Introduce students to the fascinating challenge of how Large Language Models (LLMs) interact with chess data. Explain that LLMs are designed to process human language (tokens/words), not inherently structured game data.

The Translation Layer

The key teaching point is that before an LLM can analyze a game or suggest a move, the structured chess format (like FEN or our custom 77x8x8 array) must be converted back into a sequential, text-based format that the LLM understands. This is often achieved by translating the board state into a string of tokens.

Data Diversity in AI Architectures

For example, a FEN string is a perfect, concise text input, but even our numerical 77x8x8 representation can be linearized and fed to a model. Students learn that their initial complex data engineering work—converting moves to 0-4096 indices and positions to a 77-layer tensor—is crucial for the AI training model, but the separate LLM might require a simpler text prompt to function.

This illustrates the diverse data needs of different AI architectures and shows how:

• The output of one AI component (the trained chess model) might become the input for another (the LLM used for analysis or chat).