AvantGarble - Dadaist chatbot

---

Chat with the bot here.

Context

In the early 20th century, a group of Italian and Swiss artist revolted by the horrors of WWI started making art by rejecting traditional art values, such as rationality, beauty and logic. They explored randomness and absurdity and pave the way to conceptual and surreal art. Many artists explored with incorporating randomness in the generation of their art.

One of the triumph of the current times is the development of smart chatbots, the most famous of them being ChatGPT. These chatbots use the latest advances of Artificial Intelligence (AI) to produce careful responses to any question from the user. Dada artist that often displayed readymade pieces of art taking daily objects and putting them out of context in an art gallery would probably love the tools we have at our disposal.

This project is a rejection to the smartness and rational thinking of modern chatbots by producing random (but English like) response to the users input. The work democratizes the access to Dadaist art by giving a short dadaist text. It also makes the viewer participate in the creation of the art by requiring its input.

In a Dadaist move the name "AvantGarble" was chosen by ChatGPT.

History of the project

I being working on this idea for since Jan 2024. I made two previous versions.

The first proof of concept was client side only and used a dictionary of English words to build the model. The language model was a 3 character trigrams taken directly from the dictionary. The generate text was usually a gibberish of letters that probably pass a letter frequency test for English but not much more. Also spaces in the generated text were rare, so the output often had long words. This was because a space was added to the end of each trigram but the model hat no idea of its actual distribution in English text. The code and live demo are available here

The second version had a much nicer interface developed in React and the language model was still a character based, but the trigrams now came from an actual English corpus. The trigrams are stored as separate .json files in GitHub. This meant that to generate each token (one character) a full request was needed. I tried to store the model in Google Cloud Object Storage and serve from there, but it was too expensive. The code is available here. The generated text sometimes contained real words but was mostly gibberish. This time spaces were distributed more like in English text and if you were just looking at the text without reading it felt much more English, similar to Lorem Ipsum.

This third incarnation uses word level trigrams generated from the same corpus as the second version. These makes the generated text more like a random connection of English words. The model behaves similarly to the prediction text from your phone's keyboard. The generated text is still mostly nonsense, but it contains real words and sometimes short common phrases. Also, the responses are much faster because the model is stored locally in a Unix database.

Build instructions and tech stack

Install locally

First build the front end:

cd client
npm install
npm run build

The back end runs on Python using FastAPI. To manage dependencies the project uses uv, To run activate the environment and then run fastapi run .\main.py

The front end is writing in React with Typescript and it is located in client. The backend uses FastAPI to serve the client and the answers. The answers are sent word by word to the client using websockets.

The answers are generated by a n-gram language model. The n-grams come from the Blog Authorship Corpus and are stored on a Unix database.

The live version is hosted on Google Cloud Run. To deploy follow these instructions:

1. Build and push the docker image:

docker build -t avant-garble-python .
docker push europe-west2-docker.pkg.dev/avant-garble/avant-garble/avant-garble

2. On the Google Cloud Console, go to Cloud Run and deploy a new revision of the service avant-garble-python using the new image.