I have been coding with LLMs since GPT-3.5; using it initially for relatively simple data analysis tasks in Python. Then GPT-4 came along and its coding capabilities improved remarkably. Gone was the spaghetti code, bug-fixing loops became shorter and suddenly I was able to build data wrangling solutions in Google Apps Script and Python that dramatically improved my game design workflows.
Then came the o1 preview model last year, which changed the game again. It enabled me to develop an entire AI chatbot web app prototype for a client – from scratch to deployment – in just a few days.
And now I am using the o3 model in the ChatGPT app and the agent mode in VS Code to create tools for Unity – and it just works and solves my problems! Beyond that I am experimenting with agent frameworks to find new ways to interact with the vast datasets of our F2P games. This is the science-fiction stuff I was dreaming about 40 years ago creating and playing games on my C64.
I recently talked to Lina Sophie Knees of Handelsblatt about the opportunities and risks of “Vibe Coding”. The article is in German, but hey, who cares about language barriers these days?