Solid Maldo

Inspiration

There was recently a big buzz online about GPT-3 and its capabilities, and all sorts of examples using it like https://philosopherai.com/. We all thought it was pretty exciting, and combining the idea of using a language model with a fun topic like esports pro twitter accounts was the icing on the cake.

What it does

The model we trained generates text in the style of tweets by (21) esports professionals. We curate the tweets we find the most interesting, humorous, or even impressive and schedule them for posting on the @SolidMaldo Twitter account. Of course, this also allows us to make sure no inappropriate tweets generated by the model makes it to the actual Twitter page.

How we built it

We used Google Colab as our development environment, as it offers free access to Google GPUs and has TensorFlow preinstalled- no fiddling with local requirements. We used TWINT to scrape Twitter data and gpt-2-simple as what we considered the most hassle-free way to train a GPT-2 model.

Challenges we ran into

We dealt with errors of all kinds initially, which slowed us down at the start. We had to learn the usage of two new libraries basically from scratch, so it took some time to read the documentation and understand how everything would flow together.

Accomplishments that we're proud of

As a group, we haven't worked on many projects in the past, so we're proud that we were able to work on something together and come out with a finished product. We're also proud that there weren't huge disagreements among the members about what to do or how to do it, and that communication was always smooth and friendly.

What we learned

We all took a big first step into working with deep learning models and got a lot of experience understanding how Google Colab and virtual machines work.

What's next for Solid Maldo

We hope to curate more tweets, make some people laugh, and that the experience will be valuable in our personal futures in data science, software development, etc.