The Lars Language Model

The Lars Language Model was a product of a crazy fun night with my good friend Jason Haber. We both share a very good friend named Lars, whom we have previously crafted several simple games and videos for.

Jason and I invented a huge variety of new Lars-themed micro games and apps. Whiteboards were rapidly filled and erased. We found our way to our favorite place - “the zone”.

The Lars Language Model was one of the final ideas from the evening. It was easy to do a basic test with ChatGPT. Our initial experiments indicated that it could work.

I found a popular open-source next.js app to provide a UI for the Chat GPT API. I hacked it to my liking over a week. I introduced a “Larsifier” function that augments all prompts sent to the ChatGPT API with additional instructions of a “Lars” nature. The Larsifier function will change its instructions on each subsequent prompt.

I added hooks so that all conversations are persisted in a supabase DB. I implemented the “Hall of Lars”.

The project was the perfect distraction while I had COVID. I gave it a little beta test during Christmas with my son. After belly laughing until it hurt, I knew I was on to something. I tried it out a little more with friends during a New Year’s Eve party, validating that certain friends would find it incredibly entertaining, especially when they read the responses out loud.

My son warned me that most people won’t get it, because really, it has no point. For the few of you that it resonates with, it’ll hit you right in the Lars.