In the summer of 2020, when we were all cooped up in our homes, desperate to find a new hobby, my friend Jack and I decided the best way to spend our time was to create a Twitter bot (one of the fun ones). With all the uncertainty and chaos unfolding in the world around us, we knew there was only one character we could count on to liven our moods:

You to listen to me, friend, and i still get zits. I think it's going to work every day.

— Michael Scott (@MichaelGaryBot) March 12, 2022

That's right, we based our account off Michael Scott, regional manager of the Scranton branch of Dunder Mifflin in The Office, but of course you already knew that. The account tweets thrice per day (exactly once every 8 hours), and some of the things it says are, well, strange. That's because we used a very simple n-grams, Markov chain-based model that randomly walks across sentences, guessing the next word based off randomness, but giving more likely successive words a better chance of being said. It's not as sophisticated as many modern text synthesis models out there, but it made for a fun mini project in the beginning of my coding journey. You can read more about our approach and implementation in our GitHub repository.