My first attempt at making a game with AI was tower defense, as it seemed relatively simple.
After confirming that it was running to some extent, I started creating autobots because I found it tedious to keep adjusting the balance while playing.
I instructed the autobots to jump right in and learn the game while finding the balance. However, since this wasn't a macro, they had no prior knowledge and needed to learn from scratch. So I started with Pong and trained them on various games.
I thought it would be productive to work on something else as well, so I decided to customize a model that generates stock reports. I downloaded several local models and tried using them, but they needed fine-tuning with high-quality data. I also needed to teach the model how to read stock prices and understand the correlation between news, stock prices, and events...
I chose the phi4 14b model as the main model for training, using approximately 12,000 data points. The problem was that this model had a context of 16k, while there was also a 32k model. However, Korean characters were corrupted in the 32k model. So I searched for another option and found qwen3 14b with a context of 128k. I fine-tuned it again, but the quality was lower than phi4.
I realized that I needed to use a larger quantity of high-quality data to compensate for the lower quality. While phi4 was trained on 12,000 data points, I aimed for 22,000 for qwen3 and am currently working on data refinement.
The autobots continue to play the game, but the problem is that I need to constantly fine-tune the rewards. I tried to automate this process with AI as well, but local models couldn't handle it at the moment. Fine-tuning with a cloud model worked well, so the conclusion is that since both the autobots and local models are based on AI, I can train the local models on reward data for a while. If they learn in a good direction, I can then use that data to train the autobots and create an all-in-one product.
I'm entrusting everything to my AI agents.
It seems like it will take a long time.
