Watch a Blizzard game director code a game using nothing but ChatGPT

What you need to know

  • A developer used ChatGPT to recreate the famous game "Flappy Bird."
  • ChatGPT wrote all of the code for the game, with the developer only adding prompts with natural language.
  • The result was a functional recreation of Flappy Bird without a single line of code written by a human.

OpenAI's GPT-4 is a powerful tool in the hands of creators. One developer wanted to test the capabilities of the Large Language Model (LLM) by seeing if it could code an entire video game. Wyatt Cheng, a game director at Blizzard Entertainment, shared a video on his personal YouTube channel of his experiment with ChatGPT and Flappy Bird.

Cheng used natural language to ask ChatGPT to help make the game. ChatGPT created a six-step plan to make Flappy Bird in Unity. Cheng then followed the steps, but none of the actual code was written by him. ChatGPT created every line of code and Cheng copied it over.

ChatGPT didn't act as a magic wand that made a game from scratch without some help. Cheng had to work in conjunction with the tool to troubleshoot issues, such as the bird falling down as a game started and the counter for passing pipes being timed incorrectly.

What ChatGPT did, was understand prompts written by Cheng, such as:

"The script works great for controlling the bird but the camera is not following the bird. I want the camera to follow the bird's horizontal movement."

ChatGPT then generated code matching Cheng's prompt.

It's not just Flappy Bird that developers have managed to make using ChatGPT. Ammaar Reshi, design manager at Brex, used GPT-4, MidJourney, and several other tools to create a 3-D space runner. The game is a recreation of SkyRoads, a title that was available on MS-DOS.

Reshi detailed his process in a Twitter thread:

The examples shared online are proof that you can make working games using ChatGPT, but Microsoft plans to take things further. GitHub Copilot in Visual Studio 2022 suggests lines of code and functions by using AI to understand the context of code. It can also make suggestions based on comments.

GitHub Copilot can also generate code by following prompts written in natural language, much like what ChatGPT did in Cheng's experiment.

Sean Endicott
News Writer and apps editor

Sean Endicott brings nearly a decade of experience covering Microsoft and Windows news to Windows Central. He joined our team in 2017 as an app reviewer and now heads up our day-to-day news coverage. If you have a news tip or an app to review, hit him up at sean.endicott@futurenet.com.