Roar. Bears, we've been hearing a lot about AI recently, and bears have been using AI for many kinds of operations. But a lot of us are still watching the situation and have never even touched it, let alone being actively involved in its development. So starting today, I would like to invite ideas about using AI for our games.
Now, before I go further, I must stress one important point. Deciding on the tool to use before knowing the project goal is generally not a good idea. One should always start with the project objectives and the expected outcome. The reality is, however, because we don't know AI's capabilities, we don't consider them when making expecations about our projects. This is precisely what this exercise is trying to address. Once we have an idea what AI can do, it becomes our experience, and we will natually consider that when we plan our new projects.
A few cubs have been making things - sprites, dialogue or even sound for their games, using generative AI like Copilot or Gemini. This is one use of AI, but there are many more. One website for learning AI is HuggingFace. On their model search page, there's a section called "task". Just by looking at this list, you know how much there is for us to explore. The use case we mentioned earlier was just a few items on this list ending with the word "generation".
Let me warn you though, once you leave the world of making wishes to AI using text, it's a completely different feeling. As a bear who's created an AI chat app and also taken courses on deep learning, I can tell that. It is also a different feeling when you call the model's library, called an SDK, from when you type into a text box. It is yet a different feeling when you directly feed data to the model without an SDK, or making the model part of your bigger model. And when you design the model architecture, that is, spelling out layers yourselves, the feeling is different again.
Another point I want to stress is that, do not think that we're doing something ahead of our times. In fact, we're almost the last to the party, if not already left behind. However, because of this, we've also got an advantage. The documentation of AI libraries has become significantly better than they were back in 2019. More models, more data sets, more ways to collect or generate data, and more examples are available. Just look at Scikit-Learn and Keras example galleries. We can get started much quicker than we had to a few years back. In fact, also on HuggingFace, the website I mentioned earlier, there's even a tutorial specifically about training a reinforcement learning model in games. It uses the Godot engine, not Pygame that we use, but we can learn how the AI model interacts with the engine and port things accordingly. There is a set of example environments that we can learn how to set this up from. Besides, from a software architecture perspective, model training does not depend on things like how things are drawn to the screen.
I will be reading more on AI and will keep you posted when I find something. In the meantime, if you have any idea later on, please feel free to tell me.
Gabe the Bear