Seedy

Headline and associated image.
Another headline and another associated image.

Inspiration

Seedy grew from an idea to train an AI model to discern between real vs generated images and real vs generated text. This led us to considering "real" fake text - in other words, human-written fake news, like the well-known satirical news site, The Onion. Soon, Seedy sprouted into a full-fledged idea to mimic fake news articles... to grow Onion articles.

What it does

Seedy utilizes a trained model to generate a new headline, which is then fed into OpenAI API, and DALL-E to generate an article and photograph. A user can repeatedly request new articles to be generated.

How we built it

The model used to generate headlines is from nanogpt. Using a large, already trained gpt-2 model, we fine-tuned the model with a training set of over 12,000 fake-news headlines from The Onion, Babylon Bee, and Not The Onion Reddit pages. We then hyper-tuned the model training and sampling parameters in order to get the most consistent output.

We created an API that is run on a local server. This API creates an instance of the model and can generate headlines, as well as interact with DALL-E and OpenAI API to generate the other needed items, using the headline as input.

The webpage then interacts with this API to receive and format these into a news page.

Challenges we ran into

Originally, we ran into challenges with the model learning to swear. After preventing it from that, it learned how to censor words with "*" in order to get around our blocks. It also was producing some very problematic content. We fixed this by cleaning (figuratively and literally) the training set and reducing the randomness (temperature) of the sampling.

We also ran into challenges with the webpage - API interaction. UPDATE LATER

Accomplishments that we're proud of

We are proud of the website design and front-end that none of the team had significant experience in. We are also proud of training a somewhat funny model, with somewhat consistent results. Finally, we are proud of the architecture and designing the interactions between the different aspects. Also proud of the animations

What's next for Seedy

Seedy has been planted and with some careful gardening and attention, it can flower and blossom into a field of endless onions.

Realistically, future features include more specific prompts, continued web-dev and visual improvements, and improving the model with more data and/or more tuning.