Using Generative AI in the Rush Gaming Universe
By Kush Khurana — AI/ML Team, Hike
Today on the Rush Gaming Universe, we have 14 incredibly thrilling casual games of skill. We’re at a point now where all future games on the platform will be built through AI. There are two key aspects to this:
- AI-generated Art and
- AI-generated Code
1. AI-generated Art
For the past couple of years, we have been investing heavily in leveraging AI to accelerate the creation of artwork, in particular game art. From 2D game themes and assets to 3D Rush Avatar with AI-driven personalization and animation flows, we have developed a gang of AI models to enhance and accelerate almost everything that an artist does at Hike. This has not only helped us scale quickly via TAT improvements for artwork creation but has given the artists more room for divergent thinking, iterations & experimentation, eventually resulting in improved variety at similar or better quality.
A prime example of this is the Rush Avatar itself, where once we locked in on the style and the base 3D model, we could scale things super quick by leveraging AI to customize and animate the Avatar to suit various use cases across game production and marketing. Since the available AI models had no clue about the unique handcrafted style of Rush Avatar, we first developed a fine-tuned AI model that “understood” the Rush Avatar’s art style and then designed an automation pipeline around it to enable quick creation of customized Avatars, which has now become an extension of our Avatar production team. On the animation front, we fine-tuned state-of-the-art AI models for keypoints detection and pose models to generate Avatar animations from videos. This has helped us to quickly capture interesting and trending moves (like in the video below) directly onto our 3D Avatar, making for interesting applications across gaming and marketing.
2. AI-generated Code
Coming to AI-driven code generation, GPT-4 has been a turning point. Combined with our internal Generative AI models, we’re now in a place where we can produce high-quality casual games in less than a day. Here are a couple of projects to demonstrate:
Project 1: Replicate Flappy Bird
- Time taken → 90 minutes.
- Code accuracy → ~85% of Code was generated perfectly in one shot. ~15% of code correction took 20 minutes to get right.
- Approach → This was simpler than we thought actually, as GPT-4 seemed to have a prior understanding of flappy bird’s game mechanics. Simply prompting GPT-4 for exemplary architecture and C# scripts worked for us. Iterating a couple of times on correcting the code and adjusting for a few parameters like speed, jump force, and spawn rate got us to the desired gameplay experience.
Project 2: Rush Endless Runner
- Time taken → 12 hours.
- Code accuracy → Eventually, 100% of code generated was grammatically correct and directly useable. However, we did need some ingenuity with prompt engineering to get to the code.
- Approach → First we generated the process to build the game and then iterated to generate code for each piece. We figured out that simple prompting resulted in code that was way off in this case. So, instead of correcting the code, we sought to engineer the prompts in a way that gives near perfect code in one shot. Component by component, we narrated our chain-of-thought precisely to GPT-4, and eventually got 100% correct code which we could directly integrate into Unity to test out the functionality.
Exciting Future
Fast forward 6–12 months and we’re going to be able to generate more complex gameplay as our models get better. This opens up so many incredible possibilities.
If creation is dramatically becoming easier, then should we be the only creators on our platform? What if the community could create and be incentivised to do so? Think AI + Crypto?
If you’re excited to help us build this vision; join us — we’re hiring!
