Skip to content

Latest commit

 

History

History
81 lines (57 loc) · 6.33 KB

README.md

File metadata and controls

81 lines (57 loc) · 6.33 KB
Raw

Depolarized App

Depolarized is a social application designed to encourage deep and balanced conversations on complex topics while leveraging modern LLMs to simplify (and eventually automate some of) the work of moderators and facilitators.

Quick overview

Depolarized.Small.mov

Conversation format

  • The conversations have two sides and are moderated by admin users (aka facilitators)
  • The facilitators can use LLMs in a custom-built AI playground (pre-loaded with the conversation's history) to prepare helpful feedback for contributors
  • Contributors from each side are encouraged to collaborate on producing strong arguments while demonstrating openness, curiosity and respect for the other side
  • Users can suggest specific changes, which are shown to the author as a "diff" in the UI, and the author can then accept or reject the changes
  • Users can also vote, rank and comment on arguments
  • The facilitator eventually decides manually the winning argument for each round (although, the plan would have been to automate this)

See also ./SCREENSHOTS.md for screenshots.

Examples of AI features

The AI playground can be accessed by any admin user on desktop by clicking the "AI Playground" button on the top right. This will open a side panel with some buttons corresponding to different prompts. Below is a short description of some of the prompts that we have found useful during beta testing.

  • Complexity - look for sources of complexity in the last argument
  • Completeness - check that it answers all the points made by the previous author
  • Ratings - rate the argument (is it strong, concise, respectful and factual)
  • Strong words - look for expressions or phrases which may be exaggerations
  • Platitudes - look for expressions or phrases which are too vague or generic
  • Tone & Style - how does the author "sound" (e.g. dismissive, aggressive, relax...)
  • Add references - suggest links and references to back up the claims made in the argument
  • Congratulate Win - prepare a message to congratulate the author for winning the round
  • Loozer's Prize - explain politely to the author that they didn't win this time
  • Recap - short summary of what different authors have argued so far
  • Common Ground - find some points of agreement between the last two authors
  • Next direction - propose 2-3 possible directions for the rest of the conversation
  • Next argument (X) - what might the next contributor write as a response (in style X)
  • Speaker face - (just to play with DALL-E) guess what the author might looks like

You can see and extend the list of prompts in ./src/ai/prompts.

About the code

Depolarized was built as a progressive web app with TypeScript, React, Tailwind, Firebase and ChatGPT.

Disclaimer

You're looking at a prototype, not a production-ready app. It was built in a couple of weeks in Feb 2023 to test some product ideas. Most features seemed to work during beta testing, but you should expect plenty of bugs and rough edges.

How to run

  1. Set up a Firebase project with Firestore, Functions, Phone Auth and a Blaze plan
  2. Set up an OpenAI account for ChatGPT (warning: your key is used in the front-end, so you'll want to set some strict usage limits)
  3. Look at ./src/secrets.ts to find the list of required variables and create a .env.local file to populate them
  4. Run npm start to start the app locally or firebase deploy to deploy to production
  5. Sign up with a phone number using any 6-digit code as invite code
  6. In the Firestore UI, look at the new document in the users collection and manually set the isAdmin field to true
  7. As an admin user you'll then be able to initiate conversations, moderate conversations and use the AI playground.

Technical pointers

If you came here for the code (rather than the product or the AI prompts), here are some bits you might find interesting:

  • The UI state is managed in ./src/global using Redux-like patterns (synchronous global state and actions) but without any heavy framework, and only using immer to minimize boilerplate.
  • The component ./src/components/InfoModal.tsx made it super easy to document everything in the app with smart info bubbles and carousels.
  • The UI components in ./src/components are all built with Tailwind and are fully responsive. One new trick I learnt was to extend existing React prop types to (e.g. extends React.ComponentPropsWithoutRef<"button">) to make sure that the components are fully compatible with the original HTML elements.
  • The app is aware of its current bundle version and will automatically invite user refresh the page when a new version is deployed. The magic for that is in ./src/refresh.tsx.
  • The home page (./src/pages/home) responds nicely to different screen sizes, using two separate screens on mobile and a single-screen layout with two columns on larger screens (similar to WhatsApp's desktop app).
  • Some notifications are sent to users (inside the app) when they receive feedback from others (e.g. comments or suggested edits). The code is in ./src/pages/notifications.
  • All the reviews (ratings, comments, edits) made on an argument are aggregated by the aggregateReviews function in ./functions/src/index.ts. This one took a while because debugging Firebase Functions is still painfully slow, even when using a local test environment.

Known issues

Many things could have been done better. Just to list a few:

  • The extra abstraction in the file ./src/firebase.ts was probably unnecessary. It would have been best to just use the firebase libraries directly.
  • It would have also been better to add a className prop to Tailwind components like Button to make them more flexible (instead of introducing many variations like ButtonSmall, ButtonSmallGrey,...).
  • The Firebase rules in ./firestore.rules were written in a rush but they should be improved to prevent users from accessing data they shouldn't be able to access.

Credits

This work was done in collaboration with Andrew Lewis-Pye, who made major contributions to the product's designs and recruited many of the beta testers, who in turn contributed to the project with their feedback and ideas.