Developer Builds Cyberpunk Card Game to Teach LLM Alignment Concepts

A developer has created a Gwent-inspired cyberpunk card battle game designed to visualize how large language models are fine-tuned and aligned. Players take on the role of an alignment engineer, competing against a baseline AI across three training rounds covering coding, language, and safety benchmarks. Core ML techniques such as Supervised Fine-Tuning, Direct Preference Optimization, and Model Pruning are mapped directly to in-game card mechanics. For example, playing a Model Pruning card triggers an animation that removes the lowest-value card from a row, mirroring how pruning strips low-impact weights to compress real models by 30–50%. The project also simulates risks like Catastrophic Forgetting, where fine-tuning on a narrow dataset can cause a model to lose previously learned general knowledge.