured spreadsheet context.
  • Project Scope
  • We had blocked off just 15-20 hours to complete the work, so we wanted to focus on demo-ability in our feature set. We tried to keep the UI pretty light and lean on the functionality that would play most to GPT's strengths. It turned out that GPT-3 is highly sensitive to data proximity in spreadsheets, so we focused on reformatting tasks that would have otherwise required a regex formua, and text summarization, which is impossible to do programatically in spreadsheets natively. We left behind any tasks that would reference data that was not close spatially to the task, eg. in another table or sheet.
Demo Video
Building autocomplete functionality into a spreadsheet context allows for faster workflows in one of the most common professional tools, and enables business workers to do work that previously required development expertise.

GPT-3, and other large language models, are reshaping the way that we think about interacting with technology. Used creatively, "just text" output can be leveraged for a wide variety of useful tasks.

Hackathon Results Blog Post
  • Target UserBusiness Intelligence Workers
  • Type of WorkProduct Prototyping, UI/UX Design, Front End Engineering
  • Tools UsedReact, GPT-3
  • LocationSan Francisco, CA